text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'Recently generative models have focused on combining the advantages of variational autoencoders (VAE) and generative adversarial networks (GAN) for good reconstruction and generative abilities. In this work we introduce a novel hybrid architecture, Implicit Discriminator in Variational Autoencoder (IDVAE), that combines a VAE and a GAN, which does not need an explicit discriminator network. The fundamental premise of the IDVAE architecture is that the encoder of a VAE and the discriminator of a GAN utilize common features and therefore can be trained as a shared network, while the decoder of the VAE and the generator of the GAN can be combined to learn a single network. This results in a simple two-tier architecture that has the properties of both a VAE and a GAN. The qualitative and quantitative experiments on real-world benchmark datasets demonstrates that IDVAE perform better than the state of the art hybrid approaches. We experimentally validate that IDVAE can be easily extended to work in a conditional setting and demonstrate its performance on complex datasets.'
author:
- |
Prateek Munjal[^1]\
Indian Institute of Technology\
Ropar\
[[email protected]]{}
- |
Akanksha Paul\
Indian Institute of Technology\
Ropar\
[[email protected]]{}
- |
Naraynan C Krishnan\
Indian Institute of Technology\
Ropar\
[[email protected]]{}
bibliography:
- 'ms.bib'
title: Implicit Discriminator in Variational Autoencoder
---
Introduction
============
Deep Variational Autoencoders(VAE[@VAE_kingma]) and Generative Adversarial Networks(GAN[@Goodfellow:2014:GAN:2969033.2969125]) are two recently used approaches in the generative modeling world. VAE is more stable in training but generates blurry samples. While GAN has the appealing property of generating realistic images; training a GAN is well known to be challenging leading to problems such as mode collapse.
Several recent approaches have proposed hybrid models of autoencoder and adversarial networks with a joint objective of achieving stable training like VAE and inferencing ability like GAN. In order to introduce the adversarial loss component in the objective functions most of the recent hybrid approaches include an adversary network that results in a three-tier architecture an encoder, a decoder, and an adversary network. We hypothesize that the encoder and discriminator networks can share common layers encoder itself can be reused as a discriminator, thereby assuming an overlap in the knowledge learned by the encoder and the discriminator network.
![Flow diagram of traditional hybrid approaches(left) and our proposed approach(right). We introduce the adversarial loss by collapsing the encoder into the discriminator, which we term as Implicit Discriminator. The output of discriminator is denoted by $y \in [0,1]$ where $0$ and ${1}$ represent fake and real respectively.[]{data-label="fig:approach_caption"}](images/approach.png){width="\linewidth"}
\[fig:approach\_onecol\]
The key idea behind our approach is that we would like the discriminator to provide the useful gradients to the generator if it misses any mode in the true data distribution. The traditional GAN learning does not explicitly encourage such a property in the discriminator and therefore, we suspect that it is vulnerable to the issue of mode collapse. Further, we note that the traditional L2 loss used for learning the encoder can be seen as minimizing the forward-KL divergence. The forward-KL divergence comes with the mode inclusive property (it never misses a mode in the true data distribution). Therefore, to make the discriminator aware of all the modes in the true data we propose to share the forward-KL information with the discriminator by explicitly sharing the parameters between the encoder and the discriminator network. We restrict the sharing of the parameters between the encoder and discriminator networks until the penultimate layer to facilitate the modelling of the different outputs. Figure \[fig:approach\_onecol\] illustrates our proposed two-tier architecture and contrasts it against the traditional hybrid (V)AE/GAN approaches. We propose an adversary free two-tier architecture an encoder and decoder network that has the capabilities of both a discriminator and a generator. In the proposed model the decoder network collapses into the generator, while the encoder network is merged with the discriminator resulting in a two tier architecture. We term our two tier architecture as Implicit Discriminator in Variational Autoencoder (IDVAE). VAE-GAN is a special case of IDVAE where there is no sharing of parameters between the encoder and the discriminator.
In this work we show that our proposed simpler hybrid VAE/GAN model, IDVAE outperforms the prior approaches in terms of visual fidelity measured in terms of FID score. We also show that our model sustains both reconstruction and generation ability without training an unnecessary adversary network that would result in learning more parameters. Overall the main contributions of the work are as follows :
- We introduce a novel two-tier architecture, IDVAE, which sustains the abilities of both reconstruction (VAE) and generation (GAN) without learning a separate discriminator or a generator.
- We present a training schedule that facilitates the encoder to act as as implicit discriminator while maintaining the tight coupling between encoder and decoder network.
- Empirical evaluations of IDVAE performed on benchmark datasets show that IDVAE achieves better generative ability than prior approaches. We also show that Fréchet Inception Distance, a common measure to evaluate the quality of the generations has inconsistent outputs and thereby propose an ensemble of experts for conducting quantitative evaluation.
Related Work
============
Variational Autoencoder (VAE) introduced by Kingma et al.[@VAE_kingma] minimizes the KL divergence between the real distribution ($P_x$) and the generated distribution ($P_g$) through the variational bound. Detailed analysis of VAE by Doersch[@vae_tutorial_doersch2016tutorial] shows that VAE works well in practice and is considered to model the true data distribution quite well but often generates poor quality samples the images produced by the decoder are blurred. On the other hand GAN’s [@Goodfellow:2014:GAN:2969033.2969125] generate samples that are visually more realistic through an adversarial game play between the generator and the discriminator. However, GAN’s suffer from problems like instability during training and mode collapse. Recently Bang and Shim [@DBLP:improved_training_using_representative_features] proposed RFGAN that uses pre-trained encoder features (representative features) to regularize the training of the discriminator to alleviate the problem of mode collapse. Similarly, MR-GAN[@mrgan_che2016mode] also proposes to use autoencoder features as regularizer in GAN training. Inspired from these architectures, we propose IDVAE that exploits the complementary properties of forward KL and reverse KL to capture the data distribution. While these approaches make use of a pre-trained encoder, our approach jointly and simultaneously trains a VAE and GAN achieving both the reconstruction and generation capabilities. There has also been some efforts towards utilizing the advantages of VAE for training GANs. Larsen [@vae_gan_DBLP:journals/corr/LarsenSW15] proposed VAE-GAN that collapses the decoder of the VAE into the generator of the GAN. VAE-GAN achieves sharp generations using a similarity metric learned by the intermediate representations of an explicit adversary. VAE-GAN requires an explicit discriminator, while our proposed approach overcomes this necessity by converting the encoder of the network into a discriminator. ALI[@ALI_dumoulin2017adversarially-iclr] and BiGAN[@BIGAN_DBLP:journals/corr/DonahueKD16] also propose to use three networks: the encoder, the decoder and the adversary. Unlike IDVAE, both the ALI and the BiGAN discriminator differentiates between samples from the joint distribution of observed data and latent codes. However, the reported reconstructions are of poor quality [@Mescheder2017ICML]. Akin to BiGAN discriminator setting, the AVB[@Mescheder2017ICML] model uses an additional discriminator to facilitate learning without explicitly assuming any form for posterior distribution. However, the samples generated by AVB for the CelebA dataset are observed to be blurry [@AVB_github_repo]. In contrast, the simpler IDVAE model is able generate higher quality samples with lesser parameters.
Li [@ALICE_li2017alice] propose ALICE that improves upon ALI by alleviating certain undesirable solutions (saddle points). Unlike a two tier approach of IDVAE, ALICE requires three networks and proposes to regularize the objective with cycle loss (an upper bound for conditional entropy). While in IDVAE, there is an implicit regularization on the discriminator by sharing its parameters with the encoder. The AS-VAE model of Pu [@ASVAE_NIPS2017_7020] focuses on both reverse and forward-KL between the encoder and decoder joint distributions with an objective to maximize the marginal likelihood of observations and latent codes. AS-VAE also needs two adversaries to circumvent the need of assuming an explicit form for the true intractable distribution (eqn 8 and 9 in [@Mescheder2017ICML]). IDVAE also focuses on forward-KL and reverse-KL but in a very novel way by sharing the parameters of the encoder (forward-KL) and discriminator (reverse-KL) resulting in a simpler model. $\alpha$-GAN [@rosca2017variational] fuses VAE and GAN exploiting the density ratio trick by constructing two additional discriminators for measuring the divergence between the reconstructions and the true data points, and the latent representations and the latent prior. The first discriminator minimizes the reverse-KL divergence, and the reconstruction error term minimizes the forward-KL divergence to discourage mode collapse. Training $\alpha$-GAN is difficult as it requires learning a large set of parameters (for the 4 networks). In contrast to previous approaches, Ulyanov [@AGE_DBLP:conf/aaai/UlyanovVL18] propose a two tier adversary free approach, AGE, where the encoder network is responsible for the adversarial signal. While the architectures of AGE and IDVAE appear to be similar, there are some fundamental differences in the process of learning the discriminator. The AGE discriminator compares (via divergence) the encoded real and fake distributions against a fixed reference distribution (typically, a prior in latent space). Whereas the IDVAE discriminator directly compares the real and fake data using a simple cross entropy loss, where both the reconstructions and randomly generated samples are treated as fake examples. We empirically show that IDVAE learns better as its discriminator relies on reconstructed samples as well. Importance of reconstructed samples in adversarial learning is supported in literature[@MRGAN_DBLP:journals/corr/CheLJBL16].
Methodology
===========
**Notations** Let $\textbf{x}$ be the data point in the input space $\mathcal{X}$ and $\textbf{z}$ be the code in the latent space $\mathcal{Z}$. The output of the encoder and the discriminator network for an input $\textbf{x}$ is represented as $\text{Enc}(\textbf{x})$ and $\text{Dis}(\textbf{x})$ respectively. Similarly the output of decoder network $\Tilde{\textbf{x}}$ for a latent code $\textbf{z}$ is denoted by $\text{Dec}(\textbf{z})$. The output at the $l^{\text{th}}$ layer of the encoder network for an input $\textbf{x}$ is denoted as $\text{Enc}_{l}(\textbf{x})$. This is same as the output at the $l^{\text{th}}$ layer of the discriminator network for an input $\textbf{x}$ which is denoted as $\text{Dis}_{l}(\textbf{x})$. $\text{Enc}_{l}(\textbf{x})$ and $\text{Dis}_{l}(\textbf{x})$ are used interchangeably depending on the context. In reference to Figure \[fig:onecol\], we denote the encoder specific parameters by $\theta_{enc}$ where $\theta_{enc} = \{\theta_{\mu},\theta_{\Sigma}\}$.
![The proposed architecture for IDVAE. The parameters $\theta_{\mu}$ and $\theta_{\Sigma}$ denote a single fully connected layer learning the encoder specific parameters, $\theta_{\text{enc}}$. $\theta_{\text{dis}}$ also represents a single fully connected layer which denotes discriminator specific parameters. Similarly $\theta_{\text{shared}}$ denotes the shared parameters between the encoder and the discriminator whereas $\theta_{dec}$ denotes the decoder/generator specific parameters.[]{data-label="fig:long"}](images/IDVAE.png){width="\linewidth"}
\[fig:onecol\]
We start with some preliminaries on VAE[@VAE_kingma] and GAN[@Goodfellow:2014:GAN:2969033.2969125] before describing our proposed model, which combines both of them.
Variational Autoencoder
-----------------------
A VAE comprises of learning two networks, namely, the encoder and the decoder network. In contrast to traditional autoencoders, VAE views the encoder and the decoder networks as probabilistic functions. The encoder learns a conditional distribution on the latent code $\textbf{z}$ conditioned on the input $\textbf{x}$. Similarly decoder learns a distribution on $\Tilde{\textbf{x}}$ conditioned on the latent code $\textbf{z}$.
$$\textbf{z} \thicksim \text{Enc}(\textbf{x}) = q(\textbf{z}|\textbf{x})$$
$$\Tilde{\textbf{x}} \thicksim \text{Dec}(\textbf{z}) = p(\textbf{x}|\textbf{z})$$
Vanilla VAE jointly trains over the encoder and the decoder network parameters by minimizing negative log-likelihood (reconstruction term) and divergence between prior and learned distribution in latent space $\mathcal{Z}$. The prior, $p(\textbf{z})$, over the latent space is typically assumed to be a unit Normal distribution, i.e. $\textbf{z}\thicksim \mathcal{N}(\textbf{0}, \textbf{I})$. Thus training a VAE would mean minimizing the following loss: $$\mathcal{L}_{\text{VAE}} = \mathcal{L}_{\text{recons}} + \mathcal{L}_{\text{prior}}$$ where, $$\label{eqn:eq_4}
\mathcal{L}_{\text{recons}} = -\mathds{E}_{q(\textbf{z}|\textbf{x})}[\log p(\textbf{x}|\textbf{z})]$$ $$\label{eqn:eq_5}
\mathcal{L}_{\text{prior}} = KL(q(\textbf{z}|\textbf{x})\|p(\textbf{z}))$$ and $KL(q(\textbf{z}|\textbf{x})\|p(\textbf{z}))$ is the Kullback-Leibler divergence between the distributions $q(\textbf{z}|\textbf{x})$ and $p(\textbf{z})$.
Generative Adversarial Network
------------------------------
A GAN consists of two networks, namely, the discriminator and the generator that are learned through an adversarial game play. The generator network maps a point $\textbf{z}$ in some arbitrarily low dimensional latent space $\mathcal{Z}$ to a point in a high dimensional data space $\mathcal{X}$. We denote $\text{Gen}(\textbf{z})$ as the output of the generator network when $\textbf{z}$ is the input. In a similar vein, the discriminator network maps a data point $\textbf{x}$ in the data space to a probability value $y \in [0,1]$. The objective of the discriminator is to assign the probability $y=\text{Dis}(\textbf{x})$ that $\textbf{x}$ is a sample from true distribution and the probability $1-y$ that $\textbf{x}$ is a generated sample i.e. $\textbf{x}=\text{Gen}(\textbf{z})$, with $\textbf{z}\thicksim p(\textbf{z})$. Thus, in this adversarial game play, the objective of the generator is to synthesize samples that can fool the discriminator learning the true data distribution, while the goal of the discriminator is to recognize the samples coming out of the generated(fake) distribution and the true distribution. Adversarial game play between the discriminator and the generator is formally defined by the GAN loss as $$\label{eqn:eq_6}
\mathcal{L}_{\textit{GAN}} = \log(\text{Dis}(\textbf{x})) + \log(1-\text{Dis}(\text{Gen}(\textbf{z})))$$ We want to maximize the binary cross entropy loss with respect to the discriminator (D) while minimizing it for the generator (G). Thus, the minimax objective is defined as $$\label{eqn:eq_6_2}
\underset{G}{\text{min}}\, \underset{D}{\text{max}}\,\mathcal{L}_{\textit{GAN}}$$
IDVAE
-----
Our proposed approach, **I**mplicit **D**iscriminator in **V**ariational **A**uto**E**ncoder (IDVAE), exploits the properties of both VAE and GAN. IDVAE sustains the stable training properties of VAE while generating samples of quality approaching GAN. We borrow the encoder and decoder networks from VAE with slight modifications. In particular, we collapse the VAE decoder network into the generator of the GAN and the VAE encoder network partially into the discriminator of the GAN.
We partially collapse the encoder into the discriminator following the assumption that there exists an overlap in the knowledge of encoder and discriminator network. As the encoder’s objective is to learn representational features, while the discriminator’s objective is to learn discriminative features, we restrict the weight sharing to the penultimate layer (say some $l^{th}$ layer of encoder represented as $\text{Enc}_{l}$) in encoder of the VAE. Further to encourage the encoder to learn the features of discriminator we add a single fully connected layer from $\text{Enc}_l$ to a single sigmoid node that acts as the discriminator’s output. Figure \[fig:onecol\] illustrates the proposed IDVAE network architecture.
Thus, in our model we have four sets of parameters that need to be learned, namely; $\theta_{\text{dec}}$ - the shared parameters between the decoder and the generator, $\theta_{\text{shared}}$ - the parameters shared between the encoder and the discriminator, $\theta_{\text{enc}}$ - the encoder specific parameters of the VAE, and $\theta_{\text{dis}}$ - the discriminator specific parameters of the GAN. These are updated based on the loss incurred by each of the individual networks.
The loss incurred by the encoder is used to update both $\theta_\text{shared}$ and $\theta_\text{enc}$. The encoder loss in the IDVAE consists of two components similar to a standard VAE. The first component is the reconstruction loss - $\mathcal{L}_{\text{recons}}$ and the second component is the prior discrepancy loss - $\mathcal{L}_{\text{prior}}$. It is well known that minimizing the forward-KL divergence KL($P_{data}\|P_{model}$) achieves mode coverage for generative models. Thus, we minimize forward-KL divergence by minimizing $\mathcal{L}_{\text{recons}}$ for helping IDVAE to learn the different modes in the data. The shared parameters between the encoder and the discriminator encode the forward-KL divergence information. Thus using $\mathcal{L}_{\text{recons}}$ in the encoder reduces the extent of mode collapse as the gradients from the discriminator to the generator implicitly contain the information about multiple modes. Thus the overall encoder loss ($\mathcal{L}_{enc}$) is defined as follows $$\label{eqn:eq_9}
\mathcal{L}_{\text{enc}} = \alpha\mathcal{L}_{\text{recons}} + \beta\mathcal{L}_{\text{prior}}%+ \beta\mathcal{L}_{\text{recons}}^{\text{dis}} +$$
where $\alpha$ and $\beta$ are hyper parameters controlling the contribution of each of the loss terms.
It has been shown that GAN [@Goodfellow:2014:GAN:2969033.2969125] achieves sharper images by minimizing the reverse-KL divergence. Thus, we use the implicit adversary (encoder as a shared discriminator) of IDVAE as a way to propagate reverse-KL divergence information. The discriminator loss is used to update both $\theta_\text{shared}$ and $\theta_\text{dis}$. The generated (fake) examples that are presented to the discriminator of IDVAE are the output of the decoder when viewed as a generator $\text{Dec}(\textbf{z})$, where $\textbf{z}\thicksim p(\textbf{z})$. In addition to this, we also present the synthesized sample through reconstruction, $\text{Dec}(\text{Enc}({\textbf{x})})$, for an input $\textbf{x}$. As $\text{Dec}(\text{Enc}(\textbf{x}))$ is more likely to be similar to $\textbf{x}$ than $\text{Dec}(\textbf{z})$, for an arbitrary $\textbf{z} \thicksim p(\textbf{z})$, we hypothesize that the discriminator loss corresponding to $\text{Dec}(\text{Enc}(\textbf{x}))$ encourages the generator to learn the properties of the decoder. Similarly the discriminator loss corresponding to $\text{Dec}(\textbf{z})$ encourages the decoder to learn the properties of the generator be able to generate realistic examples from the prior distribution $p(\textbf{z})$. Therefore using both the terms, $\text{Dec}(\textbf{z})$ and $\text{Dec}(\text{Enc}({\textbf{x})})$ encourages the model to learn a blend of both the generator and the decoder. Intuitively in equation \[eqn:eq\_10\] to maintain the ratio of real and fake samples shown to the discriminator the loss terms for the fake samples should be scaled by a factor of 0.5 or the real term $\text{log(Dis(\textbf{x}))}$ by 2. We observed no significant change in the performance of IDVAE when these factors are dropped, thereby giving rise to the following loss function $$\begin{gathered}
\label{eqn:eq_10}
\mathcal{L}_{\text{dis}} = -\big[\log(\text{Dis}(\textbf{x})) +\log(1-\text{Dis}(\text{Dec}(\textbf{z}))) \\
+\log(1-\text{Dis(Dec(Enc($\textbf{x}$))})\big]
\end{gathered}$$
As we have collapsed the decoder of the vanilla VAE into the generator, the loss incurred by both the decoder and generator is used to update the shared parameters between the decoder and generator ($\theta_\text{dec}$). The decoder/generator loss in IDVAE consists of two components. The first component ($\mathcal{L}^\text{dis}_\text{recons}$) is a learned similarity metric motivated by VAE-GAN [@vae_gan_DBLP:journals/corr/LarsenSW15]. Specifically, we learn a similarity metric($\mathcal{L}_{\text{recons}}^{\text{dis}}$) using an intermediate representation ( $l^{\text{th}}$ layer) of the discriminator (equivalent to the $l^{\text{th}}$ layer of the encoder) by assuming a Gaussian observation model on $\text{Dis}_{l}(\Tilde{\textbf{x}})$ with mean $\text{Dis}_{l}(\textbf{x})$ and unit covariance : $$\label{eqn:eq_7}
p(\text{Dis}_{l}(\Tilde{\textbf{x}})|\textbf{z}) = \mathcal{N}(\text{Dis}_{l}(\Tilde{\textbf{x}})|\text{Dis}_{l}(\textbf{x}),\textbf{I})$$ where for a given sample **x**, $\Tilde{\textbf{x}} = \text{Dec}(\textbf{z})$ and $\textbf{z} = \text{Enc}(\textbf{x})$. $\mathcal{L}_{\text{recons}}^{\text{dis}}$ is defined as a Gaussian observation model: $$\label{eqn:eq_8}
\mathcal{L}_{\text{recons}}^{\text{dis}} = -\mathds{E}_{q(\textbf{z}|\textbf{x})}[\text{log }p(\text{Dis}_{l}(\textbf{x})|\textbf{z})]$$ The second component is the adversarial loss which encourages the decoder to learn the properties of a generator. The adversarial loss, ($\mathcal{L}_{\text{GAN}}$), is defined as$$\label{eqn:eq_11}
\mathcal{L}_{\text{GAN}} = -log(\text{Dis(Dec($\textbf{z}'$)}))- log(\text{Dis(Dec(Enc(\textbf{$\textbf{x}$})))})$$ where $\textbf{z}' \thicksim p(\textbf{z})$.
Therefore the presence of both the reconstruction loss and the adversarial loss in objective function of decoder makes it learn a blend of the two models. The overall loss function for the decoder/generator ($\mathcal{L}_\text{dec}$) is defined as: $$\label{eqn:eq_12}
\mathcal{L}_\text{dec} = \omega\mathcal{L}_{\text{GAN}}
+\lambda\mathcal{L}_{\text{recons}}^{\text{dis}}$$
where $\omega \text{ and }\lambda$ used in $\mathcal{L}_\text{dec}$ are hyper-parameters that are learned empirically.
Training Schedule
-----------------
The $l^{th}$ shared layer between the encoder and the discriminator outputs representations for learning the parameters of the encoder’s distribution and for discriminating between samples from the true distribution and generated samples simultaneously. We need to ensure that the shared weights ($\theta_{\text{shared}}$) of the encoder and discriminator network gets the learning signal corresponding to both the encoder and discriminator objective function. In theory $\theta_{\text{shared}},\theta_{\text{dis}}$, and $\theta_{\text{enc}}$ can be updated in a single step using the joint loss of both the encoder and discriminator. However, in practise we observed training using the joint loss to be challenging in terms of hyper-parameter fine tuning. Hence, we update the shared weights ($\theta_{\text{shared}}$) in two iterations, once each with the encoder and discriminator losses respectively. In the first iteration $\theta_{\text{shared}}$ and $\theta_\text{dis}$ are updated while in the second iteration $\theta_\text{shared}$ and $\theta_\text{enc}$ are updated. Algorithm \[alg:train\_schedule\] presents the overview of the training procedure. Thus the parameters $\theta_\text{shared}$ learn the information of both reverse-KL (first iteration) and forward-KL (second iteration), which is leveraged by the decoder/generator in the third step of the algorithm. The parameters, $\theta_\text{shared}$, can be updated in any arbitrary order, we empirically found that using first the discriminator loss helps in better learning. We present the qualitative results on the other two variants (using the joint loss, and updating $\theta_{\text{shared}}$ with respect to the encoder first followed by the discriminator) in the supplementary material.
$P(z)\leftarrow \mathcal{N}(0,\textit{I}) $\
$\theta_\text{shared},\theta_\text{enc},\theta_\text{dec},\theta_\text{dis}\leftarrow \text{Initialize parameters}$ $X \gets \text{random mini batch from dataset} $ $Z \gets \text{Enc($X$)} $ $\Tilde{X} \gets \text{Dec($Z$)} $ $Z' \gets \text{samples from prior }P(Z) $ $X' \gets \text{Dec($Z'$)} $
$\theta_\text{dis},\theta_\text{shared} \xleftarrow[]{+} -\nabla_{(\theta_\text{dis},\theta_\text{shared})}(\mathcal{L}_\text{dis})$ $\theta_\text{enc},\theta_\text{shared} \xleftarrow[]{+} -\nabla_{(\theta_\text{enc},\theta_\text{shared})}(\mathcal{L}_\text{enc})$
$\theta_\text{dec} \xleftarrow[]{+} -\nabla_{\theta_\text{dec}}(\mathcal{L}_\text{dec})$
Experiments
===========
We investigate the proposed IDVAE architecture for the quality of both reconstructions and generations. We evaluate the performance of IDVAE on the following two real world benchmark datasets: i) **CIFAR10** [@cifar10_krizhevsky2009learning], which contains 60k images of which 50k are used for training and the remaining 10k for testing. (ii) **CelebA** [@celebA_liu2015faceattributes], which consists of 202,599 images. We use 1-162,770 images for training, 162,771-182,637 for validation and rest for testing. In our implementation pipeline we crop and scale the images to 64x64 for faster training. The details of encoder and decoder network architecture along with fine tuned hyper-parameters for each of the dataset are provided in the supplementary material.
For generating instances over the different datasets we randomly sample $\textbf{z}$ from the assumed prior distribution (on the latent space $\mathcal{Z}$) $\mathcal{N}(\textbf{0},\textbf{I})$. We also conducted experiments on a synthetic 2D Gaussian dataset and the MNIST digits dataset. These details can be found in the supplementary material.
We compare the performance of IDVAE against approaches that have both generative and reconstruction abilities, namely; VAE[@VAE_kingma], VAE-GAN[@vae_gan_DBLP:journals/corr/LarsenSW15], AGE [@DBLP:journals/corr/UlyanovVL17a], and $\alpha$-GAN[@rosca2017variational]. We use the pre-trained models available for AGE, while we train all the other models from scratch using the best hyper-parameters reported in the literature.
Results and Discussion
======================
We conduct both qualitative and quantitative comparison of IDVAE for generations and reconstructions against all the prior approaches.
Quantitative Analysis
---------------------
The reconstruction quality is objectively quantified using the standard square loss $\mathcal{L}_\text{recons}$. We obtain an unbiased estimate of the loss using a large test set consisting of 10k samples for both the CelebA and the CIFAR10 datasets. Thus even a small improvement on such a large set is significant considering the complexities of the dataset. It is evident from the results presented in Figure \[fig:recons\_scores\] that for the reconstruction task IDVAE performs better than or is at par with VAE-GAN, AGE and $\alpha$-GAN. However the lowest reconstruction error is obtained by VAE. This is understandable as there is no explicit penalty on the decoder for reconstructing unrealistic images. On the other hand both VAE-GAN and IDVAE strike a balance between the reconstruction loss and the generative ability. We also modify IDVAE to explicitly minimize $\mathcal{L}_{\text{recons}}$ in the decoder, which we term as IDVAE(R). The reconstruction loss of IDVAE(R) is the best among all the variants of VAE, However, this comes at the cost of quality of the generations.
![Comparing reconstruction loss (lower is better) among different generative models.[]{data-label="fig:recons_scores"}](recons.png){width="50.00000%"}
There are two popular measures for qualitatively evaluating generative models, namely the Inception Score (IS) [@inception_score_salimans2016improved] and Fréchet Inception Distance (FID) [@DBLP:FID_journals/corr/HeuselRUNKH17]. It has been shown that IS closely follows human scoring of images synthesized by generative models for the CIFAR 10 dataset[@cifar10_krizhevsky2009learning]. The IS uses the Inception v3 model pre-trained on ImageNet. The IS is a statistic on the Inception model’s output when applied to the synthesized images. This statistic captures two desirable qualities of a generative model - the synthesized images should contain an object (the image is sharp and not blurry) that is reflected in a low entropy output of the Inception model[@inception_model_szegedy2015going] and; there must be diversity in the generations that is reflected in the high entropy output of the Inception model over the entire generated set. Barratt and Sharma [@note_on_IS_barratt2018note] have recently shown that the IS suffers from suboptimalities and is an appropriate measure only for datasets that are trained on ImageNet. Thus, it is not advisable to measure the quality of generations on the CelebA dataset. The FID improves upon the IS by comparing the statistics of both the generated and true samples, instead of evaluating the generated samples in isolation. The FID is the Fréchet distance between two multivariate Gaussians estimated from the 2048-dimensional activations of the Inception-v3 pool3 layer for real and generated samples. Lower FID scores correspond to more similar real and synthetic samples.
Table \[tab:FID\_scores\], presents the Fréchet distance scores computed using Inception-15[^2] ($FID_{15}$), Inception-16[^3] ($FID_{16}$), and ResNet[@resnet_he2016deep]. As all the three experts have the same knowledge all the models are pre-trained on ImageNet[@imagenet_cvpr09] and the representations extracted from the intermediate layer have the same dimensions (2048), the relative performance of the models with respect to each expert can be compared. While the distance scores across the experts for the same model may be different, we expect the order of the goodness among the generative models to be preserved. However, as can be observed from Table \[tab:FID\_scores\] on CIFAR10 dataset, VAE-GAN appears to perform better than the AGE based on $FID_{15}$, while the trend reverses when comparing based on both $FID_{16}$ and $FRD$. Therefore, our results suggest that a generative model should be compared across a battery of experts rather than in isolation. IDVAE performs better than all the other approaches on the CelebA dataset. The result is statistically significant on both $FID_{15}$ and $FRD$ scores. On the other hand both IDVAE and $\alpha$-GAN result in the best performance on the CIFAR10 dataset. There is no significant difference between IDVAE and $\alpha$-GAN with each performing better than the other only according to a single measure. However, from Figure \[fig:recons\_scores\], it is quite apparent that $\alpha$-GAN focuses less on reconstructions whereas in IDVAE we do not observe such a bias. These results support our hypothesis that the encoder and discriminator can be a shared network. IDVAE is able to perform at par or sometimes better than VAE-GAN and $\alpha$-GAN that require a separate encoder/discriminator network. This is further verified through our qualitative results. We also observe a dip in the Fréchet distance for the IDVAE(R) model in comparison to IDVAE. As the decoder/generator of IDVAE(R) focuses on reconstructions in the image space we observe a drop in the Fréchet distance at the cost of a better reconstruction loss. Therefore IDVAE(R) model has the potential to fit within the required thresholds by tuning the hyper parameters $\omega, \lambda \text{ and } \gamma$.
[|l|c|c|c|]{} Model & $\text{FID}_{15}$ & $\text{FID}_{16}$ & FRD\
\
True Data & 3.16$\pm$0.06 & 7.41$\pm$0.82 & 26.17$\pm$1.44\
IDVAE & 23.48$\pm$0.15 & 28.15$\pm$0.39 & **105.45$\pm$0.79**\
IDVAE(R) & 43.38$\pm$0.15 & 49.9$\pm$0.85 & 191.32$\pm$5.88\
VAE-GAN & 27.04$\pm$0.12 & 33.12$\pm$0.73 & 139.95$\pm$2.71\
VAE & 85.74$\pm$0.3 & 130.38$\pm$3.47 & 626.67$\pm$8.61\
AGE & 32.19$\pm$0.3 & 29.3$\pm$0.54 & 122.43$\pm$2.61\
$\alpha$-GAN & **20.61$\pm$0.12** & 27.87$\pm$0.7 & 121.88 $\pm$ 3.09\
\
True Data & 1.58$\pm$0.02 & 2.67$\pm$0.15 & 5.77$\pm$0.35\
IDVAE & **8.53$\pm$0.12** & 9.52$\pm$0.72 & **34.47$\pm$2.41**\
IDVAE(R) & 14.81$\pm$0.17 & 16.71$\pm$0.66 & 70.99$\pm$0.91\
VAE-GAN & 9.52$\pm$0.06 & 10.5$\pm$0.9 & 38.32$\pm$1.69\
VAE & 35.27$\pm$0.04 & 55.44$\pm$0.87 & 150.02$\pm$1.41\
AGE & 12.74$\pm$0.14 & 15.27$\pm$0.36 & 82.45$\pm$0.97\
$\alpha$-GAN & 10.38$\pm$0.2 & 13.89$\pm$1.58 & 55.44$\pm$7.97\
\
Qualitative Analysis
--------------------
We present the qualitative results obtained from the different models in the Table \[tab:qualitative\] on the CIFAR10 and CelebA datasets. It is quite evident from the images that VAE results in blurry reconstructions while the rest of the approaches output sharp images, which is due to presence of adversarial loss. On both CelebA and CIFAR10 datasets, we observe IDVAE and IDVAE(R) performing on par with VAE-GAN and $\alpha$-GAN, while significantly outperforming VAE in terms of sharpness of the images. The images generated by the different models are also presented in Table \[tab:qualitative\]. The undesirable blurriness property in VAE is apparent on the CIFAR10 dataset while the performance of IDVAE is on par with both $\alpha$-GAN and VAE-GAN. The images generated by IDVAE trained on the CelebA dataset appears to capture a large diversity in background when compared to $\alpha$-GAN and VAE-GAN. We observe both IDVAE(R) and $\alpha$-GAN tend to focus more on faces than the background in these images.
-------------- ------------------------------------------------------------------------------------ --------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------
Approach
Original {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"}
VAE {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"}
IDVAE {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"}
IDVAE(R) {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"}
VAE-GAN {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"}
$\alpha$-GAN {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"} {width="\linewidth" height="3cm"}
-------------- ------------------------------------------------------------------------------------ --------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------
Conditional IDVAE
-----------------
We extend IDVAE to a conditional setting where our objective is to learn a generator/decoder whose output is controlled by some conditional information, $y$. We term this variant as **C**onditional-IDVAE (C-IDVAE). We qualitatively analyze the C-IDVAE model using MNIST[@mnist_LeCun1998GradientBasedLA] and CelebA datasets. We follow the recent work of Perarnau [@icgan_perarnau2016invertible] to provide the conditional information to the generator at the input layer, while for the discriminator this is provided after the first convolution layer. To the best of our knowledge the encoder of VAE is never made aware of the conditional information but as we have our encoder acting as a discriminator we add this conditional information after the first convolution layer.
We use the one hot encoding of the MNIST class labels as the conditional information to evaluate C-IDVAE. Figure \[fig:exp\_mnist\_conditional\_IDVAE\] illustrates the samples generated by C-IDVAE, where each row illustrates the images generated by conditioning on a unique label. We observe a large diversity in the generations in each row implying the diverse generative ability of the conditional decoder/generator.
Following the work of Perarnau [@icgan_perarnau2016invertible] we use 13 attributes that have clear visual impact out of the total 40 attributes as conditioning information while training C-IDVAE on the CelebA dataset. Figure \[fig:exp\_mnist\_conditional\_IDVAE\] presents the images generated by C-IDVAE for different conditioning information. Each row in the figure represents the images generated by C-IDVAE with the conditioning information provided in the top row, and the original image that is modified in the first column. It can be observed that the changes in each of the generations with respect to the original image are a result of the model imagining the original image on different attributes. For example, consider third and fifth rows in the Figure \[fig:exp\_mnist\_conditional\_IDVAE\], where the original image is a female face. The generations in the columns 2, 5, and 6 that are conditioned on the male attribute actually contain a face that resembles a male. Similarly in the last column the model does reasonably well in adding eyeglasses to all the generated images. Thus considering the generative ability matching with the human imagination and the complexity of the real world CelebA dataset, C-IDVAE shows the potential to model the complex distributions.
![Conditional-IDVAE on MNIST followed by celebA conditioned on visual attributes.[]{data-label="fig:exp_mnist_conditional_IDVAE"}](images/new_conditional-IDVAE/gen-img.png "fig:"){width="0.65\linewidth"} ![Conditional-IDVAE on MNIST followed by celebA conditioned on visual attributes.[]{data-label="fig:exp_mnist_conditional_IDVAE"}](images/celebA/conditional_IDVAE/conditional_inference.png "fig:"){width="0.65\linewidth"}
Summary and Future Work
=======================
In this work we introduce a novel hybrid of the variational autoencoder and the generative adversarial network, IDVAE, which does not need an explicit discriminator network. IDVAE shares a common decoder and generator network, and partially shares the encoder and the discriminator network. The qualitative and quantitative experiments on real-world benchmark datasets demonstrates that IDVAE (and its variant IDVAE(R)) performs on par and sometimes better than the state of the art hybrid approaches. We also show that IDVAE can be easily extended to work in a conditional setting, and experimentally demonstrate its performance on complex datasets. Further, our results present inadequacies of the Fréchet Inception Distance and suggests an ensemble of experts for evaluating the quality of the generations. This can be further explored to derive a measure that does not require a model that is pre-trained on data from a different domain as that of the training samples.
[^1]: Authors with equal contribution
[^2]: weights used from 2015 year model [@inception2k15]
[^3]: weights used from 2016 year model [@inception2k16]
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Rui Guo
- 'Cai-Na Hao'
- 'Xiao-Yang Xia'
date: 'Received ; accepted '
title: 'The Evolution of Advanced Merger (U)LIRGs on the Color-Stellar Mass Diagram'
---
INTRODUCTION
============
Since the discovery of the color bimodality on the color-magnitude and the color-stellar mass (color-mass, hereafter) diagrams based on the Sloan Digital Sky Survey (SDSS) (Kauffmann et al. 2003b; Baldry et al. 2004, 2006) as well as on high redshift survey samples (Bell et al. 2004; Faber et al. 2007; Ilbert et al. 2010), there are mounting works to investigate the evolutionary pathways from the blue cloud of star-forming galaxies to the red sequence of quiescent galaxies. The evolution from the blue cloud to the red sequence is suggested to be results of star formation being quenched (e.g., Faber et al. 2007). According to the quenching time-scales, the evolutionary tracks are classified into fast and slow modes (e.g., Schawinski et al. 2014; Yesuf et al. 2014). The fast track (with time-scales less than 1Gyr) has been proposed to interpret the formation of high redshift compact massive quiescent galaxies (van Dokkum et al. 2008; Kriek et al. 2009; Barro et al. 2013; Muzzin et al. 2013; Marchesini et al. 2014). These studies show that the progenitors of local massive galaxies are heavily dust-extincted starburst galaxies that are triggered by gas-rich major mergers. Therefore, investigation of local gas-rich major mergers might shed light on the fast quenching process.
In the local universe, almost all ultraluminous infrared galaxies (ULIRGs; $\LIR$[^1]$>10^{12}\Lsun$) and about half of luminous infrared galaxies (LIRGs; $10^{11}\Lsun<\LIR<10^{12}\Lsun$) were found to be gas-rich interacting/merging galaxies (e.g. Sanders & Mirabel 1996; Wang et al. 2006; Kaviraj 2009). Therefore, nearby (U)LIRGs can serve as a proper local laboratory for studying the fast quenching process. Virtually, several recent works have made efforts to identify the role of local ULIRGs in the migration from the blue cloud to the red sequence. Chen et al. (2010) and Kilerci Eser et al. (2014) have investigated the positions of local ULIRGs in the [*g-r*]{} and [*u-r*]{} color-magnitude diagrams, respectively, without applying internal extinction corrections. They found that about half of the ULIRGs sample lie outside the 90% level number density contour, and some in the green valley or the red sequence. Most recently, Guo et al. (2016b) applied both Galactic and internal dust extinction corrections for an advanced merger (adv-merger) (U)LIRGs sample and found that adv-merger (U)LIRGs are blue and more massive than the blue cloud galaxies, with $95\%\pm 2\%$ and $81\%\pm 4\%$ of them outside the blue cloud on the [*u-r*]{} and [*NUV-r*]{} color-mass diagrams, respectively, implying that the adv-merger (U)LIRGs occupy a distinct region on the color-mass diagrams. In addition, the investigation of the locus of the adv-merger (U)LIRGs in the star formation rate (SFR)-stellar mass (SFR-$M_{\ast}$) diagram showed that about three-fourths of the sample are above the star-forming main sequence, indicating that the majority of adv-merger (U)LIRGs are experiencing massive starbursts. Based on the estimates of gas exhausting time-scales, the authors suggested that the adv-merger (U)LIRGs are likely at the starting point of the fast evolutionary track.
As is well known, the equivalent width of H$\delta$$_A$ absorption line (EW(H$\delta$$_A$)) and the 4000[Å]{} break (D$_n(4000)$) measured from optical spectra are powerful indicators of the recent star formation histories. A strong H$\delta$$_A$ absorption line is a sign for a burst in the star formation history that ended about 0.1-1 Gyr ago, while the 4000[Å]{} break is an excellent age indicator for young star populations within 1 Gyr (Kauffmann et al. 2003b, 2003c). As a result, EW(H$\delta$$_A$) and D$_n(4000)$ are often used to probe recent starbursts and distinguish star formation histories dominated by starbursts from a continuous mode (Kauffmann et al. 2003b; Goto 2007; Martin et al. 2007; Yesuf et al. 2014). In addition, the equivalent width of H$\alpha$, EW(H$\alpha$), is a measure of the current to past average SFR (e.g., Knapen & James 2009; Yesuf et al. 2014).
To explore the evolutionary link between starburst galaxies and post-starbursts, Yesuf et al. (2014) proposed a classification scheme to identify transiting post-starbursts, an intermediate stage between starbursts and post-starbursts, based on EW(H$\alpha$), EW(H$\delta$$_A$), D$_n(4000)$, UV and [*Wide-field Infrared Survey Explore*]{} ([*WISE*]{}; Wright et al. 2010) photometry. Guo et al. (2016b) employed this set of criteria and identified 12 transiting post-starbursts out of the entire sample of 89 adv-merger (U)LIRGs. It is also worth carefully investigating the distributions of the spectral features of EW(H$\alpha$), EW(H$\delta$$_A$), D$_n(4000)$ on the color-mass and SFR-mass diagrams to search for evidence for possible evolutionary signs already shown in the adv-merger (U)LIRGs sample.
The paper is structured as follows. In Sections 2 we describe the sample selection and parameter estimations for our sample adv-mergers and control samples. In Section 3, we present our results and discussion. The main results are summarized in Section 4. Throughout this paper we adopt the Kroupa (2001) initial mass function and a cosmological model of ${H}_{\rm
0}=70\,{\rm km \, s^{-1}\, Mpc^{-1}}$, $\Omega_{\rm m}=0.3$ and $\Omega_{\rm
\Lambda}=0.7$.
SAMPLE SELECTION AND PARAMETER ESTIMATIONS
==========================================
Sample
------
Our adv-merger (U)LIRGs sample is directly taken from Guo et al. (2016b). We refer the readers to that paper for a detailed description of the sample selection (see also Guo et al. 2016a). In the following, we briefly describe the selection procedures.
![RGB color images of 16 example adv-mergers with single nucleus but some merger signatures, e.g., tidal tails, indicating that they are in the late stage of merging. The images are constructed from SDSS ([*g, r, i*]{}) images following Lupton et al. (2004) with 80 $\times$ 80 $\rm kpc^2$ physical size.[]{data-label="colorim.eps"}](fig1.eps){width="\textwidth"}
The parent (U)LIRGs sample in Guo et al. (2016b) was drawn from a cross-correlation analysis between the spectroscopic catalog of SDSS DR7 (York et al. 2000; Abazajian et al. 2009) and the [*Infrared Astronomical Satellite (IRAS)*]{} Point Source Catalog Redshift Survey (PSCz, Saunders et al. 2000), as well as a cross-identification between the [*IRAS*]{} 1 Jy ULIRGs sample (Kim & Sanders 1998) and the SDSS DR7 photometric catalog. We performed morphological classifications visually for the sub-sample of galaxies in the magnitude range of $14.5<r<15.9$[^2] after corrections for foreground Galactic extinction (Schlegel et al. 1998), to ensure both reliable morphological classifications and the spectroscopic observation completeness for SDSS galaxies (Fukugita et al. 2004; Kauffmann et al. 2003b). Note that galaxies classified as Seyfert I are removed from our sample, since the estimations for their extinctions, stellar masses and star formation rates are not reliable. Finally, there are 89 adv-mergers selected, which are in the late merging stage. For the purpose of this work, measurements of optical spectral indices are needed. Hence only the 79 adv-mergers with SDSS spectroscopic observations are included in this paper, among which 59 are LIRGs and 20 are ULIRGs. Figure \[colorim.eps\] shows the true color images of 16 example adv-mergers.
Meanwhile, the control samples are also taken from Guo et al. (2016b) directly. For the study of the local color-mass relation, the control sample was selected from the Oh-Sarzi-Schawinski-Yi (OSSY) catalog (Oh et al. 2011)[^3], which provides internal extinction information (E(B-V)) from stellar continuum fits. By requiring the redshift range of $0.02<z<0.05$ and absolute magnitude $M_{z,\rm Petro} < -19.5$ mag, 53604 galaxies were obtained. The absolute magnitude constraint enables us to derive an approximately mass-limited sample as described in Schawinski et al. (2014). For investigating the SFR-$M_{\ast}$ relation of star-forming galaxies, the control sample has been retrieved from the SDSS DR7 in the redshift range of $0.005<z<0.2$ and luminosity range of $14.5<r<17.77$, which consists of 152137 galaxies.
Parameters
----------
The main purpose of this paper is to investigate the distributions of EW(H$\alpha$), EW(H$\delta_A$) and D$_n(4000)$ on the [*NUV-r*]{} color-stellar mass as well as the SFR-stellar mass diagrams. Therefore, apart from the estimates of the dust extinction corrected [*NUV-r*]{} colors, SFRs and stellar masses, as described in Guo et al. (2016b), optical spectral indices are also needed.
The estimations of the dust attenuation corrected [*NUV-r*]{} colors, SFRs and stellar masses have been described in Guo et al. (2016b). We only provide a brief description here. The optical photometric data and [*NUV*]{} magnitudes were taken from the SDSS DR7 and the [*Galaxy Evolution Explorer (GALEX)*]{} satellite (Martin et al. 2005), respectively. The [*NUV*]{} photometric data are available from the [*GALEX*]{} for 69 sample adv-merger (U)LIRGs and $\sim87\%$ control sample galaxies. We further performed [*k*]{}-corrections, the Galactic extinction and internal extinction corrections to the optical and [*NUV*]{} magnitudes. The stellar masses for our sample adv-mergers and the control sample galaxies were retrieved from the Max Planck Institute for Astrophysics-Johns Hopkins University (MPA/JHU[^4]) stellar mass catalog (Kauffmann et al. 2003b), which are based on the SDSS five broad-band photometry. The SFRs for 25 star-forming and 38 composite[^5] adv-merger (U)LIRGs were calculated using their H$\alpha$ luminosities taken from the MPA/JHU catalog following Kennicutt (1998) and then were converted to a Kroupa IMF. Note that the H$\alpha$ luminosities are aperture-corrected (Hopkins et al. 2003) as well as dust extinction corrected under the assumption of the case B recombination value of intrinsic H$\alpha$/H$\beta$ as 2.86.
The D$_n(4000)$, EW(H$\alpha$) and EW(H$\delta$$_A$) were retrieved from the MPA/JHU catalog, in which the EW(H$\delta$$_A$) has been corrected for the contamination by nebular emission. We further performed the internal extinction corrections for D$_n(4000)$ and EW(H$\alpha$), using the Calzetti’s law (Calzetti et al. 2000).
RESULTS AND DISCUSSION
======================
As presented in Guo et al. (2016b), most of our sample adv-merger (U)LIRGs are located above the star-forming main sequence on the SFR-$M_{\ast}$ diagram, indicating that they are experiencing massive starbursts as triggered by gas-rich major merging. All these galaxies are blue and massive, and thus lie to the right of the blue cloud galaxies on the optical and NUV color-mass diagrams, making them distinct from the blue cloud, red sequence and green valley galaxies. These properties suggest that they are at the starting point of the fast track of galaxy evolution from the blue cloud to the red sequence. Therefore, it is interesting to ask whether they have already shown an evolutionary trend, as imprinted in their star formation history indicators. In this work, we investigate the distributions of the star formation history indicators, EW(H$\alpha$), EW(H$\delta$$_A$) and D$_n(4000)$ on the color-mass and SFR-mass diagrams to search for evidence of possible evolution among the adv-merger (U)LIRGs sample.
Distributions of EW(H$\alpha$), EW(H$\delta$$_A$) and D$_n(4000)$ in the [*NUV-r*]{} Color-Mass Diagram \[subsec:colormass\]
----------------------------------------------------------------------------------------------------------------------------
![SDSS spectra for 2 example adv-merger (U)LIRGs with weak H$\alpha$ emissions, strong H$\delta_{A}$ absorptions and deep 4000$\rm \AA$ breaks.[]{data-label="spectra.eps"}](fig2.eps){width="\textwidth"}
Although a large fraction of our sample adv-merger (U)LIRGs are undergoing massive starbursts, the optical spectra of several galaxies in our sample already present clear K+A like features. Figure \[spectra.eps\] shows the example spectra of two such galaxies, from which we can clearly see the K+A like spectroscopic signatures, i.e., weak H$\alpha$ emission, strong H$\delta$$_A$ absorption line and deep 4000[Å]{} break. These features indicate that a small fraction of adv-merger (U)LIRGs are more evolved than the others under the assumption that the adv-merger (U)LIRGs have similar star formation histories.
![Dust-corrected [*NUV-r*]{} color-mass diagrams for 69 sample adv-merger (U)LIRGs with both [*NUV*]{} photometry from the [*GALEX*]{} and spectra from the SDSS. The contours represent 9 equally spaced levels between 10% and 90% number densities of 46407 control sample galaxies with [*NUV*]{} photometry. Black solid lines show the boundaries of the green valley (Equations (7) and (8) from Guo et al. (2016b)). Dust extinction corrected EW(H$\alpha$) (left), EW(H$\delta_{A}$) (middle) and D$_n(4000)$ (right) distributions are overlaid. The color scales are shown to the right of the respective panel. []{data-label="cmd.eps"}](fig3.eps){width="\textwidth"}
To get a visual view of the evolution, we use different colors to represent the EW(H$\alpha$), EW(H$\delta$$_A$) and D$_n(4000)$ values on the [*NUV-r*]{} color-mass diagram, shown in the left, middle and right panels of Figure \[cmd.eps\], respectively. It is clear from the middle panel of Figure \[cmd.eps\] that the EW(H$\delta$$_A$) values for most sample galaxies are larger than 4[Å]{}, indicating A-type star features from a burst in the recent past. In the left and right panels of Figure \[cmd.eps\], there are obvious trends that the redder the galaxies are on the [*NUV-r*]{} color-mass diagram, the smaller the EW(H$\alpha$) and the larger the D$_n(4000)$ are. The reddest adv-merger (U)LIRGs close to the green valley mostly have D$_n(4000)$$>1.4$. We note that such trends are not caused by the correlation between the stellar mass and the EW(H$\alpha$) or D$_n(4000)$, as manifested by the apparent trends shown at a given stellar mass. Taking into account the evolutionary picture as traced by the D$_n(4000)$ and EW(H$\delta$$_A$) for an instantaneous starburst by Kauffmann et al. (2003b), the adv-merger (U)LIRGs with redder [*NUV-r*]{} colors are likely close to the end of starbursts or at the initial stage of post-starbursts according to the definitions by Goto (2007), Martin et al. (2007) and Yesuf et al. (2014). Therefore, from the distributions of the EW(H$\alpha$), EW(H$\delta$$_A$) and D$_n(4000)$ values on the [*NUV-r*]{} diagram in Figure \[cmd.eps\], we can clearly see the evolutionary trend from starbursts with blue [*NUV-r*]{} colors ($\sim 0.5$ mag) to relatively older populations with redder colors ($\sim 3$ mag) that formed in recent starbursts within the past $\sim 1$ Gyr, which is consistent with the gas exhausting time-scales as estimated in Guo et al. (2016b). This adds additional evidence for the fast evolutionary pathway to the conclusion drawn by Guo et al. (2016b).
Distributions of EW(H$\alpha$), EW(H$\delta$$_A$) and D$_n(4000)$ in the SFR-$M_{\ast}$ Diagram
-----------------------------------------------------------------------------------------------
It has been accepted that the SFR-$M_*$ relation reflects the star formation modes: galaxies on the star formation main sequence form stars in a relatively steady mode, while galaxies localized above the main sequence line form stars in a starburst mode (Rodighiero et al. 2011; Hung et al. 2013). From the analysis for adv-merger (U)LIRGs, Guo et al. (2016b) found that about three-fourths of adv-merger (U)LIRGs are located above the $1\, \sigma$ line of the local star-forming galaxy main sequence and one-fourth of the sample galaxies lie on the main sequence. Therefore, it is essential to investigate the reasons responsible for the difference between the galaxies above the main sequence and those on the main sequence. On the other hand, it is a complement to Section \[subsec:colormass\] to witness the evolution of the stellar populations on the SFR-$M_{\ast}$ diagram.
![SFR(H$\alpha$) vs. stellar mass relations for 63 star-forming and composite sample adv-merger (U)LIRGs. The black solid line indicates the local star-forming main sequence (MS) (Equation (4) from Guo et al. (2016b)) with 1 $\sigma$ dispersions (0.3 dex) shown as dashed lines. Dust extinction corrected EW(H$\alpha$) (left), EW(H$\delta_{A}$) (middle) and D$_n(4000)$ (right) distributions are overlaid. The color scales are shown to the right of the respective panel.[]{data-label="ms.eps"}](fig4.eps){width="\textwidth"}
Similar to Figure \[cmd.eps\], different colors are used to denote the EW(H$\alpha$), EW(H$\delta$$_A$) and D$_n(4000)$ values for each sample galaxy on the SFR-$M_{\ast}$ diagram, shown in the left, middle and right panels of Figure \[ms.eps\], respectively. It is clear from the left and right panels of Figure \[ms.eps\] that at a given stellar mass, there are clear trends that as the SFR decreases, the EW(H$\alpha$) decreases and the D$_n(4000)$ increases. In addition, the sample galaxies on the main sequence have relatively smaller EW(H$\delta$$_A$) on average. These results suggest that the adv-merger (U)LIRGs localized on the main sequence are more evolved counterparts of those above the main sequence, and they have started to fade from the starburst stage, as indicated by their smaller EW(H$\alpha$) and larger D$_n(4000)$. According to the starburst age estimated by D$_n(4000)$ (Kauffmann et al. 2003b), these adv-merger (U)LIRGs are going to leave the main sequence towards low SFRs in the SFR-$M_{\ast}$ diagram in $\sim 1$ Gyr. The evolution of the spectral indices on the SFR-$M_{\ast}$ diagram depicts a coherent picture with that exhibited by Figure \[cmd.eps\].
SUMMARY
=======
With the aim of studying the fast quenching process from the blue cloud to the red sequence, we retrieve 79 local adv-merger (U)LIRGs with SDSS spectroscopic observations from Guo et al. (2016b) and investigate the distributions of EW(H$\alpha$), EW(H$\delta$$_A$) and D$_n(4000)$ on the [*NUV-r*]{} color-mass and SFR-$M_{\ast}$ diagrams. The distributions of EW(H$\alpha$), EW(H$\delta$$_A$) and D$_n(4000)$ values on the [*NUV-r*]{} color-mass and SFR-$M_{\ast}$ diagrams of adv-merger (U)LIRGs show obvious evolutionary trends: at a given stellar mass, galaxies with redder [*NUV-r*]{} colors or lower SFRs have smaller EW(H$\alpha$) and larger D$_n(4000)$, indicating more evolved stellar populations. Moreover, most of the reddest adv-merger (U)LIRGs close to the green valley have D$_n(4000)$$>1.4$ and the majority of our sample galaxies have EW(H$\delta$$_A$) $> 4$Å that is a sign for A-type star features from a recent starburst in the past $0.1-1$ Gyr. These results suggest that some of the adv-merger (U)LIRGs have already shown signs of fading from the starburst stage and the reddest adv-merger (U)LIRGs in [*NUV-r*]{} color are likely at the initial stage of post-starbursts with age of $\sim 1$ Gyr. This time-scale coincides with the gas depletion time-scales derived by Guo et al. (2016b). Therefore, our results add additional support for the fast evolutionary track proposed by several groups (e.g. Muzzin et al. 2013; Barro et al. 2013, 2014a, 2014b; Marchesini et al. 2014; Schawinski et al. 2014; Belli et al. 2015; Wellons et al. 2015).
We would like to thank the anonymous referee for helpful comments that improved the manuscript. This project is supported by the NSF of China 11373027, 10973011 and 11003015. The Project-sponsored by SRF for ROCS, SEM. Funding for the creation and distribution of the SDSS Archive has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Aeronautics and Space Administration, the National Science Foundation, the U.S. Department of Energy, the Japanese Monbukagakusho, and the Max Planck Society. The SDSS Web site is http://www.sdss.org/. The SDSS is managed by the Astrophysical Research Consortium (ARC) for the Participating Institutions. The Participating Institutions are The University of Chicago, Fermilab, the Institute for Advanced Study, the Japan Participation Group, The Johns Hopkins University, the Korean Scientist Group, Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, University of Pittsburgh, Princeton University, the United States Naval Observatory, and the University of Washington. Some of the data presented in this paper were obtained from the Mikulski Archive for Space Telescopes (MAST). STScI is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. Support for MAST for non-HST data is provided by the NASA Office of Space Science via grant NNX09AF08G and by other grants and contracts.
Abazajian, K. N., Adelman-McCarthy, J. K., Ag[ü]{}eros, M. A., et al. 2009, , 182, 543 Baldry, I. K., Balogh, M. L., Bower, R. G., et al. 2006, , 373, 469 Baldry, I. K., Glazebrook, K., Brinkmann, J., et al. 2004, , 600, 681 Baldwin, J. A., Phillips, M. M., & Terlevich, R. 1981, , 93, 5 Barro, G., Faber, S. M., P[é]{}rez-Gonz[á]{}lez, P. G., et al. 2014a, , 791, 52 Barro, G., Faber, S. M., P[é]{}rez-Gonz[á]{}lez, P. G., et al. 2013, , 765, 104 Barro, G., Trump, J. R., Koo, D. C., et al. 2014b, , 795, 145 Bell, E. F., Wolf, C., Meisenheimer, K., et al. 2004, , 608, 752 Belli, S., Newman, A. B., & Ellis, R. S. 2015, , 799, 206 Calzetti, D., Armus, L., Bohlin, R. C., et al. 2000, , 533, 682 Chen, Y., Lowenthal, J. D., & Yun, M. S. 2010, , 712, 1385 Faber, S. M., Willmer, C. N. A., Wolf, C., et al. 2007, , 665, 265 Fukugita, M., Nakamura, O., Turner, E. L., Helmboldt, J., & Nichol, R. C. 2004, , 601, L127 Guo, R., Hao, C.-N., Xia, X.-Y., Wei, P., & Guo, X. 2016a, , in press (arXiv:1603.04988) Guo, R., Hao, C.-N., Xia, X. Y., Mao, S. & Shi, Y. 2016b, , in press (arXiv:1604.07122) Goto, T. 2007, , 381, 187 Hopkins, A. M., Miller, C. J., Nichol, R. C., et al. 2003, , 599, 971 Hung, C.-L., Sanders, D. B., Casey, C. M., et al. 2013, , 778, 129 Ilbert, O., Salvato, M., Le Floc’h, E., et al. 2010, , 709, 644 Kauffmann, G., Heckman, T. M., Tremonti, C., et al. 2003a, , 346, 1055 Kauffmann, G., Heckman, T. M., White, S. D. M., et al. 2003b, , 341, 33 Kauffmann, G., Heckman, T. M., White, S. D. M., et al. 2003c, , 341, 54 Kaviraj, S. 2009, , 394, 1167 Kennicutt, R. C., Jr. 1998, , 36, 189 Kewley, L. J., Dopita, M. A., Sutherland, R. S., Heisler, C. A., & Trevena, J. 2001, , 556, 121 Kilerci Eser, E., Goto, T., & Doi, Y. 2014, , 797, 54 Kim, D.-C., & Sanders, D. B. 1998, , 119, 41 Knapen, J. H., & James, P. A. 2009, , 698, 1437 Kriek, M., van Dokkum, P. G., Franx, M., Illingworth, G. D., & Magee, D. K. 2009, , 705, L71 Kroupa, P. 2001, , 322, 231 Lupton, R., Blanton, M. R., Fekete, G., et al. 2004, , 116, 133 Marchesini, D., Muzzin, A., Stefanon, M., et al. 2014, , 794, 65 Martin, D. C., Fanson, J., Schiminovich, D., et al. 2005, , 619, L1 Martin, D. C., Wyder, T. K., Schiminovich, D., et al. 2007, , 173, 342 Muzzin, A., Marchesini, D., Stefanon, M., et al. 2013, , 206, 8 Oh, K., Sarzi, M., Schawinski, K., & Yi, S. K. 2011, , 195, 13 Rodighiero, G., Daddi, E., Baronchelli, I., et al. 2011, , 739, L40 Sanders, D. B., & Mirabel, I. F. 1996, , 34, 749 Saunders, W., Sutherland, W. J., Maddox, S. J., et al. 2000, , 317, 55 Schawinski, K., Urry, C. M., Simmons, B. D., et al. 2014, , 440, 889 Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, , 500, 525 van Dokkum, P. G., Franx, M., Kriek, M., et al. 2008, , 677, L5 Wang, J. L., Xia, X. Y., Mao, S., et al. 2006, , 649, 722 Wellons, S., Torrey, P., Ma, C.-P., et al. 2015, , 449, 361 Wright, E. L., Eisenhardt, P. R. M., Mainzer, A. K., et al. 2010, , 140, 1868-1881 Yesuf, H. M., Faber, S. M., Trump, J. R., et al. 2014, , 792, 84 York, D. G., Adelman, J., Anderson, J. E., Jr., et al. 2000, , 120, 1579
\[lastpage\]
[^1]: $\LIR$ is the integrated infrared luminosity between 8-1000 $\mu$m.
[^2]: The magnitude restriction was not performed for the 1 Jy ULIRGs since they have obvious interacting or merging morphologies.
[^3]: http://gem.yonsei.ac.kr/$\sim$ksoh/wordpress
[^4]: http://www.mpa-garching.mpg.de/SDSS
[^5]: The spectral classification were performed according to the BPT diagram (Baldwin et al. 1981; Kauffmann et al. 2003a; Kewley et al. 2001).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is conjectured that every closed manifold admitting an Anosov diffeomorphism is, up to homeomorphism, finitely covered by a nilmanifold. Motivated by this conjecture, an important problem is to determine which nilmanifolds admit an Anosov diffeomorphism. The main theorem of this article gives a general method for constructing Anosov diffeomorphisms on nilmanifolds. As a consequence, we give new examples which were overlooked in a corollary of the classification of low-dimensional nilmanifolds with Anosov diffeomorphisms and a correction to this statement is proven. This method also answers some open questions about the existence of Anosov diffeomorphisms which are minimal in some sense.'
author:
- |
Jonas Deré[^1]\
KU Leuven Kulak, E. Sabbelaan 53, BE-8500 Kortrijk, Belgium
bibliography:
- 'G:/algebra/ref.bib'
title: '**A new method for constructing Anosov Lie algebras**'
---
A diffeomorphism $f: M \to M$ on a closed manifold is called Anosov if the tangent bundle splits continuously into two $df$-invariant vector bundles $E^s$ and $E^u$, such that $df$ is contracting on $E^s$ and expanding on $E^u$. The standard examples of Anosov diffeomorphisms are induced by unimodular hyperbolic automorphisms of ${\mathbb{R}}^n$ on the $n$-torus ${\mathbb{T}}^n$, considered as quotient space $\faktor{{\mathbb{R}}^n}{{\mathbb{Z}}^n}$. The first non-toral example was given by S. Smale in his paper [@smal67-1] where he also raised the question of classifying all closed manifolds admitting an Anosov diffeomorphism.
It is conjectured that every Anosov diffeomorphism is topologically conjugate to an affine infra-nilmanifold automorphism. A proof of this conjecture would thus imply that every closed manifold admitting an Anosov diffeomorphism is homeomorphic to an infra-nilmanifold admitting an Anosov diffeomorphism. More details about the definition and the conjecture can be found in [@deki11-1; @smal67-1]. In recent years, there has been quite some research about classifying infra-nilmanifolds supporting an Anosov diffeomorphism as well as constructing new examples with specific properties.
The conjecture motivates the study of infra-nilmanifolds supporting an Anosov diffeomorphism and in this paper we give a general method for constructing examples in the case of nilmanifolds. The existence of an Anosov diffeomorphism on a nilmanifold is equivalent to the existence of an Anosov automorphism on the corresponding nilpotent Lie algebra, i.e. a hyperbolic and integer-like automorphism (see Section \[intro\] for more details). One way to construct Anosov automorphisms on Lie algebras is to start from an automorphism of a free nilpotent Lie algebra and take a quotient by an ideal which is invariant under the automorphism, see for example the papers [@dm05-1; @dd03-2; @frie81-1; @payn09-1]. Another type of construction, as was used in [@laur03-1; @lw08-1; @lw09-1; @mw07-1], starts from a nilpotent Lie algebra ${\mathfrak{n}}^E$ over some field extension $E$ of ${\mathbb{Q}}$ with a hyperbolic automorphism and then gives a rational form of the Lie algebra which is invariant under the automorphism. The hard step in the latter case is to construct an explicit basis of ${\mathfrak{n}}^E$ such that the structure constants are rational and the matrix representation of the automorphism has rational entries. Note that in these papers, it is checked for each example separately that the given set of vectors is in fact a basis, that the structure constants do lie in ${\mathbb{Q}}$ and that the matrix of $f$ with respect to this basis has entries in ${\mathbb{Q}}$. The computations needed for these steps are rather cumbersome and somewhat time-consuming and these can be avoided by the main theorem of this paper.
The main theorem of this article generalizes this second method and states that given a hyperbolic and integer-like automorphism of a Lie algebra ${\mathfrak{n}}^E$ which behaves ‘nicely’ under the action of $\operatorname{Gal}(E,{\mathbb{Q}})$ (see Section \[secGC\] for the exact statement), there always exists a rational form of ${\mathfrak{n}}^E$ which is an Anosov Lie algebra. Not only does this technique significantly shorten the construction of the examples in [@laur03-1; @lw08-1; @lw09-1; @mw07-1], it also allows us to give new interesting examples of Anosov diffeomorphisms on nilmanifolds. For example, we can give a positive answer to some existence questions stated in [@laur03-1; @lw09-1].
The construction from the main theorem is also important for the classification of infra-nilmanifolds supporting an Anosov diffeomorphism. The only general classification is given by Porteous in [@port72-1] for flat manifolds and this was generalized to infra-nilmanifolds modeled on free nilpotent Lie groups in [@dd13-1], based on the work of [@dv09-1; @dv11-1]. For nilmanifolds, there is a classification of Anosov automorphisms on Lie algebras up to dimension $8$ in [@lw09-1]. One of the consequences of the main theorem is the construction of some new examples which were overlooked in this classification and we formulate a correction to this statement in Section \[verb\].
This article is built up as follows. In the first section, we give the main definitions about Anosov diffeomorphisms on nilmanifolds and state the correspondence between Anosov diffeomorphisms and hyperbolic, integer-like automorphisms of rational Lie algebras. The second part discusses the exact statement and proof of the main theorem. The last part describes the consequences of the main theorem, including the new examples answering the questions mentioned above.
Anosov diffeomorphisms on nilmanifolds {#intro}
======================================
We start with recalling the definitions and main results about Anosov diffeomorphisms on nilmanifolds. We also introduce the signature of an Anosov diffeomorphism and the type of a nilpotent Lie algebra, including some existence questions about Anosov automorphisms we will answer in this paper.
Let $N$ be a connected, simply connected, nilpotent Lie group with a lattice $\Gamma$, i.e. a discrete subgroup $\Gamma \subset N$ with compact quotient $\Gamma \backslash N$. The quotient space $\Gamma \backslash N$ is a closed manifold with fundamental group $\Gamma$ and is called a nilmanifold. The group $\Gamma$ is nilpotent, finitely generated and torsion-free and every group satisfying those three properties occurs as the fundamental group of a nilmanifold (see e.g. [@deki96-1]). Every Lie group automorphism $\alpha: N \to N$ induces a Lie algebra automorphism on the Lie algebra of $N$ and the eigenvalues of $\alpha$ are defined as the eigenvalues of this Lie algebra automorphism. If $\alpha \in \operatorname{Aut}(N)$ is a hyperbolic automorphism, meaning that it has no eigenvalues of absolute value $1$, and if $\alpha(\Gamma) = \Gamma$, then $\alpha$ induces an Anosov diffeomorphism on the nilmanifold $\Gamma \backslash N$ and the induced map is called a hyperbolic nilmanifold automorphism. For more details and the general definitions of infra-nilmanifolds and affine infra-nilmanifold automorphisms we refer to the paper [@deki11-1].
Let $\Gamma \backslash N$ be a nilmanifold and denote by $\Gamma^{\mathbb{Q}}$ the radicable hull of $\Gamma$ (also called rational Mal’cev completion). Corresponding to this radicable hull $\Gamma^{\mathbb{Q}}$, there is also a rational Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}$ and every finite dimensional rational nilpotent Lie algebra can be found in this way (see e.g. [@sega83-1]). Two lattices $\Gamma_1$ and $\Gamma_2$ of $N$ are called commensurable if there exist finite index subgroups $\Gamma_1^\prime \le \Gamma_1$ and $\Gamma_2^\prime \le \Gamma_2$ such that $\Gamma_1^\prime$ and $\Gamma_2^\prime$ are isomorphic. This is equivalent to the property that the Lie algebras of the radicable hulls $\Gamma_1^{\mathbb{Q}}$ and $\Gamma_2^{\mathbb{Q}}$ are isomorphic.
From [@deki99-1 Corollary 3.5.] and [@mann74-1] it follows that a nilmanifold $\Gamma \backslash N$ admits an Anosov diffeomorphism if and only if there exists a hyperbolic and integer-like automorphism of ${\mathfrak{n}}^{\mathbb{Q}}$. Recall that a matrix is called integer-like if its characteristic polynomial has coefficients in ${\mathbb{Z}}$ and its determinant has absolute value $1$. Since the eigenvalues and characteristic polynomial of a matrix are invariant under conjugation, the properties hyperbolic and integer-like are invariant under change of basis. An automorphism is hyperbolic respectively integer-like if the matrix representation of the automorphism for some (and thus for every) basis has the same property. This motivates the following definition:
An automorphism $\alpha \in \operatorname{Aut}({\mathfrak{n}}^E)$ of a Lie algebra ${\mathfrak{n}}^E$ over some field $E \subseteq {\mathbb{C}}$ is called *Anosov* if it is hyperbolic and integer-like. A rational Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}$ with an Anosov automorphism is called *Anosov*.
The superscripts $^{\mathbb{Q}}$ and $^E$ indicate over which field we are working. Thus a classification of all nilmanifolds admitting an Anosov diffeomorphism is equivalent to a classification of all Anosov Lie algebras. A theorem by Jacobson (see [@jaco55-1]) shows that every Anosov Lie algebra is necessarily nilpotent.
Theorem \[main\] will give us a very general way of constructing Anosov Lie algebras. By using the low-dimensional classification of Anosov Lie algebras, we can describe the isomorphism class for some of those Lie algebras in Section \[verb\], but it will be hard to do the same in general. On the other hand, there will be some properties of the Lie algebra or the Anosov automorphisms which follow immediately from the construction, e.g. the signature of the Anosov automorphism and the type of the Lie algebra, properties which we introduce below.
The definition of the signature of an Anosov diffeomorphism $f: M \to M$ makes use of the $df$-invariant splitting $TM = E^s \oplus E^u$:
The signature $\operatorname{sgn}(f)$ of an Anosov diffeomorphism $f: M \to M$ is defined as the set $\{\dim_{\mathbb{R}}E^s, \dim_{\mathbb{R}}E^u\}$.
We define the signature as a set rather than as an ordered pair since the order of the elements does not play a role. Indeed, the inverse of an Anosov diffeomorphism is again Anosov with the bundles $E^s$ and $E^u$ interchanged. For Anosov automorphisms $f: {\mathfrak{n}}^{\mathbb{Q}}\to {\mathfrak{n}}^{\mathbb{Q}}$ on a rational Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}$, the signature is defined as the set $\{p,q\}$ where $p$ is the number of eigenvalues with absolute value $<1$ and $q$ the number of eigenvalues of absolute value $> 1$. Note that the proof of [@deki99-1 Corollary 3.5.] implies that there exists an Anosov diffeomorphism $f$ with signature $\operatorname{sgn}(f)$ if and only if there exists an Anosov automorphisms with signature $\operatorname{sgn}(f)$ on the corresponding Lie algebra.
The following question is stated in [@lw09-1]:
\[q1\] Does there exist an Anosov automorphism on a non-abelian Lie algebra with signature $\{2,k\}$ for some $k \in {\mathbb{N}}$?
In Section \[sign\] it is shown that $\min(\operatorname{sgn}(f)) \geq c$ where $c$ is the nilpotency class of the Lie algebra. So a more general question is the existence of Anosov automorphisms which attain this lower bound for the signature. The main theorem gives us a positive answer in Section \[sign\].
As a consequence of a low-dimensional classification of Anosov automorphisms, the following is stated as a corollary in [@lw09-1]:
\[corlaur\] Let $\Gamma \backslash N$ be a nilmanifold of dimension $\leq 8$ which admits an Anosov diffeomorphism. Then $\Gamma \backslash N$ is a torus or the dimension is $6$ or $8$ and the signature is $\{3,3\}$ or $\{4,4\}$ respectively.
Examples illustrating that some cases were overlooked in this corollary are given in Section \[verb\] with signature $\{2,4\}$ and $\{3,5\}$. A complete list of Anosov Lie algebras admitting such signatures is given and thus this section forms a correction to the result of [@lw09-1].
If ${\mathfrak{n}}^E$ is a Lie algebra over a field $E$, then its lower central series $\gamma_i({\mathfrak{n}}^E)$ is recursively defined by $\gamma_1({\mathfrak{n}}^E) = {\mathfrak{n}}^E$ and $\gamma_i({\mathfrak{n}}^E) = [{\mathfrak{n}}^E, \gamma_{i-1}({\mathfrak{n}}^E) ]$ for all $i \geq 2$. The Lie algebra ${\mathfrak{n}}^E$ is said to be $c$-step nilpotent (or to have nilpotency class $c$) if $\gamma_{c+1}({\mathfrak{n}}^E) = 0$ and $\gamma_{c}({\mathfrak{n}}^E) \neq 0$. The type of a Lie algebra gives us information about the quotients of the lower central series:
The type of a nilpotent Lie algebra ${\mathfrak{n}}^E$ of nilpotency class $c$ is defined as the $c$-tuple $(n_1,\ldots,n_c)$, where $n_i = \dim_E \faktor{\gamma_{i}({\mathfrak{n}}^E)}{\gamma_{i+1}({\mathfrak{n}}^E)}$.
If ${\mathfrak{n}}^{\mathbb{Q}}$ is an Anosov Lie algebra of type $(3, n_2, \ldots,n_c)$, then [@payn09-1 Theorem 1.3.] states that $3 \mid n_i$ for every $i \in \{2, \ldots, c\}$. Combining this result with [@lw08-1 Proposition 2.3.] we get that every Anosov Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}$ of type $(n_1,\ldots,n_c)$ satisfies one of the following:
(i) ${\mathfrak{n}}^{\mathbb{Q}}$ is abelian or
(ii) $n_1 \geq 4$ and $n_i \geq 2$ for all $i \in \{2, \ldots, c\}$ or
(iii) $n_1 = n_2 = 3$ and $3 \mid n_i $ for all $i \in \{3, \ldots, c\}$.
In [@laur03-1] it was proved that the lower bound of (ii) occurs for every $c$ and the main theorem shows that also the lower bound of (iii) is attained for every $c$, answering the following question of [@lw08-1]:
\[q2\] Does there exist a $c$-step Lie algebra of type $(3, \ldots, 3)$ of nilpotency class $c \geq 3$ which is Anosov?
An Anosov Lie algebra which attains one of the lower bounds (ii) or (iii) for the type will be called of minimal type. In fact, in Section \[sign\] we will show that Question \[q2\] has a positive answer even if we replace $3$ by any integer $n > 2$.
To end this first section, we give a short summary of the known results about nilmanifolds supporting an Anosov diffeomorphism. In a few cases, a complete classification is given, for example in low-dimensional cases (see [@lw08-1]) and in the case of free nilpotent Lie algebras (see [@dani03-1; @dd13-1]). In [@dm05-1; @dd03-2; @frie81-1; @payn09-1], some new examples of Anosov diffeomorphisms on nilmanifolds were constructed by taking quotients of free nilpotent Lie algebras. This method is based on the work of L. Auslander and J. Scheuneman in [@as70-1]. All other examples are explicitly constructed and the main result of this article is to generalize this construction in Theorem \[main\]. This theorem avoids all the computations and allows us to construct more complicated examples of Anosov automorphisms. It seems plausible that all questions about the existence of Anosov automorphisms of specific signature on a Lie algebra of specific type can be answered in this way. The disadvantage of this method is that in general it is hard to describe the rational Lie algebra in terms of a basis and relations on this basis.
Construction of Anosov Lie algebras {#con}
===================================
The examples of Anosov Lie algebras in [@laur03-1; @lw09-1; @mw07-1] are constructed from a nilpotent Lie algebra over some field $E$ given by its decomposition into eigenspaces of an Anosov automorphism. An explicit basis is then constructed which is ‘symmetric’ under the action of the Galois group of $E$. Instead of working with a basis, we rather use this ‘symmetric’ property of the basis as the definition of a rational form of the Lie algebra. We will make this statement precise in Section \[secGC\]. For automorphisms, a similar ‘symmetric’ property can be defined such that the automorphism induces an automorphism on the rational form. In this way, no explicit calculations are needed and Theorem \[main\] allows us to easily show the existence of several new Anosov Lie algebras. This method further generalizes the work of [@payn09-1] which starts from polynomials instead of field extensions.
First let us fix some notations for this section. Let $E$ be a subfield of ${\mathbb{C}}$ such that $E$ has finite degree over ${\mathbb{Q}}$. If $n$ is the extension degree of $E \supseteq {\mathbb{Q}}$, then $E$ is called a Galois extension of ${\mathbb{Q}}$ if the group $$\operatorname{Gal}(E,{\mathbb{Q}}) = \{ \sigma: E \to E \mid \sigma \text{ is a field automorphism} \}$$ has order $n$. We will always assume that our field extensions are Galois.
A vector space $V$ over $E$ will be denoted by $V^E$ to indicate over which field we are working, as we already did above in Section \[intro\]. All the vector spaces and Lie algebras we consider are finite dimensional. If $F \supseteq E$ is a field extension, then we can consider the vector space $F \otimes_E V^E$ which we will denote as $V^F$. If $G$ is a finite group and $\rho: G \to \operatorname{GL}(V^E)$ is a representation, then there also exists a representation $\rho^F: G \to \operatorname{GL}(V^F)$ by extending the scalars, i.e. by considering $\operatorname{GL}(V^E)$ as a subgroup of $\operatorname{GL}(V^F)$. The same notations will be used for Lie algebras, e.g. a Lie algebra over the rationals ${\mathbb{Q}}$ will be denoted by ${\mathfrak{n}}^{\mathbb{Q}}$. We say that two representations $\rho_1: G \to \operatorname{GL}(V^E)$ and $\rho_2: G \to \operatorname{GL}(W^E)$ are $E$-equivalent if there exists an isomorphism $\varphi: V^E \to W^E$ such that $\rho_2(g) \circ \varphi = \varphi \circ \rho_1(g)$ for all $g \in G$.
A rational subspace $W^{\mathbb{Q}}\subseteq V^E$ is called a rational form if some (and hence every) basis of $W^{\mathbb{Q}}$ over ${\mathbb{Q}}$ is also a basis of $V^E$ over $E$. If ${\mathfrak{n}}^E$ is a Lie algebra over $E$, we call a rational subalgebra ${\mathfrak{m}}^{\mathbb{Q}}\subseteq {\mathfrak{n}}^E$ a rational form if it is a rational form seen as subspace of ${\mathfrak{n}}^E$ as vector space.
Construction of a rational form {#secGC}
-------------------------------
Instead of focusing on the basis of the rational form of a vector space, we focus on the defining property of the rational form as being ‘symmetric’ under the action of the Galois group. For this we first introduce the action of the Galois group on a vector space and define for each representation a rational subspace.
Consider the natural right action of $\operatorname{Gal}(E,{\mathbb{Q}})$ on the field $E$, given by $$\forall \sigma \in \operatorname{Gal}(E,{\mathbb{Q}}), \forall x \in E: x^\sigma = \sigma^{-1}(x).$$ By defining the action component-wise, there is also a natural right action on the vector space $E^m$. Note that the relations $$\label{relatie}
\begin{aligned}[c]
\left(\sigma(\lambda) x\right)^\sigma = \lambda x^\sigma
\end{aligned}
\text{ and }
\begin{aligned}[c]
\left(x + y\right)^\sigma = x^\sigma + y^\sigma
\end{aligned}$$ hold for all $x,y \in E^m$, $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$ and $\lambda \in E$. Let $\rho: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{GL}_m(E)$ be a representation, then it follows immediately that the subset defined as $$\begin{aligned}
V^{\mathbb{Q}}_\rho = \{ v \in E^m \mid \forall \sigma \in \operatorname{Gal}(E,{\mathbb{Q}}), \rho_\sigma(v) = v^\sigma \} \label{def}\end{aligned}$$ is a rational subspace of $E^m$. This subspace is already close to being a rational form of $E^m$ in the sense of the following lemma:
\[ratform\] If a set of vectors $v_1, \ldots, v_k$ of $V^{\mathbb{Q}}_\rho$ is linearly independent over ${\mathbb{Q}}$, then this set is also linearly independent over $E$ as vectors of $E^m$.
Assume that the lemma does not hold and take vectors $v_1, \ldots, v_k$ of $V^{\mathbb{Q}}_\rho$ with $k$ minimal which are linearly independent over ${\mathbb{Q}}$ but contradict the statement. This means that there exists $x_i \in E$ such that $$\begin{aligned}
\label{linind} \sum_{i=1}^k x_i v_i = 0.\end{aligned}$$ From the minimality of $k$ it follows that $x_1 \neq 0$ and thus by multiplying this equation by $x_1^{-1}$ we can assume that $x_1 = 1$. Since $k$ is minimal, it follows that the $x_i$ are the unique elements of $E$ such that $x_1 = 1$ and equation (\[linind\]) is true.
If we apply the map $\rho_\sigma$ to the equation, we get that $$0 = \rho_\sigma\left(\sum_{i=1}^k x_i v_i \right) = \sum_{i=1}^k x_i \rho_\sigma(v_i) = \sum_{i=1}^k x_i v_i^\sigma = \left(\sum_{i=1}^k \sigma(x_i) v_i\right)^\sigma$$ because of (\[relatie\]). We also have that $$\sum_{i=1}^k \sigma(x_i) v_i = 0.$$ Minimality of $k$ and the fact that $\sigma(x_1) = \sigma(1) = 1$ imply that $\sigma(x_i) = x_i$ for all $i$. Because this statement holds for all $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$, we conclude that the coefficients $x_i$ lie in ${\mathbb{Q}}$ and thus we get a contradiction since $v_i$ was a set of linearly independent vectors over ${\mathbb{Q}}$.
The lemma shows that the rational subspace $V^{\mathbb{Q}}_\rho$ is a rational form of $E^m$ if its dimension is maximal. This motivates the following definition:
A representation $\rho: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{GL}_m(E)$ is called *Galois compatible* (abbreviated as GC) if and only if $\dim_{\mathbb{Q}}(V^{\mathbb{Q}}_\rho) = m$. Equivalently, the representation $\rho
$ is GC if and only if $E \otimes V^{\mathbb{Q}}_\rho = E^m$ or if $V^{\mathbb{Q}}_\rho $ is a rational form of $E^m$.
The trivial representation is the easiest example of a GC representation, since $V^{\mathbb{Q}}_\rho = {\mathbb{Q}}^m$. A simple computation shows that also the regular representation is GC:
\[reg\] *Let $\rho$ be the regular representation of $\operatorname{Gal}(E,{\mathbb{Q}})$, i.e. take the vector space over $E$ spanned by the basis $\{v_\sigma \mid \sigma \in \operatorname{Gal}(E,{\mathbb{Q}})\}$ and the representation $\rho$ induced by the relations $\rho_\tau(v_\sigma) = v_{\tau\sigma}$ for all $\tau, \sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$. Every element $v$ of the rational vector space $V^{\mathbb{Q}}_\rho$ is given by $$v = \sum_{\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})} \sigma(x) v_\sigma$$ for some $x \in E$. This is a rational vector space of dimension $[E:{\mathbb{Q}}]$, which is also the order of the group $\operatorname{Gal}(E,{\mathbb{Q}})$ and thus the dimension of the regular representation $\rho$. This shows that $\rho$ is indeed GC.*
In a similar way it is easy to check that every representation which is given by permutation matrices is GC. The goal of the remaining part of this section is to show that every rational representation is GC:
\[GC\] Let $E$ be a Galois extension of ${\mathbb{Q}}$. Every representation $\operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{GL}_m({\mathbb{Q}})$ is Galois compatible.
To prove this statement we will first show that it holds for irreducible representations and then apply the following lemma:
\[som\] Let $\rho_i: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{GL}_{n_i}(E)$ with $i \in \{1,2\}$ be representations, then the following are equivalent:
1. $\rho_1$ and $\rho_2$ are Galois compatible.
2. $\rho_1 \oplus \rho_2: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{GL}_{n_1+n_2}(E)$ is Galois compatible.
Let $V_{\rho_1}^{\mathbb{Q}}$ and $V_{\rho_2}^{\mathbb{Q}}$ be the rational subspaces corresponding to $\rho_1$ and $\rho_2$, then by defition it follows that $$V_{\rho_1 \oplus \rho_2}^{\mathbb{Q}}= V_{\rho_1}^{\mathbb{Q}}\oplus V_{\rho_2}^{\mathbb{Q}}.$$ The statement of the lemma then easily follows by using Lemma \[ratform\].
The proof of the proposition is now immediate by using the previous lemma:
If two representation $\rho_1, \rho_2: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{GL}_m({\mathbb{Q}})$ are ${\mathbb{Q}}$-equivalent, then $\rho_1$ is GC if and only if $\rho_2$ is GC. So from Lemma \[som\] it follows that it is sufficient to prove the proposition for the ${\mathbb{Q}}$-irreducible representations of $\operatorname{Gal}(E,{\mathbb{Q}})$. Since every ${\mathbb{Q}}$-irreducible representation is a subrepresentation of the regular representation (see e.g. [@isaa76-1 Corollary 9.5.]), the statement follows from the example above.
Construction of automorphisms on rational forms of Lie algebras {#cons}
---------------------------------------------------------------
In the previous subsection we constructed rational forms of vector spaces $E^m$ from representations of the Galois group $\operatorname{Gal}(E,{\mathbb{Q}})$. In this subsection we use this technique to construct Lie algebras and automorphisms on them, allowing us to construct Anosov Lie algebras as well.
Let ${\mathfrak{n}}^{\mathbb{Q}}$ be a rational Lie algebra and $E$ a Galois extension of ${\mathbb{Q}}$. By taking a basis for ${\mathfrak{n}}^{\mathbb{Q}}\subseteq {\mathfrak{n}}^E$, the vector space ${\mathfrak{n}}^E$ is isomorphic to $E^m$ with $m = \dim_{\mathbb{Q}}({\mathfrak{n}}^{\mathbb{Q}})$ and thus the right action of $\operatorname{Gal}(E,{\mathbb{Q}})$ is defined on ${\mathfrak{n}}^E$. It is an exercise to check that this action does not depend on the choice of basis for ${\mathfrak{n}}^{\mathbb{Q}}$. The right action of $\operatorname{Gal}(E,{\mathbb{Q}})$ satisfies the property $$[X,Y]^\sigma = [ X^\sigma,Y^\sigma]$$ for all $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$, $X,Y \in {\mathfrak{n}}^E$. For every representation $$\rho: \operatorname{Gal}(E,{\mathbb{Q}})\to \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}}) \le \operatorname{GL}({\mathfrak{n}}^{\mathbb{Q}}),$$ with $\operatorname{GL}({\mathfrak{n}}^{\mathbb{Q}})$ the isomorphisms of ${\mathfrak{n}}^{\mathbb{Q}}$ as vector space, it follows from the definition that the rational subspace $V^{\mathbb{Q}}_\rho$ forms a subalgebra of ${\mathfrak{n}}^E$ and is thus a rational form of ${\mathfrak{n}}^E$ by Proposition \[GC\]. We will denote this subspace by ${\mathfrak{m}}^{\mathbb{Q}}_\rho$ to emphasize the fact that it is a subalgebra.
We now answer the question when a given automorphism $f: {\mathfrak{n}}^E \to {\mathfrak{n}}^E$ induces an automorphism on ${\mathfrak{m}}^{\mathbb{Q}}_\rho$. Note that $\operatorname{Gal}(E,{\mathbb{Q}})$ also acts on the right on $\operatorname{Aut}({\mathfrak{n}}^E)$, by defining for all $f \in \operatorname{Aut}({\mathfrak{n}}^E)$ the action as $$f^\sigma(v) = \left( f(v^{\sigma^{-1}}) \right)^\sigma.$$ By fixing a basis for ${\mathfrak{n}}^{\mathbb{Q}}$ and thus also for ${\mathfrak{n}}^E$, the matrix representation of $f^\sigma$ is given by applying $\sigma^{-1}$ to every entry of the matrix representation of $f$. Also every representation $\rho: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{Aut}({\mathfrak{n}}^E)$ induces a left action on $\operatorname{Aut}({\mathfrak{n}}^E)$ by conjugation. The automorphisms where this left action corresponds with the right action are exactly those that induce an automorphism on ${\mathfrak{m}}^{\mathbb{Q}}_\rho$:
\[auto\] Let ${\mathfrak{n}}^{\mathbb{Q}}$ be a rational Lie algebra and $\rho: G \to \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}})$ a representation. An element $f \in \operatorname{Aut}({\mathfrak{n}}^E)$ induces an automorphism on ${\mathfrak{m}}^{\mathbb{Q}}_\rho$ if and only if $f^\sigma = \rho_\sigma f \rho_{\sigma^{-1}}$ for all $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$.
First assume that $f$ satisfies the condition, i.e. that $f^\sigma = \rho_\sigma f \rho_{\sigma^{-1}}$ for $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$. For every $v \in {\mathfrak{m}}^{\mathbb{Q}}_\rho$ and $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$, we have that $$\begin{aligned}
\big(f(v)\big)^\sigma = f^\sigma (v^\sigma) = \rho_\sigma f \rho_{\sigma^{-1}} (\rho_\sigma(v)) = \rho_\sigma(f(v))\end{aligned}$$ and thus $f(v) \in {\mathfrak{m}}^{\mathbb{Q}}_\rho$ because $\sigma$ was taken arbitrary. Since ${\mathfrak{m}}^{\mathbb{Q}}_\rho$ is a rational form, the restriction of $f$ to ${\mathfrak{m}}^{\mathbb{Q}}_\rho$ is invertible and $f$ induces an automorphism of ${\mathfrak{m}}^{\mathbb{Q}}_\rho$.
For the other direction, fix $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$ and fix a basis $\{v_1, \ldots v_m\} \subseteq {\mathfrak{m}}^{\mathbb{Q}}_\rho$ for the Lie algebra ${\mathfrak{n}}^E$. Since $f$ induces an automorphism on ${\mathfrak{m}}^{\mathbb{Q}}_\rho$ we know that $f(v_i) \in {\mathfrak{m}}^{\mathbb{Q}}_\rho$ and thus that $\big(f(v_i)\big)^\sigma = \rho_\sigma(f(v_i))$. Since the vectors $\rho_{\sigma}(v_i)$ also form a basis for the Lie algebra ${\mathfrak{n}}^E$, it suffices to prove the relation $f^\sigma = \rho_\sigma f \rho_{\sigma^{-1}}$ on this basis. A computation shows that $$\begin{aligned}
\rho_\sigma f \rho_{\sigma^{-1}} (\rho_\sigma(v_i)) = \rho_\sigma (f(v_i)) = \big(f(v_i)\big)^\sigma = f^\sigma(v_i^\sigma) = f^\sigma(\rho_\sigma(v_i)) \end{aligned}$$ and thus the relation holds.
The main theorem of this article is the combination of Proposition \[GC\] and Lemma \[auto\]:
\[main\] Let ${\mathfrak{n}}^{\mathbb{Q}}$ be a rational Lie algebra and $\rho: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}})$ a representation. Suppose there exists a Lie algebra automorphism $f: {\mathfrak{n}}^E \to {\mathfrak{n}}^E$ such that $\rho_\sigma f \rho_{\sigma^{-1}} = f^\sigma$ for all $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$. Then there also exists a rational form ${\mathfrak{m}}^{\mathbb{Q}}\subseteq {\mathfrak{n}}^E$ such that $f$ induces an automorphism of ${\mathfrak{m}}^{\mathbb{Q}}$.
If all eigenvalues of $f$ are algebraic units of absolute value different from $1$, then ${\mathfrak{m}}^{\mathbb{Q}}$ is Anosov.
The only statement left to show is the last one. We claim that if $f$ is an automorphism of a rational Lie algebra with only algebraic units as eigenvalues, then $f$ is integer-like. Note that the coefficients of the characteristic polynomial of $f$ are formed by taking sums and products of the eigenvalues and thus all these coefficients are algebraic integers. Since these coefficients also lie in ${\mathbb{Q}}$, they are integers. The determinant of $f$ is equal to the product of all eigenvalues and therefore is an algebraic unit. Since the only algebraic units in ${\mathbb{Q}}$ are $\pm 1$, the claim now follows and this ends the proof.
Note that the type of the Lie algebra ${\mathfrak{m}}^{\mathbb{Q}}$ of the theorem is equal to the type of ${\mathfrak{n}}^{\mathbb{Q}}$ and is thus completely determined. The signature of the automorphism $f$ also does not change by restricting it to a rational form. But in general the Lie algebras ${\mathfrak{n}}^{\mathbb{Q}}$ and ${\mathfrak{m}}^{\mathbb{Q}}$ will not be isomorphic, so this theorem does not allow us to show that a specific Lie algebra is Anosov. In Section \[verb\] we determine the isomorphism class of the Lie algebra in a low-dimensional case by using the classification of low-dimensional Anosov algebras given in [@lw09-1].
In the special case where the Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}$ is given by the eigenspaces of an automorphism, the theorem becomes:
\[main2\] Let ${\mathfrak{n}}^{\mathbb{Q}}$ be a rational Lie algebra and assume there exists a decomposition of ${\mathfrak{n}}^{\mathbb{Q}}$ into subspaces $${\mathfrak{n}}^{\mathbb{Q}}= \bigoplus_{\lambda \in E} V_\lambda$$ such that $[V_\lambda,V_\mu] \subseteq V_{\lambda \mu}$. Let $\rho:\operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}})$ be a representation such that $\rho_\sigma(V_\lambda)= V_{\sigma(\lambda)}$ for all $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$. Then the linear map $f: {\mathfrak{n}}^E \to {\mathfrak{n}}^E$ given by $f(X) = \lambda X$ for all $X \in V_\lambda$ induces an automorphism on some rational form ${\mathfrak{m}}^{\mathbb{Q}}\subseteq {\mathfrak{n}}^E$.
If every $\lambda$ is an algebraic unit of absolute value different from $1$, then ${\mathfrak{m}}^{\mathbb{Q}}$ is Anosov.
We will make use of Theorem \[main\] to prove this. The condition on the Lie bracket implies that $f$ is a Lie algebra automorphism and the last condition is identical to the last condition of Theorem \[main\]. So it is left to show that $$\rho_\sigma f \rho_{\sigma^{-1}} = f^\sigma$$ for all $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$. It suffices to prove this relation for vectors $X \in V_\lambda$ for all possible $\lambda$. For such a vector, we have that $$\begin{aligned}
\rho_\sigma f \rho_{\sigma^{-1}} \left( X \right) = \rho_\sigma \Big(f \big( \rho_{\sigma^{-1}}\left( X \right) \big) \Big)= \sigma^{-1}(\lambda) \rho_\sigma \big( \rho_{\sigma^{-1}} \left( X \right) \big) = \sigma^{-1}\left( \lambda \right) X\end{aligned}$$ since $\rho_{\sigma^{-1}}(X) \in V_{\sigma^{-1}(\lambda)}$ and $$\begin{aligned}
f^\sigma \left(X \right) = \left( f(X) \right)^\sigma = \left( \lambda X \right)^\sigma = \sigma^{-1}(\lambda) X\end{aligned}$$ since $X^{\sigma^{-1}} = X$ and thus equality holds.
\[oned\] Corollary \[main2\] is the version of the main theorem we will use most of the times. In many examples, the spaces $V_\lambda$ will be one-dimensional and thus given by basis vectors $X_\lambda$ for $\lambda \in E$. In these examples, the Lie bracket is of the form $$[X_\lambda,X_\mu] = \pm X_{\lambda\mu}$$ and the representation $\rho$ is given by $\rho_\sigma(X_\lambda) = \pm X_{\sigma(\lambda)}$ for all $\lambda \in E, \sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$. To use the corollary, the only thing left to check is that $\rho$ is indeed a representation, since all other conditions are straightforward.
Consequences of Theorem \[main\] {#conseq}
================================
The main application of the theorem and corollary above lies in constructing Anosov Lie algebras of specific types and with Anosov automorphisms of specific signatures. We present here three different consequences of the main theorem, going from simplifying and correcting existing results to constructing new examples of minimal signature and minimal type.
Simplification of existing results
----------------------------------
By using our main theorem, all the examples of Anosov Lie algebras in [@laur03-1; @lw08-1; @lw09-1; @mw07-1] are now straightforward to construct. For instance, the examples in [@lw09-1; @mw07-1] follow from Remark \[oned\] since they start from a basis for the one-dimensional eigenspaces of a hyperbolic automorphism. As another example of the simplification we demonstrate how [@lw08-1 Theorem 3.2.] follows after a few lines.
Recall that a Lie algebra ${\mathfrak{n}}^E$ over a field $E$ is called graded if there exist subspaces ${\mathfrak{n}}_i^E \subseteq {\mathfrak{n}}^E$ for $i \in \{1, \ldots, k\}$ such that $$\begin{aligned}
&{\mathfrak{n}}^E = {\mathfrak{n}}_1^E \oplus {\mathfrak{n}}_2^E \oplus \ldots \oplus {\mathfrak{n}}_k^E \hspace{1mm} \text{ and} \\
&[{\mathfrak{n}}_i^E, {\mathfrak{n}}_j^E] \subseteq {\mathfrak{n}}_{i+j}^E.\end{aligned}$$ If ${\mathfrak{n}}^E$ is graded, then there exists an automorphism $f_\lambda: {\mathfrak{n}}^E \to {\mathfrak{n}}^E$ for every $\lambda \in E, \lambda \neq 0$, which is defined by $$f_\lambda (X) = \lambda^i X \hspace{2mm} \forall X \in {\mathfrak{n}}_i.$$ The proof of [@lw08-1 Theorem 3.2.] now follows immediately from the main theorem:
\[laur\] Let ${\mathfrak{n}}^{\mathbb{Q}}$ be a graded Lie algebra and consider the direct sum $$\tilde{{\mathfrak{n}}}^{\mathbb{Q}}= \underbrace{{\mathfrak{n}}^{\mathbb{Q}}\oplus \ldots \oplus {\mathfrak{n}}^{\mathbb{Q}}}_{m\ {\rm times}}$$ with $m \geq 2$. Then there exists a rational form of $\tilde{{\mathfrak{n}}}^{\mathbb{R}}$ which is Anosov.
Let $E$ be any real Galois extension of ${\mathbb{Q}}$ of degree $m$ with $\operatorname{Gal}(E,{\mathbb{Q}}) = \{ \sigma_1, \ldots, \sigma_m\}$. Every $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$ induces a permutation $\pi \in S_m$ via $$\sigma \sigma_i = \sigma_{\pi(i)} \hspace{1mm} \text{ for all } \hspace{1mm} i \in \{1, \ldots, m\},$$ and thus also an automorphism $\rho_\sigma \in \operatorname{Aut}(\tilde{{\mathfrak{n}}}^{\mathbb{Q}})$ by permuting the components of $\tilde{{\mathfrak{n}}}^{\mathbb{Q}}$ according to $\pi$. Note that $\rho$ is a representation $\operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{Aut}(\tilde{{\mathfrak{n}}}^{\mathbb{Q}})$.
Take an unit Pisot number $\lambda$ in $E$, which is possible since $m \geq 2$ (see Appendix for more details) and take the grading $ {\mathfrak{n}}^{\mathbb{Q}}= {\mathfrak{n}}_1^{\mathbb{Q}}\oplus \dots \oplus {\mathfrak{n}}_k^{\mathbb{Q}}$ for ${\mathfrak{n}}^{\mathbb{Q}}$. By writing the subspace ${\mathfrak{n}}^{\mathbb{Q}}_i$ of the $j$-th component of $\tilde{{\mathfrak{n}}}^{\mathbb{Q}}$ as $V_{\sigma_i(\lambda^j)}$, it is immediate from the construction of $\rho$ that all conditions of Corollary \[main2\] are satisfied. Since every $\sigma_i(\lambda^j)$ is an algebraic unit of absolute value different from $1$, it follows that $\tilde{{\mathfrak{n}}}^E$ (and therefore also $\tilde{{\mathfrak{n}}}^{\mathbb{R}}$) has a rational form which is Anosov.
Correction to a result of [@lw09-1] {#verb}
-----------------------------------
In [@lw09-1], a classification of all Anosov Lie algebras up to dimension $8$ is given. As one of the consequences, as we recalled in (False) Claim \[corlaur\], it is stated in [@lw09-1 Corollary 4.3.] that every Anosov diffeomorphism on a nilmanifold of dimension $\leq 8$ which is not a torus, has signature $\{3,3\}$ or $\{4,4\}$. There is not really a proof of this statement given in [@lw09-1] and in fact, by using Theorem \[main\] we can give new examples which were overlooked by the author.
First we recall some notions of [@lw09-1] about the Pfaffian form of a Lie algebra. Let ${\mathfrak{n}}^E$ be a Lie algebra over any field $E$ and take $\langle \hspace{1mm},\hspace{1mm} \rangle$ an inner product on ${\mathfrak{n}}^E$, i.e. a non-degenerate symmetric bilinear form. For every $Z \in {\mathfrak{n}}^E$, there exists a linear map $J_Z: {\mathfrak{n}}^E \to {\mathfrak{n}}^E$ defined by $$\langle J_Z X, Y \rangle = \langle [X,Y], Z \rangle \hspace{3mm} \forall X,Y \in {\mathfrak{n}}^E.$$ Note that $J_Z$ is skew-symmetric with respect to the inner product, meaning that $$\langle J_Z X, Y \rangle = - \langle X, J_Z Y \rangle$$ for all $X,Y \in {\mathfrak{n}}^E$. It is easy to check that an isomorphism $\alpha: {\mathfrak{n}}^E \to {\mathfrak{n}}^E$ is an automorphism of ${\mathfrak{n}}^E$ if and only if $$\begin{aligned}
\label{aut} \alpha^T J_Z \alpha = J_{\alpha^T (Z)}\end{aligned}$$ for all $Z \in {\mathfrak{n}}^E$, where $\alpha^T: {\mathfrak{n}}^E \to {\mathfrak{n}}^E$ is the adjoint map of $\alpha$, defined by $\langle X, \alpha^T( Y) \rangle = \langle \alpha (X), Y \rangle$.
Now assume that ${\mathfrak{n}}^E$ is a $2$-step nilpotent Lie algebra of type $(2 m,k)$ and take $V^E \subseteq {\mathfrak{n}}^E$ such that $V^E \oplus \gamma_2({\mathfrak{n}}^E) = {\mathfrak{n}}^E$ as a vector space. Take an inner product satisfying $\langle V^E , \gamma_2({\mathfrak{n}}^E)\rangle = 0$ and such that there exists an orthonormal basis for $V^E$ with respect to the inner product. Denote the vector space (in fact it is a Lie algebra) of all skew-symmetric endomorphisms of $V^E$ by ${\mathfrak{so}}(V^E)$. The construction above then induces a linear map $$\begin{aligned}
J : \gamma_2({\mathfrak{n}}^E) &\to {\mathfrak{so}}(V^E)\\ Z &\mapsto J_Z |_{V^E}.\end{aligned}$$ After taking an orthonormal basis for $V^E$, we can identify ${\mathfrak{so}}(V^E)$ with the skew-symmetric matrices. Recall that the Pfaffian on $V^E$ is the unique polynomial function $$\operatorname{Pf}: {\mathfrak{so}}(V^E) \to E$$ such that $\operatorname{Pf}(A)^2 = \det(A)$ for all $A \in {\mathfrak{so}}(V^E)$ and $\operatorname{Pf}(S) = 1$ for some fixed $S \in {\mathfrak{so}}(V^E)$ with $\det(S) = 1$ (where this last condition is needed to fix the sign). The Pfaffian obviously satisfies the relation $\operatorname{Pf}(A^T B A) = \det(A) \operatorname{Pf}(B)$ for all endomorphisms $A: V^E \to V^E$. The composition $$h = \operatorname{Pf}\circ J: \gamma_2({\mathfrak{n}}^E) \to E$$ is called the Pfaffian form of the 2-step nilpotent Lie algebra ${\mathfrak{n}}^E$. By taking another vector space $V^E$ or changing the inner product on ${\mathfrak{n}}^E$ or the basis of $V^E$, the Pfaffian form changes to a polynomial $$\begin{aligned}
\gamma_2({\mathfrak{n}}^E) &\to E \\ Z &\mapsto e h(\beta(Z))\end{aligned}$$ for some $e \in E \setminus \{ 0 \}$ and some isomorphism $\beta: \gamma_2({\mathfrak{n}}^E) \to \gamma_2({\mathfrak{n}}^E)$. Thus the Pfaffian form is uniquely determined by ${\mathfrak{n}}^E$ up to projective equivalence (see [@laur08-1 Proposition 2.4.] for the exact definition and proof of this statement). The polynomial $h$ is homogeneous of degree $m$ in $k$ variables and with coefficients in $E$. An automorphism of the Pfaffian form is an isomorphism $\beta: \gamma_2({\mathfrak{n}}^E) \to \gamma_2({\mathfrak{n}}^E)$ such that $h\circ \beta =h$.
A binary quadratic form over ${\mathbb{Q}}$ is a homogeneous polynomial of degree 2 in 2 variables, i.e. polynomials that can be written as $$h(X,Y) = a X^2 + b XY + c Y^2$$ with $a,b,c \in {\mathbb{Q}}$. The Pfaffian form of a rational Lie algebra of type $(4,2)$ is such a polynomial. The discriminant $\Delta(h)$ of $h$ is defined as $$\Delta(h) = b^2 - 4 ac.$$ In [@gss82-1] it is shown that two rational Lie algebras of type $(4,2)$ with Pfaffian forms $h_1$ and $h_2$ are isomorphic if and only if there exists $q \in {\mathbb{Q}}\setminus \{ 0 \}$ such that $\Delta(h_1) = q^2 \Delta(h_2)$. Given any $k \in {\mathbb{Z}}$, we can consider the Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}_k$ given by the basis $X_1, X_2, X_3, X_4, Z_1, Z_2$ and relations $$\begin{aligned}[c]
[X_1,X_3] &= Z_1 \\
[X_2,X_3] &= k Z_2
\end{aligned}
\hspace{2cm}
\begin{aligned}[c]
[X_1,X_4] &= Z_2 \\
[X_2,X_4] &= Z_1.
\end{aligned}$$ The Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}_k$ has Pfaffian form equal to $h(X,Y) = X^2 - k Y^2$ with discriminant $4k$. So every rational Lie algebra of type $(4,2)$ is isomorphic to ${\mathfrak{n}}^{\mathbb{Q}}_k$ for some $k \in {\mathbb{Z}}$ and ${\mathfrak{n}}^{\mathbb{Q}}_k$ is isomorphic to ${\mathfrak{n}}^{\mathbb{Q}}_{k^\prime}$ if and only if there exists a natural number $q > 0$ such that $k = q^2 k^\prime$.
The automorphisms $\beta \in \operatorname{SL}(2,{\mathbb{Z}})$ of a binary quadratic form $h(X,Y) = a X^2 + b XY + c Y^2$ are completely known and described e.g. in [@bv07-1 Theorem 2.5.5.]. Given a solution $x,y \in {\mathbb{Z}}$ of the Pell equation $$x^2 - \Delta(h) y^2 = 4,$$ the matrix $$U(x,y) = \begin{pmatrix} \frac{ x - yb}{2} & -cy \\ ay & \frac{x + yb}{2} \end{pmatrix} \in \operatorname{SL}(2,{\mathbb{Z}})$$ is an automorphism of the quadratic form $h$. The map $U$ is a bijection between the solutions of the Pell equation and the automorphisms of $h$ which lie in $\operatorname{SL}(2,{\mathbb{Z}})$. The eigenvalues of $U(x,y)$ are equal to $\frac{x \pm \sqrt{\Delta(h)} y}{2}$, so the field in which these eigenvalues lie gives us information about the discriminant of the form $h$.
Let ${\mathfrak{n}}^{\mathbb{Q}}$ be a rational Lie algebra of type $(4,2)$ and $\alpha \in \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}})$ an Anosov automorphism. By squaring $\alpha$ if necessary, we can also assume that $\det(\alpha) = 1$. From the equation $\alpha^T J_Z \alpha = J_{\alpha^T (Z)}$ (see (\[aut\])) and by applying the Pfaffian, we get that $h(\alpha^T (Z)) = h(Z)$ for all $Z \in V^{\mathbb{Q}}$. So $\alpha$ induces a hyperbolic and integer-like automorphism $\beta = \alpha^T |_{\gamma_2({\mathfrak{n}}^{\mathbb{Q}})}$ of the Pfaffian form $h$. The eigenvalues of $\beta$ (and thus also of $\alpha |_{\gamma_2({\mathfrak{n}}^{\mathbb{Q}})}$) lie in the field ${\mathbb{Q}}(\sqrt{\Delta(h)})$ and thus if we know these eigenvalues, we can determine the discriminant of the Pfaffian form of ${\mathfrak{n}}^{\mathbb{Q}}$ up to a square and therefore also the isomorphism class of ${\mathfrak{n}}^{\mathbb{Q}}$. This also implies that every Anosov Lie algebra of type $(4,2)$ is isomorphic to a Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}_k$ with $k$ a square free natural number $>1$ (since all other values of $k$ imply that the eigenvalues have absolute value $1$).
We now have all the tools to construct new Anosov automorphisms on the Anosov Lie algebras ${\mathfrak{n}}^{\mathbb{Q}}_k$:
\[count\] Let $k$ be a natural number with $k > 1$ and $k$ square free. Let ${\mathfrak{n}}^{\mathbb{Q}}_k$ be the Lie algebra with basis $X_1, X_2, X_3, X_4, Z_1, Z_2$ and relations $$\begin{aligned}[c]
[X_1,X_3] &= Z_1 \\
[X_2,X_3] &= k Z_2
\end{aligned}
\hspace{2cm}
\begin{aligned}[c]
[X_1,X_4] &= Z_2 \\
[X_2,X_4] &= Z_1.
\end{aligned}$$ Then there exists an Anosov automorphism $f$ on ${\mathfrak{n}}^{\mathbb{Q}}_k$ with $\operatorname{sgn}(f) = \{2,4\}$.
Fix the $k$ of the theorem and let $l$ be a different natural number with $l$ square free and $l>1$. Take $E = {\mathbb{Q}}(\sqrt{k},\sqrt{l})$, then $E$ is Galois over ${\mathbb{Q}}$ with $\operatorname{Gal}(E,{\mathbb{Q}}) = {\mathbb{Z}}_2 \oplus {\mathbb{Z}}_2$. Let $\tau \in \operatorname{Gal}(E,{\mathbb{Q}})$ be the unique element with $\tau(\sqrt{k}) = \sqrt{k}, \tau(\sqrt{l}) = -\sqrt{l}$ and take another $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$ such that $\operatorname{Gal}(E,{\mathbb{Q}}) = \{ 1, \sigma, \tau, \sigma \tau\}$. Take $\lambda_1 \in E$ an unit Pisot number as introduced in the Appendix. Write the Galois conjugates of $\lambda_1$ as $\lambda_1, \tau(\lambda_1) = \lambda_2, \sigma(\lambda_1) = \lambda_3, \sigma \tau(\lambda_1) = \lambda_4$
Consider the Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}$ with basis $X_{\lambda_1},X_{\lambda_2},X_{\lambda_3}, X_{\lambda_4}, Y_{\lambda_1 \lambda_2}, Y_{\lambda_3 \lambda_4}$ and Lie bracket given by $$\begin{aligned}
[X_{\lambda_1},X_{\lambda_2}] &= Y_{\lambda_1 \lambda_2} \\
[X_{\lambda_3}, X_{\lambda_4}] &= Y_{\lambda_3 \lambda_4}\end{aligned}$$ and all other brackets zero (so ${\mathfrak{n}}^{\mathbb{Q}}$ is isomorphic to the direct sum of two copies of the Heisenberg algebra of dimension $3$). Each of these basis vectors spans a $1$-dimensional subspace indexed by the same algebraic unit, corresponding to the decomposition in Corollary \[main2\]. Consider the representation $\rho: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}})$ induced by $\rho_\sigma(X_\lambda) = X_{\sigma(\lambda)}$ and $\rho_\tau(X_\lambda) = X_{\tau(\lambda)}$. A small computation shows that this is indeed a representation. By using Corollary \[main2\] we then get a rational form ${\mathfrak{m}}^{\mathbb{Q}}$ of ${\mathfrak{n}}^E$ with Anosov automorphism $f: {\mathfrak{m}}^{\mathbb{Q}}\to {\mathfrak{m}}^{\mathbb{Q}}$.
Note that $f$ has only two eigenvalues $>1$, namely $\lambda_1$ and $\lambda_1\lambda_2$ and thus $\operatorname{sgn}(f) = \{2,4\}$. Since $\tau(\lambda_1 \lambda_2) = \lambda_1 \lambda
_2$ and ${\mathbb{Q}}(\sqrt{k})$ is the unique subfield of $E$ fixed by $\tau$, the eigenvalue $\lambda_1 \lambda_2$ is in ${\mathbb{Q}}(\sqrt{k})$, showing that the Pfaffian form $h$ of ${\mathfrak{m}}^{\mathbb{Q}}$ satisfies $\Delta(h) = q^2 k$ for some $q \in {\mathbb{Q}}$. This shows that ${\mathfrak{m}}^{\mathbb{Q}}$ is isomorphic to ${\mathfrak{n}}^{\mathbb{Q}}_k$ and thus ${\mathfrak{n}}^{\mathbb{Q}}_k$ also has an Anosov automorphism of signature $\{2,4\}$.
It is also possible to start the proof from a Galois extension ${\mathbb{Q}}\subseteq E$ with $\operatorname{Gal}(E,{\mathbb{Q}}) = {\mathbb{Z}}_4$ as we show in the example below. By computing a basis for ${\mathfrak{m}}^{\mathbb{Q}}$ we give an explicit example which was overlooked in (False) Claim \[corlaur\]:
Start from the polynomial $$p(X) = X^4 -4X^3 - 4X^2+X+1,$$ which has $4$ distinct real roots, say $\lambda_1 > \lambda_2 > \lambda_3 > \lambda_4$. These roots satisfy $\lambda_1 > 1$ and $\vert \lambda_i \vert < 1$ for $i \in \{ 2, 3, 4 \}$, showing that $\lambda_1$ is an unit Pisot number. The Galois group of the field $E = {\mathbb{Q}}(\lambda_1)$ is isomorphic to ${\mathbb{Z}}_4$, which can be checked e.g. with GAP, see [@gap14-1]. A generator $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$ is given by $$\sigma(\lambda_1) = \lambda_2, \hspace{1mm} \sigma(\lambda_2) = \lambda_3, \hspace{1mm} \sigma(\lambda_3) = \lambda_4, \hspace{1mm} \sigma(\lambda_4) = \lambda_1.$$ Just as in the proof of Proposition \[count\], consider the Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}$ with basis $$X_{\lambda_1},X_{\lambda_2},X_{\lambda_3}, X_{\lambda_4}, Y_{\lambda_1 \lambda_3}, Y_{\lambda_2 \lambda_4}$$ and Lie bracket given by $$\begin{aligned}
[X_{\lambda_1},X_{\lambda_3}] &= Y_{\lambda_1 \lambda_3} \\
[X_{\lambda_2}, X_{\lambda_4}] &= Y_{\lambda_2 \lambda_4}\end{aligned}$$ Consider the representation $\rho: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}})$ induced by $\rho_\sigma(X_{\lambda_i}) = X_{\sigma(\lambda_i)}$. The main theorem guarantees us the existence of a rational form ${\mathfrak{m}}^{\mathbb{Q}}$ of ${\mathfrak{n}}^E$ which is Anosov, but we now compute this Lie algebra explicitly by giving a basis for ${\mathfrak{m}}^{\mathbb{Q}}$.
Consider the basis $U_1, U_2, U_3, U_4, V_1, V_2$ given by $$\begin{aligned}
U_i &=& \sum_{j=1}^4 \lambda_j ^{i-1} X_{\lambda_j } \\
V_i &=& \left( \lambda_3^{i} - \lambda_1^{i} \right) Y_{\lambda_1 \lambda_3} + \left(\lambda_4^{i} - \lambda_2^{i}\right) Y_{\lambda_2 \lambda_4}.\end{aligned}$$ To simplify the computations, we will use the notations $\lambda_0 = \lambda_4$ and $\lambda_5=\lambda_1$. The basis vectors $U_i$ satisfy $$U_i^\sigma = \sum_{j=1}^4 \left(\lambda_j^{i-1}\right)^\sigma X_{\lambda_{j}} = \sum_{j=1}^4 \sigma^{-1}\left(\lambda_j^{i-1}\right) X_{\lambda_{j}} = \sum_{j=1}^4 \lambda_{j-1}^{i-1} X_{\lambda_{j}}$$ and $$\rho_{\sigma}(U_i) = \sum_{j=1}^4 \lambda_j ^{i-1} \rho_\sigma\left(X_{\lambda_j }\right) = \sum_{j=1}^4 \lambda_j^{i-1} X_{\lambda_{j+1}} = \sum_{j=2}^5 \lambda_{j-1} ^{i-1} X_{\lambda_{j}}.$$ We conclude that $\rho_{\sigma}(U_i) = U_i^\sigma$ and similarly this equation also holds for the vectors $V_i$. This shows that the basis vectors $U_1, U_2, U_3, U_4, V_1, V_2$ satisfy the defining relation of the rational form given in equation (\[def\]) of Section \[secGC\] and thus they indeed span the rational form ${\mathfrak{m}}^{\mathbb{Q}}$ of Theorem \[main\]. The induced Anosov automorphism on ${\mathfrak{m}}^{\mathbb{Q}}$ guaranteed by Theorem \[main\] is given by the matrix $$\begin{aligned}
\begin{pmatrix}
0 & 0 & 0 & -1 & 0 & 0\\
1 & 0 & 0 & -1 & 0 & 0\\
0 & 1 & 0 & 4 & 0 & 0\\
0 & 0 & 1 & 4 & 0 & 0\\
0 & 0 & 0 & 0 & -\frac{1}{2} & -\frac{1}{2} \\[0.3em]
0 & 0 & 0 & 0 & -\frac{1}{2} & -\frac{5}{2}
\end{pmatrix}\end{aligned}$$ in the basis $\hspace{0.5 mm} U_1, U_2, U_3, U_4, V_1, V_2$. By using the matrix representation of this Anosov automorphism, one can compute the Lie bracket: $$\begin{aligned}[c]
\left[U_1, U_2 \right] &= V_1 \\
\left[U_2, U_3 \right] &= -\frac{1}{2} V_1 - \frac{1}{2} V_2 \\
\left[U_3, U_4 \right] &= \frac{1}{2} V_1 + \frac{3}{2} V_2
\end{aligned}
\hspace{2cm}
\begin{aligned}[c]
\left[U_1, U_3 \right] &= V_2 \\
\left[U_2, U_4 \right] &= -\frac{1}{2} V_1 - \frac{5}{2} V_2 \\
\left[U_1, U_4 \right] &= \frac{3}{2} V_1 + \frac{9}{2} V_2.
\end{aligned}$$ The discriminant of the Pfaffian form of ${\mathfrak{m}}^{\mathbb{Q}}$ is $\frac{5}{4}$, so ${\mathfrak{m}}^{\mathbb{Q}}$ is isomorphic to ${\mathfrak{n}}_5^{\mathbb{Q}}$. The characteristic polynomial of the Anosov automorphism restricted to $\gamma_2({\mathfrak{m}}^{\mathbb{Q}})$ is equal to $X^2+3X+1$ which has ${\mathbb{Q}}(\sqrt{5})$ as splitting field.
Of course, Proposition \[count\] also gives us examples of Anosov automorphisms $f: {\mathfrak{n}}^{\mathbb{Q}}_k \oplus {\mathbb{Q}}^2 \to {\mathfrak{n}}^{\mathbb{Q}}_k \oplus {\mathbb{Q}}^2$ with $\operatorname{sgn}(f) = \{3,5\}$. But by using the notion of Scheuneman duality (see [@sche67-1]) for $2$-step nilpotent Lie algebras, it is possible to give another class of Anosov automorphisms overlooked in (False) Claim \[corlaur\]. First we recall some details about this method as described in [@laur08-1].
Let $V^{\mathbb{Q}}$ be any vector space with an inner product and consider the standard inner product $B$ on ${\mathfrak{so}}(V^{\mathbb{Q}})$, given by $$B(Z_1,Z_2) = \operatorname{Tr}(Z_1^T Z_2) = - \operatorname{Tr}(Z_1 Z_2).$$ For every subspace $W^{\mathbb{Q}}$ of ${\mathfrak{so}}(V^{\mathbb{Q}})$, there exists a rational Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}= V^{\mathbb{Q}}\oplus W^{\mathbb{Q}}$ with Lie bracket $[\hspace{1mm},\hspace{1mm}]: V^{\mathbb{Q}}\times V^{\mathbb{Q}}\to W^{\mathbb{Q}}$ defined by $$B([X,Y], Z) = \langle Z(X),Y \rangle$$ for all $Z \in {\mathfrak{so}}(V^{\mathbb{Q}})$, $X,Y \in V^{\mathbb{Q}}$. This is a $2$-step nilpotent Lie algebra where the map $J: \gamma_2(V^{\mathbb{Q}}\oplus W^{\mathbb{Q}}) = W^{\mathbb{Q}}\to {\mathfrak{so}}(V^{\mathbb{Q}})$ is the inclusion. If we take the vector space $W^{\mathbb{Q}}= {\mathfrak{so}}(V^{\mathbb{Q}})$, then the result is the free $2$-step nilpotent Lie algebra on $V^{\mathbb{Q}}$. An isomorphism $\alpha: V^{\mathbb{Q}}\to V^{\mathbb{Q}}$ induces an automorphism $\bar{\alpha}: {\mathfrak{n}}^{\mathbb{Q}}\to {\mathfrak{n}}^{\mathbb{Q}}$ if and only if $\alpha^T Z \alpha \in W^{\mathbb{Q}}$ for all $Z \in W^{\mathbb{Q}}$. Let ${\mathfrak{n}}^{\mathbb{Q}}$ be a Lie algebra of type $(m,k)$ and consider the map $J: \gamma_2({\mathfrak{n}}^{\mathbb{Q}}) \to {\mathfrak{so}}(V^{\mathbb{Q}})$ as introduced above. Denote the image of $J$ as $W^{\mathbb{Q}}$, then it follows by definition that ${\mathfrak{n}}^{\mathbb{Q}}$ is isomorphic to the Lie algebra $V^{\mathbb{Q}}\oplus W^{\mathbb{Q}}$ of the previous paragraph. The dual of ${\mathfrak{n}}^{\mathbb{Q}}$ is then the Lie algebra $\tilde{n}^{\mathbb{Q}}= V^{\mathbb{Q}}\oplus \tilde{W}^{\mathbb{Q}}$ with Lie bracket as in the previous paragraph, where $\tilde{W}^{\mathbb{Q}}$ is the orthogonal complement of $W^{\mathbb{Q}}$ in ${\mathfrak{so}}(V^{\mathbb{Q}})$ relative to the inner product $B$ given above. The dual of the Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}_k$ is denoted by ${\mathfrak{h}}_k^{\mathbb{Q}}$. If $\alpha \in \operatorname{GL}(V^{\mathbb{Q}})$ induces an automorphism $\bar{\alpha}$ of ${\mathfrak{n}}^{\mathbb{Q}}$, then $\alpha^T$ induces an automorphism on $\tilde{{\mathfrak{n}}}^{\mathbb{Q}}$ since $\alpha \tilde{W}^{\mathbb{Q}}\alpha^T = \tilde{W}^{\mathbb{Q}}$ and this map is called the dual automorphism of $\bar{\alpha}$. The combined eigenvalues of $\bar{\alpha}$ and its dual on $\gamma_2({\mathfrak{n}}^{\mathbb{Q}})$ and $\gamma_2(\tilde{{\mathfrak{n}}}^{\mathbb{Q}})$ are equal to the eigenvalues of the map that $\alpha$ induces on $\gamma_2(V^{\mathbb{Q}}\oplus {\mathfrak{so}}(V^{\mathbb{Q}}))$ where $V^{\mathbb{Q}}\oplus {\mathfrak{so}}(V^{\mathbb{Q}})$ is the free $2$-step nilpotent Lie algebra on $V^{\mathbb{Q}}$.
The dual Lie algebra of ${\mathfrak{n}}^{\mathbb{Q}}_k$ is of type $(4,4)$ and denoted as ${\mathfrak{h}}^{\mathbb{Q}}_k$. The Lie algebra ${\mathfrak{h}}^{\mathbb{Q}}_k$ can also be described as the one with basis $X_1, X_2, X_3, X_4, Z_1, Z_2, Z_3, Z_4$ and relations $$\begin{aligned}[c]
[X_1,X_2] &= Z_1 \\
[X_1,X_3] &= Z_2 \\
[X_1,X_4] &= kZ_3
\end{aligned}
\hspace{2cm}
\begin{aligned}[c]
[X_2,X_3] &= -Z_3 \\
[X_2,X_4] &= -Z_2 \\
[X_3,X_4] &= Z_4,
\end{aligned}$$ see for example [@laur08-1]. From the Scheuneman duality, the following proposition is immediate:
For every $k \in {\mathbb{N}}$ with $k > 1$ and $k$ not a square, there exists an Anosov automorphism on ${\mathfrak{h}}^{\mathbb{Q}}_k$ with $\operatorname{sgn}(f) = \{3,5\}$. On ${\mathfrak{h}}^{\mathbb{Q}}_1$ every Anosov automorphism has signature $\{4,4\}$.
Note that the dual of an Anosov automorphism with $\operatorname{sgn}(f)= \{2,4\}$ is also Anosov with signature $\{3,5\}$ and vice versa. So the first part follows from the fact that ${\mathfrak{h}}^{\mathbb{Q}}_k$ is the dual of ${\mathfrak{n}}^{\mathbb{Q}}_k$. Also the second statement follows since the Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}_1$ is not Anosov.
For all other Lie algebras, (False) Claim \[corlaur\] is correct (and the arguments to prove it are the same as the ones used to prove the classification of Anosov Lie algebras up to dimension $8$). Also, the Lie algebra ${\mathfrak{h}}^{\mathbb{Q}}_k$ does not admit an Anosov automorphism of signature $(2,6)$, for example by using the same number theoretical arguments as in [@lw09-1]. So the combined results above determine completely for which Anosov Lie algebras (False) Claim \[corlaur\] is indeed false.
Note that the examples of Proposition \[count\] also answer Question \[q1\] about non-abelian examples of signature $\{2,q\}$ for some $q \in {\mathbb{N}}_0$. We give a more general approach to this question in the next section.
Anosov automorphisms of minimal signature and minimal type {#sign}
----------------------------------------------------------
In this subsection, we show how the main theorem can be used to construct Anosov automorphisms of minimal signature and Anosov Lie algebras of minimal type. These examples answer Questions \[q1\] and \[q2\] which we already mentioned in the introduction.
First we recall some basic properties of unit Pisot numbers. If $E$ is a real Galois extension of ${\mathbb{Q}}$ of degree $n$, then we call an algebraic integer a Pisot number if $ \lambda > 1$ and for all $1 \neq \sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$, it holds that $\vert \sigma(\lambda) \vert <1$. An unit Pisot number is then an algebraic unit which is also a Pisot number. We say that an algebraic unit $\lambda$ with Galois conjugates $\lambda_1= \lambda, \ldots, \lambda_n$ satisfies the full rank condition if for all $d_1, \ldots, d_n$ integers with $$\prod_{j=1}^n \lambda_j^{d_j} = \pm 1,$$ it must hold that $d_1 = d_2 = \ldots = d_n$. From [@payn09-1 Proposition 3.6.] it follows that every unit Pisot number satisfies the full rank condition.
Let ${\mathfrak{n}}^{\mathbb{Q}}$ be an Anosov Lie algebra of nilpotency class $c$ and $f: {\mathfrak{n}}^{\mathbb{Q}}\to {\mathfrak{n}}^{\mathbb{Q}}$ a hyperbolic integer-like automorphism with signature $\{p,q\}$. The characteristic polynomial $h(X)$ of $f$ has integer coefficients and constant term $\pm 1$. This implies that if $g(X)$ is a rational polynomial which divides $h(X)$, then it must have at least one root of absolute value strictly smaller than $1$. We know that $f$ induces an isomorphism on each quotient $\faktor{\gamma_{i-1}({\mathfrak{n}}^{\mathbb{Q}})}{ \gamma_i({\mathfrak{n}}^{\mathbb{Q}})}$ and thus the polynomial $h(X)$ has at least $c$ irreducible factors. Therefore $f$ has at least $c$ eigenvalues of absolute value strictly smaller than $1$ and thus $p \geq c$. By considering $f^{-1}$ as well, we get that $q \geq c$ and this shows that $\min (\operatorname{sgn}(f)) \geq c$. We say that an Anosov automorphism $f: {\mathfrak{n}}^{\mathbb{Q}}\to {\mathfrak{n}}^{\mathbb{Q}}$ has minimal signature if equality holds, i.e. if $\min(\operatorname{sgn}(f)) = c$.
Question \[q1\] asks if there exists Anosov automorphisms of minimal signature for $c= 2$ and already in Section \[verb\] we gave a positive answer to this question as a consequence of the main theorem. So the existence of Anosov automorphisms of minimal signature is a generalization of Question \[q1\] and with Theorem \[main\] we can also give a positive answer to the generalized question:
\[csig\] For every $c$, there exists an Anosov automorphism $f: {\mathfrak{n}}^{\mathbb{Q}}\to {\mathfrak{n}}^{\mathbb{Q}}$ on a Lie algebra of nilpotency class $c$ such that $f$ is of minimal signature.
Let $E$ be a real Galois extension of ${\mathbb{Q}}$ with $\operatorname{Gal}(E,{\mathbb{Q}}) \cong {\mathbb{Z}}_{2n}$ for some $n > 1$ and $\sigma$ a generator of $\operatorname{Gal}(E,{\mathbb{Q}})$. Take $\lambda$ an unit Pisot number in $E$ with the extra condition that $\vert \lambda \sigma^n(\lambda^2) \vert < 1$ (see the Appendix for more details). Since $n > 1$, we have that $ \vert \lambda \sigma^n(\lambda) \vert > \vert \prod_{i=1}^{2n} \sigma^i(\lambda) \vert = 1$. Consider the collection of algebraic integers $$\mu_{i,j} = \sigma^i(\lambda^{j-1}) \sigma^{i+n}(\lambda)$$ for all $i \in \{1, \ldots, 2n\}$ and all $j \in \{1,\ldots, c\}$. The Galois conjugates $\sigma^i(\lambda)$ of $\lambda$ are the $\mu_{i,j}$ with $j = 1$. Note that the definition implies that $\mu_{i+n,2} = \mu_{i,2}$ and all other $\mu_{i,j}$ are distinct because of the full rank condition.
Every $\mu_{i,j}$ with $i \notin \{n,2n\}$ has absolute value $<1$ since $\lambda$ is a Pisot number. The algebraic unit $\mu_{n,3} = \lambda \sigma^n(\lambda^2)$ satisfies $\vert \mu_{n,3} \vert < 1$ because of our choice of $\lambda$ and therefore also all $\mu_{n,j} = \sigma^n\left(\lambda^{j-3}\right) \mu_{n,3} $ with $j \geq 3$ have absolute value $<1$. Thus it follows that of all $\mu_{i,j}$, only $$\mu_{n,1}= \lambda,\mu_{n,2} = \mu_{2n,2} = \lambda \sigma^n(\lambda), \mu_{2n,3}= \lambda^2 \sigma^n(\lambda), \ldots, \mu_{2n,c} = \lambda^{c-1} \sigma^n(\lambda)$$ have absolute value $>1$, so in total there are $c$ of the $\mu_{i,j}$ with $\vert \mu_{i,j} \vert > 1$.
Now consider the Lie algebra ${\mathfrak{n}}^{\mathbb{Q}}$ with basis $X_{\mu_{ij}}$ for all values of $i$ and $j$, where we write the $\mu_{i,1}$ as the conjugates of $\lambda$. The Lie bracket on ${\mathfrak{n}}^{\mathbb{Q}}$ is given by on the one hand $$\begin{aligned}
[ X_{\sigma^i(\lambda)}, X_{\sigma^{i+n}(\lambda)}] = X_{\mu_{i,2}}\end{aligned}$$ for all $i \in \{1, \ldots, n\}$ and on the other hand by $$\begin{aligned}
[ X_{\sigma^i(\lambda)}, X_{\mu_{i,j}}] = X_{\mu_{i,j+1}}\end{aligned}$$ for all $i \in \{1,\ldots, 2n\}, \hspace{1mm} j \in \{2, \ldots, c\}$ (and all other brackets are $0$). It is easy to check that these relations define a Lie algebra (i.e. that the Jacobi identity holds) and that the $1$-dimensional subspaces spanned by each basis vector satisfy the conditions of Corollary \[main2\]. The map $\rho_\sigma$ given by $\rho_\sigma(X_{\mu_{i,j}}) = - X_{\sigma(\mu_{i,j})}$ for $i \in \{ n, 2n \}$ and $j \geq 2$ and $\rho_\sigma(X_{\mu_{i,j}}) = X_{\sigma(\mu_{i,j})}$ for all other $(i,j)$ defines a representation $\rho: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}})$. The minus sign in the first case comes from the relation $\mu_{i+n,2} = \mu_{i,2}$. The conditions of Corollary \[main2\] are satisfied for $\rho$ and thus this gives us a rational form ${\mathfrak{m}}^{\mathbb{Q}}$ with Anosov automorphism $f$. There are only $c$ eigenvalues of $f$ with absolute value $>1$, so $f$ is of minimal signature.
The type of the example constructed in this theorem is equal to $$\underbrace{(2n,n,2n, 2n,\ldots, 2n)}_{c\ {\rm components}}$$ for all $n \geq 2$. This is not the only possibility for Anosov automorphisms of minimal signature since one can construct examples on Lie algebras of type $$\underbrace{(2n,n,2n, n, 2n,\ldots)}_{c\ {\rm components}},$$ where the induced eigenvalues on $\faktor{\gamma_{2j}({\mathfrak{n}}^{\mathbb{Q}})}{ \gamma_{2j+1}({\mathfrak{n}}^{\mathbb{Q}})}$ are of the form $\left( \sigma^{i}\left(\lambda\right) \sigma^{n+i}\left(\lambda\right)\right)^j$. The construction of such examples is similar as in Theorem \[csig\]. We conjecture that these are the only possibilities:
Let $f: {\mathfrak{n}}^{\mathbb{Q}}\to {\mathfrak{n}}^{\mathbb{Q}}$ be an Anosov automorphism of minimal signature, then the type of ${\mathfrak{n}}^{\mathbb{Q}}$ is one of the following:
(i) $(2n,n,2n, 2n,\ldots, 2n)$ or
(ii) $(2n,n,2n, n, 2n,\ldots)$,
where $n > 1$.
The methods of this paper are useful to prove or disprove this conjecture.
In a similar way, the main theorem also gives a positive answer to Question \[q2\]. We state this theorem in a more general setting:
\[last\] For every $c \in {\mathbb{N}}_0$ and $n \in {\mathbb{N}}$ with $n > 2$, there exists an Anosov Lie algebra of type $(n,\ldots,n)$ and nilpotency class $c$.
Let $E \supseteq {\mathbb{Q}}$ be a real Galois extension with cyclic Galois group of order $n$ and let $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$ be a generator. Take $\lambda_1 \in E$ an unit Pisot number and consider the Galois conjugates $\lambda_1,\lambda_2 = \sigma(\lambda_1),\ldots, \lambda_n = \sigma^{n-1}(\lambda_1)$. Define the algebraic units $$\mu_{i,j} = \lambda_i^{j-1} \sigma(\lambda_i)$$ for all $i \in \{1,\ldots, n\}$ and $j \in \{1,\ldots, c\}$, where the algebraic units $\lambda_i$ occur as the $\mu_{i,j}$ with $j=1$. Let ${\mathfrak{n}}^{\mathbb{Q}}$ be the Lie algebra with basis $X_{\mu_{i,j}}$ for all $i\in \{1,\ldots,n\}$ and $j \in \{1,\ldots, c\}$, with Lie bracket given by $$\begin{aligned}
[X_{\lambda_i},X_{\mu_{i,j}}] &= X_{\mu_{i,j+1}}\end{aligned}$$ for all $i\in \{1,\ldots,n\}$ and $j \in \{1, \ldots, c\}$ and all other brackets $0$. It is easy to see that the Jacobi identity holds and thus that ${\mathfrak{n}}^{\mathbb{Q}}$ is indeed a Lie algebra.
The linear map $h: {\mathfrak{n}}^{\mathbb{Q}}\to {\mathfrak{n}}^{\mathbb{Q}}$ defined by $h(X_{\mu_{i,j}}) = X_{\sigma(\mu_{i,j})}$ is an automorphism of this Lie algebra of order $n$. So the map $\rho: \sigma \mapsto h$ defines a representation $\rho: \operatorname{Gal}(E,{\mathbb{Q}}) \to \operatorname{Aut}({\mathfrak{n}}^{\mathbb{Q}})$. This Lie algebra and the representation $\rho$ satisfy the conditions of Corollary \[main2\] (where the spaces $V_\lambda$ are one-dimensional and spanned by the basis vectors) and thus there exists a rational form of ${\mathfrak{n}}^E$ which is Anosov. The type of this rational form is equal to the type of ${\mathfrak{n}}^{\mathbb{Q}}$.
The case where $n = 3$ gives an answer to Question \[q2\]. It is an open question to determine all possibilities for the types $(n_1,\ldots, n_c)$ of Anosov Lie algebras with $n_1 = n_2 = 3$. To solve this problem, a careful study of the conjugates of algebraic units of degree $3$ is needed.
Appendix {#app .unnumbered}
========
An important ingredient we used during this article is the existence of unit Pisot numbers in a real Galois extension $E$ of ${\mathbb{Q}}$. In this appendix we extend some results of [@dd13-1] about $c$-hyperbolic units to unit Pisot numbers with extra conditions on them.
If $E$ is a real Galois extension of ${\mathbb{Q}}$ of degree $n$, then we call an algebraic integer a Pisot number if $ \lambda > 1$ and for all $1 \neq \sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$, it holds that $\vert \sigma(\lambda) \vert <1$. An unit Pisot number is then an algebraic unit which is also a Pisot number. Denote by $U_E$ the algebraic units of $E$ and by $\sigma_1, \ldots, \sigma_n$ all elements of the Galois group $\operatorname{Gal}(E,{\mathbb{Q}})$ with $\sigma_1 = 1$. From Dirichlet’s Unit Theorem we know that the map $$\begin{aligned}
l: U_E \to {\mathbb{R}}^{n}: \lambda \mapsto \Big( \log \big( \vert \sigma_1\left(\lambda\right)\vert\big) , \ldots, \log\big( \vert \sigma_n \left(\lambda \right) \vert \big) \Big)\end{aligned}$$ maps $U_E$ onto a cocompact lattice of the subspace $V \subseteq {\mathbb{R}}^n$, where $V$ is given by the equation $x_1 + \ldots + x_n = 0$. The unit Pisot numbers are mapped to the open subset $O \subseteq V$ given by the equations $x_1 > 0$ and $x_i < 0$ for all $i \geq 2$. So for the existence of unit Pisot numbers, one has to show that $O \cap l(U_E) \neq \emptyset$. The following lemma asserts that this is indeed the case:
\[simpel\] Let $L \subseteq {\mathbb{R}}^n$ be a cocompact lattice and $O \subseteq {\mathbb{R}}^n$ a nonempty open subset such that for all $v_1, v_2 \in O$ also $v_1 + v_2 \in O$. Then $O \cap L \neq \emptyset$.
Since $O$ is open and $L \otimes {\mathbb{Q}}$ is dense, there exists $x \in O \cap L \otimes {\mathbb{Q}}$. By taking $n x = x + \ldots + x$ for some $n \in {\mathbb{N}}_0$, we find $x \in L \cap O$.
So this lemma implies that there exist unit Pisot numbers in every real Galois extension $E \neq {\mathbb{Q}}$. The lemma also implies that every open nonempty subset of $V$ which is invariant under addition gives rise to possible algebraic units. For example, there also exists unit Pisot numbers with an extra condition on them:
\[existence\] Let $E$ be a real Galois extension of ${\mathbb{Q}}$ and fix some $\sigma \in \operatorname{Gal}(E,{\mathbb{Q}})$ with $\sigma \neq 1$. Then there always exists an unit Pisot number $\lambda \in E$ such that $ \vert \sigma(\lambda^2) \lambda \vert < 1$.
Assume that $\sigma_2$ of the map $l$ given above is equal to $\sigma$. Let $O \subseteq V$ be the open nonempty subset of $V$ given by $x_1 > 0, x_i < 0$ for all $i \geq 2$ and $x_1 + 2 x_2 <0$, then there exists $x \in O \cap l(U_E)$ because of the previous lemma. Any element of the preimage of $x$ will satisfy the conditions of the proposition.
**Acknowledgments**
I would like to thank my advisor Karel Dekimpe for his useful comments on a first version of this article and the referee for his/her remarks which have improved this paper.
[^1]: The author was supported by a Ph.D. fellowship of the Research Foundation – Flanders (FWO). Research supported by the research Fund of the KU Leuven
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Ranking alternatives is a natural way for humans to explain their preferences. It is being used in many settings, such as school choice, course allocations and residency matches. In some cases, several “items” are given to each participant. Without having any information on the underlying cardinal utilities, arguing about fairness of allocation requires extending the ordinal item ranking to ordinal bundle ranking. The most commonly used such extension is stochastic dominance (SD), where a bundle X is preferred over a bundle Y if its score is better according to all additive score functions. SD is a very conservative extension, by which few allocations are necessarily fair while many allocations are possibly fair. We propose to make a natural assumption on the underlying cardinal utilities of the players, namely that the difference between two items at the top is larger than the difference between two items at the bottom. This assumption implies a preference extension which we call diminishing differences (DD), where X is preferred over Y if its score is better according to all additive score functions satisfying the DD assumption. We give a full characterization of allocations that are necessarily-proportional or possibly-proportional according to this assumption. Based on this characterization, we present a polynomial-time algorithm for finding a necessarily-DD-proportional allocation if it exists. Using simulations, we show that with high probability, a necessarily-proportional allocation does not exist but a necessarily-DD-proportional allocation exists, and moreover, that allocation is proportional according to the underlying cardinal utilities. We also consider chore allocation under the analogous condition — increasing-differences.'
author:
- 'Erel Segal-Halevi'
- Avinatan Hassidim
- Haris Aziz
bibliography:
- 'merged.bib'
title: 'Fair Allocation with Diminishing Differences [^1] '
---
*JEL*: C62, C63, and C78
Introduction {#sec:intro}
============
Algorithms for fair assignment of indivisible items differ in the kind of information they require from the users.
Some algorithms require the users to rank bundles of items, i.e, report a total order among the bundles. Examples are the Decreasing Demand procedure of @Herreiner2002Simple, the Approximate-CEEI procedure of @budish2011combinatorial and the two-agent Undercut procedure [@Brams2012Undercut; @Aziz2015Note]. The computational and communicational burden might be large, since the number of bundles is exponential in the number of items.
Other algorithms require the users to evaluate individual items, i.e, supply a numeric monetary value for each item. Such algorithms are often termed *cardinal*. They often assume that the users’ valuations are additive, so that the value of a bundle can be calculated by summing the values of the individual items. Examples are the Adjusted Winner procedure of @Brams2000WinWin, the approximate-maximin-share procedure of @Procaccia2014Fair and the Maximum Nash Welfare procedure of @Caragiannis2016Unreasonable. In this setting the communication is linear in the number of items, but the mental burden may still be large, since assigning an exact monetary value to individual items is not easy. This is especially true when items are valued for personal reasons (such as when dividing inheritance) and do not have a market price.
This paper focuses on a third class of algorithms, which only require the users to rank individual items, i.e, report a total order among items. Such algorithms are often termed *ordinal*.
Ordinal algorithms are ubiquitous in mechanism design. They are often used in real world applications, such as the National Residency Matching Program [@Roth1997NRMP] (even when married couples insert their preferences together [@ashlagi2014stability]), school choice applications [@abdulkadiroglu2003school], and university admittance [@hassidim2016strategic; @hassidim2016redesigning]. One reason for this is that it is relatively easy for people to state ordinal preferences. Another reason is related to legacy systems: often the designer can change the allocation mechanism, but can not change the input procedure, as agents do not want to learn new ways to insert their input to the system.
Ordinal algorithms are also common in AI and in fair division. Examples are the AL two-agent procedure of @Brams2013TwoPerson, optimal-proportional procedure of @Aziz2015Fair, picking-sequence procedures of @Brams2004Dividing [@Bouveret2011General] and the envy-free procedures of @Bouveret2010Fair. Such algorithms often assume that the agents’ preferences are implicitly represented by an additive utility function, which is not known to the algorithm. This creates ambiguity in the agents’ bundle rankings. For example, if an agent ranks four items as $w\succ x \succ y\succ z$, then, based on additivity, the algorithm can know that e.g. $\{w,x\}\succ \{y,z\}$ and $\{w,y\}\succ \{x,z\}$, but cannot know the relation between $\{w,z\}$ and $\{x,y\}$. Algorithms cope with this problem in several ways.
1\. **Necessary-fairness criteria**. An allocation is called **necessarily-fair** if it is fair for *all* additive utility profiles consistent with the reported item-rankings. Here, “fair” may be substituted by any fairness criterion, such as envy-freeness or proportionality, as well as Pareto-efficiency. Necessary fairness is a strong requirement, which is not always satisfiable. For example, the AL procedure finds a necessarily-envy-free allocation, but only for two agents, and even then, it might need to discard some of the items.
2\. **Possible-fairness criteria**. An allocation is called **possibly-fair** if it is fair for *at least one* additive utility profile consistent with the reported item-rankings, Again, “fair” may be substituted by proportional or envy-free or Pareto-efficient. Possible fairness is a weak criterion; algorithms that only return possibly-fair allocations might be considered unfair by users whose actual utility function is different.
3\. **Scoring rules**. A scoring rule is a function that maps the rank of an item to a numeric score. A common example is the Borda scoring rule [@Young1974Borda], where the least desired item has a score of 1, the next item has a score of 2, and so on. The score of a bundle is the sum of the scores of its items. It is assumed that all agents have the same scoring function. I.e, even though agents may rank items differently, the mapping from the ranking to the numeric utility function is the same for all agents [@Bouveret2011General; @Kalinowski2013Social; @Baumeister2016Positional; @Darmann2016Proportional]. This strong assumption weakens the fairness guarantee. Allocation may appear unfair to agents whose actual scoring rule is different.
Contribution
------------
The present paper suggests an alternative between the strong guarantee of necessary-fairness and the weak guarantee of possible-fairness and scoring-rule-fairness.
We assume that people care more about their high-valued items than about their lower-valued items. Specifically, we assume that the utility-difference between the best item and the second-best item is at least as large as the utility between the second-best and the third-best, and so on. We call this assumption *Diminishing Differences (DD)*. The DD assumption is satisfied by the Borda scoring rule, as well as by many other scoring rules, as well as by lexicographic preferences.
DD is justified in many settings where the agents are more concerned about getting a most preferred item than about not getting a least preferred item. For example, in a matching of doctors to internships, it was reported that doctors care the most about being assigned to one of their top choices [@bronfman2015redesigning; @bronfman2015assigning].
Based on the DD assumption, we formalize several fairness notions. We call an allocation **necessarily-DD-fair** (NDD-fair) if it is fair according to all additive utility profiles satisfying the DD assumption, and **possibly-DD-fair** (PDD-fair) if it is fair according to at least one additive utility profile satisfying the DD assumption. Again, “fair” may be substituted by envy-free or proportional or Pareto-efficient. The following implications are obvious for any fairness criterion:
Necessarily-fair $\implies$ NDD-fair $\implies$ PDD-fair $\implies$ Possibly-fair
In other words, the DD-fairness criteria are intermediate in strength between necessary-fairness and possible-fairness. A formal definition of these concepts appears in **Section \[sec:dd\]**.
The first question of interest is to decide, given an item ranking and two bundles, whether the NDD or the PDD relation holds between these bundles. We prove characterizations of the NDD and PDD set relations that provide linear-time algorithms for answering these questions (**Section \[sec:NDD\]**).
Next, we prove a necessary and sufficient condition for the existence of an NDD-proportional (NDDPR) allocation. Essentially, an NDDPR allocation exists if-and-only-if it is possible to (a) give all agents the same number of items and (b) give each agent his best item. The proof is constructive and presents a simple linear-time algorithm for finding an NDDPR allocation if it exists (**Section \[sec:NDDPR\]**).
To understand the difference between NDD-fairness and necessary-fairness, it is interesting to compare the above condition to Condition D of @Brams2013TwoPerson, which is necessary and sufficient for the existence of a necessarily-proportional (NecPR) allocation for two agents. For NecPR, it is required that for every odd integer $k \in\{ 1,3,\ldots,2n-1\}$, the agents have a different set of $k$ best items; in particular, they should have a different worst item (see Example \[exm:ndd\] in Section \[sec:dd\]). For NDDPR, it is only required that the agents have a different best item. This means that an algorithm that returns NDDPR allocations may have larger ‘recall’ than an algorithm that returns only NecPR allocations. On the other hand, the ‘precision’ of such an algorithm might be lower (i.e, its output allocations may be considered unfair by some agents).
To assess the magnitude of these effects, we present a simple simulation experiment. We construct partially-correlated utility profiles at random, estimate the probability that an NDDPR/NecPR allocation exists, and check whether the NDDPR allocation given by our linear-time algorithm is indeed proportional according to the underlying cardinal utilities. We find that the increase in recall is substantial and ranges between 20% and 40%, but the decrease in precision is not substantial: when there are sufficiently many items, our simple NDDPR algorithm almost always results in an allocation which is proportional according to the cardinal utilities. This experiment indicates that there is potential for further investigation of NDDPR as a normative fairness criterion (**Section \[sec:simulations\]**).
While our main interest is in NDD-proportionality, we briefly present several extensions of our model.
First, instead of proportionality, we study the stronger property of *envy-freeness* (EF). Since every EF allocation is PR, every NDDEF allocation is NDDPR. Therefore, conditions (a) and (b) above are still necessary to NDDEF existence. When there are $n=2$ agents, EF is equivalent to PR, so NDDPR is equivalent to NDDEF. Therefore, conditions (a) and (b) are also sufficient, and when they are satisfied, an NDDEF allocation can be found in linear time. EF and PR diverge when there are three or more agents. When $n=3$, we show that an NDDEF allocation might not exist even if conditions (a) and (b) hold. We then study the computational problem of deciding whether an NDDEF allocation exists. Since the conditions of Theorem \[thm:nddpr\] are necessary for NDDEF, this decision problem is trivial whenever $M$ is not a multiple of $n$, and it is also trivial if $M=n$. Therefore the first interesting case is $M=2n$. We prove that the decision problem is NP-hard already in this case (**Section \[sec:NDDEF\]**).
Second, we study Pareto-efficiency (PE). The DD assumption has a substantial effect on fairness notions: NDD-fair allocations are easier (in terms of existence) than necessary-fair allocations and PDD-fair allocations are harder than possibly-fair allocations. Interestingly, the DD assumption does not have this effect on PE. We show that NDD-PE is equivalent to necessarily-PE and PDD-PE is equivalent to possibly-PE. So the DD assumption does not lead to a new efficiency notion (**Section \[sec:efficiency\]**).
Third, we study the allocation of *chores* — items with negative utilities. We assume that people care more about *not* getting the *worst* chore than about getting the best chore; this naturally leads to the condition of *increasing differences (ID)*. While the basic definitions and lemmas for the DD relations have exact analogues for the ID relations, our characterization for existence of NDDPR allocation of goods has no direct analogue for NIDPR allocation of chores (**Section \[sec:increasing\]**).
Finally, we compare the Diminishing-Differences assumption to another natural assumption which we call *Binary*. It is based on the assumption that each agent only cares about getting as many as possible of his $k$ best items, where $k$ is an integer that may be different for different agents. We show that, while the number of utility functions that satisfy this assumption (for a given preference relation) is much smaller than the number of DD utility functions, it does not lead to new fairness concepts: necessary-binary-fairness is equivalent to necessary-fairness and possible-binary-fairness is equivalent to possible-fairness (**Section \[sec:binary\]**).
Related Work
------------
Extending preferences over individual items to sets of items is a natural and principled way of succinctly encoding preferences [@BBP04a]. One of the most common set-extensions is *stochastic dominance* (SD). It was developed for a different but related problem — extending preferences over individual outcomes to *lotteries* on outcomes. If $X,Y$ are lotteries, then $X\succeq^{SD} Y$ iff $E[u(X)]\geq E[u(Y)]$ for every weakly-increasing utility function $u$ [@HaRu69a; @Bran17a]. In the context of fair item allocation, SD leads to the notions of necessary-fairness and possible-fairness [@Aziz2015Fair]. Other common extensions are *downward-lexicographic* (DL) and *upward-lexicographic* (UL) [@Cho2012Probabilistic; @Bouveret2010Fair; @NBR15a]. A different kind of assumption that is common in allocating computational resources is that agents have *Leontief preferences*, i.e., each agent needs resources in fixed proportions (e.g., [@kash2014no]).
The diminishing differences extension, which is the focus of this paper, is quite natural but has not been formalized in prior work. Interestingly, the DD extension is closely related to *second-order stochastic dominance* (SSD). If $X,Y$ are lotteries, then $X\succeq^{SSD} Y$ iff $E[u(X)]\geq E[u(Y)]$ for every utility function $u$ which is weakly-increasing and *weakly-concave* [@HaRu69a]. In the context of item assignment, weak concavity is equivalent to *increasing* differences — agents care more about *not* getting the *worst* item than about getting the best item. Increasing differences make sense in fair division of chores [@Aziz2017Chores] and we relate to it in Section \[sec:increasing\].
Besides fair division, set-extensions have been applied for committee voting [@ALL16a] and social choice correspondences (see e.g., [@BDS01a; @KePe84a]). Recently, set extensions have also been used in philosophic works on ethics. Suppose an ethical agent has to choose between several actions. He is unsure between two ethical theories, each of which ranks the choices differently. Thus, each choice is in fact a lottery. Using the SD set-extension, @aboodi2017intertheoretic shows that, in some cases, the agent can still choose an ethically-best action.
In social choice theory, it is common to study restricted domains of preference profiles, such as single-peaked, single-crossing or level-$r$-consensus [@Mahajne2015Level; @Nitzan2017Level1]. Many problems are much easier to solve in such restricted domains than in the domain of all preferences [@Elkind2014Detecting; @ELP17a]. The present paper focuses on a restriction to preferences satisfying the DD assumption, which has not been studied so far.
Many works on fair allocation of indivisible items look for allocations that are only approximately-fair, for example, envy-free up to at most one item [@lipton2004approximately; @budish2011combinatorial]. In contrast, we are interested in allocations that are fair without approximations. Naturally, such allocations do not always exist, so we are interested in finding conditions under which they exist.
Preliminaries {#sec:model}
=============
There is a set ${\mathcal{N}}{}$ of agents with $n=|{\mathcal{N}}|$. There is a set ${\mathcal{M}}$ of distinct items with $M=|{\mathcal{M}}|$. A *bundle* is a set of items. A *multi-bundle* is a multi-set of items, i.e, it may contain several copies of the same item.[^2]
An *allocation* $\bf{X}$ is a function that assigns to each agent $i$ a bundle $X_i$, such that ${\mathcal{M}}= X_1 \cup \cdots \cup X_n$ and the $X_i$-s are pairwise-disjoint. Each agent $i$ has a strict ranking ${\succ\xspace}_i$ on items. Each agent may also have a utility function $u_i$ on (multi-)bundles. When we deal with a single agent, we often omit the subscript $_i$ and consider an agent with ranking $\succ$ and utility-function $u$.
All utility functions considered in this paper are positive and additive, so the utility of a (multi-)bundle is the sum of the utilities of the items in it. A utility function $u$ is *consistent with ${\succ\xspace}$* if for every two items $x,y$: $$\begin{aligned}
u(\{x\}) > u(\{y\})
\iff
x{\succ\xspace}y\end{aligned}$$
We denote by $\mathcal{U}({\succ\xspace})$ the set of additive utility functions consistent with ${\succ\xspace}$.
Given $n$ rankings $\succ_1,\ldots,\succ_n$, we denote by $\mathcal{U}(\succ_1,\ldots,\succ_n)$ the set of vectors of additive utility functions $u_1,\ldots,u_n$ such that $u_i$ is consistent with ${\succ\xspace}_i$.
The following definition is well-known:
\[def:sd\] Given a ranking $\succ$ and two (multi-)bundles $X,Y$: $$\begin{aligned}
X \succsim^{Nec} Y
&&
\iff
&&
\forall u\in {\mathcal{U}}({\succ\xspace}) \text{:~~~~~} u(X)\geq u(Y).
\\
X \succsim^{Pos} Y
&&
\iff
&&
\exists u\in {\mathcal{U}}({\succ\xspace}) \text{:~~~~~} u(X)\geq u(Y)\end{aligned}$$
Given a strict ranking $\succ$, we assign to each item $x\in{\mathcal{M}}$ a *level*, denoted ${\text{\sc Lev}}(x)$, such that the level of the best item is $M$, the level of the second-best item is $M-1$, etc (this is also known as the *Borda score* of the item). we define the level of a multi-bundle as the sum of the levels of the items in it: $$\begin{aligned}
{\text{\sc Lev}}(X) := \sum_{x\in X} {\text{\sc Lev}}(x)\end{aligned}$$ where all copies of the same item have the same level.
The Diminishing-Differences Property {#sec:dd}
====================================
We define our new concept of diminishing differences (DD) in three steps: first, we define the set of DD utility functions (Definition \[def:dd-utilities\]). Based on this, we define the necessary-DD and possible-DD relations (Definition \[def:dd-relation\]). Based on this, we define the NDD-fairness and PDD-fairness concepts (Definition \[lem:dd-prop\]).
\[def:dd-utilities\] Let $\succ$ be a preference relation and $u$ a utility function consistent with $\succ$. We say that $u$ has the *Diminishing Differences (DD)* property if, for every three items with consecutive levels $x_3\succ x_2\succ x_1$ such that ${\text{\sc Lev}}(x_3)={\text{\sc Lev}}(x_2)+1={\text{\sc Lev}}(x_1)+2$, it holds that $u(x_3)-u(x_2)\geq u(x_2)-u(x_1)$.
We denote by ${\mathcal{U^{DD}}}({\succ\xspace})$ the set of all DD utility functions consistent with ${\succ\xspace}$.
Given $n$ rankings ${\succ\xspace}_1,\ldots,{\succ\xspace}_n$, We denote by ${\mathcal{U^{DD}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$ the set of all vectors of DD utility functions, $u_1,\ldots,u_n$, such that $u_i$ is consistent with ${\succ\xspace}_i$.
By definition, the Borda utility function consistent with $\succ$ is a member of ${\mathcal{U^{DD}}}(\succ)$. Another example of a member in ${\mathcal{U^{DD}}}$ is the lexicographic utility function $Lex(x) := 2^{{\text{\sc Lev}}(x)}$, by which bundles are ordered by whether they contain the best item, then by whether they contain the second-best item, etc. An equivalent definition of ${\mathcal{U^{DD}}}$ is given by the following lemma.
\[lem:u-in-DD\] $u\in {\mathcal{U^{DD}}}({\succ\xspace})$ iff, for every four items $x_2, y_2, x_1, y_1$ with $x_2\succeq x_1$ and $y_2\succeq y_1$ and $x_2\neq y_2$ and $x_1\neq y_1$: $$\begin{aligned}
\frac{u(x_2)-u(y_2)}{{\text{\sc Lev}}(x_2)-{\text{\sc Lev}}(y_2)}
\geq
\frac{u(x_1)-u(y_1)}{{\text{\sc Lev}}(x_1)-{\text{\sc Lev}}(y_1)}\end{aligned}$$
The proof is technical and we defer it to Appendix\[app:proofs\].
\[def:dd-relation\] Given a ranking $\succ$ and two (multi-)bundles $X,Y$: $$\begin{aligned}
X \succsim^{NDD} Y
&&
\iff
&&
\forall u\in {\mathcal{U^{DD}}}({\succ\xspace}) \text{:~~~~~} u(X)\geq u(Y)
\\
X \succsim^{PDD} Y
&&
\iff
&&
\exists u\in {\mathcal{U^{DD}}}({\succ\xspace}) \text{:~~~~~} u(X)\geq u(Y)
\end{aligned}$$
\[rem:dd-sd\] Comparing Definitions \[def:sd\] and \[def:dd-relation\], it is clear that: $$\begin{aligned}
X \succsim^{Nec} Y
\implies
X \succsim^{NDD} Y
\implies
X \succsim^{PDD} Y
\implies
X \succsim^{Pos} Y\end{aligned}$$
We now define the main fairness concept that we will investigate in this paper — proportionality.
\[def:dd-prop\] Given utility functions $u_1,\ldots,u_n$, an allocation $\bf{X}$ is called *proportional* if $\forall i\in{\mathcal{N}}:~~ n\cdot u_i(X_i) ~\geq~ u_i({\mathcal{M}})$.\
Given item rankings $\succ_1,\ldots,\succ_n$, an allocation $\bf{X}$ is called:
- *Necessary-DD-proportional (NDDPR)* if it is proportional for all utility profiles $u_1,\ldots,u_n$ in ${\mathcal{U^{DD}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
- *Possible-DD-proportional (PDDPR)* if it is proportional for at least one utility profile $u_1,\ldots,u_n$ in ${\mathcal{U^{DD}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
For comparison, recall that an allocation $\bf{X}$ is called:
- *necessarily-proportional (NecPR)* if it is proportional for all utility profiles $u_1,\ldots,u_n$ in $\mathcal{U}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
- *possibly-proportional (PosPR)* if it is proportional for at least one utility profile $u_1,\ldots,u_n$ in $\mathcal{U}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
Like in Remark \[rem:dd-sd\], it is clear that: necessarily-proportionality implies NDD-proportionality implies PDD-proportionality implies possibly-proportionality.
We now give equivalent definitions of NDDPR and PDDPR in terms of the NDD and PDD relations. For every integer $k$ and bundle $X_i$, define $k\cdot X_i$ as the multi-bundle in which each item of $X_i$ is copied $k$ times. Proportionality can be defined by comparing, for each agent $i$, the bundle $X_i$ copied $n$ times, to the bundle of all items ${\mathcal{M}}$.
\[lem:dd-prop\] Given item rankings $\succ_1,\ldots,\succ_n$:
- An allocation $\bf{X}$ is NDDPR iff $\forall i\in{\mathcal{N}}:~ n\cdot X_i~\succsim_i^{NDD}~{\mathcal{M}}$.
- An allocation $\bf{X}$ is PDDPR iff $\forall i\in{\mathcal{N}}:~ n\cdot X_i~\succsim_i^{PDD}~{\mathcal{M}}$.
For NDDPR, the proof involves just switching the order of two for-all quantifiers: the original definition is “for all utility profiles - for all agents” and the new definition is “for all agents - for all utility functions”. Switching for-all quantifiers yields equivalent statements.
For PDDPR, we have two different quantifiers: the original definition is “there exists a utility profile for which for every agent...” and the new definition is “for every agent there exists a utility function...”. The former definition obviously implies the latter. It remains to prove that the latter implies the former. Indeed, suppose that for every agent $i\in{\mathcal{N}}$, $n\cdot X_i \succeq^{PDD} {\mathcal{M}}$. Then for every $i$, there exists $u_i\in{\mathcal{U^{DD}}}(\succ_i)$ such that $u_i(n\cdot X_i)\geq u_i({\mathcal{M}})$. By additivity, $u_i(n\cdot X_i) = n\cdot u_i(X_i)$, so for every $i$, $n \cdot u_i(X_i)\geq u_i({\mathcal{M}})$. Therefore the allocation $\bf{X}$ is proportional according to the profile $(u_1,\ldots,u_n)\in {\mathcal{U^{DD}}}(\succ_1,\cdots,\succ_n)$.
Characterizing NDD and PDD Relations {#sec:NDD}
====================================
As a first step in finding DD-fair allocations among many agents, we study the NDD and PDD relations for a single agent. We are given a preference relation $\succ$ on items and two multi-bundles $X,Y$, and have to decide whether $X \succsim^{NDD} Y$ and/or $X \succsim^{PDD} Y$.
We begin by proving a convenient characterization of the NDD relation. For the characterization, we order the items in each multi-bundle by decreasing level, so $X=\{x_{-1},\ldots, x_{-|X|}\}$ where $x_{-1}\succeq \ldots \succeq x_{-|X|}$ (the order between different copies of the same item is arbitrary).[^3] For each $k\leq |X|$ we define $X^{-k}$ as the $k$ best items in $X$, $X^{-k}:=\{x_{-1},\ldots,x_{-k}\}$.
\[thm:X-NDD-Y\] Given a ranking $\succ$ and two (multi-)bundles $X,Y$, $X\succsim^{NDD} Y$ if-and-only-if both of the following conditions hold:
1. $|X|\geq |Y|$ and
2. for each $k\in \{1,\ldots, |Y|\}$: ${\text{\sc Lev}}(X^{-k}) \geq {\text{\sc Lev}}(Y^{-k})$.
Theorem \[thm:X-NDD-Y\] implies that there is a polynomial time algorithm to check whether $X\succsim^{NDD} Y$; see Algorithm \[algo:comparedd\].
$X,Y\subset {\mathcal{M}}$, and a ranking $\succ$ of the items in ${\mathcal{M}}$. Yes if $X\succsim^{NDD} Y$; No otherwise.
No Order the items in $X$ and $Y$ by decreasing rank: $x_{-1} \succ \cdots \succ x_{-|X|}$ and $y_{-1} \succ \cdots \succ y_{-|Y|}$. Initialize TotalLevelDiff$:= 0$. LevelDiff := $[{\text{\sc Lev}}(x_{-j}) - {\text{\sc Lev}}(y_{-j})]$ TotalLevelDiff += LevelDiff No Yes
\[algo:main\]
\[thm:X-Nec-Y\] Contrast this characterization with the following characterization of $\succeq^{Nec}$ from @Aziz2015Fair. $X\succsim^{Nec} Y$ iff:
1. $|X|\geq |Y|$ and
2. for each $k\in \{1,\ldots, |Y|\}$, the number of items in $X$ that are better than $x_k$ is at least as much as the number of items in $Y$ better than $x_k$.
Before proving Theorem \[thm:X-NDD-Y\], we give some examples.
Suppose the set of items is ${\mathcal{M}}=\{8,\ldots,1\}$ and we are given a preference-relation $8\succ\cdots \succ 1$, so that each item is represented by its level. Consider the following two bundles:
$~~~~~X = \{8,4,2\}~~~~~$$~~~~~Y = \{7,6\}~~~~~$
Note that $|X|>|Y|$, $X$ is lexicographically-better than $Y$, and even the Borda score of $X$ is higher. However, the level of $X^2$ (the two best items in $X$) is only 12 while the level of $Y^2$ is 13. Hence, by Theorem \[thm:X-NDD-Y\], $X\not \succeq^{NDD} Y$. Indeed, $X$ is not better than $Y$ according to the DD utility function $u_{square}(x) := {\text{\sc Lev}}(x)^2$, since $u_{square}(X) = 84 < 85 = u_{square}(Y)$.
Consider the following two bundles:
$~~~~~X' = \{8,5\}~~~~~$$~~~~~Y = \{7,6\}~~~~~$
Now the conditions of Theorem \[thm:X-NDD-Y\] are satisfied: $|X|\geq |Y|$ and ${\text{\sc Lev}}(X^1)\geq {\text{\sc Lev}}(Y^1)$ and ${\text{\sc Lev}}(X^2)\geq {\text{\sc Lev}}(Y^2)$. Hence the theorem implies that $X'\succeq^{NDD} Y$. In contrast, condition (ii) im Remark \[thm:X-Nec-Y\] is not satisfied since the in $X$ there is only one item better than $5$ while in $Y$ there are two such items. Therefore $X\not \succeq^{Nec} Y$. Indeed, $X'$ is not better than $Y$ by the utility function $u_{sqrt}(x):=\sqrt{{\text{\sc Lev}}(x)}$, since $u_{sqrt}(X)\approx 5.06 < 5.09 \approx u_{sqrt}(Y)$.
We assume that either (i) $|Y|>|X|$, or (ii) ${\text{\sc Lev}}(Y^{-k})>{\text{\sc Lev}}(X^{-k})$ for some $k$, and prove that $X\not\succsim^{NDD} Y$, i.e, there is a utility function $u \in {\mathcal{U^{DD}}}(\succ)$ such that $u(X) < u(Y)$.
1. If $|Y|>|X|$, then take: $$\begin{aligned}
u(x) = M|Y| + {\text{\sc Lev}}(x).\end{aligned}$$ It has diminishing-differences since the difference in utilities between items with adjacent ranks is 1.
The term $M|Y|$ is so large that the utility of a bundle is dominated by its cardinality. Formally, for every item $x$, $M|Y| < u(x) \leq M+M|Y|$, so: $$\begin{aligned}
u(X) &\leq |X|\cdot (M+M|Y|)
\\
&< M|Y| + |X|\cdot M|Y| && \text{since $|X|<|Y|$}
\\
&= (|X|+1)\cdot M|Y|
\\
&\leq |Y|\cdot M|Y| && \text{since $|X|<|Y|$}
\\
&< u(Y)
\end{aligned}$$ Hence $X\not\succsim^{NDD} Y$.
2. If for some $k$, ${\text{\sc Lev}}(Y^{-k}) > {\text{\sc Lev}}(X^{-k})$, then let $k$ be the smallest integer that satisfies this condition; hence $y_{-k} \succ x_{-k}$. Let $C := {\text{\sc Lev}}(x_{-k})-1$ and define $u$ as: $$\begin{aligned}
u(x) &= {\text{\sc Lev}}(x) && \text{for $x\prec x_{-k}$}
\\
u(x) &= [{\text{\sc Lev}}(x) - C]\cdot M|X| && \text{for $x\succeq x_{-k}$}\end{aligned}$$ so the utilities of the items worse than $x_{-k}$ are $1, 2, \ldots, C$, and the utilities of $x_{-k}$ and the items better than it are $M |X|, 2 M|X|, 3 M|X|, \ldots$.
$u$ has diminishing-differences, since the difference in utilities between adjacent items ranked weakly above $x_{-k}$ is $M|X|$, the difference between $x_{-k}$ and the next-worse item is less than $M|X|$ and more than 1, and the difference between adjacent items ranked below $x_{-k}$ is 1.
The term $M|X|$ is so large that the utility of a bundle is dominated by the level of its items that are weakly better than $x_{-k}$. Formally: $$\begin{aligned}
u(X) &= u(\{x_{-1},\ldots,x_{-k}\}) + u(\{x_{-(k+1)},\ldots,x_{-|X|}\})
\\
&= M|X|\cdot [{\text{\sc Lev}}(\{x_{-1},\ldots,x_{-k}\}) - k\cdot C]
+ {\text{\sc Lev}}(\{x_{-(k+1)},\ldots,x_{-|X|}\})\end{aligned}$$ The assumption ${\text{\sc Lev}}(Y^{-k})>{\text{\sc Lev}}(X^{-k})$ implies that ${\text{\sc Lev}}(X^{-k})\leq {\text{\sc Lev}}(Y^{-k}) -1$. Hence the leftmost term is at most $M|X|\cdot [{\text{\sc Lev}}(\{y_{-1},\ldots,y_{-k}\}) - 1 - k\cdot C]$. Since the level of an item is at most $M$, the rightmost term is less than $M|X|$. Hence: $$\begin{aligned}
u(X) &< M|X|\cdot [{\text{\sc Lev}}(\{y_{-1},\ldots,y_{-k}\}) - 1 - k\cdot C]
+ M|X|
\\
&= M|X|\cdot [{\text{\sc Lev}}(\{y_{-1},\ldots,y_{-k}\}) - k\cdot C]
\\
&\leq u(Y^{-k}) \leq u(Y)\end{aligned}$$ Hence $X\not\succsim^{NDD} Y$.
We assume that $|X|\geq |Y|$ and that $\forall k{\in\{1,\ldots,|Y|\}}: {\text{\sc Lev}}(X^{-k})\geq {\text{\sc Lev}}(Y^{-k})$. We consider an arbitrary utility function $u\in{\mathcal{U^{DD}}}(\succ)$ and prove that $\forall k{\in\{1,\ldots,|Y|\}}: u(X^{-k})\geq u(Y^{-k})$. This will imply that $u(X)\geq u(Y)$, so that $X\succsim^{NDD} Y$. We assume that for every $j{\in\{1,\ldots,|Y|\}}$: $x_j\neq y_j$. This does not lose generality, since if for some $j$ we have $x_j = y_j$, we can just remove this item from both $X$ and $Y$; this changes neither the assumptions nor the conclusion.
In the proof we use the following notation.
- $l_{k} := {\text{\sc Lev}}(x_{-k}) - {\text{\sc Lev}}(y_{-k})$.
- $L^{k} := {\text{\sc Lev}}(X^{-k}) - {\text{\sc Lev}}(Y^{-k}) = \sum_{j=1}^k l_{k}$.
- $u_{k} := u(x_{-k}) - u(y_{-k})$.
- $r_{k} := u_{k} / l_{k}$.
- $U^{k} := u(X^{-k}) - u(Y^{-k}) = \sum_{j=1}^k u_{k} = \sum_{j=1}^k r_{k} l_{k}$.
In this notation, the assumption is $\forall k{\in\{1,\ldots,|Y|\}}: L^{k}\geq 0$, and we have to prove that $\forall k{\in\{1,\ldots,|Y|\}}: U^{k}\geq 0$.
Suppose we walk on the graph of $L^{k}$ (see Figure \[fig:X-NDD-Y\]). When we move from $L^{j-1}$ to $L^{j}$, we make $l_j$ steps (upwards if $l_j> 0$ or downwards if $l_j<0$). By assumption the graph is always above zero, so to every step downwards, corresponds a previous step upwards.
Suppose we walk simultaneously on the graph of $U^{k}$. When we move from $U^{j-1}$ to $U^{j}$, we make a step of size $u_j = r_j l_j$; equivalently, we make $l_j$ steps of size $r_j$ (upwards if $l_j> 0$ or downwards if $l_j<0$). So to every step of size 1 on the graph of $L^k$ corresponds a step of size $r_j$ on the graph of $U^k$ (see Figure \[fig:X-NDD-Y\]).
![ \[fig:X-NDD-Y\] An illustration of the graphs $L^k$ and $U^k$ in the proof of Theorem \[thm:X-NDD-Y\], if part. ](dd-lemma.png){height="5cm" width="10cm"}
The items in $X$ and $Y$ are ordered by decreasing rank. Hence, by Lemma \[lem:u-in-DD\], the sequence $r_k$ is weakly decreasing with $k$. Therefore, to every step downwards of size $r_j$ on the graph of $U^k$, corresponds a previous step upwards, and its size is at least $r_j$.
Therefore, the graph of $U^{k}$, too, always remains above 0.
Our next theorem gives an analogous characterization of the PDD relation.
\[thm:X-PDD-Y\] Given a ranking $\succ$ and two (multi-)bundles $X,Y$, $Y\succsim^{PDD} X$ if-and-only-if at least one of the following conditions hold:
1. $|Y| > |X|$, or
2. for some $k\in \{1,\ldots, |Y|\}$: ${\text{\sc Lev}}(Y^{-k}) > {\text{\sc Lev}}(X^{-k})$, or
3. ${\text{\sc Lev}}(Y) \geq {\text{\sc Lev}}(X)$.
If either (i) or (ii) holds, then we can take a utility function $u\in{\mathcal{U^{DD}}}(\succ)$ exactly as in the proof of Theorem \[thm:X-NDD-Y\] \[only-if part\]. If (iii) holds, we can just take $u(x) = {\text{\sc Lev}}(x)$. Since $u\in{\mathcal{U^{DD}}}(\succ)$ and $u(Y)\geq u(X)$, this implies $Y\succsim^{PDD}X$.
We assume that none of the three conditions holds, and prove that $Y\not\succsim^{PDD} X$. So we have:
$\widehat{(i)}$ $|X|\geq |Y|$, and
$\widehat{(ii)}$ $\forall k{\in\{1,\ldots,|Y|\}}: {\text{\sc Lev}}(X^{-k})\geq {\text{\sc Lev}}(Y^{-k})$, and
$\widehat{(iii)}$ ${\text{\sc Lev}}(X) > {\text{\sc Lev}}(Y)$.
From here the proof is similar to the proof of Theorem \[thm:X-NDD-Y\] \[if part\]: by the first two conditions, the graph of $L^{k}$ is always weakly above zero, so to every step downwards corresponds a previous step upwards. Moreover, $\widehat{(iii)}$ implies that ${\text{\sc Lev}}(X)-{\text{\sc Lev}}(Y)>0$, so the graph of $L^{k}$ ends strictly above zero, so there exists a step upwards with no corresponding step downwards. Therefore the graph of $U^{k}$, too, ends strictly above 0. Therefore, $u(X) > u(Y)$.
The same is true for every $u\in{\mathcal{U^{DD}}}(\succ)$. Hence $Y\not\succsim^{PDD} X$.
Theorem \[thm:X-PDD-Y\] implies that there is a polynomial time algorithm to check whether $X\succsim^{PDD} Y$; the algorithm is similar to Algorithm \[algo:comparedd\] and we omit it.
Using Theorem \[thm:X-PDD-Y\], we illustrate the difference between PDD-fairness and possible-fairness.
\[exm:pdd\] Alice and Bob have the same preferences, for some $m\geq 3$: $$\begin{aligned}
2 m \succ 2 m - 1 \succ ... \succ 4 \succ 3 \succ 2 \succ 1\end{aligned}$$ Both Alice and Bob get $m$ items: Alice gets $2 m, 2 m - 1, ...m + 3, m + 2, 1$ and Bob gets $m + 1,m,...3,2$. Intuitively this allocation seems very unfair, since Alice gets all the $m-1$ best items. However, it is possibly-proportional, since Bob’s utility function might assign the a value near 0 to item 1 and a value near 1 to all other items.
In better accordance with our intuition, the above allocation is not PDD-proportional: by Theorem \[thm:X-PDD-Y\], Bob’s bundle is not PDD-better than Alice’s bundle, since it does not satisfy any of the conditions (i) to (iii).
Based on the two constructive theorems proved in this section, we have:
\[cor:polynomial\] It can be decided in polynomial time whether a given allocation is NDDPR or PDDPR.
Existence of NDD-Proportional Allocations {#sec:NDDPR}
=========================================
In this section, we prove a necessary and sufficient condition for existence of NDDPR allocations.
\[thm:nddpr\] An NDDPR allocation exists if-and-only-if:
- \(a) The number of items is $M=m\cdot n$, where $m$ is an integer and $n$ is the number of agents, and -
- \(b) Each agent has a different best item.
In case it exists, it can be found in time $O(M)$.
Let $X_1,\ldots,X_n$ be an NDDPR allocation. Then for every agent $i$, $n\cdot X_i \succsim^{NDD}_i {\mathcal{M}}$. By the two conditions of Theorem \[thm:X-NDD-Y\]:
\(a) For all $i$: $|n\cdot X_i| \geq |{\mathcal{M}}|
\implies n\cdot |X_i| \geq M.$ But this must be an equality since the total number of items in all $n$ bundles is exactly $M$. Therefore, the total number of items is $n\cdot |X_i|$ which is an integer multiple of $n$.
\(b) For all $i$, the level of the best item in $n\cdot X_i$ must be weakly larger than the level of the best item in ${\mathcal{M}}$. So for every $i$, $X_i$ must contain agent $i$’s best item. So the best items of all agents must be different.
Suppose the two conditions are satisfied. We prove that the *balanced round-robin* algorithm (Algorithm \[algo:roundrobin\]) produces an NDDPR allocation.[^4]
Give agent $i$ his best remaining item. Give agent $i$ his best remaining item.
Let $X_i$ be the bundle allocated to agent $i$ by balanced-round-robin. We prove that $n\cdot X_i \succsim^{NDD}_i {\mathcal{M}}$ by the two conditions of Theorem \[thm:X-NDD-Y\].
Condition (i) is satisfied with equality, since by (a) each agent gets exactly $m$ items, so $|n\cdot X_i| = n m = M = |{\mathcal{M}}|$.
Condition (ii) says that, for every $k{\in\{1,\ldots,M\}}$, the total level of the $k$ best items in the multi-bundle $n\cdot X_i$ is at least as large as the total level of the $k$ best items in ${\mathcal{M}}$. It is convenient to verify this condition following Algorithm \[algo:comparedd\]: we have to prove that, when going over the items in both bundles from best to worst, the total level-difference between them (the variable TotalLevelDiff in the algorithm) remains at least 0.
We first prove that this is true after the first round. By condition (b), in the first round each agent receives his best item, so the level of the best $n$ items in $n\cdot X_i$ is $M$. The following table shows the levels and their differences for $k\in\{1,\ldots,n\}$ (recall that all items in ${\mathcal{M}}$ are distinct):
----------------- ------- ------- ------- ---------- ------------
$k=1$ $k=2$ $k=3$ $\ldots$ $k=n$
$n\cdot X_i$ $M$ $M$ $M$ $\ldots$ $M$
${\mathcal{M}}$ $M$ $M-1$ $M-2$ … $M-n+1$
LevelDiff $0$ $1$ $2$ … $n-1$
TotalLevelDiff $0$ $1$ $3$ … $n(n-1)/2$
----------------- ------- ------- ------- ---------- ------------
We now prove that, after each round $r\geq 1$, the accumulated level-difference TotalLevelDiff is at least $n(n-1)/2$ when $r$ is odd, and at least $n(i-1)$ when $r$ is even. In particular, it is always at least 0.
The proof is by induction on $r$. We have just proved the base $r=1$.
Suppose now that $r$ is even and $r>1$. When agent $i$ gets an item, the number of items already taken is $r n - i$. Therefore, agent $i$’s best remaining item has a level of at least $M-(r n-i)$. Therefore, the level-differences for $k\in\{(r-1)n+1,~\ldots~,r n\}$ are as in the following table (where the last row uses the accumulated level-difference of $n(n-1)/2$ from the induction assumption):
----------------- ---------------------------- ----------------------------------- ---------- ----------------
$n\cdot X_i$ $\geq M-r n+i$ $\geq M-r n+i$ $\ldots$ $\geq M-r n+i$
${\mathcal{M}}$ $M-r n + n$ $M-r n + n-1$ … $M-r n + 1$
LevelDiff $\geq i-n$ $ \geq i-n+1$ … $\geq i-1$
TotalLevelDiff $\geq {n(n-1)\over 2}+i-n$ $ \geq {n(n-1)\over 2}+2 i-2 n+1$ … $\geq n(i-1)$
----------------- ---------------------------- ----------------------------------- ---------- ----------------
Since $i\geq 1$, TotalLevelDiff is always at least 0. A simple calculation shows that, after round $r$ ends, it is at least $n(i-1)$ as claimed.
Suppose now that $r$ is odd and $r>1$. When agent $i$ gets an item, the number of items already taken is $r n - (n-i+1)$. Therefore, agent $i$’s best remaining item has a level of at least $M- r n + (n-i+1)$. Therefore, the level-differences for $k\in\{(r-1)n+1,~\ldots~,r n\}$ are as in the following table:
----------------- -------------------- --------------------- ---------- --------------------
$n\cdot X_i$ $\geq M-r n+n-i+1$ $\geq M-r n+n-i+1$ $\ldots$ $\geq M-r n+n-i+1$
${\mathcal{M}}$ $M-rn+n$ $M-rn+n-1$ … $M-rn+1$
LevelDiff $\geq 1-i$ $\geq 2-i$ … $\geq n-i$
TotalLevelDiff $\geq n(i-1)+1-i$ $\geq n(i-1)+3-2 i$ … $\geq n(n-1)/2$
----------------- -------------------- --------------------- ---------- --------------------
Since $i\leq n$, TotalLevelDiff is always at least 0. A simple calculation shows that, after round $r$ ends, it is at least $n(n-1)/2$ as claimed.
Using Theorem \[thm:nddpr\], we illustrate the difference between NDD-fairness and necessary-fairness.
\[exm:ndd\] Suppose the set of items is ${\mathcal{M}}=\{2m,\ldots,1\}$, for some $m\geq 2$. Alice and Bob have almost opposite preferences: $$\begin{aligned}
\text{Alice:}&& 2 m \succ 2 m - 1 \succ ... \succ 4 \succ 3 \succ 2 \succ 1
\\
\text{Bob:}&& 2\succ 3 \succ 4 \succ... \succ 2 m - 1 \succ 2 m \succ 1\end{aligned}$$ Intuitively we would expect that opposite preferences make it easy to attain a fair division. However, in this case no necessarily-proportional allocation exists: By Remark \[thm:X-Nec-Y\], in a necessarily-fair allocation both agents must receive the same number of items ($m$). But Alice and Bob have the same worst item ($1$), so one of them must get it. Suppose it is Alice. So Alice has only $m-1$ items better than $1$, while Bob has $m$ items better than $1$. Hence, the allocation is not necessarily-proportional for Alice (her utility function might assign a value near 0 to this item and a value near 1 to all other items).
In contrast, our Theorem \[thm:nddpr\] shows that an NDD-proportional allocation exists. Intuitively, since it is possible to give each agent his/her best items, they are willing to compromise on the less important items.
\[exm:NecPR\] There are $2m$ items $\{1,\ldots,2m\}$ and two agents whose preferences are:
- Alice: $2m \succ 2m-1 \succ 2m-2 \succ 2m-3 \succ \cdots \succ 2 \succ 1$.
- Bob: $2m-1 \succ 2m \succ 2m-3 \succ 2m-2 \succ \cdots \succ 2 \succ 1$.
It is possible to give each agent his/her best item, then give each agent his/her best remaining item, and so on, until the very last step, where one agent (say, Alice) will get her worst item. The resulting allocation is NDDPR by Theorem \[thm:nddpr\], but it is not NecPR. For example, consider the utility function: $$\begin{aligned}
u_{Alice}(1) &= 0
\\
u_{Alice}(i) &= 1+i\cdot\epsilon && i>1
\end{aligned}$$ where $\epsilon$ is an infinitesimally small positive constant. According to this utility function, Alice’s utility is approximately $m-1$, which is less than her proportional share of approximately $(2m-1)/2$.
NDD-Proportionality in Simulations {#sec:simulations}
==================================
{width="110.00000%"}
We compared the various fairness criteria using a simulation experiment with $2$ agents. First, we checked to what extent the probability that an NDDPR allocations exists is higher than the probability that a NecPR allocation exists. These probabilities naturally depend on the correlation between the agents’ rankings. When the rankings are completely correlated, both NDDPR and NecPR allocations do not exist; when the rankings are completely independent, both NDDPR and NecPR allocations exist with high probability; the interesting zone is when the rankings are partially correlated. To simulate such rankings, we determined for each item a “market value” drawn uniformly at random from $[1,2]$. We determined the value of each item to each agent as the item’s market value plus noise drawn uniformly at random from $[-A,A]$, where $A$ is a parameter. Based on the cardinal values we determined the agent’s ordinal ranking. Then, we checked whether there exists an NecPR/NDDPR allocation. We did this experiment 1000 times for different values of $A\in \{0.1,\ldots,1\}$ and for different numbers of items — $2m$ items for $m\in\{2,\ldots,8\}$. Typical results are plotted in Figure \[fig:simul\]; the probability of existence of NDDPR allocations (balls) is clearly higher than that of NecPR allocations (triangles):
Since we had randomly-generated cardinal values, we used them for a secondary purpose — we checked whether, when an NDDPR allocation exists, the one found by the simple algorithm of Section \[sec:NDDPR\] is proportional according to these values (dashed lines). Note that, since the randomization we used is completely uniform and does not use the DD assumption, the probability that DD holds for both agents is very low — $1/((2m-1)!)^2$. Nevertheless, our NDDPR allocation (when it exists) is almost always proportional when the number of items or the noise size are sufficiently large, which further shows the robustness of our algorithm.
We also checked the probability of existence of PDDPR and PosPR allocations, and it was nearly $1.0$. Thus, apparently the NecPR requirement is too strong and the PDDPR and PosPR requirements are too weak, while the NDDPR requirement hits a sweet spot between ‘recall’ and ‘precision’: it allows us to solve many instances (= high ‘recall’) and most solutions are satisfactory (= high ‘precision’).[^5]
Envy-freeness {#sec:NDDEF}
=============
The following is an analogue of the definition of proportionality-related fairness concepts (Definition \[def:dd-prop\]):
\[def:dd-ef\] Given utility functions $u_1,\ldots,u_n$, an allocation $\bf{X}$ is called *envy-free (EF)* if $\forall i,j\in{\mathcal{N}}:~~ u_i(X_i) ~\geq~ u_i(X_j)$.\
Given item rankings $\succ_1,\ldots,\succ_n$, an allocation $\bf{X}$ is called:
- *Necessary-DD-envy-free (NDDEF)* if it is envy-free for all utility profiles $u_1,\ldots,u_n$ in ${\mathcal{U^{DD}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
- *Possible-DD-envy-free (PDDEF)* if it is envy-free for at least one utility profile $u_1,\ldots,u_n$ in ${\mathcal{U^{DD}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
The following is an equivalent definition of NDDEF; it is a partial analogue of Lemma \[lem:dd-prop\]:
\[lem:dd-ef\] Given item rankings $\succ_1,\ldots,\succ_n$:
- An allocation $\bf{X}$ is NDDEF iff $\forall i,j\in{\mathcal{N}}:~ X_i~\succsim_i^{NDD}~X_j$.
The proof involves switching the order of three “for all” quantifiers: the original definition is “for all utility profiles - for every agent $i$ - for every agent $j$” and the new definition is “for every agent $i$ - for every agent $j$ - for all utility functions”. Switching for-all quantifiers yields equivalent statements.
We do not have an equivalent definition of PDDEF in terms if the $\succsim_i^{PDD}$ relations.[^6] Below we focus on the NDDEF criterion, since based on our simulation results, we believe the PDDEF criterion (like most possible-fairness criteria) is too weak to be useful.
Since every NDDEF allocation is NDDPR, the two conditions of Theorem \[thm:nddpr\] are necessary for the existence of NDDEF allocations for any number of agents. In the special case of $n=2$ agents, NDDPR is equivalent to NDDEF so these conditions are also sufficient. But for $n=3$ they are no longer sufficient.
\[exm:nddef-3\] There are six items $\{1,\ldots,6\}$. The preferences of the three agents Alice Bob and Carl are:
- Alice: $6 \succ 5 \succ 3 \succ 4 \succ 2\succ 1$
- Bob: $5 \succ 4\succ 3\succ 6\succ 2\succ 1$
- Carl: $4\succ 6\succ 3\succ 5\succ 2\succ 1$
The conditions of Theorem \[thm:nddpr\] are clearly satisfied: the number of items is a multiple of 3 and the best items are all different. However, no NDDEF allocation exists. *Proof:* The preferences are the same up to a cyclic permutation between $6$ $5$ and $4$, so the agents are symmetric and it is without loss of generality to assume that Alice receives item 1. Therefore, to ensure proportionality, Alice’s bundle must be $\{6,1\}$ and her Borda score is $7$. To ensure that Alice is not envious, Bob must get $\{5,2\}$ and Carl must get $\{3,4\}$. This allocation is NDDPR but it is not NDDEF, since Bob envies Carl according to the Borda score.
We now consider the worst-case of the algorithm of Theorem \[thm:nddpr\], i.e, the case that (1) all agents have the same $n$ best items and (2) all agents have the same ranking to the other $M-n$ items. In this case, the conditions of Theorem \[thm:nddpr\] are sufficient for NDDEF existence:
\[claim:positive-3\] Suppose there are $n=3$ agents and $M=3m$ items (for some integer $m$) and (1) all agents have the same $3$ best items and (2) all agents have the same ranking to the other $M-3$ items and (3) each agent has a different best item. Then an NDDEF allocation exists.
This theorem can even be generalized to any $n$:
\[claim:positive-n\] Suppose there are $n$ agents and $M=n\cdot m$ items (for some integer $m$) and (1) all agents have the same $n\cdot k$ best items and (2) all agents have the same ranking to the other $M-k\cdot n$ items and (3) it is possible to give each agent his best item, then give each agent his best remaining item, and so on, $k$ times, where $n \leq (9 k + 3)/ 4 \approx 2.25 k$. Then, an NDDEF allocation exists.
While the conditions of Theorems \[claim:positive-3\] and \[claim:positive-n\] are quite restrictive, they are interesting from a theoretic perspective, since they show that the worst-case of proportionality is a good case for envy-freeness. Since the NDDPR characterization does not work for NDDEF allocations even for three agents, it is an open problem whether the existence of NDDEF allocations can be decided efficiently when there are 3 or more agents.
When the number of agents is not bounded, we have a hardness result:
\[thm:nddef-npcomplete\] When there are $n\geq 3$ agents and at least $2 n$ items, checking the existence of NDDEF allocations is NP-complete (as a function of $n$).
By Lemma \[lem:dd-ef\], to check whether an allocation is NDDEF, we have to do at most $n^2$ checks of the $\succsim^{NDD}$ relation. Each such check can be done in polynomial time by Theorem \[thm:X-NDD-Y\] and Algorithm \[algo:comparedd\]. Hence the problem is in NP.
The proof of NP-hardness is similar to the proof of @Bouveret2010Fair for the NP-hardness of checking existence of necessarily-envy-free allocations. The proof requires carefully checking that the reduction argument works for NDDEF as well. The details are presented in Appendix\[app:proofs\].
When the number of agents is *constant* (at least 3) and the number of items is variable, the runtime complexity of checking NDDEF existence is an open question: is it polynomial like NDDPR, or NP-hard like Necessary-EF [@Aziz2016Control]?
Pareto-efficiency {#sec:efficiency}
=================
An allocation is called *Pareto-efficient* if every other allocation is either not better for any agent, or worse for at least one agent:
\[def:dd-pe\] Given utility functions $u_1,\ldots,u_n$, an allocation $\bf{X}$ is called *Pareto-efficient (PE)* if for every other allocation $\bf{Y}$, either $\forall i\in{\mathcal{N}}: u_i(X_i) \geq u_i(Y_i)$, or $\exists i\in{\mathcal{N}}: u_i(X_i) > u_i(Y_i)$. Given item rankings $\succ_1,\ldots,\succ_n$, an allocation $\bf{X}$ is called:
- *Necessary-DD-Pareto-efficient (NDDPE)* if it is Pareto-efficient for all utility profiles $u_1,\ldots,u_n$ in ${\mathcal{U^{DD}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
- *Possible-DD-Pareto-efficient (PDDPE)* if it is Pareto-efficient for at least one utility profile $u_1,\ldots,u_n$ in ${\mathcal{U^{DD}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
For comparison, recall that an allocation $\bf{X}$ is called:
- *Necessary-Pareto-efficient (NecPE)* if it is Pareto-efficient for all utility profiles $u_1,\ldots,u_n$ in $\mathcal{U}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
- *Possible-Pareto-efficient (PosPE)* if it is Pareto-efficient for at least one utility profile $u_1,\ldots,u_n$ in $\mathcal{U}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
It is clear from the definition that necessarily-PE implies NDD-PE implies PDD-PE implies possibly-PE. With the analogous fairness concepts, these implications are strict, i.e., some possibly-fair allocations are not PDD-fair, and some NDD-fair allocations are not necessarily-fair. Interestingly, with Pareto-efficiency the situation is different:
An allocation is NecPE if-and-only-if it is NDDPE.
The implication NecPE $\implies$ NDDPE is obvious by the definition. We now consider an allocation $\bf{X}$ that is not NecPE and prove $\bf{X}$ is not NDDPE.
By @Aziz2016Optimal Theorem 9, if $\bf{X}$ is not NecPE then there are two options:
\(i) $\bf{X}$ is not possibly-PE. Then, it is certainly not NDD-PE.
\(ii) $\bf{X}$ admits a Pareto-improving one-for-two-swap. I.e, there are two agents Alice and Bob, such that $X_A$ contains an item $x$, $X_B$ contains two items $y,z$, and Bob strictly prefers the one item over the two, i.e.: $x \succ_B y$ and $x\succ_B z$.
Then $X$ is not NDD-PE, since it is not PE for the following utility profile:
- $u_A(x) = M^2+{\text{\sc Lev}}_A(x)$
- $u_B(x) = 2^{{\text{\sc Lev}}_B(x)}$
Note that both utility functions have DD. Alice’s utility is dominated by the number of items she has, so she always prefers two items to one. Bob’s utility is lexicographic, so he always prefers one good item to any number of worse items. Hence, by switching $\{x\}$ and $\{y,z\}$ we get a new allocation that is strictly better for both Alice and Bob, and does not affect any other agent.
Similarly:
An allocation is PosPE if-and-only-if it is PDDPE.
The implication PDDPE $\implies$ PosPE is obvious by definition. We now consider an allocation $\bf{X}$ that is not PDDPE and prove $\bf{X}$ is not PosPE.
Consider the *lexicographic* utility profile, by which for each $i\in{\mathcal{N}}$, $u_i(x) = 2^{{\text{\sc Lev}}_i(x)}$. Since these utilities have DD, $\bf{X}$ is not PE according to this profile. So there exists an allocation $\bf{Y}$ by which for some agent Alice: $u_A(Y_A)>u_A(X_A)$, and for all agents $B$: $u_B(Y_B)\geq u_B(X_B)$.
Since Alice prefers $Y_A$ to $X_A$ by a lexicographic utility function, there exists some integer $k\geq 1$ such that $X_A$ and $Y_A$ contain the same $k-1$ best items, but the $k$-th best item in $Y_A$ (denoted by $y_a$) is better for Alice than the $k$-th best element in $X_A$.
In allocation $\bf{X}$, item $y_a$ belonged to some other agent, say Bob. But Bob must be weakly better-off in $\bf{Y}$ than in $\bf{X}$, so $Y_B$ must contain a better item that was not in $X_B$; let’s call this item $y_b$. In allocation $\bf{X}$, item $y_b$ belonged to some other agent, say Carl. From similar considerations, Carl must have in $\bf{Y}$ an item $y_c$ that he prefers to $y_b$. Continuing this way, we end with a cycle of agents, each of whom gave an item to the previous agent and received a *better* item from the next agent.
Now consider the allocation $\bf{Z}$ which is identical to $\bf{X}$ except that the single-item exchanges in the cycle take place (so $y_a$ is given to Alice, $y_b$ is given to Bob and so on). Then $\bf{Z}$ is better than $X$ for all agents in the cycle, and this is true for any additive utility function. Hence, $\bf{X}$ is not possibly-PE.
It is interesting that whereas DD leads to new fairness notions, it does not lead to a new efficiency notion.
Chores and Increasing Differences {#sec:increasing}
=================================
In this section, we assume that we have to divide indivisible *chores*, defined as items with negative utilities. Therefore, all the utility functions we consider in this section assign strictly negative values to all items.
With chores, the Diminishing Differences condition means that the difference between the easiest to the second-easiest chore is larger than the difference between the second-hardest to the hardest chore. But usually, with chores, people care more about not getting the hardest chores than about getting the easiest chores. Therefore, we introduce the condition of *increasing* differences (ID). In many aspects, the ID condition for chores is analogous to the DD condition for goods (subsection \[sub:id-basic\]). However, finding necessarily-ID-fair allocation for chores is more difficult than necessarily-DD-fair allocation for goods (subsections \[sub:id-fair\],\[sub:id-2\],\[sub:id-3\]).
Increasing differences — basic definitions {#sub:id-basic}
------------------------------------------
The following definition is analogous to Definition \[def:dd-utilities\]:
\[def:id-utilities\] Let $\succ$ be a preference relation and $u$ a utility function consistent with $\succ$. We say that $u$ has the *Increasing Differences (ID)* property if, for every three items with consecutive levels $x_3\succ x_2\succ x_1$ such that ${\text{\sc Lev}}(x_3)={\text{\sc Lev}}(x_2)+1={\text{\sc Lev}}(x_1)+2$, it holds that $u(x_3)-u(x_2)\leq u(x_2)-u(x_1)$.
We denote by ${\mathcal{U^{ID}}}({\succ\xspace})$ the set of all ID utility functions consistent with ${\succ\xspace}$.
Given $n$ rankings ${\succ\xspace}_1,\ldots,{\succ\xspace}_n$, We denote by ${\mathcal{U^{ID}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$ the set of all vectors of ID utility functions, $u_1,\ldots,u_n$, such that $u_i$ is consistent with ${\succ\xspace}_i$.
There is a one-to-one correspondence between DD utilities and ID utilities. Given a strict ranking $\succ$, define its reverse ranking ${\succ_{\textrm{rev}}}$ as:
$$\begin{aligned}
\forall x,y\in{\mathcal{M}}: &&
y {\succ_{\textrm{rev}}} x\iff x\succ y \end{aligned}$$
Given a utility function $u$, define its reverse function ${u_{\textrm{rev}}}$ as: $$\begin{aligned}
\forall x\in{\mathcal{M}}:&&{u_{\textrm{rev}}}(x) := - u(x)\end{aligned}$$ Then:
\[lem:dd-id\] For every ranking $\succ$ and utility function $u$: $$\begin{aligned}
{u_{\textrm{rev}}} \in {\mathcal{U^{ID}}}({\succ_{\textrm{rev}}})
\iff
u\in{\mathcal{U^{DD}}}(\succ).\end{aligned}$$
Clearly ${u_{\textrm{rev}}}$ is consistent with ${\succ_{\textrm{rev}}}$, iff $u$ is consistent with $\succ$. Now:
${u_{\textrm{rev}}} \in {\mathcal{U^{ID}}}({\succ_{\textrm{rev}}})$ $\iff$
for every three consecutive items $x_3{\succ_{\textrm{rev}}} x_2{\succ_{\textrm{rev}}} x_1$:\
$~~~~~~~~{u_{\textrm{rev}}}(x_3)-{u_{\textrm{rev}}}(x_2)\leq {u_{\textrm{rev}}}(x_2)-{u_{\textrm{rev}}}(x_1)$ $\iff$
for every three consecutive items $x_1\succ x_2\succ x_3$:\
$~~~~~~~~-u(x_3)- - u(x_2)\leq - u(x_2)- - u(x_1)$ $\iff$
for every three consecutive items $x_1\succ x_2\succ x_3$:\
$~~~~~~~~~u(x_1)-u(x_2) \geq u(x_2) - u(x_3)$ $\iff$
$u\in {\mathcal{U^{DD}}}(\succ)$.
Some examples of members in ${\mathcal{U^{ID}}}(\succ)$ are:
- The negative-Borda utility function: $u_{-Borda}(x) := {\text{\sc Lev}}(x)-M-1$, by which the chore utilities, from easiest to hardest, are $-1,\ldots,-M$.
- The negative-lexicographic utility function: $u_{-Lex}(x) := - 2^{M-{\text{\sc Lev}}(x)}$. By this function, the bundles are first ranked by whether they contain the worst chore, then by whether they contain the next-worst chore, etc.
An equivalent definition of ${\mathcal{U^{ID}}}$ is given by the following lemma. It is analogous to Lemma \[lem:u-in-DD\] and proved in a similar way, so we omit the proof:
\[lem:u-in-ID\] $u\in {\mathcal{U^{ID}}}({\succ\xspace})$ iff, for every four items $x_2, y_2, x_1, y_1$ with $x_2\succeq x_1$ and $y_2\succeq y_1$ and $x_2\neq y_2$ and $x_1\neq y_1$: $$\begin{aligned}
\frac{u(x_2)-u(y_2)}{{\text{\sc Lev}}(x_2)-{\text{\sc Lev}}(y_2)}
\leq
\frac{u(x_1)-u(y_1)}{{\text{\sc Lev}}(x_1)-{\text{\sc Lev}}(y_1)}\end{aligned}$$
Analogously to Definition \[def:dd-relation\], we have the following definitions.
\[def:id-relation\] Given a ranking $\succ$ and two (multi-)bundles $X,Y$: $$\begin{aligned}
X \succsim^{NID} Y
&&
\iff
&&
\forall u\in {\mathcal{U^{ID}}}({\succ\xspace}) \text{:~~~~~} u(X)\geq u(Y)
\\
X \succsim^{PID} Y
&&
\iff
&&
\exists u\in {\mathcal{U^{ID}}}({\succ\xspace}) \text{:~~~~~} u(X)\geq u(Y)
\end{aligned}$$
The NID/PID relations are closely related to the NDD/PDD relations:
\[lem:ndd-nid\] Let $\succ$ be a ranking and ${\succ_{\textrm{rev}}}$ its inverse ranking. Then, for every two multi-bundles $X,Y$: $$\begin{aligned}
X \succsim^{NID} Y
&&
\iff
&&
Y {\succsim_{\textrm{rev}}}^{NDD} X
\\
X \succsim^{PID} Y
&&
\iff
&&
Y {\succsim_{\textrm{rev}}}^{PDD} X\end{aligned}$$
By Lemma \[lem:dd-id\]:
$X \succsim^{NID} Y$ $\iff$
$\forall u\in {\mathcal{U^{ID}}}({\succ\xspace}) \text{:~~~~~} u(X)\geq u(Y)$ $\iff$
$\forall {u_{\textrm{rev}}}\in {\mathcal{U^{DD}}}({{\succ\xspace}_{\textrm{rev}}}) \text{:~~~~~} {u_{\textrm{rev}}}(X)\leq {u_{\textrm{rev}}}(Y)$ $\iff$
$Y {\succsim_{\textrm{rev}}}^{NDD} X$.
Similarly:
$X \succsim^{PID} Y$ $\iff$
$\exists u\in {\mathcal{U^{ID}}}({\succ\xspace}) \text{:~~~~~} u(X)\geq u(Y)$ $\iff$
$\exists {u_{\textrm{rev}}}\in {\mathcal{U^{DD}}}({{\succ\xspace}_{\textrm{rev}}}) \text{:~~~~~} {u_{\textrm{rev}}}(X)\leq {u_{\textrm{rev}}}(Y)$ $\iff$
$Y {\succsim_{\textrm{rev}}}^{PDD} X$.
Thus, to check whether $X \succsim^{NID} Y$ / $X \succsim^{PID} Y$ with regards to some ranking $\succ$, we can simply use Algorithm $\ref{algo:comparedd}$ with the inverse ranking ${\succ_{\textrm{rev}}}$.
We now want to prove an analogue of Theorem \[thm:X-NDD-Y\] for chores. For this, we order the chores in each multi-bundle by *increasing* level, so $X=\{x_1,\ldots, x_{|X|}\}$ where $x_1\preceq_i \ldots \preceq_i x_{|X|}$ (the order between different copies of the same chore is arbitrary). For each $k\leq |X|$ we define $X^{k}$ as the $k$ *worst* chores in $X$, $X^{k}:=\{x_1,\ldots,x_k\}$.
\[thm:X-NID-Y\] Given a ranking $\succ$ and two (multi-)bundles $X,Y$ of chores, $X\succsim^{NID} Y$ if-and-only-if both of the following conditions hold:
1. $|X|\leq |Y|$;
2. For each $k\in \{1,\ldots, |Y|\}$: ${\text{\sc Lev}}(X^{k}) \geq {\text{\sc Lev}}(Y^{k})$.
Note that condition (i) is the opposite of condition (i) in Theorem \[thm:X-NDD-Y\]: $X$ has to have weakly *less* chores than $Y$. However, condition (ii) is identical to condition (ii) in Theorem \[thm:X-NDD-Y\].
By Lemma \[lem:ndd-nid\], $X \succsim^{NID} Y$ iff $Y {\succsim_{\textrm{rev}}}^{NDD} X$. We define the *inverse-level* of an item/bundle as its level under the inverse-ranking ${\succsim_{\textrm{rev}}}$. So the inverse-level of the hardest chore is $M$ and of the easiest chore is $1$. By Theorem \[thm:X-NDD-Y\], this is true iff both of the following conditions hold:
1. $|Y|\geq |X|$;
2. For each $k\in \{1,\ldots, |Y|\}$, the inverse-level of the $k$ chores in $Y$ that are best by ${\succ_{\textrm{rev}}}$ (i.e., worst by $\succ$), is at least as high as the inverse-level of the $k$ chores in $X$ that are worst by $\succ$.
The first condition is equivalent to $|X|\leq |Y|$ and the second condition is equivalent to ${\text{\sc Lev}}(X^{k}) \geq {\text{\sc Lev}}(Y^{k})$.
Increasing differences — fairness concepts {#sub:id-fair}
------------------------------------------
Analogously to Definition \[def:dd-prop\], we have:
\[def:id-prop\] Given item rankings $\succ_1,\ldots,\succ_n$, an allocation $\bf{X}$ is called:
- *Necessary-ID-proportional (NIDPR)* if it is proportional for all utility profiles $u_1,\ldots,u_n$ in ${\mathcal{U^{ID}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
- *Possible-ID-proportional (PIDPR)* if it is proportional for at least one utility profile $u_1,\ldots,u_n$ in ${\mathcal{U^{ID}}}({\succ\xspace}_1,\ldots,{\succ\xspace}_n)$.
Analogously to Lemma \[lem:dd-prop\], and with a similar proof that we omit, we have:
\[lem:id-prop\] Given item rankings $\succ_1,\ldots,\succ_n$:
- An allocation $\bf{X}$ is NIDPR iff $\forall i\in{\mathcal{N}}:~ n\cdot X_i~\succsim_i^{NID}~{\mathcal{M}}$.
- An allocation $\bf{X}$ is PIDPR iff $\forall i\in{\mathcal{N}}:~ n\cdot X_i~\succsim_i^{PID}~{\mathcal{M}}$.
Based on the lemmas proved in subsection \[sub:id-basic\], we have the following analogue of Corollary \[cor:polynomial\]:
\[cor:polynomial-id\] It can be decided in polynomial time whether a given allocation is NIDPR or PIDPR.
In Section \[sec:NDDPR\] we proved that an NDD-proportional allocation exists whenever the number of items is an integer multiple of the number of agents, and all agents have different best items. At first glance, the natural extension of this condition to chores is that all agents should have different worst chores. The following two examples show that this condition is neither sufficient nor necessary.
\[exm:nid-not-sufficient\] There are eight chores and four agents with rankings: $$\begin{aligned}
A:&& a\succ b\succ c\succ d
\succ w\succ x\succ y\succ z
\\
B:&& b\succ c\succ d\succ a
\succ w\succ x\succ z\succ y
\\
C:&& c\succ d\succ a\succ b
\succ w\succ z\succ y\succ x
\\
D:&& d\succ a\succ b\succ c
\succ x\succ z\succ y\succ w\end{aligned}$$ Each agent has a different best chore and each agent has a different worst chore. However, at least one agent (the one who receives $y$) has a second-worst chore. This implies that an NIDPR allocation does not exist. To see this, suppose that all agents have the same ID scoring function: $$\begin{aligned}
-996,-997,-998,-999,
-1000,-2000,-3000,-4000\end{aligned}$$ The utility of the agent who receives $y$ is at most $-3996$. However, the total value is $-13990$ and the fair share is $-13990/4 = -3497.5$.
\[exm:nid-not-necessary\] There are three chores and three agents with rankings: $$\begin{aligned}
A:&& x\succ y\succ z
\\
B:&& x\succ z\succ y
\\
C:&& x\succ z\succ y\end{aligned}$$ All agents have the same best chore, and two agents have the same worst chore. However, the following allocation is NIDPR: $$\begin{aligned}
A: \{y\} && B: \{x\} && C: \{z\}\end{aligned}$$ This is obvious for Bob since he receives his best (easiest) chore. To see that it is also true for Alice, we show that $3\cdot X_A \succsim^{NID}_A {\mathcal{M}}$ using Theorem \[thm:X-NID-Y\]. Condition (i) clearly holds since both multi-bundles have 3 chores. For Condition (ii), compare the levels of the $k$ worst chores, for $k=1,2,3$:
------------------------ ------- ------- -------
$k=1$ $k=2$ $k=3$
$3\cdot X_A$ $2$ $2$ $2$
${\mathcal{M}}$ $1$ $2$ $3$
Difference $+1$ $0$ $-1$
Accumulated difference $+1$ $+1$ $0$
------------------------ ------- ------- -------
The accumulated difference is always at least 0, so $3\cdot X_A \succsim^{NID}_A {\mathcal{M}}$. By a similar calculation, $3\cdot X_C \succsim^{NID}_C {\mathcal{M}}$. Hence the allocation is NIDPR.
Below we present a different condition that is necessary for the existence of NIDPR allocations. It is analogous to the “only-if” part of Theorem \[thm:nddpr\]. To define this condition, for each agent $i$, let $W_i$ be the set of $i$’s $\lceil{n-1\over 2}\rceil$ worst chores.
\[thm:nidpr-necessary\] If there exists a NIDPR allocation of chores among $n$ agents, then both the following conditions must hold:
\(a) The number of chores is $M = m\cdot n$, for some integer $m$.
\(b) It is possible to allocate to each agent $i$, $m$ chores that are not from $W_i$.\
(Hence, the intersection of all $\lceil{n-1\over 2}\rceil$-worst-chores sets is empty: $\cap_{i\in{\mathcal{N}}} W_i = \emptyset$).
Let $(X_1,\ldots,X_n)$ be an NIDPR allocation. Then for every agent $i$,\
$n\cdot X_i ~\succsim^{NID}_i~ {\mathcal{M}}$. By Theorem \[thm:X-NID-Y\].
\(a) For every $i\in{\mathcal{N}}$: $|n\cdot X_i| \leq |{\mathcal{M}}|
\implies n\cdot |X_i| \leq M.$ But this must be an equality since the total number of items in all $n$ bundles is exactly $M$. So the total number of items is $n\cdot |X_i|$ which is an integer multiple of $n$.
\(b) For every $i\in{\mathcal{N}}$, the level of the $n$ worst chores in $n\cdot X_i$ must be weakly larger than the level of the $n$ worst chores in ${\mathcal{M}}$. The $n$ worst chores in ${\mathcal{M}}$ have levels $1,\ldots,n$, so their total level is ${n(n+1)\over 2}$. The $n$ worst chores in $n\cdot X_i$ are just $n$ copies of the worst chore in $X_i$. Thus, the level of this chore must be at least ${n(n+1)\over 2} / n = {n+1\over 2}$. Since levels are integers, the smallest level in $X_i$ must be at least $\lceil {n+1\over 2} \rceil$. So the agent must not get any of his $\lceil {n-1\over 2} \rceil$ worst chores. In other words, agent $i$ must not get any chore from the set $W_i$. Since all chores must be allocated, no chore may be in the intersection of all $W_i$.
In Example \[exm:nid-not-sufficient\], $\lceil {n-1\over 2}\rceil = 2$, and the intersection of the 2-worst-chores sets is not empty (it contains chore $y$), so an NIDPR allocation does not exist. In Example \[exm:nid-not-necessary\], $\lceil {n-1\over 2}\rceil = 1$, the intersection of the worst-chore sets is empty (since not all three agents have the same worst chore), and an NIDPR allocation exists.
We do not know if the condition of Theorem \[thm:nidpr-necessary\] is also sufficient for the existence of NIDPR allocations in general. Below we prove that they are sufficient in two special cases: two agents, and three agents with “almost” identical rankings.
NIDPR allocation for two agents {#sub:id-2}
-------------------------------
With two agents, for each $i\in\{1,2\}$, the set $W_i$ contains just the worst chore of agent $i$, so the necessary condition of Theorem \[thm:nidpr-necessary\] simply says that each agent has a different worst chore. We show that this condition is also sufficient for the existence of NIDPR allocations. The following is a partial analogue to the “if” part of Theorem \[thm:nddpr\].
\[thm:nidpr-sufficient-2\] There exists a NIDPR allocation of chores among $n=2$ agents whenever the following conditions both hold:
\(a) The number of chores is $M = m\cdot n$, for some integer $m$.
\(b) The worst chores of the agents are different.
In case it exists, it can be found in time $O(M)$.
We prove that the balanced-round-robin algorithm (Algorithm \[algo:roundrobin\]), used in the proof of Theorem \[thm:nddpr\], produces an NIDPR allocation in our case. Let $X_i$ be the bundle allocated to agent $i$ by this algorithm. We prove that $n\cdot X_i \succsim^{NID}_i {\mathcal{M}}$ by the two conditions of Theorem \[thm:X-NID-Y\].
Condition (i) is satisfied with equality, since by (a) each agent gets exactly $m$ items, so $|n\cdot X_i| = n m = M = |{\mathcal{M}}|$.
Condition (ii) says that, for every $k{\in\{1,\ldots,M\}}$, the total level of the $k$ worst chores in the multi-bundle $n\cdot X_i$ is at least as large as the total level of the $k$ worst chores in ${\mathcal{M}}$. In other words, the level-difference ${\text{\sc Lev}}_i((n\cdot X_i)^{k}) - {\text{\sc Lev}}_i({\mathcal{M}})$ must be at least 0 for every $k$. Note that each agent receives chores from best to worst. We are interested in levels of chores from worst to best, so we analyze the algorithm from the last round towards the first. We check that the total level-difference (the variable TotalLevelDiff in Algorithm \[algo:comparedd\]) is always at least 0.
The worst chore in $X_i$ is the chore given to agent $i$ in the last round. In the last round, only two chores remain to be allocated. By condition (b) the worst chores of the agents are different, hence the best remaining chores of the agents are also different. Hence, in the last round each agent receives a chore with a level of at least 2, so the level of the worst $2$ chores in $n\cdot X_i$ is at least $2$. The following table shows the levels and their differences for $k\in\{1,2\}$:
----------------- ---------- ----------
$k=1$ $k=2$
$n\cdot X_i$ $\geq 2$ $\geq 2$
${\mathcal{M}}$ $1$ $2$
LevelDiff $\geq 1$ $\geq 0$
TotalLevelDiff $\geq 1$ $\geq 1$
----------------- ---------- ----------
Clearly the level difference is weakly-positive and there is a total accumulated difference of at least 1.
We now prove that, for each round $r\geq 1$ (counting from the last one), the accumulated level-difference for $k\in\{2 r - 1, 2 r \}$ is at least $1$ when $r$ is odd, and at least $4 - 2 i$ when $r$ is even. In particular, it is always at least 0. The proof is by induction on $r$. We have just proved the base $r=1$.
Suppose now that $r$ is even and $r>1$. When agent $i$ gets a chore, the number of remaining chores is $r n + 1 - i$, so agent $i$’s best remaining chore has a level of at least $r n + 1 - i = 2 r + 1 - i$. Therefore, the level-differences for $k\in\{2 r - 1, 2 r \}$ are as in the following table (the last row takes into account an accumulated difference of at least $+1$ from the induction assumption):
----------------- -------------------- --------------------
$n\cdot X_i$ $\geq 2 r + 1 - i$ $\geq 2 r + 1 - i$
${\mathcal{M}}$ $2 r - 1$ $2 r$
LevelDiff $\geq 2 - i$ $ \geq 1 - i$
TotalLevelDiff $\geq 3-i$ $\geq 4 - 2 i$
----------------- -------------------- --------------------
So for both $i$, the accumulated differences are positive and their total is at least $4 - 2 i$, as claimed.
Suppose now that $r$ is odd and $r>1$. When agent $i$ gets a chore, the number of remaining chores is $rn + i - 2$. Therefore, agent $i$’s best remaining item has a level of at least $rn + i - 2 = 2 r + i - 2$. Therefore, the level-differences for $k\in\{2 r-1,r\}$ are as in the following table(the last row takes into account an accumulated difference of at least $4 - 2 i$ from the induction assumption):
----------------- -------------------- --------------------
$n\cdot X_i$ $\geq 2 r + i - 2$ $\geq 2 r + i - 2$
${\mathcal{M}}$ $2 r - 1$ $2 r$
LevelDiff $\geq i-1$ $\geq i-2$
TotalLevelDiff $\geq 3-i$ $\geq 1$
----------------- -------------------- --------------------
So for both $i$, the accumulated differences are positive and their total is at least $1$, as claimed.
NIDPR allocations for three agents {#sub:id-3}
----------------------------------
Theorem \[thm:nidpr-sufficient-2\] does not generalize to three or more agents. The balanced-round-robin protocol does not necessarily find a NIDPR allocation, even if it exists. In Example \[exm:nid-not-necessary\], the rankings satisfy the necessary condition of Theorem \[thm:nidpr-necessary\], and an NIDPR allocation exists, but the round-robin protocol (in the order A B C) yields the allocation: $$\begin{aligned}
A: \{x\} && B: \{z\} && C: \{y\}\end{aligned}$$ which is not NIDPR since it gives Carl his worst chore.
For three agents, we consider the following special case:
- All agents have the same $n$ worst chores;
- All agents have the same $M-n$ best chores, and rank them identically.
In some sense this is a “worst case” of fair allocations, since the agents’ preferences are as similar as they can be without violating the necessary condition.
We prove that, in this “worst case”, the necessary condition of Theorem \[thm:nidpr-necessary\] is also sufficient.
\[thm:nidpr-sufficient-3\] There exists a NIDPR allocation of chores among $n=3$ agents whenever the following conditions hold:
\(a) The number of chores is $M = m\cdot n$, for some integer $m$.
\(b) Not all agents have the same worst chore;
\(c) All agents have the same $n$ worst chores;
\(d) All agents have the same $M-n$ worst chores and rank them identically.
In this case, it can be found in time $O(M)$.
We first allocate the $n$ worst chores. By condition (b), it is possible to give each agent a chore with a level of at least 2. Moreover, by simple case analysis it is possible to see that it is always possible to give at least one agent a chore with a level of at least 3. Hence, after this step, the total level-differences of all agents are at least 0:
----------------- ------- ------- -------
$k=1$ $k=2$ $k=3$
$n\cdot X_i$ $2$ $2$ $2$
${\mathcal{M}}$ $1$ $2$ $3$
LevelDiff $1$ $0$ $-1$
TotalLevelDiff $1$ $1$ $0$
----------------- ------- ------- -------
and the total level-difference of at least one agent is $3$:
----------------- ------- ------- -------
$k=1$ $k=2$ $k=3$
$n\cdot X_i$ $3$ $3$ $3$
${\mathcal{M}}$ $1$ $2$ $3$
LevelDiff $2$ $1$ $0$
TotalLevelDiff $2$ $3$ $3$
----------------- ------- ------- -------
We now have $M-n$ remaining chores. By condition (d), the levels of these chores are the same for all agents, namely, $4,\ldots, M$. We allocate them from worst ($4$) to best ($M$), using a round-robin protocol. There are $m-1$ allocation rounds; in each round, the first (worst) chore is given to an agent whose TotalLevelDiff is at least 3. We prove by induction that, indeed, when each round ends, there is at least one agent with TotalLevelDiff at least 3, while all other agents have TotalLevelDiff at least 0.
The induction base ($r=1$) was already proved above. Assume the claim is true until the beginning of some round $r$. The level of the next chore to allocate is $3 r - 2$. It is given to an agent with TotalLevelDiff at least 3, so his levels change as follows:
----------------- ------------- ------------- -----------
$k=3 r - 2$ $k=3 r - 1$ $k=3 r$
$n\cdot X_i$ $3 r - 2$ $3 r - 2$ $3 r - 2$
${\mathcal{M}}$ $3 r - 2$ $3 r - 1$ $3 r$
LevelDiff $0$ $-1$ $-2$
TotalLevelDiff $\geq 3$ $\geq 2$ $\geq 0$
----------------- ------------- ------------- -----------
The next chore is $3 r - 1$. It is given to an agent with TotalLevelDiff at least 0, so his levels change as follows:
----------------- ------------- ------------- -----------
$k=3 r - 2$ $k=3 r - 1$ $k=3 r$
$n\cdot X_i$ $3 r - 1$ $3 r - 1$ $3 r - 1$
${\mathcal{M}}$ $3 r - 2$ $3 r - 1$ $3 r$
LevelDiff $1$ $0$ $-1$
TotalLevelDiff $\geq 1$ $\geq 1$ $\geq 0$
----------------- ------------- ------------- -----------
The next chore is $3 r1$. It is given to an agent with TotalLevelDiff at least 0, so his levels change as follows:
----------------- ------------- ------------- ----------
$k=3 r - 2$ $k=3 r - 1$ $k=3 r$
$n\cdot X_i$ $3 r$ $3 r$ $3 r$
${\mathcal{M}}$ $3 r - 2$ $3 r - 1$ $3 r$
LevelDiff $2$ $1$ $0$
TotalLevelDiff $\geq 2$ $\geq 3$ $\geq 3$
----------------- ------------- ------------- ----------
As claimed, after round $r$ ends, all agents have TotalLevelDiff at least 0, and one agent has TotalLevelDiff at least 3.
Hence, the resulting allocation is NIDPR.
Currently we do not know how to generalize Theorem \[thm:nidpr-sufficient-3\] for agents with general valuations, or for more than 3 agents. In particular, the following natural protocol does not work. Suppose we divide, instead of chores, *exemptions* from chores. For each of the $M$ chores, create $(n-1)M$ exemptions. In each round, give each agent an exemption from the worst chore for which he is not yet exempted. Finally, allocate to each agent the chores for which he has no exemption. This protocol guarantees that, if the necessary conditions of Theorem \[thm:nidpr-necessary\] hold, each agent will not receive any chore from the set $W_i$. In particular, this protocol finds the NIDPR allocation in Example \[exm:nid-not-necessary\]. However, in general it might yield infeasible allocations. For example, with three agents, six chores and the following rankings: $$\begin{aligned}
A:&& a \succ b\succ c\succ x\succ y\succ z
\\
B:&& c \succ b\succ z\succ a\succ x\succ y
\\
C:&& c \succ b\succ z\succ a\succ x\succ y\end{aligned}$$ when dividing exemptions in the order A B C, C B A, A B C, C B A, they are taken in the order: $z y y, x x c, b a a, z b ?$ where in the last step, the only remaining exemption is for chore $c$, from which Alice is already exempted. Theorem \[thm:nidpr-sufficient-3\] can be extended to more than 3 agents. Whenever the worst $n$ chores can be allocated such that the total level-difference of all agents is at least 0 and the total level-difference of some agents is sufficiently high, it is possible to allocated the other chores such that the total level-difference of all agents remains at least 0. Moreover, instead of requiring that all agents have exactly the same ranking to their $M-n$ best chores, it is sufficient that all agents have the same worst $n$ chores (levels $1,\ldots,n$), the same next-worst $n$ chores (levels $n+1,\ldots,2n$), etc. We omit these results since we believe that the main interesting challenge is generalizing the theorem to arbitrary rankings. Finding a general sufficient condition and protocol for NIDPR allocation of chores remains an interesting open question.
Binary Utilities {#sec:binary}
================
In this section, we compare the diminishing/increasing differences assumptions to another natural assumption, which we call *Binary*. It is based on the assumption that each agent only cares about getting as many as possible of his $k$ best items, where $k$ is an integer that may be different for different agents. The binary assumption was also studied by @bouveret2008efficiency [Proposition 21], who proved that finding an efficient envy-free allocations with such preferences is NP-complete.
The following definition is analogous to Definitions \[def:dd-utilities\] and \[def:id-utilities\]:
\[def:bin-utility\] Let $\succ$ be a preference relation and $u$ a utility function consistent with $\succ$. We say that $u$ is *Binary* if, for some integer $k\geq 1$: $$\begin{aligned}
u(x) =
\begin{cases}
1 & \text{when~} {\text{\sc Lev}}(x)\geq k
\\
0 & \text{when~} {\text{\sc Lev}}(x)< k
\end{cases}\end{aligned}$$ We denote by ${\mathcal{U^{BIN}}}({\succ\xspace})$ the set of all binary utility functions consistent with ${\succ\xspace}$.
The following definition is analogous to Definitions \[def:dd-relation\] and \[def:id-relation\]:
\[def:bin-relation\] For (multi-)bundles $X_i,Y_i$: $$\begin{aligned}
X_i \succsim^{NBIN}_i Y_i
&&
\iff
&&
\forall u_i\in {\mathcal{U^{BIN}}}({\succ\xspace}_i) \text{:~~~~~} u_i(X_i)\geq u_i(Y_i)
\\
X_i \succsim^{PBIN}_i Y_i
&&
\iff
&&
\exists u_i\in {\mathcal{U^{BIN}}}({\succ\xspace}_i) \text{:~~~~~} u_i(X_i)\geq u_i(Y_i)
\end{aligned}$$
For every fairness concept, NBIN-fairness and PBIN-fairness can be defined based on the $\succsim^{NBIN}$ and $\succsim^{PBIN}$ relations, analogously to Definitions \[def:dd-prop\] and \[def:dd-ef\] and \[def:id-prop\].
\[def:bin-prop\] An allocation $X$ is called
- *Necessary-Binary-proportional (NBINPR)* if:\
$\forall i\in{\mathcal{N}}:~ n\cdot X_i~\succsim_i^{NBIN}~{\mathcal{M}}$.
- *Possible-Binary-proportional (PBINPR)* if:\
$\forall i\in{\mathcal{N}}:~ n\cdot X_i~\succsim_i^{PBIN}~{\mathcal{M}}$.
The Binary assumption appears much more restrictive than the DD assumption. For every ${\succ\xspace}$, the set ${\mathcal{U^{BIN}}}({\succ\xspace})$ contains only $M$ utility functions — much less than ${\mathcal{U^{DD}}}({\succ\xspace})$. Therefore, one could expect NBIN-fairness to be easier to satisfy than NDD-fairness. But this is not the case: NBIN-fairness is equivalent to necessary fairness and PBIN-fairness is equivalent to possible fairness. This follows from the following theorem.
For every item-ranking $\succ$ and every multi-bundles $X,Y$:
\(a) $X \succeq^{Nec} Y$ if-and-only-if $X \succeq^{NBIN} Y$ and
\(b) $X \succeq^{Pos} Y$ if-and-only-if $X \succeq^{PBIN} Y$.
It is sufficient to prove the following directions:
\(a) If $X \succeq^{NBIN} Y$ then $X \succeq^{Nec} Y$;
\(b) If not $X \succeq^{PBIN} Y$ then not $X \succeq^{Pos} Y$.
For the proof, we use the following notation.
- The $M$ items are denoted by their level, so the best item is $M$ and the worst is $1$.
- For a multi-bundle $X$ and an item $j$, the number of copies of $j$ in $X$ is denoted $X[j]$.
- The $M$ utility functions in ${\mathcal{U^{BIN}}}(\succ)$ are denoted by $U_k$, for $k{\in\{1,\ldots,M\}}$.
In this notation, for every $k{\in\{1,\ldots,M\}}$ and multi-bundle $X$: $$\begin{aligned}
\label{eq:U_k}
U_k(X) =
\sum_{j=k}^M X[j]
$$ so $U_M(X)=X[M]$ (the agent cares only about the best item), $U_{M-1}(X)=X[M]+X[M-1]$ (the agent cares only about the two best items), etc.
Moreover, for every function $u$ in ${\mathcal{U}}(\succ)$ and multi-bundle $X$: $$\begin{aligned}
\label{eq:u}
u(X) &=
\sum_{j=1}^M u(j)\cdot X[j]
$$ Substituting the $X[j]$ in using gives: $$\begin{aligned}
u(X) &=
u(M)\cdot U_M(X) + \sum_{j=2}^{M}
u(j-1)\cdot \bigg(U_{j-1}(X)-U_{j}(X)
\bigg)
\\
&=
\sum_{j=2}^{M}
\bigg(u(j) - u(j-1)\bigg)\cdot U_j(X)
+
u(1) \cdot U_1(X)\end{aligned}$$ so every additive function $u$ is a linear combination of the functions $U_k$. Note that all coefficients in this linear combination are non-negative. Hence:
- If $\forall k{\in\{1,\ldots,M\}}$: $U_k(X)\geq U_k(Y)$, then $\forall u\in\mathcal{U}(\succ): u(X)\geq u(Y)$. This implies (a).
- If $\forall k{\in\{1,\ldots,M\}}$: $U_k(X) < U_k(Y)$, then $\forall u\in\mathcal{U}(\succ): u(X)<u(Y)$. This implies (b).
Conclusions and Future Work
===========================
We formalized natural ways to compare sets by using the DD (diminishing differences) assumption for goods and ID (increasing differences) assumption for chores. The relations lead to new fairness concepts which we studied in detail. Two main open questions remain for future work: one about envy-free allocation (Section \[sec:NDDEF\]), and one about allocation of chores (Section \[sec:increasing\]). Besides these questions, it may be interesting to extend the results to the case where agents may express weak preferences between items.
Additionally, it can be interesting to identify other interesting set extensions that correspond to classes of utility functions. For example, suppose that agents care both about getting a best item and about not getting a worst item, but do not care much about intermediate items (so the differences in utilities are decreasing at first and then increasing). What can be said of fair allocations under this assumption?
Acknowledgments {#acknowledgments .unnumbered}
===============
We acknowledge the Dagstuhl Seminar 16232 on Fair Division where this project was initiated. We are grateful to four anonymous IJCAI reviewers for their very helpful comments.
Haris Aziz is supported by a Scientia Fellowship. Erel Segal-Halevi was supported by the ISF grant 1083/13, the Doctoral Fellowships of Excellence Program and the Mordecai and Monique Katz Graduate Fellowship Program at Bar-Ilan University. Avinatan Hassidim is supported by ISF grant 1394/16.
Proof of technical lemmas {#app:proofs}
=========================
**Lemma \[lem:u-in-DD\].** Given a ranking $\succ$ and a utility function $u$, the following are equivalent:
1. \[item:1\] $u\in {\mathcal{U^{DD}}}({\succ\xspace})$;
2. \[item:2\] For every four items $x_4, x_3, x_2, x_1$ with $x_4\succeq x_2$ and $x_3\succeq x_1$ and $x_4\neq x_3$ and $x_2\neq x_1$:
$$\begin{aligned}
\frac{u(x_4)-u(x_3)}{{\text{\sc Lev}}(x_4)-{\text{\sc Lev}}(x_3)}
\geq
\frac{u(x_2)-u(x_1)}{{\text{\sc Lev}}(x_2)-{\text{\sc Lev}}(x_1)}\end{aligned}$$
**\[item:1\] $\implies$ \[item:2\]**: Let $k:={\text{\sc Lev}}(x_4)-{\text{\sc Lev}}(x_3)$. Then there are $k+1$ items whose level is between $x_4$ and $x_3$ (inclusive). Denote these items by $y_j$ for $j\in\{0,\ldots,k\}$, such that $x_4 = y_{k} \succ y_{k-1} \succ \ldots \succ y_0 = x_3$. Then, the left-hand size of \[item:2\] can be written as: $$\begin{aligned}
\frac{
\sum_{j=1}^{k} u(y_j) - u(y_{j-1})
}{
k
}\end{aligned}$$ This is an arithmetic mean of the $k$ differences $u(y_j) - u(y_{j-1})$, for $j\in\{1,\ldots,k\}$.
Similarly, the right-hand size of \[item:2\] is an arithmetic mean of $k'$ utility-differences of items whose level is between $x_2$ and $x_1$.
By assumption, $x_4\succeq x_2$ and $x_3\succeq x_1$. By the DD property \[item:1\], to each difference in the left-hand side corresponds a weakly-smaller difference in the right-hand size. Therefore, the arithmetic mean in the left-hand side is weakly larger.
**\[item:2\] $\implies$ \[item:1\]**: in \[item:2\], let $x_3$ be the element ranked immediately below $x_4$, let $x_2=x_3$, and let $x_1$ be the element ranked immediately below $x_2$. Then the denominators both equal 1, and $u$ satisfies the DD definition.
**Theorem \[thm:nddef-npcomplete\].** When there are $n\geq 3$ agents and at least $2 n$ items, checking the existence of NDDEF allocations is NP-hard (as a function of $n$).
The proof is similar to the proof of @Bouveret2010Fair for the NP-hardness of checking existence of NecEF allocations. We now present their reduction and show that it works for NDDEF as well.
The proof is by reduction from the *exact-3-cover* problem, whose inputs are:
- A base set of $3 q$ elements;
- A set-family containing $n\geq q$ triplets, $C_1,\ldots,C_n$, each of which contains exactly 3 elements from the base-set.
The question is whether there exist $q$ pairwise-disjoint triplets whose union is the base-set. Given an instance of exact-3-cover, an instance of fair item allocation is constructed as follows:
- To the $3 q$ base elements correspond $3 q$ *main items*, denoted by $Main$. To each triplet $C_i$ corresponds a set of three main items, denoted by $Main_i$, such that $\forall i{\in\{1,\ldots,n\}}: Main_i\subseteq Main$. The sets $Main_i$, like the triplets $C_i$, are *not* necessarily disjoint. We denote by $Main_{-i}$ the main items not in $Main_i$.
- There are also $3 n$ *dummy items* denoted by $Dummy$. To each triplet $C_i$ corresponds a set of three dummy items, denoted by $Dummy_i:=\{d_{i},d_{i'},d_{i''}\}$. All such sets are pairwise disjoint. We denote by $Dummy_{-i}$ the dummy items not in $Dummy_i$.
- There are $3 (n-q)$ *auxiliary items*, denoted by $Aux$. They are partitioned to $n-q$ pairwise-disjoint triplets, denoted by $Aux_j:=\{x_{j},x_{j'},x_{j''}\}$, for $j{\in\{q+1,\ldots,n\}}$. All in all, there are $6 n$ items.
- To each triplet $C_i$ corresponds a set of three *agents*, $Agents_i = \{i,i',i''\}$. The sets $Agents_i$ are pairwise disjoint. All in all, there are $3 n$ agents.
- The preferences of the three agents in $Agents_i$ are, in general: $$\begin{aligned}
Dummy_i \succ Main_i \succ Aux_{q+1} \succ \cdots \succ Aux_{n} \succ Dummy_{-i} \succ Main_{-i}\end{aligned}$$ Their preferences over the three items in $Dummy_i$ are “cyclic”, i.e, for agent $i$ it is $d_{i}\succ d_{i'}\succ d_{i''}$, for agent $i'$ it is $d_{i'}\succ d_{i''}\succ d_{i}$, and for agent $i''$ it is $d_{i''}\succ d_{i}\succ d_{i'}$. Their preferences over the three items in $Main_i$ are cyclic in a similar way. Their preferences over the three items in $Aux_j$, for each $j{\in\{q+1,\ldots,n\}}$, are cyclic in a similar way. Their preferences over $Dummy_{-i}$ and $Main_{-i}$ are arbitrary.
@Bouveret2010Fair prove that there exists a NecEF allocation iff there exists an exact-3-cover. The proof involves three arguments:
1. \[item:nef\] In a NecEF allocation, each agent must receive the same number of items. Here there are $6 n$ items and $3 n$ agents so each agent must get exactly two items. One of these items must be its top dummy item, which is easy to do since the top dummy items of all agents are different. So, it remains to prove that there is an exact-3-cover, if-and-only-if the second items can be allocated in a NecEF way, i.e., such that each agent prefers the worst item in his bundle to the worst item in any other bundle.
2. \[item:cover-nef\] $Cover \implies allocation$: Suppose there is an exact-3-cover, e.g, with the triplets $C_1,\ldots,C_q$. Then, the sets of main items $Main_1,\ldots,Main_q$ are pairwise-disjoint and their union is exactly $Main$. Then, for each $j{\in\{1,\ldots,q\}}$, it is possible to allocate the three items in $Main_j$ to the three agents in $Agents_j$, giving each agent his favorite main item. Let’s call these $3 q$ agents in the triplets $Agents_1,\ldots,Agents_q$, the “lucky agents”. The allocation is NecEF for the lucky agents since their worst item is their 4th-best item while the worst item in any other bundle is at most their 5th-best item (since their three best items are the dummy items and they are already allocated). It remains to determine an allocation for the $3(n-q)$ “unlucky” agents, $Agents_{q+1},\cdots,Agents_n$. For each $j{\in\{q+1,\ldots,n\}}$, give to the three agents in $Agents_j$, the three items in $Aux_j$, giving each agent his favorite item from that triplet. This item is better for them than the worst items in the other bundles, which are from $Main_{-i}$ or $Dummy_{-i}$, so the allocation is NecEF for them too.
3. \[item:nef-cover\] $Allocation \implies cover$: Suppose there is a NecEF allocation. For each $i{\in\{1,\ldots,n\}}$, consider the three agents in $Agents_i$. We claim that either all of them receive a main item from $Main_i$, or none of them does. *Proof:* suppose e.g. that agent $i$ receives item $m_i\in Main_i$ but agent $i'$ does not receive any item from $Main_i$. Then, the allocation of $i$ is $\{d_i,m_i\}$ and the best possible allocation for $i'$ is $\{d_{i'},x_{i'}\}$, where $x_{i'}$ is the auxiliary item preferred by agent $i'$. But for agent $i'$, both items allocated to agent $i$ are better than $x_{i'}$. Therefore agent $i'$ might envy $i$, so the division is not NecEF. Since each main item must be allocated to exactly one agent, there exists an exact-3-cover: the triplet $C_i$ is in the cover if-and-only-if the agents in $Agents_i$ receive the items in $Main_i$.
We now show that the reduction also works for NDDEF. Claim \[item:nef\] works for NDDEF by Theorem \[thm:X-NDD-Y\]. Claim \[item:cover-nef\] clearly works for NDDEF since every NEF allocation is NDDEF. It remains to prove claim \[item:nef-cover\]. Suppose there exists an NDDEF allocation. This allocation is, in particular, envy-free according to the Borda score. We claim that, for each $i{\in\{1,\ldots,n\}}$, either all three agents in $Agents_i$ receive an item from $Main_i$, or none of them does. *Proof:* consider the following two cases:
- One agent, say $i$, receives his main item $m_i$, but the other agents $i',i''$ do not receive their main items. The dummy items give agent $i''$ a Borda-advantage of 1 over $i$. The best second item that can be allocated to $i''$ is his best auxiliary item, but this leaves him with a Borda-disadvantage of 2 relative to $i$, so $i''$ Borda-envies $i$.
- Two agents, say $i',i''$, receive their main items $m_{i'},m_{i''}$, but agent $i$ does not receive his main item. The dummy items give agent $i$ a Borda-advantage of 1 over $i'$. The best second item that can be allocated to $i$ is his best auxiliary item, but this leaves him with a Borda-disadvantage of 2 relative to $i'$, so $i$ Borda-envies $i'$.
Therefore the reduction is valid for NDDEF too, and the theorem is proved.
Existence of NDDEF allocations
==============================
\[claim:positive-3\] Suppose there are $n=3$ agents and $M=3m$ items (for some integer $m$) and (1) all agents have the same $3$ best items and (2) all agents have the same ranking to the other $M-3$ items and (3) each agent has a different best item. Then an NDDEF allocation exists.
We apply Theorem \[thm:X-NDD-Y\]. The first condition is satisfied since there are $3 m $ items. For the second condition, we have to ensure that for each pair of agents $i,j$ and for every integer $k\geq 1$, the Borda difference (according to agent $i$’s ranking) between the $k$ best items of $i$ and the $k$ best items of $j$ is weakly positive. We call this difference the *Borda advantage* of agent $i$.
We focus on one of the agents which we call “agent 1”. By condition (3), agent 1 receives his top item, whose level is $M$. By condition (1), the other two agents have the same three best items as agent 1, so one of them receives an item whose level for agent 1 is $M-1$ and the other receives an item whose level for agent 1 is $M-2$. We call the former “agent 2” and the latter “agent 3”. From the previous discussion, we know that agent 1 has a Borda advantage of $1$ on agent 2 and $2$ on agent 3. Similarly, agent 2 has a Borda advantage of at least $1$ on agent 3. This can be summarized as follows:
> For every pair of agents $i,I$ such that $i<I$, the Borda advantage of agent $i$ on agent $I$ is at least $I-i$.
Now we divide the remaining items (after the 3 best items) as follows:
Repeat as long as there are items:
For $i = 3\ldots 1$: give agent $i$ the best remaining item.
For $i = 1\ldots 3$: give agent $i$ the best remaining item.
The following are true for every two agents $I>i$:
- In the first loop, agent $I$ gains a positive Borda advantage of $I-i$ on agent $i$ and agent $i$ gains a negative Borda advantage of $i-I$ on agent $I$
- In the second loop, agent $i$ gains a positive Borda advantage of $i-I$ on agent $I$ and agent $I$ gains a negative Borda advantage of $I-i$ on agent $i$.
Therefore, the Borda advantage of each agent on every other agent remains weakly positive and the final division is envy-free.
The following example shows that the three conditions of Theorem \[claim:positive-3\] are insufficient when there are $4$ or more agents:
\[exm:nddef-4\] There are 8 items $\{1,\ldots.8\}$ and 4 agents with rankings:
- Alice: $8 \succ 7 \succ 6 \succ 5 \succ 4 \succ 3 \succ 2 \succ 1$
- Bob: $7 \succ 5 \succ 8 \succ 6 \succ 4 \succ 3 \succ 2 \succ 1$
- Carl: $6 \succ 7 \succ 8 \succ 5 \succ 4 \succ 3 \succ 2 \succ 1$
- Dana: $5 \succ 6 \succ 8 \succ 7 \succ 4 \succ 3 \succ 2 \succ 1$
Someone must receive $1$, and his/her Borda score will be exactly $9$. So, to prevent envy, this agent must be allowed to determine exactly who of the other three agents receives which of the other four items. This will necessarily make one of them envious. In particular, there are four cases:
- Alice receives 1, Bob — 2, Carl — 3, Dana — 4, so Bob envies Dana.
- Bob receives 1, Dana — 2, Alice — 3, Carl — 4, Dana envies Carl.
- Carl receives 1, Bob — 2, Alice – 3, Dana — 4, Bob envies Dana.
- Dana receives 1, Carl — 2, Alice — 3, Bob — 4, Carl envies Bob.
We now extend Theorem \[claim:positive-3\] to any number of agents.
Let $k\geq 1$ be an integer. Suppose we give each agent its top item, then give each agent its top remaining item and so on. If we manage to do this $k$ times with no conflict, then the agents’ preferences are called **$k$-good**.
Theorem \[claim:positive-3\] is about 3 agents with $1$-good preferences. Example \[exm:nddef-4\] shows that it is not true for 4 agents with $1$-good preferences. This raises the following question: for what pairs $(n,k)$ an analogue of Theorem \[claim:positive-3\] is true for $n$ agents with $k$-good preferences? Below we give a lower and an upper bound.
\[claim:positive-n\] Suppose there are $n$ agents and $M=n\cdot m$ items (for some integer $m$) and (1) all agents have the same $n\cdot k$ best items and (2) all agents have the same ranking to the other $M-k\cdot n$ items and (3) the agents have $k$-good preferences, where $n \leq (9 k + 3)/ 4 \approx 2.25 k$. Then, an NDDEF allocation exists.
### Phase 1: Initial Borda Advantages
For every $t\in\{1,\dots,k\}$, consider the $t$-th iteration of giving agents their top item.
In iteration $t$, Alice receives her top remaining item. Since the number of items taken so far is $(t-1)n$, the item given to Alice in this iteration is ranked for her $(t-1)n+1$ or higher. Therefore, she gains an additional Borda advantage of at least $+1$ on each of the other $n-1$ agents, at least $+2$ on at least $n-2$ other agents, at least $+3$ on at least $n-3$ other agents, and so on.
At the last iteration, $t=k$ and Alice gains the following Borda advantages:
- At least $k$ on all other $n-1$ agents. [^7]
- At least $(3k/2)$ on at least $n-2$ other agents. [^8]
- At least $2 k$ on at least $n-3$ other agents. [^9]
- ... At least $k + (q+1)\cdot (k/2)$ on at least $n-q$ other agents.
Each of the other agents has similar advantages.
### Phase 2: Agent ordering
The goal of this phase is to determine the order of agents for receiving the remaining items (which are ranked the same by all agents).
The order is determined by the advantages from the previous phase, according to the following procedure:
- Call Alice “agent 1”.
- For $i = 1 \text{ to } n - k - 1$, ask agent $i$: “who is the unnumbered agent on which you have the smallest advantage?”. Call that agent “agent $i+1$” (after this loop, there are $n-k$ numbered agents).
- For $i = 1 \text{ to } n - (3k/2) - 1$, ask agent $i$: “who is the unnumbered agent on which you have the smallest advantage?”. Call that agent “agent $i+(n-k)$”. Let $k'$ denote the number of numbered agents after this loop. Note that $k' \leq (n-k)+(n-3k/2-1) = 2n-(5k/2)-1 \leq (9k+3)/2 - (5k)/2 - 1 = 2k + 1/2$; since $k'$ is integer, $k'\leq 2k$.
- For $i = 1 \text{ to } n - 2k - 1$, ask agent $i$: “who is the unnumbered agent on which you have the smallest advantage?”. Call that agent “agent $i+k'$”. Let $k''$ denote the number of numbered agents after this loop. Note that $k'' = k' + (n-2k-1) \leq (2k+1/2)+(n-2k-1) < n$.
- Number the remaining agents arbitrarily with the remaining numbers (from $k''+1$ to $n$).
The advantages from the first phases can be summarized as follows:
> For every pair of agents $i,I$ such that $i<I$, the Borda advantage of agent $i$ on agent $I$ is at least $I-i\cdot k$.
We prove the above claim by cases based on the difference $I-i$:
- $(I-i) \leq k$. This case is easy since, as explained in Phase 1, every agent has an advantage of at least $k$ on *all* other agents (regardless of their number).
- $k < (I-i) \leq (3 k/2)$. Since $I\leq n$, this case is possible only when: $i < n-k$, so agent $i$ has participated in the first agent-numbering loop. As explained in Phase 1, every agent has an advantage of at least $(3k/2)$ on all other agents except (maybe) a single agent, and this single agent was numbered $i+1$, so $I-i=1$ and it is not in the range $k<(I-i)$. Therefore, in the given range, the advantage of $i$ is at least $(3k/2)$ which is at least $(I-i)$.
- $(3 k/2) < (I-i) \leq 2 k$. Since $I\leq n$, this case is possible only when: $i < n - (3k/2)$, so agent $i$ has participated in the second agent-numbering loop. As explained in Phase 1, every agent has an advantage of at least $2k$ on all other agents except (maybe) two agents, one of whom was numbered $i+1$ and the other was numbered $i+(n-k)$. By assumption, $n-k \leq (5k+3)/4$, and it is at most $3k/2$ for all relevant integers $k,n$. [^10] So both cases are not in the range $(3 k/2)<(I-i)$. Therefore, in the given range, the advantage of $i$ is at least $(2k)$ which is at least $(I-i)$, as claimed.
- $2 k < (I-i)$. This case is possible only when: $i < n - 2k$, so agent $i$ has participated in the third agent-numbering loop. As explained in Phase 1, every agent has an advantage of at least $(5k/2)$ on all other agents except (maybe) three agents: one of them was numbered $i+1$, the other was numbered $i+(n-k)\leq i+(5k+3)/4$, and the third one was numbered $i+k'$ where $k'\leq 2k$. In all cases, $I-i\leq 2k$, so all three cases are not in the range $2 k < (I-i)$. Therefore, in the given range, the advantage of $i$ is at least $(5k/2)$. Since $I\leq n$, the difference $I-i\leq n-1 \leq (9k-1)/4 < 5k/2$, so the advantage is at least $(I-i)$, as claimed.
### Phase 3: Dividing the remaining items
The last phase works exactly as in the case $n=3$ (Theorem \[claim:positive-3\]), and with the same proof:
Repeat as long as there are items:
For $i = n\ldots 1$: give agent $i$ the best remaining item.
For $i = 1\ldots n$: give agent $i$ the best remaining item.
Theorem \[claim:positive-n\] is a positive result: it shows that, with $k$ steps of top-item allocation, we can handle at least $\approx 2.25k$ agents. We are going to prove a corresponding (but not matching) negative result. This negative result relies on the existence of a *vertex expander graph*:
\[lem:expander\] For sufficiently large $n$, there exists a *vertex expander graph* — a graph that satisfies the following properties, for some constants $d\in \mathbb{N}$ and $f\in(0,0.5)$ to be specified later:
- The out-degree of each node is $d$ (we denote the outgoing neighbors of node $i$ by $G_{i,1},\dots,G_{i,d}$);
- For every fraction $g\leq f$, every subset of $g n$ nodes has at least $g n +1$ outgoing neighbors (besides themselves).
We construct a random graph and prove that it has the desired properties with positive probability. To construct the graph, we choose the neighbors of each node independently at random. For simplicity, we assume that the choice is done with replacement, so some nodes might have the same neighbor twice, and that a node may be selected as its own neighbor. Since our goal is to prove a lower bound on the number of common neighbors, this assumption is fine.
Given a specific subset of $g n$ nodes, their neighbors are $d g n$ independent random selections from $n$ elements. A subset is *bad* if, among these selections, at most $2 g n$ are unique. We now upper-bound the probability that a given subset is bad.
In general, when $m$ items are selected from $n$ items with replacement, the probability that the number of unique items is exactly $k$ is approximately: [^11] $$\begin{aligned}
P(k)\approx
{k^m\over n^m}{n\choose k}
\approx
{k^{m-k} n^{n-m} \over (n-k)^{n-k}}\end{aligned}$$ $P(k)$ is apparently increasing with $k$ (at least when $k\leq n/2$, which is the case here), so the probability that *at most* $2 g n$ items are unique is: $$\begin{aligned}
P_{bad}(g) = \sum_{i=1}^{2 g n} P(k) \leq (2 g n)\cdot P(2 g n)\end{aligned}$$ Here, we have $m=d\cdot g$, so: $$\begin{aligned}
P(2 g n) &\approx
{(2 g n)^{(d-2)g n} \cdot n^{n-d g n} \over (n-2 g n)^{n-2 g n}}
\\
&=
{(2 g)^{(d-2)g n}\cdot n^{n - 2 g n} \over (1-2 g)^{n-2 g n} \cdot n^{n - 2 g n}}
\\
&=
\bigg(
{(2 g)^{(d-2)g} \over (1-2 g)^{1-2 g}}
\bigg)^n
\\
P_{bad}(g) &\leq 2 g n \cdot \bigg(
{(2 g)^{(d-2)g} \over (1-2 g)^{1-2 g}}
\bigg)^n\end{aligned}$$ By the union bound, the probability that there exists a bad subset with $g n$ items is at most $P_{bad}$ times the number of subsets of $g n $ items, which is ${n\choose gn}$: $$\begin{aligned}
{n\choose g n} P_{bad}
&\approx {n^n\over (g n)^{g n} (n-g n)^{n-g n}} P_{bad}
\\
&=\bigg({1 \over (g)^{g} (1-g)^{1-g}}\bigg)^n P_{bad}
\\
&\leq
2 (g n) \cdot \bigg(
{(2 g)^{(d-2)g} \over (1-2 g)^{1-2 g} g^g (1-g)^{1-g}}
\bigg)^n\end{aligned}$$ Define the expression in parentheses by $Q(g,d)$: $$\begin{aligned}
Q(g,d) := {(2 g)^{(d-2)g} \over (1-2 g)^{1-2 g} g^g (1-g)^{1-g}}\end{aligned}$$ Our goal is to find $d$ and $f$ such that, for every $g<f$, $Q(g,d)<1$. Numeric calculations suggest several solutions, such as: $$\begin{aligned}
d&=4 ,&& f=0.0128 > 1/79
\\
d&=5 ,&& f=0.088 > 1/12
\\
d&=6 ,&& f=0.1714 > 1/6
\\
d&=7 ,&& f=0.2367 > 1/5
\\
d&=8 ,&& f=0.2852 > 1/4
\\
d&=9 ,&& f=0.3211
\\
d&=10 ,&& f=0.3481 > 1/3
\\
d&=100 ,&& f=0.4922\end{aligned}$$ For each such pair $(d,f)$, for sufficiently large $n$, the probability that there exists a bad subset with $g n$ items is less than 1. Then, there is positive probability that no subset with $g n$ items is bad. This means that there exists a graph in which every subset of $g n$ nodes has at least $2 g n + 1$ neighbors, which is at least $g n +1$ neighbors besides themselves, as required.
We now return to NDDEF allocations.
\[claim:negative-n-2\] Suppose there are $n$ agents and $M=n\cdot m$ items (for some integer $m$) and (1) all agents have the same $k\cdot n$ best items and (2) all agents have the same ranking to the other $M-k\cdot n$ items and (3) the agents have $k$-good preferences, where $n \geq 15.6 k$. Then, it is possible that an NDDEF allocation does *not* exist.
We apply Lemma \[lem:expander\] with $d=9$ and $f=0.3211$. Let $G$ be a graph whose existence is guaranteed by that lemma, and let $K:= (d+1) k / 2 = 5 k$. We construct valuation profiles as follows.
There are $M=n(k+1)$ items. The best item of agent $i$ is item $i$, so the first iteration of giving top items succeeds. Moreover, for every agent $i$, the second-best item is $G_{i,1}$, the third-best item is $G_{i,2}$, ..., and the $(d+1)$-best item is $G_{i,d}$. Hence, after the first iteration, each agent $i$ has a Borda advantage of exactly $j$ over $G_{i,j}$, for $j\in\{1,\dots,d\}$.
After the first iteration is done, the best remaining item of agent $i$ is item $n+i$, so the second iteration of giving top items succeeds, too. Moreover, for every agent $i$, the second-best item is $n+G_{i,d}$, the third-best item is $n+G_{i,d-1}$, ..., and the $(d+1)$-best is $n+G_{i,1}$. Hence, after the second iteration, each agent $i$ has a Borda advantage of exactly $d+1$ over $G_{i,j}$, for $j\in\{1,\dots,d\}$.
The valuations over the top $n k$ items continue in a similar way. In each two iterations, agent $i$ gains an advantage of $d+1$ over his $d$ friends. Therefore, after $k$ iterations, each agent $i$ has a Borda advantage of exactly $(d+1) k / 2 = K$ over all his friends.
For the last $n$ items, all agents have exactly the same ranking. Each allocation of these items corresponds to a permutation $\sigma$ of the integers $\{1,\dots,n\}$, with the interpretation that agent $i$ receives the item with Borda score $\sigma(i)$ (put differently: the agent with $\sigma(i)=n$ picks first, the agent with $\sigma(i)=n-1$ picks second, and so on).
Given a graph $G$, we say that a permutation $\sigma$ is *good for agent $i$* if: $$\begin{aligned}
\forall j\in\{1,\dots,d\}:
\sigma(i) \geq \sigma(G_{i,j}) - K\end{aligned}$$ In words, $\sigma$ is good for agent $i$ if it gives all his friends a Borda advantage of at most $K$ over him.
We say that a permutation is *good for $G$* if, given $G$, it is good for all agents.
If the permutation $\sigma$ corresponds to an NDDEF allocation given $G$, then $\sigma$ must be good for $G$ (goodness is necessary, though not sufficient, for envy-freeness). Therefore, to prove an impossibility result, it is sufficient to prove that the graph $G$ has no good permutations.
By the assumptions of the theorem, $n\geq 15.6 k$. Therefore, $K \leq {5\over 15.6} n < f n$. Therefore, by the guarantees of Lemma \[lem:expander\], every set of $K$ agents has at least $K+1$ friends (besides themselves). We now prove that, in this situation, no permutation can be good.
Let $\sigma$ be any permutation. Define the *poor agents* as the $K$ agents to whom $\sigma$ gives the smallest values ($1,\dots,K$). These poor agents have at least $K +1$ friends besides themselves; at least one of these friends receives in $\sigma$ a value of at least $2K+1$. This is, in particular, more than $i+K$ for every poor agent $i\in\{1,\dots,K\}$. Therefore, $\sigma$ is not good for $G$.
[^1]: A preliminary version of this paper was published in the proceedings of IJCAI 2017 [@segal2017fair]. In the present version, the proofs are substantially revised and simplified, omitted proofs are included, and there are two new sections: one on division of chores (section \[sec:increasing\]) and one on binary utilities (section \[sec:binary\]).
[^2]: Multi-bundles are used mainly as a technical tool during the proofs; our primary results concern simple bundles, that contain (at most) a single copy of each item.
[^3]: We use negative indices so that the order of indices is the same as the order of levels.
[^4]: For simplicity we assume the agents truthfully report their rankings to the algorithm. Even if one or more agents are not truthful, the allocation will still be NDDPR according to the revealed rankings, as long as the revealed rankings satisfy condition (b).
[^5]: The experiment code is available at GitHub:\
https://github.com/erelsgl/PrefLib-Tools/tree/master/experiments/diminishing-differences
[^6]: The original definition of PDDEF is: “there exists a utility profile for which, for every $i$ and $j$, $i$ does not envy $j$.”. This implies that, for every $i$ and $j$, there exists a utility function $u_i$ by which $i$ does not envy $j$, i.e., $X_i \succsim^{PDD}_i X_j$. However, the implication is not true in the other direction when there are three or more agents. For example, if $X_1 \succsim^{PDD}_1 X_2$ and $X_1 \succsim^{PDD}_1 X_3$, then it is possible that agent 1 does not envy agent 2 by some DD function $u_{1,2}$, and does not envy agent 3 by some other DD function $u_{1,3}$, but there is no *single* DD function by which agent 1 envies neither agent 2 nor agent 3.
[^7]: In each iteration, Alice wins an advantage of at least $1$ on every other agent.
[^8]: In each iteration, there is at most one agent on which Alice wins an advantage of at most $1$. Call this agent the “dangerous agent” of iteration $t$. In all $k$ iterations, there is at most one agent who is dangerous for more than $k/2$ iterations. All other $n-2$ agents are dangerous for at most $k/2$ iterations, and in the other iterations they are not dangerous, which means that Alice wins an advantage of at least $2$ over them. All in all, Alice’s advantage over these $n-2$ agents is at least: $(k/2) + (k/2)\cdot 2 = (3k/2)\cdot d_k$.
[^9]: The proof is similar to the one in the previous footnote.
[^10]: For $k=1$, $n\leq 3$ and this was already handled by the previous claim. For $k=2$, $n\leq 5$ and $n-k \leq 3 = 3k/2$. For $k\geq 3$, $(5k+3)/4 \leq (5k+3)/4 + (k/4)-(3/4) = 6k/4 = 3k/2$.
[^11]: see http://math.stackexchange.com/a/1957820/29780 . Based on a calculation by Henry.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a declarative and modular specification of an automated trading system (ATS) in the concurrent linear framework CLF. We implemented it in Celf, a CLF type checker which also supports executing CLF specifications. We outline the verification of two representative properties of trading systems using generative grammars, an approach to reasoning about CLF specifications.'
author:
- Iliano Cervesato Sharjeel Khan Giselle Reis Dragiša Žunić
bibliography:
- 'references.bib'
title: |
Formalization of Automated Trading Systems\
in a Concurrent Linear Framework[^1]
---
Introduction
============
Trading systems are platforms where buy and sell orders are automatically matched. Matchings are executed according to the operational specification of the system. In order to guarantee trading fairness, these systems must meet the requirements of regulatory bodies, in addition to any internal requirement of the trading institution. However, both specifications and requirements are presented in natural language which leaves space for ambiguity and interpretation errors. As a result, it is difficult to guarantee regulatory compliance [@DeBel93]. For example, the main US regulator, the Securities and Exchanges Commission (SEC), has fined several companies, including Deutsche Bank (37M in 2016), Barclay’s Capital (70M in 2016), Credit Suisse (84M in 2016), UBS (19.5M in 2015) and many others [@Freedman2015] for non-compliance.
Modern trading systems are complex pieces of software with intricate and sensitive rules of operation. Moreover they are in a state of continuous change as they strive to support new client requirements and new order types. Therefore it is difficult to attest that they satisfy all requirements at all times using standard software testing approaches. Even as regulatory bodies recently demand that systems must be “fully tested” [@FCA2018], experience has shown that (possibly unintentional) violations often originate from unforeseen interactions between order types [@PassmoreIgnatovich2017].
Formalization and formal reasoning can play a big role in mitigating these problems. They provide methods to verify properties of complex and infinite state space systems with certainty, and have already been applied in fields ranging from microprocessor design [@Jones2001], avionics [@Souyris2009], election security [@Pattison15AJCAI], and financial derivative contracts [@PeytonJones2000; @BahrElsman2015]. Trading systems are a prime candidate as well.
In this paper we use the logical framework CLF [@clftechrep1] to specify and reason about trading systems. CLF is a linear concurrent extension of the long-established LF framework [@Harper93]. Linearity enables natural encoding of state transition, where facts are consumed and produced thereby changing the system’s state. The concurrent nature of CLF is convenient to account for the possible orderings of exchanges.
The contributions of this research are twofold: (1) We formally define an archetypal automated trading system in CLF [@clftechrep1] and implement it as an executable specification in Celf. (2) We demonstrate how to prove some properties about the specification using generative grammars [@Simmons12], a technique for meta-reasoning in CLF.
The paper is organized as follows: Section \[sec:CLF\] introduces the concurrent logical framework CLF. Section \[sec:ATS\] introduces the core concepts related to automated trading systems (ATS), followed by Section \[sec:formlimit\], which presents the formalization of an ATS in CLF/Celf. Section \[sec:proof\] contains proofs of two properties based on generative grammars, going towards automated reasoning in CLF. We conclude and outline possible further developments in Section \[sec:conclusion\].
Concurrent Linear Logic and Celf {#sec:CLF}
================================
The logical framework CLF [@clftechrep1] is based on a fragment of intuitionistic linear logic. It extends the logical framework LF [@Harper93] with the linear connectives $\multimap$, $\binampersand$, $\top$, $\otimes$, $1$ and $!$ to obtain a resource-aware framework with a satisfactory representation of concurrency. The rules of the system impose a discipline on when the synchronous connectives $\otimes$, $1$ and $!$ are decomposed, thus still retaining enough determinism to allow for its use as a logical framework. Being a type-theoretical framework, CLF unifies implication and universal quantification as the dependent product construct. For simplicity we present only the logical fragment of CLF (i.e., without terms) needed for our encodings. A detailed description of the full framework can be found in [@clftechrep1].
We divide the formulas in this fragment of CLF into two classes: *negative* and *positive*. Negative formulas have right invertible rules and positive formulas have left invertible rules. Their grammar is: [$$\begin{array}{rllr}
N, M & ::= & a \multimap N \mid a \rightarrow N \mid \{ P \} \mid \forall x. N \mid a & \text{(negative formulas)}\\
P, Q & ::= & P \otimes Q \mid 1 \mid ! a \mid a & \text{(positive formulas)}
\end{array}$$]{}where $a$ is an atom (i.e., a predicate). Positive formulas are enclosed in the lax modality $\{\cdot\}$, which ensures that their decomposition happens atomically.
$$\infer[1_l]
{{\Gamma; \Delta; {\color{blue} \Psi, 1} \vdash P_0}}
{{\Gamma; \Delta; {\color{blue} \Psi} \vdash P_0}}
\quad
\infer[\otimes_l]
{{\Gamma; \Delta; {\color{blue} \Psi, P \otimes Q} \vdash P_0}}
{{\Gamma; \Delta; {\color{blue} \Psi, P, Q} \vdash P_0}}
\quad
\infer[!_l]
{{\Gamma; \Delta; {\color{blue} \Psi, !a} \vdash P_0}}
{{\Gamma, a; \Delta; {\color{blue} \Psi} \vdash P_0}}
\quad
\infer[\mathsf{st}]
{{\Gamma; \Delta; {\color{blue} \Psi, a} \vdash P_0}}
{{\Gamma; \Delta, a; {\color{blue} \Psi} \vdash P_0}}
\quad
\infer[\mathsf{L}]
{{\Gamma; \Delta; {\color{blue} \cdot} \vdash P_0}}
{{\Gamma; \Delta \vdash P_0}}$$\
$$\infer[1_r]
{{\Gamma; \cdot \vdash {\color{red} 1}}}
{}
\quad
\infer[\otimes_r]
{{\Gamma; \Delta_1, \Delta_2 \vdash {\color{red} P \otimes Q}}}
{{\Gamma; \Delta_1 \vdash {\color{red} P}}
&{\Gamma; \Delta_2 \vdash {\color{red} Q}}}
\quad
\infer[!_r]
{{\Gamma; \cdot \vdash {\color{red} !a}}}
{{\Gamma; \cdot \vdash {\color{red} a}}}
\quad
\infer[\mathsf{R}]
{{\Gamma; \Delta \vdash {\color{red} a}}}
{{\Gamma; \Delta \vdash a}}
\quad
\infer[\mathsf{init}]
{{\Gamma; a \vdash a}}
{}$$\
$$\infer[\multimap_l]
{{\Gamma; \Delta_1, \Delta_2, a \multimap N \vdash F}}
{{\Gamma; \Delta_1 \vdash {\color{red} a}}
&{\Gamma; \Delta_2, N \vdash F}}
\quad
\infer[\rightarrow_l]
{{\Gamma; \Delta, a \rightarrow N \vdash F}}
{{\Gamma; \cdot \vdash {\color{red} a}}
&{\Gamma; \Delta, N \vdash F}}
\quad
\infer[\forall_l]
{{\Gamma; \Delta, \forall x. N \vdash F}}
{{\Gamma; \Delta, N[x \mapsto t] \vdash F}}
\quad
\infer[\{\}_l]
{{\Gamma; \Delta, \{ P \} \vdash P_0}}
{{\Gamma; \Delta; {\color{blue} P} \vdash P_0}}$$\
$$\infer[\multimap_r]
{{\Gamma; \Delta \vdash a \multimap N}}
{{\Gamma; \Delta, a \vdash N}}
\quad
\infer[\rightarrow_r]
{{\Gamma; \Delta \vdash a \rightarrow N}}
{{\Gamma, a; \Delta \vdash N}}
\quad
\infer[\forall_r]
{{\Gamma; \Delta \vdash \forall x. N}}
{{\Gamma; \Delta \vdash N[x \mapsto \alpha]}}
\quad
\infer[\{\}_r]
{{\Gamma; \Delta \vdash \{ P \}}}
{{\Gamma; \Delta \vdash {\color{red} P}}}
\quad
\infer[\mathsf{cont}]
{{\Gamma, N; \Delta \vdash C}}
{{\Gamma, N; \Delta, N \vdash C}}$$
The sequent calculus proof system for this fragment of CLF is presented in Figure \[fig:clf-seq\]. The sequents make use of either two or three contexts on the left: $\Gamma$ contains unrestricted formulas, $\Delta$ contains linear formulas and $\Psi$, when present, contains positive formulas. The decomposition phase of a positive formula is indicated in [red]{} on the right and in [blue]{} on the left. These phases end (by means of rules or ) after the formula is completely decomposed.
Since CLF has both the linear and intuitionistic implications, we can specify computation in two different ways. Simplifying somewhat, linear implication formulas are interpreted as multiset rewriting: the bounded resources on the left are consumed and those on the right are produced. State transitions can be modeled naturally this way. Intuitionistic implication formulas are interpreted as backward-chaining rules *à la* Prolog, providing a way to compute solutions for a predicate by matching it with the head (rightmost predicate) of a rule and solving the body. In this paper, predicates defined by backward-chaining rules are written in [green]{}.
The majority of our encoding involves rules in the following shape (for atomic $p_i$ and $q_i$):\
$p_1 \otimes ... \otimes p_n \multimap \{ q_1 \otimes ... \otimes q_m \}$ which is the uncurried version of: $p_1 \multimap ... \multimap p_n \multimap \{ q_1 \otimes ... \otimes q_m \}$.
This framework is implemented as the tool Celf (<https://clf.github.io/celf/>) which we used for the encodings. Following the tool’s convention, variable names start with an upper-case letter.
{width=".75\columnwidth"}
Automated Trading System (ATS) {#sec:ATS}
==============================
Real life trading systems differ in the details of how they manage orders (there are hundreds of order types in use [@Mackintosh2014]). However, there is a certain common core that guides all those trading systems, and which embodies the market logic of trading on an exchange. We have formalized those elements in what we call an automated trading system, or an ATS. In what follows we introduce the basic notions.
An *order* is an investor’s instruction to a broker to buy or sell securities (or any asset type which can be traded). They enter an ATS sequentially and are *exchanged* when successfully matched against opposite order(s). In this paper, we will be concerned with *limit*, *market* and *cancel* orders.
A *limit order* has a specified limit price, meaning that it will trade at that price or better. In the case of a limit order to sell, a limit price $P$ means that the security will be sold at the best available price in the market, but no less than $P$. And dually for buy orders. If no exchange is possible, the order stays in the market waiting to be exchanged – these are called *resident orders*. A *market order* does not specify the price, and will be immediately matched against opposite orders in the market. If none are available, the order is discarded. A *cancel order* is an instruction to remove a resident order from the market.
A *matching algorithm* determines how resident orders are prioritized for exchange, essentially defining the mode of operation of a given ATS. The most common one is *price/time priority*. Resident orders are first ranked according to their price (increasingly for sell and decreasingly for buy orders); orders with the same price are then ranked depending on when they entered.
Figure \[fig:marketView\] presents a visualization of a (Bitcoin) market. The left-hand side (green) contains resident buy orders, while the right-hand side (pink) contains resident sell orders. The price offered by the most expensive buy order is called *bid* and the cheapest sell order is called *ask*. The point where they (almost) meet is the *bid-ask* spread, which, at that particular moment, was around 2,468 USD.
Standard regulatory requirements for real world trading systems include: the bid price is always strictly less than the ask price (i.e., no locked – ${\mathit{bid}}$ is equal to ${\mathit{ask}}$ – or crossed – ${\mathit{bid}}$ is greater than ${\mathit{ask}}$ – states), the trade always takes place at either ${\mathit{bid}}$ or ${\mathit{ask}}$, the price/time priority is always respected when exchanging orders, the order priority function is transitive, among others.
Formalization of an ATS {#sec:formlimit}
=======================
We have formalized the most common components of an ATS in the logical framework CLF and implemented them in Celf. The formalization of an ATS is divided into three parts. First, we represent the market infrastructure using some auxiliary data structures. Then we determine how to represent the basic order types and how they are organized for processing. Finally we encode the exchange rules which act on incoming orders.
Since we are using a linear framework, the state of the system is naturally represented by a set of facts which hold at that point in time. Each rule consumes some of these facts and generates others, thus reaching a new state. Many operations are dual for buy and sell orders, so, whenever possible, predicates and rules are parameterized by the action ( or , generically denoted $A$). The machinery needed in our formalization includes natural numbers, lists and queues. Their encoding relies on the backward-chaining semantics of Celf.
The full encoding can be found at <https://github.com/Sharjeel-Khan/financialCLF>.
Infrastructure
--------------
The trading system’s infrastructure is represented by the following four linear predicates: [$${{\text{\ttfamily\upshape queue}}(Q)} \qquad {{\text{\ttfamily\upshape priceQ}}(A,~P,~Q)} \qquad {{\text{\ttfamily\upshape actPrices}}(A,~L)} \qquad {{\text{\ttfamily\upshape time}}(T)}$$]{}Predicate ${{\text{\ttfamily\upshape queue}}(Q)}$ represents the queue in which orders are inserted for processing. As orders arrive in the market, they are assigned a timestamp and added to $Q$. For an action $A$ and price $P$, the queue $Q$ in ${{\text{\ttfamily\upshape priceQ}}(A,~P,~Q)}$ contains all resident orders with those attributes. Due to how orders are processed, the queue is sorted in ascending order of timestamp. We maintain the invariant that price queues are never empty. Price queues correspond to columns in the graph of Figure \[fig:marketView\]. For an action $A$, the list $L$ in ${{\text{\ttfamily\upshape actPrices}}(A,~L)}$ contains the exchange prices available in the market, i.e., all the prices on the $x$-axis of Figure \[fig:marketView\] with non-empty columns. Note that the bid price is the maximum of $L$ when $A$ is and the ask price is the minimum when $A$ is . The time is represented by the fact ${{\text{\ttfamily\upshape time}}(T)}$ and increases as the state changes.
The fact is the entry point in our formalization. This fact starts the ATS. It is rewritten to an empty order queue, empty active price lists for and , and the zero time:
[$${\text{\ttfamily\upshape begin}} \multimap \{ {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape empty}})} \otimes {{\text{\ttfamily\upshape actPrices}}({\text{\ttfamily\upshape buy}},~{{\text{\ttfamily\upshape nil}}})} \otimes {{\text{\ttfamily\upshape actPrices}}({\text{\ttfamily\upshape sell}},~{{\text{\ttfamily\upshape nil}}})} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape z}})} \}$$ ]{}
Orders’ Structure
-----------------
An order is represented by a linear fact ${{\text{\ttfamily\upshape order}}(O,~A,~P,~ID,~N)}$, where $O$ is the type of order, $A$ is an action, $P$ is the order price, $\mathit{ID}$ is the identifier of the order and $N$ is the quantity. $P$, $\mathit{ID}$ and $N$ are natural numbers. In this paper, $O$ is one of , , or . An order predicate in the context is consumed and added to the order queue for processing via the following rule:
[$$\begin{array}{rl}
&{{\text{\ttfamily\upshape order}}(O,~A,~P,~ID,~N)} \otimes {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape time}}(T)} \otimes {{\text{\ttfamily\upshape {\color{{Green}} enq}}}(Q,~{{\text{\ttfamily\upshape ordIn}}(O,~A,~P,~ID,~N,~T)},~Q')} \\
\multimap &\{ {{\text{\ttfamily\upshape queue}}(Q')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \}
\end{array}$$ ]{}
The predicate is transformed into a term ${{\text{\ttfamily\upshape ordIn}}(O,~A,~P,~ID,~N,~T)}$ containing the same arguments plus the timestamp $T$. This term is added to the order queue $Q$ by the (backward-chaining) predicate . This queue allows the sequential processing of orders given their time of arrival in the market, thus simulating what happens in reality. The timestamp is also used to define resident order priority. Sequentiality is guaranteed as all state transition rules act only on the first order in the queue. Nevertheless, due to Celf’s non-determinism, orders are added to the queue in an arbitrary order.
Limit Orders
------------
According to the matching logic, there are two basic actions for every limit order in the queue: exchange (partially or completely) or add to the market (becomes resident). The action taken depends on the order’s limit price (at which it is willing to trade), the bid and ask prices, as well as the quantity of resident orders[^2].
#### Adding orders to the market
An order is added to the market when its limit price $P$ is such that it cannot be exchanged against opposite resident orders. Namely when $P < {\mathit{ask}}$ in the case of a buy order, and when $P > {\mathit{bid}}$ in the case of a sell order. There are two rules for adding an order, depending on whether there are resident orders at that price in the market or not (see Figure \[fig:limitOrdersAdd\]). The (backward chaining) predicate is provable when the order cannot be exchanged.
[$$\begin{array}{rl}
{\text{\ttfamily\upshape limit/empty:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes \phantom{.} \\
& {{\text{\ttfamily\upshape {\color{{Green}} store}}}(A,~L',~P)} \otimes {{\text{\ttfamily\upshape actPrices}}(A,~L)} \otimes {{\text{\ttfamily\upshape {\color{{Green}} notInList}}}(L,~P)} \otimes {{\color{{Green}} {\text{\ttfamily\upshape insert}}}(L,~P,~LP)} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes {{\text{\ttfamily\upshape priceQ}}(A,~P,~{{\text{\ttfamily\upshape consP}}(ID,~N,~T,~{\text{\ttfamily\upshape nilP}})})} \\
& \phantom{\{} \otimes {{\text{\ttfamily\upshape actPrices}}(A,~LP)} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\
\\
{\text{\ttfamily\upshape limit/queue:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes \phantom{.} \\
& {{\text{\ttfamily\upshape {\color{{Green}} store}}}(A,~L',~P)} \otimes {{\text{\ttfamily\upshape priceQ}}(A,~P,~PQ)} \otimes {{\text{\ttfamily\upshape {\color{{Green}} extendP}}}(PQ,~ID,~N,~T,~PQ')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap &\{ {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes {{\text{\ttfamily\upshape priceQ}}(A,~P,~PQ')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \}
\end{array}$$ ]{}
The first line is the same for both. Given the order at the of the order queue, the predicate will bind $A'$ to the dual action of $A$ (i.e., if $A$ is , $A'$ will be and vice-versa). Then [($A'$, $L'$)]{} binds $L'$ to the list of active prices of $A'$. The incoming order can be added to the market only if there is no dual resident order at an acceptable price. For example, if $A$ is at price $P$, any resident order with price $P$ *or less* would be an acceptable match. The predicate holds iff there is no acceptable match.
The second line of each rule distinguishes whether the new order to be added is the first one at that price ( rule – ${{\text{\ttfamily\upshape {\color{{Green}} notInList}}}(L,~P)}$) or not. If it is, the active price list is updated by backward chaining on ${{\color{{Green}} {\text{\ttfamily\upshape insert}}}(L,~P,~LP)}$, and rewriting ${{\text{\ttfamily\upshape actPrices}}(A,~L)}$ to ${{\text{\ttfamily\upshape actPrices}}(A,~LP)}$. Additionally, a new price queue is created with that order alone: ${{\text{\ttfamily\upshape priceQ}}(A,~P,~{{\text{\ttfamily\upshape consP}}(ID,~N,~T,~{\text{\ttfamily\upshape nilP}})})}$. If there are resident orders at the same price (and action), the existing price queue is extended with the new order by backward chaining on ${{\text{\ttfamily\upshape {\color{{Green}} extendP}}}(PQ,~ID,~N,~T,~PQ')}$ and by rewriting ${{\text{\ttfamily\upshape priceQ}}(A,~P,~PQ)}$ to ${{\text{\ttfamily\upshape priceQ}}(A,~P,~PQ')}$. Both rules increment the time by one unit.
#### Exchanging orders
The rules for exchanging orders are presented in Figure \[fig:limitOrdersEx\]. A limit order is exchanged when its limit price $P$ satisfies $P \leq {\mathit{bid}}$, in the case of sell orders, or $P \geq
{\mathit{ask}}$ for buy orders.
[$$\begin{array}{rl}
{\text{\ttfamily\upshape limit/1:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes \phantom{a} \\
& {{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A,~L',~P,~X)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~{\text{\ttfamily\upshape nilP}})})} \otimes {{\color{{Green}} {\text{\ttfamily\upshape remove}}}(L',~X,~L'')}~ \otimes\\
& {{\text{\ttfamily\upshape {\color{{Green}} nat-equal}}}(N,~N')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L'')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\
\\
{\text{\ttfamily\upshape limit/2:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes \phantom{a} \\
& {{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A,~L',~P,~X)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~{{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)})})} ~\otimes\\
& {{\text{\ttfamily\upshape {\color{{Green}} nat-equal}}}(N,~N')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)})} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\
\\
{\text{\ttfamily\upshape limit/3:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes \phantom{a} \\
& {{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A,~L',~P,~X)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~{\text{\ttfamily\upshape nilP}})})} \otimes {{\color{{Green}} {\text{\ttfamily\upshape remove}}}(L',~X,~L'')} \otimes \phantom{a} \\
& {{\text{\ttfamily\upshape {\color{{Green}} nat-great}}}(N,~N')} \otimes {{\text{\ttfamily\upshape {\color{{Green}} nat-minus}}}(N,N',~N'')} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N'',~T)},Q))} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L'')} \} \\
\\
{\text{\ttfamily\upshape limit/4:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes \phantom{a} \\
& {{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A,~L',~P,~X)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~{{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)})})} \otimes \phantom{a} \\
& {{\text{\ttfamily\upshape {\color{{Green}} nat-great}}}(N,~N')} \otimes {{\text{\ttfamily\upshape {\color{{Green}} nat-minus}}}(N,N',~N'')} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N'',~T)},Q))} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes \phantom{a} \\
& {{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)})} \} \\
\\
{\text{\ttfamily\upshape limit/5:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes \phantom{a} \\
& {{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A,~L',~P,~X)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~L)})} ~\otimes\\
& {{\text{\ttfamily\upshape {\color{{Green}} nat-less}}}(N,~N')} \otimes {{\text{\ttfamily\upshape {\color{{Green}} nat-minus}}}(N',N,~N'')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID',~N'',~T',~L)})} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \}
\end{array}$$ ]{}
The (backward chaining) predicate binds $X$ to the exchange price (either ${\mathit{bid}}$ or ${\mathit{ask}}$). We distinguish between an incoming order that “consumes” all the quantity available at price $X$, or only a part of the combined quantity available. The arithmetic comparison and operations are implemented in the usual backward-chaining way using a unary representation of natural numbers.
All five rules start the same way: the first line binds $L'$ to the list of active prices of the dual orders. The backward-chaining predicate ${{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A,~L',~P,~X)}$ holds iff there is a matching resident order. In this case, it binds $X$ to the best available market price. The first order in the price queue for $X$ has priority and will be exchanged. Let $N$ be the quantity in the incoming order and $N'$ be the quantity of the resident order with highest priority. There are three cases:
- $N = N'$ (rules and ): Both orders will be completely exchanged. The incoming order is removed from the order queue by continuing with ${{\text{\ttfamily\upshape queue}}(Q)}$. The resident order is removed from its price queue and we distinguish two cases:
- It is the last element in the queue (): then the fact ${{\text{\ttfamily\upshape priceQ}}(\_,~\_,~\_)}$ is not rewritten and $X$ is removed from the list of active prices by the backward chaining predicate ${{\color{{Green}} {\text{\ttfamily\upshape remove}}}(L',~X,~L'')}$. This list is rewritten from ${{\text{\ttfamily\upshape actPrices}}(A',~L')}$ to ${{\text{\ttfamily\upshape actPrices}}(A',~L'')}$.
- Otherwise (), the resident order is removed from the price queue by rewriting ${{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID',~N,~T',~{{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)})})}$ to ${{\text{\ttfamily\upshape priceQ}}(A',~X,~{{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)})}$.
- $N > N'$ (rules and ): The incoming order will be partially exchanged, and not leave the order queue as long as there are matching resident orders. At each exchange its quantity is updated to $N''$, the difference between $N$ and $N'$, computed by the backward-chaining predicate ${{\text{\ttfamily\upshape {\color{{Green}} nat-minus}}}(N,N',~N'')}$. The order queue is rewritten to\
${{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N'',~T)},Q))}$. The resident order will be completely consumed and removed from the market. Therefore, we need to distinguish two cases as before (last element in its price queue – – or not – ).
- $N < N'$ (rule ): The incoming order is completely exchanged and removed from the order queue (rewritten to ${{\text{\ttfamily\upshape queue}}(Q)}$). The resident order is partially exchanged and its quantity is updated to $N''$, computed by the backward chaining rule ${{\text{\ttfamily\upshape {\color{{Green}} nat-minus}}}(N',N,~N'')}$. Notice that the price queue is rewritten with the order in the same position, so its priority does not change.
By convention, the time is only updated once an order is completely processed and removed from the order queue.
Market Orders {#subs:market}
-------------
A market order is meant to be exchanged immediately at current market prices. As long as there are available sellers or buyers, market orders are exchanged. The remaining part of a market order is discarded. Certainty of execution is a priority over the price of execution.
In this case, there are no rules for adding/storing of market orders. The exchange rules are similar to the ones for limit orders, with subtle differences. Market orders do not have a desired price as an attribute, they only have a desired quantity. Therefore exchanging is continued as long as the quantity was not reached and there are available resident orders (the price $P$ is nominally presented but it is never used). In the rules, the predicate ${{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A,~L,~P,~X)}$ is replaced with ${{\color{{Green}} {\text{\ttfamily\upshape mktExchange}}}(A,~L,~X)}$, which simply binds $X$ to the best offer in the market. In the unlikely event that there are no more dual resident orders (verified by checking if $L$ in ${{\text{\ttfamily\upshape actPrices}}(A',~L)}$ is empty), the order is removed from the order queue.
[$$\begin{array}{rl}
{\text{\ttfamily\upshape market/empty:}}
& {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape market}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~{{\text{\ttfamily\upshape nilN}}})} ~\otimes \\
& {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~{{\text{\ttfamily\upshape nilN}}})} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\[2mm]
{\text{\ttfamily\upshape market/1:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape market}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} ~\otimes\\
& {{\color{{Green}} {\text{\ttfamily\upshape mktExchange}}}(A,~L',~Y)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~Y,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~{{\text{\ttfamily\upshape nilP}}})})} \otimes {{\color{{Green}} {\text{\ttfamily\upshape remove}}}(L',~Y,~L'')} ~\otimes \\
& {{\text{\ttfamily\upshape {\color{{Green}} nat-equal}}}(N,~N')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{{{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L'')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\
\\
{\text{\ttfamily\upshape market/2:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape market}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} ~\otimes \\
& {{\color{{Green}} {\text{\ttfamily\upshape mktExchange}}}(A,~L',~Y)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~Y,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~({{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)})}))} ~\otimes \\
& {{\text{\ttfamily\upshape {\color{{Green}} nat-equal}}}(N,~N')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~Y,~({{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)}))} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\
\\
{\text{\ttfamily\upshape market/3:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape market}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes\phantom{a} \\
& {{\color{{Green}} {\text{\ttfamily\upshape mktExchange}}}(A,~L',~Y)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~Y,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~{\text{\ttfamily\upshape nilP}})})} \otimes {{\color{{Green}} {\text{\ttfamily\upshape remove}}}(L',~Y,~L'')} ~\otimes\\
&{{\text{\ttfamily\upshape {\color{{Green}} nat-great}}}(N,~N')} \otimes {{\text{\ttfamily\upshape {\color{{Green}} nat-minus}}}(N,N',~N'')} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N'',~T)},Q))} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L'')} \} \\
\\
{\text{\ttfamily\upshape market/4:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape market}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes\phantom{a} \\
& {{\color{{Green}} {\text{\ttfamily\upshape mktExchange}}}(A,~L',~Y)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~Y,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~({{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)})}))} ~\otimes\\
&{{\text{\ttfamily\upshape {\color{{Green}} nat-great}}}(N,~N')} \otimes {{\text{\ttfamily\upshape {\color{{Green}} nat-minus}}}(N,N',~N'')} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~A,~P,~ID,~N'',~T)},Q))} ~\otimes \\
& {{\text{\ttfamily\upshape priceQ}}(A',~Y,~({{\text{\ttfamily\upshape consP}}(ID1,~N1,~T1,~L)}))} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \} \\
\\
{\text{\ttfamily\upshape market/5:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape market}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape {\color{{Green}} dual}}}(A,~A')} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} ~\otimes\\
& {{\color{{Green}} {\text{\ttfamily\upshape mktExchange}}}(A,~L',~Y)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~Y,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~L)})} ~\otimes\\
&{{\text{\ttfamily\upshape {\color{{Green}} nat-less}}}(N,~N')} \otimes {{\text{\ttfamily\upshape {\color{{Green}} nat-minus}}}(N',N,~N'')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{ {{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape priceQ}}(A',~Y,~({{\text{\ttfamily\upshape consP}}(ID',~N'',~T',~L)}))} \otimes {{\text{\ttfamily\upshape actPrices}}(A',~L')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \}
\end{array}$$ ]{}
The rules for exchanging market orders are given in Figure \[fig:marketOrdersEx\]. The first rule, , addresses the situation when a market order is in the order queue, but there are no opposite resident orders to be matched against. The order is then removed from the order queue. Rules and address the case when the incoming market order’s quantity is the same as the quantity of the best available opposite resident order. In this case these orders are simply exchanged, and again we distinguish cases whether the resident order was the last in the queue () or not (). Rules and address the case when the incoming market order’s quantity is greater than the quantity of the best available opposite resident order, i.e., when $N>N'$. In this case the resident order is exchanged and the rest of the market order remains in the order queue. The two rules distinguish whether the resident order was the last in the price queue () or not (). Finally, rule describes the situation when an incoming market order is strictly less (quantity-wise) compared to the best opposite resident order. The considered market order is exchanged completely whereas the resident order only partially.
Cancel Orders {#subs:cancel}
-------------
Cancel order is an instruction to remove a particular resident order from the trading system. Cancel orders refer to a resident order by its identifier. If, by chance, the order to be canceled is not in the market, nothing happens and the cancel order is removed from the order queue. If it is there, it is removed from the price queue. Similarly as in exchanging limit orders this results in two sub-cases: the order is the last one in its price queue or not.
The rules for performing order canceling are given in Figure \[fig:cancelOrders\].
[$$\begin{array}{rl}
{\text{\ttfamily\upshape cancel/inListNil:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape cancel}},~A,~P,~ID,~N,~T)},Q))} \otimes {{\text{\ttfamily\upshape actPrices}}(A,~L')} ~\otimes \\
& {{\text{\ttfamily\upshape priceQ}}(A,~P,~{{\text{\ttfamily\upshape consP}}(ID,~N,~T',~{\text{\ttfamily\upshape nilP}})})} \otimes {{\color{{Green}} {\text{\ttfamily\upshape remove}}}(L',~P,~L'')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{{{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}(A,~L'')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\
\\
{\text{\ttfamily\upshape cancel/inListCons:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape cancel}},~A,~P,~ID,~N,~T)},Q1))}~\otimes \\
& {{\text{\ttfamily\upshape priceQ}}(A,~P,~Q)} \otimes {{\text{\ttfamily\upshape {\color{{Green}} inListF}}}(Q,~ID)} \otimes {{\color{{Green}} {\text{\ttfamily\upshape removeF}}}(Q,~ID,~Q')} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{{{\text{\ttfamily\upshape queue}}(Q1)} \otimes {{\text{\ttfamily\upshape priceQ}}(A,~P,~Q')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\
\\
{\text{\ttfamily\upshape cancel/notInListQueue:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape cancel}},~A,~P,~ID,~N,~T)},Q1))}~\otimes \\
& {{\text{\ttfamily\upshape priceQ}}(A,~P,~Q)} \otimes {{\text{\ttfamily\upshape {\color{{Green}} notInListF}}}(Q,~ID)} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{{{\text{\ttfamily\upshape queue}}(Q1)} \otimes {{\text{\ttfamily\upshape priceQ}}(A,~P,~Q)} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \\
\\
{\text{\ttfamily\upshape cancel/notInListActive:}} & {{\text{\ttfamily\upshape queue}}({\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape cancel}},~A,~P,~ID,~N,~T)},Q1))} \otimes {{\text{\ttfamily\upshape actPrices}}(A,~L')} ~\otimes \\
& {{\text{\ttfamily\upshape {\color{{Green}} notInList}}}(L',~P)} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
\multimap & \{{{\text{\ttfamily\upshape queue}}(Q1)} \otimes {{\text{\ttfamily\upshape actPrices}}(A,~L')} \otimes {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \}
\end{array}$$ ]{}
Towards a Mechanized Verification of ATS Properties {#sec:proof}
===================================================
Using our formalization we are able to check that this combination of order-matching rules does not violate some of the expected ATS properties. Although CLF is a powerful logical framework fit for specifying the syntax and semantics of concurrent systems, stating and proving properties about these systems goes beyond its current expressive power. For this task, one needs to consider states of computation, and the execution traces that lead from one state to another. Recent developments show that CLF contexts (the states of computation) can be described in CLF itself through the notion of generative grammars [@Simmons12]. These are grammars whose language consists of all possible CLF contexts which satisfy the property being considered. The general idea is to show that every reachable state consists of a context in this language.
Using such grammars plus reasoning on steps and traces of computation, it is possible to state and prove meta-theorems about CLF specifications. This method is structured enough to become the meta-reasoning engine behind CLF [@CervesatoS13], and therefore it is used for the proofs in this paper.
No Locked or Crossed Market
---------------------------
Here we show that the ${\mathit{bid}}$ price $B$ is always less than the ${\mathit{ask}}$ price $S$ in any reachable state using the rules presented in the previous sections. In other words, we show the following invariant:
\[prop.1\] If ${{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B)}$ and ${{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S)}$ and ${{\text{\ttfamily\upshape {\color{{Green}} maxP}}}(L_B,~B)}$ and ${{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L_S,~S)}$, then $B < S$.
Definition \[grammarNLC\] shows a grammar that generates contexts, or states, satisfying Property \[prop.1\]. This is achieved by the guards ${{\text{\ttfamily\upshape {\color{{Green}} maxP}}}(L_B,~B)}$, ${{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L_S,~S)}$, $B < S$ on the rewriting rule for the start symbol ${{\text{\ttfamily\upshape gen}}(Q,~L_B,~L_S,~T)}$.
\[grammarNLC\] The following generative grammar ${\Sigma_{NLC}}$[^3] produces only contexts where ${\mathit{bid}}< {\mathit{ask}}$.
[$$\begin{array}{lcl}
{\text{\ttfamily\upshape gen/0}} &:& {{\text{\ttfamily\upshape gen}}(Q,~L_B,~L_S,~T)} \otimes {{\text{\ttfamily\upshape {\color{{Green}} maxP}}}(L_B,~B)} \otimes {{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L_S,~S)} \otimes B < S \\
& & \multimap \{{{\text{\ttfamily\upshape queue}}(Q)} \otimes {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B)} \otimes {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S)} \otimes {{\text{\ttfamily\upshape time}}(T)} \\
& & \qquad \otimes~ {\text{\ttfamily\upshape gen-buy}}(L_B) \otimes {\text{\ttfamily\upshape gen-sell}}(L_S)\}.\\[1.5mm]
{\text{\ttfamily\upshape gen/buy1}} &:& {\text{\ttfamily\upshape gen-buy}}({{\text{\ttfamily\upshape nilP}}}) \multimap \{1\}.\\[1.5mm]
{\text{\ttfamily\upshape gen/buy2}} &:& {\text{\ttfamily\upshape gen-buy}}(P::L_B) \multimap \{{{\text{\ttfamily\upshape priceQ}}({{\text{\ttfamily\upshape buy}}},~P,~L)} \otimes {\text{\ttfamily\upshape gen-buy}}(L_B)\}.\\[1.5mm]
{\text{\ttfamily\upshape gen/sell1}} &:& {\text{\ttfamily\upshape gen-sell}}({{\text{\ttfamily\upshape nilP}}}) \multimap \{1\}.\\[1.5mm]
{\text{\ttfamily\upshape gen/sell2}} &:& {\text{\ttfamily\upshape gen-sell}}(P::L_S) \multimap \{{{\text{\ttfamily\upshape priceQ}}({{\text{\ttfamily\upshape sell}}},~P,~L)} \otimes {\text{\ttfamily\upshape gen-sell}}(L_S)\}.\\[1.5mm]
\end{array}$$ ]{}
Intuitively, to show that the market is never in a locked-or-crossed state, we show that, given a context generated by the grammar in Definition \[grammarNLC\], the application of an ATS rule (one step) will result in another context that can also be generated by this grammar. Coupled with the fact that computation starts at a valid context, this shows that the property is always preserved. More formally, we will show the theorem:
\[thm.1\] For every $\Delta \in L({\Sigma_{NLC}})$ and rule $\sigma$, if $\Delta \xrightarrow{\sigma} \Delta'$, then $\Delta' \in L({\Sigma_{NLC}})$.
This theorem can be represented visually as: $$\xymatrix{
{{\text{\ttfamily\upshape gen}}(Q,~L_B,~L_S,~T)} \ar[d]^{\epsilon} & {{\text{\ttfamily\upshape gen}}(Q',~L_B',~L_S',~T')} \ar[d]^{\epsilon'}\\
\Delta \ar[r]^{\sigma} & \Delta'
}$$ The proof consists in showing the existence of $\epsilon'$.
The proof proceeds by case analysis on $\sigma$. We consider only rules that change the linear facts ${{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B)}$ and ${{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S)}$, since otherwise we can simply take $\epsilon' = \epsilon$ (possibly with different instantiations for the variables $L$ in ${\text{\ttfamily\upshape gen/sell2}}$ and ${\text{\ttfamily\upshape gen/buy2}}$). Moreover, we restrict ourselves to the case of incoming ${{\text{\ttfamily\upshape buy}}}$ orders. The case for ${{\text{\ttfamily\upshape sell}}}$ is analogous.
####
This rule rewrites $L_B$, the list of buy prices, to a list $L_B'$ which extends $L_B$ by a new price $P$. Since ${\color{{Green}}{\text{\ttfamily\upshape store}}}$ was provable, we know that $P$ is less than the minimum sell price in the market. For to be applicable, we need:
[$$\Delta = \{ {{\text{\ttfamily\upshape queue}}(Q)}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B)}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S)}, {{\text{\ttfamily\upshape time}}(T)} \} \cup \Delta_1$$ ]{} In which case we conclude that:
[$$\begin{array}{rrl}
Q &= &{\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape {{\text{\ttfamily\upshape buy}}}}},~A,~P,~ID,~N,~T)},Q') \\
\epsilon &= &{\text{\ttfamily\upshape gen/0}}(Q_g, L_B, L_S, T); \epsilon_1; \epsilon_2
\end{array}$$]{}where $\epsilon_1; \epsilon_2$ rewrite ${\text{\ttfamily\upshape gen-buy}}(L_B)$ and ${\text{\ttfamily\upshape gen-sell}}(L_S)$ to the ${{\text{\ttfamily\upshape priceQ}}(\_,~\_,~\_)}$ facts that form $\Delta_1$.
After applying the rule, the context is modified to:
[$$\begin{array}{rcl}
\Delta' &=& \{ {{\text{\ttfamily\upshape queue}}(Q')}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B')}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S)}, {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))}\}\\
&& \cup~ \{{{\text{\ttfamily\upshape priceQ}}({{\text{\ttfamily\upshape buy}}},~P,~{{\text{\ttfamily\upshape consP}}(ID,~N,~T,~{\text{\ttfamily\upshape nilP}})})}, \Delta_1\}
\end{array}$$]{}where $L_B'$ is computed by the ${{\color{{Green}} {\text{\ttfamily\upshape insert}}}(L_B,~P,~L_B')}$ rule and consists of $L_B$ augmented by $P$.
A derivation of $\Delta'$ can be obtained via the following steps:
[$$\epsilon' = {\text{\ttfamily\upshape gen/0}}(Q, L_B', L_S, {\text{\ttfamily\upshape s}}(T)); \epsilon_1; {\text{\ttfamily\upshape gen/sell2}}; \epsilon_2$$]{}where one extra step ${\text{\ttfamily\upshape gen/sell2}}$, with $L = {{\text{\ttfamily\upshape consP}}(ID,~N,~T,~{\text{\ttfamily\upshape nilP}})}$, generates ${{\text{\ttfamily\upshape priceQ}}({{\text{\ttfamily\upshape buy}}},~P,~L)}$. Observe that the guard $B < S$ in ${\text{\ttfamily\upshape gen/0}}$ still holds: if ${{\text{\ttfamily\upshape {\color{{Green}} maxP}}}(L_B',~B)}$ and $B \neq P$, then $B < S$ was part of the assumption. In case ${{\text{\ttfamily\upshape {\color{{Green}} maxP}}}(L_B',~P)}$, observe that ${{\text{\ttfamily\upshape {\color{{Green}} store}}}({{\text{\ttfamily\upshape buy}}},~L_S,~P)}$ only holds if $P < S$, where ${{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L_S,~S)}$. This property is related only to backward chaining predicates and can be proved in the LF framework Twelf using standard techniques.
####
This rule rewrites $L_S$, the list of sell prices, to a list $L_S'$ which consists of $L_S$ without a price $X$. For ${\text{\ttfamily\upshape limit/1}}$ to be applicable, we need:
[$$\begin{array}{rcl}
\Delta &=& \{{{\text{\ttfamily\upshape queue}}(Q)}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B)}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S)}, {{\text{\ttfamily\upshape time}}(T)}\}\\
&& \cup~ \{{{\text{\ttfamily\upshape priceQ}}({{\text{\ttfamily\upshape sell}}},~X,~{{\text{\ttfamily\upshape consP}}(ID',~N,~T',~{\text{\ttfamily\upshape nilP}})})}, ~\Delta_1\}
\end{array}$$]{}where $X \in L_S$ is computed by ${{\color{{Green}} {\text{\ttfamily\upshape exchange}}}({{\text{\ttfamily\upshape buy}}},~L_S,~P,~X)}$.
This can be derived by:
[$$\begin{array}{rrl}
Q &= &{\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~{{\text{\ttfamily\upshape buy}}},~P,~ID,~N,~T)},Q')\\
\epsilon &= &{\text{\ttfamily\upshape gen/0}}(Q_g, L_B, L_S, T); \epsilon_1; {\text{\ttfamily\upshape gen/sell2}}; \epsilon_2
\end{array}$$]{}where $\epsilon_1; {\text{\ttfamily\upshape gen/sell2}}; \epsilon_2$ rewrite ${\text{\ttfamily\upshape gen-buy}}(L_B)$ and ${\text{\ttfamily\upshape gen-sell}}(L_S)$ to the ${{\text{\ttfamily\upshape priceQ}}(\_,~\_,~\_)}$ facts that form $\Delta_1$, with the explicit ${\text{\ttfamily\upshape gen/sell2}}$ generating the fact ${{\text{\ttfamily\upshape priceQ}}({{\text{\ttfamily\upshape sell}}},~X,~{{\text{\ttfamily\upshape consP}}(ID',~N,~T',~{\text{\ttfamily\upshape nilP}})})}$.
After applying the rule, the context is modified to:
[$$\Delta' = \{ {{\text{\ttfamily\upshape queue}}(Q')}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B)}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S')}, {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \cup \Delta_1$$]{}where $L_S'$ is computed by the ${{\color{{Green}} {\text{\ttfamily\upshape remove}}}(L_S,~X,~L_S')}$ rule and consists of $L_S$ without $X$.
A derivation of $\Delta'$ can be obtained via the following steps:
[$$\epsilon' = {\text{\ttfamily\upshape gen/0}}(Q, L_B, L_S', {\text{\ttfamily\upshape s}}(T)); \epsilon_1; \epsilon_2$$]{}Since $L_S' \subset L_S$, then, considering ${{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L_S,~S)}$ and ${{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L_S',~S')}$, it is the case that $S \leq S'$. Thus $B < S$ implies $B < S'$. This can be proved in Twelf given the specification of the appropriate relations (such as $\subset$).
####
This case is analogous to ${\text{\ttfamily\upshape limit/1}}$, except that the incoming order is only partially exchanged because its quantity $N$ is greater than the quantity $N'$ of the matching resident order. The initial context is:
[$$\begin{array}{rcl}
\Delta &=& \{{{\text{\ttfamily\upshape queue}}(Q)}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B)}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S)}, {{\text{\ttfamily\upshape time}}(T)}\}\\
&& \cup~ \{{{\text{\ttfamily\upshape priceQ}}({{\text{\ttfamily\upshape sell}}},~X,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~{\text{\ttfamily\upshape nilP}})})},\Delta_1\}
\end{array}$$]{}where $X \in L_S$ is computed by ${{\color{{Green}} {\text{\ttfamily\upshape exchange}}}({{\text{\ttfamily\upshape buy}}},~L_S,~P,~X)}$.
Which can be derived as before:
[$$\begin{array}{rrl}
Q &= &{\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~{{\text{\ttfamily\upshape buy}}},~P,~ID,~N,~T)},Q''),~N>N' \\
\epsilon &= &{\text{\ttfamily\upshape gen/0}}(Q, L_B, L_S, T); \epsilon_1; {\text{\ttfamily\upshape gen/sell2}}; \epsilon_2
\end{array}$$ ]{}
where $\epsilon_1; {\text{\ttfamily\upshape gen/sell2}}; \epsilon_2$ rewrite ${\text{\ttfamily\upshape gen-buy}}(L_B)$ and ${\text{\ttfamily\upshape gen-sell}}(L_S)$ to the ${{\text{\ttfamily\upshape priceQ}}(\_,~\_,~\_)}$ facts that form $\Delta_1$, with the explicit ${\text{\ttfamily\upshape gen/sell2}}$ generating the fact ${{\text{\ttfamily\upshape priceQ}}({{\text{\ttfamily\upshape sell}}},~X,~{{\text{\ttfamily\upshape consP}}(ID',~N',~T',~{\text{\ttfamily\upshape nilP}})})}$.
After applying the rule, the context is modified to:
[$$\Delta' = \{ {{\text{\ttfamily\upshape queue}}(Q')}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape buy}}},~L_B)}, {{\text{\ttfamily\upshape actPrices}}({{\text{\ttfamily\upshape sell}}},~L_S')}, {{\text{\ttfamily\upshape time}}({\text{\ttfamily\upshape s}}(T))} \} \cup \Delta_1$$ ]{} where $L_S'$ is computed by the ${{\color{{Green}} {\text{\ttfamily\upshape remove}}}(L_S,~X,~L_S')}$ rule and consists of $L_S$ without $X$.
A derivation of $\Delta'$ can be obtained via the following steps (note that time does not change for this rule):
[$$\begin{array}{rcl}
Q' &=& {\text{\ttfamily\upshape front}}({{\text{\ttfamily\upshape ordIn}}({\text{\ttfamily\upshape limit}},~{{\text{\ttfamily\upshape buy}}},~P,~ID,~N-N',~T)},Q'')\\
\epsilon' &=& {\text{\ttfamily\upshape gen/0}}(Q', L_B, L_S', T); \epsilon_1; \epsilon_2
\end{array}$$]{}Since $L_S' \subset L_S$, then, considering ${{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L_S,~S)}$ and ${{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L_S',~S')}$, it is the case that $S \leq S'$. Thus $B < S$ implies $B < S'$. As before, this argument can be developed in Twelf.\
The cases for market orders, $\sigma = {\text{\ttfamily\upshape market/\_}}$, are analogous to the ${\text{\ttfamily\upshape limit/\_}}$ cases above. As for the case $\sigma = {\text{\ttfamily\upshape cancel/inListNil}}$ - this rule rewrites $L_B$, the list of buy prices, to $L_B'$ which is $L_B$ without a price $P$. This case is analogous to ${\text{\ttfamily\upshape limit/1}}$ and ${\text{\ttfamily\upshape limit/3}}$, except that $\epsilon$ contains an application of ${\text{\ttfamily\upshape gen/buy2}}$ which is deleted to obtain $\epsilon'$.
Exchanges happen at ${\mathit{bid}}$ or ${\mathit{ask}}$
--------------------------------------------------------
Every incoming order will either be exchanged or, if this is not possible, stored as a resident order to be exchanged when it is matched (if ever). The predicate will be provable exactly when the incoming order can be exchanged, i.e., there exists an opposite resident order at an “acceptable” price. The acceptable price is: lower than the price of an incoming order, or greater than the price of an incoming order. Since orders have an associated quantity, the exchange may partially or totally consume the orders.
The price at which the exchange takes place is the best possible with respect to the *incoming* order. Consequently, an incoming order is exchanged at the minimal sell price (${\mathit{ask}}$), while an incoming order is exchanged against the maximal available buy price (${\mathit{bid}}$). Therefore whenever an order is exchanged it happens at either ${\mathit{bid}}$ or ${\mathit{ask}}$ price, and *only* at that price. In this section we show that this is indeed the case for our encoding. We will consider only the rules specifying exchange of *limit* orders (Figure \[fig:limitOrdersEx\]). The case for *market* orders follows in a similar fashion, but considering the predicate [mktExchange]{} instead of [exchange]{}.
\[prop.2\] All exchanges happen at and only at the price ${\mathit{bid}}$ or ${\mathit{ask}}$.
Note that Property \[prop.1\] is a property of the reachable *states*, while Property \[prop.2\] is concerns *transitions* between states. We can split it into two parts: (1) in every exchange, only one resident order of price $X$ is consumed; and (2) $X$ is ${\mathit{bid}}$ or ${\mathit{ask}}$.
Part (1) can be more formally stated as:
\[thm:bidask.1\] Let $\Delta, \Delta' \in L({\Sigma_{NLC}})$, and $\sigma$ be one of the exchange rules ${\text{\ttfamily\upshape limit/i}}$ for $1 \leq \mathtt{i} \leq 5$. If $\Delta \xrightarrow{\sigma} \Delta'$, then for all ${{\text{\ttfamily\upshape priceQ}}(A,~Y,~L)} \in \Delta$ if $Y \neq X$ then ${{\text{\ttfamily\upshape priceQ}}(A,~Y,~L)} \in \Delta'$, where $X$ is determined by ${{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A',~L,~P,~X)}$ on the left side of rule $\sigma$.
By inspection of the rules ${\text{\ttfamily\upshape limit/i}}$, we observe that the only facts of the shape ${{\text{\ttfamily\upshape priceQ}}(A,~Z,~L)}$ involved in the rewriting are those where $Z = X$, where $X$ is bound by ${{\color{{Green}} {\text{\ttfamily\upshape exchange}}}(A',~L,~P,~X)}$.
Part (2) of the property can be stated more precisely as:
\[thm:bidask.2\] If ${{\color{{Green}} {\text{\ttfamily\upshape exchange}}}({{\text{\ttfamily\upshape buy}}},~L,~P,~X)}$ then ${{\text{\ttfamily\upshape {\color{{Green}} minP}}}(L,~X)}$.\
If ${{\color{{Green}} {\text{\ttfamily\upshape exchange}}}({{\text{\ttfamily\upshape sell}}},~L,~P,~X)}$ then ${{\text{\ttfamily\upshape {\color{{Green}} maxP}}}(L,~X)}$.
The statement concerns only a backward chaining predicate, so the proof follows standard meta-reasoning techniques from LF, and can be implemented in a few lines in Twelf.
Taken together, Theorems \[thm:bidask.1\] and \[thm:bidask.2\] guarantee Property \[prop.2\].
Conclusion and Future Work {#sec:conclusion}
==========================
We have formalized the core rules guiding the trade on exchanges worldwide. We have done this by formalizing an archetypal automated trading system in the concurrent logical framework CLF, with an implementation in Celf.
Encoding orders in a market as linear resources results in straightforward rules that either consume such orders when they are bought/sold, or store them in the market as resident orders. Moreover the specification is modular and easy to extend with new order types, which is often required in practice. This was our experience when adding market and immediate-or-cancel types of orders to the system. The concurrent aspect of CLF simulates the non-determinism when orders are accumulated in the order queue, but, as explained, orders from the queue are processed sequentially[^4].
Using our formalization we were able to prove two standard properties about a market working under these rules. First we proved that at any given state the ${\mathit{bid}}$ price is smaller than the ${\mathit{ask}}$, i.e., the market is never in a locked-or-crossed state. Secondly we showed that the trade always take place at ${\mathit{bid}}$ or ${\mathit{ask}}$. The first property was proved using generative grammars, an approach motivated by our goal to automate meta reasoning on CLF specifications (not implemented in the current version of Celf). Recent investigations indicate that this approach can handle many meta-theorems [@Simmons12; @CervesatoS13] related to *state* invariants, and ours is yet another example. The second property is a combination of: (1) a property of a backward chaining predicate; and (2) a *transition* invariant. The former can be proved using established methods in LF (in fact, we have proved the desired property in Twelf). The second can be verified by inspection of the rules: the only linear facts that change in the next state, are those rewritten on the right side of $\multimap$.
This specification is an important case study for developing the necessary machinery for automated reasoning in CLF. It is one more evidence of the importance of quantification over steps and traces of a (forward-chaining) computation. It is interesting to note that our example combines forward and backward-chaining predicates, but the generative grammar approach still behaves well. In part because we are only concerned with a linear part of the context. Nevertheless, the proof still relied on some properties of backward chaining predicates. In the second proof, this is even more evident. This indicates, unsurprisingly, that meta-reasoning of CLF specifications must include the already developed meta-reasoning of LF specifications. In the meantime, we are investigating other properties of financial systems that present interesting challenges for meta-reasoning, such as showing that the price/time priority is respected upon exchange.
The difference between the proofs presented might be an indication that we need to follow a more general approach than the one used in LF. In that framework, most theorems that motivated the work have the same shape and their proofs follow the same strategy. Therefore, it is possible to save the user a lot of work by asking them to specify only the necessary parts that fills in a “proof template”, which is then checked mechanically. When working with more ad-hoc systems, such as the case of financial exchanges, the properties and proofs are less regular, making it harder to figure out a good “template” that fits all properties of interest. In this case, it may be beneficial to leave more freedom (and consequently more work) to the user, in order to allow more flexible meta-reasoning.
Concurrently, we plan to formalize other models of financial trading systems, as this is a relevant application addressing some critical challenges.
[^1]: This paper was made possible by grant NPRP 7-988-1-178 from the Qatar National Research Fund (a member of the Qatar Foundation). The statements made herein are solely the responsibility of the authors.
[^2]: Sometimes an incoming limit order will be partially filled, with the remainder (once resident orders that match this limit price are filled) becoming a new resident order.
[^3]: $NLC$ stands for no locked-or-crossed.
[^4]: As far as we know, no real life trading system performs parallel order matching and execution.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents a fully automated procedure for controller synthesis for multi-agent systems under coupling constraints. Each agent is modeled with dynamics consisting of two terms: the first one models the coupling constraints and the other one is an additional bounded control input. We aim to design these inputs so that each agent meets an individual high-level specification given as a Metric Interval Temporal Logic (MITL). First, a decentralized abstraction that provides a space and time discretization of the multi-agent system is designed. Second, by utilizing this abstraction and techniques from formal verification, we propose an algorithm that computes the individual runs which provably satisfy the high-level tasks. The overall approach is demonstrated in a simulation example conducted in MATLAB environment.'
author:
-
bibliography:
- 'references.bib'
title: 'On the Timed Temporal Logic Planning of Coupled Multi-Agent Systems'
---
[^1]
Multi-Agent Systems, Cooperative Control, Hybrid Systems, Formal Verification, Timed Logics, Abstractions, Discrete Event Systems.
Introduction
============
Cooperative control of multi-agent systems has traditionally focused on designing distributed control laws in order to achieve global tasks such as consensus and formation control, and at the same time fulfill properties such as network connectivity and collision avoidance. Over the last few years, the field of control of multi-agent systems under high-level specifications has been gaining attention. In this work, we aim to additionally introduce specific time bounds into these tasks, in order to include specifications such as: “Robot 1 and robot 2 should visit region $A$ and $B$ within 4 time units respectively or “Both robots 1 and 2 should periodically survey regions $A_1$, $A_2$, $A_3$, avoid region $X$ and always keep the longest time between two consecutive visits to $A_1$ below 8 time units".
The qualitative specification language that has primarily been used to express the high-level tasks is Linear Temporal Logic (LTL) (see, e.g., [@muray_2010_receding]). There is a rich body of literature containing algorithms for verification and synthesis of multi-agent systems under high level specifications ([@guo_2015_reconfiguration; @zavlanos_2016_multi-agent_LTL; @pappas_2016_implan]). Controller synthesis under timed specifications has been considered in [@liu_MTL; @murray_2015_stl; @topcu_2015; @baras_MTL_2016_new]. In [@liu_MTL], the authors addressed the problem of designing high-level planners to achieve tasks for switching dynamical systems under Metric Temporal Logic (MTL) specification and in [@murray_2015_stl], the authors utilized a counterexample-guided synthesis for cyber-physical systems subject to Signal Temporal Logic (STL) specifications. In [@topcu_2015], an optimal control problem for continuous-time stochastic systems subject to objectives specified in MITL was studied. In [@baras_MTL_2016_new], the authors focused on motion planning based on the construction of an efficient timed automaton from a given MITL specification. However, all these works are restricted to single agent planning and are not extendable to multi-agent systems in a straightforward way. The high-level coordination of multiple vehicles under timed specifications has been considered in [@frazzoli_MTL], by solving an optimization problem over the tasks’ execution time instances.
An automata-based solution for multi-agent systems was proposed in our previous work [@alex_2016_acc], where Metric Interval Temporal Logic (MITL) formulas were introduced in order to synthesize controllers such that every agent fulfills an individual specification and the team of agents fulfills a global specification. Specifically, the abstraction of each agent’s dynamics was considered to be given and an upper bound of the time that each agent needs to perform a transition from one region to another was assumed. Furthermore, potential coupled constraints between the agents were not taken into consideration. Motivated by this, in this work, we aim to address the aforementioned issues. We assume that the dynamics of each agent consists of two parts: the first part is a consensus type term representing the coupling between the agent and its neighbors, and the second one is an additional control input which will be exploited for high-level planning. Hereafter, we call it a free input. A decentralized abstraction procedure is provided, which leads to an individual Transition System (TS) for each agent and provides a basis for high-level planning. Additionally, this abstraction is associated to a time quantization which allows us to assign precise time durations to the transitions of each agent. Abstractions for both single and multi-agent systems can be found in [@alur_2000_discrete_abstractions; @zamani_2012_symbolic; @abate_2014_finite_abstractions; @PJ_tac; @tabuada_compositional_abstractions; @boskos_cdc_2015]. Compositional frameworks are provided in [@PJ_tac] for safety specifications of discrete time systems, and [@tabuada_compositional_abstractions], which is focused on feedback linearizable systems with a cascade interconnection. Therefore, these results are not applicable to the systems we consider, which evolve in continuous time and do not require a specific network interconnection.
Motivated by our previous work [@boskos_cdc_2015], we start from the consensus dynamics of each agent and we construct a Weighted Transition System (WTS) for each agent in a decentralized manner. Each agent is assigned an individual task given in MITL formulas. We aim to design the free inputs so that each agent performs the desired individual task within specific time bounds. In particular, we provide an automatic controller synthesis method for coupled multi-agent systems under high-level tasks with timed constraints. A motivation for this framework comes from applications such as the deployment of aerial robotic teams. In particular, the consensus coupling allows the robots to stay sufficiently close to each other and maintain a connected network during the evolution of the system. Additionally, individual MITL formulas are leveraged to assign area monitoring tasks to each robot individually. The MITL formalism enables us to impose time constraints on the monitoring process. Compared to existing works on multi-agent planning under temporal logic specifications, the proposed approach considers dynamically coupled multi-agent systems under timed temporal specifications in a distributed way. To the best of the authors’ knowledge, this is the first time that a fully automated framework for multi-agent systems consisting of both constructing an abstraction and conducting high-level timed temporal logic planning is considered.
This paper is organized as follows. In Section \[sec: preliminaries\] a description of the necessary mathematical tools, the notations and the definitions are given. Section \[sec: prob\_formulation\] provides the dynamics of the system and the formal problem statement. Section \[sec: solution\] discusses the technical details of the solution. Section \[sec: simulation\_results\] is devoted to a simulation example. Finally, conclusions and future work are discussed in Section \[sec: conclusions\].
Notation and Preliminaries {#sec: preliminaries}
==========================
Notation
--------
Denote by $\mathbb{R}, \mathbb{Q}_+, \mathbb{N}$ the set of real, nonnegative rational and natural numbers including 0, respectively. Define also $\mathbb{T}_{\infty} = \mathbb{T} \cup \{\infty\}$ for a set $\mathbb{T} \subseteq \mathbb{R}$; $\mathbb{R}_{\ge 0}$ is the set of real numbers with all elements nonnegative. Given a set $S$, we denote by $|S|$ its cardinality, by $S^N = S \times \dots \times S$, its $N$-fold Cartesian product and by $2^S$ the set of all its subsets. For a subset $S$ of $\mathbb{R}^n$, denote by $\text{cl}(S), \text{int}(S)$ and $\partial S = \text{cl}(S) \backslash \text{int}(S)$ its closure, interior and boundary, respectively, where $\backslash$ is used for set subtraction. The notation $\|x\|$ is used for the Euclidean norm of a vector $x \in \mathbb{R}^n$ and $\|A\| = \text{max} \{\|Ax\| : \|x\| = 1\}$ for the induced norm of a matrix $A \in \mathbb{R}^{m \times n}$. Given a matrix $A$ denote by $\lambda_{\text{max}}(A) = \text{max} \{|\lambda| : \lambda \in \sigma(A) \}$ the spectral radius of $A$, where $\sigma(A)$ is the set of all the eigenvalues of $A$; $A \otimes B$ denotes the Kronecker product of the matrices $A, B \in \mathbb{R}^{m \times n}$ (see [@horn1994topics]). Define also by $I_n \in \mathbb{R}^{n \times n}$ the identity matrix.
Multi-Agent Systems {#sec:prelims:system}
-------------------
An *undirected graph* $\mathcal{G}$ is a pair $(\mathcal{I}, \mathcal{E})$, where $\mathcal{I} = \{1,\dots,N\}$ is a finite set of nodes, representing a team of agents, and $\mathcal{E} \subseteq \{\{i,j\}: i,j \in \mathcal{I}, i \neq j\}$, is the set of edges that model the communication capability between the neighboring agents. For each agent, its neighbors’ set $\mathcal{N}(i)$ is defined as $\mathcal{N}(i) = \{j_1, \ldots, j_{N_i} \} = \{ j \in \mathcal{I} : \{i,j\} \in \mathcal{E}\}$ where $N_i = |\mathcal N(i)|$. The Laplacian matrix $L(\mathcal{G}) \in \mathbb{R}^{N \times N}$ of the graph $\mathcal{G}$ is defined as $L(\mathcal{G}) = D(\mathcal{G}) D(\mathcal{G})^{\top}$ where $D(\mathcal{G})$ is the $N \times |\mathcal{E}|$ incidence matrix, as it is defined in [@mesbahi2010graph Chapter 2]. The graph Laplacian $L(\mathcal{G})$ is positive semidefinite and symmetric. If we consider an ordering $0 = \lambda_1(\mathcal{G}) \leq \lambda_2(\mathcal{G}) \leq \ldots \leq \lambda_N(\mathcal{G}) = \lambda_{\text{max}}(\mathcal{G})$ of the eigenvalues of $L(\mathcal{G})$ then we have that $\lambda_2(\mathcal{G}) > 0$ iff $\mathcal{G}$ is connected ([@mesbahi2010graph Chapter 2]).
Given a vector $x_i = (x^1_i, \ldots, x^n_i) \in \mathbb{R}^n$, the component operator $c(x_i, \ell) = x_i^\ell \in \mathbb{R}, \ell = 1, \ldots, n$ gives the projection of $x_i$ onto its $\ell$-th component (see [@mesbahi2010graph Chapter 7]). Similarly, for the stack vector $x = (x_1, \ldots, x_N) \in \mathbb{R}^{Nn}$ the component operator is defined as $c(x, \ell) = (c(x_1, \ell), \ldots, c(x_N, \ell)) \in \mathbb{R}^{N}, \ell = 1, \ldots, n$. By using the component operator, the norm of a vector $x \in \mathbb{R}^{Nn}$ can be evaluated as $\|x\| = \left\{ \displaystyle \sum_{\ell = 1}^{n} \| c(x, \ell) \|^2 \right\}^{\frac{1}{2}}$.
Denote by $\widetilde{x} \in \mathbb{R}^{|\mathcal{E}|n}$ the stack column vector of the vectors $x_i-x_j, \{i, j\} \in \mathcal{E}$ with the edges ordered as in the case of the incidence matrix $D(\mathcal{G})$. Then, the following holds: $$\label{eq:xtilde}
\widetilde{x} = \left(D(\mathcal{G})^{\top} \otimes I_n \right) x.$$
Cell Decompositions
-------------------
In the subsequent analysis a discrete partition of the workspace into cells will be considered which is formalized through the following definition.
\[def: cell\_decomposition\] A *cell decomposition* $S = \{S_\ell\}_{\ell \in \mathbb{I}}$ of a set $\mathcal{D} \subseteq \mathbb{R}^n$, where $\mathbb{I} \subseteq \mathbb{N}$ is a finite or countable index set, is a family of uniformly bounded convex sets $S_\ell, \ell \in \mathbb{I}$ such that $\text{int}(S_\ell) \cap \text{int}(S_{\hat{\ell}}) = \emptyset$ for all $\ell, \hat{\ell} \in \mathbb{I}$ with $\ell \neq \hat{\ell}$ and $\cup_{\ell \in \mathbb{I}} S_\ell = \mathcal D$ We assume that the interiors of the cells are non-empty.
Time Sequence, Timed Run and Weighted Transition System
-------------------------------------------------------
In this section we include some definitions from computer science that are required to analyze our framework.
An infinite sequence of elements of a set $X$ is called an *infinite word* over this set and it is denoted by $\chi = \chi(0)\chi(1) \ldots$. The $i$-th element of a sequence is denoted by $\chi(i)$. For certain technical reasons that will be clarified in the sequel, we will assume hereafter that $\mathbb{T} = \mathbb{Q}_+$.
([@alur1994]) A *time sequence* $\tau = \tau(0) \tau(1) \ldots$ is an infinite sequence of time values $\tau(j) \in \mathbb{T}$, satisfying the following properties:
- Monotonicity: $\tau(j) < \tau(j+1)$ for all $j \geq 0$.
- Progress: For every $t \in \mathbb{T}$, there exists $\ j \ge 1$, such that $\tau(j) > t$.
An *atomic proposition* $p$ is a statement that is either True $(\top)$ or False $(\bot)$.
([@alur1994]) Let ${\mathit{AP}}$ be a finite set of atomic propositions. A *timed word* $w$ over the set ${\mathit{AP}}$ is an infinite sequence $w^t = (w(0), \tau(0)) (w(1), \tau(1)) \ldots$ where $w(0) w(1) \ldots$ is an infinite word over the set $2^{{\mathit{AP}}}$ and $\tau(0) \tau(1) \ldots$ is a time sequence with $\tau(j) \in \mathbb{T}, \ j \geq 0$.
\[def: WTS\] A Weighted Transition System (*WTS*) is a tuple $(S, S_0, Act, \longrightarrow, d, AP, L)$ where $S$ is a finite set of states; $S_0 \subseteq S$ is a set of initial states; $Act$ is a set of actions; $\longrightarrow \subseteq S \times Act \times S$ is a transition relation; $d: \longrightarrow \rightarrow \mathbb{T}$ is a map that assigns a positive weight to each transition; ${\mathit{AP}}$ is a finite set of atomic propositions; and $L: S \rightarrow 2^{AP}$ is a labeling function. The notation $s \overset{\alpha}{\longrightarrow} s'$ is used to denote that $(s, \alpha, s') \in \longrightarrow$ for $s, s' \in S$ and $\alpha \in Act$. For every $s \in S$ and $\alpha \in Act$ define $\text{Post}(s, \alpha) = \{s' \in S : (s, \alpha, s') \in \longrightarrow\}$.
\[run\_of\_WTS\] A *timed run* of a WTS is an infinite sequence $r^t = (r(0), \tau(0))(r(1), \tau(1)) \ldots$, such that $r(0) \in S_0$, and for all $j \geq 1$, it holds that $r(j) \in S$ and $(r(j), \alpha(j), r(j+1)) \in \longrightarrow$ for a sequence of actions $\alpha(1) \alpha(2) \ldots$ with $\alpha(j) \in Act, \forall \ j \geq 1$. The *time stamps* $\tau(j), j \geq 0$ are inductively defined as:
1. $\tau(0) = 0$.
2. $\displaystyle \tau(j+1) = \tau(j) + d(r(j), \alpha(j), r(j+1)), \ \forall \ j \geq 1.$
Every timed run $r^t$ generates a *timed word* $w(r^t) = (w(0), \tau(0)) \ (w(1), \tau(1)) \ldots$ over the set $2^{{\mathit{AP}}}$ where $w(j) = L(r(j))$, $\forall \ j \geq 0$ is the subset of atomic propositions that are true at state $r(j)$.
Metric Interval Temporal Logic
------------------------------
The syntax of *Metric Interval Temporal Logic (MITL)* over a set of atomic propositions $AP$ is defined by the grammar: $$\varphi := p \ | \ \neg \varphi \ | \ \varphi_1 \wedge \varphi_2 \ | \ \bigcirc_I \varphi \ | \ \Diamond_I \varphi \mid \square_I \varphi \mid \varphi_1 \ \mathcal{U}_I \ \varphi_2,$$ where $p \in {\mathit{AP}}$, and $\bigcirc$, $\Diamond$, $\square$ and $\mathcal U$ are the next, eventually, always and until temporal operator, respectively; $I = [a, b] \subseteq \mathbb{T}$ where $a, b \in [0, \infty]$ with $a < b$ is a non-empty timed interval. MITL can be interpreted either in continuous or point-wise semantics [@pavithra_expressiveness]. In this paper, the latter approach is utilized, since the consideration of point-wise (event-based) semantics is more suitable for the automata-based specifications considered in a discretized state-space. The MITL formulas are interpreted over timed words like the ones produced by a WTS which is given in Def. \[run\_of\_WTS\].
\[def:mitl\_semantics\] ([@pavithra_expressiveness], [@quaknine_decidability]) Given a timed word $w^t = (w(0),\tau(0))(w(1),\tau(1)) \dots$, an MITL formula $\varphi$ and a position $i$ in the timed word, the satisfaction relation $(w^t, i) \models \varphi$, for $\ i \geq 0$ (read $w^t$ satisfies $\varphi$ at position $i$) is inductively defined as follows: $$\begin{aligned}
\label{eq: for1}
&(w^t, i) \models p \Leftrightarrow p \in w(i), \\
&(w^t, i) \models \neg \varphi \Leftrightarrow (w^t, i) \not \models \varphi, \\
&(w^t, i) \models \varphi_1 \wedge \varphi_2 \Leftrightarrow (w^t, i) \models \varphi_1 \ \text{and} \ (w^t, i) \models \varphi_2, \\
&(w^t, i) \models \bigcirc_I \ \varphi \Leftrightarrow (w^t, i+1) \models \varphi \ \text{and} \ \tau(i+1) - \tau(i) \in I, \\
&(w^t, i) \models \Diamond_I \varphi \Leftrightarrow \exists j \ge i, \ \text{such that} \ (w^t, j) \models \varphi, \tau(j)-\tau(i) \in {I}, \\
&(w^t, i) \models \square_I \varphi \Leftrightarrow \forall j \ge i, \ \tau(j)-\tau(i) \in {I} \Rightarrow (w^t, j) \models \varphi, \\
&(w^t, i) \models \varphi_1 \ \mathcal{U}_I \ \varphi_2 \Leftrightarrow \exists j\ge i, \ \text{s.t. } (w^t, j) \models \varphi_2, \tau(j)-\tau(i) \in I \notag \\
&\hspace{65mm} \text{and} \ (w^t, k) \models \varphi_1, \forall \ i \leq k < j.
\end{aligned}$$ We say that a timed run $r^t = (r(0),\tau(0))(r(1),\tau(1)) \dots$ satisfies the MITL formula $\varphi$ (we write $r^t \models \varphi$) if and only if the corresponding timed word $w(r^t) = (w(0),\tau(0))(w(1),\tau(1)) \dots$ with $w(j) = L(r(j)), \forall j \ge 0$, satisfies the MITL formula ($w(r^t) \models \varphi$).
It has been proved that MITL is decidable in infinite words and point-wise semantics, which is the case considered here (see [@alur_mitl; @reynold] for details). The model checking and satisfiability problems are *EXPSPACE*-complete. It should be noted that in the context of timed systems, EXSPACE complexity is fairly low [@bouyer_phd].
Consider the WTS with $S = \{s_0, s_1, s_2\}$, $S_0 = \{s_0\}$, $Act = \emptyset$, $\longrightarrow = \{(s_0, \emptyset, s_1)$, $(s_1, \emptyset, s_2)$, $(s_1, \emptyset, s_0)$, $(s_2, \emptyset, s_1)\}$, $d((s_0, \emptyset, s_1)) = 1.0$, $d((s_1, \emptyset, s_2)) = 1.5$, $d((s_1, \emptyset, s_0)) = 2.0$, $d((s_2, \emptyset, s_1)) = 0.5$, $AP = \{green\}$, $L(s_0) = \{green\}$, $L(s_1) = L(s_2) = \emptyset$ depicted in Figure 1.
\[fig:wts\_example\]
(pseudo1) at (-1.2,0); (0) \[line width = 1.0\] at (0,0)\[shape=circle,draw\]\[fill=green!20\] [$s_0$]{}; (1) \[line width = 1.0\] at (2.5,0)\[shape=circle,draw\] [$s_1$]{}; (5) \[line width = 1.0\] at (5.0,0)\[shape=circle,draw\] [$s_2$]{};
\(0) edge \[bend left = 15\] node \[above\] [$1.0$]{} (1) (1) edge \[bend right = -15\] node \[below\] [$2.0$]{} (0) (1) edge \[bend right = -15\] node \[above\] [$1.5$]{} (5) (5) edge \[bend right = -15\] node \[below\] [$0.5$]{} (1)
(pseudo1) edge (0);
Let two timed runs of the system be: $$\begin{aligned}
r_1^t &= (s_0, 0.0)(s_1, 1.0)(s_0, 3.0)(s_1, 4.0)(s_0, 6.0) \ldots, \\
r_2^t &= (s_0, 0.0)(s_1, 1.0)(s_2, 2.5)(s_1, 3.0)(s_0, 5.0) \ldots,
\end{aligned}$$ and two MITL formulas $\varphi_1 = \Diamond_{[2, 5]} \{green\}, \varphi_2 = \square_{[0, 5]} \{green\}$. According to the MITL semantics, it follows that the timed run $r_1^t$ satisfies $\varphi_1$ ($r_1^t \models \varphi_1$), since at the time stamp $3.0 \in [2, 5]$ we have that $L(s_0) = \{green\}$ so the atomic proposition $green$ occurs at least once in the given interval. On the other hand, the timed run $r_2^t$ does not satisfy $\varphi_2$ ($r_2^t \not \models \varphi_2$) since the atomic proposition $green$ does not hold at every time stamp of the run $r_2^t$ (it holds only at the time stamp $0.0$).
Timed Büchi Automata {#sec: timed_automata}
--------------------
*Timed Büchi Automata (TBA)* were originally introduced in [@alur1994]. In this work, we partially adopt the notation from [@bouyer_phd; @tripakis_tba]. Let $C = \{c_1, \ldots, c_{|C|}\}$ be a finite set of *clocks*. The set of *clock constraints* $\Phi(C)$ is defined by the grammar: $$\phi := \top \mid \ \neg \phi \ | \ \phi_1 \wedge \phi_2 \ | \ c \bowtie \psi,$$ where $c \in C$ is a clock, $\psi \in \mathbb{T}$ is a clock constant and $\bowtie \ \in \{ <, >, \geq, \leq, = \}$. A clock *valuation* is a function $\nu: C \rightarrow\mathbb{T}$ that assigns a value to each clock. A clock $c_i$ has valuation $\nu_i$ for $i \in \{1, \ldots, |C|\}$, and $\nu = (\nu_1, \ldots, \nu_{|C|})$. We denote by $\nu \models \phi$ the fact that the valuation $\nu$ satisfies the clock constraint $\phi$.
A *Timed Büchi Automaton* is a tuple $\mathcal{A} = (Q, Q^{\text{init}}, C, Inv,
E, F, AP, \mathcal{L})$ where $Q$ is a finite set of locations; $Q^{\text{init}} \subseteq Q$ is the set of initial locations; $C$ is a finite set of clocks; $Inv: Q \rightarrow \Phi(C)$ is the invariant; $E \subseteq Q \times \Phi(C) \times 2^C \times Q$ gives the set of edges of the form $e = (q, \gamma, R, q')$, where $q$, $q'$ are the source and target states, $\gamma$ is the guard of edge $e$ and $R$ is a set of clocks to be reset upon executing the edge; $F \subseteq Q$ is a set of accepting locations; ${\mathit{AP}}$ is a finite set of atomic propositions; and $\mathcal{L}: Q \rightarrow 2^{AP}$ labels every state with a subset of atomic propositions.
A state of $\mathcal{A}$ is a pair $(q, \nu)$ where $q \in Q$ and $\nu$ satisfies the *invariant* $Inv(q)$, i.e., $\nu \models Inv(q)$. The initial state of $\mathcal{A}$ is $(q(0), (0,\ldots,0))$, where $q(0) \in Q^{\text{init}}$. Given two states $(q, \nu)$ and $(q', \nu')$ and an edge $e = (q, \gamma, R, q')$, there exists a *discrete transition* $(q, \nu) \overset{e}{\longrightarrow} (q', \nu')$ iff $\nu \models \gamma$, $\nu' \models Inv(q')$, and $R$ is the *reset set*, i.e., $\nu'_i = 0$ for $c_i \in R$ and $\nu'_i = \nu_i$ for $c_i \notin R$. Given a $\delta \in \mathbb{T}$, there exists a *time transition* $(q, \nu) \overset{\delta}{\longrightarrow} (q', \nu')$ iff $q = q', \nu' = \nu+\delta$ ($\delta$ is summed component-wise) and $\nu' \models Inv(q)$. We write $(q, \nu) \overset{\delta}{\longrightarrow} \overset{e}{\longrightarrow} (q', \nu')$ if there exists $q'', \nu''$ such that $(q, \nu) \overset{\delta}{\longrightarrow} (q'', \nu'')$ and $(q'', \nu'') \overset{e}{\longrightarrow} (q', \nu')$ with $q'' = q$.
An infinite run of $\mathcal{A}$ starting at state $(q(0), \nu)$ is an infinite sequence of time and discrete transitions $(q(0), \nu(0))\overset{\delta_0}{\longrightarrow} (q(0)', \nu(0)') \overset{e_0}{\longrightarrow} (q(1), \nu(1)) \overset{\delta_1}{\longrightarrow} (q(1)', \nu(1)') \ldots$, where $(q(0),\nu(0))$ is an initial state. This run produces the timed word $w = (\mathcal{L}(q(0)), \tau(0))(\mathcal{L}(q(1)), \tau(1)) \ldots$ with $\tau(0) = 0$ and $\tau(i+1) = \tau(i) +\delta_i$, $\forall \ i \geq 1$. The run is called *accepting* if $q(i) \in F$ for infinitely many times. A timed word is *accepted* if there exists an accepting run that produces it. The problem of deciding the language emptiness of a given TBA is PSPACE-complete [@alur1994]. In other words, an accepting run of a given TBA can be synthesized, if one exists. In other words, an accepting run of a given TBA can be synthesized, if one exists. Any MITL formula $\varphi$ over $AP$ can be algorithmically translated into a TBA with the alphabet $2^{\mathit{AP}}$, such that the language of timed words that satisfy $\varphi$ is the language of timed words produced by the TBA ([@alur_mitl; @maler_MITL_TA; @nickovic_timed_aut; @MITL_2_TA_tool]).
\[ex: example\_1\] A TBA with $Q = \{q_0, q_1, q_2\}, Q^{\text{init}} = \{q_0\}, C = \{c\}, Inv(q_0) = Inv(q_1) = Inv(q_2) = \emptyset, E = \{(q_0, \{c \leq c_2\}, \emptyset, q_0), (q_0, \{c \leq c_1 \vee c > c_2\}, c, q_2), (q_0, \{c \geq c_1 \wedge c \leq c_2\}, c, q_1), (q_1, \top, c, q_1), (q_2, \top,c, q_2)\}, F = \{q_1\}, AP = \{green\}, \mathcal{L}(q_0) = \mathcal{L}(q_2) = \emptyset, \mathcal{L}(q_1) = \{green\} $ that accepts all the timed words that satisfy the formula $\varphi_3 = \Diamond_{[c_1, c_2]} \{green\}$ is depicted in Figure \[fig:TBA\_example\]. This formula will be used as reference for the following examples and simulations.
(pseudo) at (-1.5,0); (0) \[line width = 1.0\] at (0,0)\[shape=circle,draw\] [$\ q_0 \ $]{}; (1) \[line width = 1.0\] at (4.5,0)\[shape=circle,draw, double\]\[fill=green!20\] [$\ q_1 \ $]{}; (2) \[line width = 1.0\] at (0,-2)\[shape=circle,draw\] [$\ q_2 \ $]{};
\(0) edge \[bend left = 0\] node \[above\] (1) (0) edge \[loop above\] node \[above\] () (1) edge \[loop above\] node \[above\] [$\top, c := 0$]{} () (2) edge \[loop below\] node \[below\] [$\top, c := 0$]{} () (0) edge \[bend left = 0\] node \[above\] (2)
(pseudo) edge (0);
at (-0.2, 1.6) [$c \leq c_2, \emptyset$]{}; at (2.2, 0.3) [$c \geq c_1 \wedge c \leq c_2$]{}; at (2.3, -0.3) [$\tiny c := 0$]{}; at (-2.0, -0.7) [$\tiny c < c_1 \vee c > c_2$]{}; at (-1.9, -1.1) [$\tiny c := 0$]{};
at (4.5,-0.8) [$\{green\}$]{};
An example of a timed run of this TBA is $(q_0, 0) \overset{\delta = \alpha_1}{\longrightarrow} (q_0, \alpha_1) \overset{e = (q_0, \{c \geq c_1 \wedge c \leq c_2\}, c, q_1)}{\longrightarrow} (q_1, 0) \ldots$ with $c_1 \leq \alpha_1 \leq c_2 $, which generates the timed word $w^t = (\mathcal{L}(q_0), 0)(\mathcal{L}(q_0), \alpha_1)(\mathcal{L}(q_1), \alpha_1) \ldots = (\emptyset, 0)(\emptyset, \alpha_1) (\{green\}, \alpha_1) \ldots$ that satisfies the formula $\varphi_3$. The timed run $(q_0, 0) \overset{\delta = \alpha_2}{\longrightarrow} (q_0, \alpha_2) \overset{e = (q_0, \{c \leq c_1 \vee c > c_2\}, c, q_2)}{\longrightarrow} (q_2, 0) \ldots$ with $\alpha_2 < c_1$, generates the timed word $w^t = (\mathcal{L}(q_0), 0)(\mathcal{L}(q_0), \alpha_2) (\mathcal{L}(q_2), \alpha_2) \ldots = (\emptyset, 0)(\emptyset, \alpha_2)(\emptyset, \alpha_2) \ldots$ that does not satisfy the formula $\varphi_3$.
Traditionally, the clock constraints and the TBAs are defined with $\mathbb{T} = \mathbb{N}$. However, they can be extended to accommodate $\mathbb{T} = \mathbb{Q}_+$, by multiplying all the rational numbers that are appearing in the state invariants and the edge constraints with their least common multiple.
Problem Formulation {#sec: prob_formulation}
===================
System Model
------------
We focus on multi-agent systems with coupled dynamics of the form: $$\label{eq: system}
\dot{x}_i = -\sum_{j \in \mathcal{N}(i)}^{} (x_i - x_j)+v_{i}, x_i \in \mathbb{R}^n, i \in \mathcal{I}.$$ The dynamics consists of two parts; the first part is a consensus protocol representing the coupling between the agent and its neighbors, and the second one is a control input which will be exploited for high-level planning and is called free input. In this work, it is assumed that the free inputs are bounded by a positive constant $v_{\text{max}}$, i.e., $\| v_i(t) \| \leq v_{\text{max}}, \ \forall \ i \in \mathcal{I}, t \geq 0$.
The topology of the multi-agent network is modeled through an undirected graph $\mathcal{G} = (\mathcal{I},\mathcal{E})$, where $\mathcal{I} = \{1, \ldots, N\}$ and the following assumption is made.
The communication graph $\mathcal{G} = (\mathcal{I},\mathcal{E})$ of the system is undirected, connected and static i.e., every agent preserves the same neighbors for all times.
Specification
-------------
Our goal is to control the multi-agent system so that each agent’s behavior obeys a desired individual specification $\varphi_i$ given in MITL. In particular, it is required to drive each agent to a desired subset of the *workspace* $\mathbb R^n$ within certain time limits and provide certain atomic tasks there. Atomic tasks are captured through a finite set of *services* $\Sigma_i, i \in \mathcal{I}$. The position $x_i$ of each agent $i \in \mathcal{I}$ is labeled with services that are offered there. Thus, a *service labeling function*: $$\label{eq:label_lambda}
\Lambda_i:\mathbb{R}^n\to 2^{\Sigma_i},$$ is introduced for each agent $i \in \mathcal{I}$ which maps each state $x_i \in \mathbb{R}^n$ to the subset of services $\Lambda_i(x_i)$ which hold true at $x_i$ i.e., the subset of services that the agent $i$ *can* provide in position $x_i$. It is noted that although the term service labeling function it is used, these functions are not necessarily related to the labeling functions of a WTS as in Definition \[def: WTS\]. Define also by $\Lambda(x) = \bigcup_{i \in \mathcal I} \Lambda_i(x)$ the union of all the service labeling functions. We also assume that $\Sigma_i \cap \Sigma_j = \emptyset$, for all $i,j \in \mathcal{I}, i \neq j$ which means that the agents do not share any services. Let us now introduce the following assumption which is necessary for formally defining the problem.
\[assumption: AP\_cell\_decomposition\] There exists a decomposition $S = \{S_\ell\}_{\ell \in \mathbb I}$ of the workspace $\mathbb{R}^n$ which forms a cell decomposition according to Def. \[def: cell\_decomposition\] and respects the labeling function $\Lambda$ i.e., for all $S_\ell \in S$ it holds that $\Lambda(x) = \Lambda(x'), \forall \ x, x' \in S_\ell$. This assumption implies that the same services hold at all the points that belong to the same cell of the decomposition.
Define for each agent $i \in \mathcal{I}$ the labeling function: $$\label{eq:label_mathcal_lambda}
\mathcal{L}_i: S \to 2^{\Sigma_i},$$ which denotes the fact that when agent $i$ visits a region $S_\ell \in S$ and it can choose to *provide* a subset of the services that are being offered there i.e. it chooses to satisfy a subset of $\mathcal{L}_i(S_\ell)$.
The trajectory of each agent $i$ is denoted by $x_i(t), t \geq 0, i \in \mathcal{I}$. The trajectory $x_i(t)$ is associated with a unique sequence $r_{x_i}^t = (r_i(0), \tau_i(0))(r_i(1), \tau_i(1))(r_i(2), \tau_i(2))\ldots$, of regions that the agent $i$ crosses, where for all $j \ge 0$ it holds that: $x_i(\tau_i(j)) \in r_i(j)$ and $\Lambda_i(x_i(t)) = \mathcal{L}_i(r_i(j)), \forall \ t \in [\tau_i(j), \tau_i(j+1))$ for some $r_i(j) \in S$ and $r_i(j) \ne r_i(j+1)$. The timed word $w_{x_i}^t = (\mathcal{L}_i(r_i(0)), \tau_i(0))(\mathcal{L}_i(r_i(1)), \tau_i(1))(\mathcal{L}_i(r_i(2)), \tau_i(2))\ldots$, where $w_i(j) = \mathcal{L}_i(r_i(j)), j \ge 0, i \in \mathcal{I}$, is associated uniquely with the trajectory $x_i(t)$, and represents the sequence of services that *can be provided* by the agent $i$ following the trajectory $x_i(t), t \ge 0$.
Define a *timed service word* as: $$\begin{aligned}
\label{eq:time_serviced_word}
\widetilde{w}_{x_i}^t &= (\beta_i(z_0), \widetilde{\tau}_i(z_0))(\beta_i(z_1), \widetilde{\tau}_i(z_1))(\beta_i(z_2), \widetilde{\tau}_i(z_2)) \ldots,\end{aligned}$$ where $z_0 =0 < z_1 < z_2 < \ldots$ is a sequence of integers, and for all $j \ge 0$ it holds that $\beta_i(z_j) \subseteq \mathcal{L}_i(r_i(z_j))$ and $\widetilde{\tau}_i(z_j) \in [\tau_i(z_j), \tau_i(z_j+1))$. The timed service word is a sequence of services that are actually provided by agent $i$ and is compliant with the trajectory $x_i(t), t \ge 0$ by construction.
The specification task $\varphi_i$ given as an MITL formula over the set of services $\Sigma_i$ as in Def. \[def:mitl\_semantics\], captures requirements on the services to be provided by agent $i$, for each $i \in \mathcal{I}$. We say that a trajectory $x_i(t)$ satisfies a formula $\varphi_i$ given in MITL over the set $\Sigma_i$, and formally write $x_i(t) \models \varphi_i, \forall t \ge 0$, if and only if there exists a *timed service word* $\widetilde{w}_{x_i}^t$ that complies with $x_i(t)$ and satisfies $\varphi_i$ according to the semantics of Def. \[def:mitl\_semantics\].
\[ex: example\_01\] Consider $N = 2$ agents performing in the partitioned environment of Figure \[fig: example\_01\]. Both agents have the ability to pick up, deliver and throw two different balls. Their sets of services are $\Sigma_1 = \{\rm pickUp1, deliver1, throw1\}$ and $\Sigma_2 = \{\rm pickUp2, deliver2, throw2\}$, respectively, and satisfy $\Sigma_1 \cap \Sigma_2 = \emptyset$. Three points of the agents’ trajectories that belong to different cells with different services are captured. Assume that $t_1 < t_1' < t_2 < t_2 < t_2' < t_3 < t_3'.$ The trajectories $x_1(t), x_2(t), t \ge 0$ are depicted with the red lines. According to Assumption \[assumption: AP\_cell\_decomposition\], the cell decomposition $S = \{S_\ell\}_{\ell \in \mathbb I} = \{S_1, \ldots, S_6\}$ is given where $\mathbb I = \{1, \ldots, 6\}$ respects the labeling functions $\Lambda_i, \mathcal{L}_i, i \in \{1,2\}$. In particular, it holds that: $$\begin{aligned}
&\Lambda_1(x_1(t)) = \mathcal{L}_1(r_1(0)) = \{\rm pickUp1\}, t \in [0, t_1), \notag \\
&\Lambda_1(x_1(t)) = \mathcal{L}_1(r_1(1)) = \{\rm throw1\}, t \in [t_1, t_2), \notag \\
&\Lambda_1(x_1(t)) = \mathcal{L}_1(r_1(2)) = \{\rm deliver1\}, t \in [t_2, t_3), \notag \\
&\Lambda_1(x_1(t)) = \mathcal{L}_1(r_1(3)) = \emptyset, t \ge t_3. \notag \\
&\Lambda_2(x_2(t)) = \mathcal{L}_2(r_2(0)) = \{\rm pickUp2\}, t \in [0, t_1'), \notag \\
&\Lambda_2(x_2(t)) = \mathcal{L}_2(r_2(1)) = \{\rm deliver2\}, t \in [t_1', t_2'), \notag \\
&\Lambda_2(x_2(t)) = \mathcal{L}_2(r_2(2)) = \{\rm throw2\}, t \in [t_2', t_3'), \notag \\
&\Lambda_2(x_2(t)) = \mathcal{L}_2(r_2(3)) = \emptyset, t \ge t_3'. \notag
\end{aligned}$$ By the fact that $w_i(j) = \mathcal{L}(r_i(j)), \forall \ i \in \{1,2\}, j \in \{1,2,3\}$, the corresponding individual timed words are given as: $$\begin{aligned}
w^t_{x_1} &= (\{\rm pickUp1\}, 0)(\{\rm throw1\}, t_1)(\{\rm deliver1\}, t_2)(\emptyset, t_3), \notag \\
w^t_{x_2} &= (\{\rm pickUp2\}, 0)(\{\rm deliver2\}, t_1')(\{\rm throw2\}, t_2')(\emptyset, t_3'). \notag
\end{aligned}$$ According to , two time service words (depicted with in Figure \[fig: example\_01\]) are given as: $$\begin{aligned}
\widetilde{w}_1^t &= (\beta_1(z_0), \widetilde{\tau}_1(z_0))(\beta_1(z_1), \widetilde{\tau}_1(z_1)), \notag \\
\widetilde{w}_2^t &= (\beta_2(z'_0), \widetilde{\tau}_2(z'_0))(\beta_2(z'_1), \widetilde{\tau}_2(z'_1)), \notag \\
\end{aligned}$$ where for agent 1 we have: $z_0 = 0, z_1 = 2, \beta_1(z_0) = \{\rm pickUp1\} \subseteq \mathcal{L}_1(r_1(z_0)), \beta_1(z_1) = \{\rm deliver1\} \subseteq \mathcal{L}_1(r_1(z_1))$. The corresponding elements for agent 2 are $z'_0 = 0, z'_1 = 2, \beta_2(z'_0) = \{\rm pickUp2\} \subseteq \mathcal{L}_2(r_2(z'_0)), \beta_2(z'_1) = \{\rm throw2\} \subseteq \mathcal{L}_2(r_2(z'_1))$. The time stamps $\widetilde{\tau}_1(z_0), \widetilde{\tau}_1(z_1)$ should satisfy the following conditions: $$\begin{aligned}
\widetilde{\tau}_1(z_0) &\in [\tau_1(z_0), \tau_1(z_0+1)) = [0, t_1), \\
\widetilde{\tau}_1(z_1) &\in [\tau_1(z_1), \tau_1(z_1+1)) = [t_2, t_3), \\
\widetilde{\tau}_2(z'_0) &\in [\tau_2(z'_0), \tau_2(z'_1)) = [0, t_1'), \\
\widetilde{\tau}_2(z_1') &\in [\tau_1(z_1'), \tau_1(z_1'+1)) = [t_2', t_3').
\end{aligned}$$
(-2.5, -5.0) grid (0,0); (-7.5,0.0) – (-2.5,0.0); (-7.5,-2.5) – (-2.5,-2.5); (-7.5,-5.0) – (-2.5,-5.0); (-10.0, -5.0) grid (-7.5,0);
(-10, -2.5) rectangle +(2.5, 2.5); (-10, -5.0) rectangle +(2.5, 2.5); (-7.5, -2.5) rectangle +(5.0, 2.5); (-2.5, -2.5) rectangle +(2.5, 2.5); (-2.5, -5.0) rectangle +(2.5, 2.5); (-7.5, -5.0) rectangle +(5.0, 2.5);
(-8.8, -3.7) .. controls (-7.50, -0.0) .. (-5.8, -3.2);
(-5.1, -3.7) .. controls (-0.2, -2.5) .. (-3.0, -1.1);
(-8.8, -3.7) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$x_1(0)$]{}]{}\] (A1) ; (-8.35, -2.5) node\[circle, inner sep=0.8pt, fill=black\] (B1) ; (-7.50, -0.85) node\[circle, inner sep=0.8pt, fill=black, label=[left:[$x_1(t_2)$]{}]{}\] (C1) ; (-6.2, -2.5) node\[circle, inner sep=0.8pt, fill=black\] (D1) ;
(-1.15, -2.5) node\[circle, inner sep=0.8pt, fill=black\] (A2) ; (-2.5, -3.05) node\[circle, inner sep=0.8pt, fill=black\] (B2) ; (-5.1, -3.7) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$x_2(0)$]{}]{}\] (C2) ; (-2.5, -1.35) node\[circle, inner sep=0.8pt, fill=black\] (D2) ;
at (-8.7, 0.5) [$S_1$]{}; at (-5.0, 0.5) [$S_2$]{}; at (-1.3, 0.5) [$S_3$]{}; at (-1.3, -5.5) [$S_4$]{}; at (-5.0, -5.5) [$S_5$]{}; at (-8.7, -5.5) [$S_6$]{}; at (-9,-2.2) [$x_1(t_1)$]{}; at (-5.5,-2.2) [$x_1(t_3)$]{}; at (-1.8,-3.3) [$x_2(t_1')$]{}; at (-0.57,-2.1) [$x_2(t_2')$]{}; at (-1.9,-1.1) [$x_2(t_3')$]{};
Problem Statement
-----------------
We are now ready to define the problem treated in this paper formally as follows:
\[problem: basic\_prob\] Given $N$ agents that are governed by dynamics as in , modeled by the communication graph $\mathcal{G}$, $N$ task specification formulas $\varphi_1, \ldots, \varphi_N$ expressed in MITL over the sets of services $\Sigma_1, \ldots, \Sigma_{{N}}$, respectively, service labeling functions $\Lambda_1, \ldots, \Lambda_N$, as in , a cell decomposition $S = \{S_\ell\}_{\ell \in \mathbb I}$ as in Assumption \[assumption: AP\_cell\_decomposition\] and the labeling functions $\mathcal{L}_1, \ldots, \mathcal{L}_N$ given by , assign control laws to the free inputs $v_1, \ldots, v_N$ such that each agent fulfills its individual specification i.e., $x_i(t) \models \varphi_i, \forall i \in \mathcal{I}, t \ge 0$, given the upper bound $v_{\text{max}}$.
It should be noted that, in this work, the dependencies between the agents are induced through the coupled dynamics and not in the discrete level, by allowing for couplings between the services (i.e., $\Sigma_i \cap \Sigma_j \ne \emptyset$, for some $i, j \in \mathcal I$). Hence, even though the agents do not share atomic propositions, the constraints on their motion due to the dynamic couplings may restrict them to fulfill the desired high-level tasks. Treating additional couplings through individual atomic propositions in the discrete level, constitutes a far from trivial problem, which is a topic of current work.
In our previous work on the multi-agent controller synthesis framework under MITL specifications [@alex_2016_acc], the multi-agent system was considered to have fully-actuated dynamics. The only constraints on the system were due to the presence of time constrained MITL formulas. In the current framework, we have two types of constraints: the constraints due to the coupling dynamics of the system , which constrain the motion of each agent, and, the timed constraints that are inherently imposed from the time bounds of the MITL formulas. Thus, there exist formulas that cannot be satisfied either due to the coupling constraints or the time constraints of the MITL formulas. These constraints, make the procedure of the controller synthesis in the discrete level substantially different and more elaborate than the corresponding multi-agent LTL frameworks in the literature ([@guo_2015_reconfiguration; @frazzoli_vehicle_routing; @belta_2010_product_system; @belta_cdc_reduced_communication]).
The motivation for introducing the cell decomposition $S = \{S_\ell\}_{\ell \in \mathbb I}$ in this Section, comes from the requirement to know a priori which services hold in each part of the workspace. As will be clarified through the problem solution, this is necessary since the abstraction of the workspace (which is part of our proposed solution) may not be compliant with the initial given cell decomposition, and thus, new cell decompositions might be required.
Proposed Solution {#sec: solution}
=================
In this section, a systematic solution to Problem \[problem: basic\_prob\] is introduced. Our overall approach builds on abstracting system though a WTS for each agent and exploiting the fact that the timed runs in the $i$-th WTS project onto the trajectories of agent $i$ while preserving the satisfaction of the individual MITL formulas $\varphi_i, i \in \mathcal{I}$. The following analysis is performed:
1. Initially, the boundedness of the agents’ relative positions is proved, in order to guarantee boundedness of the coupling terms $-\sum_{j \in \mathcal{N}(i)}^{} (x_i - x_j)$. This property is required for the derivation of the symbolic models. (Section \[sec: boundedeness\]).
2. We utilize decentralized abstraction techniques for the multi-agent system, i.e., a discretization of both the workspace and time in order to model the motion capabilities of each agent by a WTS $\mathcal{T}_i, \ i \in \mathcal{I}$ (Section \[sec: abstration\]).
3. Given the WTSs, consistent runs are defined in order to take into consideration the coupling constraints among the agents. The computation of the product of the individual WTSs is also required (Section \[sec: runs consistency\]).
4. A five-step automated procedure for controller synthesis which serves as a solution to Problem \[problem: basic\_prob\] is provided in Section \[sec: synthesis\].
5. Finally, the computational complexity of the proposed approach is discussed in Section \[sec:complexity\].
The next sections provide the proposed solution in detail.
Boundedness Analysis {#sec: boundedeness}
--------------------
\[theorem: theorem\_1\] Consider the multi-agent system modeled by the undirected communication graph $\mathcal{G}$. Assume that the network graph is connected (i.e. $\lambda_2(\mathcal{G}) > 0$) and let $v_i, i \in \mathcal{I}$ satisfy $\| v_i(t) \| \leq v_{\text{max}}, \ \forall \ i \in \mathcal{I}, t \geq 0$. Furthermore, let $\bar{R} > K_2 v_{\text{max}}$ be a positive constant, where $K_2 = \frac{2 \sqrt{N} (N-1) \left\| D(\mathcal{G})^{\top} \right\|}{\lambda_2^2(\mathcal{G})} > 0$ and where $D(\mathcal{G})$ is the network adjacency matrix. Then, for each initial condition $x_i(0) \in \mathbb{R}^n$, there exists a time $T > 0$ such that $\widetilde{x}(t) \in \mathcal{X}, \ \forall t \geq T$, where $\mathcal{X} = \{x \in \mathbb{R}^{Nn} : \|\widetilde{x}\| \leq \bar{R}\}$ and with $\widetilde{x}(t)$ as given in .
The proof can be found in Appendix \[app:proof\_theorem\_1\].
It should be noticed that the relative boundedness of the agents’ positions guarantees a global bound on the coupling terms $-\sum_{j \in \mathcal{N}(i)}^{} (x_i - x_j)$, as defined in . This bound will be later exploited in order to capture the behavior of the system in $\mathcal{X} = \{x \in \mathbb{R}^{Nn} : \|\widetilde{x}\| \leq \bar{R}\}$, by a discrete state WTS.
Abstraction {#sec: abstration}
-----------
In this section we provide the abstraction technique that is adopted from our previous work [@boskos_cdc_2015] in order to capture the dynamics of each agent into Transition Systems. Thereafter, we work completely in discrete level, which is necessary in order to solve Problem \[problem: basic\_prob\].
Firstly, some additional notation is introduced. Given an index set $\mathbb{I}$ and an agent $i \in \mathcal{I}$ with neighbors $j_1, \ldots, j_{N_i} \in \mathcal{N}(i)$, define the mapping $\text{pr}_i : \mathbb{I}^N \rightarrow \mathbb{I}^{N_i+1}$ which assigns to each $N$-tuple ${\bf{l}}= (l_1, \ldots, l_N) \in \mathbb{I}^N$ the $N_i+1$ tuple ${\bf{l}}_i = (l_i, l_{j_1}, \ldots, l_{j_{N_i}}) \in \mathbb{I}^{N_i+1}$ which denotes the indices of the cells where the agent $i$ and its neighbors belong.
### Well-Posed Abstractions
Loosely speaking, an abstraction is characterized by a discretization of the workspace into cells, which we denote by $\bar{S} = \{\bar{S}_{l}\}_{l \in \bar{\mathbb{I}}}$, a time step $\delta t$ and selection of feedback laws in place of the free inputs $v_i(t), \forall i \in \mathcal{I}$. The time step $\delta t$ models the time that an agent needs to transit from one cell to another, and $v_i(t)$ is the controller that guarantees such a transition. Note that the time step $\delta t$ is the same for all the agents. Let us denote by $(\bar{S},\delta t)$ the aforementioned *space-time discretization*.
Before defining formally the concept of well-posed abstractions, an intuitive graphical representation is provided. Consider a cell decomposition $\bar{S} = \{\bar{S}_{l}\}_{l \in \bar{\mathbb{I}} = \{1, \ldots, 12\}}$ as depicted in Figure \[fig: well\_possed\_abstraction\] and a time step $\delta t$. The tails and the tips of the arrows in the figure depict the initial cell and the endpoints of agent’s $i$ trajectories at time $\delta t$ respectively. In both cases in the figure we focus on agent $i$ and consider the same cell configuration for $i$ and its neighbors. By configuration we mean the cell that the agent $i$ and its neighbors belong at a current time. However, different dynamics are considered for Cases (i) and (ii). In Case (i), it can be observed that for the three distinct initial positions in cell $\bar{S}_{l_i}$, it is possible to drive agent $i$ to cell $\bar{S}_{l_i'}$ at time $\delta t$. We assume that this is possible for all initial conditions in this cell and irrespectively of the initial conditions of $i$’s neighbors in their cells and the inputs they choose. It is also assumed that this property holds for all possible cell configurations of $i$ and for all the agents of the system. Thus, we have a *well-posed discretization* for system (i). On the other hand, for the same cell configuration and system (ii), the following can be observed. For three distinct initial conditions of $i$ the corresponding reachable sets at $\delta t$, which are enclosed in the dashed circles, lie in different cells. Thus, it is not possible given this cell configuration of $i$ to find a cell in the decomposition which is reachable from every point in the initial cell and we conclude that discretization is not well-posed for system (ii).
\[scale=.80\]
(0,0) – (4,0); (0,1) – (4,1); (0,2) – (4,2); (0,3) – (4,3);
(0,0) – (0,3); (1,0) – (1,3); (2,0) – (2,3); (3,0) – (3,3); (4,0) – (4,3);
(0,0) – (1,0) – (1,1) – (0,1) – (0,0); (2,1) – (3,1) – (3,2) – (2,2) – (2,1); (3,2) – (4,2) – (4,3) – (3,3) – (3,2); (0,2) – (1,2) – (1,3) – (0,3) – (0,2);
(0.3,2.5) circle (0.3cm); (0.4,2.7) circle (0.3cm); (0.5,2.6) circle (0.3cm);
(0.3,2.5) circle (0.3cm); (0.4,2.7) circle (0.3cm); (0.5,2.6) circle (0.3cm);
(0.2,0.3) .. controls (0.3,1.5) .. (0.3,2.5); (0.2,0.3) circle (1.5pt);
(0.5,0.8) .. controls (0.5,1.5) .. (0.4,2.7); (0.5,0.8) circle (1.5pt);
(0.7,0.6) .. controls (0.7,1.5) .. (0.5,2.6); (0.7,0.6) circle (1.5pt);
\(A) at (0,0.5); (A) at (0.85,0.15); (A) at (3,1.5); (2.3,1.4) circle (1.5pt) node\[right\][$x_{j_{1}}$]{}; (A) at (4,2.5); (3.4,2.8) circle (1.5pt) node\[below\][$x_{j_{2}}$]{}; (A) at (0,2.5);
(5,0) – (9,0); (5,1) – (9,1); (5,2) – (9,2); (5,3) – (9,3);
(5,0) – (5,3); (6,0) – (6,3); (7,0) – (7,3); (8,0) – (8,3); (9,0) – (9,3);
(5,0) – (6,0) – (6,1) – (5,1) – (5,0); (7,1) – (8,1) – (8,2) – (7,2) – (7,1); (8,2) – (9,2) – (9,3) – (8,3) – (8,2); (5,2) – (6,2) – (6,3) – (5,3) – (5,2); (6,2) – (7,2) – (7,3) – (6,3) – (6,2); (6,1) – (7,1) – (7,2) – (6,2) – (6,1);
(5.3,2.5) circle (0.3cm); (6.4,2.7) circle (0.3cm); (6.5,1.7) circle (0.3cm);
(5.3,2.5) circle (0.3cm); (6.4,2.7) circle (0.3cm); (6.5,1.7) circle (0.3cm);
(5.2,0.3) .. controls (5.3,1.5) .. (5.3,2.5); (5.2,0.3) circle (1.5pt);
(5.5,0.8) .. controls (5.5,1.5) .. (6.4,2.7); (5.5,0.8) circle (1.5pt);
(5.7,0.6) .. controls (5.7,1) .. (6.5,1.7); (5.7,0.6) circle (1.5pt);
\(A) at (2,-0.5); (A) at (7,-0.5);
\(A) at (5,0.5); (A) at (5.85,0.15); (A) at (8,1.5); (7.3,1.4) circle (1.5pt) node\[right\][$x_{j_{1}}$]{}; (A) at (9,2.5); (8.4,2.8) circle (1.5pt) node\[below\][$x_{j_{2}}$]{};
(0.4,2.7) – (1.5,3.5); (A) at (1.5,3.5);
(5.3,2.5) – (6.5,3.5); (A) at (6.5,3.5);
More specifically, consider a $(\bar{S},\delta t)$-space-time discretization which is the outcome of the abstraction technique that is designed for the problem solution and will be presented in Section \[sec:discrete\_system\_abstraction\]. Let $\bar{S} = \{\bar{S}_{l}\}_{l \in \bar{\mathbb{I}}}$ be a cell decomposition in which the agent $i$ occupies the cell $\bar{S}_{l_i}$, $\delta t$ be a time step and $\bar{d}_{\text{max}} = \text{sup} \{ \|x - y\| : x,y \in \bar{S}_l, l \in \mathbb{I} \}$ be the diameter of the cell decomposition $\bar{S}$. It should be noted that this decomposition is not necessarily the same cell decomposition $S$ from Assumption \[assumption: AP\_cell\_decomposition\] and Problem \[problem: basic\_prob\]. Through the aforementioned space and time discretization $(\bar{S},\delta t)$ we aim to capture the reachability properties of the continuous system , in order to create a WTS for each agent. If there exists a free input for each state in $\bar{S}_{l_i}$ that navigates the agent $i$ into the cell $\bar{S}_{l_i'}$ precisely in time $\delta t$, regardless of the locations of the agent $i$’s neighbors within their current cells, then a transition from $l_i$ to $\l_i'$ is enabled in the WTS. This forms the well-possessedness of transitions. A more detailed mathematical derivation as well as feedback laws $v_i(t), i \in \mathcal{I}$ which guarantee a well-posed space-time discretization $(\bar{S},\delta t)$ can be found in our previous work [@boskos_cdc_2015].
### Sufficient Conditions
We present at this point the sufficient conditions that relate the dynamics of the multi-agent system , the time step $\delta t$ and the diameter $\bar{d}_{\text{max}}$, and guarantee the existence of the aforementioned well-posed transitions for each cell. Based on our previous work [@boskos_cdc_2015] (Section III, inequality (3), Section IV, inequalities (28, 29)), in order to derive well-posed abstractions, a nonlinear system of the form: $$\label{eq:boskos_dynamics}
\dot{x}_i = f_i(x_i, \mathbf{x}_j)+v_i, i \in \mathcal{I},$$ where $\mathbf{x}_j = (x_{j_1}, \ldots, x_{j_{N_i}}) \in \mathbb{R}^{N_i n}$, should fulfill the following sufficient conditions:
**(C1)** There exists $M > v_{\text{max}} > 0$ such that $\|f_i(x_i, \mathbf{x}_j) \| \leq M, \ \forall i \in \mathcal{I}, \forall \ x \in \mathbb{R}^{Nn} : \text{pr}_i(x) = (x_i, \mathbf{x}_j) \ \text{and} \ \widetilde{x} \in \mathcal{X}$, by applying the projection operator $\text{pr}_i$ for $\mathbb{I} = \mathbb{R}^n$.
**(C2)** There exists a Lipschitz constant $L_1 > 0$ such that: $$\begin{aligned}
\| f_i(x_i, \mathbf{x}_j) - f_i(x_i, \mathbf{y}_j) \| \leq L_1 \|(x_i, \mathbf{x}_j) - (x_i, \mathbf{y}_j) \|, \forall \ i \in \mathcal{I}, x_i, y_i \in \mathbb{R}^n, \mathbf{x}_j, \mathbf{y}_j \in \mathbb{R}^{N_in}.
\label{eq:lipsitch_1}\end{aligned}$$
**(C3)** There exists a Lipschitz constant $L_2 > 0$ such that: $$\begin{aligned}
\| f_i(x_i, \mathbf{x}_j) - f_i(y_i, \mathbf{x}_j) \| \leq L_2 \|(x_i, \mathbf{x}_j) - (y_i, \mathbf{x}_j) \|, \forall \ i \in \mathcal{I}, x_i, y_i \in \mathbb{R}^n, \mathbf{x}_j, \mathbf{y}_j \in \mathbb{R}^{N_i n}.\end{aligned}$$
From and we get $f_i(x_i, \mathbf{x}_j) = \displaystyle -\sum_{j \in \mathcal{N}(i)}^{}(x_i-x_j)$. By checking all the conditions one by one for $f_i(x_i, \mathbf{x}_j)$ as in , it can be shown that our system satisfies all the conditions **(C1)**-**(C3)**. The proof can be found in Appendix \[app:suff\_conditions\].
Based on the sufficient condition for well posed abstractions in [@boskos_cdc_2015], the diameter $\bar{d}_{\max}$ and the time step $\delta t$ of the discretization $\bar{S}, \delta t$ can be selected as:
$$\begin{aligned}
\bar{d}_{\max} & \in \left(0,\frac{(1-\lambda)^2 v_{\max}^{2}}{4ML}\right], \label{dmax} \\
\delta t \in \Bigg[ & \frac{(1-\lambda)v_{\max}-\sqrt{(1-\lambda)^2 v_{\max}^{2}-4ML \bar{d}_{\max}}}{2ML}, \notag \\
& \frac{(1-\lambda)v_{\max}+\sqrt{(1-\lambda)^2 v_{\max}^{2}-4ML \bar{d}_{\max}}}{2ML} \Bigg], \label{deltat}
\end{aligned}$$
where $L=\max\{3L_{2}+4L_{1}\sqrt{N_i},i\in\mathcal{I}\}$ and with the dynamics bound $M$ and the Lipschitz constants $L_1$, $L_2$ as previously defined. Furthermore, $\lambda\in(0,1)$ is a design parameter which quantifies the part of the free input that is additionally exploited for reachability purposes. In particular, given an agent’s initial cell configuration, the agent can reach any point inside an appropriate ball at $\delta t$ through a parameterized feedback law in place of the free input $v_i$. The radius of this ball increases proportionally to the value of $\lambda$, and thus, also the number of the agent’s successor cells, which are the ones intersecting the ball. It is noted that an increasing choice of $\lambda$ results in finer discretizations, therefore providing a quantifiable trade-off between the discrete model’s accuracy and complexity. Furthermore, it follows from the acceptable values of $\bar d_{\max}$ that the cells can be selected coarser, when (i) the available control $v_{\max}$ is larger, and, (ii) the coupling term bound $M$ together with the dynamics’ variation, which is captured through the parameter $L$, are smaller. Analogous restrictions need to hold for the time step $\delta t$. In particular, the time step cannot be selected very large, because the required control for the manipulation of the coupling terms increases due to the evolution of the agent’s neighbors during the transition interval. Finally, the time step cannot be selected very small either, because controlling the agent to the same point from each initial condition in its cell, will require a large control effort over a very short transition interval.
\[remark:d\_max\_remark\] Assume that a cell-decomposition of diameter $\bar{d}_{\text{max}}$ and a time step $\delta t$ which guarantee well-posed transitions, namely, which satisfy and , have been chosen. Then, it is also possible to choose any other cell-decomposition with diameter $\hat{d}_{\text{max}} \leq \bar{d}_{\text{max}}$ since, by , the range of acceptable $\delta t$ increases.
Having shown that the dynamics of system satisfy the sufficient conditions **(C1)**-**(C3)**, a well-posed space-time discretization $(\bar{S}, \delta t)$ has been obtained. Recall now Assumption \[assumption: AP\_cell\_decomposition\]. It remains to establish the compliance of the cell decomposition $S = \{S_\ell\}_{\ell \in \mathbb{I}}$, which is given in the statement of Problem \[problem: basic\_prob\], with the cell decomposition $\bar{S} = \{\bar{S}_l\}_{l \in \bar{\mathbb{I}}}$, which is the outcome of the abstraction. By the term of compliance, we mean that: $$\label{eq:cell_decomposition_compliance}
\bar{S}_l \cap S_\ell \in S \cup \{\emptyset\}, \forall \ \bar{S}_l \in \bar{S}, S_\ell \in S, l \in \bar{\mathbb{I}}, \ell \in \mathbb{I}.$$ In order to address this problem, define: $$\label{eq:final_decomposition}
\hat{S} = \{ \hat{S}_{\hat{l}} \}_{\hat{l} \in \hat{\mathbb{I}}} = \{ \bar{S}_l \cap S_\ell : \ l \in \bar{\mathbb{I}}, \ell \in \mathbb{I} \} \backslash \{ \emptyset \},$$ which forms a cell decomposition and is compliant with the cell decomposition $S$ from Problem \[problem: basic\_prob\] with diameter $\hat{d}_{\text{max}} = \text{sup} \{ \|x - y\| : x,y \in \hat{S}_{\hat{l}}, \hat{l} \in \hat{\mathbb{I}} \} \leq \bar{d}_{\text{max}}$ and serves as the abstraction solution of this problem. By taking all the intersections $\bar{S}_l \cap S_\ell, \forall \ l \in \bar{\mathbb{I}}, \forall \ \ell \in \mathbb{I}$ and enumerating them through the index set $\hat{\mathbb{I}}$, the cells $\{ \hat{S}_{\hat{l}} \}_{\hat{l} \in \hat{\mathbb{I}}}$ are constructed. For the cells $\{ \hat{S}_{\hat{l}} \}_{\hat{l} \in \hat{\mathbb{I}}}$, the following holds: $\forall \ \ell \in \mathbb{I}, \exists \ l \in \bar{\mathbb{I}}$ such that $\hat{S}_{\hat{l}} = \bar{S}_l \cap S_\ell$ and $\text{int}(\hat{S}_{\hat{l}}) \cap \text{int}(\hat{S}_{\hat{l}'}) \neq \emptyset$ for all $\hat{l}' \in \hat{\mathbb{I}} \backslash \{\hat{l}\}$. After all the intersections we have $\cup_{\hat{l} \in \hat{\mathbb{I}}} \hat{S}_{\hat{l}} = \mathcal{X}$. The diameter of the cell decomposition $\hat{S} = \{ \hat{S}_{\hat{l}} \}_{\hat{l} \in \hat{\mathbb{I}}}$ is defined as $\hat{d}_{\text{max}} = \text{sup} \{ \|x - y\| : x,y \in \hat{S}_{\hat{l}}, \hat{l} \in \hat{\mathbb{I}} \} \leq \bar{d}_{\text{max}}$. Hence, according to Remark \[remark:d\_max\_remark\], we have a well-posed abstraction. The following Example is an illustration of these derivations.
\[ex: example\_03\] Let $S = \{S_\ell\}_{\ell \in \{1,\ldots,6\}}$ be the cell decomposition of Problem \[problem: basic\_prob\], which is depicted in Figure \[fig: example\_03\] by the red rectangles. In the same figure, we illustrate the cell decomposition $\bar{S} = \{\bar{S}_l\}_{l \in \bar{\mathbb{I}} = \{1, \ldots, 6\}}$. $\bar{S}$ serves as potential solution of this Problem satisfying all the abstraction properties that have been mentioned in this Section. It can be observed that the two cell decompositions are not compliant according to . However, by using the methodology below Remark \[remark:d\_max\_remark\], a new cell decomposition $\hat{S} = \{ \hat{S}_{\hat{l}} \}_{\hat{l} \in \hat{\mathbb{I}} = \{1, \ldots, 15\}}$, which is compliant with $S$ and has $15$ regions, can be obtained. $\hat{S}$ forms the final cell decomposition solution of Problem \[problem: basic\_prob\] and is depicted in Figure \[fig:example\_04\]. Let also $\bar{d}_{\text{max}}, \hat{d}_\text{max}$ be the diameters of the cell decompositions $\bar{S}, \hat{S}$ respectively. Then, it holds that $\hat{d}_\text{max} \leq \bar{d}_{\text{max}}$, which is in accordance with Remark \[remark:d\_max\_remark\].
(-2.5, -5.0) grid (0,0); (-7.5,0.0) – (-2.5,0.0); (-7.5,-2.5) – (-2.5,-2.5); (-7.5,-5.0) – (-2.5,-5.0); (-10.0, -5.0) grid (-7.5,0);
(0,0) rectangle (-3.5,-3.5); (0,-3.5) rectangle (-3.5,-5.0); (-3.5,-3.5) rectangle (-6.5,-5); (-3.5,0) rectangle (-6.5,-3.5); (-6.5,0) rectangle (-10.0,-3.5); (-6.5,-3.5) rectangle (-10.0,-5.0);
at (-8.7, -1.2) [$S_1$]{}; at (-5.0, -0.9) [$S_2$]{}; at (-1.3, -1.2) [$S_3$]{}; at (-1.3, -4.0) [$S_4$]{}; at (-5.0, -4.0) [$S_5$]{}; at (-8.7, -4.0) [$S_6$]{};
(-1.8, -1.8) – (2.0, -1.8); at (0.80, -1.75) [$\bar{S}_3$]{};
(-1.8, -4.2) – (1.8, -5.2); at (0.65, -4.9) [$\bar{S}_4$]{};
(-4.8, -1.8) – (-3.2, 2.0); at (-3.6, 0.8) [$\bar{S}_2$]{};
(-4.8, -4.3) – (-4.8, -7.0); at (-4.8, -5.9) [$\bar{S}_5$]{};
(-8.5, -4.3) – (-8.5, -7.1); at (-8.5, -5.9) [$\bar{S}_6$]{};
(-8.5, -1.7) – (-12.2, -1.7); at (-11, -1.7) [$\bar{S}_1$]{};
(-2.5,0) – (-7.5,-2.5); (A) at (-6.2,-1.6);
(-2.5, -5.0) grid (0,0); (0,0) rectangle (-3.5,-3.5); (-7.5,0.0) – (-2.5,0.0); (-7.5,-2.5) – (-2.5,-2.5); (-7.5,-5.0) – (-2.5,-5.0); (-10.0, -5.0) grid (-7.5,0);
(0,-3.5) rectangle (-3.5,-5.0); (-3.5,-3.5) rectangle (-6.5,-5); (-3.5,0) rectangle (-6.5,-3.5); (-6.5,0) rectangle (-10.0,-3.5); (-6.5,-3.5) rectangle (-10.0,-5.0);
(0,0) rectangle (-2.5,-2.5); (0,-2.5) rectangle (-2.5,-3.5); (0,-3.5) rectangle (-2.5,-3.5); (0,-3.5) rectangle (-2.5,-5.0); (-2.5,0) rectangle (-3.5,-2.5); (-2.5,-2.5) rectangle (-3.5,-3.5); (-2.5,0) rectangle (-3.5,-2.5); (-2.5,-3.5) rectangle (-3.5,-5.0); (-3.5,0.0) rectangle (-6.5,-2.5); (-6.5,0.0) rectangle (-7.5,-2.5); (-7.5,0.0) rectangle (-7.5,-2.5); (-7.5,0.0) rectangle (-10.0,-2.5); (-3.5,-2.5) rectangle (-6.5,-3.5); (-6.5,-2.5) rectangle (-7.5,-3.5); (-7.5,-2.5) rectangle (-10.0,-3.5); (-3.5,-3.5) rectangle (-6.5,-5.0); (-6.5,-3.5) rectangle (-7.5,-5.0); (-7.5,-3.5) rectangle (-10.0,-5.0);
at (-8.9, -0.8) [$\hat{S}_1$]{}; at (-7.0, -1.2) [$\hat{S}_2$]{}; at (-5.0, -1.2) [$\hat{S}_3$]{}; at (-3.0, -1.2) [$\hat{S}_4$]{}; at (-1.3, -1.2) [$\hat{S}_5$]{}; at (-1.3, -2.95) [$\hat{S}_6$]{}; at (-3.0, -2.95) [$\hat{S}_7$]{}; at (-5.0, -2.95) [$\hat{S}_8$]{}; at (-7.0, -2.95) [$\hat{S}_{9}$]{}; at (-8.7, -2.95) [$\hat{S}_{10}$]{}; at (-8.7, -4.20) [$\hat{S}_{11}$]{}; at (-7.0, -4.20) [$\hat{S}_{12}$]{}; at (-5.0, -4.20) [$\hat{S}_{13}$]{}; at (-3.0, -4.20) [$\hat{S}_{14}$]{}; at (-1.3, -4.20) [$\hat{S}_{15}$]{};
(-7.5,0) – (-10.0,-2.5); at (-8.8, -2.10) [$\hat{d}_{\text{max}}$]{};
### Discrete System Abstraction {#sec:discrete_system_abstraction}
For the solution to Problem \[problem: basic\_prob\], the WTS of this agent which corresponds to the cell decomposition $\hat{S}$ with diameter $\hat{d}_{\text{max}}$ and the time step $\delta t$ will be exploited. Thus, the WTS of each agent is defined as follows:
\[def: indiv\_WTS\] The motion of each agent $i \in \mathcal{I}$ in the workspace is modeled by the WTS $\ \mathcal{T}_i = (S_i, S_i^{\text{init}}, Act_i, \longrightarrow_i, d_i, AP_i, \hat L_i)$ where:
- $S_i = \hat{\mathbb{I}}$ is the set of states of each agent which is the set of indices of the cell decomposition.
- $S_i^{\text{init}} \subseteq S_i$ is a set of initial states defined by the agents’ initial positions in the workspace.
- $Act_i = {\hat{\mathbb{I}}}^{N_i+1}$, the set of actions representing where agent $i$ and its neighbors are located.
- For a pair $(l_i, {\bf{l}_i}, l'_i)$ we have that $(l_i, {\bf{l}_i}, l'_i) \in \longrightarrow_i$ iff $l_i \overset{\bf{l}_i}{\longrightarrow_i} l_i'$ is well-posed for each $l_i, l'_i \in S_i$ and ${\bf{l}_i} = (l_i, l_{j_1}, \ldots, l_{j_{N_i}}) \in Act_i$.
- $d_i: \longrightarrow_i \rightarrow \mathbb{T}$, is a map that assigns a positive weight (duration) to each transition. The duration of each transition is exactly equal to $\delta t > 0$.
- ${\mathit{AP}}_i = \Sigma_i$, is the set of atomic propositions which are inherent properties of the workspace.
- $L_i: S_i \rightarrow 2^{AP_i}$, is the labeling function that maps every state $s \in S_i$ into the services that can be provided in this state.
The individual WTSs of the agents will allow us to work completely in the discrete level and design sequences of controllers that solve Problem \[problem: basic\_prob\].
Every WTS $\mathcal{T}_i, i \in \mathcal{I}$ generates timed runs and timed words of the form $r_i^t = (r_i(0)$, $\tau_i(0))(r_i(1)$, $\tau_i(1))(r_i(2)$, $\tau_i(2)) \ldots$, $w_i^t = (L_i(r_i(0))$, $\tau_i(0))(L_i(r_i(1))$, $\tau_i(1))(L_i(r_i(2))$, $\tau_i(2)) \ldots$ respectively, over the set $2^{AP_i}$ according to Def. \[run\_of\_WTS\] with $\tau_i(j) = j \delta t, \forall \ j \ge 0$. The relation between the timed words that are generated by the WTSs $\mathcal{T}_i, i \in \mathcal{I}$ with the timed service words produced by the trajectories $x_i(t), i \in \mathcal{I}, t \ge 0$ is provided through the following remark:
\[lemma:compliant\_WTS\_runs\_with\_trajectories\] By construction, each timed word produced by the WTS $\mathcal{T}_i$ is a service timed word associated with the trajectory $x_i(t)$ of the system . Hence, if we find a timed word of $\mathcal{T}_i$ satisfying a formula $\varphi_i$ given in MITL, we also find for each agent $i$ a desired timed word of the original system, and hence trajectories $x_i(t)$ that are a solution to the Problem \[problem: basic\_prob\]. Therefore, the produced timed words of $\mathcal{T}_i$ are compliant with the service timed words of the trajectories $x_i(t)$.
Runs Consistency {#sec: runs consistency}
----------------
Due to the coupled dynamics between the agents, it is required that each individual agent’s run is compliant with the corresponding discrete trajectories of its neighbors, which determine the actions in the agent’s run. Therefore, even though we have the individual WTS of each agent, the runs that the latter generates may not be performed by an agent due to the constrained motion that is imposed by the coupling terms. Hence, we need to synchronize the agents at each time step $\delta t$ and determine which of the generated runs of the individual WTS can be performed by the agent. Hereafter, they will be called *consistent runs*. In order to address the aforementioned issue, we provide a centralized product WTS which captures the behavior of the coupled multi-agent system as a team, and the generated product run (see Def. \[def: consistent\_runs\]) can later be projected onto consistent individual runs. The following two definitions deal with the product WTS and consistent runs respectively.
\[def:product\_TS\] Given the individual WTSs $\mathcal{T}_i, i \in \mathcal{I}$ from Def. \[def: indiv\_WTS\], the product WTS $\mathcal{T}_p = (S_p, S_p^{\text{init}}, \longrightarrow_p, L_p)$ is defined as follows:
- $S_p = \hat{\mathbb{I}}^N$;
- $(s_1, \ldots, s_N) \in S^{\text{init}}$ if $s_i \in S_i^{\text{init}}, \forall \ i \in \mathcal{I}$;
- $(\bf{l}, \bf{l}') \in \longrightarrow_p$ iff $l_i' \in \text{Post}_i(l_i, \text{pr}_i(\bf{l})), \forall \ i \in \mathcal{I}, \forall \ \bf{l} = (l_1, \ldots, l_N), \bf{l}' = (l'_1, \ldots, l'_N)$;
- $L_p:\hat{\mathbb I}^N\to 2^{\cup_i^N\Sigma_i}$ defined as $L_p({\bf l})=\cup_{i=1}^NL_i(l_i)$;
- $d_p: \longrightarrow_p \rightarrow \mathbb{T}$: as in the individual WTS’s case, with transition weight $d_p(\cdot) = \delta t$.
\[def: consistent\_runs\] Given a timed run: $$\begin{aligned}
&r_p^t = ((r_p^1(0), \ldots , r_p^N(0)),\tau_p(0)) \noindent ((r_p^1(1),\ldots,r_p^N(1)),\tau_p(1))\ldots, \notag
\end{aligned}$$ that is generated by the product WTS $\mathcal T_p$, the induced set of projected runs $$\{r_i^t = (r_p^i(0), \tau_p(0))(r_p^i(1), \tau_p(1)) \ldots : i \in \mathcal I\},$$ of the WTSs $\mathcal T_1,\ldots, \mathcal T_N$, respectively will be called *consistent runs*. Since the duration of each agent’s transition is $\delta t$ it holds that $\tau_p(j) = j \delta t, j \geq 0$.
Therefore, through the product WTS $\mathcal{T}_p$, we can always generate individual consistent runs for each agent. It remains to provide a systematic approach of how to determine consistent runs $\widetilde{r}_1, \ldots, \widetilde{r}_N$ which are associated with the corresponding time serviced words $\widetilde{w}_1^t, \ldots, \widetilde{w}_N^t$. Note that we use the tilde accent to denote timed runs and words that correspond to the problem solution. The corresponding compliant trajectories $x_1(t), \ldots, x_N(t)$ of the timed words $\widetilde{w}_1^t, \ldots, \widetilde{w}_N^t$ satisfy the corresponding MITL formulas $\varphi_1, \ldots, \varphi_N$, and they are a solution to Problem \[problem: basic\_prob\]. This follows from the fact that the product transition system is simulated by the $\delta t$-sampled version of the continuous system (see [@tabuada_book_verification] for the definition of a simulation relation). In particular, let $\mathcal{T}_{\delta t}$ be the $\delta t$-sampled WTS of system (1), as defined in [@tabuada_book_verification Def. 11.4], with labeling function $L_{\delta t}:\mathbb R^{Nn}\to 2^{\cup_{i=1}^N\Sigma_i}$ given as $L_{\delta t}(x_1,\ldots,x_N)=\cup_{i=1}^N \Lambda_i(x_i)$ and $\Lambda_i$ as defined in Section 2. Consider also the WTS $\mathcal{T}_{p}$ and the relation $\mathcal{R} \subseteq S_{p} \times \mathcal{X}^N$ given as $({\bf l},(x_1,\ldots,x_N))\in\mathcal{R}$, iff $(x_1,\ldots,x_N)\in S_{l_1}\times\cdots\times S_{l_N}$, where ${\bf l}=(l_1,\ldots,l_N)$. Then, from the definition of the agent’s individual transitions in each WTS $\mathcal T_i$ and the fact that for all points in a cell the same atomic proposition hold true, it can be deduced that $\mathcal{R}$ is a simulation relation from $\mathcal{T}_{p}$ to the $\delta t$-sampled WTS $\mathcal{T}_{\delta t}$.
We chose to utilize decentralized abstractions, to generate the individual WTSs $\mathcal{I}, i \in \mathcal{I}$ for each agent and to compute the synchronized-centralized product WTS $\mathcal{T}_p$ for the following reasons:
1. The state space of the centralized system to be abstracted is $\mathcal{X}^N \subseteq \mathbb{R}^{Nn}$, which is harder to visualize and handle as well as not naturally related to the individual specifications. Thus, it is more “natural” to define the specifications through the individual transition system of each agent corresponding to a discretization of $\mathcal{X}$ and then generate the product in order to obtain potential consistent satisfying plans.
2. Additionally, many centralized abstraction frameworks are based on approximations of the system’s reachable sets from a given cell over the transition time interval. These frameworks, require in the general nonlinear case, global dynamics properties and may avoid taking into account the finer dynamics properties of the individual entities, which can lead to more conservative estimates for large scale systems.
We provide here an example that explains the notation that has been introduced until now.
\[ex: example\_2\]
(0, 0) rectangle +(3, 3.0); (-7.5, -3) rectangle +(3, 3.0); (-4.5, 0) rectangle +(4.5, 3.0); (-7.5, 0) rectangle +(3, 3.0); (0, -3) rectangle +(3, 3.0); (-4.5, -3) rectangle +(4.5, 3);
(-7.5, -3) grid (3,3);
(-6.8,-0.6) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$i$]{}]{}\] (A) ; (-3.8, 0.7) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (B) ; (-0.8, -0.7) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (C) ; (0.8, 0.7) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (D) ; (-5.3, -2.1) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$j_2$]{}]{}\] (E) ; (-6.8, 2.3) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (F) ; (-5.0, 2.2) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (G) ; (-0.8, 2.2) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (H) ; (2.2, 2.2) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (I) ; (-3.8, -0.85) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (K) ; (-0.8, -2.30) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (L) ; (0.75, -0.8) node\[circle, inner sep=0.8pt, fill=black, label=[below:[$ $]{}]{}\] (M) ;
at (-6.8, 2.69) [$j_1$]{};
\(A) to \[bend left=35\] (B); (B) to \[bend left=-35\] (C); (C) to \[bend left=35\] (D); (F) to \[bend left=15\] (G); (G) to \[bend left=15\] (H); (H) to \[bend right=15\] (I); (E) to \[bend right=15\] (K); (K) to \[bend right=15\] (L); (L) to \[bend right=15\] (M);
(1.5,-3) – (3,-1.5); (N) at (2.2,-2);
at (1.8, 2.7) [$22$]{}; at (1.9, -0.3) [$8$]{}; at (-7.2, 1.20) [$15$]{}; at (-7.2, -1.9) [$1$]{}; at (-2.3, 3.5) [$S_5$]{}; at (1.7, 3.4) [$S_4$]{}; at (1.7, -3.7) [$S_3$]{}; at (-2.3, -3.7) [$S_2$]{}; at (-6.3, -3.7) [$S_1$]{}; at (-6, 3.5) [$S_6$]{};
at (-5.6, 0.95) [$\delta t$]{}; (A) – (E); (A) – (F); (B) – (G); (C) – (H); (D) – (I); (K) – (B); (L) – (C); (M) – (D);
Consider $N = 3$ agents performing in the workspace with $\mathcal{N}(i) = \{1,2\}$ as depicted in Figure \[fig: example\_1\]. $S = \{S_\ell\}_{\ell \in \mathbb{I} = \{1, \ldots, 6\}}$ is the given cell decomposition from Problem \[problem: basic\_prob\] and ${S} = \{\bar{S}_{l}\}_{l \in \mathbb{I} = \{1, \ldots, 28\}}$ is the cell decomposition which is the outcome of the proposed abstraction technique. Let the atomic propositions be $\{p_1,\ldots,p_6\} = \{\rm orange, $ $green, blue, yellow, red, grey\}$. The red arrows represent both the transitions of the agent $i$ and its neighbors. The dashed lines indicate the edges in the network graph. For the atomic propositions we have that $L_i(14) = \{p_1\}, L_i(17) = \{p_5\}, L_i(10) = \{p_2\}, L_i(20) = \{p_4\}, L_{j_1}(28) = \{p_6\} = L_{j_1}(27), L_{j_1}(24) = \{p_5\}, L_{j_1}(22) = \{p_4\}, L_{j_2}(2) = \{p_1\}, L_{j_2}(12) = \{p_2\} = L_{j_2}(5), L_{j_2}(9) = \{p_3\}$. Note also the diameter of the cells $\hat{d}_{\text{max}} = \bar{d}_{\text{max}}$. For the cell configurations we have: $$\begin{aligned}
Init \ (t = 0)&:
\begin{cases}
{\bf{l}_i} = (14, 28, 2), \\
{\bf{l}_{j_1}} = (28, 14), \\
{\bf{l}_{j_2}} = (2, 14), \\
\end{cases}
Step \ 1 \ (t = \delta t) :
\begin{cases}
{\bf{l}_i} = (17, 27, 13), \\
{\bf{l}_{j_1}} = (27, 17), \\
{\bf{l}_{j_2}} = (13, 17), \\
\end{cases} \notag \\
Step \ 2 \ (t = 2 \delta t)&:
\begin{cases}
{\bf{l}_i} = (10, 24, 5), \\
{\bf{l}_{j_1}} = (24, 10), \\
{\bf{l}_{j_2}} = (5, 10), \\
\end{cases}
Step \ 3 \ (t = 3 \delta t) :
\begin{cases}
{\bf{l}_i} = (20, 22, 9), \\
{\bf{l}_{j_1}} = (22, 20), \\
{\bf{l}_{j_2}} = (9, 20), \\
\end{cases} \notag
\end{aligned}$$ which are actions to the corresponding transitions. Three consistent timed runs are given as: $$\begin{aligned}
r_i^t & = (r_i(0) = 14, \tau_i(0) = 0) (r_i(1) = 17, \tau_i(1) = \delta t)(r_i(2) = 10, \tau_i(2) = 2 \delta t) \notag \\
& \hspace{60mm} (r_i(3) = 20, \tau_i(3) = 3 \delta t), \notag \\
r_{j_1}^t & = (r_{j_1}(0) = 28, \tau_{j_1}(0) = 0) (r_{j_1}(1) = 27, \tau_{j_1}(1) = \delta t) (r_{j_1}(2) = 24, \tau_{j_1}(2) = 2 \delta t) \notag \\
& \hspace{60mm} (r_{j_1}(3) = 22, \tau_{j_1}(3) = 3\delta t), \notag \\
r_{j_2}^t & = (r_{j_2}(0) = 2, \tau_{j_2}(0) = 0) (r_{j_2}(1) = 13, \tau_{j_2}(1) = \delta t)(r_{j_2}(2) = 5, \tau_{j_2}(2) = 2 \delta t) \notag \\
& \hspace{60mm} (r_{j_2}(3) = 9, \tau_{j_2}(3) = 3\delta t). \notag
\end{aligned}$$ It can be observed that $r_i^t \models (\varphi_i = \Diamond_{[0, 6]}\{yellow\})$ if $3 \delta t \in [0, 6]$, $r_{j_1}^t \models (\varphi_{j1} = \Diamond_{[3, 10]}\{red\})$ if $2 \delta t \in [3, 10]$ and $r_{j_2}^t \models (\varphi_{j2} = \Diamond_{[3, 9]}\{blue\})$ if $3 \delta t \in [3, 9]$.
Controller Synthesis {#sec: synthesis}
--------------------
The proposed controller synthesis procedure is described with the following steps:
1. $N$ TBAs $\mathcal{A}_i, \ i \in \mathcal{I}$ that accept all the timed runs satisfying the corresponding specification formulas $\varphi_i, i \in \mathcal{I}$ are constructed.
2. A Büchi WTS $\tilde{\mathcal T}_i = \mathcal{T}_i \otimes \mathcal{A}_i$ (see Def. \[def: buchi\_WTS\] below) for every $i \in \mathcal{I}$ is constructed. The accepting runs of $\tilde{\mathcal T}_i$, computed using standard graph search algorithms, are the individual runs of the $\mathcal{T}_i$ that satisfy the corresponding MITL formula $\varphi_i, \ i \in \mathcal{I}$.
3. We pick a set of accepting runs $\{\widetilde{r}^t_1, \ldots, \widetilde{r}^t_N\}$ from Step 2. We check if they are consistent according to Def. \[def: consistent\_runs\]. If this is true then we proceed with Step 5. If this is not true then we repeat Step 3 with a different set of accepting runs. At worst case, we perform a finite predefined number of selections $R_{\text{selec}}$; if a consistent set of accepting runs is not found, we proceed with the less efficient centralized procedure in Step 4, which however searches through all sets of all possible accepting runs.
4. We create the product $\widetilde{\mathcal{T}}_p = \mathcal{T}_p \otimes \mathcal{A}_p$ where $\mathcal{A}_p$ is the TBA that accepts all the words that satisfy the formula $\varphi = \varphi_1 \wedge \ldots \wedge \varphi_N$. An accepting run $\widetilde{r}_p$ of the product is projected into the accepting runs $\{\widetilde{r}_1, \ldots, \widetilde{r}_N\}$. If there is no accepting run found in $\mathcal{T}_p \otimes \mathcal{A}_p$, then Problem \[problem: basic\_prob\] has no solution.
5. The abstraction procedure allows to find an explicit feedback law for each transition in $\mathcal T_i$. Therefore, an accepting run $\widetilde{r}^t_i$ in $\mathcal T_i$ that takes the form of a sequence of transitions is realized in the system in via the corresponding sequence of feedback laws.
(-15.6, 6.0) rectangle +(2.3, 0.9); at (-14.45, 6.45) [$\text{MITL2TBA}$]{}; (-15.6, 0.80) rectangle +(2.3, 0.9); at (-14.45, 1.25) [$\text{MITL2TBA}$]{};
(-14.5,6.0) – (-14.5,4.6); (-14.5,7.6) – (-14.5,6.9); (-14.5,0.1) – (-14.5,0.8); (-14.5,1.7) – (-14.5,3.1);
at (-14.5, 7.80) [$\varphi_1$]{}; at (-14.5, -0.1) [$\varphi_N$]{};
at (-14.0, 5.35) [$\mathcal{A}_1$]{}; at (-14.0, 2.40) [$\mathcal{A}_N$]{};
at (-14.5, 3.25) [$\otimes$]{}; at (-14.5, 4.45) [$\otimes$]{};
at (-14.5, 3.95) [$\vdots$]{};
(-14.30,4.45) – (-12.6,4.45); (-14.30,3.25) – (-12.6,3.25);
at (-12.3, 4.47) [$\widetilde{\mathcal{T}}_1$]{}; at (-12.3, 3.95) [$\vdots$]{}; at (-12.3, 3.26) [$\widetilde{\mathcal{T}}_N$]{};
(-16.35,4.45) – (-14.7,4.45); (-16.35,3.25) – (-14.7,3.25);
at (-16.7, 4.47) [$\mathcal{T}_1$]{}; at (-16.7, 3.95) [$\vdots$]{}; at (-16.7, 3.26) [$\mathcal{T}_N$]{};
(-12.00,4.45) – (-11.3,4.45); (-12.00,3.25) – (-11.3,3.25);
(-11.3, 2.85) rectangle +(1.85, 0.70); (-11.3, 4.15) rectangle +(1.85, 0.70); at (-10.32, 4.50) [$\text{synthesis}$]{}; at (-10.32, 3.20) [$\text{synthesis}$]{};
(-9.35,4.45) – (-8.65,4.45); (-9.35,3.25) – (-8.65,3.25);
at (-8.35, 4.47) [$\widetilde{r}_1^t$]{}; at (-8.35, 3.95) [$\vdots$]{}; at (-8.35, 3.26) [$\widetilde{r}_N^t$]{};
(-17.70,4.47) – (-17.00,4.47); (-17.70,3.26) – (-17.00,3.26);
(-19.6, 2.90) rectangle +(1.8, 0.7); (-19.6, 4.20) rectangle +(1.8, 0.7); at (-18.70, 4.60) [$\tiny \text{abstraction}$]{}; at (-18.70, 3.30) [$\tiny \text{abstraction}$]{};
(-20.35,4.47) – (-19.65,4.47); (-20.35,3.26) – (-19.65,3.26);
at (-22.70, 4.47) [$\displaystyle \dot{x}_1 = -\sum_{j \in \mathcal{N}(1)}^{} (x_1 - x_j)+v_{1}$]{}; at (-22.70, 3.95) [$\vdots$]{}; at (-22.70, 3.26) [$\displaystyle \dot{x}_N = -\sum_{j \in \mathcal{N}(N)}^{} (x_N - x_j)+v_{N}$]{};
(-8.35, 4.80) – (-8.35, 8.50); (-8.35, 8.50) – (-18.00, 8.50); (-19.00, 8.50) – (-23.00, 8.50); (-23.00, 8.50) – (-23.00, 5.1);
(-8.35, 2.90) – (-8.35, -0.80); (-8.35, -0.80) – (-18.00, -0.80); (-19.00, -0.8) – (-23.00, -0.8); (-23.00, -0.8) – (-23.00, 2.6);
(-19.0, 7.95) rectangle +(1.0, 1.0); at (-18.5, 8.40) [$v_1(t)$]{}; (-19.0, -1.25) rectangle +(1.0, 1.0); at (-18.5, -0.80) [$v_N(t)$]{};
In order to construct the Buchi WTSs $\widetilde{\mathcal{T}}_p$ and $\widetilde{\mathcal{T}}_i, i \in \mathcal{I}$ that were presented in Steps 2 and 4, consider the following generic definition:
\[def: buchi\_WTS\] Given a WTS $\mathcal{T}_i =(S_i, S_{i}^{\text{init}}, Act_i, \longrightarrow_i, d_i, AP_i, L_i)$, and a TBA ${\mathcal{A}}_i = (Q_i, Q^\text{init}_i, C_i, Inv_i, E_i, F_i, \\ AP_i, \mathcal{L}_i)$ with $|C_i|$ clocks and let $C^{\mathit{max}}_i$ be the largest constant appearing in ${\mathcal{A}}_i$. Then, their *Büchi WTS* $\widetilde{{\mathcal{T}}}_i = \mathcal{T}_i \otimes {\mathcal{A}}_i = (\widetilde{S}_i, \widetilde{S}_{i}^{{\mathit{init}}}, \widetilde{Act}_i, {\rightsquigarrow}_{i}, \widetilde{d}_i, \widetilde{F}_i, AP_i, \widetilde{L}_i)$ is defined as follows:
- [$\widetilde{S}_i \subseteq \{(s_i, q_i) \in S_i \times Q_i : {L}_i(s_i) = \mathcal{L}_i(q_i)\} \times \mathbb{T}_\infty^{|C_i|} $.]{}
- $\widetilde{S}_{i}^{{\mathit{init}}} = S_i^{{\mathit{init}}} \times Q_i^{{\mathit{init}}} \times \{0\}^{|C_i|}$.
- $\widetilde{Act}_i = Act_i$.
- $(\widetilde{q}, act_i, \widetilde{q} ') \in {\rightsquigarrow}_i$ iff
- $\widetilde{q} = (s, q, \nu_1, \ldots, \nu_{|C_i|}) \in \widetilde{S}_i$,\
$\widetilde{q} ' = (s', q', \nu_1', \ldots, \nu_{|C_i|}') \in \widetilde{S}_i$,
- $act_i \in Act_i$,
- $(s, act_i, s') \in \longrightarrow_i$, and
- there exists $\gamma, R$, such that $(q, \gamma, R, q') \in E_i$, $\nu_1,\ldots,\nu_{|C_i|} \models \gamma$, $\nu_1',\ldots,\nu_{|C_i|}' \models Inv_i(q')$, and for all $i \in \{1,\ldots, |C_i|\}$ it holds that: $$\nu_i' =
\begin{cases}
0, & \text{if } c_i \in R \\
\nu_i + d_i(s, s'), & \text{if } c_i \not \in R \text{ and } \\ & \nu_i + d_i(s, s') \leq C^{\mathit{max}}_i \\ \infty, & \text{otherwise}.
\end{cases}$$
Then, $\widetilde{d}_i(\widetilde{q}, \widetilde{q}') = d_i(s, s')$.
- $\widetilde{F}_i = \{(s_i, q_i,\nu_1,\ldots,\nu_{|C_i|}) \in Q_i : q_i \in F_i\}$.
- $\widetilde{L}_i(s_i, q_i, \nu_1, \ldots, \nu_{|C_i|}) = {L}_i(s_i)$.
The Buchi WTS $\widetilde{\mathcal{T}}_p$ is constructed in a similar way to Def. \[def: buchi\_WTS\] by using the product of $\mathcal{T}_p$ and $\mathcal{A}_p$. Each Büchi WTS $\widetilde{\mathcal{T}}_i, i \in \mathcal I$ is in fact a WTS with a Büchi acceptance condition $\widetilde{F}_i$. A timed run of $\widetilde{\mathcal{T}}_i$ can be written as $\widetilde{r}_i^t = (q_i(0), \tau_i(0))(q_i(1), \tau_i(1)) \ldots$ using the terminology of Def. \[run\_of\_WTS\]. It is *accepting* if $q_i(j) \in \widetilde F_i$ for infinitely many $i \geq 0$. An accepting timed run of $\widetilde{\mathcal{T}}_i$ projects onto a timed run of $\mathcal{T}_i$ that satisfies the local specification formula $\varphi_i$ by construction. Formally, the following lemma, whose proof follows directly from the construction and and the principles of automata-based LTL model checking (see, e.g., [@katoen]), holds:
\[eq: lemma\_1\] Consider an accepting timed run $\widetilde{r}_i^t = (q_k(0), \tau_i(0))(q_i(1), \tau_i(1)) \ldots$ of the Büchi WTS $\widetilde{\mathcal{T}}_i$ defined above, where $q_i(j) = (r_i(j), s_i(j), \nu_{i, 1}, \ldots, \nu_{i, |C_i|})$ denotes a state of $\mathcal{\widetilde T}_i$, for all $j \geq 0$. The timed run $\widetilde{r}_i^t$ projects onto the timed run $r_i^t = (r_i(0), \tau_i(0))(r_i(1), \tau_i(1)) \ldots$, of the WTS $\mathcal{T}_i$ that produces the timed word $w(r_i^t) = (L_i(r_i(0))$, $\tau_i(0))(L_i(r_i(1))$, $\tau_i(1)) \ldots$ accepted by the TBA $\mathcal{A}_i$ via its run $\rho_i = s_i(0)s_i(1) \ldots$. Vice versa, if there exists a timed run $r_i^t = (r_i(0),\tau_i(0))(r_i(1),\tau_i(1))\ldots,$ of the WTS $\mathcal{T}_k$ that produces a timed word $w(r_i^t) = (L_i(r_i(0))$, $\tau_i(0))(L_i(r_i(1))$, $\tau_i(1)) \ldots$ accepted by the TBA $\mathcal{A}_i$ via its run $\rho_i = s_i(0)s_i(1)\ldots$ then there exists the accepting timed run $\widetilde{r}_i^t = (q_i(0)$, $\tau_i(0))(q_i(1)$, $\tau_i(1)) \ldots$ of $\widetilde{\mathcal{T}}_i$, where $q_i(j) = (r_i(j),s_i(j),\nu_{i,1}, \ldots, \nu_{i,|C_i|})$, in $\widetilde{\mathcal{T}}_i$.
The proposed framework is depicted in Figure \[fig:solution\_scheme\]. The dynamics of each agent $i$ is abstracted into a WTS $\mathcal{T}_i$ (orange rectangles). Then the product between each WTS $\mathcal{T}_i$ and the $TBA$ $\mathcal{A}_i$ is computed according to Def. \[def: buchi\_WTS\]. The TBA $\mathcal{A}_i$ accepts all the words that satisfy the formula $\varphi_i$ (blue rectangles). For every Büchi WTS $\widetilde{\mathcal{T}}_i$ the controller synthesis procedure that was described in this Section (red rectangles) is performed and a sequence of accepted runs $\{\widetilde{r}_1^t, \dots, \widetilde{r}_N^t\}$ is designed. Every accepted run $\widetilde{r}_i^t$ maps into a decentralized controller $v_i(t)$ which is a solution to Problem \[problem: basic\_prob\].
A solution obtained from Steps 1-5, gives a sequence of controllers $v_1, \dots, v_N$ that guarantees the satisfaction of the formulas $\varphi_1, \dots, \varphi_N$ of the agents $1, \dots, N$ respectively, governed by the dynamics as in , thus, they are a solution to Problem \[problem: basic\_prob\].
Complexity {#sec:complexity}
----------
Denote by $|\varphi|$ the length of an MITL formula $\varphi$. A TBA $\mathcal{A}_i, i \in \mathcal{I}$ can be constructed in space and time $2^{\mathcal{O}(|\varphi_i|)}, i \in \mathcal{I}$. Let $\varphi_{\text{max}} = \underset{i \in \mathcal{I}}{\text{max}} \{ |\varphi_i|\}$ be the MITL formula with the longest length. Then, the complexity of Step 1 is $N 2^{\mathcal{O}(|\varphi_{\text{max}})|}$. Step 2 costs $\mathcal{O}(N 2^{|\varphi_i|}|\mathcal{S}_i|)$, where $|\mathcal{S}_i| = |\hat{\mathbb{I}}|$ is the number of states of the WTS $\mathcal{T}_i$. We have the best case complexity as $\mathcal{O}(N R_{\text{selec}} 2^{|\varphi_{\text{max}}|} |\hat{\mathbb{I}}|)$, since the Step 3 is more efficient than Step 4. The worst case complexity of our proposed framework is when Step 4 is followed, which is $\mathcal{O}(2^{|\varphi_{\text{max}}|}|\hat{\mathbb{I}}|^N)$.
[0.45]{}
[0.45]{}
[|C[0.9cm]{}||C[2.3cm]{}|C[1.6cm]{}||C[2.3cm]{}|C[1.6cm]{}|]{}\
&\
Step & Reachable States & Time & Reachable States & Time\
$\delta t$ & $2$ & $0.13 \sec$ & $24$ & $0.15 \sec$\
$2 \delta t$ & $12$ & $0.05 \sec$ & $235$ & $0.11 \sec$\
$3 \delta t$ & $18$ & $0.05 \sec$ & $1006$ & $0.88 \sec$\
$4 \delta t$ & $54$ & $1.0 \sec$ & $3104$ & $5.17 \sec$\
$5 \delta t$ & $250$ & $0.90 \sec$ & $7983$ & $20.80 \sec$\
$6 \delta t$ & $250$ & $1.27 \sec$ & $12545$ & $48.77 \sec$\
$7 \delta t$ & $250$ & $1.36 \sec$ & $13493$ & $70.21 \sec$\
$8 \delta t$ & $475$ & $1.31 \sec$ & $16078$ & $74.56 \sec$\
$9 \delta t$ & $875$ & $2.21 \sec$ & $23690$ & $107.79 \sec$\
$10 \delta t$ & $1100$ & $3.09 \sec$ & $33171$ & $185.91 \sec$\
$ $ & &\
Simulation Results {#sec: simulation_results}
==================
In order to show how the proposed framework scales with respect to the number of agents and the time solution horizon, we consider two simulation examples with two simulation scenarios each.
**Simulation Example 1 :** Consider a system of three agents with $x_i \in \mathbb{R}^2, \ i \in \mathcal{I} = \{1, 2, 3\}, \mathcal{N}(1) = \{2\} = \mathcal{N}(3), \mathcal{N}(2) = \{1, 3\}$ is considered. According to , the dynamics are given as: $\dot{x}_1 = -(x_1-x_2)+v_1, \dot{x}_2 = -(x_2-x_1)-(x_2-x_3)+v_2$ and $\dot{x}_3 = -(x_3-x_2)+v_3$. The simulation parameters are set to $L_1 = \sqrt{2}$, $L_2 = 2$ and $\delta t = 0.1$. The initial agents’ positions are set to $(-4,4), (0,6)$ and $(7,0)$ respectively. We consider Scenario 1 and Scenario 2 with $\lambda = 0.14$, $\bar{d}_{\max} = 0.25$ and $\lambda = 0.21$, $\bar{d}_{\max} = 0.20$, as is depicted in Figure \[fig:1a\] and Figure \[fig:1b\], respectively. The cell decomposition presented in this paper and the reachable cells of each agent are depicted in Figure \[fig:1a\]-\[fig:1b\]. The reachable cells of each agent are depicted with purple, cyan and yellow respectively. In Figure \[fig:1a\]-\[fig:1b\] we can observe the evolution of the reachable sets of each agent at time $t = 10 \delta t$. It can be observed that the agents are not necessarily moving between neighboring cells and not all the individual runs satisfy the desired specification. The simulation statistics are depicted in Table \[table1\]. The simulations were carried out in MATLAB Environment on a desktop with 8 cores, 3.60GHz CPU and 16GB of RAM.
**Simulation Example 2 :** Consider a multi-agent system with $x_i \in \mathbb{R}^2$, $i \in \mathcal{I} = \{1, 2, 3, 4\}$, $\mathcal{N}(1) = \{2\}, \mathcal{N}(2) = \{1,3\}$, $\mathcal{N}(3) = \{2, 4\}$, $\mathcal{N}(4) = \{3\}$. According to , the dynamics are given as: $\dot{x}_1 = -(x_1-x_2)+v_1, \dot{x}_2 = -(x_2-x_1)-(x_2-x_3)+v_2$, $\dot{x}_3 = -(x_3-x_2)-(x_2-x_4)+v_3$ and $\dot{x}_4 = -(x_4-x_3)+v_4$. The simulation parameters are set to $L_1 = \sqrt{2}$, $L_2 = 2$, $\delta t = 0.1$. The workspace is decomposed into square cells, which are depicted with blue color in Figure \[fig:simulation2\]. The initial agents’ positions are set to $(-4,4)$, $(0,6)$, $(7,0)$ and $(4,-5)$, respectively. We consider Scenario 1 and Scenario 2 with $\lambda = 0.14$, $\bar{d}_{\max} = 0.25$ and $\lambda = 0.21$, $\bar{d}_{\max} = 0.20$, as is depicted in Figure \[fig:1a\] and Figure \[fig:1b\], respectively. The specification formulas for the Scenario $1$ are set to $\varphi_1 = \Diamond_{[0.2, 1.0]} \{\rm green\}$, $\varphi_2 = \Diamond_{[0.1, 1.0]} \{\rm orange\}$, $\varphi_3 = \Diamond_{[0.5, 1.5]} \{\rm black\}$ and $\varphi_4 = \Diamond_{[0.3, 2.3]} \{\rm gray\}$, respectively. The cell decomposition presented in this paper, the reachable cells of each agent up to time $t = 10 \delta t$ and the goal regions are depicted in Figure \[fig:simulation2\]. The reachable cells of each agent are depicted with purple, cyan and yellow respectively. Note that the agents’ transitions are not necessarily performed between neighboring cells. The individual consistent runs $\widetilde{r}_1^t$, $\widetilde{r}_2^t$, $\widetilde{r}_3^t$ and $\widetilde{r}_4^t$ of agents $1$, $2$, $3$ and $4$ that satisfy the formulas $\varphi_1$, $\varphi_2$, $\varphi_3$ and $\varphi_4$, respectively are depicted in Figure \[fig:simulation2\] with black arrows. Each arrow represents a transition from a state to another according to Def. \[def: indiv\_WTS\]. The product WTS $\mathcal{T}_p$ has $52877$ reachable states in case of $\lambda = 0.14$ and $1255547$ when $\lambda = 0.21$ since, according to Section 4.2.2, larger values of $\lambda$ lead to finer discretization. Agent $1$ satisfies $\varphi_1$ in $5 \delta t$, agent $2$ satisfies $\varphi_2$ in $2 \delta t$, agent $3$ satisfies $\varphi_3$ in $8 \delta t$ and agent $4$ satisfies $\varphi_4$ in $10 \delta t$. The simulation is performed in a horizon of $10$ steps and it takes $668.80 \sec$ ($609.5 \sec$ for the abstraction and $59.30 \sec$ for the graph search) for Scenario 1.
[0.45]{}
[0.45]{}
[|C[0.9cm]{}||C[2.3cm]{}|C[1.6cm]{}||C[2.3cm]{}|C[2.1cm]{}|]{}\
&\
Step & Reachable States & Time & Reachable States & Time\
$\delta t$ & $8$ & $0.22 \sec$ & $24$ & $0.32 \sec$\
$2 \delta t$ & $39$ & $0.11 \sec$ & $208$ & $0.24 \sec$\
$3 \delta t$ & $216$ & $0.16 \sec$ & $6702$ & $5.92 \sec$\
$4 \delta t$ & $1610$ & $17.97 \sec$ & $26843$ & $74.95 \sec$\
$5 \delta t$ & $3168$ & $16.73 \sec$ & $89817$ & $91.68 \sec$\
$6 \delta t$ & $5346$ & $21.87 \sec$ & $133904$ & $252.12 \sec$\
$7 \delta t$ & $5808$ & $31.55 \sec$ & $222037$ & $451.61 \sec$\
$8 \delta t$ & $10168$ & $37.39 \sec$ & $358941$ & $1427.65 \sec$\
$9 \delta t$ & $23004$ & $93.98 \sec$ & $644489$ & $4803.12 \sec$\
$10 \delta t$ & $52877$ & $389.50 \sec$ & $1255547$ & $19321.07 \sec$\
$ $ & &\
Conclusions and Future Work {#sec: conclusions}
===========================
A systematic method for controller synthesis of dynamically coupled multi-agent path-planning has been proposed, in which timed constraints of fulfilling a high-level specification are imposed to the system. The solution involves a boundedness analysis, the abstraction of each agent’s motion into WTSs, TBAs as well as Büchi WTSs construction. The simulation example demonstrates our solution approach. Future work includes further computational improvement of the abstraction method and more complicated high-level tasks being imposed to the agents in order to exploit the expressiveness of MITL formulas.
Proof of Theorem 1 {#app:proof_theorem_1}
==================
Consider the following candidate Lyapunov function $V: \mathbb{R}^{Nn} \rightarrow \mathbb{R}_{\ge 0}$ $$V(x) = \frac{1}{2} \sum_{i=1}^{N} \sum_{j \in \mathcal{N}(i)}^{} \| x_i-x_j\|^2 = \|\widetilde{x}\|^2 > 0.$$ The time derivative of $V$ along the trajectories of , can be computed as $$\begin{aligned}
&\hspace{-1mm}\dot{V} = \left[ \nabla V(x) \right]^{\top} \ \dot{x} \nonumber \\
&\hspace{-1mm}= \sum_{k=1}^{n} \left\{ c\left(\frac{\partial V}{\partial x} ,k\right)^{\top} \left[ -L(\mathcal{G}) \ c(x,k)+c(v, k) \right] \right\}, \label{lyap1}\end{aligned}$$ where $\displaystyle c\left(\frac{\partial V}{\partial x} ,k \right) = \begin{bmatrix} \frac{\partial V}{\partial x_1^k} & \dots & \frac{\partial V}{\partial x_N^k} \end{bmatrix}^{\top}$. By computing the partial derivative of the Lyapunov function with respect to vector $x_i, \ i \in \mathcal{I}$ we get $\displaystyle \frac{\partial V}{\partial x_i} = \sum_{j \in \mathcal{N}(i)}^{} (x_i-x_j), \ i \in \mathcal{I}$ from which we have that $\displaystyle c\left(\frac{\partial V}{\partial x} ,k\right)^{\top} = c(x,k)^{\top} \ L(\mathcal{G}), \ k = 1,...,n$. Thus, by substituting the last in we get $$\begin{aligned}
\dot{V} &= -\sum_{k=1}^{n} \left\{ c(x,k)^{\top} \ \left[ L(\mathcal{G})\right]^2 \ c(x,k) \right\} + \sum_{k=1}^{n} \left\{ c(x,k)^{\top} \ \left[ L(\mathcal{G})\right]^2 c(v, k) \right\} \nonumber \\
&\leq -\sum_{k=1}^{n} \left\{ c(x,k)^{\top} \ \left[ L(\mathcal{G})\right]^2 \ c(x,k)) \right\} + \left\| \sum_{k=1}^{n} \left\{ c(x,k)^{\top} \ L(\mathcal{G}) \ c(v, k) \right\} \right\|. \label{eq: lyap3}\end{aligned}$$
For the first term of we have that $$\begin{aligned}
\sum_{k=1}^{n} \left\{ c(x,k)^{\top} \ L(\mathcal{G})^2 \ c(x,k)) \right\} = \sum_{k=1}^{n} \left\| L(\mathcal{G}) \ c(x,k) \right\|^2. \notag\end{aligned}$$
For the second term of we get $$\begin{aligned}
\left\| \sum_{k=1}^{n} \left\{ c(x,k)^{\top} \ L(\mathcal{G}) \ c(\nu,k) \right\} \right\| & = \left\| \sum_{k=1}^{n} \left\{ c(x,k)^{\top} \ D(\mathcal{G}) \ D(\mathcal{G})^{\top} \ c(\nu,k) \right\} \right\| \notag \\
& \leq \sum_{k=1}^{n} \left\{ \left\| D(\mathcal{G}) ^{\top} \ c(x,k) \right\| \ \left\| D(\mathcal{G})^{\top} \right\| \ \left\| c(\nu,k) \right\|\right\} \notag \\
& = \left\| D(\mathcal{G})^{\top} \right\| \ \sum_{k=1}^{n} \left\{ \left\| c(\widetilde{x},k) \right\| \ \left\| c(\nu,k) \right\| \right\}. \label{eq: lyap4}\end{aligned}$$ By using the Cauchy-Schwarz inequality in we get $$\begin{aligned}
\left\| \sum_{k=1}^{n} \left\{ c(x,k)^{\top} \ L(\mathcal{G}) \ c(v, k) \right\} \right\| & \leq \left\| D(\mathcal{G})^{\top} \right\| \ \left( \sum_{k=1}^{n} \left\| c(\widetilde{x},k) \right\|^2 \right)^{\frac{1}{2}} \ \left( \sum_{k=1}^{n} \left\| c(v, k) \right\|^2 \right)^{\frac{1}{2}} \notag \\
& = \left\| D(\mathcal{G})^{\top} \right\| \|\widetilde{x}\| \|v\| \leq \left\| D(\mathcal{G})^{\top} \right\| \|\widetilde{x}\| \sqrt{N} \|v\|_\infty, \notag\end{aligned}$$
where $\|v\|_\infty = \text{max} \left\{\|v_i\| : i \in \mathcal{I} \right\} \leq v_{\text{max}}$. Thus, by combining the previous inequalities, is written $$\dot{V} \leq -\sum_{k=1}^{n} \left\{ \left\| L(\mathcal{G}) c(x,k) \right\|^2 \right\} + \sqrt{N} \left\| D(\mathcal{G})^{\top} \right\| \|\widetilde{x}\| v_{\text{max}}. \label{eq: lyap8}$$ By exploiting Lemma \[lemma: lemma\_1\], is written as: $$\begin{aligned}
\dot{V} &\leq -\lambda_2^2(\mathcal{G}) \sum_{k=1}^{n} \left\{\left\| c(x^{\perp},k) \right\|^2 \right\} + \sqrt{N} \left\| D(\mathcal{G})^{\top} \right\| \|\widetilde{x}\| v_{\text{max}} \notag \\
&= -\lambda_2^2(\mathcal{G}) \ \|x^\perp\|^2 + \sqrt{N} \left\| D(\mathcal{G})^{\top} \right\| \|\widetilde{x}\| v_{\text{max}} \notag \\
&\leq -\frac{\lambda_2^2(\mathcal{G})}{2(N-1)}\|\widetilde{x}\|^2 + \sqrt{N} \left\| D(\mathcal{G})^{\top} \right\| \|\widetilde{x}\| v_{\text{max}} \notag \\
&\leq -K_1 \|\widetilde{x}\| \left(\|\widetilde{x}\|-K_2 v_{\text{max}} \right). \label{eq:concl_1}\end{aligned}$$ where $K_1 = \frac{\lambda_2^2(\mathcal{G})}{2(N-1)} > 0$. By using the following implication $\widetilde{x} = D^{\top}(\mathcal{G})x \Rightarrow \|\widetilde{x}\| = \|D(\mathcal{G})^\top x\| \leq \|D(\mathcal{G})^\top\| \|x\|$, apparently, we have that $0 < V(x) = \|\widetilde{x}\|^2 \leq \|D(\mathcal{G})^\top\|^2 \|x\|^2 \ \text{and} \ \dot{V}(x) < 0$ when $\|\widetilde{x}\| \geq \bar{R} > K_2 v_{\text{max}} $. Thus, there exists a finite time $T > 0$ such that the trajectory will enter the compact set $\mathcal{X} = \{x \in \mathbb{R}^{Nn} : \|\widetilde{x}\| \leq \bar{R}\}$ and remain there for all $t \geq T$ with $\bar{R} > K_2 v_{\text{max}}$. This can be extracted from the following. Let us define the compact set $$\Omega = \left\{ x \in \mathbb{R}^{Nn} : K_2 v_{\text{max}} < \bar{R} \leq \|\widetilde{x}\| \leq \bar{M} \right\},$$ where $\bar{M} = V(x(0)) = \|\widetilde{x}(0)\|^2$. Without loss of generality it is assumed that it holds $\bar{M} > \bar{R}$. Let us define the compact sets $\mathcal{S}_{1} = \left\{ x \in \mathbb{R}^{Nn} : \|\widetilde{x}\| \leq \bar{M} \right\}, \mathcal{S}_{2} = \left\{ x \in \mathbb{R}^{Nn} : \|\widetilde{x}\| \leq K_2 v_{\text{max}} \right\}$. From the equivalences $\forall \ x \in S_1 \Leftrightarrow V(x) = \|\widetilde{x}\|^2 \leq \bar{M}^2, \forall \ x \in S_2 \Leftrightarrow V(x) = \|x\|^2 \leq K^2_2 v^2_{\text{max}}$, we have that the boundaries $\partial S_1, \partial S_2$ of sets $S_1, S_2$ respectively, are two level sets of the Lyapunov function $V$. By taking the above into consideration we have that $\partial S_2 \subsetneq \partial S_1$. Hence, we get from that: $$\dot{V}(x) < 0, \forall \ x \in \Omega = S_1 \backslash S_2, \label{eq:concl_2}$$ In view of and the fact that the sets $\mathcal{S}_1, \mathcal{S}_2$ are defined in terms of level sets of $V$, we conclude that both $S_1$ and $S_2$ are invariant with respect to the system . Consequently, according to [@liberzon_switching Lemma 5.1] the trajectory that starts inside the set $S_1$ has to enter the interior of the set of $S_2$ in finite time $T > 0$ and remain there for all time $t \geq T$.
Proof of Lemma 2
================
\[lemma: lemma\_1\] Let $x^\perp$ be the projection of the vector $x \in \mathbb{R}^{Nn}$ to the orthogonal complement of the subspace $H = \{x \in \mathbb{R}^{Nn}: x_1 = \ldots = x_N\}$. Then, the following properties hold: $$\begin{aligned}
\| L(\mathcal{G}) \ c(x, k) \| &\geq \lambda_2 (\mathcal{G}) \ \|c(x^{\perp}, k) \|, \ \forall \ k \in \mathcal{I}, \\
\|x^\perp\| &\geq \frac{1}{\sqrt{2(N-1)}} \|\widetilde{x}\|.
\end{aligned}$$
The proof can be found in [@boskos_cdc_connectivity Appendix A].
Sufficient Conditions for Well-Possessedness of the Abstraction {#app:suff_conditions}
===============================================================
We investigate here if the system satisfies the sufficient conditions **(C1)**-**(C3)** for well-posed abstractions.
**(C1)** For every $i \in \mathcal{I}, \forall \ x \in \mathbb{R}^{Nn} : \widetilde{x} \in \mathcal{X}$ and $\text{pr}_i(x) = (x_i, \mathbf{x}_j)$ it holds that: $$\begin{aligned}
\| f_i(x_i, \mathbf{x}_j) \| &= \left\| -\sum_{j \in \mathcal{N}(i)}^{}(x_i-x_j) \right\| \leq \sum_{j \in \mathcal{N}(i)}^{} \| x_i-x_j \| \notag \\
&\leq \sum_{ (i,j) \in \mathcal{E} }^{} \| x_i-x_j \| = \Delta x \leq \bar{R}.\end{aligned}$$ Thus, $M = \bar{R}$. We have also that $\|D(\mathcal{G})^\top\| = \sqrt{\lambda_{\text{max}} ( D(\mathcal{G}) D(\mathcal{G})^\top)} = \sqrt{\lambda_{\text{max}}(\mathcal{G})}$ and $\lambda_2(\mathcal{G}) \leq \frac{N}{N-1} \min \{N_i : i \in \mathcal{I} \}$ from [@fiedler1973algebraic]. For $N > 2$ it holds that $\lambda_2(\mathcal{G}) < N$. From Theorem we have that $\bar{R} > K_2 v_{\text{max}} \Leftrightarrow M > K_2 v_{\text{max}}$. It holds that $M > v_{\text{max}}$ since
$$\begin{aligned}
K_2 &= \frac{2 \sqrt{N} (N-1) \left\| D(\mathcal{G})^{\top} \right\|}{\lambda_2^2(\mathcal{G})} = \frac{2 \sqrt{N} (N-1) \sqrt{\lambda_{\text{max}}(\mathcal{G})}}{\sqrt{\lambda_2^3(\mathcal{G})} \sqrt{\lambda_2(\mathcal{G})}} \notag \\
&\geq \frac{2 \sqrt{N} (N-1)}{\sqrt{N^3}} \sqrt{\frac{\lambda_{\text{max}}(\mathcal{G})}{\lambda_2(\mathcal{G})}} \geq \frac{2 \sqrt{N} (N-1)}{\sqrt{N^3}} > 1. \notag\end{aligned}$$
**(C2)** Starting from the left hand side of we get: $$\begin{aligned}
\| f_i(x_i, \mathbf{x}_j) - f_i(x_i, \mathbf{y}_j) \| & = \Big\| -\sum_{j \in \mathcal{N}(i)}^{}(x_i-x_j) + \sum_{j \in \mathcal{N}(i)}^{}(x_i-y_j) \Big\| \notag \\
& \leq \text{max}\{\sqrt{N_i} : i \in \mathcal{I}\} \ \|(x_i, \mathbf{x}_j) - (x_i, \mathbf{y}_j) \|. \notag\end{aligned}$$ Thus, the condition **(C2)** holds and the Lipschitz constant is $L_1 = \text{max}\{\sqrt{N_i} : i \in \mathcal{I}\} >0$, where the inequality $\displaystyle \left( \sum_{i = 1}^{\rho} \alpha_i \right)^2 \leq \rho \ \left( \sum_{i=1}^\rho \alpha_i^2 \right)$ is used.\
**(C3)** By using the same methodology as in **(C2)**, we conclude that $L_2 = \text{max} \{N_i : i \in \mathcal{I}\} > 0$.
[^1]: This work was supported by the H2020 ERC Starting Grant BUCOPHSYS, the EU H2020 Research and Innovation Programme under GA No. 731869 (Co4Robots), the SSF COIN project, the Swedish Research Council (VR) and the Knut och Alice Wallenberg Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
This paper aims at setting out the basics of ${\ensuremath{\mathbb{Z}}}$-graded manifolds theory. We introduce ${\ensuremath{\mathbb{Z}}}$-graded manifolds from local models and give some of their properties. The requirement to work with a completed graded symmetric algebra to define functions is made clear. Moreover, we define vector fields and exhibit their graded local basis. The paper also reviews some correspondences between differential ${\ensuremath{\mathbb{Z}}}$-graded manifolds and algebraic structures.
**Keywords.** Supergeometry, graded manifold, differential graded manifold, $Q$-manifold.
**2010 MSC.** 58A50, 51-02
address: |
School of Mathematics\
University of Leeds\
Leeds LS2 9JT\
United Kingdom
author:
- Maxime Fairon
bibliography:
- 'ITGG.bib'
title: Introduction to graded geometry
---
Introduction {#SectIntro}
============
In the $70$s, supermanifolds were introduced and studied to provide a geometric background to the developing theory of supersymmetry. In the Berezin-Leites [@BerLei] and Kostant [@Kost77] approach, they are defined as ${\ensuremath{\mathbb{Z}}}_2$-graded locally ringed spaces. More precisely, this means that a supermanifold is a pair $(|N|,\mathcal{O}_N)$ such that $|N|$ is a topological space and $\mathcal{O}_N$ is a sheaf of ${\ensuremath{\mathbb{Z}}}_2$-graded algebras which satisfies, for all sufficiently small open subsets $U$, $$\mathcal{O}_N(U) \simeq \mathcal{C}^\infty_{{\ensuremath{\mathbb{R}}}^n}(U)\otimes \bigwedge {\ensuremath{\mathbb{R}}}^m\,\,, \nonumber$$ where $\bigwedge {\ensuremath{\mathbb{R}}}^m$ is endowed with its canonical ${\ensuremath{\mathbb{Z}}}_2$-grading. Accordingly, a local coordinate system on a supermanifold splits into $n$ smooth coordinates on $U$ and $m$ elements of a basis on ${\ensuremath{\mathbb{R}}}^m$. Both types of coordinates are distinguished via the parity function $p$ with values in ${\ensuremath{\mathbb{Z}}}_2=\{0,1\}$. Namely, the first $n$ coordinates are said to be even ($p=0$), while the others are odd ($p=1$).
From the late $90$s, the introduction of an integer grading was necessary in some topics related to Poisson geometry, Lie algebroids and Courant algebroids. These structures could carry the integer grading only, as it appeared in the work of Kontsevitch [@Konts03], Roytenberg [@Roy02] and Severa [@SeveraHomotopy], or could be endowed with both ${\ensuremath{\mathbb{Z}}}_2$- and ${\ensuremath{\mathbb{Z}}}$-gradings, as it was introduced by Voronov [@Voronov02]. The first approach, that we follow, led to the definition of a ${\ensuremath{\mathbb{Z}}}$-graded manifold given by Mehta [@ThesisPing]. Similarly to a supermanifold, a ${\ensuremath{\mathbb{Z}}}$-graded manifold is a graded locally ringed space. Its structure sheaf $\mathcal{O}_N$ is a sheaf of ${\ensuremath{\mathbb{Z}}}$-graded algebras which is given locally by $$\mathcal{O}_N(U) \simeq \mathcal{C}^\infty_{{\ensuremath{\mathbb{R}}}^n}(U)\otimes
\overline{\mathbf{S}\mathcal{W}}\,\,, \nonumber$$ where $\mathcal{W}=\oplus_i \mathcal{W}_i$ is a real ${\ensuremath{\mathbb{Z}}}$-graded vector space whose component of degree zero satisfies $\mathcal{W}_0=\{0\}$. Here, $\overline{\mathbf{S}\mathcal{W}}$ denotes the algebra of formal power series of supercommutative products of elements in $\mathcal{W}$. Therefore, the local coordinate system of the ${\ensuremath{\mathbb{Z}}}$-graded manifold can be described by $n$ smooth coordinates and a graded basis of $\mathcal{W}$. In particular, it might admit generators of even degree but not of smooth type. This implies that such coordinates do not square to zero and can generate polynomials of arbitrary orders. Their existence requires the introduction of formal power series to obtain the locality of every stalk of the sheaf. This condition, which has not always been accurately considered in past works, also occurs with ${\ensuremath{\mathbb{Z}}}^n_2$-grading for $n\geqslant 2$, as discussed in [@Covolo].
Complete references on supermanifolds can be found (*e.g.* [@CCFsupersym; @DelMor99; @Kost77; @Varadarajansupersym]), but mathematicians have not yet prepared monographs on graded manifolds. The aim of this paper is to show what would appear at the beginning of such a book and it aspires to be a comprehensive introduction to ${\ensuremath{\mathbb{Z}}}$-graded manifolds theory for both mathematicians and theoretical physicists.
The layout of this article is inspired by [@CCFsupersym; @Varadarajansupersym] and the reader needs a little knowledge of sheaf theory from \[SubSectGradedRingSp\] onward. We first introduce graded vector spaces, graded rings and graded algebras in \[SectBasis\]. These objects allow us to define graded locally ringed spaces and their morphisms. After that, we introduce graded domains and we show that their stalks are local. In particular, we notice that this property follows from the introduction of formal power series to define sections. The graded domains are the local models of graded manifolds, which are studied in \[SectionGradedManifold\]. Specifically, we give the elementary properties of graded manifolds and define their vector fields. In particular, we prove that there exists a local graded basis of the vector fields which is related to the local coordinate system of the graded manifold. Finally, we illustrate in \[SectionApplication\] the theory of ${\ensuremath{\mathbb{Z}}}$-graded manifolds with a few theorems stating the correspondence between differential graded manifolds and algebraic structures. These important examples are usually referred to as $Q$-manifolds. Other applications can be found in [@CattaneoSchatz; @Qiu2011].
**Conventions.** In this paper, ${\ensuremath{\mathbb{N}}}$ and ${\ensuremath{\mathbb{Z}}}$ denote the set of nonnegative integers and the set of integers, respectively. We write ${\ensuremath{\mathbb{N}}}^\times$ and ${\ensuremath{\mathbb{Z}}}^\times$ when we consider these sets deprived of zero. Notice that some authors use the expression *graded manifolds* to talk about supermanifolds with an additional ${\ensuremath{\mathbb{Z}}}$-grading (see [@Voronov02]), but we only use this expression in the present paper to refer to ${\ensuremath{\mathbb{Z}}}$-graded manifolds.
Preliminaries {#SectBasis}
=============
Graded algebraic structures {#SubSectGradedAlgebra}
---------------------------
### Graded vector space
A *${\ensuremath{\mathbb{Z}}}$-graded vector space* is a direct sum $\mathcal{V}=\bigoplus_{i}\mathcal{V}_i$ of a collection of ${\ensuremath{\mathbb{R}}}$-vector spaces $(\mathcal{V}_i)_{i\in{\ensuremath{\mathbb{Z}}}}$. If a nonzero element $v\in \mathcal{V}$ belongs to one of the $\mathcal{V}_i$, one says that it is *homogeneous* of degree $i$. We write $|\cdot|$ for the application which assigns its degree to a homogeneous element. Moreover, we only consider ${\ensuremath{\mathbb{Z}}}$-graded vector spaces of *finite type*, which means that $\mathcal{V}=\bigoplus_{i}\mathcal{V}_i$ is such that $\text{dim}(\mathcal{V}_i)<\infty$ for all $i\in {\ensuremath{\mathbb{Z}}}$. A *graded basis* of $\mathcal{V}$ is a sequence $(v_\alpha)_\alpha$ of homogeneous elements of $\mathcal{V}$ such that the subsequence of all elements $v_\alpha$ of degree $i$ is a basis of the vector space $\mathcal{V}_i$, for all $i\in{\ensuremath{\mathbb{Z}}}$.
If $\mathcal{V}$ is a ${\ensuremath{\mathbb{Z}}}$-graded vector space, $\mathcal{V}[k]$ denotes the ${\ensuremath{\mathbb{Z}}}$-graded vector space $\mathcal{V}$ lifted by $k\in{\ensuremath{\mathbb{Z}}}$ : $(\mathcal{V}[k])_i=\mathcal{V}_{i-k}$ for all $i\in {\ensuremath{\mathbb{Z}}}$. The ${\ensuremath{\mathbb{Z}}}$-graded vector space $\mathcal{V}$ with reversed degree is denoted by $\Pi\mathcal{V}$ and satisfies $(\Pi\mathcal{V})_i=\mathcal{V}_{-i}$ for all $i \in {\ensuremath{\mathbb{Z}}}$.
Given two ${\ensuremath{\mathbb{Z}}}$-graded vector spaces $\mathcal{V}$ and $\mathcal{W}$, their direct sum $\mathcal{V}\oplus \mathcal{W}$ can be defined with the grading $(\mathcal{V}\oplus \mathcal{W})_i=\mathcal{V}_i\oplus \mathcal{W}_i$, as well as the tensor product $\mathcal{V}\otimes \mathcal{W}$ with $(\mathcal{V}\otimes \mathcal{W})_i = \oplus_{j\in{\ensuremath{\mathbb{Z}}}} \mathcal{V}_j \otimes \mathcal{W}_{i-j}$. Both constructions are associative. A *morphism of ${\ensuremath{\mathbb{Z}}}$-graded vector spaces* $T: \mathcal{V} \rightarrow \mathcal{W}$ is a linear map which preserves the degree : $T(\mathcal{V}_i) \subseteq \mathcal{W}_i$ for all $i\in{\ensuremath{\mathbb{Z}}}$. We write $\text{Hom}(\mathcal{V},\mathcal{W})$ for the set of all morphisms between $\mathcal{V}$ and $\mathcal{W}$.
The category of ${\ensuremath{\mathbb{Z}}}$-graded vector spaces is a symmetric monoidal category. It is equipped with the nontrivial commutativity isomorphism $\mathbf{c}_{\mathcal{V},\mathcal{W}}:\mathcal{V}\otimes \mathcal{W}\to\mathcal{W}\otimes\mathcal{V}$ which, to any homogeneous elements $v \in \mathcal{V}$ and $w \in \mathcal{W}$, assigns the homogeneous elements $\mathbf{c}_{\mathcal{V},\mathcal{W}}(v\otimes w) = (-1)^{|v||w|} w\otimes v$.
Finally, duals of ${\ensuremath{\mathbb{Z}}}$-graded vector spaces can be defined. Given a ${\ensuremath{\mathbb{Z}}}$-graded vector space $\mathcal{V}$, $(\mathcal{V},{\ensuremath{\mathbb{R}}})$ denotes the ${\ensuremath{\mathbb{Z}}}$-graded vector space which contains all the linear maps from $\mathcal{V}$ to ${\ensuremath{\mathbb{R}}}$. Its grading is defined by setting $\left(\text{\underline{Hom}}(\mathcal{V},{\ensuremath{\mathbb{R}}})\right)_i$ to be the set of linear maps $f$ such that $f(v)\in {\ensuremath{\mathbb{R}}}$ if $v\in \mathcal{V}_{-i}$. It is the dual of $\mathcal{V}$ and we write $\mathcal{V}^*=$$(\mathcal{V},{\ensuremath{\mathbb{R}}})$. One can show that the latter satisfies $(\mathcal{V}^*)_i=(\mathcal{V}_{-i})^*$, for all $i\in{\ensuremath{\mathbb{Z}}}$.
From now on, we will refer to ${\ensuremath{\mathbb{Z}}}$-graded objects simply as graded objects. However, we will keep the complete writing in the definitions or when it is needed, to keep it clear that the grading is taken over ${\ensuremath{\mathbb{Z}}}$.
The different notions introduced for ${\ensuremath{\mathbb{Z}}}$-graded vector spaces extend to ${\ensuremath{\mathbb{Z}}}$-graded $R$-modules over a ring $R$. In particular, taking $R={\ensuremath{\mathbb{Z}}}$, these constructions apply to any ${\ensuremath{\mathbb{Z}}}$-graded abelian group (which is a direct sum of abelian groups) seen as a direct sum of ${\ensuremath{\mathbb{Z}}}$-modules.
### Graded ring {#sSubSectRings}
A *${\ensuremath{\mathbb{Z}}}$-graded ring* $\mathcal{R}$ is a ${\ensuremath{\mathbb{Z}}}$-graded abelian group $\mathcal{R}=\bigoplus_{i}\mathcal{R}_i$ with a morphism $\mathcal{R}\otimes \mathcal{R}\to \mathcal{R}$ called the *multiplication*. By definition, it satisfies $\mathcal{R}_i \mathcal{R}_j \subseteq \mathcal{R}_{i+j}$. A graded ring $\mathcal{R}$ is *unital* if it admits an element $1$ such that $1\,r=r=r\, 1$ for any $r\in \mathcal{R}$. In that case, the element $1$ satisfies $1\in \mathcal{R}_0$. The graded ring $\mathcal{R}$ is *associative* if $(ab)c=a(bc)$ for every $a,b,c\in \mathcal{R}$. The graded ring $\mathcal{R}$ is *supercommutative* when $ab=(-1)^{|a||b|}ba$ for any homogeneous elements $a,b\in \mathcal{R}$. This means that the multiplication is invariant under the commutativity isomorphism $\mathbf{c}$.
We can introduce the definitions of left ideal, right ideal and two-sided ideal (which is referred to as *ideal*) of a graded ring in the same manner as in the non-graded case. It is easy to see that, in a supercommutative associative unital graded ring, a left (or right) ideal is an ideal. A *homogeneous ideal* of $\mathcal{R}$ is an ideal $I$ such that $I=\bigoplus_kI_k$, for $I_k=I\cap \mathcal{R}_k$. Equivalently, a homogeneous ideal $I$ is an ideal generated by a set of homogeneous elements $H\subseteq \cup_k \mathcal{R}_k$. In that case, we write $I=\langle H \rangle$ when we want to emphasize the generating set of $I$. An homogeneous ideal $I\subsetneq \mathcal{R}$ is said to be *maximal* if, when $I\subseteq J$ with $J$ another homogeneous ideal, then either $J=I$ or $J=\mathcal{R}$. A *local* graded ring $\mathcal{R}$ is a graded ring which admits a unique maximal homogeneous ideal.
Define $\mathcal{J}_\mathcal{R}$ as the ideal generated by the elements of $\mathcal{R}$ with nonzero degree. We can consider the quotient $\mathcal{R}/\mathcal{J}_\mathcal{R}$ and the projection map $\pi : \mathcal{R} \to \mathcal{R}/\mathcal{J}_\mathcal{R}$. We say that $\mathcal{R}$ is a *$\pi$-local graded ring* if it admits a unique maximal homogeneous ideal $\mathfrak{m}$ such that $\pi(\mathfrak{m})$ is the unique maximal ideal of $\mathcal{R}/\mathcal{J}_\mathcal{R}$. Note that a local graded ring is always $\pi$-local, but the converse is not true (see Remark \[RemLocality\] below).
The different notions introduced for ${\ensuremath{\mathbb{Z}}}$-graded rings extend to any object with an underlying ${\ensuremath{\mathbb{Z}}}$-graded ring structure.
### Graded algebra
A ${\ensuremath{\mathbb{Z}}}$-*graded algebra* is a ${\ensuremath{\mathbb{Z}}}$-graded vector space $\mathcal{A}$ endowed with a morphism $\mathcal{A}\otimes \mathcal{A}\to \mathcal{A}$ called *multiplication*. Alternatively, it is a graded ring with a structure of ${\ensuremath{\mathbb{R}}}$-module.
Consider a ${\ensuremath{\mathbb{Z}}}$-graded vector space $\mathcal{W}$. The *free ${\ensuremath{\mathbb{Z}}}$-graded associative algebra* generated by $\mathcal{W}$ is $\bigotimes \mathcal{W} = \bigoplus_{k\in {\ensuremath{\mathbb{N}}}}\otimes^k \mathcal{W}$. The product on $\bigotimes \mathcal{W}$ is the concatenation, while its degree is induced by the degree on $\mathcal{W}$. Then, the *symmetric ${\ensuremath{\mathbb{Z}}}$-graded associative algebra* generated by $\mathcal{W}$ is given by $\mathbf{S} \mathcal{W}=\bigotimes \mathcal{W} / J$, where $J= \left\langle\{ u\otimes v -(-1)^{|u||v|} v \otimes u \, \, |\,\, u,v\in \mathcal{W} \, \text{homogeneous} \}\right\rangle$.
This object gathers the non-graded exterior and symmetric algebras. Indeed, split $\mathcal{W}$ as $\mathcal{W}=\mathcal{W}_{\text{even}}\oplus\mathcal{W}_{\text{odd}}$ with $\mathcal{W}_{\text{even}}=\oplus_i \mathcal{W}_{2i}$ and $\mathcal{W}_{\text{odd}}=\oplus_i \mathcal{W}_{2i+1}$. Then we obtain that $\mathbf{S}\mathcal{W}\simeq \mathbf{S}\mathcal{W}_{\text{even}} \otimes \bigwedge \mathcal{W}_{\text{odd}}$ as ${\ensuremath{\mathbb{Z}}}_2$-graded algebras (with the two operations on the right considered on non-graded vector spaces).
The graded algebra $\mathbf{S} \mathcal{W}$ can be seen as the set of polynomials in a graded basis of the graded vector space $\mathcal{W}$. We write $\overline{\mathbf{S} \mathcal{W}}$ to indicate its completion by allowing formal power series. Assume that $(w_\alpha)_\alpha$ is a graded basis of $\mathcal{W}$. Then, any element $s\in \overline{\mathbf{S} \mathcal{W}}$ admits a unique decomposition with respect to the $(w_\alpha)_\alpha$ : $$\label{EqDecompoSW}
s=s_0 + \sum_{K=1}^\infty \sum_{\alpha_1\leqslant \ldots \leqslant \alpha_K}
s_{\alpha_1\ldots \alpha_K} w_{\alpha_1}\ldots w_{\alpha_K},$$ where $s_0$ and all the $s_{\alpha_1\ldots \alpha_K}$ are real numbers. We use the multiplicative notation $w_{\alpha_1}\ldots w_{\alpha_K}$ to denote $w_{\alpha_1} \otimes \ldots \otimes w_{\alpha_K}$ modulo the supercommutativity in $\overline{\mathbf{S}\mathcal{W}}$.
Graded ringed space {#SubSectGradedRingSp}
-------------------
A *${\ensuremath{\mathbb{Z}}}$-graded ringed space* $S$ is a pair $(|S|,\mathcal{O}_S)$ such that $|S|$ is a topological space and $\mathcal{O}_S$ is a sheaf of associative unital supercommutative ${\ensuremath{\mathbb{Z}}}$-graded rings, called the *structure sheaf* of $S$. A *${\ensuremath{\mathbb{Z}}}$-graded locally ringed space* is a ${\ensuremath{\mathbb{Z}}}$-graded ringed space $S=(|S|,\mathcal{O}_S)$ whose stalks $\mathcal{O}_{S,x}$ are local graded rings for all $x \in |S|$.
Any locally ringed space is a graded locally ringed space whose sheaf has only elements of degree zero.
A *morphism of ${\ensuremath{\mathbb{Z}}}$-graded ringed spaces* $\phi:(|M|,\mathcal{F})\rightarrow(|N|,\mathcal{G})$ is a pair $(|\phi|,\phi^*)$, where $|\phi|:|M| \rightarrow |N|$ is a morphism of topological spaces and $\phi^* : \mathcal{G}\rightarrow \phi_* \mathcal{F}$ is a morphism of sheaves of associative unital supercommutative ${\ensuremath{\mathbb{Z}}}$-graded rings, which means that it is a collection of morphisms $\phi_V :\mathcal{G}(V)\rightarrow \mathcal{F}(|\phi|^{-1}(V))$ for all $V \in |N|$.
A *morphism of ${\ensuremath{\mathbb{Z}}}$-graded locally ringed spaces* $\phi:(|M|,\mathcal{F})\rightarrow(|N|,\mathcal{G})$ is a morphism of ${\ensuremath{\mathbb{Z}}}$-graded ringed spaces such that, for all $x\in |M|$, the induced morphism on the stalk $\phi_x :\mathcal{G}_{|\phi|(x)}\rightarrow \mathcal{F}_x$ is local, which means that $\phi_x^{-1}\left( \mathfrak{m}_{_{M,x}} \right) = \mathfrak{m}_{_{N,|\phi|(x)}}$, where $\mathfrak{m}_{_{M,x}}$ (respectively $\mathfrak{m}_{_{N,|\phi|(x)}}$) is the maximal homogeneous ideal of $\mathcal{F}_x$ (resp. $\mathcal{G}_{|\phi|(x)}$).
Let $(|M|,\mathcal{F})$ and $(|N|,\mathcal{G})$ be two graded locally ringed spaces. Assume that for all $x\in |M|$, there exists an open set $V\subseteq|M|$ containing $x$ and an open set $\widetilde{V}\subseteq |N|$ such that there exists an isomorphism of graded locally ringed spaces $\phi_V:(V,\restriction{\mathcal{F}}{V})\to(\widetilde{V},\restriction{\mathcal{G}}{\widetilde{V}})$. Then, we say that the graded locally ringed space $(|M|,\mathcal{F})$ is *locally isomorphic to* $(|N|,\mathcal{G})$.
A smooth manifold is a locally ringed space locally isomorphic to $({\ensuremath{\mathbb{R}}}^n,\mathcal{C}^\infty_{{\ensuremath{\mathbb{R}}}^{n}})$. This can be rephrased as a local isomorphism of graded locally ringed spaces whose sheaves have only elements of degree zero.
Graded domain {#SubSectGradedDomain}
-------------
\[DefDomGrad\] Let $(p_j)_{j\in {\ensuremath{\mathbb{Z}}}}$ be a sequence of non-negative integers and $\mathcal{W}=\bigoplus_{j} \mathcal{W}_j$ be a ${\ensuremath{\mathbb{Z}}}$-graded vector space of dimension $(p'_j)_{j \in {\ensuremath{\mathbb{Z}}}}$ with $p'_0=0$ and $p'_j=p_j$ otherwise. We say that $U^{(p_j)}=(U,\mathcal{O}_U)$ is a *${\ensuremath{\mathbb{Z}}}$-graded domain* of dimension $(p_j)_{j\in {\ensuremath{\mathbb{Z}}}}$, if $U$ is an open subset of ${\ensuremath{\mathbb{R}}}^{p_0}$ and for all $V\subseteq U$, $\mathcal{O}_U(V) =\mathcal{C}^\infty_{{\ensuremath{\mathbb{R}}}^{p_0}}(V)\otimes\overline{\mathbf{S}\mathcal{W}}$.
Let $(p_j)_{j\in {\ensuremath{\mathbb{Z}}}}$ be a sequence of non-negative integers. We write ${\ensuremath{\mathbb{R}}}^{(p_j)}=\left({\ensuremath{\mathbb{R}}}^{p_0}\right)^{(p_j)}$ to indicate the graded domain of dimension $(p_j)_{j\in {\ensuremath{\mathbb{Z}}}}$ and topological space ${\ensuremath{\mathbb{R}}}^{p_0}$.
A *global coordinate system* on $U^{(p_j)}$ is given by a global coordinate system $(t_1,\ldots,t_{p_0})$ on $U$, and by a graded basis $(w_\alpha)_\alpha$ of $\mathcal{W}$. We write such a system on a graded domain as $(t_i,w_\alpha)_{i,\alpha}$. From Equation , we see that any section $f\in \mathcal{O}_U(V)$, with $V\subseteq U$, admits a unique decomposition in the global coordinate system : $$\label{EqDecompo}
f=f_0 + \sum_{K=1}^\infty \sum_{\alpha_1\leqslant \ldots \leqslant \alpha_K}
f_{\alpha_1\ldots \alpha_K} w_{\alpha_1}\ldots w_{\alpha_K},$$ where $f_0$ and all the $f_{\alpha_1\ldots \alpha_K}$ are smooth functions on $V$.
Let $\mathcal{J}(V)$ be the *ideal generated by all sections with nonzero degree* on the open subset $V$. The map $\pi_1:\mathcal{O}_U(V)\to\mathcal{C}^\infty_{U}(V):f\mapsto f_0$, with $f_0$ defined in Equation , admits $\mathcal{J}(V)$ as kernel. Moreover, it is a left inverse for the embedding $\mathcal{C}^\infty_{U}(V)\to\mathcal{O}_U(V)$.
If we write $\pi$ for the canonical projection $\mathcal{O}_U(V)\to \mathcal{O}_U(V)/\mathcal{J}(V)$, then $\pi_1$ implies the existence of an isomorphism between $\mathcal{C}_U^\infty(V)$ and $\mathcal{O}_U(V)/\mathcal{J}(V)$ as shown in the following diagram :
\^\_[U]{}(V) && &&\_U(V)\
&(3,2)\_ &&&\_\
& &&&\_U(V)/(V)
The map $\pi_1$ can be used for evaluating a section at a point $x\in V$. It is the *value* at $x$.
Let $U^{(p_j)}=(U,\mathcal{O}_U)$ be a ${\ensuremath{\mathbb{Z}}}$-graded domain of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$ and $V\subseteq U$ be an open subset. Let $x\in V$ and $f\in \mathcal{O}_U(V)$. The *value* (or *evaluation*) of $f$ at $x$ is defined as $f(x)=\pi_1(f)(x)\in{\ensuremath{\mathbb{R}}}$.
The *ideal of sections with null value* at $x\in V$, written as $\mathcal{I}_x(V)$, is given by $\mathcal{I}_x(V)=\left\{ f\in\mathcal{O}_U(V) \big{|}f(x)=0\right\}$. Remark that $\mathcal{J}(V)\subseteq\mathcal{I}_x(V)$. For any $r\in {\ensuremath{\mathbb{N}}}^\times$, we can define $\mathcal{I}_x^r(V)$ as $\left\{f\in \mathcal{O}_{U}(V)\,\big{|}\,\exists f_1,\ldots,f_r\in\mathcal{I}_x(V), \, f=f_1\ldots f_r\right\}$.
*(Hadamard’s lemma).*\[LemHadamardDomGrad\] Let $U^{(p_j)}=(U,\mathcal{O}_U)$ be a ${\ensuremath{\mathbb{Z}}}$-graded domain of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$ with global coordinate system $(t_i,w_\alpha)_{i,\alpha}$. Consider an open subset $V\subseteq U$, a point $x\in V$ and a section $f\in \mathcal{O}_U(V)$. Then, for all $k\in {\ensuremath{\mathbb{N}}}$, there exists a polynomial $P_{k,x}$ of degree $k$ in the variables $\big(t_i-t_i(x)\big)_{i}$ and $(w_\alpha)_{\alpha}$ such that $f - P_{k,x} \in \mathcal{I}_x^{k+1}(V)$.
Recall that, if $g\in\mathcal{C}_U^\infty(V)$ is a smooth function in $t=(t_1,\ldots,t_{p_0})$, one can take its Taylor series of order $r\in{\ensuremath{\mathbb{N}}}$ at $x$, written as $T_{x}^r(g;t)$. It is a polynomial of order $r$ that satisfies $g(t)=T_{x}^r(g;t)+S_{x}^r(g;t)$ for $S_{x}^r(g;t)$ a sum of elements of the form $h(t) Q_x^r(t)$ with $h$ smooth and $Q_x^r(t)$ a homogeneous polynomial of order $r+1$ in the variables $(t_i-t_i(x))_i$.
From Equation , we can write $f\in \mathcal{O}_U(V)$ as $$f=f_0 + \sum_{K=1}^\infty \sum_{\alpha_1\leqslant \ldots \leqslant \alpha_K}
f_{\alpha_1\ldots \alpha_K} w_{\alpha_1}\ldots w_{\alpha_K}.$$
Since $f_0\in \mathcal{C}^\infty(V)$ and $f_{\alpha_1\ldots \alpha_K}\in \mathcal{C}^\infty(V)$ for all indices, we can develop them in Taylor series at $x$ in the variables $t=(t_1,\ldots,t_{p_0})$. If, for all $k\in{\ensuremath{\mathbb{N}}}$, we define $P_{k,x}$ as $$P_{k,x}(t)= T^k_{x}(f_0;t) + \sum_{K=1}^k \sum_{\alpha_1\leqslant \ldots \leqslant \alpha_K}
T^{k-K}_x(f_{\alpha_1\ldots \alpha_K};t) w_{\alpha_1}\ldots w_{\alpha_K},$$ then $P_{k,x}$ is a polynomial of order $k$. In particular, we obtain that $$\begin{aligned}
f-P_{k,x}(t)&=& S^k_{x}(f_0;t) + \sum_{K=1}^k \sum_{\alpha_1\leqslant \ldots \leqslant \alpha_K}
S^{k-K}_x(f_{\alpha_1\ldots \alpha_K};t) w_{\alpha_1}\ldots w_{\alpha_K} \\
&&+ \sum_{K= k+1}^\infty \sum_{\alpha_1\leqslant \ldots \leqslant \alpha_K}
f_{\alpha_1\ldots \alpha_K}(t) \, w_{\alpha_1}\ldots w_{\alpha_K} \,.\end{aligned}$$ As $w_\alpha\in \mathcal{I}_x(V)$ for all $\alpha$ and $S_x^r(g;t)\in\mathcal{I}_x^{r+1}(V)$ for any function $g$, we deduce that $f - P_{k,x} \in \mathcal{I}_x^{k+1}(V)$.
\[CorHadamard\] Let $U^{(p_j)}=(U,\mathcal{O}_U)$ be a ${\ensuremath{\mathbb{Z}}}$-graded domain of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$ and $V\subseteq U$ be an open subset. Then $$\bigcap_{k\geqslant1} \bigcap_{x\in V} \mathcal{I}_x^k(V) \, = \, \{0\}\,.$$
We show that, if $f,g\in\mathcal{O}_U(V)$ are two sections such that for all $k\in{\ensuremath{\mathbb{N}}}$, $f-g\in \mathcal{I}_x^{k+1}(V)$ for all $x\in V$, then $f=g$.
Consider a global coordinate system $(t_i,w_\alpha)_{i,\alpha}$ on $U^{(p_j)}$. Set $h=f-g$ and decompose this section as in Equation : $$h=h_0+ \sum_{K=1}^\infty \sum_{\alpha_1\leqslant \ldots \leqslant \alpha_K}
h_{\alpha_1\ldots \alpha_K} w_{\alpha_1}\ldots w_{\alpha_K}.$$ Fix $k\in {\ensuremath{\mathbb{N}}}$. For all $K\leqslant k$, the condition $h\in I_x^{k+1}(V)$ implies that $h_{\alpha_1\ldots \alpha_K} \in I_x^{k+1-K}(V)$. Since this is true for all $x\in V$, we have that $h_0=0$ and $h_{\alpha_1\ldots \alpha_K}=0$ on $V$ for all indices with $K\leqslant k$. As $k$ is arbitrary, we conclude that $h=0$ on $V$.
\[PropCaracTige\] Let $U^{(p_j)}=(U,\mathcal{O}_U)$ be a ${\ensuremath{\mathbb{Z}}}$-graded domain of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$ with global coordinate system $(t_i,w_\alpha)_{i,\alpha}$. Consider an open subset $V\subseteq U$ and a point $x\in V$. Then $\mathcal{I}_x(V)=\langle\{t_i-t_i(x),w_\alpha\}_{_{i,\alpha}}\rangle$.
We write $A=\langle \{t_i-t_i(x),w_\alpha\}_{_{i,\alpha}}\rangle$ for the proof. Consider a section $f\in \mathcal{I}_x(V)$. Decomposing $f$ as in Equation , we have $$f=f_0+ \sum_{K\geqslant 1} \sum_{\alpha_1\leqslant \ldots \leqslant \alpha_K}
f_{\alpha_1\ldots \alpha_K} w_{\alpha_1}\ldots w_{\alpha_K}.$$ From the definition of $A$, we only need to show that $f_0 \in A$. By assumption, $f\in \mathcal{I}_x(V)$, so that its value at $x$ is zero. Therefore, considering the Taylor series of $f_0$ of order $1$ gives that $f_0(t)=\sum_i h_i(t) (t_i-t_i(x))$ for some smooth functions $h_i$ with $i\in\{1,\ldots,p_0\}$. Hence $f_0\in A$ since $A$ contains the sections $(t_i-t_i(x))_i$. The inclusion $A\subseteq \mathcal{I}_x(V)$ is direct by definition of $A$.
\[PropEvaluationValeur\] Consider an open subset $V\subseteq U$ and a point $x\in V$. Then, $\mathcal{O}_{U}(V)={\ensuremath{\mathbb{R}}}\oplus \mathcal{I}_x(V)$. Moreover, the projection on constant sections ${\operatorname{ev}}_x:\mathcal{O}_{U}(V)\to {\ensuremath{\mathbb{R}}}$ calculates the value of every section.
Remark that a section $f\in\mathcal{O}_U(V)$ can be written as $f=(f-f(x))+f(x)$, where $f(x)=f(x)\cdot 1_V$ is a multiple of the unit section $1_V\in\mathcal{O}_U(V)$. Besides, one can see that the section $f-f(x)$ has value $0$ at $x$. Hence $f$ can be written as a sum of a constant section and an element of $\mathcal{I}_x(V)$ and we obtain that $\mathcal{O}_{U}(V)\subseteq {\ensuremath{\mathbb{R}}}\oplus \mathcal{I}_x(V)$. Under these notations, we have ${\operatorname{ev}}_x(f)=f(x)\cdot 1_V$. The second statement follows directly.
At a point $x$ of a graded domain $U^{(p_j)}$, we write the elements of the stalk, called *germs*, as $[f]_x\in\mathcal{O}_{U,x}$. The evaluation is defined on germs in the same manner as it is done for sections. We denote the ideal generated by all germs with nonzero degree by $\mathcal{J}_x$, and the ideal of germs with null value by $\mathcal{I}_x$. These two ideals can be obtained by inducing the ideals $\mathcal{J}(V)$ and $\mathcal{I}_x(V)$ on the stalk.
\[RemPropStalk\] The statements in Lemma \[LemHadamardDomGrad\] as well as Propositions \[CorHadamard\], \[PropCaracTige\] and \[PropEvaluationValeur\] can be reformulated for germs. In particular, we get that $\mathcal{I}_x$ is a homogeneous ideal such that $\mathcal{O}_{U,x}={\ensuremath{\mathbb{R}}}\oplus\mathcal{I}_x$.
\[LemMaximality\] $\mathcal{I}_x$ is the unique maximal homogeneous ideal of $\mathcal{O}_{U,x}$.
The decomposition $\mathcal{O}_{U,x}={\ensuremath{\mathbb{R}}}\oplus\mathcal{I}_x$ ensures that $\mathcal{I}_x$ is a maximal homogeneous ideal.
Assume that $\mathfrak{m}$ is another maximal homogeneous ideal of $\mathcal{O}_{U,x}$. Then, there exists an homogeneous element $[f]_x$ which is in $\mathfrak{m}$ but not in $\mathcal{I}_x$. Since $\mathcal{I}_x$ contains the ideal $\mathcal{J}_x$ generated by all germs of nonzero degree, $[f]_x$ is a germ of degree zero.
By maximality of $\mathcal{I}_x$, the ideal $\mathcal{O}_{U,x} [f]_x + \mathcal{I}_x$ is $\mathcal{O}_{U,x}$. Therefore, there exist $[h]_x\in\mathcal{O}_{U,x}$ and $[g]_x\in \mathcal{I}_x$, both of degree zero, such that $1=[h]_x[f]_x+[g]_x$. If we take the value of the germs at $x$, the equality gives that $1=[h]_x(x)[f]_x(x)$ since $[g]_x(x)=0$. Thus, setting $a=[f]_x(x)$, we have $a\in {\ensuremath{\mathbb{R}}}\setminus\{0\}$. According to the decomposition $\mathcal{O}_{U,x}={\ensuremath{\mathbb{R}}}\oplus\mathcal{I}_x$, we have that $[f]_x=a+[\tilde{f}]_x$ with some $[\tilde{f}]_x\in\mathcal{I}_x$ of degree zero.
Take a section $\tilde{f}$ defined in a neighbourhood of $x$ which represents $[\tilde{f}]_x$ and set $F=a^{-1}\sum\limits_{k=0}^\infty (-a^{-1}\tilde{f})^k$. On the stalk $\mathcal{O}_{U,x}$, we find that $[F]_x[f]_x=1$. Since $\mathfrak{m}$ is an ideal, this equality implies that $1\in\mathfrak{m}$. Thus, we have $\mathfrak{m}=\mathcal{O}_{U,x}$. Hence $\mathcal{I}_x$ is the unique maximal homogeneous ideal of $\mathcal{O}_{U,x}$.
A ${\ensuremath{\mathbb{Z}}}$-graded domain is a ${\ensuremath{\mathbb{Z}}}$-graded locally ringed space.
By Definition \[DefDomGrad\], a graded domain is a graded ringed space. The locality of the stalks follows from Lemma \[LemMaximality\].
\[RemLocality\] An alternative definition of graded domain can be given. We can set that the sheaf is $\mathcal{O}_U(V)=\mathcal{C}^\infty_U(V)\otimes \mathbf{S}\mathcal{W}$, namely we do not consider formal power series in the graded coordinates but only polynomials. Then, in general, the ideal $\mathcal{I}_x$ of germs with null value is not the unique maximal homogeneous ideal of $\mathcal{O}_{U,x}$. For example, the homogeneous ideal $\langle 1-[w]_x \rangle$ (where $w$ is a a non-nilpotent element of degree $0$ which is a product of local coordinates of nonzero degree) is contained in a maximal homogeneous ideal different from $\mathcal{I}_x$, as the germ $1-[w]_x$ has value $1$. Thus, the stalks of this alternative graded domain are not local. Nevertheless, they are $\pi$-local: $\mathcal{I}_x$ is the only maximal homogeneous ideal of $\mathcal{O}_{U,x}=\mathcal{C}^\infty_{U,x}\otimes \mathbf{S}\mathcal{W}$ which projects onto the maximal ideal of $\mathcal{C}^\infty_{U,x}$. Though this definition can be chosen, it limits the study of ${\ensuremath{\mathbb{Z}}}$-graded manifolds since a local expression of the form $f(t+w)$ would not always admit a finite power series in $w$, which is an obstacle to introduce differentiability.
Graded manifold {#SectionGradedManifold}
===============
Definition
----------
\[DefVarieteGraduee\] Let $|M|$ be a Hausdorff secound-countable topological space and $\mathcal{O}_{M}$ be a sheaf of associative unital supercommutative ${\ensuremath{\mathbb{Z}}}$-graded algebras on $|M|$, such that $M=(|M|,\mathcal{O}_{M})$ is a ${\ensuremath{\mathbb{Z}}}$-graded locally ringed space. Moreover, let $(p_j)_{j\in {\ensuremath{\mathbb{Z}}}}$ be a sequence of non-negative integers. We say that $M$ is a *${\ensuremath{\mathbb{Z}}}$-graded manifold* of dimension $(p_j)_{j\in {\ensuremath{\mathbb{Z}}}}$, if there exists a local isomorphism of ${\ensuremath{\mathbb{Z}}}$-graded locally ringed spaces $\phi$ between $M$ and ${\ensuremath{\mathbb{R}}}^{(p_j)}$. We say that an open subset $V\subseteq|M|$ is a *trivialising open set* if the restriction of $\phi$ to $V$, written $\phi_V$, is an isomorphism of graded locally ringed spaces between $(V,\restriction{\mathcal{O}_{M}}{V})$ and its image $\big(\widetilde{V},\restriction{\mathcal{C}^\infty_{{\ensuremath{\mathbb{R}}}^{p_0}}}{\widetilde{V}}\big)$, for $\widetilde{V}=\phi_V(V)$.
If a sequence $(t_i,w_\alpha)_{i,\alpha}$ is a global coordinate system on the ${\ensuremath{\mathbb{Z}}}$-graded domain ${\ensuremath{\mathbb{R}}}^{(p_j)}$, its pullbacks on trivialising open sets form a local coordinate system on $M$. By abuse of notation, one says that $(t_i,w_\alpha)_{i,\alpha}$ is a *local coordinate system* on $M$. Morphisms of graded manifolds are defined as morphisms of graded locally ringed spaces.
\[RemLocGradMan\] Alternatively, one can define a graded manifold as a graded $\pi$-locally ringed space (*i.e.* stalks are $\pi$-local graded rings), locally isomorphic to the alternative graded domains defined in Remark \[RemLocality\]. Full development of such a definition is left to the interested reader.
\[ExVarGradFibreQlcq\] Let $M$ be a smooth manifold of dimension $n$ and write $|M|$ for its underlying topological space. If $E\to M$ is a vector bundle of rank $m$ and $k$ is a nonzero integer, we define the sheaf $\mathcal{O}_{E[k]}$ on $|M|$ by
- $\mathcal{O}_{E[k]}(V)$ equals $\Gamma(V,\restriction{\bigwedge E^*}{V})$ if $k$ is odd,
- $\mathcal{O}_{E[k]}(V)$ equals $\Gamma(V,\restriction{\overline{\mathbf{S} E^*}}{V})$ if $k$ is even,
for all open subsets $V\subseteq |M|$. Here, the algebra bundles $\bigwedge E^*$ and $\overline{\mathbf{S}E^*}$ are considered in the usual non-graded sense. Then $\mathfrak{G}_{E[k]}=(|M|,\mathcal{O}_{E[k]})$ is a graded manifold of dimension $(p_j)_{j\in {\ensuremath{\mathbb{Z}}}}$, with $p_0=n$, $p_k=m$ and $p_j=0$ if $j\in {\ensuremath{\mathbb{Z}}}\setminus\{0,k\}$. The grading on the sections is defined by $|v|=kl$ if $v\in\Gamma\left(M,\mathbf{S}^l E^*\right)$ or $v\in \Gamma\big(M,\bigwedge ^l E^*\big)$.
The graded manifold in Example \[ExVarGradFibreQlcq\] is usually denoted by $E[k]$. We prefer to use a different notation in these notes so that such a graded manifold obtained from a shift is written differently than a lifted graded vector space.
One can define ${\ensuremath{\mathbb{N}}}$-graded manifolds as ${\ensuremath{\mathbb{Z}}}$-graded manifolds whose dimension is indexed by ${\ensuremath{\mathbb{N}}}$. Therefore, any homogeneous section has nonnegative degree. This leads to an interesting property: on a ${\ensuremath{\mathbb{N}}}$-graded manifold, there does not exist a section of degree zero which can be obtained as a product of sections of nonzero degree. This means that the locality and the $\pi$-locality of the stalks are equivalent conditions. Hence, ${\ensuremath{\mathbb{N}}}$-graded manifolds do not require the introduction of formal series to be studied, unlike ${\ensuremath{\mathbb{Z}}}$-graded manifolds (see Remarks \[RemLocality\] and \[RemLocGradMan\]). Another difference between ${\ensuremath{\mathbb{N}}}$- and ${\ensuremath{\mathbb{Z}}}$-graded manifolds is that there exists a Batchelor-type theorem on ${\ensuremath{\mathbb{N}}}$-graded manifolds [@BonavolontaPoncin]. It is not known if an analogous result holds for ${\ensuremath{\mathbb{Z}}}$-graded manifolds.
Properties {#SubSectProperties}
----------
Let $M=(|M|,\mathcal{O}_{M})$ be a ${\ensuremath{\mathbb{Z}}}$-graded manifold and take a point $x\in |M|$. By definition, the stalk $\mathcal{O}_{M,x}$ is a local graded ring which admits a maximal homogeneous ideal $\mathfrak{m}_x$. Using the isomorphism $\phi$ of a trivialising open set around $x$ with a graded domain $(U,\mathcal{O}_U)$, we get two canonical algebra isomorphisms : $\mathcal{O}_{M,x}/\mathfrak{m}_{x}\simeq \mathcal{O}_{U,\phi(x)}/\mathcal{I}_{\phi(x)}\simeq {\ensuremath{\mathbb{R}}}$. We set ${\operatorname{ev}}_x:\mathcal{O}_{M,x}\to {\ensuremath{\mathbb{R}}}$ to denote the composition of the projection $\mathcal{O}_{M,x}\to\mathcal{O}_{M,x}/\mathfrak{m}_{x}$ with the above isomorphism.
\[DefValeur\] Let $M=(|M|,\mathcal{O}_M)$ be a ${\ensuremath{\mathbb{Z}}}$-graded manifold of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$, $V\subseteq |M|$ an open subset and $x\in V$. Set $\mathfrak{m}_x$ to indicate the maximal ideal of $\mathcal{O}_{M,x}$. For every section $f\in \mathcal{O}_M(V)$, the *value* (or *evaluation*) of $f$ at $x$ is given by $f(x)={\operatorname{ev}}_x([f]_x)$.
From the definition above, the ideal of sections with null value $\mathcal{I}_x$ at $x$ can be defined on the graded manifold as the kernel of ${\operatorname{ev}}_x$. Moreover, if $V\subseteq |M|$ is a trivialising open set, Hadamard’s Lemma \[LemHadamardDomGrad\] and Proposition \[CorHadamard\] hold on $V$.
Although it is not used for defining the value on a graded manifold, the ideal generated by all sections with nonzero degree exists. We write $\mathcal{J}(V)$ to denote this ideal on an open subset $V$ of a graded manifold $M$. If $V$ is a trivialising open set, for any section $f\in\mathcal{O}_{M}(V)$ we set $\tilde{f}$ to indicate the element $\pi_1(f)\in \mathcal{C}^\infty(V)$ (see \[SubSectGradedDomain\]). Under this notation, the following partition of unity holds true :
\[PropPartUnite\] Let $M=(|M|,\mathcal{O}_M)$ be a ${\ensuremath{\mathbb{Z}}}$-graded manifold of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$ and $(V_\beta)_{\beta}$ be an open covering of $|M|$ by trivialising open subsets. Then there exists a sequence $(g_\beta)_{\beta}\subset \mathcal{O}_M(|M|)$ such that for all $\beta$,
1. $g_\beta \in \left(\mathcal{O}_M(|M|)\right)_0$ ,
2. $\text{\emph{supp}}\left( g_\beta \right)\subset V_\beta$ ,
3. $\sum g_\beta=1$ and $\tilde{g}_\beta \geqslant 0$.
The proof of this proposition is exactly the same that the one for the partition of unity on a supermanifold [@CCFsupersym].
Let $M=(|M|,\mathcal{O}_M)$ be a graded manifold. We can use Proposition \[PropPartUnite\] to show that the presheaf $\mathcal{O}_M/\mathcal{J}:V\mapsto \mathcal{O}_M(V)/\mathcal{J}(V)$, where $V\subseteq |M|$, is a sheaf. Note that $\mathcal{O}_M$ is $\pi$-local since it is local. Therefore the stalks of $\mathcal{O}_M/\mathcal{J}$ are local. Hence, $\mathcal{O}_M/\mathcal{J}$ is a sheaf locally isomorphic to $\mathcal{C}^\infty_{{\ensuremath{\mathbb{R}}}^{p_0}}$ and we get that $(|M|,\mathcal{O}_M/\mathcal{J})$ is a smooth manifold. Thus, one can assume from the beginning that a graded manifold has an underlying structure of smooth manifold. This justifies the frequent definition of graded manifold found in the literature, which considers graded manifolds as smooth manifolds with an additional *glued* structure.
\[RemZeroDeg\] In recent developments (see [@Vogl16; @MehtaStiXu]) formal polynomial functions on graded manifolds have been considered with respect to a ${\ensuremath{\mathbb{Z}}}$-graded vector space that has a non-trivial part of degree zero. This means that, in Definition \[DefVarieteGraduee\], we could choose $\mathcal{W}$ such that $\text{dim}(\mathcal{W}_0)\geqslant 1$. In that case, the underlying structure of smooth manifold still exists, but one has to obtain it from the morphism $\pi_1$ instead of $\pi$ (see \[SubSectGradedDomain\]) as their images are no longer isomorphic.
Vector fields
-------------
Let $M=(|M|,\mathcal{O}_M)$ be a ${\ensuremath{\mathbb{Z}}}$-graded manifold. We say that a *derivation of degree $k$* on an open subset $V\subseteq|M|$ is an application $X_V^k:\mathcal{O}_M(V)\to \mathcal{O}_M(V)$ which maps every section of degree $m$ onto a section of degree $m+k$ and which satisfies $$\label{EqDerivation}
X_V^k(fg)=X_V^k(f)\,g+(-1)^{k |f|}f\,X_V^k(g), \text{ }\forall f,g\in \mathcal{O}_M(V).$$
A *vector field* $X$ on a ${\ensuremath{\mathbb{Z}}}$-graded manifold $M=(|M|,\mathcal{O}_M)$ of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$ is a ${\ensuremath{\mathbb{R}}}$-linear derivation of $\mathcal{O}_M$, *i.e.* a family of mappings $X_V:\mathcal{O}_M(V)\to \mathcal{O}_M(V)$ for all open subsets $V\subseteq|M|$, compatible with the sheaf restriction morphism and that satisfies $X_V=\sum_{k\in {\ensuremath{\mathbb{Z}}}} X_V^k$, where $X_V^k$ is a derivation of degree $k$. We say that $X$ is *graded* of degree $k$ if $X_V=X_V^k$ for all $V\subseteq |M|$.
The *graded tangent bundle* ${\operatorname{Vec}}_M$ of $M$ is the sheaf whose sections are the vector fields. It assigns to every open subset $V\subseteq|M|$ the graded vector space of derivations on $\restriction{\mathcal{O}_M}{V}$.
First, we study the graded tangent bundle on a graded domain. Assume that $U^{(p_j)}=(U,\mathcal{O}_U)$ is a graded domain of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$ with global coordinate system $(t_i,w_\alpha)_{i,\alpha}$. We have that $\mathcal{O}_U=\mathcal{C}^\infty_U\otimes\overline{\mathbf{S}\mathcal{W}}$ for some graded vector space $\mathcal{W}$ with graded basis $(w_\alpha)_{\alpha}$. Consider the basis $\left(\frac{\partial}{\partial t_i}\right)_i$ of the non-graded tangent bundle ${\operatorname{Vec}}_{U}$ and set $\left(\frac{\partial}{\partial w_\alpha}\right)_\alpha$ for the dual graded basis of $(w_\alpha)_\alpha$ in $\mathcal{W}^*$. These objects satisfy, for all indices $i,j,\alpha,\beta$, $$\begin{aligned}
\quad\frac{\partial}{\partial t_i}\, (t_j)=\delta_{ij}\,,\quad \quad\frac{\partial}{\partial t_i}\, (w_\beta)=0 \,, \label{EqCDVti}\\
\quad\frac{\partial}{\partial w_\alpha}\, (w_\beta)=\delta_{\alpha\beta}\,,\quad \quad\frac{\partial}{\partial w_\alpha}\,(t_j)=0
\label{EqCDVwalpha}\,.\end{aligned}$$ Moreover, we can extend them as ${\ensuremath{\mathbb{R}}}$-linear derivations on the whole sheaf by Leibniz’s rule . These are graded vector fields and the degree of $\frac{\partial}{\partial t_i}$ is $0$ while the degree of $\frac{\partial}{\partial w_\alpha}$ is $-|w_\alpha|$.
\[LemDecompoCDV\] Let $U^{(p_j)}=(U,\mathcal{O}_U)$ be a ${\ensuremath{\mathbb{Z}}}$-graded domain of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$. Then its graded tangent bundle ${\operatorname{Vec}}_{U^{(p_{j})}}$ is a free sheaf of $\mathcal{O}_U$-modules such that $$\label{EqBaseCDV}
{\operatorname{Vec}}_{U^{(p_{j})}} \, \simeq \, \bigg(\mathcal{O}_U \underset{\mathcal{C}^\infty_U}{\otimes}
{\operatorname{Vec}}_{U}\bigg) \,\, \oplus \,\, \left(\mathcal{O}_U \underset{{\ensuremath{\mathbb{R}}}}{\otimes} \mathcal{W}^*\right) \,.$$ If $(t_i,w_\alpha)_{i,\alpha}$ is a global coordinate system on $U^{(p_j)}$, the vector field $X$ admits the following splitting : $$X=\sum_{i}\, X_{0,i}\frac{\partial}{\partial t_i}
+\sum_{\alpha}\, X_{1,\alpha}\frac{\partial}{\partial w_\alpha}\,,$$ where $X_{0,i}=X(t_i)\in\mathcal{O}_U(U)$ and $X_{1,\alpha}=X(w_\alpha)\in \mathcal{O}_U(U)$ for all $i,\alpha$.
Let $X\in {\operatorname{Vec}}_{U^{(p_{j})}}$ be a vector field. For all open subsets $V\subseteq U$ and every $f\in\mathcal{O}_U(V)$, $f X$ is a derivation on $\mathcal{O}_U(V)$. Therefore ${\operatorname{Vec}}_{U^{(p_{j})}}$ is a sheaf of $\mathcal{O}_U$-modules. The fact that it is free and Equation are consequences of the existence of a graded basis, which is obtained below.
Assume that the global coordinates on $U^{(p_j)}$ are $(t_i,w_\alpha)_{i,\alpha}$ and let $X\in {\operatorname{Vec}}_{U^{(p_{j})}}$ be a vector field. We define $$\hat{X}=\sum \limits_{i} X(t_i) \frac{\partial}{\partial t_i} + \sum \limits_{\alpha}
X(w_\alpha) \frac{\partial}{\partial w_\alpha},$$ where $X(z)$ is the section in $\mathcal{O}_{U}(U)$ given by applying $X$ to the section $z$. Since $\hat{X}$ is a derivation, $D=X-\hat{X}$ is also one and it vanishes on the global coordinates $(t_i,w_\alpha)_{i,\alpha}$. By Leibniz’s rule , $D$ is equal to zero on all polynomials generated by the global coordinates.
Our aim is to show that $D$ is null on all sections. Take an open subset $V\subseteq U$ and a section $f\in\mathcal{O}_U(V)$. From Lemma \[LemHadamardDomGrad\] we have that, for all $n\geqslant 0$, there exists a polynomial $P_{n,x}$ of degree $n$ in the global coordinates such that $f-P_{n,x}\in \mathcal{I}_x^{n+1}(V)$. We set $h=f-P_{n,x}$.
Notice that for any vector field $Y$ we have $Y(\mathcal{I}_x^n(V))\subset\mathcal{I}_x^{n-1}(V)$ for $n\geqslant1$. Since $D$ vanishes on every polynomial in $(t_i)_i$ and $(w_\alpha)_\alpha$, we have by linearity that $D(f)=D(h)\in \mathcal{I}_x^n(V)$. Since this is true for all $n\geqslant 1$ and any point $x\in V$, Proposition \[CorHadamard\] gives that $D(f)=0$ on $U$. As $f$ is any section of $\mathcal{O}_U(V)$ and $V$ is any open set, we have proved that $D=0$.
It remains to show that the decomposition is unique. Assume that $\tilde{X}$ is another decomposition of $X$ such that $\tilde{X}=\sum\limits_{i}f_i\frac{\partial}{\partial t_i}+\sum\limits_{\alpha}g_\alpha\frac{\partial}{\partial w_\alpha}.$ The relations $(\tilde{X}-\hat{X})(t_i)=0$ and $(\tilde{X}-\hat{X})(w_\alpha)=0$ implies that $f_i=X(t_i)$ and $g_\alpha=X(w_\alpha)$.
The following result is a consequence of Definition \[DefVarieteGraduee\] and Lemma \[LemDecompoCDV\] :
\[CorCaracBaseCDV\] Let $M=(|M|,\mathcal{O}_M)$ be a ${\ensuremath{\mathbb{Z}}}$-graded manifold of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$. Then ${{\operatorname{Vec}}_M}$ is a locally free sheaf of $\mathcal{O}_M$-modules such that, for any trivialising open subset $V\subseteq |M|$, its restriction to $V$ satisfies $$\restriction{{\operatorname{Vec}}_{M}}{V} \, \simeq \, \bigg(\restriction{\mathcal{O}_M}{V} \underset{\mathcal{C}^\infty_V}{\otimes}
{\operatorname{Vec}}_{V}\bigg) \,\, \oplus \,\, \left(\restriction{\mathcal{O}_M}{V} \underset{{\ensuremath{\mathbb{R}}}}{\otimes} \mathcal{W}^*\right) \,.$$ If $(t_i,w_\alpha)_{i,\alpha}$ is a local coordinate system on $M$, the vector field $X$ admits the following splitting on a trivialising open set $V$ : $$X=\sum_{i}\, X_{0,i}\frac{\partial}{\partial t_i}
+\sum_{\alpha}\, X_{1,\alpha}\frac{\partial}{\partial w_\alpha}\,,$$ where $X_{0,i}=X(t_i)\in \mathcal{O}_M(V)$ and $X_{1,\alpha}=X(w_\alpha)\in \mathcal{O}_M(V)$ for all $i,\alpha$.
The graded tangent bundle of a graded manifold can be turned into a sheaf of graded Lie algebras if it is endowed with the graded commutator $[a,b]=ab-(-1)^{|a||b|}ba$.
Given a graded manifold $M=(|M|,\mathcal{O}_M)$, the map ${\operatorname{ev}}_x:\mathcal{O}_{M,x}\to {\ensuremath{\mathbb{R}}}$ allows us to define tangent vectors at a point $x\in M$. To see it, first remark that any vector field $X$ can be induced on the stalk. Then, if $X$ is defined locally by $X=\sum_{i}\, f_i\frac{\partial}{\partial t_i}+\sum_{\alpha}\, g_\alpha \frac{\partial}{\partial w_\alpha}$, its restriction to the stalk $[X]_x:\mathcal{O}_{M,x}\to \mathcal{O}_{M,x}$ is given by $$\begin{aligned}
[X]_x= \sum_i [f_i]_x \,\, \left[\frac{\partial}{\partial t_i}\right]_{x}
+\sum_{\alpha}\, [g_\alpha]_x \,\, \left[\frac{\partial}{\partial w_\alpha}\right]_{x},\end{aligned}$$ where $\left[\frac{\partial}{\partial t_i}\right]_{x}$ and $\left[\frac{\partial}{\partial w_\alpha}\right]_{x}$ satisfy Equations and induced on $\mathcal{O}_{M,x}$, respectively. The *tangent vector associated to* $X$, written $X_x$, is the application $X_x={\operatorname{ev}}_x\circ [X]_x:\mathcal{O}_{M,x}\to {\ensuremath{\mathbb{R}}}$. More generally, tangent vectors are defined from derivations of the stalk by composition with ${\operatorname{ev}}_x$. The graded vector space of tangent vectors at $x$ is the *graded tangent space*, denoted by $T_xM$.
It is usually not possible to recover a vector field from the tangent vectors that it defines at all points. Indeed, assume that a graded manifold $M$ admits $(t_1,w_1)$ as local coordinates of respective degrees $0$ and $1$. Consider the vector field $X$ given locally by $X=\frac{\partial}{\partial t_1}+w_1 \frac{\partial}{\partial w_1}$. Then the induced tangent vector at $x\in M$ satisfies $X_x=\restriction{\frac{\partial}{\partial t_1}}{x}$, where $\restriction{\frac{\partial}{\partial t_1}}{x}={\operatorname{ev}}_x\circ \left[\frac{\partial}{\partial t_1}\right]_{x}$.
Application to differential graded manifolds {#SectionApplication}
============================================
Let $M=(|M|,\mathcal{O}_M)$ be a ${\ensuremath{\mathbb{Z}}}$-graded manifold of dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$. A vector field $Q \in {\operatorname{Vec}}_M$ is *homological* if it has degree $+1$ and commutes with itself under the graded commutator : $[Q,Q]=2Q^2=0$.
A ${\ensuremath{\mathbb{Z}}}$-graded manifold $M$ endowed with a homological vector field $Q$ is said to be a *differential graded manifold* or a *$Q$-manifold*. We write it $(M,Q)$ and we refer to it as a *dg-manifold*.
Let $M$ be a smooth manifold of dimension $n$ and let $\Omega$ be the sheaf of its differential forms. The graded manifold $\mathfrak{G}_{TM[1]}=(|M|,\Omega)$ has dimension $(p_j)_{j\in{\ensuremath{\mathbb{Z}}}}$, with $p_j=n$ if $j=0,1$ and $p_j=0$ otherwise. If $(x_i)_{i=1}^n$ are local coordinates on $M$, then $(x_i,dx_i)_{i=1}^n$ is a local coordinate system on $\mathfrak{G}_{TM[1]}$ whose degrees are respectively $0$ and $1$. The exterior differential $d$ is a vector field on $\mathfrak{G}_{TM[1]}$, since its restriction on any open set $V\subseteq |M|$ is a derivation of $\Omega(V)$. It is given locally by $d=\sum_i dx_i \frac{\partial}{\partial x_i}$. One can compute that $|d|=1$ and $[d,d]=2\, d^2=0$. Therefore, $(\mathfrak{G}_{TM[1]},d)$ is a dg-manifold.
Several geometrico-algebraic structures can be encoded in terms of dg-manifolds. The easiest example is certainly the Lie algebra structures on a vector space.
\[ThmLieAlg\] Let $\mathcal{V}_0$ be a vector space. Then the structures of Lie algebra on $\mathcal{V}_0$ are in one-to-one correspondence with the homological vector fields on $\mathfrak{G}_{\mathcal{V}_0[1]}=\left(\{\ast\},\mathbf{S} \left(\mathcal{V}_0^*[1]\right) \right)$.
If $\mathcal{V}_0$ is a Lie algebra, one can show that its Chevalley-Eilenberg differential is a vector field of degree $1$ which squares to zero. It endows $\mathfrak{G}_{\mathcal{V}_0[1]}$ with a homological vector field. Conversely, one can construct a Lie bracket, known as the derived bracket, from the homological vector field. This construction is detailed in [@KosmannYvette].
In fact, Theorem \[ThmLieAlg\] can be seen as a corollary of two more general correspondences. If we consider an arbitrary ${\ensuremath{\mathbb{Z}}}$-graded vector space, we have
\[ThmSHLieAlg\] Let $\mathcal{V}$ be a ${\ensuremath{\mathbb{Z}}}$-graded vector space such that $\text{dim}(\mathcal{V}_i)=0$ for all $i\leqslant 0$. Then the structures of strongly homotopy Lie algebra on $\mathcal{V}$ are in one-to-one correspondence with the homological vector fields on $\mathfrak{G}_{\mathcal{V}}=\big(\{\ast\},\overline{\mathbf{S} \left(\Pi \mathcal{V}^*\right)} \big)$.
Our definition of $\mathcal{V}$ being a strongly homotopy Lie algebra is that $\mathcal{V}[-1]$ admits a sh-Lie structure as stated by Lada and Stasheff in [@LadaStasheff]. The proof for Theorem \[ThmLieAlg\] strictly follows the correspondence of such structures with degree $+1$ differentials on the free graded algebra $\mathbf{S} \left(\Pi \mathcal{V}^*\right)$. It was proved in [@LadaStasheff] that each strongly homotopy Lie algebra implies the existence of a differential, while the converse can be found in [@LadaMarkl]. We refer to [@SatiSchreiStash §6.1] for additional details on this relation. Note that Voronov has provided an analogous definition in ${\ensuremath{\mathbb{Z}}}_2$-settings, which gives an equivalent result for supermanifolds [@VoronovDerivedBrackets; @VoronovHigherBrackets].
An alternative generalization of Theorem \[ThmLieAlg\] is to allow a non-trivial underlying topological space :
\[ThmLieAlgebroid\] Let $E\to M$ be a real vector bundle. Then the structures of real Lie algebroid on $E$ are in one-to-one correspondence with the homological vector fields on $\mathfrak{G}_{E[1]}=\big(|M|,\Gamma(|M|,\bigwedge E^*)\big)$.
This theorem is due to Vaĭntrob [@Vaintrob] in the framework of supermanifolds. The proof is exactly the same in the ${\ensuremath{\mathbb{Z}}}$-graded case, since there are no local coordinates of even nonzero degree on $\mathfrak{G}_{E[1]}$ (see *e.g.* [@KiselevVanDeLeur]).
To conclude, the interested reader is invited to find in [@BonavolontaPoncin] another general equivalence. It links Lie $n$-algebroids to ${\ensuremath{\mathbb{N}}}$-graded manifolds (with a homological vector field and generators of degree at most $n$), for all $n\in{\ensuremath{\mathbb{N}}}^\times$.
**Acknowledgments.** This work is based on the master’s thesis submitted by the author at the Université Catholique de Louvain in June 2015. He is very grateful to Jean-Philippe Michel for lots of useful discussions and for his remarks on the different versions of this paper, and to Yannick Voglaire for bringing the material covered in Remark \[RemZeroDeg\] to his attention. The author also thanks the referees for useful suggestions. Part of the writing up of this paper was done with the support of a *University of Leeds 110 Anniversary Research Scholarship*.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We provide conditions on the $p$-adic Galois representation of a smooth proper variety over a complete nonarchimedean extension of $\QQ_p$ to have (potentially) good ordinary reduction.'
author:
- Sanath Devalapurkar
bibliography:
- '../../main.bib'
title: Galois representations and ordinary reduction
---
Introduction
============
Let $K/\QQ_p$ be a finite extension with perfect residue field $k$ of characteristic $p$, and let $C$ be a completion of an algebraic closure of $K$. A smooth proper variety $X_0/K$ has *good reduction* if there it admits a smooth proper model $\fr{X}_0$ over $\co_K$. Let $X = X_0\otimes_K\ol{K}$. In this case:
1. The $\gal{K}$-representation on $\H^n_\et(X;\QQ_\ell)$ is unramified for all $n$ and all primes $\ell\neq p$; moreover,
2. The $\gal{K}$-representation $\H^n_\et(X;\QQ_p)$ is crystalline for all $n$.
In many cases, there is a converse to this result. For instance, when $X_0$ is an abelian variety, the Néron–Ogg–Shafarevich theorem (see [@serre-tate]) provides a converse: $H^1_\et(X;\QQ_\ell)$ is an unramified $\gal{K}$-representation for one (hence all) primes $\ell\neq p$ if and only if $X_0$ has good reduction. The $p$-adic analogue of this result was proved by Coleman–Iovita ([@coleman-iovita Theorem II.4.7]): $X_0$ has good reduction if and only if $H^1_\et(X;\QQ_p)$ is a crystalline $p$-adic $\gal{K}$-representation. In [@neron-ogg-k3], Chiarellotto, Lazda, and Liedtke obtained analogous statements for certain K3 surfaces.
Let $A$ be an abelian variety over $K$ of dimension $g$ with good reduction, so there exists a smooth proper model $\cA$ over $\co_K$ whose generic fiber is $K$. The special fiber $\cA_\sp$ is an abelian variety over a perfect field of characteristic $p$. The associated $p$-divisible group $\cA_\sp[p^\infty]$ has height $2g$; however, its connected component $\cA_\sp[p^\infty]^\circ$ need not have the same height. The height of the formal group $\cA_\sp[p^\infty]^\circ$ provides a stratification of the moduli space of abelian varieties. The subset with maximal height (namely, $g$) is dense in this moduli space; in this case, the abelian variety has ordinary reduction.
In light of the above discussion, it is natural to ask if there is a Galois-theoretic criterion on the étale cohomology of $A\otimes_K\ol{K}$ which ensures that $A$ has ordinary reduction. The answer is yes:
\[thm1\] Let $A$ be an abelian variety with good reduction, and let $V$ denote the $p$-adic $\gal{K}$-representation $\H^n_\et(A\otimes_K \ol{K};\QQ_p)$. Then $A$ is ordinary iff there is a complete filtration $F^i V$ of $V$ which splits, such that the inertia subgroup of $\gal{K}$ acts trivially on $F^i
V/F^{i+1} V$ for $i>\dim A$, and by the $p$-adic cyclotomic character for $i\leq \dim A$.
This result is probably well-known to the experts; however, we were unable to find a reference in the literature, so we include a proof in this paper.
Note that as ordinarity is a condition on the $p$-divisible group of $\cA_\sp$, one only expects such a criterion on the $p$-adic étale cohomology of $A\otimes_K\ol{K}$, and not its $\ell$-adic étale cohomology.
Motivated by this result, one can ask if an analogue of Theorem \[thm1\] holds for smooth proper varieties which are not necessarily abelian varieties. In order to do this, one of course needs an appropriate definition of the term “ordinary” for general smooth proper varieties. This was provided by Bloch and Kato in [@bloch-kato]. Let $X$ be a smooth proper variety over $K$ with good reduction, and let $\fr{X}$ denote a smooth proper model for $X$ over $\co_K$.
Let $d\Omega^j_{\fr{X}_\sp}$ denote the sheaf of exact differentials on $\fr{X}_\sp$. We say that $X$ has *ordinary reduction* if $H^i(\fr{X}_\sp, d\Omega^j_{\fr{X}_\sp})$ is $0$ for all $i,j$.
In analogy with the condition imposed in Theorem \[thm1\], we also make the following definition.
\[ordinary-gal-rep\] A $\gal{K}$-representation $V$ is *ordinary* if there is a finite filtration by $\gal{K}$-stable vector spaces $F^i V$ such that the inertia subgroup of $\gal{K}$ acts on $F^i V/F^{i+1} V$ by some power $\chi^{n_i}$ of the cyclotomic character.
The main goal of this paper is to prove the following analogues of Theorem \[thm1\]:
\[thm2\] Let $X_0$ be a smooth proper variety over $K$ with a smooth proper model $\fr{X}$ over $\co_K$ such that $H^\ast_\cris(\fr{X}_\sp/W(k))$ and $H^\ast(\fr{X}_\sp,W\Omega^\ast_{\fr{X}_\sp})$ are torsion-free. The étale cohomology $H^i_\et(X;\QQ_p)$ is an ordinary $\gal{K}$-representation whose associated filtration (from Definition \[ordinary-gal-rep\]) splits for all $i$ if and only if $X_0$ has ordinary reduction.
\[thm3\] Let $X_0$ be a K3 surface over $K$ with good reduction. The étale cohomology $H^2_\et(X;\QQ_p)$ is an ordinary $\gal{K}$-representation whose associated filtration (from Definition \[ordinary-gal-rep\]) splits if and only if $X_0$ has ordinary reduction.
Combined with [@matsumoto-k3], we obtain conditions on the $\gal{K}$-representation associated to a K3 surface guaranteeing potentially good ordinary reduction.
We will deduce these results from the following more general theorem.
\[thm4\] Suppose $X_0$ is a smooth proper variety over $K$ with good reduction such that the special fiber of some model $\fr{X}$ of $X_0$ over $\co_K$ has torsion-free crystalline cohomology. If $H^n_\et(X;\QQ_p)$ is an ordinary $\gal{K}$-representation whose filtration (from Definition \[ordinary-gal-rep\]) splits, then the Newton polygon for $\fr{X}_\sp$ has integer slopes.
Acknowledgements {#acknowledgements .unnumbered}
----------------
Thanks to Matthew Emerton, Mark Kisin, Sameera Vemulapalli, and David Zureick-Brown for chatting with me. Thanks also to the anonymous reader who pointed out errors in an earlier draft of this writeup, and to Bjorn Poonen for forwarding these comments to me.
Background
==========
Proving Theorem \[thm1\]
------------------------
Our goal is to prove Theorem \[thm1\], which we will recall here for the reader’s convenience.
Let $A$ be an abelian variety with good reduction, and let $V$ denote the $p$-adic $\gal{K}$-representation $\H^n_\et(A\otimes_K \ol{K};\QQ_p)$. Then $A$ is ordinary iff there is a complete filtration $F^i V$ of $V$ which splits, such that the inertia subgroup of $\gal{K}$ acts trivially on $F^i
V/F^{i+1} V$ for $i>\dim A$, and by the $p$-adic cyclotomic character for $i\leq \dim A$.
In this case, we only have to consider the case $n=1$. Assume that $A$ is ordinary. Since $A$ has good reduction, we lift it to a smooth proper model $\cA$ over $\co_K$. Let $\cA[p^\infty]$ denote the $p$-divisible group of its special fiber. The connected-étale sequence runs $$0\to \cA[p^\infty]^\circ \to \cA[p^\infty]\to \cA[p^\infty]^\et\to 0;$$ taking the rational $p$-adic Tate module gives $$0\to V_p(\cA[p^\infty]^\circ) \to V_p(\cA[p^\infty]) \to
V_p(\cA[p^\infty]^\et) \to 0.$$ The $\gal{K}$-representation $V_p(\cA[p^\infty]^\et)$ is unramified, while the inertia subgroup of $\gal{K}$ acts on $V_p(\cA[p^\infty]^\circ)$ by a direct sum of copies of the $p$-adic cyclotomic character. By [@neron-ogg-k3 Theorem 2.4], we know that as $\gal{K}$-representations, we have $$H^n_\et(A\otimes_K\ol{K}; \QQ_p) \simeq
H^n_\et(\cA\otimes_{\co_K}\widehat{\ol{K}}; \QQ_p).$$ Note that $V_p(A) = H^1_\et(A\otimes_K\ol{K};\QQ_p)^\vee$, so $V_p(\cA[p^\infty])$ can be identified with $V_p(A) =: V$ as $\gal{K}$-representations. As $A$ is ordinary, we know that $\cA[p^\infty]$ is an extension of an étale $p$-divisible group by a multiplicative $p$-divisible group; taking the rational Tate module of the resulting filtration on $\cA[p^\infty]$ gives a filtration on $V$ (with $F^{\dim A} V
= V_p(\cA[p^\infty]^\circ)$) satisfying the conditions listed in Theorem \[thm1\].
For the converse, suppose that the $p$-adic $\gal{K}$-representation $\H^1_\et(A\otimes_K\ol{K};\QQ_p) =: V$ satisfies the conditions in Theorem \[thm1\]. Then $V$ contains a $\gal{K}$-stable $\Z_p$-lattice $\Lambda:=H^1_\et(A\otimes_K\ol{K};\Z_p)$. Under the correspondence of [@brinon-conrad Proposition 7.2.2], this lattice comes from the $p$-divisible group $A[p^\infty]$. Since $A$ has good reduction, this can be lifted to a $p$-divisible group over $\co_K$.
Let $W\subsetneq V$ denote the $\dim A$-th step of the filtration on $V$. We get a $\gal{K}$-stable $\Z_p$-lattice $W \cap \Lambda$ of $W$ which is contained inside $\Lambda$. Again referring to [@brinon-conrad Proposition 7.2.2], we can find a $p$-divisible subgroup $\GG$ of $A[p^\infty]$. This in turn can be lifted to a subgroup $\wt{\GG}$ of the $p$-divisible group over $\co_K$. By our assumptions on $V$, the associated $\gal{K}$-representation itself admits a complete split filtration, with the inertia subgroup of $\gal{K}$ acting on each quotient by the $p$-adic cyclotomic character. Thereby identifying $\wt{\GG}$ with a multiplicative $p$-divisible group, we find that the special fiber $\cA_\sp$ is an ordinary abelian variety, as desired.
Breuil–Kisin modules
--------------------
In order to prove Theorem \[thm4\], we will need to recall some of the theory of Breuil–Kisin modules. Let $\fr{S} = W(k)[[T]]$. There is a Frobenius $\varphi$ on $\fr{S}$ defined by the usual Frobenius on $W(k)$ and the map $T\mapsto T^p$. The map $\fr{S}\to \co_K$ sending $T$ to a uniformizer $\pi$ has kernel generated by $E(T)$, the Eisenstein polynomial for $\pi$. Recall:
A Breuil–Kisin module $M$ is a $\fr{S}$-module along with an isomorphism $$\varphi_M: M\otimes_\fr{S}^\varphi \fr{S}[\frac{1}{E}] \simeq
M[\frac{1}{E}].$$
Let $\fr{S}\{1\}$ denote the Breuil–Kisin module with underlying $\fr{S}$-module $\fr{S}$ and Frobenius $\varphi_{\fr{S}\{1\}}(x) =
\frac{u}{E(T)}\varphi(x)$, with $u$ some explicit unit in $\fr{S}$. We write $\fr{S}\{n\} = \fr{S}\{1\}^{\otimes n}$.
In analogy with Definition \[ordinary-gal-rep\], we say that a Breuil–Kisin module $(M,\varphi_M)$ is ordinary if there is a filtration by submodules $(F^i M,\varphi_M|_{F^i M})$ such that each successive quotient is a finite free Breuil–Kisin module of rank $1$.
We will need the following result (see [@bms-i Theorem 4.4]):
\[bk\] There is a fully faithful tensor functor $M$ from $\Z_p$-lattices $\Lambda$ in crystalline $\gal{K}$-representations $V$ to finite free Breuil–Kisin modules, characterized by the property that there is a $\varphi,\gal{K_\infty}$-equivariant identification $$M(\Lambda)\otimes_\fr{S} W(C^\flat) \simeq \Lambda\otimes_{\Z_p}
W(C^\flat).$$
\[tate\] We have $M(\Z_p(n)) \simeq \fr{S}\{n\}$.
The functor of Theorem \[bk\] is not generally exact. However, we have:
\[ordinary-bk\] Let $V$ be an ordinary crystalline $\gal{K}$-representation, and let $\Lambda$ be a $\gal{K}$-stable $\Z_p$-lattice in $V$. Then $M(\Lambda)$ is an ordinary Breuil–Kisin module.
Let $F^i V$ be the filtration on $V$ (from Definition \[ordinary-gal-rep\]). Then $F^i V \cap \Lambda =: F^i \Lambda$ forms a filtration of $\gal{K}$-stable $\Z_p$-lattices inside $\Lambda$. As $W(C^\flat)$ is torsion-free, the characterization of $M(\Lambda)$ from Theorem \[bk\] proves that the rank of $M(\Lambda)$ as a $\fr{S}$-module is the rank of $\Lambda$ as a $\Z_p$-lattice. It follows that $M(F^i
\Lambda)/M(F^{i+1} \Lambda)$ is a finite free Breuil–Kisin module of rank $1$.
\[split\] If the filtration on $V$ splits, then so does the filtration on $M(\Lambda)$. Indeed, there is a canonical map $M(F^i \Lambda)/M(F^{i+1} \Lambda)\to
M(F^i \Lambda/F^{i+1} \Lambda)$; by functoriality, we get a map $M(F^i
\Lambda/F^{i+1} \Lambda)$ which splits the exact sequence defining $M(F^i
\Lambda)/M(F^{i+1} \Lambda)$.
The proof of Theorem \[thm4\]
=============================
We recall Theorem \[thm4\] for the reader’s convenience.
Suppose $X_0$ is a smooth proper variety over $K$ with good reduction such that the special fiber of some model $\fr{X}$ of $X_0$ over $\co_K$ has torsion-free crystalline cohomology. If $H^n_\et(X;\QQ_p)$ is an ordinary $\gal{K}$-representation whose filtration splits, then the Newton polygon for $\fr{X}_\sp$ has integer slopes.
Suppose that $X_0$ is as above, and that $V = H^n_\et(X;\QQ_p)$. Since $H^\ast_\cris(\fr{X}_\sp/W(k))$ is torsion-free, we learn from [@bms-i Theorem 1.1(ii)] and [@neron-ogg-k3 Theorem 2.4] that $H^n_\et(X;\Z_p) =: T$ is a $\gal{K}$-stable $\Z_p$-lattice inside $V$. The discussion preceding [@bms-i Theorem 1.4] implies that $$\label{cris-compare}
M(T)\otimes_{\fr{S}} W(k) = H^n_\cris(\fr{X}_\sp/W(k))$$ Suppose $V$ is an ordinary $\gal{K}$-representation whose filtration splits. Then by Lemma \[ordinary-bk\], Remark \[split\], and Equation , we find that $H^n_\cris(\fr{X}_\sp/W(k))$ is a $F$-crystal over $k$ which splits into a direct sum of rank one $F$-crystals. By the Dieudonné–Manin classification, rank one $F$-crystals are of the form $M_{r/1} = W(k)\langle T\rangle/(T = p^r)$ for $r\in \Z_{\geq 0}$. In particular, the Newton polygon for $\fr{X}_\sp$ has integer slopes, as desired.
The slopes of the Hodge polygon of a smooth proper variety are always integers.
This follows from Theorem \[thm4\] and [@bloch-kato Proposition 7.3(7)].
In order to prove Theorem \[thm3\], we will need the following result from [@katz-products]:
\[ordinary-cond\] Let $X_0$ be a smooth proper variety over $K$ with good reduction. Let $\fr{X}$ be a lift of $X_0$ to $\co_K$. Suppose that $H^\ast_\cris(\fr{X}_\sp/W(k))$ is torsion-free. Then $X_0$ is ordinary if and only if the following conditions are satisfied:
- the Hodge–de-Rham spectral sequence $$E_1^{s,t} = H^t(\fr{X}_\sp;\Omega^s_{\fr{X}_\sp/k}) \Rightarrow
H^{s+t}_\mathrm{dR}(\fr{X}_\sp/k)$$ collapses at the $E_1$-page.
- the Newton and Hodge polygons of $H^\ast_\cris(\fr{X}_\sp/W(k))$ coincide.
It is a general fact about K3 surfaces that $H^\ast_\cris(\fr{X}_\sp/W(k))$ is torsion-free. By [@liedtke Proposition 2.5], the conditions of Theorem \[thm4\] are satisfied for every K3 surface having good reduction. By Theorem \[thm4\], we find that the slopes of the Newton polygon of $X_0$ must be integers. Moreover, our hypotheses on $X_0$ imply that it cannot be a supersingular K3 surface. By the Artin–Mazur classification of the heights of K3 surfaces, we conclude that the slopes of the Newton polygon of $X_0$ can be integers if and only if the height of $X_0$ is $1$; this implies that the Hodge and Newton polygons coincide. By Proposition \[ordinary-cond\], we conclude that $X_0$ has potentially good ordinary reduction, as desired.
The same argument gives another proof of Theorem \[thm1\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We review the Raman shift method as a non-destructive optical tool to investigate the thermal conductivity and demonstrate the possibility to map this quantity with a micrometer resolution by studying thin film and bulk materials for thermoelectric applications. In this method, a focused laser beam both thermally excites a sample and undergoes Raman scattering at the excitation spot. The temperature dependence of the phonon energies measured is used as a local thermometer. We discuss that the temperature measured is an effective one and describe how the thermal conductivity is deduced from single temperature measurements to full temperature maps, with the help of analytical or numerical treatments of heat diffusion. We validate the method and its analysis on 3- and 2-dimensional single crystalline samples before applying it to more complex Si-based materials. A suspended thin mesoporous film of phosphorus-doped laser-sintered $\rm{Si}_{\rm{78}}\rm{Ge}_{\rm{22}}$ nanoparticles is investigated to extract the in-plane thermal conductivity from the effective temperatures, measured as a function of the distance to the heat sink. Using an iterative multigrid Gauss-Seidel algorithm the experimental data can be modelled yielding a thermal conductivity of after normalizing by the porosity. As a second application we map the surface of a phosphorus-doped 3-dimensional bulk-nanocrystalline Si sample which exhibits anisotropic and oxygen-rich precipitates. Thermal conductivities as low as are found in the regions of the precipitates, significantly lower than the in the surrounding matrix. The present work serves as a basis to more routinely use the Raman shift method as a versatile tool for thermal conductivity investigations, both for samples with high and low thermal conductivity and in a variety of geometries.'
author:
- 'B. Stoib'
- 'S. Filser'
- 'J. Stötzel'
- 'A. Greppmair'
- 'N. Petermann'
- 'H. Wiggers'
- 'G. Schierning'
- 'M. Stutzmann'
- 'M. S. Brandt'
title: 'Spatially Resolved Determination of Thermal Conductivity by Raman Spectroscopy *accepted in Semicond. Sci. Technol.* '
---
Introduction {#sec:Introduction}
============
In many fields of materials research and development, heat management becomes increasingly important. Both, exceptionally high or low thermal conductivities may be required for optimum device functionality. For example, in microelectronics it is required to efficiently cool integrated circuits to avoid diffusion or electromigration, thus making high thermal conductances on a sub-micrometer scale necessary.[@Tong2011; @Moore2014; @Balandin2009; @Balandin2011] Graphene or isotopically purified crystals have been proposed as useful high thermal conductivity materials for such applications.[@Morelli2002; @Balandin2011] On the other hand, thermoelectric devices or sensors based on micro-calorimetry benefit from materials with a low thermal conductivity, capable of sustaining temperature differences.[@Niklaus2007; @Snyder2008; @Kanatzidis2010; @Nielsch2011; @Schierning2014] To this end, material inhomogeneities on the micro- and nanometer scale can help to efficiently block heat transport by phonons due to wavelength selective scattering.[@Hochbaum2008; @Tang2010; @Biswas2012]
Along with the increasing importance of thermal management, advanced techniques are being developed to experimentally measure thermal conductivities. Standard methods used today include the laser flash method for samples of rather large dimensions and well defined thickness,[@Cape1963] the $3\omega$ method for flat thin films with a good thermal junction to the underlying substrate,[@Borca-Tasciuc2001] micro-electromechanical measurement platforms, e.g., for individual samples of nanowires,[@Voelklein2010] or time or frequency domain thermoreflectance measurements for samples with well defined specular and temperature dependent reflectivity.[@Cahill2004] Hardly any of these techniques are free of challenges, such as limited throughput, unknown heat capacity, rough sample surface, highly diffusive reflection, high electrical conductivity, poorly defined sample thickness or spurious thermal conductance by contacts, substrates or the ambient.[@Tritt2004]
Especially in the regime of materials with low thermal conductivity, micro- and nanostructures offer the possibility to reduce thermal transport.[@Yang2012] Thus, obtaining information on the local thermal conductivity is key to understanding and optimising materials properties, but is also rather demanding. Force microscopy methods are suitable to extract local differences of the thermal conductivities, but the quantification remains difficult.[@Nonnenmacher1992; @Fiege1999; @Meckenstock2008; @Majumdar1999; @Gomes2007; @Zhang2010] Local measurements of the thermal conductivity have also been reported using thermoreflectance methods.[@Huxtable2004; @Zhao2012; @Zheng2007; @Wei2013]
Another optical method which is capable of measuring the thermal conductivity of materials is the Raman shift method, which is also called Raman thermography, micro Raman method or optothermal Raman measurement technique.[@Cai2010; @Lee2011; @Perichon2000; @Balandin2008; @Huang2009; @Soini2010; @Li2009; @Doerk2010; @Balandin2011] Using a strongly focused laser beam, this technique potentially offers a spatial resolution on the micrometer scale. Although this technique based on Raman spectroscopy was applied already to porous low-thermally conducting materials quite a few years ago,[@Perichon2000] it only became popular after the work of Balandin and co-workers for measuring the thermal conductivity of suspended graphene.[@Balandin2008; @Ghosh2009; @Teweldebrhan2010; @Ghosh2010; @Balandin2011; @Chen2012; @Nika2012; @Yan2013] It has now been used by many groups and extended to other materials, such as carbon nanotubes, Si, SiGe, Ge or GaAs.[@Cai2010; @Lee2011; @Liu2013; @Stoib2014; @Soini2010; @Chavez-Angel2014; @Liu2011] The method uses the fact that the energy of Raman active phonon modes usually is dependent on temperature. If this dependence is known, the Raman spectrum obtained contains quantitative information on how strongly the sample was heated by the Raman excitation laser during the measurement, which, for a known excitation power, contains explicit information on the thermal conductance of the structure or device investigated. Together with sufficient knowledge about the sample geometry and the path of heat flow in the sample, it is possible to obtain the thermal conductivity $\kappa$, the material specific intensive quantity of interest.
The present work summarizes the theoretical and analytical basis of the Raman shift method and applies it to some complex structures and sample morphologies. In section \[sec:RamanShift\] we start by discussing in detail how a temperature can be measured by Raman spectroscopy and how it can be simulated numerically. We present the principles of the Raman shift method by means of a one-dimensional model and introduce two-dimensional Raman shift mapping. In section \[sec:ModelSystems\] we obtain $\kappa$ of a homogeneous bulk material and of a thin suspended membrane. These examples are a preparation for section \[sec:ApplicationofRSM\], where we apply the methods presented to structurally more complex systems, such as inhomogeneous bulk-nanocrystalline Si and a thin suspended mesoporous film made from SiGe nanocrystals, before closing with some concluding remarks.
The Principle of the Raman Shift Method {#sec:RamanShift}
=======================================
\[sec:PrincipleRSM\] We start our introduction into the fundamentals of the Raman shift method by a discussion of how a temperature can be determined with Raman spectroscopy. Then, we use an illustrative one-dimensional system to determine the thermal conductivity from such Raman temperature measurements, that serves as a model for our future studies of more complex sample structures.
Measurement of an Effective Raman Temperature {#sec:HowtomeasureT}
---------------------------------------------
In harmonic approximation the energy of atomic vibrations in a solid is determined by the mass of the atoms and by the force constants between the masses. The anharmonicity of the potential leads to a change in the force constants with temperature and usually a crystal *softens* with increasing temperature. In Raman scattering light interacts with these vibrations. Hence, the energy shift $\Delta k$ of Stokes and anti-Stokes scattered light also usually decreases with an increasing temperature of the sample studied.[@Cardona1983] In fact, the Stokes shift follows a distinct material specific dependence on temperature and can thus be used as a non-contact thermometer. The Stokes/anti-Stokes intensity ratio yields similar temperature information,[@Compaan1984a] but is, however, often more difficult to measure.[@Herman2011] As a typical example, the dependence of the Stokes shift $\Delta k$ of crystalline Si is shown in figure \[fig:SiLO\] for the longitudinal optical phonon mode.[@Cowley1965; @Hart1970; @Balkanski1983; @Menendez1984; @Burke1993; @Brazhkin2000; @Doerk2009] The choice of the phonon mode to be evaluated for temperature measurements depends mostly on the signal-to-noise ratio, but may also be influenced by the substrates available when investigating, e.g., thin films, since the Raman signal from the substrate should not interfere. Dependencies similar to figure \[fig:SiLO\] are observed in other solids as well,[@Menendez1984; @Liu1999; @Cui1998; @Li2009a; @Sahoo2013] making the Raman shift method applicable to a large variety of materials systems.
. Copyright 2014, AIP Publishing LLC.[]{data-label="fig:SiLO"}](Fig01_SiLOsmall.pdf)
In the great majority of Raman spectroscopy experiments, the temperature distribution $T(\vec{r})$ is not homogeneous in the sample region where the laser light is Raman scattered. This means that the Raman spectrum collected will contain contributions of hotter (e.g., in the beam centre) and colder (edge of the laser beam) regions of the sample, caused by the inhomogeneous excitation via, e.g., a gaussian laser beam, *and* the thermal conductance of the device studied. Thus, care must be taken when deducing a temperature from a Raman spectrum and the spectrum collected should be interpreted as a weighted average.[@Herman2011; @Liu2011] We will call the temperature deduced from the Stokes shift $\Delta k$ measured an effective Raman temperature $T_{\rm{Raman}}$, to distinguish it from the local temperature $T(\vec{r})$ of the sample. In a very simple approach we assume that every location $\vec{r}$ on the sample contributes to $T_{\rm{Raman}}$ by its local temperature $T(\vec{r})$ weighted by the local excitation power density $H(\vec{r})$. We add up all those contributions in the sample volume and normalize it by the total absorbed laser power $P$ to obtain $$T_{\rm{Raman}}=\frac{1}{P}\int{H\left(\vec{r}\right) T(\vec{r})} c(T(\vec{r})) g(\vec{r})\text{d}\vec{r},\label{eq:weightingcomplicated}$$ where $c(T(\vec{r}))$ is the (in principle temperature dependent) Raman scattering cross section and $g(\vec{r})$ is a function that accounts for the effect that Raman scattering of weakly absorbed light takes place deep in the sample and that such scattered light is less efficiently collected by the objective. In all following calculations and experiments we will assume $c(T)$ to be constant. We further assume full surface near absorption, so that $g(\vec{r})=1$. Then, equation (\[eq:weightingcomplicated\]) simplifies to $$T_{\rm{Raman}}=\frac{1}{P}\int{H(\vec{r}) T(\vec{r})} \text{d}S,\label{eq:weighting}$$ where $\text{d}S$ is a surface element on the sample. This approach to determine $T_{\rm{Raman}}$ does also neither include the line-shape of the Raman signal nor its temperature dependence,[@Liu2011] but nevertheless improves the understanding of the Raman shift method in comparison to most analyses in literature and corrects the effects of different temperatures beneath the laser beam to first order.
![Weighting of the local temperature distribution with the excitation heating power distribution to obtain the Raman temperature $T_{\rm{Raman}}$. Panel (a) shows a colour coded plot of a simulated temperature distribution on a $30 \times \unit{30}{\micro\meter^2}$ grid. A laser beam with a total absorbed power of and a standard deviation of excites a hypothetical 2-dimensional film with $\kappa=\unit{400}{\watt\per\meter\usk\kelvin}$ at $(x|y)$=($|$). At the border of the film, the heat sink forces the temperature to . Panel (b) shows a colour coded plot of the gaussian heating power density. The temperature distribution in panel (a) is the result of the excitation in panel (b). Panel (c) and (d) show cross sections of the colour plots at $y=\unit{15}{\micro\meter}$. Notably, the temperature is not constant in the area of excitation. The effective Raman temperature is a weighted average of the temperature distribution on the surface and the excitation laser power density and is shown in panel (c) by the dashed line.[]{data-label="fig:Faltung"}](Fig02_Faltungonecolumnsmall.pdf)
Figure \[fig:Faltung\] shows an illustrative example of a hypothetical 2-dimensional square sample with $\kappa=\unit{400}{\watt\per\meter\usk\kelvin}$, which is heated by a gaussian laser beam at ($x|y$)=($|$), having a standard deviation of $w=\unit{2}{\micro\meter}$. Panel (a) shows a simulation of the temperature distribution $T(\vec{r})$ which is established in equilibrium on the square sample when exciting with the heating power density $H(\vec{r})$ shown in panel (b). Panel (c) and (d) are sections along the dashed lines in panel (a) and (b), respectively. In the case shown, the Raman temperature according to equation (\[eq:weighting\]) at $(x|y)=(\unit{5}{\micro\meter}|\unit{15}{\micro\meter})$ is $T_{\rm{Raman}}=\unit{313}{\kelvin}$, which is significantly lower than the maximum temperature of .
The temperature distribution in figure \[fig:Faltung\](a) obeys the stationary heat diffusion equation[@Carslaw1986] $$- H(\vec{r})=\kappa(\vec{r}) \Delta T(\vec{r}) + \vec{\nabla}T(\vec{r}) \cdot \vec{\nabla}\kappa(\vec{r}).\label{eq:HDE}$$ It can be used because typically the minimum acquisition time of a Raman spectrum is of the time scale of a second, so that for small scale samples the measurement conditions are close to equilibrium. In equation (\[eq:HDE\]), a locally varying thermal conductivity $\kappa(\vec{r})$, e.g., due to a temperature dependent thermal conductivity, is considered.
For most sample geometries the temperature distribution for a given excitation cannot be calculated analytically. Whenever this is not possible, we use a numerical approach, where the field of interest is discretized in a rectangular grid and the discretized stationary heat diffusion equation is solved on every grid point. As an example we discuss a two-dimensional quadratic grid of dimension $a$, divided into $n$ grid points in each direction, so that one pixel has a width of $h =\frac{a}{n}$. The spatial coordinates $x$ and $y$ can be expressed by two indices $i$ and $j$ $$(x,y) \rightarrow (i \times h,j \times h).$$ Derivatives in equation (\[eq:HDE\]) are expressed in terms of discrete differences, e.g. $$\frac{\partial^2 T}{\partial x^2} \rightarrow \frac{T_{i+1,j}+T_{i-1,j}-2T_{i,j}}{h^2} .$$ The boundary conditions of a constant temperature $T_{\rm{sink}}$ outside of the simulation area and the continuity of heat flow, so that the heat introduced by $H(\vec{r})$ and the heat flowing into the heat sink are the same, are included. A thermal resistance $R_{\rm{th}}$ to the heat sink can also be considered. The problem to be solved can then be written as $$\underline{A} \cdot T_{i,j}=H_{i,j},\label{eq:MatrixFormDiscreteProblem}$$ which is a linear set of equations with the matrix $\underline{A}$ containing all thermal conductivities and contact resistances.
Instead of directly solving equation (\[eq:MatrixFormDiscreteProblem\]), computation speed is enhanced by implementing a solver, based on an iterative Gauss-Seidel algorithm, where the computation effort only scales almost proportionally to the number of grid points.[@Briggs2000] In this algorithm the differential equation is not solved for all grid points simultaneously, but for each grid point in successive cycles, so that the discretized heat equation on each point is solved for $T_{i,j}$ with the values of neighboring points inserted from the previous cycle. This is repeated until a desired accuracy is achieved. Since spatially slowly varying temperature distributions converge only weakly, we use a multigrid algorithm on several grid sizes, first approximating the global temperature distribution on a coarse grid, and then refining this grid by factors of 2 and interpolating the temperature distribution stepwise.[@Briggs2000] Between all steps, Gauss-Seidel iterations are performed. The use of different grid sizes drastically speeds up the convergence of the method.
Determination of the Thermal Conductivity {#sec:Howtoobtainkappa}
-----------------------------------------
By using the example of an effective 1-dimensional bar which is attached to a heat sink at one end, we will now discuss how the thermal conductivity of a material under test can be obtained based in the measurement of $T_{\rm{Raman}}$ introduced above. Figure \[fig:SchemaDidaktik\] schematically shows the focused Raman laser hitting the bar at its end and acting as the heat source. The heat generated at the right end will propagate through the bar to the heat sink on the left. For simplicity, let us assume that $\kappa$ in the bar is neither dependent on temperature nor position. Then, outside the laser beam where $H(x)=0$, equation (\[eq:HDE\]) can be written as $$0=\kappa \frac{\partial^2 T}{\partial x^2}. \label{eq:oneDimHDE}$$ Thus, the temperature decreases linearly from the excitation spot to the heat sink, as shown by the solid line in figure \[fig:SchemaDidaktik\].
![Measuring the thermal conductivity of a bar-shaped material by the Raman shift method. The Raman laser acts both as the heating source and, together with the Raman spectrometer, as the thermometer. The beam of the laser is directed by mirrors (M) to the microscope objective (O), which focuses the light on the sample of length $l$ and cross section $A$. Raman scattered light is directed via a beam splitter (B) to the Raman spectrometer and $T_{\rm{Raman}}$ is measured. For vanishing contact resistance to the heat sink, the temperature distribution drawn as the black solid line is established in equilibrium. The grey dashed line considers a finite contact resistance to the heat sink and a lower thermal conductivity, so that the same Raman temperature would be measured at the end of the bar.[]{data-label="fig:SchemaDidaktik"}](Fig03_SchemaDidaktiksmall.pdf)
To quantitatively obtain the thermal conductivity from equation (\[eq:oneDimHDE\]) and from the experimental value of $T_{\rm{Raman}}$, appropriate boundary conditions have to be set. As already pointed out, the continuity equation requires that the total heat generated at the bar’s right end has to propagate to the heat sink. Neglecting the extension of the laser beam and a thermal contact resistance between the bar and the heat sink, the temperature of the bar at its left end is equal to the temperature of the heat sink $T_{\rm{sink}}$, so that $$P=\frac{A}{l}\kappa \left(T_{\rm{Raman}}-T_{\rm{sink}}\right),\label{eq:oneDimCont}$$ where $P$ is the absorbed power, $A$ is the cross section and $l$ the length of the bar. This directly leads to $$\kappa=\frac{l}{A}\frac{P}{\left(T_{\rm{Raman}}-T_{\rm{sink}}\right)}.\label{eq:kappaOneDim}$$ In the example discussed so far $\kappa$ can be determined by only a single temperature measurement at the right end of the bar. If a thermal contact resistance has to be considered, at least a second temperature measurement needs to be performed at a different spot along the bar and equation (\[eq:kappaOneDim\]) has to be suitably changed. A possible temperature distribution along the bar for the case of a finite contact resistance is shown as a grey dashed line in figure \[fig:SchemaDidaktik\]. Eventually, performing many measurements along the bar, together with modelling the heat transport for the given sample geometry, significantly improves the accuracy of the method.
In general, such spatially resolved temperature measurements can be performed in two dimensions scanning the whole sample surface, leading to what we will call a Raman temperature map. Figure \[fig:MappingDidaktik\] illustrates the generation of such a map by simulation. For each laser position on the sample surface the local temperature distribution $T(\vec{r})$ has to be calculated and weighted with $H(\vec{r})$ to obtain $T_{\rm{Raman}}$ at this spot. Experimentally, at each position a Raman spectrum is collected, and, using a relation such as the one shown in figure \[fig:SiLO\], the corresponding effective temperature is deduced. By mapping the sample, enough information is collected to model both the thermal conductivity and a thermal contact resistance to the heat sink. Because the excitation as well as the temperature measurement are performed with a single laser beam, it is important to note, that such a Raman temperature map is not a temperature distribution, which via Raman scattering could be obtained only by using two lasers.[@Reparaz2014] There, the temperature distribution excited by a strong laser would be probed using a rather weak second laser, keeping the additional heating by the second laser to a minimum.
![Simulation of a Raman temperature map. The laser beam is scanned across a sample and on every position, the effective Raman temperature is obtained by weighting the equilibrium temperature distribution $T(\vec{r})$ (left) with the local heating power density $H(\vec{r})$ of the excitation laser (right).[]{data-label="fig:MappingDidaktik"}](Fig04_MappingDidaktiksmall.pdf)
Model Systems {#sec:ModelSystems}
=============
Before applying the Raman shift method to two material systems relevant for thermoelectrics, we first validate the method using bulk and thin film samples of single-crystalline Si and Ge.
Heat Conduction Into a Semi-Infinite Half Space {#sec:3D}
-----------------------------------------------
The first model system is a homogeneous and semi-infinite bulk material, filling the half-space $z>0$. We want to analyse this system analytically and use cylindrical coordinates $r$, $\phi$ and $z$ to describe it. The Raman laser beam exhibits a radially gaussian shaped excitation power density $$H(r,z=0)=\frac{P}{2\pi w^2} e^{-\frac{r^2}{2 w^2}},$$ with absorbed power $P$ and standard deviation $w$. Here, $r$ is the radius from the center of the beam and $z$ points into the material. The steps presented to deduce the effective Raman temperature in this case are developed following Carslaw and Jaeger.[@Carslaw1986] We assume that the heat supplied by the Raman laser beam is only introduced in the plane $z=0$, which corresponds to a model where the excitation power is strongly absorbed at the surface. Within the material the temperature $T(\vec{r})$ must obey the stationary heat equation in cylindrical coordinates without heat sources $$\frac{\partial^2 T}{\partial r^2}+\frac{1}{r}\frac{\partial T}{\partial r}+ \frac{\partial^2 T}{\partial z^2}=0,\label{eq:RadHDE}$$ which is satisfied by $$T \propto e^{-|\lambda| z} J_0(\lambda r)$$ for any $\lambda$ with $J_0(\lambda r)$ being the Bessel function of first kind and zeroth order. Circular heat flow in direction of the azimuthal angle $\phi$ can be neglected due to the symmetry of the problem. Equation \[eq:RadHDE\] is also satisfied by $$T = \int\limits_0^{\infty} e^{-|\lambda| z} J_0(\lambda r) f(\lambda) \text{d} \lambda,\label{eq:T}$$ where $f(\lambda)$ is chosen to fulfil the boundary conditions. In our problem the Neumann boundary condition is given by the energy flow from the surface into the volume, introduced by the laser power density $H(r,z=0)$, $$-\kappa \left. \frac{\partial T}{\partial z}\right|_{z=0+} = \frac{P}{2\pi w^2} e^{-\frac{r^2}{2 w^2}}.\label{eq:bound}$$ Inserting equation (\[eq:T\]) into equation (\[eq:bound\]) leads to the condition $$\kappa \int\limits_0^{\infty} \lambda J_0(\lambda r) f(\lambda) \text{d} \lambda=\frac{P}{2\pi w^2} e^{-\frac{r^2}{2 w^2}}$$ for $f(\lambda)$. For the solution, the relation $$\int\limits_0^{\infty} x J_0(x r) e^{-\frac{w^2 x^2}{2}} \text{d} x= \frac{1}{w^2}e^{-\frac{r^2}{2w^2}}\label{eq:watson}$$ is needed.[@Watson1995] Therefore, we can insert the function $$f(\lambda)= \frac{P}{2\pi \kappa} e^{-\frac{w^2\lambda^2}{2}}\label{eq:f}$$ into equation (\[eq:T\]), resulting in $$T(r) = \frac{P}{2\pi \kappa} \int \limits_0^{\infty} J_0(\lambda r) e^{-\frac{w^2\lambda^2}{2}} \text{d} \lambda. \label{eq:T2}$$ The effective Raman temperature $T_{\rm{Raman}}$ can then be obtained from equation (\[eq:weighting\]) and (\[eq:watson\]) as $$\begin{aligned}
T_{\rm{Raman}} &= T_{\rm{sink}}+\frac{1}{P} \int\limits_{\phi=0}^{2 \pi} \int\limits_{r=0}^{\infty} T(r) H(r) \text{d} \phi r \text{d} r \nonumber \\
&=T_{\rm{sink}}+ \frac{P}{4\sqrt{\pi} \kappa w}. \label{eq:T3}\end{aligned}$$ For a homogeneous semi-infinite sample, excited by a gaussian shaped laser beam with strong absorption, the spatially constant thermal conductivity $\kappa$ is then given by $$\kappa=\frac{P}{4\sqrt{\pi} \left( T_{\rm{Raman}}-T_{\rm{sink}}\right) w}. \label{eq:kap}$$
To test the validity of equation (\[eq:kap\]) we investigate single-crystalline Si and Ge wafers. All Raman experiments in this work are performed using a Dilor spectrometer equipped with a grating and a liquid nitrogen cooled CCD detector. To map samples, an $x$-$y$ stage is used. An Ar ion laser operating at a wavelength of $\unit{514.5}{\nano\meter}$ excites Raman scattering. Various objectives are used for the micro Raman experiments, and their nearly gaussian spots were characterized by scanning the laser beam across the sharp edge of an evaporated Au film on top of a Si wafer, recording the decreasing Raman intensity of the Si LO mode. The spot width was obtained by deconvolution. For the experiment on the wafers we use a $10\times$ objective with a spot standard deviation of $w=\unit{0.73}{\micro\meter}$. To enhance the accuracy we not only measure a single Raman spectrum for one excitation power, but perform series of measurements with different excitation powers. Then, (\[eq:kap\]) changes to $$\kappa=\frac{\frac{\partial \Delta k}{\partial T}}{4\sqrt{\pi} w\frac{\partial \Delta k}{\partial P}}. \label{eq:kap2}$$ Due to the high thermal conductivity of the single-crystalline wafers, only a small temperature increase of less than is observed during the experiment. Thus, we linearize the relation in figure \[fig:SiLO\] near room temperature and obtain $\frac{\partial \Delta k}{\partial T}=\unit{-0.0214}{cm^{-1}\per K}$ for Si. From the recorded power series on the single-crystalline Si wafer we obtain $\frac{\partial \Delta k}{\partial P}=\unit{-0.0245}{cm^{-1}\per mW}$, yielding $\kappa=\unit{168}{\watt\per\meter\usk\kelvin}$. With this result we only slightly overestimate literature values of $\kappa=\unit{156}{\watt\per\meter\usk\kelvin}$ and $\kappa=\unit{145}{\watt\per\meter\usk\kelvin}$, reported for Si around room temperature by references and , respectively. We have performed a similar experiment on a single-crystalline Ge wafer using $\frac{\partial \Delta k}{\partial P}=\unit{-0.0186}{cm^{-1}\per mW}$ from reference and obtained $\kappa=\unit{49}{\watt\per\meter\usk\kelvin}$, in similarly good agreement with the value of $\unit{60}{\watt\per\meter\usk\kelvin}$ reported in reference
These results show that by using the Raman shift method applying equation (\[eq:kap2\]) one can measure the thermal conductivity of a homogeneous and 3-dimensional material with an accuracy of at least 10%. The assumption of surface-near absorption of the excitation light, made during the deduction of equation (\[eq:kap\]), is fulfilled better for Ge, where the absorption coefficient for light at $\unit{514.5}{\nano\meter}$ is around $\alpha_{\rm{Ge}}=\unit{63\times10^4}{cm^{-1}}$,[@Humlicek1989] compared to Si with literature values around $\alpha_{\rm{Si}}=\unit{2\times10^4}{cm^{-1}}$.[@Humlicek1989; @Sik1998; @Aspnes1983] For a penetration depth in the range of or larger than the excitation laser beam, the effective area through which the heat is introduced into the material is enhanced, so that the thermal conductivity is over-estimated when equation (\[eq:kap\]) is applied. This may explain the tendency for our experiments on Si versus Ge wafers.
In-Plane Conduction of Heat {#sec:2D}
---------------------------
In the previous example the sample investigated was uniform so that Raman spectra taken at different positions on the bulk sample yield identical Raman temperatures. In this section we discuss an example where $T_{\rm{Raman}}$ depends on the position where it is measured due to the fact that although the thermal conductivity $\kappa$ can be expected to be homogeneous, the conductance is not. The sample is a thin and $10\times\unit{10}{mm^2}$ wide membrane of single-crystalline Si, which is carried by a thick Si support at the border. An optical micrograph of the sample in transmission is shown in the inset in figure \[fig:2micrometerSi\]. The membrane is freely suspended on an area of $4.8\times\unit{4.8}{mm^2}$. Using a $10\times$ objective resulting in a spot with a standard deviation of and a laser power of , the solid symbols in figure \[fig:2micrometerSi\] show an experimental Raman temperature scan across the sample, which was measured in vacuum. As soon as the excitation spot is on the freely suspended part of the membrane $T_{\rm{Raman}}$ increases. The heat absorbed in the membrane has to flow in-plane, which increases $T_{\rm{Raman}}$ when the excitation spot is moved away from the underlying support acting as the heat sink. In the center region of the membrane $T_{\rm{Raman}}$ is rather independent of the exact position. The variation of the experimental data in figure \[fig:2micrometerSi\] corresponds to an uncertainty of the determination of the thermal conductivity of the order of $10\%$.
![Raman temperature scan across a thin crystalline Si membrane of thickness. The full symbols show experimentally determined Raman temperatures as the excitation laser beam is scanned along the dashed line in the inset. $T_{\rm{Raman}}$ is increased on the suspended part and only weekly depends on the exact position in the center region. The dashed line is the result of a simulation with $\kappa_{\unit{300}{K}}=\unit{122}{\watt\per\meter\usk\kelvin}$, decreasing with temperature according to a power law with an exponent of $-1.15$[@Asheghi1997]. In the inset the bright region of the transmission optical microscopy image is the freely suspended part of the membrane, whereas in the dark region the membrane is supported by a $\unit{0.5}{mm}$ thick Si substrate.[]{data-label="fig:2micrometerSi"}](Fig05_2micrometerSismall.pdf)
In contrast to the 3-dimensional heat flow problem in section \[sec:3D\], here only 2-dimensional transport in the plane of the thin membrane is taken into account. Although the absorption follows an exponential dependence, the fact that the thickness of the film is of the order of $\alpha_{\rm{Si}}^{-1}$ allows to assume a homogeneous heating independent of the depth in the membrane, so that in our simulation no heat transport perpendicular to the membrane has to be considered. The high ratio of beam diameter and lateral size of the suspended membrane necessitates a large number of grid points in the simulation to correctly cover the temperature distribution at the excitation spot. Assuming a reflectivity of $38\%$,[@Humlicek1989; @Sik1998; @Aspnes1983] neglecting the temperature dependence of $\kappa$ would yield $\kappa=\unit{88}{\watt\per\meter\usk\kelvin}$ (not shown). However, in the temperature range relevant for this measurement and in the regime of thin films with a thickness of the order of micrometer, the thermal conductivity should be modelled by a power law dependence on temperature.[@Asheghi1997] The dashed line in figure \[fig:2micrometerSi\] is the result of our simulation of the Raman temperature across the suspended membrane with a room temperature thermal conductivity of $\kappa_{\unit{300}{K}}=\unit{122}{\watt\per\meter\usk\kelvin}$ and an exponent of approximately $-1.15$.[@Asheghi1997]
In thin films, phonon confinement effects decrease the thermal in-plane conductivity.[@Aksamija2010; @Turney2010; @McGaughey2011; @Tang2011; @Maznev2013] For films of the thickness investigated in our work, confinement is expected to reduce the conductivity at room temperature by a few percent.[@Asheghi1997; @Asheghi1998; @Ju1999; @Liu2005; @Maldovan2011; @Chavez-Angel2014] This is in good agreement with the value of $\kappa$ obtained by our combination of Raman spectroscopy and simulation. In comparison to the bulk value obtained in the previous section this suggests that the Raman shift technique is applicable to thin films similarly well.
Application of the Raman Shift Method {#sec:ApplicationofRSM}
=====================================
In this section we will present applications of the Raman shift method to two Si-based samples, a bulk-like 3-dimensional material of dense nanocrystalline Si with considerable oxygen content and of a porous thin film of $\rm{Si}_{\rm{78}}\rm{Ge}_{\rm{22}}$.[@Stoib2012; @Stoib2013; @Stoib2014; @Stein2011; @Petermann2011; @Kessler2013; @Schierning2014] Both material systems are fabricated from the same type of raw material, which is a powder of gas phase synthesized nanocrystals of Si and SiGe, respectively. Among other possible applications, these materials are of considerable interest within the framework of thermoelectrics.[@Petermann2011; @Stoib2012; @Schierning2014]
First, we show how to obtain $\kappa$ for a thin film which is suspended over a trench. Here, we assume a spatially constant thermal conductivity and set appropriate boundary conditions at the trench edges in the simulation. In such a case, the Raman temperature depends on the distance of the excitation spot to the heat sink as already seen in section \[sec:2D\]. Then, we investigate a bulk-nanocrystalline Si sample with a morphology suggesting a spatially varying $\kappa$ and analyse this with the model of a semi-infinite homogeneous material treated in section \[sec:3D\]. We take equation (\[eq:kap\]) as a basis for the evaluation and attribute different Raman temperatures to different local values of $\kappa$.
Thin Laser-Sintered Nanoparticle Films {#sec:LaserSinteredFilms}
---------------------------------------
. Copyright 2014, AIP Publishing LLC.[]{data-label="fig:SiGebeides"}](Fig06_SiGebeidessmall.pdf)
The thin film sample is a thin mesoporous film of $\rm{Si}_{\rm{78}}\rm{Ge}_{\rm{22}}$. It is fabricated by spin-coating a dispersion of diameter SiGe alloy nanoparticles to obtain films of thickness. The particles are heavily doped with 2% P during their microwave plasma gas synthesis.[@Knipping2004; @Stein2011; @Petermann2011] This high doping level is typical for Si-based thermoelectric materials to optimise the power factor.[@Slack1991; @Dismukes1964; @Snyder2008; @Schierning2014] After removal of the native oxide by hydrofluoric acid, the film is sintered in vacuum by a pulsed Nd:YAG laser operating at $\unit{532}{nm}$ with a fluence of . The resulting mesoporous morphology is shown in a scanning electron microscopy (SEM) image in figure \[fig:SiGebeides\](a). Further information on the fabrication and (thermo-)electric properties of such films can be found in references and .
In-plane measurements of $\kappa$ of as-fabricated thin films are hampered by the significant contribution of the substrate to the thermal conductance. Therefore, the film is transferred onto a support structure. This is a single-crystalline Ge wafer, into which trenches have been etched by reactive ion etching. Germanium was chosen as a heat sink material because its Raman spectrum does not overlap with the Si-Si phonon mode to be investigated. After the transfer, a focused laser beam was scanned along the border of the trench with a high fluence. This compacted the film and ensured a firmer attachment to the heat sink. Figure \[fig:SiGebeides\](b) shows a detail of an SEM side view of such a suspended laser-sintered nanoparticle film.
In the Raman shift experiment of this sample we use an absorbed power of $P_{\rm{absorbed}}=\unit{12}{\micro\watt}$ for excitation and a $20\times$ objective with a spot standard deviation of . Due to the high surface area of the porous film the measurements are carried out in a vacuum chamber of a pressure of $p=\unit{10^{-1}}{mbar}$, which was found to be necessary to rule out spurious thermal conductance by contact to the surrounding ambient gas atmosphere. Because of the high Si content in the alloy the phonon mode that was used to extract $T_{\rm{Raman}}$ was the Si-Si vibration. Its temperature dependence in such SiGe alloys is very similar to that in pure Si and can be linearized in the region from room temperature to by $\frac{\partial \Delta k}{\partial T}=\unit{-0.0229}{cm^{-1}\per K}$.[@Burke1993]
A Raman temperature map of the film with the suspended part in the center and the trench in $y$ direction is shown in figure \[fig:SiGebeides\](c). In the suspended part of the film, the Raman temperature is as high as , whereas the part of the film which is in direct contact with the supporting Ge wafer can efficiently conduct the heat introduced to the underlying heat sink, so that the Raman temperature stays close to room temperature. In figure \[fig:SiGebeides\](d), the same data are shown in a more quantitative way. For different scans across the trench, the central Raman temperature varies by approximately 10%, which gives the same estimate for the accuracy of determining $\kappa$ by the Raman shift method as discussed in section \[sec:ModelSystems\]. We also plot the mean value of all scans as filled squares, which is quite symmetric with respect to the center axis of the trench.
In contrast to the measurement shown in figure \[fig:2micrometerSi\], where the suspended part was wide in $x$ and $y$ direction, the suspended part of the laser-sintered thin film has a width of only (trench width) and a length of more than . Thus, the simulation grid used had an aspect ratio of 3:1, which was found to be of sufficient accuracy, neglecting the small fraction of heat transport in $y$ direction.
The complete simulated profile of Raman temperatures across the trench assuming a temperature independent value of $\kappa$ is shown as solid line in figure \[fig:SiGebeides\](d) and describes well the mean value of the experimental data within their experimental variation. In our simulation the porous film is treated as an effective medium and the solid line corresponds to an effective in-plane thermal conductivity of $\kappa_{\rm{eff}}=\unit{0.05}{\watt\per\meter\usk\kelvin}$ and a negligible contact resistance. In small grained and doped SiGe alloy thin films, it is justified to neglect the temperature dependence of the thermal conductivity,[@Stein2011; @Steigmeier1964; @McConnell2001] so that despite the large temperature differences in our experiment we obtain useful values for $\kappa_{\rm{eff}}$. Normalizing $\kappa_{\rm{eff}}$ by a factor $(1-\rm{porosity})$ with a typical porosity of 50% for these laser-sintered thin films, we obtain the in-plane thermal conductivity usually given in literature.[@Boor2011; @Tang2010] In the present case, this yields $\kappa_{\rm{normalized}}=\unit{0.1}{\watt\per\meter\usk\kelvin}$. Estimating all uncertainties entering the simulation we obtain a maximum thermal conductivity of $\kappa_{\rm{normalized}}^{\rm{max}}=\unit{0.3}{\watt\per\meter\usk\kelvin}$.
It is generally believed that the mean free path of phonons is drastically reduced in materials with a hierarchy of scattering centers. The present sample exhibits such a disorder of different length scales, starting at the atomic scale due to alloy scattering in the SiGe alloy. Also some nanocrystals with a diameter of the order of survived the sintering process and are incorporated in the matrix. Typical for this type of laser-sintered material, grain boundaries between the grains of typically constitute larger scattering centers. For long wavelength phonons the mesoporous structure with typical structure sizes of is the relevant scatterer. In the laser-sintered mesoporous thin films a reduction of $\kappa$ by a factor of approximately 10-20 is observed, compared to nanograined but dense SiGe materials.[@Wang2008; @Stein2011] Our value for $\kappa$ in mesoporous n-type doped SiGe is approximately a factor of 20 lower. Most likely percolation effects, which were intensively studied in pure Si materials and also affect electrical transport, are responsible for this additional reduction.[@Boor2011; @Tang2010]
Bulk-Nanocrystalline Silicon {#sec:NCSi}
----------------------------
In this second application of the Raman shift method to Si-based materials, we investigate the local variation of $\kappa$ for bulk-nanocrystalline Si. The sample studied is synthesized from a powder of microwave plasma grown Si nanoparticles, which are doped with 1% P in the gas phase and have a diameter of $22-\unit{25}{\nano\meter}$. The Si nanocrystal powder used for this sample was exposed on purpose to ambient oxygen for three weeks to obtain a significant oxygen content known to impact the microstructure. The powder was then pre-compacted and solidified by current-activated pressure-assisted densification, resulting in a slight increase in crystallite size to approximately .[@Petermann2011; @Stein2011; @Schierning2011] The direction of current in this sintering method leads to an anisotropy of the resulting material.[@Meseth2012] During densification, oxygen relocates within the nanoparticle network and forms mainly two types of oxygen-rich precipitates, small and rather spherical precipitates of approximately in size and larger agglomerates of such small precipitates forming larger structures of $\rm{SiO}_x$.[@Schierning2011; @Meseth2012] The latter are shaped like a disc, with their axis pointing in the direction of the sinter current, and have diameters of several tens of microns and a thickness of approximately . The enriched oxygen content in the larger precipitates is accompanied by an enhanced porosity in this region.[@Schierning2011] Both, the different elemental composition and the different microstructure of the precipitates compared to the surrounding matrix, suggest a non-uniform thermal conductivity of the material. After densification, the sample investigated here was cut and polished by ion milling, so that the surface was flat on a tens of nanometer scale. Figure \[fig:SkizzeDiffRamanLaserFlash\] illustrates the orientation of the precipitates within the sample investigated. The Raman experiments were carried out on the polished top surface. Additional laser flash measurements of $\kappa$ were conducted from the orthogonal direction, due to geometrical restrictions of the sample. The direction of the sinter current was parallel to the direction of laser flash measurements.
![Differences of the measurement geometries of the Raman shift method and the laser flash method, applied to bulk-nanocrystalline Si. The oxygen-rich areas of precipitates (grey) are disc shaped and lie perpendicularly to the laser flash measurement direction. For the microscopic Raman shift method, these precipitates play a less important role as barriers for thermal transport. The direction of the sinter current was the same as for the laser flash measurement.[]{data-label="fig:SkizzeDiffRamanLaserFlash"}](Fig07_RamanvsLaserFlash.pdf){width="30.00000%"}
The investigation of local variations of the thermal conductivity of this sample is based on the following procedure: Applying the Raman shift method, we first extract a Raman temperature map. Using an incident laser power of the sample is partly heated up to , so that we use $\frac{\partial \Delta k}{\partial T}=\unit{-0.0255}{cm^{-1}\per K}$ as a linear interpolation in figure \[fig:SiLO\]. The high signal-to-noise ratio in the Raman experiments allows to include the contribution of free charge carriers, introduced by the high amount of P and the strong illumination, in the evaluation of the Raman spectrum. Therefore, in contrast to the other experiments of this work, the Raman temperature is not determined experimentally from the maximum of the Raman line, but rather from a fit of a Fano lineshape to the spectra.[@Cerdeira1972; @Chandrasekhar1978] Although the material is not homogeneous, we assume it to be homogeneous in the near field of the excitation laser beam. In these study we use a $100\times$ objective resulting in a gaussian beam of standard deviation, which is much smaller than average distances of the oxygen-rich precipitates. We again assume a reflectivity of $38\%$.[@Humlicek1989; @Sik1998; @Aspnes1983] By using equation (\[eq:kap\]) we then calculate a map of local thermal conductivities.
Figure \[fig:Duisburg\] shows such a map of the thermal conductivity, using a different colour scale than in the maps of $T_{\rm{Raman}}$ discussed earlier, and the corresponding microstructure of the bulk-nanocrystalline Si sample as observed by SEM. In panel (a) an overview map of $\kappa$ is shown. The map exhibits anisotropic structures which are elongated in $y$ direction and have a lower thermal conductivity compared to the surrounding Si matrix. The dashed rectangle in (a) is shown in panel (b) with a higher resolution compared to panel (a). Structures on the length scale of a micrometer can be discerned, which demonstrates that the measurement is capable to detect local variations in $\kappa$ close to the resolution limit given by the spot size. As a guide to the eye, the green dot in the lower right corner illustrates the full gaussian width of the laser excitation spot.
The thermal conductivities obtained by the Raman shift experiment are in the range between 11 and . This is an order of magnitude lower compared to the values reported for undoped single-crystalline Si.[@Glassbrenner1964; @Maycock1967] The extremely high content of P and the small grained nanostructure on a scale of resulting from sintering the small nanoparticles can be made responsible for this reduced thermal conductivity.[@Petermann2011; @Schierning2014; @Stein2011; @Schwesig2011] The thermal conductivity of the very sample investigated here has also been characterized as a function of temperature using the laser flash method. At room temperature the laser flash method yields a thermal conductivity of $\kappa=\unit{9.5}{\watt\per\meter\usk\kelvin}$, which decreases to $\kappa=\unit{6.5}{\watt\per\meter\usk\kelvin}$ at . Thus, the temperature dependence is not pronounced, justifying the neglect of a temperature dependence of $\kappa$ when deducing equation (\[eq:kap\]) also for this type of sample. However, the values obtained for $\kappa$ obtained by the laser flash method are roughly a factor of 2 lower, compared to the results obtained by the Raman shift method. The most likely reason for this difference is the measurement geometry. As sketched in figure \[fig:SkizzeDiffRamanLaserFlash\], for the laser flash measurements the heat flow was perpendicular to the disc shaped precipitates, making them a maximum barrier for heat transport. In the Raman shift method, the heat is spread radially into the material, with heat transport suffering only little from the alignment of the precipitates. Further reasons for the slightly different result are spurious thermal conduction by air during the measurements and the finite absorption coefficient of Si at the wavelength used, which leads to a slight over-estimation of $\kappa$ using equation (\[eq:kap\]) as discussed before.
To attribute the local variations in $\kappa$ observed in this material to structural features, we show SEM micrographs of the areas investigated by the Raman shift method in panel (c) to (e) of figure \[fig:Duisburg\]. Panel (c) shows the region investigated in panel (b). The large structure on the right half of the panel can clearly be recovered in the SEM image. Also the smaller feature in the upper left corner of panel (b) can be found in panel (c), and is magnified in panel (d). In contrast to the surrounding area, the surface of this feature is less flat and shows a porous interior. The same conclusion can be drawn from panel (e), which shows the second rectangular area marked with dashed lines in panel (b). A similar porosity as in the small feature can be found here. Energy dispersive X-ray scans across the structure in panel (e) confirm that the oxygen content in the porous region is enhanced by at least a factor of 4.[@Meseth2012] Correlating the SEM image in panel (e) to the thermal conductivity map in panel (b) suggests that the porous regions clearly visible in SEM exhibit a lower thermal conductivity compared to the surrounding area. At least in principle, this apparently lower thermal conductivity could arise from the local increase of the absorbed laser power, which in turn could be caused by the roughness of the surface visible in the SEM micrographs.[@Algasinger2013] However, since strong variations in $\kappa$ are also found for flat parts of the bulk-nanocrystalline sample studied, it can be concluded that the contrast in the maps of thermal conductivity originates to a significant part from the locally varying thermal conductivity.
Summary and Conclusion {#sec:Conclusion}
======================
We showed that by performing a micro Raman scattering experiment where the laser simultaneously acts as a thermal excitation source and as a thermometer, using the temperature dependence of the energy of Raman active phonon modes, one can determine the thermal conductance of a specimen. Knowing or simulating the geometry of heat propagation from the excitation spot to the heat sink is key to obtaining reliable data on the thermal conductivity. We discussed that it is necessary to take the non-homogeneous temperature distribution beneath the excitation spot into account to correctly interpret the effective temperature deduced from the Raman spectrum. Applying this Raman shift method to both 3-dimensional heat flow into a semi-infinite homogeneous material and to 2-dimensional heat transport in a suspended thin film we experimentally validated the technique and its analysis. Finally, we used the Raman shift method to determine the thermal in-plane conductivity for laser-sintered thin films of $\rm{Si}_{\rm{78}}\rm{Ge}_{\rm{22}}$ nanoparticles. Assuming a spatially constant $\kappa$ and attributing an increased temperature solely to the locally varying distance to the heat sink, we demonstrated that the Raman shift method can measure porosity-normalized values of the thermal conductivity as low as $\unit{0.1}{\watt\per\meter\usk\kelvin}$. As a second application, we investigated local variations of the thermal conductivity of a 3-dimensional bulk-nanocrystalline Si sample exhibiting microscopic $\rm{SiO}_x$ precipitates in SEM investigations. Here, the Raman shift method is able to measure local variations of the thermal conductivity by more than 40% between oxygen-rich porous regions and dense regions with reduced oxygen content with a spatial resolution of the spot size of the exciting laser beam.
Obtaining reliable quantitative information on the local thermal conductivity requires sound knowledge on three major parameters, which are the intensity profile of the exciting laser beam, the geometry of thermal transport and the absorbed optical heating power. The latter turns out to be the most critical parameter for samples with complex microstructure and can be challenging to determine. Rough surfaces or porous materials, often accompanied by a spatial variation of the elemental composition, can require to base the evaluation of the results of the Raman shift method on assumptions on absorption coefficients and reflectivities, since the direct measurement of reflected and transmitted excitation power is difficult in many sample geometries. The vector field of heat propagation can only be calculated in very rare cases analytically. Therefore, numerical simulations need to be performed to solve the heat diffusion equation which involves considerable computation effort, especially when the problem cannot be reduced to two dimensions. Finally, although the intensity profile of the laser beam used can easily be accessed experimentally, the implications of a spatially inhomogeneous excitation combined with a non-homogeneous temperature distribution on the resulting Raman spectrum measured can be manifold. This includes Raman scattering cross sections, line shapes, absorption profiles or collection efficiencies. However, a set of reasonable assumptions can make the Raman shift method a straight forward method.
This study demonstrated that the variety of materials systems and sample geometries that can be investigated by the Raman shift method without mechanical contact makes the method a versatile and powerful tool to obtain thermal information of small scale and complex materials systems. With that, the method complements well more traditional and established tools and enables insight into thermal transport on a micrometer scale.
Acknowledgments {#acknowledgments .unnumbered}
===============
We acknowledge funding by the German Research Foundation DFG via the priority program SPP 1386 “Nanostructured Thermoelectrics” and additional support by the Bavarian State Ministry of the Environment and Consumer Protection via the project “Umwelt Nanotech”.
[98]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [**]{}, ed. (, , ) [****, ()](\doibase http://dx.doi.org/10.1016/j.mattod.2014.04.003) @noop [**** ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevB.66.195304) [****, ()](\doibase 10.1117/12.755128) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1002/aenm.201100207) [****, ()](\doibase 10.1002/pssa.201300408) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1063/1.1729711) [****, ()](\doibase http://dx.doi.org/10.1063/1.1353189) @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1063/1.1819431) @noop [**]{} (, , ) [****, ()](\doibase
http://dx.doi.org/10.1063/1.4773462) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1063/1.2908445) @noop [****, ()]{} [****, ()](http://stacks.iop.org/0022-3727/40/i=21/a=029) [****, ()](\doibase http://dx.doi.org/10.1063/1.3300826) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1016/j.actamat.2007.05.037) [****, ()](\doibase
http://dx.doi.org/10.1063/1.4815867) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1063/1.3532848) @noop [****, ()]{} @noop [****, ()]{} [****, ()](http://stacks.iop.org/1367-2630/11/i=9/a=095012) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](http://stacks.iop.org/0953-8984/24/i=23/a=233203) [****, ()](\doibase
http://dx.doi.org/10.1063/1.4833250) [****, ()](\doibase
http://dx.doi.org/10.1063/1.4801495) [****, ()](\doibase
http://dx.doi.org/10.1063/1.4873539) @noop [****, ()]{} [****, ()](\doibase
http://dx.doi.org/10.1063/1.3583603) , ed., @noop [**]{} (, , ) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1051/jphys:019650026011065900) [****, ()](\doibase 10.1103/PhysRevB.1.638) [****, ()](\doibase 10.1103/PhysRevB.28.1928) [****, ()](\doibase 10.1103/PhysRevB.29.2051) [****, ()](\doibase 10.1103/PhysRevB.48.15016) [****, ()](\doibase
10.1134/1.1320111) [****, ()](\doibase 10.1103/PhysRevB.80.073306) [****, ()](\doibase
http://dx.doi.org/10.1063/1.124083) [****, ()](\doibase
http://dx.doi.org/10.1063/1.367972) [****, ()](\doibase 10.1103/PhysRevB.80.054304) @noop [****, ()]{} @noop [**]{} (, , ) @noop [**]{}, ed. (, , ) [****, ()](\doibase http://dx.doi.org/10.1063/1.4867166) @noop [**]{} (, , ) [****, ()](\doibase 10.1103/PhysRev.134.A1058) [****, ()](\doibase http://dx.doi.org/10.1016/0038-1101(67)90069-X) [****, ()](\doibase http://dx.doi.org/10.1063/1.342720) [****, ()](\doibase http://dx.doi.org/10.1063/1.368951) [****, ()](\doibase 10.1103/PhysRevB.27.985) [****, ()](\doibase 10.1063/1.119402) @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1063/1.3296394) [****, ()](\doibase http://dx.doi.org/10.1063/1.3644163) [****, ()](\doibase
http://dx.doi.org/10.1063/1.3622317) [****, ()](\doibase http://dx.doi.org/10.1063/1.4795601) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1063/1.2149497) [****, ()](\doibase http://dx.doi.org/10.1063/1.3607295) [****, ()](\doibase 10.1063/1.4726041) [****, ()](\doibase 10.1002/pssa.201228392) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1002/adem.201200233) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRev.136.A1149) [****, ()](\doibase 10.1109/84.946782) @noop [****, ()]{} [****, ()](\doibase http://dx.doi.org/10.1063/1.3027060) [****, ()](\doibase 10.1063/1.3658021) [****, ()](\doibase
http://dx.doi.org/10.1016/j.scriptamat.2012.04.039) [****, ()](\doibase 10.1103/PhysRevB.5.1440) [****, ()](\doibase 10.1103/PhysRevB.17.1623) @noop [****, ()]{} [****, ()](\doibase
10.1002/aenm.201201038)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[**The shape of urban settlements plays a fundamental role in their sustainable planning. Properly defining the boundaries of cities is challenging and remains an open problem in the Science of Cities. Here, we propose a worldwide model to define urban settlements beyond their administrative boundaries through a bottom-up approach that takes into account geographical biases intrinsically associated with most societies around the world, and reflected in their different regional growing dynamics. The generality of the model allows to study the scaling laws of cities at all geographical levels: countries, continents, and the entire world. Our definition of cities is robust and holds to one of the most famous results in Social Sciences: Zipf’s law. According to our results, the largest cities in the world are not in line with what was recently reported by the United Nations. For example, we find that the largest city in the world is an agglomeration of several small settlements close to each other, connecting three large settlements: Alexandria, Cairo, and Luxor. Our definition of cities opens the doors to the study of the economy of cities in a systematic way independently of arbitrary definitions that employ administrative boundaries.**]{}'
author:
- |
Erneson A. Oliveira$^{1,2,3}\footnote{Correspondence to:
[email protected]}$, Vasco Furtado$^1$, José S. Andrade Jr.$^{3}$, Hernán A. Makse$^{3,4}$
title: A worldwide model for boundaries of urban settlements
---
Introduction
============
What are cities? In [*The Death and Life of the Great American Cities*]{}, Jacobs argues that human relations can be seen as a proxy for places within cities [@jacobs1961]. A modern view of cities establishes that they can be defined by the interactions among several types of networks [@batty2013; @west2017], from infrastructure networks to social networks. In recent years, an increasing number of studies have been proposed to define cities through consistent mathematical models [@makse1995; @makse1998; @rozenfeld2008; @rozenfeld2011; @rybski2013; @murcio2013; @hernando2014; @frasco2014; @arcaute2014; @masucci2015; @arcaute2016; @fluschnik2016] and to investigate urban indicators at inter- and intra-city scales, in order to shed some light on problems faced by decision makers [@bettencourt2007; @arbesman2009; @bettencourt2010a; @bettencourt2010b; @gomez-lievano2012; @gallos2012a; @bettencourt2013; @fragkias2013; @oliveira2014; @melo2014; @louf2014; @alves2015; @li2015; @hanley2016; @caminha2017; @operti2017]. Despite the efforts of such studies, to properly defining the boundaries of urban settlements remains an open problem in the Science of Cities. A minimum criterion of acceptability for any model of cities seems to be the one that retrieves a conspicuous scaling law found for United States (US), United Kingdom (UK), and other countries, known as Zipf’s law [@rozenfeld2008; @rozenfeld2011; @gabaix1999a; @gabaix1999b; @ioannides2003; @cordoba2008; @giesen2010; @giesen2011; @jiang2011; @giesen2012; @ioannides2013; @duraton2014; @watanabe2015].
In 1949, Zipf [@zipf1949] observed that the frequency of words used in the English language obeys a natural and robust power law behavior, [*i.e.*]{} a few words are used many times, while many words are used just a few times. Zipf’s law can be represented generically by the following relation between the size $S$ of objects from a given set and its rank $R$: $$\label{eq1}
R \propto S^{-\zeta},$$
where $\zeta=1$ is Zipf’s exponent. The size of objects is, in the original context, the frequency of used words. On the other hand, if such objects are cities, then the sizes stand for the population of each city, taking into account Zipf’s law and reflecting the fact that there are more small towns than metropolises in the world. We emphasize that it is not straightforward that the Zipf’s law, despite its robustness, should hold independently of the city definition, since other scaling relations are not, such as the allometric exponents for CO$_2$ emissions and light pollution [@oliveira2014; @operti2017]. Many other man-made and natural phenomena also exhibit the same persistent result, [*e.g.*]{} earthquakes and incomes [@sornette1996; @okuyama1999].
Here, we propose a worldwide model to define urban settlements beyond their usual administrative boundaries through a bottom-up approach that takes into account cultural, political, and geographical biases naturally embedded in the population distribution of continental areas. After all, it is not surprising that two regions, [*e.g.*]{} one in western Europe and another one in eastern Asia, spatially contiguous in population or in commuting level have different cultural, political or geographical characteristics. Thus, it is also not surprising that such issues yield different stages of the same mechanics of growth. The main goal of our model is to be successful in defining cities even in large regions. Our conjecture is straightforward: there are hierarchical mechanisms, similar to those present in previous studies of cities in the UK [@arcaute2016] and brain networks [@gallos2012b], behind the growth and innovation of urban settlements. These mechanisms are ruled by a combination of general measures, such as the population and the area of each city, and intrinsic factors which are specific to each region, [*e.g.*]{} topographical heterogeneity, political and economic issues, and cultural customs and traditions. In other words, if political turmoil or economic recession plagues a metropolis for a long time, all of its satellites are affected too, [*i.e.*]{} the entire region ruled by the metropolis will be negatively impacted.
The models
==========
City Clustering Algorithm (CCA)
-------------------------------
In 2008, Rozenfeld [*et al.*]{} [@rozenfeld2008] proposed a model to define cities beyond their usual administrative boundaries using a notion of spatial continuity of urban settlements, called the City Clustering Algorithm (CCA) [@rozenfeld2008; @rozenfeld2011; @rybski2013; @oliveira2014; @frasco2014; @fluschnik2016; @caminha2017; @operti2017]. The CCA is defined for discrete or continuous landscapes [@rozenfeld2011] by two parameters: a population density threshold $D^*$ and a distance threshold $\ell$. These parameters describe the populated areas and the commuting distance between areas, respectively. Here, we adopt the following strategy to improve the discrete CCA performance: (i) Supposing a regular rectangular lattice $L_x \times L_y$ of sites where the population density of the $k$-th site is $D_k$, we perform an initial agglomeration by $D^*$ to identify all clusters. If $D_k>D^*$, then the $k$-th site is populated and we aggregate it with its populated nearest neighbors. Otherwise, the $k$-th site is unpopulated. (ii) For each populated cluster, we define its [*shell sites*]{}, [*i.e.*]{} sites in the interface between populated and unpopulated areas. (iii) Lastly, we perform a final agglomeration by $\ell$, taking into account only the shell sites. If $d_{ij}<\ell$, where $d_{ij}$ is the distance between the $i$-th and $j$-th shell sites, and if they belong to different clusters, then the $i$-th and $j$-th sites belong to the same CCA cluster, even with spatial discontinuity. Otherwise, they indeed belong to different CCA clusters. This simple strategy improves the algorithm’s computational performance because the number of shell sites is proportional to $L$, where $L=L_x \approx L_y$ is a linear measure of the lattice.
City Local Clustering Algorithm (CLCA)
--------------------------------------
We propose a worldwide model based on the CCA, called the City Local Clustering Algorithm (CLCA), not only to define cities beyond their usual administrative boundaries, but also to take into account the intrinsic cultural, political and geographical biases associated with most societies and reflected in their particular growing dynamics. The traditional CCA, with fixed $\ell$ and $D^*$, when applied to a large population density map, can introduce biases defining a lot of clusters in some regions, while in others just a few. We present the CLCA with the aim of defining cities even in large regions in order to overcome such CCA weakness. Hence, it is possible that other models, such as the models based on street networks proposed by Masucci [*et al.*]{} [@masucci2015] and Arcaute [*et al.*]{} [@arcaute2016], carry the same CCA burden and that local adaptations are necessary for their applications into large regions.
The main idea of our model is to analyze the change of the CCA clusters through the variation of $D^*$ under the perspective of different regions. First, we define a regular rectangular lattice $L_x \times L_y$ of sites, where the population density of the $k$-th site is $D_k$. We sort all the sites in a list according to the population density, in descending order. Therefore, the site with the greatest population density is the first entry in this list, which we call the first [*reference site*]{}. The reference site can be considered as the current core of the analyzed region. Second, we apply the CCA to the lattice, keeping a fixed value of $\ell$, for a range of $D^*$ decreasing from a maximum value $D^{(max)}$ to a minimum value $D^{(min)}$ with a decrement $\delta$. During the decreasing of $D^*$, clusters are formed and they spread out to all regions of the lattice. Eventually, the cluster that contains the reference site (from now on the [*reference cluster*]{}), together with one or more of the other clusters, will merge from $D^{(i)}$ to $D^{(i+1)}$, where $D^{(i+1)} = D^{(i)}-\delta$. In order to accept or deny the merging of these clusters, we introduce three conditions:
(i) If the area $A_r(D^{(i)})$ of the reference cluster $r$, [*i.e.*]{} the cluster that contains the $r$-th reference site at $D^{(i)}$, obeys $$\label{eq2}
A_r(D^{(i)}) < A^*,$$
then the reference cluster $r$ always merges with other clusters, because it is still considered very small. In this context, the area $A^*$ can be understood as the minimal area of a metropolis.
(ii) If the difference between the areas of the reference cluster $r$ at $D^{(i+1)}$ and $D^{(i)}$ obeys $$\label{eq3}
A_r(D^{(i+1)})-A_r(D^{(i)}) > H^* A_r(D^{(i)}),$$
then the reference cluster $r$ has grown without merging (Fig.\[clca\]a) or there is a merging of at least two large clusters (Fig.\[clca\]b). In the last case, we emphasize that if there are more than two clusters involved in the merging process, the reference cluster $r$ may not be one of the largest. As the first case is not desirable, we can avoid it by reducing the value of $\delta$ and keeping the value of $H^*$ relatively high. The parameter $H^*$ can be understood as the percentage of the area of the reference cluster $r$ at $D^{(i)}$. If the second case happens, we consider the entire region inside of the reference cluster $r$ at $D^{(i+1)}$, but the clusters of this region (which we call the [*usual clusters*]{}) are defined by those at $D^{(i)}$. The usual clusters are the CCA clusters at the imminence of the merging process between $D^{(i)}$ and $D^{(i+1)}$. This includes the reference cluster $r$ itself and one or more of the other clusters before the merging (Fig.\[clca\]b). Furthermore, all of the sites of the reference cluster $r$ at $D^{(i+1)}$ are removed from the initial list of reference sites. This condition is necessary because we should not merge two large metropolises.
(iii) In condition (ii), when a reference cluster $r$ is merging with another cluster that covers one or more regions already defined by previous reference clusters at different values of $D^*$, there is a strong likelihood of the emergence of a [*forbidden region*]{} within that cluster. In this case, we force the region already defined by the largest value of $D^*$ to grow to the limits of the forbidden region (Fig.\[clca\]c). The forbidden regions are the complementary areas of the reference clusters already defined within the usual clusters. As a consequence of this procedure, some CCA clusters that were hidden after the analysis of the previous reference cluster arise in this forbidden region. We justify this condition by the idea that a metropolis rules the growth of its satellites, since it plays a fundamental role in their socioeconomic relations.
We apply the same procedure to the second reference cluster, to the third reference cluster, and so on. Finally, we also define the [*isolated clusters*]{} with the minimum value of $D^*$ for all the cases accepted in condition (ii). In order to make our model clearer, we chose the descending order to sort the population density for one reason: To favor the merging process of the high density clusters that rose from the decreasing of the $D^*$. In practice, we run our revised discrete CCA just once for the entire range of input parameters and store all of the outputs in order to improve the performance of the model. The apparent simplicity of this task hides a RAM memory management problem of storing all of the outputs in a medium-performance computer. We overcome such a barrier through the [*zram module*]{} [@zram2017], available in the newest linux kernels. The zram module creates blocks which compress and store information dynamically in the RAM memory itself, at the cost of processing time.
The dataset
===========
We use the [*Global Rural-Urban Mapping Project*]{} (GRUMPv1) [@grump2011], available from the [*Socioeconomic Data and Applications Center*]{} (SEDAC) at Columbia University, to apply the CLCA to a single global dataset. The GRUMPv1 dataset is composed of georeferenced rectangular population grids for 232 countries around the world in the year 2000 (Fig. \[grumpv1\]). Such a dataset is a compilation of gridded census and satellite data for the populations of urban and rural areas. These data are provided at a high resolution of $30\;\text{arc-seconds}$, equivalent to $30/3600\;\text{degrees}$ or a grid of $0.926\;km \times 0.926\;km$ at the Equator. We note that despite of the heterogeneous population distributions that built the GRUMPv1, its overall resolution is tolerable to the CLCA, since we can identify well-defined clusters around all continents in the raw data.
We calculate the area of each site by the composition of two spherical triangles [@snyder1987]. The area of a spherical triangle with edges $a$, $b$ and $c$ is given by $$\label{eq4}
A=4 R_e^2 \tan^{-1}\left[\tan\left(\frac{s}{2}\right)\tan\left(\frac{s_a}{2}\right)\tan\left(\frac{s_b}{2}\right)\tan\left(\frac{s_c}{2}\right)\right]^{1/2},$$
where $s=(a/R_e+b/R_e+c/R_e)/2$, $s_a=s-a/R_e$, $s_b=s-b/R_e$, and $s_c=s-c/R_e$. In this formalism, $R_e=6,378.137\;km$ is the Earth’s radius and the edge lengths are calculated by the great circle (geodesic) distance between two points $i$ and $j$ on the Earth’s surface: $$\label{eq5}
d_{ij}=R_e\cos^{-1}[\sin(\phi_i) \sin(\phi_j)+\cos(\phi_i)\cos(\phi_j)\cos(\lambda_j-\lambda_i)].$$
The values of $\lambda_i$ ($\lambda_j$) and $\phi_i$ ($\phi_j)$, measured in radians, are the longitude and latitude, respectively, of the point $i$ ($j$). Thus, we are able to define the population density for each site of the lattice, since its population and area are known.
We also pre-process the GRUMPv1 dataset, dividing all countries and continents — and even the entire world — into large regions which we call [*clusters of regions*]{}, to apply our model in a feasible computational time using medium-performance computers. These regions are defined by the CCA with lower and upper bound parameters $D^*=50\;\text{people}/km^2$ and $\ell=10\;km$, respectively. We believe that such large clusters can hold the socioeconomic and cultural relations among different urban settlements of a territory. Fig. \[us\_east\]a shows the largest clusters of regions in the US; as we can see, all of the Eastern US is considered a single cluster.
Results
=======
To show the relevance of our model, we apply the CLCA to the GRUMPv1 dataset at three different geographic levels: countries, continents and the entire world. For each case, we consider only a single set of CLCA parameters. We justify our choices with the following assumptions: (i) $D^{(min)}=100\;\text{people}/km^2$, a value slightly greater than the lower bound CCA parameter ($D^*=50\;\text{people}/km^2$) used to define the regions of clusters; (ii) $D^{(max)}=1000\;\text{people}/km^2$, a loosened value of $D^{(max)}=\infty$; (iii) $\delta=10\;\text{people}/km^2$, a small enough value to avoid the reference clusters growing without merging; (iv) $\ell=3\;km$, the critical distance threshold, already extensively analyzed by previous CCA studies [@rozenfeld2008; @rozenfeld2011; @oliveira2014]; (v) $A^*=50\;km^2$, the minimum area of a metropolis, as it is required that $A^*$ be reasonably greater than the minimum unit of area from the dataset and smaller than a metropolis’ area; and (vi) $H^*=0.05$, a large enough value to favor the merging of clusters which are similar in size. The Fig. \[us\_east\]b shows the CLCA cities defined by the single set of CLCA parameters. For other regions, see the Supplementary Information (SI).
We study the population distribution using the Maximum Likelihood Estimator (MLE) proposed by Clauset [*et al.*]{} [@clauset2009]. Their approach combines maximum-likelihood fitting methods with goodness-of-fit tests based on Kolmogorov-Smirnov (KS) statistic. The Fig. \[g\_countries\_panel\] shows the log-log behavior of the Cumulative Distribution Function (CDF) for the population of the CLCA cities, considering only the countries with the highest number of CLCA cities for each continent (for other countries, see the SI). The $Pr(\mathcal{P} \geq P)$ represents the probability that a random population $\mathcal{P}$ takes on a value greater or equal to the population $P$. In all CDF plots, we also show the maximum likelihood power-law fit, as well as, the value of the exponent $\zeta=\alpha-1$, where $\alpha$ is the MLE exponent, and the value of $P_{min}$, the lower bound of the MLE.
In Fig. \[g\_countries\_hist\], we show a normalized histogram, with frequency $F$, of the $\zeta$ exponents for all countries ( 145 out of 232) with at least 10 CLCA cities in the region covered by the maximum likelihood power-law fit. The mean value of the $\zeta$ exponents is $\bar{\zeta}=0.98$, with variance $\sigma^2=0.09$. The dashed red line stands for the normal distribution $\mathcal{N}(\bar{\zeta},\sigma^2)$. In spite of the $\zeta$ exponent heterogeneity illustrated by Fig. \[g\_countries\_hist\], Zipf’s law holds for most countries around the globe. We emphasize that such results corroborate with previous studies performed for one country or a small number of countries [@rozenfeld2008; @rozenfeld2011; @gabaix1999a; @gabaix1999b; @ioannides2003; @cordoba2008; @giesen2010; @giesen2011; @jiang2011; @giesen2012; @ioannides2013; @duraton2014; @watanabe2015]. In special, the Fig. \[g\_countries\_hist\] also endorses an astute meta-analysis performed by Cottineau [@cottineau2017]. Cottineau provided a comparison among the Zipf’s law exponents found in 86 studies. Our results strongly corroborate those presented in such study, except that our exponents are ranged between 0 and 2.
Furthermore, we challenge the robustness of our model at higher geographic levels: continents and the entire world. We performed the same analyses and find that our results persist on both scales, [*i.e.*]{} the CLCA cities follows Zipf’s law for continents and the entire world, as illustrated in Figs. \[g\_continents\_panel\] and \[g\_world\].
We summarize our results in a set of 7 tables: Tables \[t\_countries\_af\], \[t\_countries\_as\], \[t\_countries\_eu\], \[t\_countries\_na\] , \[t\_countries\_oc\], and \[t\_countries\_sa\], for countries from Africa; Asia; Europe; North America; Oceania; and South America, respectively. Table \[t\_continents\_world\] contains similar information for all continents and the entire world. In all cases, we show the name of the considered region (country, continent or globe), the ISO 3166-1 alpha-3 code associated (only for countries), the number of cities obtained by the CLCA and those covered by the MLE, the lower bound $P_{min}$, and the Zipf’s exponent $\zeta$.
It is remarkable that the top CLCA city, with a population of 63,585,039 people, is composed of three large urban settlements (Alexandria, Cairo, and Luxor) connected by several small ones. Figs. \[cairo\]a, \[cairo\]b, and \[cairo\]c show the largest cluster of regions in Egypt for the GRUMPv1 dataset; CLCA cities; and night-time lights from the National Aeronautics and Space Administration (NASA) [@nasa2016], respectively. We believe the main reason for this finding has been present in the Northeast of Africa since before the beginning of ancient civilization — namely, the Nile river. Actually, it is well known that almost the entire Egypt population lives in a strip along the Nile river, in the Nile delta, and in the Suez canal on 4% of the total country area ($~10^6\;km^2$), where there are arable lands to produce food [@gehad2003]. The river and delta regions are composed by some large cities and a lot of small villages, making them extremely dense. Therefore, our results raise the hypothesis that the cities and villages across the Nile can be seen as a kind of “megacity”, despite spatially non-contiguous, due to the socioeconomic relation, reflected on the high commuting levels, among close subregions.
The Table \[t\_world\] shows the top 25 CLCA cities in the entire world by population, and their associated areas. After the top CLCA city, Alexandria-Cairo-Luxor, we emphasize that the 13 next-largest CLCA cities are in Asia. Indeed, we can see that the shape of the tail end of the entire world population distribution (in Fig. \[g\_world\]) is roughly ruled by the greater CLCA city in Africa and several CLCA cities in Asia.
These facts are not in line with what was recently reported by the [**United Nations (UN)**]{} [@un2016], [*e.g.*]{} the largest CLCA city, Alexandria-Cairo-Luxor, is just the 9th largest city according to the UN, and the largest UN city, Tokyo, is just the 4th largest according to our analyses.
Conclusions
===========
We propose a model to define urban settlements through a bottom-up approach beyond their usual administrative boundaries, and moreover to account for the intrinsic cultural, political, and geographical biases associated with most societies and reflected in their particular growing dynamics. We claim that such a property qualifies our model to be applied worldwide, without any regional restrictions. We also propose an alternative strategy to improve the computational performance of the discrete CCA. We emphasize that the CCA can still be used to define cities; however, it depends upon a different tuning of its parameters for each large region without direct socioeconomic and political relations. Furthermore, we show that the definition of cities proposed by our approach is robust and holds to one of the most famous results in Social Science, Zipf’s law, not only for previously studied countries, [*e.g.*]{} the US, the UK, or China, but for all countries (145 from 232 provided by GRUMPv1) around the world. We also find that Zipf’s law emerges at different geographic levels, such as continents and the entire world. Another highlight of our study is the fact that our model is applied upon one single dataset to define all cities. Furthermore, we find that the most populated cities are not the major players in the global economy (such as New York City, London, or Tokyo). The largest CLCA city, with a population of 63,585,039 people, is an agglomeration of several small cities close to each other which connects three large cities: Alexandria, Cairo, and Luxor. Finally, after the top CLCA city of Alexandria-Cairo-Luxor, we find that the next-largest 13 CLCA cities are in Asia. These facts are not in full agreement with a recent UN report [@un2016]. According to our results, the largest CLCA city, Alexandria-Cairo-Luxor, is just the 9th largest city according to the UN, while the largest UN city, Tokyo, is just the 4th largest according to our analyses.
Data Accessibility
==================
The dataset supporting this article are available at <http://sedac.ciesin.columbia.edu/data/collection/grump-v1>. More specifically, the reader can click on “Data sets” and, after that, on “Population Count Grid, v1 (1990,1995,2000)”. We also provide the codes for the proposed model that are available at <https://doi.org/10.5061/dryad.968nq8n>.
Competing Interests
===================
We have no competing interests.
Author’s Contributions
======================
EAO performed the data analysis, the algorithm of the proposed model, and the statistical analysis. He also participated in the design of the study and drafted the manuscript. VF carried out the funding acquisition and helped draft the manuscript. JSA participated in the design of the study, carried out the funding acquisition, and helped draft the manuscript. HAM conceived, designed, and coordinated the research, as well as, carried out the funding acquisition and helped draft the manuscript. All authors approved the manuscript.
Acknowledgements
================
We thank the Global Rural-Urban Mapping Project (GRUMPv1) team for the dataset provided. Furthermore, we would like to thank X. Gabaix for helpful comments and discussions.
Funding
=======
We gratefully acknowledge funding by CNPq, CAPES, FUNCAP, NSF, and ARL Cooperative Agreement Number W911NF-09-2-0053 (the ARL Network Science CTA).
[10]{}
Jacobs J. 1961 [*The Death and Life of the Great American Cities*]{}. New York: Random House.
Batty M. 2013 [*The New Science of Cities*]{}. London: MIT Press.
West G. 2017 [*Scale: The Universal Laws of Growth, Innovation, Sustainability, and the Pace of Life in Organisms, Cities, Economies, and Companies*]{}. New York: Penguin Press.
Makse HA, Havlin S, Stanley H E. 1995 Modelling urban growth patterns. [*Nature*]{} [**377**]{}, 608-612. (<http://dx.doi.org/10.1038/377608a0>)
Makse HA, Andrade Jr. JS, Batty M, Havlin S, Stanley HE. 1998 Modeling urban growth patterns with correlated percolation. [*Phys. Rev. E*]{} [*58*]{}, 7054. (<http://dx.doi.org/10.1103/PhysRevE.58.7054>)
Rozenfeld HD, Rybski D, Andrade Jr. JS, Batty M, Stanley HE, Makse HA. 2008 Laws of population growth. [*Proc. Natl. Acad. Sci. USA*]{} [**105**]{}, 18702-18707. (<http://dx.doi.org/10.1073/pnas.0807435105>)
Rozenfeld HD, Rybski D, Gabaix X, Makse HA. 2011 The area and population of cities: new insights from a different perspective on cities. [*Am. Econ. Rev.*]{} [**101**]{}, 2205–2225. (<http://dx.doi.org/10.1257/aer.101.5.2205>)
Rybski D, Garc[í]{}a Cant[ú]{} Ros A, Kropp JP. 2013 Distance-weighted city growth. [*Phys. Rev. E*]{} [**87**]{}, 042114. (<http://dx.doi.org/10.1103/PhysRevE.87.042114>)
Murcio R, Sosa-Herrera A, Rodriguez-Romo S. 2013 Second-order metropolitan urban phase transitions. [*Chaos Solitons Fractals*]{} [**48**]{}, 22-31. (<http://dx.doi.org/10.1016/j.chaos.2013.01.001>)
Hernando A, Hernando R, Plastino A. 2014 Space-time correlations in urban sprawl. [*J. R. Soc. Interface*]{} [**11**]{}, 20130930. (<http://dx.doi.org/10.1098/rsif.2013.0930>)
Frasco GF, Sun J, Rozenfeld HD, ben Avraham D. 2014 Spatially distributed social complex networks. [*Phys. Rev. X*]{} [**4**]{}, 011008. (<http://dx.doi.org/10.1103/PhysRevX.4.011008>)
Arcaute E, Hatna E, Fergunson P, Youn H, Johansson A, Batty M. 2014 Constructing cities, desconstructing scaling laws. [*J. R. Soc. Interface*]{} [**12**]{}, 20140745. (<http://dx.doi.org/10.1098/rsif.2014.0745>)
Masucci AP, Arcaute E, Hatna E, Stanilov K, Batty M. 2015 On the problem of boundaries and scaling for urban street networks. [*J. R. Soc. Interface*]{} [**12**]{}, 20150763. (<http://dx.doi.org/10.1098/rsif.2015.0763>)
Arcaute E, Molinero C, Hatna E, Murcio R, Vargas-Ruiz C, Masucci AP, Batty M. 2016 Cities and regions in Britain through hierarchical percolation. [*Open Science*]{} [**3**]{}, 150691. (<http://dx.doi.org/10.1098/rsos.150691>)
Fluschnik T, Kriewald S, Garc[í]{}a Cant[ú]{} Ros A, Zhou B, Reusser DE, Kropp JP, Rybski D. 2016 The Size Distribution, Scaling Properties and Spatial Organization of Urban Clusters: A Global and Regional Percolation Perspective. [*ISPRS Int J Geoinf*]{} [**5**]{}, 110. (<http://dx.doi.org/10.3390/ijgi5070110>)
Bettencourt LM, Lobo J, Helbing D, K[ü]{}hnert C, West GB. 2007 Growth, innovation, scaling, and the pace of life in cities. [*PNAS*]{} [**104**]{}, 7301–7306. (<http://dx.doi.org/10.1073/pnas.0610172104>)
Arbesman S, Kleinberg JM, Strogatz SH. 2009 Superlinear scaling for innovation in cities. [*Phys. Rev. E*]{} [**79**]{}, 016115. (<http://dx.doi.org/10.1103/PhysRevE.79.016115>)
Bettencourt LM, Lobo J, Strumsky D, West GB. 2010 Urban scaling and its deviations: Revealing the structure of wealth, innovation and crime across cities. [*PLoS ONE*]{} [**11**]{}, e13541. (<http://dx.doi.org/10.1371/journal.pone.0013541>)
Bettencourt LM, West GB. 2010 A unified theory of urban living. [*Nature*]{} [**467**]{}, 912-913. (<http://dx.doi.org/10.1038/467912a>)
Gomez-Lievano A, Youn H, Bettencourt LMA. 2012 The Statistics of Urban Scaling and Their Connection to Zipf’s Law. [*PLoS ONE*]{} [**7**]{}, 0040393. (<https://doi.org/10.1371/journal.pone.0040393>)
Gallos LK, Barttfeld P, Havlin S, Sigman M, Makse HA. 2012 Collective behavior in the spatial spreading of obesity. [*Sci. Rep.*]{} [**2**]{}, 454. (<http://dx.doi.org/10.1038/srep00454>)
Bettencourt LM. 2013 The Origins of Scaling in Cities. [*Science*]{} [**340**]{}, 6139. (<http://dx.doi.org/10.1126/science.1235823>)
Fragkias M, Lobo J, Strumsky D, Seto KC. 2013 Does size matter? Scaling of CO2 emissions and U.S. urban areas. [*PLoS ONE*]{} [**8**]{}, e64727. (<http://dx.doi.org/10.1371/journal.pone.0064727>)
Oliveira EA, Andrade Jr. JS, Makse HA. 2014 Large cities are less green. [*Sci. Rep.*]{} [**4**]{}, 4235. (<http://dx.doi.org/10.1038/srep04235>)
Melo HPM, Moreira AA, Batista E, Makse HA, Andrade Jr. JS. 2014 Statistical Signs of Social Influence on Suicides. [*Sci. Rep.*]{} [**4**]{}, 6239. (<http://dx.doi.org/10.1038/srep06239>)
Louf R, Barthelemy M. 2014 How congestion shapes cities: from mobility patterns to scaling. [*Sci. Rep.*]{} [**4**]{}, 5561. (<http://dx.doi.org/10.1038/srep05561>)
Alves LGA, Mendes RS, Lenzi EK, Ribeiro HV. 2015 Scale-Adjusted Metrics for Predicting the Evolution of Urban Indicators and Quantifying the Performance of Cities. [*PLoS ONE*]{} [**10**]{}, e0134862. (<http://dx.doi.org/10.1371/journal.pone.0134862>)
Li X, Wang X, Zhang J, Wu L. 2015 Allometric scaling, size distribution and pattern formation of natural cities. [*Palgrave Commun*]{} [**1**]{}, 15017. (<http://dx.doi.org/10.1057/palcomms.2015.17>)
Hanley QS, Lewis D, Ribeiro HV. 2016 Rural to Urban Population Density Scaling of Crime and Property Transactions in English and Welsh Parliamentary Constituencies. [*PLoS ONE*]{} [**11**]{}, e0149546. (<http://dx.doi.org/10.1371/journal.pone.0149546>)
Caminha C, Furtado V, Pequeno THC, Ponte C, Melo HPM, Oliveira EA, Andrade Jr. JS. 2017 Human mobility in large cities as a proxy for crime. [*PloS ONE*]{} [**12**]{}, e0171609. (<http://dx.doi.org/10.1371/journal.pone.0171609>)
Operti FG, Oliveira EA, Carmona HA, Machado JC, Andrade Jr. JS. 2017 The light pollution as a surrogate for urban population of the US cities. [*arXiv*]{} 1706.05139.
Gabaix X. 1999 Zipf’s Law and the Growth of Cities. [*Am. Econ. Rev.*]{} [**89**]{}, 129-132. (<http://dx.doi.org/10.1257/aer.89.2.129>)
Gabaix X. 1999 Zipf’s law for cities: An explanation. [*Q. J. Econ.*]{} [**114**]{}, 738-767. (<http://dx.doi.org/10.1162/003355399556133>)
Ioannides YM, Overman HG. 2003 Zipf’s law for cities: An empirical examination. [*Reg. Sci. Urban. Econ.*]{} [**33**]{}, 127-137. (<http://dx.doi.org/10.1016/S0166-0462(02)00006-6>)
C[ó]{}rdoba JC. 2008 On the distribution of city sizes. [*J. Urban Econ.*]{} [**63**]{}, 177-197. (<http://dx.doi.org/10.1016/j.jue.2007.01.005>)
Giesen K, Zimmermann A, S[ü]{}dekum J. 2010 The size distribution across all cities Double Pareto lognormal strikes. [*J. Urban Econ*]{} [**68**]{}, 129-137. (<http://dx.doi.org/10.1016/j.jue.2010.03.007>)
Giesen K, S[ü]{}dekum J. 2011 Zipf’s Law for Cities in the Regions and the Country. [*Econ. Geogr.*]{} [**11**]{}, 667-686. (<http://dx.doi.org/10.1093/jeg/lbq019>)
Jiang B, Jia T. 2011 Zipf’s law for all the natural cities in the United States: a geospatial perspective. [*Int. J. Geogr. Inform. Sci.*]{} [**25**]{}, 1269-1281. (<http://dx.doi.org/10.1080/13658816.2010.510801>)
Giesen K, S[ü]{}dekum J. 2012 The French overall city size distribution. [*R[é]{}gion et D[é]{}veloppement*]{} [**36**]{}, 107-126. (<https://EconPapers.repec.org/RePEc:tou:journl:v:36:y:2012:p:107-126>)
Ioannides Y, Skouras S. 2013 US city size distribution: Robustly Pareto, but only in the tail. [*J. Urban Econ.*]{} [**73**]{}, 18-29. (<http://dx.doi.org/10.1016/j.jue.2012.06.005>)
Duranton G, Puga D. 2014 The Growth of Cities. [*Handbook of Economic Growth*]{} [**2**]{}, 781-853.
Watanabe T, Uesugi I, Ono A. 2015 [*The economics of interfirm networks*]{}. Tokyo: Springer.
Zipf GK. 1949 [*Human behavior and the principle of least effort: an introduction to human ecology*]{}. Cambridge: Addison-Wesley.
Sornette D, Knopoff L, Kagan YY, Vanneste C. 1996 Rank-ordering statistics of extreme events: application to the distribution of large earthquakes. [*J. Geophys. Res.*]{} [**101**]{}, 13883-13893. (<http://dx.doi.org/10.1029/96JB00177>)
Okuyama K, Takayasu M, Takayasu H. 1999 Zipf’s law in income distribution of companies. [*Physica A*]{} [**269**]{}, 125-131. (<http://dx.doi.org/10.1016/S0378-4371(99)00086-2>)
Gallos LK, Makse HA, Sigman M. 2012 A small world of weak ties provides optimal global integration of self-similar modules in functional brain networks. [*PNAS*]{} [**109**]{}, 2825-2830. (<http://dx.doi.org/10.1073/pnas.1106612109>)
Linux Kernel Organization. 2017 [*Zram: Compressed RAM based block devices*]{}. See <https://www.kernel.org/doc/Documentation/blockdev/zram.txt> (Accessed 01-08-2017).
Center for International Earth Science Information Network (CIESIN)/Columbia University, International Food Policy Research Institute (IFPRI), The World Bank, Centro Internacional de Agricultura Tropical (CIAT). 2011 [*Global Rural-Urban Mapping Project Version 1*]{}. See <http://sedac.ciesin.columbia.edu/data/collection/grump-v1> (Accessed 01-08-2017).
Snyder JP. 1987 [*Map Projections - A Working Manual*]{} Washington: United States Government Printing Office.
Clauset A, Shalizi RC, Newman MEJ. 2009 Power-Law Distributions in Empirical Data. [*SIAM Review*]{} [**51**]{}, 661-703. (<https://dx.doi.org/10.1137/070710111>)
Cottineau C. 2017 MetaZipf. A dynamic meta-analysis of city size distributions. [*PLoS ONE*]{} [**12**]{}, e0183919. (<https://doi.org/10.1371/journal.pone.0183919>)
National Aeronautics and Space Administration (NASA). 2016 [*Visible Infrared Imaging Radiometer Suite (VIIRS)*]{}. See <http://npp.gsfc.nasa.gov/viirs.html> (Accessed: 01-08-2017).
United Nations. 2016 [*The world’s cities in 2016*]{}. See <http://www.un.org/en/development/desa/population/publications/pdf/urbanization/the_worlds_cities_in_2016_data_booklet.pdf> (Accessed 01-08-2017).
Gehad A. 2003 [*Deteriorated Soils in Egypt: Management and Rehabilitation*]{}. See <http://www.fao.org/tempref/agl/agll/ladadocs/detsoilsegypt.doc> (Accessed 01-08-2017).
![[**(Color online) City Local Clustering Algorithm (CLCA): Representation of the conditions (ii) and (iii)**]{}. (a) The growth of the reference cluster without the merging process. (b) The rising of the usual clusters. The usual clusters are the CCA clusters at the imminence of the merging process between $D^{(i)}$ and $D^{(i+1)}$. (c) For $t$-th, $s$-th, and $r$-th reference clusters ($t$-th is prior to $s$-th which is prior to $r$-th), the merging processes are performed as described in (b), even though there are clusters already defined close to and within the current analysed region in the second and third case, respectively. In the latter, there is the emergence of a forbidden region. The forbidden regions are the complementary areas of the reference clusters already defined within the usual clusters. In order to define the clusters inside those areas, we force the region defined by the largest value of $D^*$ to grow to the limits of the forbidden region. Here, we suppose that $D^{(j)}>D^{(k)}$. The filled dots stand for the reference sites.[]{data-label="clca"}](clca.pdf){width="100.00000%"}
![[**(Color online) The Global Rural-Urban Mapping Project (GRUMPv1) dataset.**]{} The population map of the entire world from the GRUMPv1 dataset in logarithmic scale.[]{data-label="grumpv1"}](grumpv1.pdf){width="80.00000%"}
![[**(Color online) The largest cluster of regions for the United States (US).**]{} (a) The single population density cluster from the Eastern US is defined by the City Clustering Algorithm (CCA) with lower and upper bound parameters $D^*=50\;\text{people}/km^2$ and $\ell=10\;km$, respectively. The population, provided by the Global Rural-Urban Mapping Project (GRUMPv1) dataset, is shown in logarithmic scale within each populated area. (b) Application of the City Local Clustering Algorithm (CLCA) for the cluster of regions of the Eastern US. The CLCA cities are represented in several colors, [*e.g.*]{} New York in mustard, Philadelphia in light brown, Washington-Baltimore in light green, Boston in green and Chicago in red. The CLCA parameters used were $D^{(min)}=100\;\text{people}/km^2$, $D^{(max)}=1000\;\text{people}/km^2$, $\delta=10\;\text{people}/km^2$, $\ell=3\;km$, $A^*=50\;km^2$, and $H^*=0.05$.[]{data-label="us_east"}](us_east_ab.pdf){width="80.00000%"}
![[**(Color online) Cumulative Distribution Function (CDF) $Pr(\mathcal{P} \geq P)$ versus population $P$, in log-log scale, for the countries with the highest number of cities in each continent (for other countries, see the SI).**]{} (a)-(f) Cities proposed by the City Local Clustering Algorithm (CLCA) are represented by light blue circles. The solid black line is the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009]. The value of the lower bound $P_{min}$ and the exponent $\zeta$ are also shown. The CLCA parameters used were $D^{(min)}=100\;\text{people}/km^2$, $D^{(max)}=1000\;\text{people}/km^2$, $\delta=10\;\text{people}/km^2$, $\ell=3\;km$, $A^*=50\;km^2$ and $H^*=0.05$.[]{data-label="g_countries_panel"}](g_countries_panel.pdf){width="80.00000%"}
![[**(Color online) Normalized histogram, with frequency $F$, of the $\zeta$ exponent at the country level.**]{} The plot shows those countries (145 out of 232) with at least 10 cities defined by the City Local Clustering Algorithm (CLCA) in the region covered by the maximum likelihood power-law fit. We find the mean value of the Zipf’s exponents $\bar{\zeta}=0.98$ and its variance $\sigma^2=0.09$. The dashed red line stands for the normal distribution $\mathcal{N}(\bar{\zeta},\sigma^2)$. Therefore, Zipf’s law holds for the most countries.[]{data-label="g_countries_hist"}](g_countries_hist.pdf){width="80.00000%"}
![[**(Color online) Cumulative Distribution Function (CDF) $Pr(\mathcal{P} \geq P)$ versus population $P$, in log-log scale, for the continents.**]{} (a)-(f) Cities proposed by the City Local Clustering Algorithm (CLCA) are represented by light blue circles. The solid black line is the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009]. The value of the lower bound $P_{min}$ and the exponent $\zeta$ are also shown. The CLCA parameters used were $D^{(min)}=100\;\text{people}/km^2$, $D^{(max)}=1000\;\text{people}/km^2$, $\delta=10\;\text{people}/km^2$, $\ell=3\;km$, $A^*=50\;km^2$ and $H^*=0.05$.[]{data-label="g_continents_panel"}](g_continents_panel.pdf){width="80.00000%"}
![[**(Color online) Cumulative Distribution Function (CDF) $Pr(\mathcal{P} \geq P)$ versus population $P$, in log-log scale, for the entire world.**]{} (a)-(f) Cities proposed by the City Local Clustering Algorithm (CLCA) are represented by light blue circles. The solid black line is the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009]. The value of the lower bound $P_{min}$ and the exponent $\zeta$ are also shown. The CLCA parameters used were $D^{(min)}=100\;\text{people}/km^2$, $D^{(max)}=1000\;\text{people}/km^2$, $\delta=10\;\text{people}/km^2$, $\ell=3\;km$, $A^*=50\;km^2$ and $H^*=0.05$.[]{data-label="g_world"}](g_world.pdf){width="80.00000%"}
![[**(Color online) Northeastern region of Egypt**]{}. (a) The cluster of regions defined by the pre-processing of the GRUMPv1 dataset for the Northeastern region of Egypt. (b) The largest city defined by the City Local Clustering Algorithm (CLCA) in the entire world is formed by several cities, including Alexandria, Cairo, and Luxor. (c) Night-time lights of the Northeast of Egypt provided by National Aeronautics and Space Administration (NASA). The CLCA cities found exhibit a remarkable similarity with the lights across the Nile.[]{data-label="cairo"}](cairo.pdf){width="80.00000%"}
=
Country ISO CLCA cities CLCA cities$^\dagger$ $P_{min}$ $\zeta$
----------------------------------- ----- ------------- ----------------------- ----------- -------------------
Angola AGO 20 16 43,937 0.780 $\pm$ 0.195
Benin BEN 40 30 12,607 0.780 $\pm$ 0.142
Burkina Faso BFA 139 78 12,314 1.256 $\pm$ 0.142
Botswana BWA 79 58 1,674 0.785 $\pm$ 0.103
Central African Republic CAF 37 11 14,868 1.230 $\pm$ 0.371
Ivory Coast CIV 83 47 18,400 0.962 $\pm$ 0.140
Cameroon CMR 143 93 7,478 0.711 $\pm$ 0.074
Democratic Republic of the Congo COD 191 47 25,996 0.764 $\pm$ 0.111
Congo COG 21 18 17,673 1.050 $\pm$ 0.248
Comoros COM 16 15 4,167 0.922 $\pm$ 0.238
Cape Verde CPV 16 11 5,205 1.083 $\pm$ 0.327
Algeria DZA 273 112 24,192 0.910 $\pm$ 0.086
Egypt EGY 19 12 11,967 0.511 $\pm$ 0.147
Eritrea ERI 27 12 6,559 0.730 $\pm$ 0.211
Ethiopia ETH 244 147 6,638 0.688 $\pm$ 0.057
Gabon GAB 33 27 3,108 0.844 $\pm$ 0.162
Ghana GHA 95 25 54,662 1.145 $\pm$ 0.229
Guinea GIN 34 13 40,118 1.234 $\pm$ 0.342
Gambia GMB 35 33 1,186 0.610 $\pm$ 0.106
Guinea-Bissau GNB 26 14 9,148 1.139 $\pm$ 0.305
Kenya KEN 179 20 72,756 1.383 $\pm$ 0.309
Liberia LBR 42 19 6,468 0.604 $\pm$ 0.139
Libyan Arab Jamahiriya LBY 30 18 40,273 1.180 $\pm$ 0.278
Lesotho LSO 14 11 1,999 0.651 $\pm$ 0.196
Morocco (includes Western Sahara) MAR 58 50 26,325 0.763 $\pm$ 0.108
Madagascar MDG 138 74 14,867 1.340 $\pm$ 0.156
Mali MLI 152 146 4,463 1.161 $\pm$ 0.096
Mozambique MOZ 127 14 128,214 1.861 $\pm$ 0.497
Malawi MWI 179 72 4,194 0.779 $\pm$ 0.092
Namibia NAM 31 17 12,467 1.637 $\pm$ 0.397
Niger NER 58 36 10,717 0.753 $\pm$ 0.126
Nigeria NGA 144 80 89,587 0.893 $\pm$ 0.100
Sudan SDN 77 56 39,764 1.031 $\pm$ 0.138
Senegal SEN 42 34 13,475 0.798 $\pm$ 0.137
Sierra Leone SLE 62 52 1,899 0.612 $\pm$ 0.085
Chad TCD 75 14 19,574 1.086 $\pm$ 0.290
Togo TGO 54 11 82,964 1.667 $\pm$ 0.503
Tunisia TUN 46 36 16,130 1.014 $\pm$ 0.169
United Republic of Tanzania TZA 114 33 73,621 0.936 $\pm$ 0.163
Uganda UGA 155 33 30,587 1.386 $\pm$ 0.241
South Africa ZAF 1,915 97 53,320 1.270 $\pm$ 0.129
Zambia ZMB 55 34 7,118 0.666 $\pm$ 0.114
Zimbabwe ZWE 28 24 13,411 0.746 $\pm$ 0.152
: [**African countries.**]{} We show the name, the ISO 3166-1 alpha-3 code, the number of cities obtained by the City Local Clustering Algorithm (CLCA) and the number of those covered by the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009] (represented by $\dagger$), the lower bound $P_{min}$, and the Zipf’s exponent $\zeta$.[]{data-label="t_countries_af"}
=
Country ISO CLCA cities CLCA cities$^\dagger$ $P_{min}$ $\zeta$
-------------------------------------- ----- ------------- ----------------------- ----------- -------------------
Afghanistan AFG 95 38 29,242 0.809 $\pm$ 0.131
Armenia ARM 41 19 17,088 1.256 $\pm$ 0.288
Azerbaijan AZE 34 21 17,169 0.776 $\pm$ 0.169
Bangladesh BGD 103 58 26,586 0.581 $\pm$ 0.076
Bhutan BTN 19 15 893 0.469 $\pm$ 0.121
China CHN 4,782 2,706 29,467 0.941 $\pm$ 0.018
Cyprus CYP 17 15 626 0.486 $\pm$ 0.126
Georgia GEO 52 38 6,526 0.765 $\pm$ 0.124
Indonesia IDN 2,416 542 12,876 0.894 $\pm$ 0.038
India IND 1,040 299 94,976 0.786 $\pm$ 0.045
Iran IRN 169 56 100,763 1.194 $\pm$ 0.160
Israel ISR 24 20 877 0.448 $\pm$ 0.100
Jordan JOR 13 11 15,253 0.803 $\pm$ 0.242
Japan JPN 270 33 289,039 1.011 $\pm$ 0.176
Kazakhstan KAZ 77 22 103,289 1.505 $\pm$ 0.321
Kyrgyz Republic KGZ 134 37 9,117 0.991 $\pm$ 0.163
Cambodia KHM 84 24 34,495 1.735 $\pm$ 0.354
Korea KOR 131 23 126,819 0.750 $\pm$ 0.156
Lao Peoples Democratic Republic LAO 35 20 12,595 0.958 $\pm$ 0.214
Sri Lanka LKA 23 20 8,573 0.466 $\pm$ 0.104
Maldives MDV 149 40 1,498 1.799 $\pm$ 0.285
Myanmar MMR 115 37 69,935 1.190 $\pm$ 0.196
Mongolia MNG 24 19 13,179 1.419 $\pm$ 0.325
Malaysia MYS 119 15 157,843 1.286 $\pm$ 0.332
Nepal NPL 39 22 15,396 0.560 $\pm$ 0.119
Oman OMN 28 12 34,956 1.519 $\pm$ 0.438
Pakistan PAK 96 45 90,356 0.790 $\pm$ 0.118
Philippines PHL 352 38 106,854 1.195 $\pm$ 0.194
Democratic Peoples Republic of Korea PRK 53 20 174,121 1.502 $\pm$ 0.336
Saudi Arabia SAU 57 15 156,672 0.861 $\pm$ 0.222
Syrian Arab Republic SYR 39 20 29,908 0.647 $\pm$ 0.145
Thailand THA 100 24 23,482 0.718 $\pm$ 0.147
Tajikistan TJK 39 13 17,660 0.740 $\pm$ 0.205
Turkmenistan TKM 30 14 26,319 0.883 $\pm$ 0.236
East Timor TLS 23 15 1,220 0.547 $\pm$ 0.141
Turkey TUR 338 244 18,389 0.926 $\pm$ 0.059
Taiwan TWN 16 13 2,186 0.344 $\pm$ 0.095
Uzbekistan UZB 56 36 15,865 0.574 $\pm$ 0.096
Viet Nam VNM 345 72 35,980 0.876 $\pm$ 0.103
Yemen YEM 46 22 38,276 1.059 $\pm$ 0.226
: [**Asian countries.**]{} We show the name, the ISO 3166-1 alpha-3 code, the number of cities obtained by the City Local Clustering Algorithm (CLCA) and the number of those covered by the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009] (represented by $\dagger$), the lower bound $P_{min}$, and the Zipf’s exponent $\zeta$.[]{data-label="t_countries_as"}
=
Country ISO CLCA cities CLCA cities$^\dagger$ $P_{min}$ $\zeta$
----------------------- ----- ------------- ----------------------- ----------- -------------------
Albania ALB 46 32 6,030 0.783 $\pm$ 0.139
Austria AUT 116 74 4,383 0.754 $\pm$ 0.088
Belgium BEL 43 31 9,800 0.706 $\pm$ 0.127
Bulgaria BGR 56 29 33,338 1.308 $\pm$ 0.243
Bosnia-Herzegovina BIH 57 17 15,708 1.186 $\pm$ 0.288
Belarus BLR 36 17 73,682 1.123 $\pm$ 0.272
Switzerland CHE 71 15 55,878 1.167 $\pm$ 0.301
Czech Republic CZE 206 33 41,254 1.393 $\pm$ 0.243
Germany DEU 331 242 13,926 0.811 $\pm$ 0.052
Denmark DNK 134 85 2,248 0.682 $\pm$ 0.074
Spain ESP 358 36 133,759 1.192 $\pm$ 0.199
Estonia EST 51 13 14,041 1.178 $\pm$ 0.327
Finland FIN 72 22 27,831 1.444 $\pm$ 0.308
France FRA 1,253 114 42,160 1.087 $\pm$ 0.102
United Kingdom GBR 214 22 229,133 0.983 $\pm$ 0.210
Greece GRC 320 93 7,639 0.930 $\pm$ 0.096
Croatia HRV 88 40 9,672 1.085 $\pm$ 0.172
Hungary HUN 143 25 34,474 1.189 $\pm$ 0.238
Ireland IRL 189 62 4,775 1.093 $\pm$ 0.139
Iceland ISL 15 12 708 0.560 $\pm$ 0.162
Italy ITA 400 157 19,724 0.885 $\pm$ 0.071
Lithuania LTU 76 32 10,654 1.007 $\pm$ 0.178
Latvia LVA 75 28 9,276 1.107 $\pm$ 0.209
Republic of Moldova MDA 31 23 6,609 0.570 $\pm$ 0.119
Macedonia MKD 45 23 11,001 0.981 $\pm$ 0.205
Netherlands NLD 69 16 112,058 1.288 $\pm$ 0.322
Norway NOR 105 18 21,795 1.214 $\pm$ 0.286
Poland POL 236 160 17,390 0.903 $\pm$ 0.071
Portugal PRT 139 32 17,110 1.027 $\pm$ 0.182
Romania ROU 522 385 3,129 0.740 $\pm$ 0.038
Russia RUS 622 384 31,964 0.893 $\pm$ 0.046
Serbia and Montenegro SCG 60 27 38,415 1.340 $\pm$ 0.258
Slovakia SVK 88 20 35,068 1.468 $\pm$ 0.328
Slovenia SVN 88 32 3,273 0.730 $\pm$ 0.129
Sweden SWE 168 61 11,449 1.008 $\pm$ 0.129
Ukraine UKR 164 107 36,515 0.833 $\pm$ 0.081
: [**European countries.**]{} We show the name, the ISO 3166-1 alpha-3 code, the number of cities obtained by the City Local Clustering Algorithm (CLCA) and the number of those covered by the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009] (represented by $\dagger$), the lower bound $P_{min}$, and the Zipf’s exponent $\zeta$.[]{data-label="t_countries_eu"}
=
Country ISO CLCA cities CLCA cities$^\dagger$ $P_{min}$ $\zeta$
--------------- ----- ------------- ----------------------- ----------- -------------------
Canada CAN 1,135 308 4,879 0.815 $\pm$ 0.046
Costa Rica CRI 14 11 20,751 1.195 $\pm$ 0.360
Cuba CUB 113 46 34,673 1.327 $\pm$ 0.196
Guatemala GTM 25 14 28,353 0.948 $\pm$ 0.253
Honduras HND 236 35 17,120 1.290 $\pm$ 0.218
Haiti HTI 23 18 21,953 0.897 $\pm$ 0.211
Mexico MEX 474 284 11,992 0.726 $\pm$ 0.043
Nicaragua NIC 31 28 9,802 0.821 $\pm$ 0.155
Panama PAN 40 12 17,717 1.089 $\pm$ 0.314
El Salvador SLV 25 13 21,323 0.816 $\pm$ 0.226
United States USA 22,893 1,624 9,874 0.876 $\pm$ 0.022
: [**North American countries.**]{} We show the name, the ISO 3166-1 alpha-3 code, the number of cities obtained by the City Local Clustering Algorithm (CLCA) and the number of those covered by the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009] (represented by $\dagger$), the lower bound $P_{min}$, and the Zipf’s exponent $\zeta$.[]{data-label="t_countries_na"}
=
Country ISO CLCA cities CLCA cities$^\dagger$ $P_{min}$ $\zeta$
------------------ ----- ------------- ----------------------- ----------- -------------------
Australia AUS 177 145 5,332 0.788 $\pm$ 0.065
Fiji FJI 15 14 936 0.807 $\pm$ 0.216
Marshall Islands MHL 28 27 44 0.760 $\pm$ 0.146
New Zealand NZL 108 79 3,077 0.776 $\pm$ 0.087
Papua New Guinea PNG 30 13 13,828 1.479 $\pm$ 0.410
: [**Oceanian countries.**]{} We show the name, the ISO 3166-1 alpha-3 code, the number of cities obtained by the City Local Clustering Algorithm (CLCA) and the number of those covered by the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009] (represented by $\dagger$), the lower bound $P_{min}$, and the Zipf’s exponent $\zeta$.[]{data-label="t_countries_oc"}
=
Country ISO CLCA cities CLCA cities$^\dagger$ $P_{min}$ $\zeta$
----------- ----- ------------- ----------------------- ----------- -------------------
Argentina ARG 749 227 10,880 0.994 $\pm$ 0.066
Bolivia BOL 83 57 6,729 0.841 $\pm$ 0.111
Brazil BRA 966 613 18,555 1.057 $\pm$ 0.043
Chile CHL 59 19 93,915 1.422 $\pm$ 0.326
Colombia COL 402 163 12,890 0.886 $\pm$ 0.069
Ecuador ECU 94 54 12,717 0.832 $\pm$ 0.113
Peru PER 417 153 8,279 0.867 $\pm$ 0.070
Paraguay PRY 29 26 4,928 0.700 $\pm$ 0.137
Uruguay URY 79 16 23,346 1.310 $\pm$ 0.327
Venezuela VEN 81 28 82,323 1.254 $\pm$ 0.237
: [**South American countries.**]{} We show the name, the ISO 3166-1 alpha-3 code, the number of cities obtained by the City Local Clustering Algorithm (CLCA) and the number of those covered by the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009] (represented by $\dagger$), the lower bound $P_{min}$, and the Zipf’s exponent $\zeta$.[]{data-label="t_countries_sa"}
=
Continent/Globe CLCA cities CLCA cities$^\dagger$ $P_{min}$ $\zeta$
--------------------------- ------------- ----------------------- ----------- -------------------
Africa 4,860 660 61,569 0.940 $\pm$ 0.037
Asia 10,953 1,167 169,588 0.947 $\pm$ 0.028
Europe 6,118 1,489 33,951 0.895 $\pm$ 0.023
Oceania 180 103 2,668 0.745 $\pm$ 0.073
N. America 24,919 1,364 20,373 0.883 $\pm$ 0.024
S. America 2,934 522 39,514 0.929 $\pm$ 0.041
World (except Antarctica) 50,314 8,019 35,725 0.871 $\pm$ 0.010
: [**Continents and the entire world.**]{} We show the name, the number of cities obtained by the City Local Clustering Algorithm (CLCA) and the number of those covered by the maximum likelihood power-law fit defined by the Maximum Likelihood Estimator (MLE) [@clauset2009] (represented by $\dagger$), the lower bound $P_{min}$, and the Zipf’s exponent $\zeta$.[]{data-label="t_continents_world"}
=
CLCA City Country CLCA population (people) CLCA Area ($km^2$)
--------------------------- ------------- -------------------------- --------------------
Alexandria-Cairo-Luxor Egypt 63,585,039 34,434
Dhaka Bangladesh 48,419,117 26,963
Guangzhou-Macau-Hong Kong China 44,384,647 12,896
Tokyo Japan 34,318,072 9,189
Kolkota India 28,876,910 10,408
Patna India 28,484,380 18,670
Xi’an China 25,370,875 39,736
Jakarta-Bekasi-Banten Indonesia 23,814,197 5,862
Hanoi-Hai Phong Vietnam 22,480,083 19,128
New Delhi India 22,136,675 6,914
Seoul South Korea 20,318,881 3,610
Mumbai India 18,431,960 2,443
Manila Philippines 17,591,794 4,039
Mexico City Mexico 17,190,725 2,845
São Paulo Brazil 16,984,627 2,840
Kyoto-Osaka-Kobe Japan 16,398,829 4,608
New York City US 16,364,109 4,471
Shangai China 15,291,143 2,529
Kochi-Kottayam-Kollam India 14,551,809 8,091
Surabaya-Gresik-Malang Indonesia 14,289,547 6,891
Los Angeles US 13,615,610 5,167
Cirebon-Tegal-Kebumen Indonesia 12,758,617 6,818
Semarang-Klaten-Surakarta Indonesia 12,456,408 6,418
Moscow Russia 11,894,034 1,448
Buenos Aires Argentina 11,132,081 2,653
: [**Top 25 cities, by population, in the world.**]{} We emphasize that, after the top CLCA city (Alexandria-Cairo-Luxor), the 13 next-largest CLCA cities are in Asia. The largest United Nation (UN) city, Tokyo, is just the 4th according to our analyses.[]{data-label="t_world"}
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'L. Ya. Glozman'
date: 'Received: date / Revised version: date'
title: Chiralspin symmetry and QCD at high temperature
---
[leer.eps]{} gsave 72 31 moveto 72 342 lineto 601 342 lineto 601 31 lineto 72 31 lineto showpage grestore
\[sec:intro\]Introduction
=========================
A structure of the QCD phase diagram as well as a nature of the strongly interacting matter in different regimes attracts enormous experimental and theoretical efforts. It is established in QCD calculations on the lattice that there is a transition to the chirally symmetric regime at large temperatures and low densities, where the quark condensate, an order parameter of $SU(N_F)_L \times SU(N_F)_R$ chiral symmetry, vanishes. In addition there is a strong evidence that above the critical temperature also the $U(1)_A$ symmetry gets restored [@Cossu:2013uua; @Tomiya:2016jwr; @Bazavov:2012qja]. Very recently $N_F=2$ lattice simulations with the domain-wall Dirac operator have demonstrated emergence of $SU(2)_{CS}$ and $SU(4)$ symmetries [@Glozman:2014mka; @Glozman:2015qva] at increasing temperature [@R]. These symmetries have been observed earlier in dynamical lattice simulations upon artificial truncation of the near-zero modes of the Dirac operator at zero temperature [@Denissenya:2014poa; @Denissenya:2014ywa; @Denissenya:2015mqa; @Denissenya:2015woa]. The $SU(2)_{CS} \supset U(1)_A$ and $SU(2N_F) \supset SU(N_F)_L \times SU(N_F)_R \times U(1)_A$ symmetries are symmetries of the chromo-electric interaction in QCD. In addition to the chiral transformations the $SU(2)_{CS}$ and $SU(2N_F)$ rotations mix the left- and right-handed components of quark fields. The chromo-magnetic interaction as well as the quark kinetic term break these symmetries down to $SU(N_F)_L \times SU(N_F)_R \times U(1)_A$.
Here we demonstrate that the quark chemical potential term in the QCD Lagrangian is $SU(2)_{CS}$ and $SU(2N_F)$ symmetric, i.e. it has the same symmetry as the confining chromo-electric interaction. Consequently the quark chemical potential term can only impose the $SU(2)_{CS}$ and $SU(2N_F)$ symmetries of confinement. This means that QCD at high temperature beyond the chiral symmetry restoration line on the $\mu -T$ plane, where the quark condensate vanishes, should have approximate $SU(2)_{CS}$ and $SU(2N_F)$ symmetries with increasing accuracy with temperature and chemical potential.
\[sec:su4\]$SU(2)_{CS}$ and $SU(2N_F)$ symmetries [@Glozman:2014mka; @Glozman:2015qva]
======================================================================================
The $SU(2)_{CS}$ chiralspin transformations, defined in the Dirac spinor space are
$$\label{V-def}
\Psi \rightarrow \Psi^\prime = e^{i {\bf{\varepsilon} \cdot \bf{\Sigma}}/{2}} \Psi \; ,$$
with the following generators
$$\vec \Sigma = \{\gamma_k,-i \gamma_5\gamma_k,\gamma_5\},
\label{eq:su2cs_}$$
$k=1,2,3,4$. Different $k$ define different irreducible representations of dim=2. $U(1)_A$ is a subgroup of $SU(2)_{CS}$. The $su(2)$ algebra $[\Sigma_\alpha,\Sigma_\beta]=2i\epsilon^{\alpha\beta\gamma}\Sigma_\gamma$ is satisfied with any Euclidean gamma-matrix, obeying the following anticommutation relations
$$\gamma_i\gamma_j + \gamma_j \gamma_i =
2\delta^{ij}; \qquad \gamma_5 = \gamma_1\gamma_2\gamma_3\gamma_4.
\label{eq:diracalgebra}$$
The $SU(2)_{CS}$ transformations mix the left- and right-handed fermions. The free massless quark Lagrangian does not have this symmetry.
An extension of the $SU(2)_{CS} \times SU(N_F)$ product leads to a $SU(2N_F)$ group. This group has the chiral symmetry of QCD $SU(N_F)_L \times SU(N_F)_R \times U(1)_A$ as a subgroup. Its transformations and generators are given by $$\label{W-def}
\Psi \rightarrow \Psi^\prime = e^{i \bf{\epsilon} \cdot \bf{T}/2} \Psi\; ,$$
$$\begin{aligned}
\{
(\tau_a \otimes \mathds{1}_D),
(\mathds{1}_F \otimes \Sigma_i),
(\tau_a \otimes \Sigma_i)
\}\end{aligned}$$
where $\tau$ are flavour generators with flavour index $a$ and $i=1,2,3$ is the $SU(2)_{CS}$ index.
The fundamental vector of $SU(2N_F)$ at $N_F=2$ is
$$\Psi =\begin{pmatrix} u_{\textsc{L}} \\ u_{\textsc{R}} \\ d_{\textsc{L}} \\ d_{\textsc{R}} \end{pmatrix}.$$
The $SU(2N_F)$ transformations mix both flavour and chirality.
Symmetries of different parts of the QCD Lagrangian and $SU(2)_{CS}$, $SU(2N_F)$ emergence at high temperatures.
================================================================================================================
The interaction of quarks with the gluon field in Minkowski space-time can be splitted into a temporal and a spatial part:
$$\label{Lagrangian}
\overline{\Psi} \gamma^{\mu} D_{\mu} \Psi = \overline{\Psi} \gamma^0 D_0 \Psi
+ \overline{\Psi} \gamma^i D_i \Psi\; .$$
The first (temporal) term includes an interaction of the color-octet quark charge density $\bar \Psi (x) \gamma^0 \vec \lambda \Psi(x) = \Psi (x)^\dagger \vec \lambda \Psi(x)$ with the chromo-electric part of the gluonic field ($\vec \lambda$ are color Gell-Mann matrices). It is invariant with respect to any unitary transformation that can be defined in the Dirac spinor space, in particular it is invariant under the chiral transformations, the $SU(2)_{CS}$ transformations (1) as well as the transformations (4).
The spatial part contains a quark kinetic term and an interaction of the chromo-magnetic field with the color-octet spatial current density. This spatial part is invariant only under chiral $SU(N_F)_L \times SU(N_F)_R \times U(1)_A$ transformations and does not admit higher $SU(2)_{CS}$ and $SU(2N_F)$ symmetries. Consequently the QCD Lagrangian has, in the chiral limit, only the $U(N_F)_L \times U(N_F)_R$ chiral symmetry.
It was found on the lattice with chirally-invariant fermions in $N_F=2$ dynamical simulations that truncation of the near-zero modes of the Dirac operator results in emergence of the $SU(2)_{CS}$ and $SU(4)$ symmetries in hadrons [@Denissenya:2014poa; @Denissenya:2014ywa; @Denissenya:2015mqa; @Denissenya:2015woa].
The emergence of the $SU(2)_{CS}$ and $SU(4)$ symmetries upon truncation of the lowest modes of the Dirac operator means that the effect of the chromo-magnetic interaction in QCD is located exclusively in the near-zero modes. At the same time the confining chromo-electric interaction, which is $SU(2)_{CS}$- and $SU(4)$-symmetric, is distributed among all modes of the Dirac operator.
To conclude, the low-lying modes of the Dirac operator are responsible not only for chiral symmetry breaking, as it is seen from the Banks-Casher relation [@BC], but also for the $SU(2)_{CS}$ and $SU(4)$ breaking via the magnetic effects. The magnetic effects are linked exclusively to the near-zero modes.
Given this insight one could expect emergence of the $SU(2)_{CS}$ and $SU(4)$ symmetries at high temperatures, because at high temperature the near-zero modes of the Dirac opperator are suppressed. This expectation has been confirmed very recently in lattice simulations with chiral fermions [@R]. It was found that indeed above the critical temperature at vanishing chemical potential the approximate $SU(2)_{CS}$ and $SU(4)$ symmetries are seen in spatial correlation functions and by increasing the temperature the $SU(2)_{CS}$ and $SU(4)$ breaking effects decrease rapidly; at the highest available temperature 380 MeV these breaking effects are at the level of 5%.
Symmetries of the quark chemical potential.
===========================================
Will a non-zero chemical potential break this symmetry? Consider the quark part of the Euclidean QCD action at a temperature $T=1/\beta$ in a medium with the quark chemical potential $\mu$:
$$S = \int_{0}^{\beta} d\tau \int d^3x
\overline{\Psi} [ \gamma_{\mu} D_{\mu} + \mu \gamma_4 ] \Psi,$$
where $\Psi$ and $\overline{\Psi}$ are independent integration variables. The field $\overline{\Psi}$ is defined such that it transforms like $ \Psi^\dagger \gamma_4$, i.e. like Minkowskian $\overline{\Psi}$.
This means that the quark chemical potential term
$$\mu \overline{\Psi} \gamma_4 \Psi$$
transforms under chiral, $SU(2)_{CS}$ and $SU(2N_F)$ transformations as
$$\mu {\Psi}^\dagger \Psi,$$
i.e. it is invariant under all these unitary groups. In other words, the dense QCD matter not only does not break the $SU(2)_{CS}$ and $SU(2N_F)$ symmetries, but in a sense imposes them since the chemical potential $\mu$ is an external parameter that can be arbitrary large. The chemical potential term is a color-singlet. Consequently this term can only reinforce the $SU(2)_{CS}$ and $SU(2N_F)$ symmetries at high temperature and zero chemical potential arising from the chromo-electric color-octet term and a compensation is impossible.
{width="0.9\linewidth"}
We conclude that at high temperature $T \sim 400$ MeV at any chemical potential the QCD matter is approximately $SU(2)_{CS}$- and $SU(4)$-symmetric. The $SU(2)_{CS}$ and $SU(4)$ symmetries emerge due to yet unknown microscopic dynamics. This dynamics suppresses (screens) the chromo-magnetic field while the chromo-electric interaction between quarks is still active.[^1]
The elementary objects in the high temperature QCD matter are chiral quarks connected by the chromo-electric field, without any magnetic effects, a kind of a string [@GS]. These objects cannot be described as bound states in some nonrelativistic potential. With the nonrelativistic Schrödinger equation appearance of chiral as well as of $SU(2)_{CS}$ and $SU(4)$ symmetries is impossible. Consequently the QCD matter at high temperature and low chemical potential could be named a “stringy matter”, see Fig. 1.
Conclusions
===========
The main new insight of this short note is that the approximate $SU(2)_{CS}$ and $SU(2N_F)$ symmetries emerge at a temperature $\sim 2 T_c$ on the $T -\mu$ phase diagram and their breaking decreases with increased chemical potential. So we can consider the QCD matter at these temperatures as at least approximately $SU(2)_{CS}$- and $SU(2N_F)$-symmetric. These symmetries rule out the asymptotically free deconfined quarks: free quarks are incompatible with these symmetries. Note that these symmetries cannot be obtained in perturbation theory which relies on a symmetry of a free Dirac equation, i.e. on chiral symmetry. The elementary objects in the QCD matter at these temperatures are chiral quarks connected by the chromo-electric field. Such a matter is not a quark-gluon plasma (the plasma notion is defined in physics as a system of free charges with Debye screening of the electric field) and could be more adequately named as a stringy fluid.
I am thankful to M. Chernodub who provoked this analysis by asking the author about the fate of the chiralspin symmetry in a dense medium. Partial support from the Austrian Science Fund (FWF) through the grant P26627-N27 is acknowledged.
[30]{}
G. Cossu, S. Aoki, H. Fukaya, S. Hashimoto, T. Kaneko, H. Matsufuru and J. I. Noaki, Phys. Rev. D [**87**]{}, (2013) 114514. Erratum: \[Phys. Rev. D [**88**]{}, (2013) 019901\]. A. Tomiya, G. Cossu, S. Aoki, H. Fukaya, S. Hashimoto, T. Kaneko and J. Noaki, Phys. Rev. D [**96**]{}, (2017) 034509. A. Bazavov [*et al.*]{} \[HotQCD Collaboration\], Phys. Rev. D [**86**]{}, (2012) 094503. L. Y. Glozman, Eur. Phys. J. A [**51**]{}, (2015) 27. L. Y. Glozman and M. Pak, Phys. Rev. D [**92**]{},(2015) 016001. C. Rohrhofer, Y. Aoki, G. Cossu, H. Fukaya, L. Y. Glozman, S. Hashimoto, C. B. Lang and S. Prelovsek, Phys. Rev. D [**96**]{},(2017) 094501. M. Denissenya, L. Y. Glozman and C. B. Lang, Phys. Rev. D [**89**]{}, (2014) 077502. M. Denissenya, L. Y. Glozman and C. B. Lang, Phys. Rev. D [**91**]{}, (2015) 034505. M. Denissenya, L. Y. Glozman and M. Pak, Phys. Rev. D [**91**]{}, (2015) 114512. M. Denissenya, L. Y. Glozman and M. Pak, Phys. Rev. D [**92**]{}, (2015) 074508 Erratum: \[Phys. Rev. D [**92**]{}, (2015) 099902\].
T. Banks and A. Casher, Nucl. Phys. B [**169**]{} (1980) 103. L. Y. Glozman, Phys. Lett. B [**541**]{}, (2002) 115.
[^1]: A plausible microscopic explanation of this phenomenon could be related to suppression at high T of the local topological fluctuations of the gluonic field, like instantons, monopols etc. According to the Atiyah-Singer theorem difference of the number of the left- and right-handed zero modes of the Dirac operator is related to the topological charge $Q$ of the gauge configuration. Consequently with $|Q| \geq 1$ amount of the right- and left-handed zero modes is not equal which manifestly breaks the $SU(2)_{CS}$ symmetry since the $SU(2)_{CS}$ transformations mix the left- and right-handed components of quarks. The topological configurations contain the chromo-magnetic field. What would be exact zero modes become the near-zero modes of the Dirac operator in the global gauge configuration that contain local topological fluctuations, like in the Shuryak-Diakonov-Petrov theory of chiral symmetry breaking in the instanton liquid. Consequently all effects of the chromo-magnetic field are localised in the near-zero modes, while confining chromo-electric field is distributed among all modes. At $T > T_C$ the local toplogical fluctuations are melt what leads first to restoration of chiral symmetry and then to $SU(2)_{CS}$ emergence.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We construct Lagrange interpolating polynomials for a set of points and values belonging to the algebra of real quaternions ${{\mathbb}{H}}\simeq {{\mathbb}{R}}_{0,2}$, or to the real Clifford algebra ${{\mathbb}{R}}_{0,3}$. In the quaternionic case, the approach by means of Lagrange polynomials is new, and gives a complete solution of the interpolation problem. In the case of ${{\mathbb}{R}}_{0,3}$, such a problem is dealt with here for the first time. Elements of the recent theory of slice regular functions are used. Leaving apart the classical cases ${{\mathbb}{R}}_{0,0} \simeq{{\mathbb}{R}}$, ${{\mathbb}{R}}_{0,1}\simeq{{\mathbb}{C}}$ and the trivial case ${{\mathbb}{R}}_{1,0} \simeq {{\mathbb}{R}}\oplus {{\mathbb}{R}}$, the interpolation problem on Clifford algebras ${{\mathbb}{R}}_{p,q}$ with $(p,q) \neq (0,2),(0,3)$ seems to have some intrinsic difficulties.'
address: 'Department of Mathematics, University of Trento, I–38123, Povo-Trento, Italy'
author:
- RICCARDO GHILONI
- ALESSANDRO PEROTTI
title: Lagrange polynomials over Clifford numbers
---
[^1]
Introduction
============
The aim of this work is to define Lagrange interpolating polynomials for a set of points and values belonging to a real Clifford algebra. We make some preliminary considerations to select the Clifford algebras on which the construction can be performed. We then restrict to two cases, the Clifford algebra of signature $(0,2)$ (isomorphic to the algebra of real quaternions) and the one of signature $(0,3)$.
Let ${{\mathbb}{R}}_{p,q}$ denote the real Clifford algebra with signature $(p,q)$, equipped with the usual Clifford anti-involution $x\mapsto x^c$ defined by $$x^c=([x]_0+[x]_1+[x]_2+[x]_3+[x]_4+\cdots)^c=[x]_0-[x]_1-[x]_2+[x]_3+[x]_4-\cdots,$$ where $[x]_k$ denotes the $k$–vector component of $x\in {{\mathbb}{R}}_{p,q}$ (cf. e.g. [@CSSSbook §4.1] or [@GHS §3.2]). For every element $x$ of ${{\mathbb}{R}}_{p,q}$, the *trace* of $x$ is $t(x):=x+x^c$ and the (squared) *norm* of $x$ is $n(x):=xx^c$. Let $m:=p+q$. An element $x$ of ${{\mathbb}{R}}_{p,q}$ can be represented in the form $x=\sum_K x_Ke_K$, with $K=(i_1,\ldots,i_k)$ an increasing multiindex of length $k$, $0\le k \le m$, $e_K=e_{i_1}\cdots e_{i_k}$, $e_\emptyset=1$, $x_K\in{{\mathbb}{R}}$, $x_\emptyset=x_0$, $e_1,\ldots,e_m$ basis elements (with $e_i^2=1$ for $i\le p$, $e_i^2=-1$ for $i>p$). The (real vector) subspace generated by $1,e_1,e_2,\ldots,e_m$ is called the set of *paravectors* in ${{\mathbb}{R}}_{p,q}$ and denoted by ${{\mathbb}{R}}^{(m+1)}$. We identify the field of real numbers with the subspace of ${{\mathbb}{R}}_{p,q}$ generated by the unit of the algebra.
In a non–commutative setting, the ring of polynomials is usually defined by fixing the position of the coefficients w.r.t. the indeterminate $X$ (e.g. on the right) and by imposing commutativity of $X$ with the coefficients when two polynomials are multiplied together (cf. e.g. [@Lam §16]). Given two polynomials $P(X)$ and $Q(X)$, let $P{\boldsymbol{\cdot}}Q$ denote the product obtained in the way we just described. If $P$ has *real* coefficients, then $(P{\boldsymbol{\cdot}}Q)(x)=P(x)Q(x)$. In general, a direct computation (cf. [@Lam §16.3]) shows that if $P(x)$ is invertible, then $$\label{product}
(P{\boldsymbol{\cdot}}Q)(x)=P(x)Q(P(x)^{-1}xP(x)).$$
In this setting, a [(left) root]{} of a polynomial $P(X)=\sum_{h=0}^dX^h a_h$ is an element $x\in{{\mathbb}{R}}_{p,q}$ such that $P(x)=\textstyle\sum_{h=0}^dx^h a_h=0$. As shown in [@GhPe_AIM] and [@GhPe_Trends], in order to obtain a good structure for the zero locus of a polynomial, it is necessary to restrict the domain where roots are looked for, and to impose some conditions on the polynomial. We recall from [@GhPe_AIM] the definition of the *quadratic cone* ${{\mathcal}{Q}}_{p,q}$ of ${{\mathbb}{R}}_{p,q}$: $${{\mathcal}{Q}}_{p,q}:={{\mathbb}{R}}\cup\{x\in{{\mathbb}{R}}_{p,q}\;|\; t(x)\in{{\mathbb}{R}},n(x)\in{{\mathbb}{R}},4n(x)>t(x)^2\}.$$ The quadratic cone coincides with the whole Clifford algebra only when ${{\mathbb}{R}}_{p,q}$ is a division algebra, i.e. for ${{\mathbb}{R}}_{0,0}\simeq{{\mathbb}{R}}$, ${{\mathbb}{R}}_{0,1}\simeq{{\mathbb}{C}}$ and ${{\mathbb}{R}}_{0,2}\simeq{{\mathbb}{H}}$. If we restrict roots to the quadratic cone, then an *admissible* polynomial $P(X)$ with Clifford coefficients (cf. [@GhPe_AIM] for this notion) satisfies a version of the Fundamental Theorem of Algebra. In the construction of Lagrange polynomials for points $x_1,\ldots, x_n$, we are lead to consider only elements $x_i$ in the quadratic cone. Moreover, the procedure requires the invertibility of differences of the form $x_i-x_j$ ($i\ne j$), and also of the form $x_i'-x_j'$, with $x_i'$ and $x_j'$ in the same conjugacy classes of $x_i$ and $x_j$, respectively.
These conditions impose severe restrictions on the Clifford algebras in which the procedure can be done. In every Clifford algebra ${{\mathbb}{R}}_{p,q}$ with $p\ge2$, one can find elements $x,y\in {{\mathcal}{Q}}_{p,q}$, not belonging to the same conjugacy class, such that the difference $x-y$ is non–invertible (e.g. $x=e_{12}$ and $y=1/3e_1+2/3e_{12}$). Due to the isomorphism between ${{\mathbb}{R}}_{2,0}$ and ${{\mathbb}{R}}_{1,1}$, also in the algebras ${{\mathbb}{R}}_{1,q}$, with $q\ge1$, one can find pairs of elements with the same properties. As we will see below, this fact has consequences also on the number of roots of polynomials, and therefore on the uniqueness of interpolating polynomials.
In ${{\mathbb}{R}}_{1,0}\simeq{{\mathbb}{R}}\oplus{{\mathbb}{R}}$, the quadratic cone reduces to the real line ${{\mathbb}{R}}$, where the construction of the Lagrange polynomials is well–known. Therefore, leaving apart the classical cases ${{\mathbb}{R}}_{0,0}\simeq{{\mathbb}{R}}$, ${{\mathbb}{R}}_{0,1}\simeq{{\mathbb}{C}}$, we are left with the algebras ${{\mathbb}{R}}_{0,q}$, with $q\ge2$. In this case, the quadratic cone is simply $${{\mathcal}{Q}}_{0,q}={{\mathbb}{R}}\cup\{x\in{{\mathbb}{R}}_{0,q}\;|\; t(x)\in{{\mathbb}{R}},n(x)\in{{\mathbb}{R}}\}.$$ Let ${{\mathbb}{S}}$ denote the set of square roots of $-1$ in ${{\mathcal}{Q}}_{0,q}$. ${{\mathbb}{S}}$ is the set of elements $J$ such that $t(J)=0$, $n(J)=1$. Every $x\in{{\mathcal}{Q}}_{0,q}$ has a decomposition $x=\alpha+\beta J$ with $\alpha,\beta\in{{\mathbb}{R}}$, $\beta\ge0$ and $J\in{{\mathbb}{S}}$.
For $q=2,3$, it can be shown that the conjugacy class of $y=\alpha+\beta J\in{{\mathcal}{Q}}_{0,q}$ is the set ${{\mathbb}{S}}_y=\{\alpha+\beta K\;|\;K\in{{\mathbb}{S}}\}\subset{{\mathcal}{Q}}_{0,q}$. This comes from the fact that ${{\mathbb}{S}}$ forms a unique conjugacy class (cf. [@HHA]). Since $t(y)=2\alpha$, $n(y)=\alpha^2+\beta^2$, ${{\mathbb}{S}}_x={{\mathbb}{S}}_y$ if and only if $t(x)=t(y)$ and $n(x)=n(y)$. The set ${{\mathbb}{S}}_y$ coincides with the zero locus in ${{\mathcal}{Q}}_{0,q}$ of the *characteristic polynomial* of $y$, i.e. the polynomial with real coefficients $$\Delta_y(X):=(X-y)\cdot (X-y^c)=X^2-X t(y)+n(y).$$ If $q\ge4$, the conjugacy class of an element $x\in{{\mathcal}{Q}}_{0,q}$ is not necessarily contained in the quadratic cone. For example, the class of $e_4$ contains also $(2+e_{123})^{-1}e_4(2+e_{123})=5/3e_4-4/3 e_{1234}\notin{{\mathcal}{Q}}_{0,q}$. For this reason, we will restrict to the cases $q=2,3$.
Preliminary results
===================
For $q=2$ (the quaternionic case), the trace and the norm of an element are always real. For $q=3$, $t$ and $n$ take values in the center of the algebra, i.e. in the subspace generated by $1$ and $e_{123}$: $t(x)=2(x_0+x_{123})$, $n(x)=|x|^2 +\phi(x)e_{123}$, where $|x|=(x\cdot x)^{1/2}$ is the euclidean norm of $x\in{{\mathbb}{R}}_{0,3}\simeq{{\mathbb}{R}}^8$ and $\phi(x)=x\cdot (xe_{123})=2(x_0x_{123}-x_1x_{23}+x_2x_{13}-x_3x_{12})$. Since $\phi(x)=\phi(x^c)$, we have that $n(x)=n(x^c)$ for each $x\in{{\mathbb}{R}}_{0,3}$. We have ${{\mathcal}{Q}}_{0,2}={{\mathbb}{R}}_{0,2}\simeq{{\mathbb}{H}}$, ${{\mathcal}{Q}}_{0,3}=\{x\in{{\mathbb}{R}}_{0,3}\;|\; x_{123}=0,\phi(x)=0\}$.
In ${{\mathbb}{H}}$, every non–zero element is invertible. In ${{\mathbb}{R}}_{0,3}$, there are non–invertible non–zero elements.
\[pro1\] Let $x\in{{\mathbb}{R}}_{0,3}$. Then the following facts hold:
1. $x$ is invertible if and only if its norm $n(x)$ is invertible.
2. $x$ is invertible if and only if $\psi_+(x)\psi_-(x)\ne0$, where $$\psi_\pm(x)=(x_0\pm x_{123})^2+(x_1\mp x_{23})^2+(x_2\pm x_{13})^2+(x_3\mp x_{12})^2.$$
1. If $x$ is invertible, then $(x^{-1})^c=(x^c)^{-1}$ and $(x^{-1})^cx^{-1}$ is the inverse of $n(x)=xx^c$. Conversely, if $n(x)$ is invertible, then $n(x)^{-1}x^c=x^c n(x)^{-1}$ is the inverse of $x$.
2. The center of ${{\mathbb}{R}}_{0,3}$ is isomorphic to the algebra ${{\mathbb}{R}}\oplus{{\mathbb}{R}}$. Therefore, an element $a+be_{123}$ is invertible if and only if $a^2-b^2\ne0$. From the first part it follows that $x\in{{\mathbb}{R}}_{0,3}$ is invertible if and only if $|x|^4-\phi(x)^2\ne0$. The thesis follows from the fact that $\psi_+(x)=\frac12|x+xe_{123}|^2=|x|^2+\,x\cdot (xe_{123})=|x|^2+\phi(x)$ and $\psi_-(x)=|x|^2-\phi(x)$.
In ${{\mathcal}{Q}}_{0,3}$, there exist distinct elements whose difference is not invertible. For example, $e_1,e_{23}$ belong to ${{\mathbb}{S}}$ but $e_1-e_{23}$ is not invertible.
\[invertible\] Let $x,y\in{{\mathcal}{Q}}_{0,3}$, with ${{\mathbb}{S}}_x\ne{{\mathbb}{S}}_y$. Then $x-y$ is invertible.
If $x-y$ is not invertible, from Proposition \[pro1\] we get $\psi_+(x-y)=0$ or $\psi_-(x-y)=0$. Assume $\psi_+(x-y)=0$. Then $x_0+x_{123}=y_0+y_{123}$ and $\psi_+(x)=\psi_+(y)$. Since $x,y\in{{\mathcal}{Q}}_{0,3}$, $\phi(x)=\phi(y)=0$. It follows that $n(x)=|x|^2=\psi_+(x)=n(y)$. Moreover, $x_{123}=y_{123}=0$. Therefore $t(x)=2x_0=2y_0=t(y)$. But then $x$ and $y$ are in the same conjugacy class. The same conclusion is obtained if $\psi_-(x-y)=0$.
The previous results can be obtained also by using an explicit form of the isomorphism ${{\mathbb}{R}}_{0,3}\simeq{{\mathbb}{H}}\oplus{{\mathbb}{H}}$ (cf. for example [@GHS] for such an isomorphism).
Let $V(P)$ denote the set of roots of a polynomial $P(X)=\sum_{h=0}^dX^h a_h$ belonging to the quadratic cone ${{\mathcal}{Q}}_{0,q}$: $$V(P)=\{x\in{{\mathcal}{Q}}_{0,q}\;|\; P(x)=\textstyle\sum_{h=0}^dx^h a_h=0\}.$$
We now prove an analogue of the Gordon–Motzkin Theorem ([@GordonMotzkin], see also [@Lam §16.4]), which in his original form is valid for polynomials over division rings (e.g. over the quaternions).
\[GordonMotzkin\] Let $P(X)=\sum_{h=0}^dX^h a_h$ be a polynomial of positive degree $d$, with coefficients $a_h\in{{\mathbb}{R}}_{0,3}$. Then its roots belong to at most $d$ distinct conjugacy classes in ${{\mathcal}{Q}}_{0,3}$.
We proceed by induction on $d$. If $d=1$, and $xa_1+a_0=ya_1+a_0=0$, with $x,y\in{{\mathcal}{Q}}_{0,3}$, then $(x-y)a_1=0$. From Proposition \[invertible\], it follows that $x$ and $y$ must belong to the same conjugacy class. For $d\ge2$, let $y\in V(P)$. Applying the non–commutative version of the Remainder Theorem (cf. [@Lam §16.2]), we can find a polynomial $Q(X)=\sum_{h=0}^{d-1}X^hb_h$, of degree $d-1$, such that $$P(X)=(X-y){\boldsymbol{\cdot}}Q(X).$$ If $x\in V(P)$, with ${{\mathbb}{S}}_x\ne{{\mathbb}{S}}_y$, then $a:=x-y$ is invertible from Proposition \[invertible\]. Since $0=P(x)=aQ(a^{-1}xa)$, we get $Q(a^{-1}xa)=0$. Therefore $x':=a^{-1}xa\in{{\mathbb}{S}}_x\cap V(Q)$. From the inductive hypothesis, $x$ belongs to the union of at most $d-1$ conjugacy classes. Therefore the roots of $P$ belong to at most $d$ conjugacy classes.
It is not clear if the preceding theorem holds on every Clifford algebra ${{\mathbb}{R}}_{0,q}$. The same proof can not be repeated, since for $q>3$ the element $x'=a^{-1}xa$ does not necessarily belongs to the quadratic cone, and then Proposition \[invertible\] can not be applied. Surely the result is not valid on Clifford algebras ${{\mathbb}{R}}_{p,q}$ with $p\ge2$: for example, the degree one polynomial $X(e_1-e_{12})+e_2-1$ has roots $e_{12}$ and $1/3e_1+2/3e_{12}$ in the quadratic cone, belonging to two distinct conjugacy classes.
The preceding theorem is valid in any Clifford algebra ${{\mathbb}{R}}_{0,q}$ for *admissible* (see [@GhPe_AIM]) polynomials. In particular, it holds for polynomials with paravector coefficients, a case considered in [@YangQianActa].
A polynomial of degree $d$ can have more than $d$ roots if two or more of these are allowed to belong to the same conjugacy class. In ${{\mathbb}{R}}_{0,3}$, this can happen also for degree one polynomials. For example, the (not admissible) polynomial $X(e_1+e_{23})+1-e_{123}$ has two distinct roots $e_1$ and $e_{23}$ in the same conjugacy class ${{\mathbb}{S}}$. Observe that $e_2\in{{\mathbb}{S}}$ but is not a root of the polynomial.
Since the quadratic cone ${{\mathcal}{Q}}_{0,3}$ contains the paravector space ${{\mathbb}{R}}^{(4)}$, one can consider the subset of paravector roots of $P=\sum_{h=0}^dX^h a_h$: $$V^{(4)}(P)=\{x\in{{\mathbb}{R}}^{(4)}\;|\; P(x)=\textstyle\sum_{h=0}^dx^h a_h=0\}\subset V(P).$$ Let $r$ be the number of real roots of $P$ (counted with multiplicity). As shown in [@CoSaSt2009Israel], if $V^{(4)}(P)$ contains two distinct roots in the same conjugacy class ${{\mathbb}{S}}_y$, then ${{\mathbb}{S}}_y\cap{{\mathbb}{R}}^{(4)}\subset V^{(4)}(P)$. We call these roots *spherical roots* of $P$. In this case, $P$ is divisible by the characteristic polynomial $\Delta_y(X)$ of $y$. Let $s_y$ be the maximum exponent of a power of $\Delta_y$ dividing $P$. Let $s$ be the sum of integers $s_y$ when $y$ varies, without repetitions, in the conjugacy classes of non–real roots contained in $V^{(4)}(P)$. From Theorem \[GordonMotzkin\], the number of these classes is at most $d$. We can get a more precise estimate, similar to what obtained in [@PogoruiShapiro] in the quaternionic case.
\[GordonMotzkinParavectors\] Let $P(X)=\sum_{h=0}^dX^h a_h$ be a polynomial of positive degree $d$, with coefficients $a_h\in{{\mathbb}{R}}_{0,3}$. Let $r$ and $s$ be as before. Let $k$ be the number of non–real, non–spherical paravector roots of $P$. Then $r+2s+k\le d$.
We can factor out from $P$ a polynomial with real coefficients, of degree $r+2s$. The quotient is a polynomial of degree $d-r-2s$, to which Theorem \[GordonMotzkin\] applies. Therefore $k\le d-r-2s$, and the estimate is proved.
Main results
============
Lagrange interpolation on the quaternionic space {#lagrange-interpolation-on-the-quaternionic-space .unnumbered}
------------------------------------------------
The problem of polynomial interpolation on quaternions has already been considered, usually studying properties of a quaternionic Vandermonde matrix. In [@Lam1995 §16] (in the general setting of division rings) and in [@RMXQLT1999], it is proved that the problem has a unique solution if and only if the interpolation points are distinct and every conjugacy class contains at most two of the points. Here we define the Lagrange interpolating polynomials, and give the supplementary condition that must be satisfied by the data to assure the existence of the solution when more than two points belong to the same conjugacy class. This is a *collinearity* condition involving also the values to be taken, coming from a property of polynomials with right coefficients (shared with the larger class of *slice regular* functions, cf. [@GeSt2007Adv; @GhPe_AIM]): their restriction to each sphere ${{\mathbb}{S}}_y$ is an affine function.
\[LagrangeH\] Let ${{\mathbb}{S}}_1,\ldots,{{\mathbb}{S}}_n$ be pairwise distinct conjugacy classes of ${{\mathbb}{H}}$ and, for every $j \in \{1,\ldots,n\}$, let $x_{j1},\ldots,x_{jd_j}$ be pairwise distinct elements of ${{\mathbb}{S}}_j$ with $d_j \geq 1$, let $w_{j1},\ldots,w_{jd_j}$ be arbitrary elements of ${{\mathbb}{H}}$ and let $d_j':=\min\{d_j,2\}$. Define $d:=-1+\sum_{j=1}^nd_j'$. Then there exists, and is unique, a quaternionic polynomial $P(X)=\sum_{h=0}^dX^h a_h$ of degree at most $d$ such that $P(x_{jk})=w_{jk}$ for each $j \in \{1,\ldots,n\}$ and $k \in \{1,\ldots,d_j\}$ if and only if, for every $j \in \{1,\ldots,n\}$ with $d_j \geq 3$, the following *quaternionic collinear condition* $({\mathcal}{C}_j)$ holds: $$(x_{j2}-x_{j1})^{-1}(w_{j2}-w_{j1})=(x_{jh}-x_{j1})^{-1}(w_{jh}-w_{j1})\quad \forall h \in \{3,\ldots,d_j\}.$$
Up to reordering indices, we can assume that $d_1=d_2=\ldots=d_m=1$ and $d_{m+1} \geq 2$, …, $d_n \geq 2$ for some $m \in \{0,1,\ldots,n\}$.
Given $T \in {{\mathbb}{H}}[X]$ and $y\in {{\mathbb}{H}}$, with $V(T)\cap{{\mathbb}{S}}_y=\emptyset$, it follows from formula that the polynomial $$S(X):=T(X){\boldsymbol{\cdot}}(X-T(y)^{-1}y T(y))$$ vanishes exactly on $V(T) \cup \{y\}$. Then we can find, for each $j \in \{1,\ldots,m\}$ and for each $k \in \{m+1,\ldots,n\}$, polynomials $P_j$, $P_{k,1}$, $P_{k,2} \in {{\mathbb}{H}}[X]$ such that $$V(P_j)=\{x_{11},\ldots,x_{(j-1)1},x_{(j+1)1},\ldots,x_{m1}\},$$ $$V(P_{k1})=\{x_{11},\ldots,x_{m1}\} \cup \{x_{k2}\},\quad
V(P_{k2})=\{x_{11},\ldots,x_{m1}\} \cup \{x_{k1}\}.$$ For each $j \in \{1,\ldots,m\}$ and $k \in \{m+1,\ldots,n\}$, define the quaternionic Lagrange polynomials $$\begin{aligned}
L_j(X)&:=\Delta_{x_{(m+1)1}}(X)\cdots\Delta_{x_{n1}}(X)\, P_j(X)\, a_j,
\\
L'_k(X)&:=\Delta_{x_{(m+1)1}}(X)\cdots\Delta_{x_{(k-1)1}}(X)\Delta_{x_{(k+1)1}}(X)\cdots\Delta_{x_{n1}}(X)\,P_{k1}(X)\, b_{k1},
\\
L''_k(X)&:=\Delta_{x_{(m+1)1}}(X)\cdots\Delta_{x_{(k-1)1}}(X)\Delta_{x_{(k+1)1}}(X)\cdots\Delta_{x_{n1}}(X) \,P_{k2}(X)\, b_{k2},\end{aligned}$$ where $a_j:=\left(\Delta_{x_{(m+1)1}}(x_{j1})\cdots\Delta_{x_{n1}}(x_{j1})P_j(x_{j1})\right)^{-1}$ and $$b_{k\ell}:=\left(\Delta_{x_{(m+1)1}}(x_{k\ell})\cdots\Delta_{x_{(k-1)1}}(x_{k\ell})\Delta_{x_{(k+1)1}}(x_{k\ell})\cdots\Delta_{x_{n1}}(x_{k\ell})P_{k\ell}(x_{k\ell})\right)^{-1}$$ with $\ell \in \{1,2\}$. Then $$\textstyle
P:=\sum_{j=1}^mL_j w_{j1}+\sum_{k=m+1}^n(L'_k w_{k1}+L''_k w_{k2})$$ is an interpolating polynomial for points $\{x_{i1}\}_{i=1,\ldots, m}\cup\{x_{jh}\}_{j=m+1,\ldots, n}^{h=1,2}$. If there are conjugacy classes containing more than two data points, it remains to prove that $P$ interpolates also at the points $\{x_{jh}\}_{j=m+1,\ldots, n}^{h=3,\ldots,d_j}$. Since the restriction of $P$ to each sphere ${{\mathbb}{S}}_y$ is an affine function, there exist $a,b\in{{\mathbb}{H}}$ such that $P(x)=xa+b$ for every $x\in{{\mathbb}{S}}_y$. For the sphere ${{\mathbb}{S}}_{x_{jh}}$ the constants $a$ and $b$ are given by $$a=(x_{j2}-x_{j1})^{-1}(w_{j2}-w_{j1}),\quad b=w_{j1}-x_{j1}a\,.$$ Therefore, the set of equalities $P(x_{jh})=x_{jh}a+b=w_{jh}$ for $h\in\{3,\ldots, d_j\}$, is equivalent to the collinearity condition (${\mathcal}C_j$).
The uniqueness of the interpolating polynomial comes from the estimate on the number of roots of a quaternionic polynomial proved by Pogorui and Shapiro [@PogoruiShapiro].
We give an example of the procedure described in the proof for five points in ${{\mathbb}{H}}\times{{\mathbb}{H}}$ satisfying the collinearity conditions.
Let $x_{11}=0$, $x_{21}=1+i$, $x_{31}=i$, $x_{32}=j$, $x_{33}=k$. Consider the values $w_{11}=1$, $w_{21}=-1$, $w_{31}=1$, $w_{32}=k$, $w_{33}=-j$. Note that $x_{31},x_{32},x_{33}$ belong to the same conjugacy class ${{\mathbb}{S}}$, with characteristic polynomial $X^2+1$. The relative collinearity condition is satisfied: $$(x_{32}-x_{31})^{-1}(w_{32}-w_{31})=-i=(x_{33}-x_{31})^{-1}(w_{33}-w_{31}).$$ We construct the Lagrange polynomials $L_1,L_2,L_3',L_3''$. We set: $$\begin{aligned}
P_1(X)&:=X-x_{21}=X-1-i\\
L_1^*(X)&:=(X^2+1)P_1(X)=X^3-X^2(1+i)+X-1-i\\
L_1(X)&:=L_1^*(X)L_1^*(x_{11})^{-1}=L_1^*(X)(-1-i)^{-1}=X^3\left(\tfrac{i-1}2\right)+X^2+X\left(\tfrac{i-1}2\right)+1\\
\text{and}\\
P_2(X)&:=X-x_{11}=X\\
L_2^*(X)&:=(X^2+1)P_2(X)=X^3+X\\
L_2(X)&:=L_2^*(X)L_2^*(x_{21})^{-1}=L_2^*(X)(-1+3i)^{-1}=X^3\left(\tfrac{-1-3i}{10}\right)+X\left(\tfrac{-1-3i}{10}\right).\end{aligned}$$ Moreover, let $$\begin{aligned}
Q(X)&:=(X-x_{11}){\boldsymbol{\cdot}}(X-x_{21})=X^2-X(1+i)\\
P_{31}(X)&:=Q(X){\boldsymbol{\cdot}}(X-Q(x_{32})^{-1}x_{32}Q(x_{32}))=X^3+X^2\left(\tfrac{-3-i-j+2k}3\right)-X\left(\tfrac{-2+2i-3j+k}3\right)\\
L_3'(X)&:=P_{31}(X)P_{31}(x_{31})^{-1}=X^3\left(\tfrac{5-2j-k}{10}\right)+X^2\left(\tfrac{-1+k}2\right)+X\left(\tfrac{5-5i+3j-k}{10}\right)\\
\text{and}\\
P_{32}(X)&:=Q(X){\boldsymbol{\cdot}}(X-Q(x_{31})^{-1}x_{31}Q(x_{31}))=X^3-X^2(1+2i)+X(-1+i)\\
L_3''(X)&:=P_{32}(X)P_{32}(x_{32})^{-1}=X^3\left(\tfrac{1-2i+2j+k}{10}\right)-X^2\left(\tfrac{1+k}2\right)+X\left(\tfrac{1+3i-3j+k}{10}\right).\end{aligned}$$ Finally, we get the unique interpolating polynomial of degree 3 $$\begin{aligned}
P(X)&:=L_1(X)w_{11}+L_2(X)w_{21}+L_3'(X)w_{31}+L_3''(X)w_{32}=\\
&=L_1(X)-L_2(X)+L_3'(X)+L_3''(X)k=X^3i+X^2+1.\end{aligned}$$
Lagrange interpolation on the Clifford algebra ${{\mathbb}{R}}_{0,3}$ {#lagrange-interpolation-on-the-clifford-algebra-mathbbr_03 .unnumbered}
---------------------------------------------------------------------
We now perform the construction of Lagrange polynomials for a set of points in ${{\mathbb}{R}}_{0,3}$. Due to the presence of zero–divisors, we must restrict to points belonging to different conjugacy classes.
\[Lagrange1\] Let $x_{1},\ldots,x_{m}$ be pairwise distinct elements of ${{\mathcal}{Q}}_{0,3}$. Assume that $x_{j}\in{{\mathbb}{S}}_j$ for each $j \in \{1,\ldots,m\}$, with ${{\mathbb}{S}}_1,\ldots,{{\mathbb}{S}}_m$ pairwise distinct conjugacy classes. Let $w_{1},\ldots,w_{m}$ be arbitrary elements of ${{\mathbb}{R}}_{0,3}$. Define $d:=m-1$. Then there exists a unique polynomial $P(X)=\sum_{h=0}^dX^h a_h$ with coefficients $a_h\in{{\mathbb}{R}}_{0,3}$, of degree at most $d$, such that $P(x_{j})=w_{j}$ for each $j \in \{1,\ldots,m\}$.
In order to prove the theorem, we need a preliminary result.
\[poly\] Given a polynomial $P(X)=\sum_{h=0}^dX^h a_h$ with coefficients $a_h\in{{\mathbb}{R}}_{0,3}$ and $y \in{{\mathcal}{Q}}_{0,3}$, with $V(P)\cap{{\mathbb}{S}}_y=\emptyset$ and $P(y)$ invertible, the polynomial $Q(X):=P(X){\boldsymbol{\cdot}}(X-P(y)^{-1}y P(y))$ vanishes on $V(P) \cup \{y\}$, and $Q(x)$ is invertible for each $x\notin{{\mathbb}{S}}_y$ such that $P(x)$ is invertible.
The first part follows from the equality $$Q(X)=X P(X)-P(X)P(y)^{-1}y P(y).$$ For each $x\in{{\mathcal}{Q}}_{0,3}$ such that $P(x)$ is invertible, it holds $$Q(x)=P(x)(P(x)^{-1}xP(x)-P(y)^{-1}y P(y)).$$ If $x\notin{{\mathbb}{S}}_y$, the latter equality and Proposition \[invertible\] give the invertibility of $Q(x)$.
Let $P^{(1)}_m(X):=X-x_1$ and define recursively, for $k=2,\ldots,m-1$, the polynomials $$P^{(k)}_m(X):=P^{(k-1)}_m(X){\boldsymbol{\cdot}}(X-P_m^{(k-1)}(x_k)^{-1}x_k P_m^{(k-1)}(x_k)).$$ Note that $P_m^{(k-1)}(x_l)$ is invertible for every $l=k,\ldots,m$, as can be seen applying inductively Lemma \[poly\]. The polynomial $P_m:=P_m^{(m-1)}$ vanishes at $x_1,\ldots,x_{m-1}$ and $P_m(x_m)$ is invertible. We can then define the $m$-th Lagrange polynomial $L_m(X):=P_m(X)P_m(x_m)^{-1}$.
For each $j\in\{1,\ldots,m-1\}$, we can define similarly the Lagrange polynomials $L_1,\ldots, L_{m-1}$. Finally, we set $P(X):=\sum_{j=1}^m L_j(X)w_j$.
The uniqueness of the interpolating polynomial follows immediately from Theorem \[GordonMotzkin\].
As an illustration of the procedure described in the proof of Theorem \[Lagrange1\], we give an example of Lagrange interpolation for three points in ${{\mathbb}{R}}_{0,3}\times{{\mathbb}{R}}_{0,3}$.
Let $x_1=e_1$, $x_2=e_2+e_{23}$, $x_3=-1$ in ${{\mathcal}{Q}}_{0,3}$ and let $w_1=1$, $w_2=2e_{23}$, $w_3=e_1$. We construct the Lagrange polynomial $L_3$: $$\begin{aligned}
P_3^{(1)}(X)&:=X-e_1\\
P_3(X)&:=P_3^{(2)}(X):=(X-e_1){\boldsymbol{\cdot}}(X-(e_2+e_{23}-e_1)^{-1}(e_2+e_{23})(e_2+e_{23}-e_1))=\\
&=X^2+X\left(\frac{e_1}{5}-\frac{3 e_2}{5}+\frac{2 e_1 e_3}{5}-\frac{e_2 e_3}{5}\right)+\left(\frac{6}{5}+\frac{3 e_1 e_2}{5}+\frac{2 e_3}{5}+\frac{1}{5} e_1 e_2 e_3\right)\\
L_3(X)&:=P_3(X)P_3(-1)^{-1}=X^2a_3^2+Xa_3^1+a_3^0, \quad\text{with}\\
a_3^0&=\tfrac1{30}(16+4 e_1-3 e_2+3 e_1 e_2+2 e_3+2 e_1 e_3+e_2 e_3+e_1 e_2 e_3),\\
a_3^1&=\tfrac1{30}(-3+5 e_1-6 e_2+4 e_1 e_3+2 e_1 e_2 e_3),\\
a_3^2&=\tfrac1{30}(11+e_1-3 e_2-3 e_1 e_2-2 e_3+2 e_1 e_3-e_2 e_3+e_1 e_2 e_3).\end{aligned}$$ Similarly, we compute the Lagrange polynomials $L_1$ and $L_2$: $$\begin{aligned}
L_1(X)&:=X^2a_1^2+Xa_1^1+a_1^0, \quad\text{with}\\
a_1^0&=\tfrac1{20}(8-8 e_1+6 e_2-6 e_1 e_2-4 e_3-4 e_1 e_3-2 e_2 e_3-2 e_1 e_2 e_3),\\
a_1^1&=\tfrac1{20}(6-10 e_1+12 e_2-8 e_1 e_3-4 e_1 e_2 e_3),\\
a_1^2&=\tfrac1{20}(-2-2 e_1+6 e_2+6 e_1 e_2+4 e_3-4 e_1 e_3+2 e_2 e_3-2 e_1 e_2 e_3),\\
\text{and\quad}L_2(X)&:=X^2a_2^2+Xa_2^1+a_2^0, \quad\text{with}\\
a_2^0&=\tfrac1{15}(1+4 e_1-3 e_2+3 e_1 e_2+2 e_3+2 e_1 e_3+e_2 e_3+e_1 e_2 e_3),\\
a_2^1&=\tfrac1{15}(-3+5 e_1-6 e_2+4 e_1 e_3+2 e_1 e_2 e_3),\\
a_2^2&=\tfrac{1}{15} (-4+e_1-3 e_2-3 e_1 e_2-2 e_3+2 e_1 e_3-e_2 e_3+e_1 e_2 e_3).\end{aligned}$$ Finally, we get the interpolating polynomial $$\begin{aligned}
P(X)&:=\sum_{j=1}^m L_j(X)w_j:=X^2a_2+Xa_1+a_0, \quad\text{with}\\
a_0&=\tfrac{1}{15} (6 e_1+8 e_2-5 e_1 e_2-10 e_3-2 e_1 e_3+4 e_2 e_3+5 e_1 e_2 e_3),\\
a_1&=\tfrac1{15}(-6-9 e_1+13 e_2+3 e_1 e_2-8 e_3-12 e_1 e_3-e_2 e_3+9 e_1 e_2 e_3),\\
a_2&=\tfrac1{15}(-6+5 e_2+8 e_1 e_2+2 e_3-10 e_1 e_3-5 e_2 e_3+4 e_1 e_2 e_3).\end{aligned}$$ Observe that the expansion of $P(X)$ w.r.t. the eight real coordinates contains 239 terms.
[10]{}
F. Colombo, I. Sabadini, F. Sommen, and D.C. Struppa. , volume 39 of [*Progress in Mathematical Physics*]{}. (Birkhäuser Boston Inc., Boston, MA, 2004).
F. Colombo, I. Sabadini, and D.C. Struppa. Slice monogenic functions. (2009) 385–403.
G. Gentili and D.C. Struppa. A new theory of regular functions of a quaternionic variable. (2007) 279–301.
R. Ghiloni and A. Perotti. A new approach to slice regularity on real algebras. In [*Hypercomplex analysis and its Applications*]{}, Trends Math., pages 109–124. (Birkhäuser, Basel, 2011).
R. Ghiloni and A. Perotti. Slice regular functions on real alternative algebras. (2011) 1662–1691.
B. Gordon and T. S. Motzkin. On the zeros of polynomials over division rings. (1965) 218–226.
K. G[ü]{}rlebeck, K. Habetha, and W. Spr[ö]{}[ß]{}ig. . (Birkhäuser Verlag, Basel, 2008).
E. Hitzer, J. Helmstetter, and R. Ablamowicz. Square roots of $-1$ in real clifford algebras. In [*Quaternion and Clifford Fourier transforms and wavelets*]{}, Trends Math. (Birkhäuser, Basel, 2013).
R.M. Hou, X.Q. Zhao, and L.T. Wang. The double determinant of [V]{}andermonde’s type over quaternion field. (1999) 977–984.
T. Y. Lam. , volume 131 of [ *Graduate Texts in Mathematics*]{}. (Springer-Verlag, New York, 1991).
T. Y. Lam. . Problem Books in Mathematics. (Springer-Verlag, New York, 1995).
A. Pogorui and M. Shapiro. On the structure of the set of zeros of quaternionic polynomials. (2004) 379–389.
Y. Yang and T. Qian. On sets of zeroes of [C]{}lifford algebra–valued polynomials. (2010) 1004–1012.
[^1]: Work partially supported by GNSAGA of INdAM, MIUR-PRIN project “Varietà reali e complesse: geometria, topologia e analisi armonica" and MIUR-FIRB project “Geometria Differenziale e Teoria Geometrica delle Funzioni"
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A new “self-calibrated” statistical analysis method has been developed for the reduction of nulling interferometry data. The idea is to use the statistical distributions of the fluctuating null depth and beam intensities to retrieve the astrophysical null depth (or equivalently the object’s visibility) in the presence of fast atmospheric fluctuations. The approach yields an accuracy much better (about an order of magnitude) than is presently possible with standard data reduction methods, because the astrophysical null depth accuracy is no longer limited by the magnitude of the instrumental phase and intensity errors but by uncertainties on their probability distributions. This approach was tested on the sky with the two-aperture fiber nulling instrument mounted on the Palomar Hale telescope. Using our new data analysis approach alone - and no observations of calibrators - , we find that error bars on the astrophysical null depth as low as a few $10^{-4}$ can be obtained in the near infrared, which means that null depths lower than $10^{-3}$ can be reliably measured. This statistical analysis is not specific to our instrument and may be applicable to other interferometers.'
author:
- 'C. Hanot, B. Mennesson, S. Martin, K. Liewer, F. Loya, D. Mawet, P. Riaud, O. Absil and E. Serabyn'
bibliography:
- 'aeos\_bib.bib'
title: 'Improving interferometric null depth measurements using statistical distributions: theory and first results with the Palomar Fiber Nuller'
---
Introduction
============
Since the first discovery of an exoplanet around a solar-type star [@Mayor95], the quest to find earth-like exoplanets and, even more importantly detect the presence of life on them became a major topic in astrophysics. However, the direct imaging of such systems is very challenging because of the high spatial resolution and dynamic range required. One possible way to overcome these difficulties is to use nulling interferometry [@Bracewell78]. In this approach, one destructively combines the light coming from two or more apertures in order to dim the bright on-axis starlight and reveal faint objects or structures in the immediate vicinity.
The analysis of interferometric data in general, and nulling interferometric data in particular [@Colavita09], is a complex task because accurate calibration of the instrument is needed to extract the scientific information. In the case of interferometric nulling, the quantity of interest is the astrophysical null depth ($N_a$), which is the inverse of the rejection ratio, and directly relates to the target’s spatial brightness distribution. In practice however, the measured interferometric null depth is not strictly equal to the astrophysical null depth, because of the effects of instrumental noise and error sources such as phase differences, intensity mismatch and global intensity fluctuations. It had been thought that a proper determination of the astrophysical null depth requires the mean values of these instrumental error sources to be accurately known, e.g. [@Serabyn00; @Lay04]. The classical method used for deriving astrophysical null depths - and visibilities - has therefore been to average different sequences recorded on the science star and estimate the instrumental bias by observing a calibrator star [@Colavita09]. This technique has been extensively used for years for both classical and nulling interferometry, but suffers from well known limitations: (i), the final accuracy depends on the scientific knowledge of the calibrator star, (ii), the accuracy is limited by the stability of the observing conditions and (iii), calibrator observations are time consuming.
To circumvent these limitations, we describe here a new method of calibrating astrophysical null depths, based on measuring the properties of the observed null depth [*[distribution]{}*]{}. The basic idea is to record a time sequence of the rapid null depth fluctuations, and then retrieve the underlying astrophysical information by modeling the observed statistics of the null depth distribution. Using such a statistical analysis, we show in the following that it is possible to retrieve astrophysical null depths with much better accuracy than classical approaches allow. Moreover, our initial stellar observations indicate that this statistical approach does not require any observation of calibrator stars, at least down to null depth measurement accuracies as low as a few $10^{-4}$ (the exact number depends on the instrument set-up being used). In this paper, we first explain the principle and theory of this new statistical data analysis strategy, and then apply it to initial astronomical null data obtained with the Palomar Fiber Nuller (PFN) [@Serabyn06a; @Serabyn06b; @Mennesson06; @Martin08], a nulling-based interferometric “coronagraph” developed at the Jet Propulsion Laboratory. However, we emphasize that the new reduction method can potentially be applied to any null and/or visibility measurements in general.\
The statistics of the null depth
================================
The expression for the null
---------------------------
We begin from the expression for the observed null depth of a two beam interferometer for a point source in the presence of error sources, as given by @Serabyn00. In the case of two planar monochromatic wavefronts, the combined stellar intensity measured at constructive interference (+) and destructive interference (-) at time $t$ is given by : $$\begin{aligned}
I_\pm^{*}(t) &=& \frac{1}{2}\left[ I_1^{*}(t)+I_2^{*}(t)\pm 2 \cos\left(\Delta\phi(t)\right)\cos\left(\alpha_{rot}\right)\sqrt{I_1^{*}(t)I_2^{*}(t)}\right] \\
&=& \langle I^{*}(t) \rangle \left[ 1\pm\cos\left(\Delta\phi(t)\right)\cos\left(\alpha_{rot}\right)\sqrt{1-\left(\delta I(t)\right)^2}\right]
\label{intpm}\end{aligned}$$ where $I_1^{*}(t)$ and $I_2^{*}(t)$ are the individual stellar intensities of beams 1 and 2 at the beam combiner, respectively, $ \langle I^{*}(t) \rangle =(I_1^{*}(t)+I_2^{*}(t))/2$ is the average input beam intensity, $\delta I(t)=(I_1^{*}(t)-I_2^{*}(t))/ (I_1^{*}(t)+I_2^{*}(t))$ is the fractional deviation from the mean intensity, $\Delta \phi(t)=\phi_1(t)-\phi_2(t)$ is the relative phase delay, and $\alpha_{rot}$ the relative polarization rotation angle.\
The null depth, defined as the inverse of the rejection ratio, is given by: $$\begin{aligned}
N(t)=\frac{I_-^{*}(t)}{I_+^{*}(t)}=\frac{I_-(t)-I_{b}(t)}{I_+(t)-I_{b}(t)}
\label{nullgen}\end{aligned}$$ where $I_{\pm}(t)$ are the constructive and destructive interference intensities including the background level and $I_{b}(t)$ is the background intensity collected by the interferometer. If $\Delta \phi(t)$, $\delta I(t)$ and $\alpha_{rot}$ are all $<<$ 1, the null depth for a point source in the absence of background can be approximated by: $$\begin{aligned}
N(t) &\simeq& {1 \over 4}
\left[ {(\delta I(t))^2 + (\Delta\phi (t))^2 + \alpha_{rot}(t)^2 } \right]
\label{null_func_0}\end{aligned}$$ For a source of finite extent, the observed null depth also depends on the astrophysical null depth $N_a$, determined by the leakage of the source spatial brightness distribution through the null fringe pattern[^1]. For small values of $N_{a}$, the observed null depth can be expressed as [@Serabyn00] [^2]: $$\begin{aligned}
N(t) &\simeq \ N_{a} + {1\over 4}\left[ ({\delta I(t))^2} + ({\Delta\phi (t))^2} + {\alpha_{rot}(t)^2} \right]
\label{null_func_1}\end{aligned}$$ Sometimes, interferometers do not measure the background intensity $I_{b}(t)$ nor the constructive interference term $I_{+}(t)$ at the same time as the destructive signal $I_{-}(t)$, but the observing procedure provides some estimates of their values which we denote as $\hat{I}_{b}(t)$ and $\hat{I}_{+}(t)$, while $\hat{I}_{+}^{*}(t)=\hat{I}_{+}(t)-\hat{I}_{b}(t)$. This means that one does not access the actual null, but an estimate of it given by: $$\begin{aligned}
\hat{N}(t) = \frac{I_{-}(t)-\hat{I}_{b}(t)}{\hat{I}_{+}(t)-\hat{I}_{b}(t)} = N(t)\frac{I_{+}^{*}(t)}{\hat{I}_{+}^{*}(t)}+\frac{I_{b}(t)-\hat{I}_{b}(t)}{\hat{I}_{+}^{*}(t)}
\label{null_func_2}\end{aligned}$$ or $$\begin{aligned}
\hat{N}(t) = I_{r}(t)N(t)+N_{b}(t)
\label{null_func_3}\end{aligned}$$ where $N_{b}(t)=(I_{b}(t)-\hat{I}_{b}(t))/\hat{I}_{+}^{*}(t)$ is the background induced instantaneous error in the estimated null and where $I_{r}(t)=I_{+}^{*}(t)/\hat{I}_{+}^{*}(t)$ is the relative intensity deviation at time t.\
For small values of $N_{a}$, $\delta I(t)$, $\Delta\phi(t)$ and $\alpha_{\rm rot}(t)$, one can use Equation\[null\_func\_1\] for $N(t)$, and the estimated null $\hat{N}(t)$ can be approximated by: $$\begin{aligned}
\hat{N}(t) & \simeq & I_r(t) \left[ N_{a} + {1\over 4}[({\delta I(t))^2} + ({\Delta\phi (t))^2} + ({\alpha_{rot}(t))^2}] \right]+N_{b}(t)
\label{null_func_4}\end{aligned}$$ Although it does not correspond exactly to the actual instantaneous null level (which we cannot measure unless all peak and background measurements are made simultaneously), $\hat{N}(t)$ is the basic measured quantity derived from the observations which is used in this paper. All that matters for the accuracy of our data analysis is that we have: (i) the correct description of $\hat{N}(t)$ as a function of the astrophysical null and instrumental noise terms, i.e. Eq.\[null\_func\_4\] and (ii) some way to evaluate these various noise terms (or more exactly their distributions), which is the object of the following section.
Analytical model for the statistical distribution of null values {#Mathematical_considerations}
----------------------------------------------------------------
Because it would be extremely difficult to zero out or perfectly calibrate all instantaneous error terms, we take here the opposite tack, and ask what can be learned from the observed distribution of the null depth fluctuations. We thus begin by deriving the mathematical expression for the probability distribution corresponding to the null depth estimate given by equations \[null\_func\_2\], \[null\_func\_3\] and \[null\_func\_4\] when the relative phase, the intensity mismatch, the background, and the relative intensity all fluctuate randomly with small amplitudes.
We first assume that the polarization term, $\alpha_{rot}(t)$, is constant, so that we can neglect its time variability in the statistical analysis. For symmetrically placed beams within a common aperture, polarization mismatches should be small compared to phase and intensity errors, and this approximation is valid down to null levels of $10^{-4} $ or lower [@Haguenauer06; @Martin08][^3].
Neglecting this term, the measured null (Eq. \[null\_func\_4\]) then consists of the sum of three terms multiplied by a fourth, and then the product is added to a fifth term. Of these, only the astrophysical null term is fixed (for a given baseline vector). We next assume that the remaining error terms in Eq. \[null\_func\_4\] - the relative intensity uncertainty $ I_r(t) $, the beam intensity mismatch $\delta I(t)$ , the beam differential phase $\Delta\phi(t)$ and the background uncertainty terms are uncorrelated random variables (this assumption is justified in Sect. \[section\_correlation\]). We further assume here that each of these have normal distributions (see sections \[section\_intensities\], \[section\_background\] and \[section\_opd\] for a complete description of the probability distributions of these terms), with means $\mu_i$ and standard deviations $\sigma_i$. Each individual probability density function (PDF) is then given by: $$\begin{aligned}
f_{i}(z_i) = \frac{1}{\sqrt{2\pi}\sigma_i} e^{\frac{-(z_i-\mu_i)^2}{2\sigma_i^2}}
\label{pdf}\end{aligned}$$ where the index $i$ refers equally to the $I_r(t)$, $\delta I(t)$, $\Delta\phi (t)$ and $N_b(t)$ distributions, and where $z_i$ is the corresponding random variable.
However the $\delta I(t)$ and $\Delta\phi(t)$ distributions do not appear linearly but quadratically in the null distribution. In the case where $\delta I(t)$ and $\Delta\phi(t)$ both follow normal distributions, the PDFs of $(\delta I(t))^2 / 4$ and $(\Delta\phi(t))^2/4$ are given by: $$\begin{aligned}
f_{i}(\frac{z_i^2}{4})=\frac{1}{\sqrt{2\pi} \sigma_i }\frac{e^{-(4 z_i+\mu_i^2)/2\sigma^2_i}} {\sqrt{4 z_i}} \cosh{\left(\frac{\mu_i\sqrt{4 z_i}}{\sigma^2_i}\right)}
\label{pdf_square}\end{aligned}$$\
The two distributions, $(\delta I)^{2}/4$ and $(\delta\phi)^{2}/4$, are illustrated in Fig. \[1\] for realistic values of their means and standard deviations. The next step in building the nulling PDF is to sum the phase and the intensity mismatch distributions. If $(\Delta\phi(t))^2$ and $(\delta I(t))^2$ are two independent random variables, the density function $f_{\delta I^2/4+\Delta\phi^2/4}(y)$ is given by the convolution of their respective density functions [@Rohatgi76]. If we denote $y$ as being $(z_{\Delta\phi^2}^2+z_{\delta I^2}^2)/4$, this convolution can be expressed as follows:
$$\begin{aligned}
f_{\frac{\delta I^2}{4}+\frac{\Delta\phi^2}{4}}(y) & = & ( f_{\frac{\Delta\phi^2}{4}} \otimes f_{\frac{\delta I^2}{4}} )(y) \\
& = & \int_{-\infty}^{+\infty} f_{\frac{\Delta\phi^2}{4}}(y_1)f_{\frac{\delta I^2}{4}}(y-y_1) \, \rm d \it y_1
\label{pdf_sum}\end{aligned}$$
Adding the astrophysical null term, $N_{a}$, in Eq.\[pdf\_sum\] then corresponds to a further convolution of Eq.\[pdf\_sum\] with a Dirac function $\delta(N_{a})$. The result is simply a translation of the density function by $N_{a}$ (see Fig\[1\]): $$\begin{aligned}
f_{\frac{\delta I^2}{4}+\frac{\Delta\phi^2}{4}+N_{a}}(y)=f_{\frac{\delta I^2}{4}+\frac{\Delta\phi^2}{4}}(y-N_{a})
\label{pdf_na}\end{aligned}$$\
Now folding the effect of the relative intensity uncertainty $I_r(t)$ into the expression for the measured null (Eq. \[null\_func\_2\]), one computes the distribution of the product of $I_r(t)$ with $\left( (\delta I(t))^2/4+(\Delta\phi(t))^2/4+N_{a}\right)$. Assuming these are uncorrelated random variables [@Rohatgi76], the resulting null depth distribution is: $$\begin{aligned}
f_{\hat{N}}(z_{I_r}) = \int_{-\infty}^{+\infty}\frac{1}{|y|} f_{\frac{\delta I^2}{4}+\frac{\Delta\phi^2}{4}+N_{a}}(y) f_{I_r}\left( \frac{z_{I_r}}{y} \right) \rm d\it y
\label{eq12}\end{aligned}$$
The analytical solution for this integral exists for phase and intensity fluctuations following Gaussian distributions. However, the distribution $f_{\hat{N}}(z_{I_r})$ displays a singularity for $y=0$.
The final expression of the measured null distribution (see Eq. \[null\_func\_4\]) is obtained by convolving Eq.\[eq12\] with the equivalent background null depth distribution $f_{N_b}$: $$\begin{aligned}
f_{\hat{N}}(N) = f_{\hat{N}}(z_{I_r})\otimes f_{N_b}(z_{N_b})
\label{pdf_final_bkg}\end{aligned}$$ Summarizing all the steps described in this section, the final analytical expression for the measured null depth can be retrieved from the individual distributions as follows: $$\begin{aligned}
f_{\hat{N}}(N) = f_{N_b}\otimes \left[ \int_{-\infty}^{+\infty}\frac{1}{|y-N_a|}f_{I_r}\left( f_{\frac{\delta I^2}{4}} \otimes f_{\frac{\Delta \phi^2}{4}} \otimes N_a \right) \right]
\label{pdf_summary}\end{aligned}$$\
The measured null distribution expressed by Eq. \[pdf\_summary\] depends on 9 independent parameters: the means and standard deviations of the 4 error distributions, and the astrophysical null. In the simpler case of a system where only a phase error impacts the measured null distribution, the null PDF depends only on three parameters: the mean and standard deviation of the phase error and the astrophysical null. Figure\[1\] illustrates that $\mu_{\Delta\phi}$ and $\sigma_{\Delta\phi}$ define together the Full Width at Half Maximum (FWHM) and the skewness of the PDF, while $N_a$ only changes its horizontal position.
Fitting strategies {#section_fit}
------------------
Two methods can be used to generate null depth distributions to be fitted to the data. In the first one, referred to hereafter as the “analytical method”, the distribution is generated analytically using the measured means and standard deviations of the background and intensity distributions. The second method, called the numerical method, generates simulated distributions using the measured intensity and background distributions, together with simulated phase error sequences having normal distributions, according to Eq.\[null\_func\_4\]. For illustration, we apply our techniques to data obtained with the Palomar Fiber Nuller (PFN), a deployable near infrared ($\simeq$ 2.16 microns) interferometric coronagraph developed at the Jet Propulsion Laboratory and recently installed at the Palomar Hale telescope [@Serabyn06a; @Serabyn06b; @Mennesson06; @Martin08]. As described in the following sections, this method strongly reduces both statistical and systematic errors and can avoid the observation of calibrator stars (depending on the instrument). Therefore we call them respectively the analytical and numerical self-calibrated methods.
### Analytical Self Calibrated Method (ASC) {#sect:anal_method}
The first strategy makes use of theoretical expression \[pdf\_summary\] to calculate null depth distributions, assuming that all instrumental terms follow gaussian statistics. No temporal data sequences are simulated, only the null probability distribution, which depends on nine parameters: the astrophysical null, and the mean and standard deviations of the four gaussian error terms (relative phase, intensity mismatch, relative intensity and background variations). The number of unknown parameters depends on the exact interferometric configuration and must be kept as small as possible in order to give a unique solution to the problem. In most interferometers, the individual beam intensities and the background intensity are monitored as part of the observing sequence, which leaves only three free parameters to be fitted: the mean and standard deviation of the phase error and the astrophysical null depth. In the case of the PFN, two symmetrically placed sub-apertures on the primary mirror are interfered, and using a rapidly spinning wheel, interleaved ($<$ 200 milli-sec) sequences of the interferometrically combined (nulled) signal, the individual beam intensities and the background are recorded. Using this data, we fit the recorded relative intensity mismatch $\delta I(t)$, relative intensity $I_r(t)$ and background $N_{b}(t)$ with Gaussian profiles (see Sect.\[sect:limitations\] for validation of this hypothesis). The resultant mean and standard deviation values are then injected into Eq.\[pdf\_summary\]. The remaining three free parameters of Eq.\[pdf\_summary\], i.e. the differential phase parameters $\mu_{\Delta \phi}$ , $\sigma_{\Delta \phi}$, and the astrophysical null $N_a$, are then adjusted so as to fit the calculated curve to the observed null (Eq.\[null\_func\_2\]) distribution. As detailed in section \[sect:unique\], except for the marginal case where the phase fluctuations are close to 0 (typically $\sigma_{\Delta\phi}<0.005 \,\rm rad$), only one combination of these 3 parameters provides a suitable fit to the observed null data distribution. The pair ($\mu_{\Delta \phi_d}$ , $\sigma_{\Delta \phi_d}$) defines both the FWHM and the skewness of the modeled null distribution, while $N_a$ adjusts the horizontal position of the distribution peak (see Fig.\[1\]) and only one combination of $\mu_{\Delta\phi}$, $\sigma_{\Delta\phi}$ and $N_a$, fits the distribution.\
Figure\[1\] illustrates the construction of the analytical null depth distribution from the individual distributions while Fig.\[2\] (left) illustrates the analytical fitting strategy on a nulling sequence measured on the sky with the PFN. After being fitted by Gaussian distributions, $\mu_{\delta I}$, $\sigma_{\delta I}$, $\mu_{I_N}$, $\sigma_{I_N}$, $\mu_{N_B}$ and $\sigma_{N_B}$ are injected into Eqs. \[pdf\] and \[pdf\_square\] to compute their impact on the measured null distribution. The influence of the intensity mismatch is represented by the grey long dashed curve in Fig.\[1\]. Assuming a Gaussian distribution of the phase error as well, different values of $\mu_{\Delta\phi}$, $\sigma_{\Delta\phi}$ and $N_a$ are used to generate distribution curves. The impact of their distributions is also illustrated in Fig.\[1\] by the grey dashed and dotted curve for the phase error and by the grey plain curve for the astrophysical null. All these distributions are finally combined together according to Eq.\[pdf\_summary\]. The resulting black curve can then be compared to the measured distribution (Fig.\[2\]).
### Numerical Self Calibration Method (NSC) {#section_num}
![*Left*: Fit using the analytical self calibration method on a dataset obtained on $\alpha$ Boo with the PFN in July 2009. The astrophysical null corresponding to the best fit is $0.0136\pm0.0002$. *Right*: Same fit but using the numerical self calibration approach. Note that the simulated distribution now presents more structure, as it integrates the actual distributions of background and intensity terms. The astrophysical null corresponding to the best fit is $0.0137\pm0.0003$.[]{data-label="2"}](0011_final_anal.eps "fig:"){height="7cm"} ![*Left*: Fit using the analytical self calibration method on a dataset obtained on $\alpha$ Boo with the PFN in July 2009. The astrophysical null corresponding to the best fit is $0.0136\pm0.0002$. *Right*: Same fit but using the numerical self calibration approach. Note that the simulated distribution now presents more structure, as it integrates the actual distributions of background and intensity terms. The astrophysical null corresponding to the best fit is $0.0137\pm0.0003$.[]{data-label="2"}](0011_final.eps "fig:"){height="7cm"}
Unlike the ASC, the numerical self-calibration approach (NSC) does not make any assumption about the distributions of the intensity mismatch, background and total intensity terms, which are assumed to be measured within the null sequence or close in time. Instead of fitting the distribution of these three measured signals by gaussian distributions, we use the data - and hence actual distributions - recorded for each of these quantities and inject them directly into Eq.\[null\_func\_4\].
In the case of the PFN for instance, interleaved ($<$100milli-sec) measurements of the individual beams, interferometric (close to null) and background intensities are recorded over sequences of a few minutes. Although the background and individual beam signals are not recorded exactly at the same time as the null, their distributions can be measured with very high fidelity. In order to fit a sequence of observed null values, we combine these observed distributions with a generated random phase error (with a normal distribution) of the same size (same number of data points) according to (Eq.\[null\_func\_4\]). We only make two assumptions when using this method: (i) the differential phase follows a Gaussian distribution, (ii) the individual beam intensities are uncorrelated. The latter condition, which seems valid for the PFN measurements (see Sect.\[sect:limitations\]), implies that the distribution of the differential intensity term $\delta I(t)$ derived from $I_1^*(t)$ and $I_2^*(t')$ measured at different times, is the same as if the individual intensities were measured simultaneously. The main advantage of the numerical technique is that the data monitored by the instrument (i.e. the individual beam intensities and the background) are directly injected into the model. Therefore, no matter what the real distributions are for those terms, no bias is introduced into the modeled probability distribution. However, as the random generation of the differential phase vector produces slightly different distributions and best fit parameters for different seeds, the numerical method adds some intrinsic uncertainty. This “fitting noise” is computed by generating many random phase errors and measuring the standard deviation of the resulting best fit astrophysical null depths. This uncertainty adds quadratically to the statistical error defined in the next section. Consequently, the final error bar quoted on the astrophysical null derived by the numerical method is slightly larger than in the analytical case, but the potential sources of bias are reduced. Figure\[2\], right panel, shows an example of probability distribution fitting (same $\alpha$ Boo sequence as above) using the numerical method.
Error bars and residual comparison
----------------------------------
In this section, we compare the results obtained with the two fitting approaches. To perform this comparison, we make use of the retreived parameter corresponding to the best fits, the goodness of the fit and the relative residuals between the models and the data. To compute the goodness of the fit and derive the optimum fit parameters, we minimize a reduced Pearson $\chi^2$ quantity, defined as: $$\begin{aligned}
\chi^2 = \frac{1}{N_{\rm bins}-4} \sum_{i=1}^{n} \frac{\left( f_{\hat{N}}^{Obs}(i)- f_{\hat{N}}^{Theo}(i)\right)^2}{f_{\hat{N}}^{Theo}(i)}
\label{pearson}\end{aligned}$$ where $f_{\hat{N}}^{Obs}$ and $f_{\hat{N}}^{Theo}$ are respectively the observed and theoretical null probability distributions and $N_{\rm bins}-4$ is the number of independent degrees of freedom. Following usual recommendations for robust fitting of probability distributions [@Cochran54; @Rayson04], we use a number of histogram bins equal to $\simeq\sqrt{N_{\rm pts}}$, where $N_{\rm pts}$ is the number of measurement points over the full range of observed null values. Also, only the largest null depth interval for which the occurence within each bin is $\ge5$ is used for the fitting. Unlike the NSC, the probability distribution obtained with the ASC method must be re-scaled prior to computing the $\chi^2$ to ensure that the total number of occurences in the theoretical distribution corresponds to the total number of measurements within the dataset. Mathematically, it comes down to introducing a scaling factor $C$ to match the integral of the observed and theoretical distributions over the domain of definition. i.e. $$\begin{aligned}
C\, . \int_{N_{min}}^{1}f_{\hat{N}}^{Theo}(n) \, \rm d\it n = \int_{N_{min}}^{1} f_{\hat{N}}^{Obs}(n) \, \rm d\it n\end{aligned}$$ where $N_{min}$ is the minimum observed null value of the distribution[^4].
Overall, the analysis of different datasets with both fitting methods provided similar results, with reduced $\chi^2$ ranging between 1 and 1.5, meaning reasonably good statistical agreement between the model and the observations. The computation of realistic error bars must combine two different components which add quadratically: (i) statistical (random) errors and (ii) systematic errors. Systematic errors, such as those arising from slow drifts in the experimental set-up (quasi-statics, e.g. [@Colavita09]), are not captured by the statistical analysis of a single sequence, and will be discussed in Sect.\[section\_comparison\]. A thorough description of the different sources of quasi-static errors will also be presented in Sect.\[sect:limitations\]. We only compute and quote statistical errors in this section.
For an individual sequence, the statistical uncertainty $\sigma_{\rm stat}$ on the derived astrophysical null is assessed using the $\chi^{2}$ statistical properties (see [@num-recip], §15.6.4). $N_{a}$ is varied around its optimal value while the $\chi^{2}$ is minimized by adjusting the other two parameters. The error bar on $N_a$ corresponds to the $N_a$ variation required to increase the reduced $\chi^2$ by a tabulated increment based on the desired confidence level and the number of degrees of freedom in the fit. A 68.3% confidence level was adopted on the quoted error bar, and the analysis of the covariance of the fit with the other two parameters (i.e. $\mu_{\Delta\phi}$ and $\sigma_{\Delta\phi}$) is presented in the Appendix. This estimation of the retrieved parameters error bars is only valid if the observed null values are affected by zero mean gaussian noise. As a sanity check, we then also conducted a bootstrapping analysis -independent of the actual noise properties -, resampling and replacing the observed null values to generate many (500) “fake” sequences. Analyzing the corresponding histograms yields astrophysical null (68.3% confidence interval) statistical uncertainties similar to those derived using the $\chi^{2}$ approach.
As an illustration, the left panel of Figure \[2\] shows the best analytical self-calibrated fit (black curve) to the null distribution observed (grey dashed line and squares) on the bright star $\alpha$ Boo with the PFN (one particular two minute long sequence). The reduced $\chi^2$ is 1.17 and the derived astrophysical null calculated with a $1 \sigma$ confidence interval is $0.0136\pm 0.0002$ (see Appendix \[app:a\] for more details).The error bar quoted here is the statistical error only. The main advantage of this analytical fitting method is its mathematical consistency and precision. However, it assumes normal and uncorrelated noise distributions for all noise sources, instead of injecting their observed distributions into the model. These assumptions will be justified and explained in Section \[sect:limitations\]. Another characteristic of this approach is that because of the $1/|y|$ term in Eq. \[eq12\], the distribution is not defined for a null depth N = 0. However this issue can be solved by simply removing the bin containing $N=0$ during the fitting process.
![Relative error ($(f^{Obs}_{\hat{N}}-f^{Theo}_{\hat{N}})/f^{Obs}_{\hat{N}}$) between the fitted null distributions and the measured one as a function of the null depth. The grey curve with circular markers, represents the relative error relative to the analytical self calibrated method while the black curve with square markers represents the relative error obtained with the statistical method. Both relative residuals are similar with rms values $\simeq$ 0.05. []{data-label="rel_diff"}](relative_error.eps){width="8cm"}
The right panel of Figure\[2\] represents the best fit obtained with the numerical approach on the same $\alpha$ Boo dataset. The reduced Pearson’s $\chi^2$ is 1.23. The derived astrophysical null depth is $N_a=0.0137\pm0.0003$, in excellent agreement with the value obtained using the analytical approach and gaussian statistics for all instrumental terms. The quoted error bar accounts for both the statistical uncertainty and the numerical “fitting noise” discussed in section Sect.\[section\_num\]
To complete this comparison, Figure \[rel\_diff\] shows the relative difference between the measured null distribution and the distribution obtained with both the analytical method (grey curve, circular markers) and the numerical fitting method (black curve, square markers). As can be seen on this figure, the two different statistical data reduction methods are very equivalent in terms of accuracy: the relative residuals between their distributions and the measured one are similar. This is particularly true for small null depth values ($N<0.05$) where most of the astrophysical information is located. Overall, this comparison shows that very similar results are obtained on $\alpha$ Boo with the analytical an numerical methods.
Amplitude of the fluctuations {#sect:unique}
-----------------------------
In this section, we demonstrate the conditions that must be fulfilled by the error fluctuations in order to produce a distribution that can be fitted by a unique combination of the parameters. For that, we consider the simpler case where only phase errors are present.\
First, let us consider the extreme case of a perfectly stable system ($\sigma_{\Delta\phi}=0$) but with an error on the phase shift ($\mu_{\Delta\phi}\neq 0$). The measured null distribution is then a Dirac function that peaks at $\hat{N}= N_a + (\mu_{\Delta\phi})^{2}/4$ (see Fig.\[fig:phase\_distr\]). Therefore, only the sum, $N_a + (\mu_{\Delta\phi})^{2}/4$, can be determined but not the astrophysical null.
\
Of course, using a statistical approach for analyzing perfectly constant data does not make much sense and is not realistic. However, it shows that the phase fluctuations must have a minimal amplitude to make a statistical approach applicable. Now, let us consider the more realistic case of a system having both a phase fixed bias and phase fluctuations. Eq. \[pdf\_square\] expresses the impact of phase fluctuations on the null depth distribution. From this equation, it can be seen that the larger the fluctuations, the broader the corresponding distribution (see Fig.\[fig:phase\_distr\]). If the FWHM of this distribution is smaller than the bin size used for computing the null distribution, it appears as a Dirac function (which corresponds to a fixed phase error) and the fitted parameters cannot be found. Now, if the phase distribution can be properly sampled in several bins, the three parameters that must be fitted ($\mu_{\Delta\phi}$, $\sigma_{\Delta\phi}$ and $N_a$) can be retreived. More importantly, the solution found is unique. For small phase fluctuations, Eq. \[pdf\_square\] can be approximated by a Gaussian function whose FWHM is $2\sqrt{2 \ln 2} \times \sigma_{\Delta\phi}\sqrt{\mu_{\Delta\phi}}$. The criterion for a unique solution to our fit is therefore that this FWHM is larger that a few ($k$) times the histogram bin size (i.e. $k$bin size $< 2\sqrt{2 \ln 2} \times \sigma_{\Delta\phi}\sqrt{\mu_{\Delta\phi}}$). From this equation, it can be seen that for larger mean phase offsets, the minimum phase fluctuation required to meet this criterion decreases. This is due to the fact that the null depth depends quadratically on the phase error. In practice, simulations have shown that the phase distribution must be sampled over at least 6 bins ($k\ge 6$). Fig.\[7\] shows, on simulated datasets, the minimum amplitude of the phase fluctuations required as a function of the mean phase error for a bin size of 0.001. This bin size directly depends on the number of data points available within a dataset (see Sect. \[sect:anal\_method\]). It means that increasing the observing time (and therefore the number of data points within a dataset) allow the use of smaller bin sizes and hence even easier parameter retrieval. In this figure, a fit is considered successful when all three parameters are found within some a priori tolerable error. For $N_a$, it means that the error is smaller than the histogram bin size. For the two other parameters, it means that their effect on the null depth ($\Delta\phi^2/4$) is also smaller than the histogram bin size. It is interesting to note that even small fluctuations compared to the mean phase errors are sufficient to retrieve the astrophysical null depth with a very good accuracy. This also means that it is possible to measure $N_a$ even if the fluctuating phase error never reaches zero, and so even when the true astrophysical null value is never reached. Finally it is important to note that the results obtained with the PFN illustrating this paper correspond to parameter combinations located well within the *“parameters retrieved”* zone of Fig.\[7\].
On sky performance: classical vs. statistical reduction methods {#section_comparison}
===============================================================
In order to investigate the validity and accuracy of our statistical data reduction approach, we applied it to astronomical data obtained with the PFN during a July 2009 observing run. In order to evaluate the astrophysical null accuracy achieved with our statistical analysis, we present here the results obtained on a series of consecutive independent measurements of $\alpha$ Boo with the PFN. We explore both the repeatability of the results (precision assessment), and their consistency with values previously reported by long baseline interferometry (accuracy and bias assessment).
We use here a set of five independent null sequences recorded on $\alpha$ Boo with the PFN in July 2009, and compare the astrophysical nulls, $N_a$, and precisions derived from (i) the “classical” null (or visibility) data reduction method and (ii) from the probability distribution analysis. We then compare our results with the stellar diameter measurement obtained on this same star with long baseline interferometry (LBI), discussing the aspects of accuracy and systematic errors.
Classical reduction method
--------------------------
Very few nulling data from ground based telescopes have been analyzed so far, as only two nulling interferometers are operating: the Keck Interferometer Nuller [@Colavita09] and the BLINC Nuller [@Hinz00]. Until now, the method used to analyse nulling data was analogous to that used for calibrating visibility measurements. The principle is to first evaluate the null depth observed on the science target by averaging the fluctuating instantaneous null depth over a significant number of points. This measurement is biased due to the fast fluctuations of phase and intensity errors. The same measurement is then conducted on a calibrator star of well known diameter, located close to the science target and with a similar magnitude at the wavelength of observation [@Merand05]. For both stars, the measured null depth $\langle N(t) \rangle$ is the sum of the astrophysical null $N_{a}$ and the mean instrumental null $\langle N_{\rm i}(t) \rangle$ averaged over the sequence: $$\begin{aligned}
\langle N(t) \rangle & = & N_{a}+ \langle N_{\rm i}(t) \rangle \\
\langle N_{\rm cal}(t + \Delta t) \rangle & = & N_{a, \, \rm cal}+ \langle N_{\rm i, \,cal}(t + \Delta t) \rangle
\label{null_cal}\end{aligned}$$ where the astrophysical null depth on the calibrator star ($N_{a, \,cal}$) is assumed to be known thanks to an accurate photosphere model or from independent interferometric observations. Therefore, assuming that the instrumental null is constant, one estimates the scientific target’s astrophysical null as: $$\begin{aligned}
N_a = N_{a,\, \rm cal} + \langle N(t) \rangle - \langle N_{\rm cal}(t+\Delta t) \rangle
\label{null_sci}\end{aligned}$$ Obviously, the accuracy on $N_a$ depends both on the calibrator’s astrophysical null uncertainty and on the stability of the instrumental null (or the ability to extrapolate its value accurately based on bracketing calibrator observations).\
The method used to emulate a “classical analysis” of our null data is the following. First the “bad” (large instantaneous nulls) data points within each dataset are rejected, both for the target and the calibrator. Only the data having null values between the minimum measured null $N_{min}$ and $N_{min}+\sigma_N$ are kept, where $\sigma_N$ is the rms of the null data (see Fig \[fig:null\_seq\], dash-dot line). This is also called the sigma clipping method. The null depth of an individual object sequence is then computed as the mean of the remaining data points (see Fig \[fig:null\_seq\], dashed line). The same approach is applied to both the scientific target and the calibrator data, and the calibrated astrophysical null depth is then computed using Eq.\[null\_sci\]. The black stars in Figure\[5\], left panel, represent the calibrated null depths obtained with this classical data analysis on five consecutive $\rm \alpha \, Boo$ datasets. These data have been calibrated using five datasets obtained on $\alpha$ Her. The error bar on the individual measurements is given by the quadratic sum of the target statistical error, the systematic error and the calibrator total error (statistical error and diameter uncertainty). The individual statistical error bars are computed from the variance of the null depth fluctuations within each dataset (after applying data clipping) and are equal to 0.018 on $\alpha$ Boo. The systematic error is more difficult to calculate and can be assessed both by comparing the measured null depth with the one expected from previous measurements of $\alpha$ Boo’s stellar diameter and by comparing the individual statistical errors with the variance of the null over the 5 datasets. Assuming no/low systematics and averaging over the 5 data points, the astrophysical leakage measured on $\rm \alpha \, Boo$ is $0.0123\pm 0.008$ (see Table\[tab\_summary\]). The significant slow drift of the measured nulls in Fig.\[5\], left panel clearly illustrates that the classical method is very sensitive to the instrumental/ seeing conditions and to the fact that the calibrator was only observed [*[after]{}*]{} the five $\alpha$ Boo sequences and not in between them. The large error bars derived -even in the quite optimistic case of no systematics- demonstrate that in fact, with the short PFN interferometric baseline and when using the classical data reduction method, $\alpha$ Boo’s near infrared diameter can not be measured reliably .
Statistical reduction method
----------------------------
On the other hand, our statistical data analysis approach uses the whole range of null values recorded and neither uses nor requires any calibration star. Using the same five $\rm \alpha \, Boo$ datasets, the individual astrophysical nulls measured using statistics have much smaller individual error bars (0.0003), and are very stable over the whole two hours of observation (Fig \[5\], right panel). Using the 5 datasets obtained on $\alpha$ Boo, a set of $(N_{a,i}, \sigma_{\rm{stat},i})$ best fitting values is derived. From that ensemble, we compute the weighted mean value of $N_{a}$ with weights $w_i=1/\sigma_{\rm stat,\it i}^2$. The weighted mean astrophysical null value derived over the full sequence is 0.0132. Assuming [*[no systematic errors]{}*]{} and simply propagating the individual error bars ($\sigma_{\rm stat,\it i}$), the final statistical error bar is given by $\sigma_{\rm stat}^{-2}=\sum_{i}\sigma_{\rm stat,\it i}^{-2}$ and amounts to 0.00013. This yields an astrophysical null estimate of $N_{a}= 0.0132 \pm 0.00013$ for $\alpha$ Boo (see Table\[tab\_summary\]).
Of course, systematic errors can be present in the data, for instance arising from slow drifts in the experimental set-up (quasi-statics) which are not captured by the statistical analysis of a single sequence. However, conversely to the classical method case, no obvious long term drift is visible versus time. The weighted standard deviation computed over the sequence is $0.0004$, in fairly good agreement with the quoted individual error of 0.0003, pointing to small systematics if any. This weighted standard deviation can also serve as an estimate of the systematic error per individual measurement, e.g. [@Colavita09]. The systematic error on the mean is likely smaller than that per individual measurement, but we do not have enough data to check for such reduction of the systematics wrt the number of measurements. Consequently, we estimate the final error bar on $\alpha$ Boo’s measured astrophysical null to be at the few $10^{-4}$ or lower.
Another way to estimate systematics and constant biases is to compare the astrophysical null derived by the statistical method with previous measurements obtained by long baseline interferometry (LBI). This is the object of the following section. A detailed description of the potential sources of quasi-static errors as well as their impact on the null depth is also presented in section \[sect:limitations\]. Finally, observations of calibrators can obviously be used in conjunction with the statistical data analysis to further reduce the effect of residual biases.
Comparison to LBI data
----------------------
For a naked star represented by a limb darkened disk of diameter $\theta_{LD}$ with a limb darkening coefficient $A_\lambda$, the observed astrophysical null is given by [@Absil06a; @Absil11] $$\begin{aligned}
N_{a,LD} = \left( \frac{\pi B \theta_{LD}} {4\lambda}\right)^2 \left( 1-\frac{7A_\lambda}{15} \right) \left( 1-\frac{A_\lambda}{3} \right)^{-1}
\label{limb-darkened}\end{aligned}$$ where $\lambda$ is the central wavelength of observation and $B$ the baseline length. For the PFN, these values are 2.16 $\mu m$ and 3.20m, respectively. This expression can be simplified in the case of uniform disk models by setting $A_\lambda=0$.
LBI measurements of $\alpha$ Boo in the K band (where limb darkening effects and corresponding uncertainties are reduced) provide very accurate results. We use the value of 20.91 $\pm$ 0.08 mas derived by FLUOR/IOTA [@Perrin98; @Lacour08]. This value is also very consistent with the previous measurement of 20.95 $\pm$ 0.20 mas obtained at I2T [@diBenedetto87].
Using the 0.350 linear limb darkening coefficient predicted in the K band for a 4300K giant star with $\rm log\, g = 2.0$ [@Claret95], we get an astrophysical null depth of 0.01314 $\pm$ 0.00010 at the PFN baseline. This is excellent agreement with or measured value of 0.001320 $\pm$ 0.00013 ( or $\pm$ 0.0004 when being conservative wrt systematics) reported above, which corresponds to a limb darkened diameter of 20.96 $\pm$ 0.09 mas (see Table \[tab\_summary\]). The discrepancy between the PFN and LBI $\alpha$ Boo measurements is then at the $10^{-4}$ level, and within the error bars of each measurement. This demonstrates that in the illustrative case of $\alpha$ Boo, our measurement is not only precise but also very accurate. It suggests that if any bias is present in our calibrator-free measurements of $\alpha$ Boo, they are at the few $10^{-4}$ level or below. A similar analysis of PFN data using the statistical reduction method confirms this result on a larger sample on 8 bright giants/ supergiants (Mennesson et al. in preparation). In comparison, the very best 1$\sigma$ null accuracy reported by long baseline interferometry is $\simeq$ 0.002 in the mid-infrared [@Colavita09], and $\simeq$ 0.0025 in the near infrared [@Kervella04], (equivalent to a visibility accuracy of 0.005 for an unresolved source). This indicates that using the self-calibrated data reduction approach, a gain of an order of magnitude in null (or visibility) accuracy can be achieved.
In fact there is little that is specific to the PFN instrument in our approach, and the statistical data reduction method could in principle be applied to any 2-beam interferometer working around null with a fringe tracker. Since null and visibility measurements are equivalent, the statistical analysis may thus also prove useful to regular long baseline visibility interferometry (Mennesson et al. in prep).
Method Name $N_a$ $\theta [\rm mas]$
---------------- -------------------- ---------------------- ----------------------- -- -- --
Classic. nulls $\rm \alpha\, Boo$ $0.0123\pm 0.008$ $20.25^{+6.4}_{-9.8}$
Stat. nulls. $\rm \alpha\, Boo$ $0.0132\pm 0.00013 $ $20.96\pm 0.09$
LBI vis. $\rm \alpha\, Boo$ $0.0131\pm 0.00010 $ $20.91\pm0.08$
: Comparison between limb darkened (LD) diameters found by the PFN using both the classical and the numerical statistical data reduction method and by long baseline interferometry. Note that the null depth value given for LBI is an equivalent null on a 3.4m baseline derived from the measurement of the angular diamter.
\[tab\_summary\]
Possible limitations {#sect:limitations}
====================
We explore in this section some possible limitations of the statistical data reduction technique, which may appear when trying to measure very deep astrophysical null depths. Limitations arise from well identified sources: temporal effects, chromatic effects, and deviations from the assumptions used in the modeling. There are only two assumptions made in the self calibration technique: no temporal correlation between the individual beam intensities, and Gaussian distribution of the error sources [^5]. In the following, we investigate these different effects, assess their contributions to the final null depth estimates, and suggest some mitigation techniques.
Intensity distributions {#section_intensities}
-----------------------
Conversely to the numerical method, where the measured relative intensity uncertainty $I_r(t,\Delta t)$ and intensity mismatch $\delta I(t)$ are directly injected into the model, the analytical approach assumes these distributions to be Gaussian and computes their mean and standard deviation to feed the analytical expression of the measured null distribution (see Sect. \[Mathematical\_considerations\]). Therefore a possible limitation of the analytical approach could occur if these distributions are not Gaussian.\
Fig.\[6\] shows typical relative intensity uncertainty and intensity mismatch distributions measured with the PFN. While the left hand panel compares the $I_r(t)$ measured distribution (grey crosses) with a Gaussian distribution (black curve), the right hand one does the same for the $\delta I(t)$ distribution. As can be seen, both distributions can be reliably fitted by a Gaussian distribution. The goodness of the fit reduced $\chi^2$ values are $\simeq 0.99$ for both $I_r(t)$ and $\delta I(t)$. The bottom panel of these two figures illustrates the relative residuals between the observed distribution and the best Gaussian fit. For both fits, the residuals are close to zero for the whole central part where most of the information is located. Such a good agreement between the measured distributions and Gaussian distributions makes us confident these assumptions are justified and can be used but we note that a very slight skewness may be present.
Background distribution {#section_background}
-----------------------
{width="9.5cm"}
The analytical self-calibration method (unlike the NSC which uses the recorded background level), make the assumption that the distribution of the background level is normal and fits a gaussian profile on the recorded data to feed the analytical expression of the estimated null depth (Eq.\[pdf\_summary\]). However, background drifts can occur during observations either because of instrumental (e.g. electronic drifts) or observational reasons (the background depending on the sky position and time of observation) and can cause biases in the determination of the null depth. Figure\[fig\_background\] represent the distribution of the equivalent background null measured on $\alpha$ Boo over 2 minutes. The grey squares represent the actual measured distribution while the black curve is the best Gaussian fit corresponding to this distribution. Once again, the goodness of the fit is excellent with a $\chi^2\simeq 0.98$. However it must be stressed that this assumption is only verified for the particular PFN observations illustrated in this paper, and must be checked when using other instruments.
Correlation issues {#section_correlation}
------------------
In our statistical (both numerical and analytical) self-calibrated method (section 2.2.1), we made the assumptions that the different noise terms (background, differential intensity, overall intensity and differential phase) were temporally uncorrelated, so we could compute the theoretical null distribution from the individual noise distributions.\
The cross-correlation of the intensity and phase terms is difficult to estimate. However, the optical/ near infrared coherence length of the atmosphere is generally much smaller than the distance between an interferometers sub-apertures. Consequently, as the interferometric baseline increases, an even smaller correlation is expected between differential phase and intensity. Even with the compact PFN system, the typical value for the Fried’s radius is 70 cm [@Roddier83] at 2.2 microns, to be compared with an interferometric baseline of 3.4 m. In the case of single-mode fiber injection, the intensities of the individual beams are primarily driven by the local tip-tilt and overall phase corrugations of the individual apertures, and have no relation to the differential phase between the apertures. This suggests that the absence of correlation between the different noise terms is to first order justified both for the PFN, and long baseline interferometry in general.\
![Typical intensity correlation measured during an observation with the PFN. The dashed dark-grey line correspond to the correlation between the beam1 intensity at time $t$ ($I_1(t)$) and the same beam intensity at time $t+\Delta t$ ($I_1(t+\Delta t)$). The dashed light-grey line represent the same correlation but computed for beam 2 and the black line is the correlation between the two different beams intensities for different time delays.[]{data-label="fig:int_correlation"}](correlation_a_b_0030.eps){width="10cm"}
The actual amount of correlation between the two beam intensities can be assessed by comparing the correlation of $I_1(t)$ with $I_1(t+\Delta t)$, $I_2(t)$ with $I_2(t+\Delta t)$ and $I_1(t)$ with $I_2(t+\Delta t)$. Figure \[fig:int\_correlation\] illustrates such a comparison for a typical dataset obtained with the PFN. For time delays close to zero both beam intensities are of course perfectly correlated with themselves (dark and light grey dashed curves). The correlation then decreases following a gaussian like curve until typical time delays of $\sim0.2$s are reached. The correlation is then very close to zero ($<$ to a couple of percent). This information directly gives us an indication of the atmospherical conditions. Indeed, as long as the turbulent cells stay above the individual apertures, some correlation will remain between the beam intensity measurements at times $t$ and $t+\Delta t$. Considering that our apertures are 1.5m wide, we expect to lose completely the correlation between $I_1(t)$ (resp. $I_2(t)$) and $I_1(t+\Delta t)$ (resp. $I_2(t+\Delta t)$) when $\Delta t$ is such that the turbulent cell has moved by more than $1.5$m. Given the correlation time obtained from Fig.\[fig:int\_correlation\], we can infer a wind speed during the observations of $\sim7.5$m/s which is consistent with typical conditions at Palomar Observatory. On the other hand, the profile of the correlation between $I_1(t)$ and $I_2(t+\Delta t)$ is completely different. Indeed, the measured values are always under $5\%$, even at short time delays. We can therefore quantitively confirm that even for interferometric observations with small baselines and operated under good atmospheric conditions, no significant correlation exists between the two beam intensities.
There is no physical reason why the background should correlate with any of the other terms. However it is possible that background intensity and the beam intensities are correlated to some extent if they are measured sequentially on the same detector (remanence). Such an effect depends on the hardware used for each instrument. We have computed its effect on the PFN measurements by computing the correlation between the mean beam intensities and background measurements over each chop cycle. We find that the correlation, if any, is smaller than $5\%$.
Differential phase distribution {#section_opd}
-------------------------------
For both statistical reduction techniques presented, the differential phase - computed at the central observing wavelength, see section \[sect:chrom\]- is assumed to exhibit a Gaussian distribution over the recorded nulling sequence. The validity of this assumption is difficult to assess from our data. As long as the instrument tracks around a constant optical path difference (OPD) position, it seems a reasonable assumption. In the case of the PFN, the two beams come from the same AO corrected wavefront. Tracking a single OPD comes down to the fact that the AO system, which essentially acts as a fringe tracker, tries to maintain the same reference flat wavefront over the sequence. Some studies have shown that indeed, the phase residuals after a AO system are Gaussian, which supports our assumption [@Cagigal00]. If for some reason the fringe tracker or AO system loses lock, or if the OPD is obviously oscillating between several distinct positions, the resulting distribution will no longer be Gaussian, and the corresponding data should be discarded. The reduced Pearson $\chi^{2}$ defining the quality of the probability distribution fit (Eq. \[pearson\]) is a good quantitative tool to assess the validity of the gaussian OPD distribution. If the measured $\chi^2$ are much larger than one, the error bars on the final astrophysical ND should be increased accordingly. Determining the potential bias caused by any departure from a Gaussian OPD distribution is beyond the scope of this paper, but can likely play a role for measuring reliable nulls at very low levels.
Temporal effects
----------------
The nulling expression established above (Eq. 8) is valid for instantaneous nulls. However, a photometer or camera will work with a limited frequency response or a finite individual integration time $\delta t$. In practice, this means that even when all of the instrumental terms of Eq 4 go through zero instantaneously, the measured null will in general be higher. Assuming that the polarization mismatch term is negligible, the best measurable null at any time t will be limited to:
$$N_{min}= \frac{ \sigma^2_{\delta I(t, dt)} + \sigma^2_{\Delta \phi(t, dt)} } {4}$$
\
where $\sigma^2_{\delta I(t, dt)} $ and $\sigma^2_{\Delta \phi(t, dt)}$ are respectiveley the variance of the intensity mismatch and of the differential phase, both measured over a time interval $\delta t$. The effect of finite temporal integration is then to cause a (positive) bias to the observed null depth. If the individual integrations are short enough compared to the typical fluctuation timescale, this bias can be kept to a very low level. Moreover, it could be at least partially calibrated via observations of reference stars. In the case of the PFN for instance, we use 2-10 ms individual integrations, to be compared with $\simeq$ 100 ms for the typical coherence time of atmospheric turbulence at K band. Using a Kolmogorov spectrum for the turbulence, and using the PFN short baseline, we find for instance that the atmospheric phase rms is less than 1 nm over 10 ms, limiting the minimum null depth $\simeq 2 \times 10^{-6}$ . Similarly, the intensity mismatch term follows atmospheric timescales, and its variance over 10 ms is not expected to cause any significant bias either. Laboratory nulling experiments with fiber nuller set-ups have already produced 10 ms nulls at the $\simeq 10^{-6}$ level with visible laser light [@Haguenauer06], and $10^{-4}$ nulls with dual polarization broad-band light over the full K band. In the latter case, dispersive and/ or polarization effects are likely dominating the error budget, and the effect of finite integration is not found to play a role up to 50 ms. Finally, astrophysical nulls at the 0.001 (or even slightly lower) level have been measured on Vega with the PFN (Mennesson et al. , in prep.), showing experimentally that temporal effects are at most at this level (and probably much smaller) on the PFN. The optimum individual integration time is thus a trade-off between sensitivity and dynamic range.
Chromatic effects {#sect:chrom}
-----------------
Usually, interferometric / nulling observations are conducted over a finite spectral bandwidth. We concentrate here on the effects of the chromatic phase term, expected to dominate over the chromatic aspects of intensity or polarization mismatch. For a polychromatic observation, the phase error ($\Delta \phi(t)$) is the sum of the piston error calculated at band center ($\Delta\phi_c(t)$) and the chromatic phase error ($\Delta\phi_\lambda(\lambda,t)$). @Serabyn00 has demonstrated that the influence of these phase errors on a polychromatic null depth measurement is given by
$$\begin{aligned}
N_\phi(t) = \frac{\Delta \phi_c^2(t)}{4} + N_{\rm chrom}
\label{eq_phase_error3}\end{aligned}$$
\
where $N_{chrom}$ = $\int_{\lambda_{min}}^{\lambda_{max}}\frac{\Delta\phi_\lambda^2(\lambda, t)}{4}\, d\lambda$ is the chromatic null bias . So even in the case where the differential phase at the center of the band is zero, a positive bias is present (either constant or slowly varying, see below) , and one measures $N_\phi(t)= N_{chrom}$. This additive bias directly impacts the astrophysical null depth measurement.
In the case of the PFN, this chromatic term is minimized by inserting glass plates of different thickness in each of the two beams. The chromatic bias is experimentally found to be lower than $10^{-4}$ in the laboratory. On the sky, the dispersive phase is no longer a strictly static term coming from the instrument. It is also impacted by differential atmospheric refraction across the band, and varies over the night according to the target’s position with respect to zenith. Detailed calculations are beyond the scope of this paper, but this effect is small ($<10^{-4}$) across the K band with the PFN short baseline when observing within 20 degrees of zenith. Additionally, solutions exist to strongly reduce or completely eliminate this effect: disperse over several spectral bins, always orient the interferometric baseline perpendicular to the refraction direction (trivial on a single-dish interferometer with multiple sub-apertures), or use atmospheric dispersion compensators at the telescope. Moreover, this refraction effect is fortunately very repeatable, and can be precisely calibrated by observing reference stars at the same zenith angle.
Summary of limitations
----------------------
The assumptions proper to the analytical method (gaussian distribution of background and intensity terms, correlation issues) seem all individually valid in the case of the PFN. The analytical method also provides very similar results to those obtained by the numerical method, which makes fewer assumptions. The assumption that the differential phase follows a gaussian distribution can not be directly checked with the PFN data, but seems reasonable wrt theoretical expectations.
Temporal and chromatic effects (as well as polarization effects, which we completely ignored for the PFN), may come into play at the $10^{-4}$ level, even more when considering the application to LBI which uses very long non common beam paths. However, these systematic effects - slowly varying for the most part-, can be either minimized by instrumental design, or strongly reduced via observations of calibrator stars.
A more serious limitation to the reduction method presented is that its ultimate sensitivity may be limited by the small integration times needed to freeze the phase and intensity fluctuations. Infrared cameras with very low read noise will definitely help. Taking long sequences will also help, up to the point where systematics will dominate. More work is clearly needed to understand the trade-off between individual integration time, sensitivity and final accuracy.
Conclusions
===========
The theory of a new data reduction method for interferometric nulling (or visibility) observations has been presented in this paper. Based on the analysis of null distributions, this technique allows the retrieval of high dynamic range astrophysical null depth measurements, at contrast levels far exceeding the usual limits set by mean instrumental performance and fluctuations. The ultimate performance of this statistical data reduction depends on the specific design of the interferometric instrument and on the observing strategy. This technique is potentially applicable to any interferometric set-up using a fringe tracking capability and any type of beam recombination (co-axial or multi-axial) into a single-mode waveguide. Applying our data reduction method to stellar observations obtained at K-band ($\simeq$ 2.2 microns) with the first generation fiber nulling instrument installed at the Palomar 5m (Hale) telescope, we demonstrated for the first time that: (i) deep and accurate nulling is not restricted to mid-infrared wavelengths but may be extended to the near infrared domain, providing substantial gains in resolution and sensitivity and (ii) nulling accuracies significantly lower than $10^{-3}$ (systematics and statistical errors included) can be achieved, without any observation of calibrator stars. Although this remains to be further validated with an optimized instrument, simulations suggest that this new analysis will enable direct detection of faint structures at the $\simeq10^{-4}$ level within the near diffraction limit of large AO equipped ground based telescopes, i.e at angular separations ranging from $\simeq$ 20 to 150 mas. Implications for high accuracy long baseline interferometry, both from the ground and from space, remain to be quantitatively explored. But since the statistical approach allows the detection of astrophysical signals well below the mean contrast level [*[and its rms fluctuations]{}*]{}, we anticipate that the instrumental stability requirements could be strongly relaxed. This implies that the constraints on intensity and phase fluctuations may be strongly reduced. This is a most attractive prospect for deep nulling interferometry from space. A similar statistical analysis may also be conducted successfully for regular coronagraphic instruments [@Riaud10].
Confidence intervals and covariance {#app:a}
===================================
In section \[section\_fit\], we described our fitting strategies, and developed the minimization process used to fit the distribution corresponding to an individual null sequence. The statistical error bar $\sigma_{\rm stat}$ on the derived astrophysical null depth is then determined by applying small fluctuations to $N_a$ around its best fit value. For every new value of $N_a$, the two other parameters ($\mu_{\Delta\phi}$ and $\sigma_{\Delta\phi}$) are adjusted to minimize the $\chi^2$. As the number of degrees of freedom of our system is known and is $N_{\rm bins}-4$, it is possible to calculate the $\Delta\chi^2$ relative to a certain confidence level. The error bars are generally evaluated for $1\sigma$ confidence levels, and so we use this criteria. $\sigma_{\rm stat}$ corresponds to the increment in $N_a$ required to increase the reduced $\chi^2$ from its minimum value $\chi^{2}_{\rm min}$ to $\chi^{2}_{\rm min}+\Delta\chi^2$. For the dataset obtained on $\alpha$Boo with the PFN, the $1\sigma$ error bar corresponds to a very small $\chi^2$ increment, $\Delta\chi^2=0.07$ and we find $\sigma_{\rm stat}= 3\times10^{-4}$ for the NSC (see Fig.\[fig:covariance\], left). Another way of calculating the error bars consist in using bootstrapping methods. We double-checked our confidence interval using this technique and found similar error bars ($\simeq 3\times10^{-4}$). This error bar takes into account the fitting noise that is not present for the ASC.
![Left: variation of the reduced $\chi^2$ – measuring the goodness of the fit to the observed data – as a function of the astrophysical null depth $N_{a}$. The mean and standard deviation of the phase error are left as free parameters, and adjusted to minimize the $\chi^{2}$ for each new value of $N_a$. Center: projected $\chi^2$ map of our model in the $N_a$ vs $\mu_{\Delta\phi}$ plane. For each point, $\sigma_{\Delta\phi}$ is chosen to minimize the $\chi^2$. Right: same map but projected in the $N_a$ vs $\sigma_{\Delta\phi}$ plane. For these two maps, the contours are over-plotted for each $\chi^2$ intervals of 0.5.[]{data-label="fig:covariance"}](chi_carre_null2.eps "fig:"){height="4.5cm"} ![Left: variation of the reduced $\chi^2$ – measuring the goodness of the fit to the observed data – as a function of the astrophysical null depth $N_{a}$. The mean and standard deviation of the phase error are left as free parameters, and adjusted to minimize the $\chi^{2}$ for each new value of $N_a$. Center: projected $\chi^2$ map of our model in the $N_a$ vs $\mu_{\Delta\phi}$ plane. For each point, $\sigma_{\Delta\phi}$ is chosen to minimize the $\chi^2$. Right: same map but projected in the $N_a$ vs $\sigma_{\Delta\phi}$ plane. For these two maps, the contours are over-plotted for each $\chi^2$ intervals of 0.5.[]{data-label="fig:covariance"}](chi_square_map_na_muphi.eps "fig:"){height="4.5cm"} ![Left: variation of the reduced $\chi^2$ – measuring the goodness of the fit to the observed data – as a function of the astrophysical null depth $N_{a}$. The mean and standard deviation of the phase error are left as free parameters, and adjusted to minimize the $\chi^{2}$ for each new value of $N_a$. Center: projected $\chi^2$ map of our model in the $N_a$ vs $\mu_{\Delta\phi}$ plane. For each point, $\sigma_{\Delta\phi}$ is chosen to minimize the $\chi^2$. Right: same map but projected in the $N_a$ vs $\sigma_{\Delta\phi}$ plane. For these two maps, the contours are over-plotted for each $\chi^2$ intervals of 0.5.[]{data-label="fig:covariance"}](chi_square_map_na_suphi.eps "fig:"){height="4.5cm"}
\
The central and right panels of Fig. \[fig:covariance\] represent the normalized $\chi^2$ of our fits projected in two different parameters planes (i.e. $N_a$ vs. $\mu_{\Delta\phi}$ for the central panel and $N_a$ vs. $\sigma_{\Delta\phi}$ for the right one). The contours on these maps are displayed for increments of the $\chi^2$ of $\Delta\chi^2=0.5$. These maps illustrate the covariance of the fits with the two free parameters: the mean and standard deviation of the phase error fluctuations. They show that relatively large variations on the fitted values of these phase parameters, between 0.05 and 0.1 rad, only produce a marginal effect on the measured astrophysical null, smaller than $10^{-3}$, but produce very large effects on the fit quality. This result is important as it clearly illustrates the resilience of our approach to possible error on the assessment of the phase fluctuations.
[^1]: For a given baseline orientation, the astrophysical null $N_a$ can be expressed in terms of the source complex visibility $\mathcal{V}$ as $N_a = (1-|\mathcal{V}|)/(1+|\mathcal{V}|).$
[^2]: The theory we present here can be extended for larger values of $N_{a}$, and of the error sources by keeping the full expression of $I_{\pm}^{*}$ in the definition of the null depth.
[^3]: For long baseline interferometers, the polarization effect can be measured on calibrator stars and accurately corrected as it generally varies slowly over time.
[^4]: The null depth in interferometry is generally considered to be defined between 0 and 1. However, the instantaneous measured null can be $<0$ because of the background fluctuations. This is why the limit of integration must be defined between the minimum measured null depth and 1.
[^5]: Note that in the case of the numerical method, only the phase error distribution must be assumed to be Gaussian.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is argued that the hollowness effect (depletion in the absorptive part of the scattering cross section at small values of the impact parameter) in the proton-proton scattering at the the LHC energies finds its origin in the quantum nature of the process, resulting in large values of the real part of the eikonal phase. The effect cannot be reconciled with an incoherent superposition of the absorption from the proton constituents, thus suggests the change of this basic paradigm of high-energy scattering.'
address:
- |
${}^{1}$The H. Niewodniczański Institute of Nuclear Physics,\
Polish Academy of Sciences, PL-31342 Cracow, Poland
- '${}^{2}$Institute of Physics, Jan Kochanowski University, PL-25406 Kielce, Poland'
- |
${}^{3}$Departamento de Física Atómica, Molecular y Nuclear and\
Instituto Carlos I de Fisica Teórica y Computacional, Universidad de Granada, E-18071 Granada, Spain
author:
- 'Wojciech Broniowski$^{1,2}$[^1] and Enrique Ruiz Arriola$^3$[^2]'
bibliography:
- 'NN-high-energy.bib'
title: 'Hollowness in $pp$ scattering[^3] [^4]'
---
=1
PACS: [13.75.Cs, 13.85.Hd]{}
In this talk we discuss the significance of the recent $pp$ scattering results from the Large Hadron Collider for our understanding of the underlying physical processes in highest-energy collisions. In particular, we argue that the [*hollowness*]{} in the inelastic cross section treated as a function of the impact parameter $b$, i.e., its depletion at low $b$, must necessarily originate from quantum coherence, precluding a probabilistic folding interpretation. More details of our analysis can be found in [@Arriola:2016bxa; @RuizArriola:2016ihz], where we also analyze the effect in 3-dimensions via the optical potential interpretation.
The TOTEM [@Antchev:2013gaa] and ATLAS (ALFA) [@Aad:2014dca] Collaborations have measured the differential elastic cross section for the $pp$ collisions at $\sqrt{s}=7$ TeV, later repeated for $\sqrt{s}=8$ TeV [@Antchev:2013paa; @Aaboud:2016ijx]. When the data are used to obtain the inelastic cross section in the impact-parameter representation, a striking feature appears: there is more inelasticity when the two protons are separated by about half a fermi in the traverse direction than for the head-on collisions. We term this phenomenon [*hollowness*]{}. This unusual feature has been brought up and interpreted by other authors [@Alkin:2014rfa; @Dremin:2014eva; @Dremin:2014spa; @Dremin:2016ugi; @Dremin:2017ylm; @Anisovich:2014wha; @Troshin:2016frs; @Troshin:2017zmg; @Troshin:2017ucy]. A model realization of the effect was implemented via hot-spots in [@Albacete:2016pmp].
We use the parametrization of the $pp$ scattering data [@Fagundes:2013aja] based on the Barger-Phillips model (modified BP2) [@Phillips:1974vt], with the form $$\begin{aligned}
{\cal A} (s,t) &\equiv& \frac{f(s,t)}{p} =
\sum_{n} c_n(s) F_n(t) s^{\alpha_n (t)} =\frac{i \sqrt{A} e^{\frac{B t}{2}}}{\left(1-\frac{t}{t_0}\right)^4}+i \sqrt{C} e^{\frac{D t}{2}+i \phi },
\nonumber \\ \label{eq:mBP2}\end{aligned}$$ where $f(s,t)$ is the quantum mechanical scattering amplitude. The modified BP2 model deals with the $t$ dependence, and the $s$-dependent parameters are fitted separately to the differential elastic $pp$ cross sections at $\sqrt{s}= 23.4$, $30.5$, $44.6$, $52.8$, $62.0$, and $7000~{\rm GeV}$. A typical quality of the fit, from the ISR [@Amaldi:1979kd] at $\sqrt{s}=23.4$ GeV to the LHC at $\sqrt{s}=7$ TeV, can be appreciated from Fig. \[fig:data\](a). These fits are not sensitive to the [*phase*]{} of the scattering amplitude.
The $\rho(s)$ parameter is defined as the ratio of the real to imaginary parts of the amplitude at $t=0$: $$\begin{aligned}
\rho(s) = \frac{{\rm Re}{\cal A} (s,0) }{{\rm Im}{\cal A} (s,0)}
$$ This parameter has been recently determined for the LHC energy of $\sqrt{s}=8$ TeV in [@Antchev:2016vpy]. To agree with this experimental constraint we replace the parametrization of the scattering amplitude of Eq. (\[eq:mBP2\]) with $$\begin{aligned}
{\cal A} (s,t) \to \frac{i+\rho(s)}{\sqrt{1+\rho(s)^2}}|{\cal A} (s,t)|. \label{eq:rhoind}\end{aligned}$$ This procedure assumes a $t$-independent ratio of the real to imaginary parts of the scattering amplitude for all $t$-values, which is the simplest choice. More general prescriptions have been analyzed in detail in Ref. [@Antchev:2016vpy]. Our results presented below are similar if we take, e.g., the Bailly et al. [@Bailly:1987ki] parametrization $\rho(s,t)=\rho_0(s)/(1-t/t_0(s))$, where $t_0(s)$ is the position of the diffractive minimum. However, admittedly, there is some dependence on the choice of the model of $\rho(s,t)$. Moreover, the problem is linked to the separation of the Coulomb and strong amplitudes. The issue is crucial for the proper extraction of the physical results and the ambiguity has a long history since the early diagrammatic work of West and Yennie [@West:1968du], which is consistent with the eikonal approximation [@Cahn:1982nr; @Block:1984ru] but becomes sensitive to internal structure from electromagnetic information such as form factors (see, e.g., [@Prochazka:2016wno] and references therein).
![(a) The data for the ISR energy of $\sqrt{s}=23.4$ GeV [@Amaldi:1979kd] and the LHC energy of $\sqrt{s}=7$ TeV [@Antchev:2013gaa] with overlaid fits according to Eq. \[eq:mBP2\]. (b) Plot of the integrand of Eq. (\[eq:invf\]), showing that that the range of the experimental data in $q$ is sufficient to carry out the Fourier-Bessel transform for the values of $b$ of interest. \[fig:data\]](exp.pdf "fig:"){width="51.00000%"} ![(a) The data for the ISR energy of $\sqrt{s}=23.4$ GeV [@Amaldi:1979kd] and the LHC energy of $\sqrt{s}=7$ TeV [@Antchev:2013gaa] with overlaid fits according to Eq. \[eq:mBP2\]. (b) Plot of the integrand of Eq. (\[eq:invf\]), showing that that the range of the experimental data in $q$ is sufficient to carry out the Fourier-Bessel transform for the values of $b$ of interest. \[fig:data\]](int_2.pdf "fig:"){width="48.00000%"}
$\sqrt{s}$ \[GeV\] $\sigma_{\rm el}$ \[mb\] $\sigma_{\rm in}$ \[mb\] $\sigma_{\rm T}$ \[mb\] $B~[{\rm GeV}^{-2}]$ $\rho$
------------------------------------- -------------------------- -------------------------- ------------------------- ---------------------- ------------
23.4 6.6 31.2 37.7 11.6 0.00
[@Amaldi:1979kd] 6.7(1) 32.2(1) 38.9(2) 11.8(3) 0.02(5)
200 10.0 40.9 50.9 14.4 0.13
[@Aielli:2009ca; @Bueltmann:2003gq] 54(4) 16.3(25)
7000 25.3 73.5 98.8 20.5 0.140
[@Antchev:2013gaa] 25.4(11) 73.2(13) 98.6(22) 19.9(3) 0.145(100)
: Basic scattering observables for several collision energies obtained from Eq. (\[eq:rhoind\]), compared to experimental vales (lower rows). $B$ is the slope parameter of the differential elastic cross section.
\[tab:Fag-Mod\]
Our prescription (\[eq:rhoind\]) maintains by construction the quality of the fits shown in Fig. \[fig:data\], but also the experimental values for $\rho(s)$ are reproduced, which would not be the case if Eq. (\[eq:mBP2\]) were used. Basic physical quantities stemming from our method are listed in Table \[tab:Fag-Mod\], with good agreement with the data supporting the used parametrization.
We now recall the relevant formulas from scattering theory: The $pp$ elastic differential cross section is given by $$\begin{aligned}
\frac{d\sigma_{\rm el}}{dt}= \frac{\pi}{p^2} \frac{d \sigma_{\rm el}}{d \Omega} =
\frac{\pi}{p^2} |f(s,t) |^2 = \pi | {\cal A} (s,t) |^2\, , \end{aligned}$$ with $p= \sqrt{s/4-M^2}$ the CM momentum and the partial wave expansion of the scattering amplitude (we neglect spin effects) equal to $$\begin{aligned}
f(s,t) =\sum_{l=0}^ \infty (2l+1) f_l(p) P_l(\cos \theta). \label{eq:PWA}\end{aligned}$$ The total cross section is given by the optical theorem, $\sigma_T =
4 \pi {\rm Im} f(s,0)/p$, and Coulomb effects are negligible at $|t| >
8 \pi \alpha /\sigma_T$, where $\alpha \simeq 1/137$ is the QED fine structure constant and $\sigma_T$ is the total strong scattering cross section. For $p a \gg 1$, with $a$ denoting the interaction range, one can use the eikonal approximation with $bp = l +1/2 + {\cal
O}(s^{-1})$, where $b$ is the impact parameter. The $b$ representation the scattering amplitude can be straightforwardly obtained from a Fourier-Bessel transform of $f(s,t)$, known from the data parametrization. Explicitly, $$\begin{aligned}
2ph(b,s)={i} \left [ 1-e^{i \chi(b)} \right ] =2p f_l(p) + {\cal O}(s^{-1}) = 2 \int_0^\infty q dq J_0(bq) f(s,-q^2).
\nonumber \\ \label{eq:invf}\end{aligned}$$ In Fig. \[fig:data\](b) we demonstrate that the range of the TOTEM data in $q$ is sufficient to carry out this transform to a satisfactory accuracy needed in our analysis.
![Real (lower curves) and imaginary (upper curves) parts of the eikonal scattering amplitude $2ph(b)$ for several collision energies. We note that for the LHC energies, at the origin $2p {\rm Im} h(0)>1$. \[fig:amp\]](ph2.pdf){width="49.00000%"}
![The total (a), elastic (b), and inelastic (c) cross section, as well as the edge function, plotted as functions of the impact parameter at various collision energies. \[fig:sigmas\]](sigtot.pdf "fig:"){width="49.00000%"} ![The total (a), elastic (b), and inelastic (c) cross section, as well as the edge function, plotted as functions of the impact parameter at various collision energies. \[fig:sigmas\]](plela.pdf "fig:"){width="49.00000%"}\
![The total (a), elastic (b), and inelastic (c) cross section, as well as the edge function, plotted as functions of the impact parameter at various collision energies. \[fig:sigmas\]](inel2.pdf "fig:"){width="49.00000%"} ![The total (a), elastic (b), and inelastic (c) cross section, as well as the edge function, plotted as functions of the impact parameter at various collision energies. \[fig:sigmas\]](edge2.pdf "fig:"){width="50.10000%"}\
The standard formulas for the total, elastic, and inelastic cross sections (in our analysis we treat all the components to the inelastic scattering jointly, not discriminating, e.g., the diffractive components) in the $b$ representation can be parameterized with the eikonal phase $\chi(b)$ and have the form [@Blankenbecler:1962ez] $$\begin{aligned}
\sigma_T &=& \frac{4 \pi}p {\rm Im} f(s,0) = 4 p \int d^2 b {\rm Im} h(\vec b,s) = 2 \int d^2 b \left [ 1- {\rm Re} \, e^{i \chi(b)} \right ] \, \label{eq:st} ,\\
\sigma_{\rm el} &=& \int d\Omega |f(s,t)|^2 = 4 p^2 \int d^2 b |h(\vec b,s)|^2 = \int d^2 b | 1- e^{i \chi(b)} |^2 \, \label{eq:sel}, \\
\sigma_{\rm in} &\equiv& \sigma_T - \sigma_{\rm el} = \int d^2 b \sigma_{\rm in} (b) = \int d^2 b \left [ 1- e^{- 2 {\rm Im} \chi (b)} \right ] , \label{eq:sin}\end{aligned}$$ with the integrands $\sigma_{\rm in} (b)$, $\sigma_{\rm el} (b)$ and $\sigma_{\rm T} (b)$ being dimensionless quantities that can be interpreted as the corresponding $b$-dependent relative number of collisions. For instance, accordingly to Eq. (\[eq:sin\]), the inelasticity profile is defined as $$\begin{aligned}
\sigma_{\rm in} (b) = 4p {\rm Im} h(b,s) - 4p^2|h(b,s)|^2. \label{eq:prof}\end{aligned}$$ While unitarity implies $\sigma_{\rm in}(b)> 0$, one also has $\sigma_{\rm in}(b) \le 2 k(b,s) - k(b,s)^2$, with $k(b,s) \equiv 2 p
{\rm Im h(b,s)}$, and hence one also has the upper bound .
. \[fig:closeup\]](nin2.pdf){width="49.00000%"}
Now we come to our results. In Fig. \[fig:amp\] we present the real and imaginary parts of the eikonal amplitude $2ph(b)$ for several collision energies. The real parts are smaller from the corresponding imaginary parts, as their ratio is given by the (constant) $\rho$ parameter. The important observation here is that the imaginary parts go above 1 near the origin for the LHC collision energies. We will come back to this issue shortly.
![Imaginary (a) and real (b) part of the eikonal scattering phase, plotted as functions of the impact parameter for several collision energies. We note that at the LHC energies ${\rm Re}\chi(b=0)$ goes above $\pi/2$. \[fig:chi\]](imchi.pdf "fig:"){width="49.00000%"} ![Imaginary (a) and real (b) part of the eikonal scattering phase, plotted as functions of the impact parameter for several collision energies. We note that at the LHC energies ${\rm Re}\chi(b=0)$ goes above $\pi/2$. \[fig:chi\]](rechi.pdf "fig:"){width="49.00000%"}
In Fig. \[fig:sigmas\] we collect the results for the impact-parameter representations of the total, elastic, and inelastic cross sections, as well as for the [*edge*]{} function [@Block:2014lna; @Block:2015sea], defined as $\sigma_{\rm in}(b)-\sigma_{\rm el}(b)$. The most important feature, visible from Fig. \[fig:sigmas\] and more accurately form the close-up of Fig. \[fig:closeup\], is the [*hollowness*]{}: the inelastic cross section develops a minimum at $b=0$ at the LHC collision energies.
To better understand these results, one should resort to the formulas expressed with the eikonal phase, plotted in Fig. \[fig:chi\]. We have $$\begin{aligned}
2p{\rm Im}\, h(b) &=& 1-e^{-{\rm Im}\chi(b)}\cos {\rm Re}\chi(b), \label{eq:eik2} \\
2p{\rm Re}\, h(b) &=& e^{-{\rm Im}\chi(b)}\sin {\rm Re}\chi(b), \nonumber \\
\sigma_T(b)&=& 2 - 2 e^{-{\rm Im}\chi(b)} \cos {\rm Re}\chi(b), \nonumber \\
\sigma_{\rm el}(b)&=& 1 + e^{-2 {\rm Im}\chi(b)} - 2 e^{-{\rm Im}\chi(b} \cos {\rm Re}\chi(b), \nonumber \\
\sigma_{\rm in}(b)&=& 1 - e^{-2 {\rm Im}\chi(b)}, \nonumber \\
\sigma_{\rm el}(b)-\sigma_{\rm in}(b)&=& 2 e^{-{\rm Im}\chi(b)} \left[ \cos{\rm Re} \chi(b)-e^{-{\rm Im}\chi(b)} \right]. \nonumber\end{aligned}$$ We note several facts following from the above relations:
1. Going of $2p{\rm Im}\, h(b)$ above 1 and $\sigma_T(b)$ above 2 are caused by ${\rm Re} \chi(b)>\pi/2$, where $\cos {\rm Re}\chi(b)<0$ (cf. Figs. \[fig:amp\], \[fig:sigmas\](a), and \[fig:chi\](b)).
2. In addition, if $\chi(b)>\pi/2$, the edge function is negative and $\sigma_{\rm
el}(b)>1$.
3. The departure of $2p{\rm Im}\, h(b)$ from 1 is of similar order as $2p{\rm Re}\, h(b)$, with both suppressed with $e^{-{\rm Im}\chi(b)}$.
We see that this is the [*real*]{} part of the eikonal phase which controls the behavior related to hollowness.
![Illustration of Eq. (\[eq:ga\]). The solid line corresponds to the experimental values of the ratio of the elastic to total $pp$ cross section. Values of $\sigma_{\rm el}/\sigma_T>1/4$ correspond to hollowness in the Gaussian model. \[fig:parab\]](parab2.pdf){width="49.00000%"}
One may give a simple criterion for $\sigma_{\rm in}(b)$ to develop a minimum at $b=0$. From Eqs. (\[eq:prof\]) and (\[eq:rhoind\]) we get $$\begin{aligned}
\frac{d\sigma_{\rm in}(b)}{db^2}= 2p \frac{d{\rm Im}\, h(b)}{db^2} \left [ 1-(1+\rho^2) 2p {\rm Im}\, h(b) \right ],\end{aligned}$$ which is negative at the origin if $$\begin{aligned}
2 p {\rm Im} h (0)>\frac{1}{1+\rho^2}\sim 1. \label{eq:crit}\end{aligned}$$ Since $\rho=0.14$ at the LHC, the departure of $1/(1+\rho^2)$ from 1 is at a level of $2\%$.
We also find from Eq. (\[eq:eik2\]) that $$\begin{aligned}
\frac{d\sigma_{\rm in}(b)}{db^2} = 2 e^{-2 {\rm Im}\chi(b)} \frac{d {\rm Im}\chi(b)}{db^2},\end{aligned}$$ thus the appearance of the dip at the origin in $\sigma_{\rm in}(b)$ is associated with the dip in ${\rm Im}\chi(b)$. This is manifest between Fig. \[fig:closeup\] and Fig. \[fig:chi\](a).
Dremin [@Dremin:2014eva; @Dremin:2014spa; @Dremin:2016ugi] proposed a simple Gaussian model of the amplitude which one may adapt to the presence of the real part of the amplitude (which is crucial for maintaining unitarity with the hollowness effect). One can parametrize the amplitude at low values of $b$ (which is the numerically relevant region) as $$\begin{aligned}
&& {{\rm Im}(2p \, h(p)) = A e^{-\frac{2b^2}{2B}} }, \;\;
A=\frac{4 \sigma_{\rm el}}{\left(1+\rho^2\right) \sigma_{\rm tot}}, \;\; B=\frac{\left(1+\rho^2\right) \sigma_{\rm tot}^2}{16 \pi \sigma_{\rm el}}.\end{aligned}$$ The curvature of the inelasticity profile at the origin is $$\begin{aligned}
\frac{1}{2} \left . \frac{d^2 n_{\rm in}(b)}{db^2} \right |_{b=0} =
\frac{64 \pi \sigma_{\rm el}^2 (4\sigma_{\rm el}- \sigma_{\rm
tot})}{\left(\rho^2+1\right)^2 \sigma_{\rm tot}^4},\end{aligned}$$ We note it changes sign when [$\sigma_{\rm el}=\frac{1}{4}\sigma_{\rm tot}$]{}, with the value at the origin $$\begin{aligned}
\sigma_{\rm in}(0)= \frac{8\sigma_{\rm el}}{(1+\rho^2) \sigma_{\rm tot}} \left ( 1 - 2 \frac{\sigma_{\rm el}}{\sigma_{\rm tot}} \right). \label{eq:ga}\end{aligned}$$ As predicted by Dremin, the hollowness effect emerges when $\sigma_{\rm el}>\frac{1}{4}\sigma_{\rm tot}$, which is the case of the LHC collision energies. We illustrate relation (\[eq:ga\]) in Fig. \[fig:parab\].
The final point, very important from the conceptual point of view and for the understanding of the effect, is the impossibility of hollowness to emerge from incoherent folding of inelasticities of collisions of the protons’ partonic constituents. In many models incoherent superposition is assumed, i.e., the inelasticity of the $pp$ process is obtained from the folding formula shown below. These ideas have been implemented in microscopic models based on intuitive geometric interpretation [@Chou:1968bc; @Chou:1968bg; @Cheng:1987ga; @Bourrely:1978da; @Block:2006hy; @Block:2015sea]. Folding involves $$\begin{aligned}
\sigma_{\rm in}(b) &\propto& \int d^2 b_1 d^2 b_2
\rho( \vec b_1 + \vec b/2) w(\vec b_1 - \vec b_2 ) \rho(\vec b_2-\vec b/2) \nonumber \\ &=& \int d^3 b_1 d^3 b_2 \rho(\vec b_1) w(\vec b_1 - \vec b_2 ) \rho(\vec b_2) \nonumber \\
&-& \frac12 \int d^3 b_1 d^3 b_2 [\vec b \cdot \nabla \rho(\vec b_1)]
w(\vec b_1-\vec b_2) [\vec b \cdot \nabla \rho(\vec b_2)]+ \dots,\end{aligned}$$ where $w(\vec b_1-\vec b_2)$ is a positive-definite kernel (folding models usually take $w(\vec b_1-\vec b_2) \propto \delta (\vec b_1
-\vec b_2)$) and $\rho(\vec{b})$ describes the (possibly correlated) transverse distribution of components in the proton. By passing to the Fourier space it is simple to show that [$\sigma_{\rm
in}(b)=\alpha^2 - \beta^2 b^2+\dots$]{}, with real constants $\alpha$ and $\beta$, therefore $\sigma_{\rm in}(b)$ has necessarily a [local maximum]{} at $b=0$, in contrast to the phenomenological hollowness result at the LHC energies. An analogous argument holds for the 3D-hollowness unveiled in our work [@Arriola:2016bxa; @RuizArriola:2016ihz], which takes place already at lower energies.
In conclusion, we stress that the hollowness effect in $pp$ scattering at the LHC energies has necessarily a quantum origin. As just shown, it cannot be obtained by an incoherent folding of inelasticities of collisions of partonic constituents. Moreover, we have demonstrated that the real part of the scattering amplitude plays a crucial role in generating hollowness: the effect appears when the real part of the eikonal phase becomes larger than $\pi/2$. Per se, there is nothing unusual in that fact. If coherence occurs, the phases of amplitudes from the constituents may add up (as is the case, e.g., in the Glauber model [@glauber1959high]) and at some point the value of $\pi/2$ may be crossed. A microscopic realization of this quantum mechanism remains, however, a challenge. Finally, we note that in [@Arriola:2016bxa; @RuizArriola:2016ihz] we have presented a three-dimensional interpretation of the effect, which offers an even more pronounced hollowness feature.
[^1]: [email protected]
[^2]: [email protected]
[^3]: Talk presented by WB at XXIII Cracow EPIPHANY Conference, 9-12 January 2017.
[^4]: Supported by Polish National Science Center grant 2015/19/B/ST2/00937, by Spanish Mineco Grant FIS2014-59386-P, and by Junta de Andalucía grant FQM225-05.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work we focus on a group of Galactic double neutron star (DNS) systems with long orbital periods of $ \gtrsim 1$ day and low eccentricities of $\lesssim 0.4$. The feature of these orbital parameters is used to constrain the evolutionary processes of progenitor binaries and the supernova (SN) kicks of the second born NSs. Adopting that the mass transfer during primordial binary evolution is highly non-conservative (rotation-dependent), the formation of DNS systems involves a double helium star binary phase, the common envelope (CE) evolution initiates before the first NS formation. During the CE evolution the binary orbital energy is obviously larger when using a helium star rather than a NS to expel the donor envelope, this can help explain the formation of DNS systems with long periods. SN kicks at NS birth can lead to eccentric orbits and even the disruption of binary systems, the low eccentricities require that the DNSs receive a small natal kick at the second collapse. Compared with the overall distribution of orbital parameters for observed DNS binaries, we propose that the second born NSs in most DNS systems are subject to small natal kicks with the Maxwellian dispersion velocity of less than $ 80 \,\rm km\,s^{-1} $, which can provide some constraints on the SN explosion processes. The mass distribution of DNS binaries is also briefly discussed. We suggest that the rotation-dependent mass transfer mode and our results about SN kicks should be applied to massive binary evolution and population synthesis studies.'
author:
- 'Yong Shao$^{1,2}$ and Xiang-Dong Li$^{1,2}$'
title: On the role of supernova kicks in the formation of Galactic double neutron star systems
---
Introduction
============
More than 40 years has passed since the discovery of the first double neutron star (DNS) system PSR B1913+16 [@ht75], and there are now over a dozen known in our Galaxy (see Table 1). The orbital periods $ P_{\rm orb} $ of the DNS binaries are distributed in a wide range of $ \sim 0.1- 45 $ days, around half are close binaries ($ P_{\rm orb} < 0.5$ day) and the other half wide binaries ($ P_{\rm orb} > 1$ day). The orbital eccentricities $ e $ vary in a range of $ \sim 0.1-0.8 $, and $ \sim 80\% $ of them have a relatively low value of $ e < 0.4 $. The components of the DNS systems generally have a similar mass of $ \sim 1.2-1.4 M_{\odot} $ except for PSR J0453+1559. The observed pulsars are usually mildly recycled, with the spin periods $ P_{\rm spin} $ in the range of $ 17-185 \rm\, ms$ and the surface magnetic fields of the order $ 10^{8}-10^{10} \rm\,G$.
Many previous works [e.g., @fv75; @ty93; @l97; @dp03; @i03; @t17] have investigated the formation and evolution of DNSs which involve a variety of binary evolutionary stages. The canonical channel for the DNS formation has already been established [@bv91; @tv06 for reviews]. In a primordial binary with both components more massive than $ 8 M_{\odot} $, the primary star evolves to firstly fill its Roche lobe and transfer material to the secondary star. After that the primary becomes a helium star, which finally explodes with a supernova (SN) to be a NS. Then the secondary star evolves and transfers mass to the NS, and the binary becomes a high-mass X-ray binary (HMXB). Because of the large mass ratio, the mass transfer is dynamically unstable and binary system enters a common envelope (CE) phase [see recent review by @i13] . During the CE phase, the orbital energy of the NS is used to dispel the donor’s envelope, and the orbital period is dramatically decreased. As the stripped donor (a helium star) finally becomes a NS through SN explosion, a DNS is formed if the binary is not disrupted. The subsequent evolution of the DNS system is driven by gravitational wave radiation, and its orbital decay can lead to a DNS merger event associated with a short gamma-ray burst [@a17].
Modelling the DNS formation is subject to a number of uncertainties, especially those in the processes of CE evolution, SN explosions and related natal kicks. CE evolution plays a vital role in the orbital evolution [@w84]. During SN explosions the natal kicks imparted to newborn NS can cause the orbit to be eccentric or even disrupt the binary. It has been shown that most of the massive binaries are disrupted if the first NS receives a sufficiently large kick, and very few systems can survive to become HMXBs [@bz09; @sl14]. From the orbital parameter distribution of the observed HMXBs, @p02 suggested that the NSs in wide and nearly circular binaries likely received small natal kicks, implying that they originated from electron-capture supernovae [ECS, @p04; @v04] rather than core-collapse supernovae (CCS). After the second SN explosion, the survived DNSs are generally in eccentric orbits. Based on the distributions of the DNS masses and their orbital parameters, @bp16 proposed that the second collapse in the majority of the observed DNS systems involves small mass ejection and a low kick velocity [see also @t17].
Since the recent discoveries of gravitational wave transients by the advanced LIGO detector [e.g. @a16; @a17], a study of gravitational wave astronomy enters a golden age. Based on massive binary evolution, the rates of the merger events in the local universe can be theoretically estimated by a binary population synthesis method. However, the input parameters and physical assumptions should be carefully verified before the method is used to predict the merger rates in the local universe. Population synthesis studies [see e.g., @py98; @a15; @db15; @b18; @c18; @k18; @v18; @gm18a; @gm18b; @mg18] can be used to evaluate these uncertainties by adopting different input parameters to match the calculated distributions of the DNS population with those in our Galaxy. In this paper, we model the population of the Galactic DNSs. Compared to previous works, we have updated some important treatments of binary evolution, paying particular attention to the stability of mass transfer in primordial binaries, CE evolution, SN explosions and natal kicks. Our results may provide helpful constraints on the input physics of massive binary evolution. The remainder of this paper is organized as follows. In Section 2, we introduce the binary population synthesis method and the adopted assumptions. We present our calculated results and discussions in Section 3. We make a comparison with previous works in Section 4 and give a summary in Section 5.
Binary population synthesis method
==================================
To explore the distribution of the Galactic DNS systems, we use the population synthesis code *BSE* originally developed by @h02. With *BSE* we can simulate the evolution of a large of binary stars with different initial parameters. The modelling of binary evolution involves detailed treatments of stellar evolution, tidal interactions, mass and angular momentum loss, and mass and angular momentum transfer. In the code some modifications have been made by @sl14. Several important physical inputs are described as follows.
Mass transfer in the primordial binaries
----------------------------------------
Mass transfer begins in a primordial binary when the primary star evolves to fill its Roche lobe. The mass transfer efficiency (i.e. the fraction of the accreted matter by the secondary among the transferred matter) is a vital factor to determine the fate of the binary system. Rapid mass accretion can drive the secondary star to get out of thermal equilibrium and the responding expansion may also cause the secondary star to fill its own Roche lobe, leading to the formation of a contact binary [@ne01]. The contact binary will evolve to merge if the primary is a main sequence star, otherwise the binary system will enter a CE phase [see e.g., @d13; @sl14].
@p81 showed that only a small amount of accreted mass can accelerate a star to reach its critical rotation. When the secondary star is a rapid rotator, the mass transfer efficiency is subject to large uncertainty [@l98; @p05; @se09]. When modelling the evolution of the primordial binaries, @sl14 built three mass transfer modes. In Mode I, the mass accretion rate onto the secondary star is assumed to be dependent on its rotating rate, equal to the mass transfer rate multiplied by a factor of $ 1-\Omega/\Omega_{\rm cr}$ (where $ \Omega $ and $ \Omega_{\rm cr} $ are the angular velocity of the secondary and the corresponding critical value, respectively); in Mode II, the mass accretion efficiency is simply assumed to be 0.5; in Mode III, the mass accretion rate is assumed to be the mass transfer rate multiplied by a factor of $ \min (10\frac{\tau_{\dot{M}}}{\tau_{KH2}},1)$, where $\tau_{\dot{M}}$ is the mass transfer timescale and $\tau_{KH2}$ the secondary$'$s thermal timescale [@h02]. In the case of rapid mass accretion, the secondary star will get out of thermal equilibrium, expand and become overluminous. The thermal timescale of the secondary is greatly decreased, so the mass transfer process in Mode III is generally quasi-conservative, which is disfavored in massive binary evolution [@sl14] and will not be considered in this work. In comparison the mass transfer efficiency in Mode I can be $ \lesssim 0.2 $ for Case B binaries [@sl16]. The maximal initial mass ratios for avoiding the contact phase can reach as high as $ \sim 6 $ in Mode I, and drops to $ \sim 2 $ in Mode II and III [@sl14].
In our calculation we adopt Mode I to treat the mass transfer in the primordial binaries, unless specified otherwise. The reasons are briefly described as follows. When modelling the evolution of massive binaries, @sl16 demonstrated that the Galactic Wolf-Rayet/O binaries can be produced only if the mass transfer is highly non-conservative. It is also noted @d07 discussed the mass transfer efficiency in massive close binaries and found a tendency that systems with long initial orbital periods evolve in a more non-conservative way than narrow systems, which might be caused by the coupling of spin-up and tidal interaction during the mass transfer phase.
Highly non-conservative mass transfer in Mode I can prevent the binary entering a CE phase and result in the formation of a long period system [see Figure 6 of @sl14]. MWC 656 is the first confirmed Be/black hole binary with an orbital period of $ \sim 60 $ days [@c14]. Assuming that the rapid rotating Be star has been accelerated by stable mass transfer during the primordial binary evolution, @sl14 showed that non-conservative mass transfer is more likely to lead to the formation of MWC 656-like binaries. Recently @k18 suggested that the binary AS 386 may be a candidate Be/black hole system, which contains a Be star of mass $ 6-8 M_{\odot} $ in a $ \sim 131 $ day orbit. Such a long period system is hardly produced if the progenitor binary has experienced CE evolution [@p03; @g15].
By comparing the fraction of runaway Be stars among the whole Be star population, @be18 concluded that Be stars could be explained by an origin of mass transfer in binaries, that is, most of the runway Be stars are the products of disrupted Be/NS binaries. For the identified runway Be stars [@bg01], the stellar spectral types can be late to be B8 (the corresponding stellar mass $ \sim 3M_{\odot} $). This feature can be well produced by inefficient mass accretion which has spun up the secondary star without significant mass growth.
Now that we assume that the mass transfer is highly inefficient, for the matter that is not accreted by the secondary star, we let it escape the binary system in the form of isotropic wind, taking away the specific angular momentum of the secondary star. In addition, we follow @h02 to treat the process of rejuvenation, during which the secondary star appears even younger when fresh hydrogen is mixed into its growing convective core due to mass accretion.
Supernova explosion mechanisms and natal kicks
----------------------------------------------
A NS can be formed through both ECS and CCS. In the *BSE* code, the helium core mass at the base of the asymptotic giant branch is used to set the limits for identifying various CO cores [@h02]. If the helium core mass is lower than $ 1.83M_{\odot} $, the star will form a degenerate CO core and finally leave a CO white dwarf. Above this mass the star develops a partially degenerate CO core. If the core reaches a critical mass of $ 1.08M_{\odot} $, it will non-explosively burn into a degenerate ONe core. Stars with an ONe core more massive than $ 1.38M_{\odot} $ are believed to collapse into NSs through ECS, otherwise form an ONe white dwarf. When the helium core is massive enough to form a non-degenerate CO core, stable nuclear burning will continue until a CCS occurs. However, investigations show that the trigger of ECS explosions is subject to big uncertainties, and mass transfer in binary systems makes it more complicated [see e.g., @n84; @p04; @wh15; @p17]. In the calculations, we consider several possible helium core mass ranges $ \Delta M_{\rm ecs} $ ($ 1.83-2.25 M_{\odot} $, $ 1.83-2.5 M_{\odot} $, $ 1.83-2.75 M_{\odot} $), within which the star may eventually explode in an ECS explosion. The upper masses of these ranges can directly distinguish whether a star ends its life in an ECS or a CCS explosion. The larger the mass range, the more NSs formed through ECS.
We adopt the rapid SN model of @f12 to deal with the mass of NSs formed from CCS, which varies in the range of $ \sim 1.1 - 2M_{\odot} $. The gravitational mass of ECS NSs is set to be $ 1.3M_{\odot} $ considering the fact that some baryonic mass of the $1.38 M_{\odot} $ ONe core is converted into the binding energy of the NS [@t96; @f12]. A newborn NS is assumed to be imparted by a natal kick, which may lead to the disruption of the binary system. Based on the observations of the pulsars’ proper motions, @h05 demonstrated that the pulsar birth velocities can be described by a Maxwellian distribution with a dispersion velocity of $\sigma = 265 \rm\, km\, s^{-1} $. More recently, @vic17 argued that the birth velocities of pulsars can be better fitted by a bimodal distribution with $\sigma \simeq 80 $ and $ 320 \rm\, km\, s^{-1} $, which may respectively be related to ECS and CCS. Analysing the nature of observed DNSs reveals that the second formed NSs in most systems are likely to receive low kicks [@bp16; @t17]. In our calculations, the NS kick velocities are assumed to obey Maxwellian distributions. We adopt $\sigma_{\rm k,ecs} = 20 $, 40 and $80 \rm\, km\, s^{-1} $ for ECS NSs and $\sigma_{\rm k,ccs} = 150 $ and $300 \rm\, km\, s^{-1} $ for CCS NSs to examine their possible influence on the DNS formation.
Mass transfer in NS binaries
----------------------------
After the birth of the first NS, the binary evolves into the HMXB phase. Modelling the evolution of NS HMXBs reveals that the maximal mass ratio for stable mass transfer is $ \sim 3.5 $ [@t00; @prp02; @sl12]. If the donor is a helium star, we assume that the binary enters CE evolution only when the orbital period is less than 0.06 day [@t15]. In the case of unstable mass transfer that leads to CE evolution, we calculate the orbital energy of the binary system to examine whether it is sufficient to expel the stellar envelope. We use the results of @xl10 and @w16 for the binding energy parameter $ \lambda $ of the envelope, and take the CE efficiency $ \alpha_{\rm CE} $ to be 1.0 in the calculations.
For stellar wind mass loss, we employ the fitting formula of @nd90, except for hot OB stars, for which we use the simulated rates of @v01. For the stripped helium stars and Wolf-Rayet stars, we adopt the rates of @v17[^1]. During the CE phase, we ignore any mass accretion because of its very short duration. In NS binaries with a helium donor star, mass transfer proceeds rapidly for a short ($ < 10^{5} \,\rm yr$) time through Roche lobe overflow [@t15], the accretion rate of the NS is assumed to be limited by its Eddington rate.
Results and discussions
=======================
We have performed a series of Monte Carlo simulations for a large number of binary systems with different initial parameters. For each of the considered models, we simulate the evolution of $ 10^{8} $ binary systems. We follow @h02 to set the initial parameters of the primordial binaries. The primary masses obey the initial mass function given by @k93, and the mass ratios of the secondary to the primary satisfy a flat distribution between 0 and 1. The distribution of the orbital separations is assumed to be flat in the logarithm in the range of $ 3-10^{4} R_{\odot}$. All binaries are assumed to have circular orbits, since the outcome of the interactions of binary systems with the same semilatus rectum is almost independent of eccentricity [@h02]. Several authors [@s78; @d06; @rw10] have proposed that the star formation rate in our Galaxy is $ \sim1-5 M_{\odot}\,\rm yr^{-1} $. Here we adopt a moderate rate of $ 3 M_{\odot}\,\rm yr^{-1} $ over a period of 10 Gyr. When investigating the star formation history of the Galaxy, @r00 showed that the star formation rate is not likely to change by much more than a factor of two over the past 10 Gyr. So adopting a constant star formation rate should be an acceptable approximation.
From each of the population synthesis calculations, we obtain a few to several ten thousand initial DNS systems. The corresponding birthrates can be calculated according to the parameters of their primordial binaries [@h02]. After the DNS formation, we adopt the formulae of @p64 to trace the subsequent orbital evolution with time due to gravitational wave radiation. The DNS merger rate within the lifetime of the Galaxy is mainly dependent on the initial orbital parameter of the DNS binary[^2]. In the calculations, we assume that the DNS binaries evolve to merge when the orbital periods shrink to be less than 0.001 day. Since we adopt a constant star formation rate of $ 3 M_{\odot}\,\rm yr^{-1} $, the birthrate of the Galactic DNS systems does not vary with time, while the merger rate tends to gradually increase over the past 10 Gyr. The total number of the DNS binaries in our Galaxy can be obtained by accumulating the product of the birthrate and the evolutionary time for every binary. In Table 2, we show the merger rate $ R_{\rm merger} $, the birthrate $ R_{\rm birth} $ and the total number of the Galactic DNSs at present, and the fraction $ f $ of DNS systems with the second NS being formed through ECS among all DNSs. Obviously the larger values of both $ \sigma_{\rm k,ecs} $ and $ \sigma_{\rm k,ccs} $ result in the decreases of $ R_{\rm merger} $ and $ R_{\rm birth} $. We obtain $ R_{\rm merger} \sim 2-10 \,\rm Myr^{-1} $, $ R_{\rm birth} \sim 3-27 \,\rm Myr^{-1} $, and the total number of DNSs in our Galaxy is of the order $ 10^{4} $. It is also seen that wider $ \Delta M_{\rm ecs} $ leads to higher $ f $, which varies in the range of $ 0.06-0.85 $.
The $ P_{\rm orb}-e $ diagram
-----------------------------
Since the orbital evolution of DNS systems is solely driven by gravitational wave radiation, we can synthesize the orbital parameters of the simulated DNS population into a diagram that represents the snapshot after 10 Gyr star formation and evolution. We show in Figure 1 the calculated distribution of Galactic DNS systems (with the numbers given in Table 2) in the $ P_{\rm orb}-e $ plane when adopting $ \Delta M_{\rm ecs} = (1.83-2.25 M_{\odot}) $. The left, middle and right panels correspond to $\sigma_{\rm k,ecs} =$ 20, 40 and $80 \rm\, km\, s^{-1} $ for ECS NSs, and the top and bottom panels correspond to $\sigma_{\rm k,ccs} =$ 150 and $300 \rm\, km\, s^{-1} $ for CCS NSs, respectively. Each panel includes $ 20\times 20 $ image elements, in which $ P_{\rm orb} $ varies logarithmically from 0.01 to $ 10^{4} $ days with steps of 0.3, and $ e $ linearly from 0 to 1 with steps of 0.05. Until now there exist 14 observed DNSs whose orbital periods and eccentricities are precisely measured (see Table 1). They are plotted as blue triangles in the figure. Figure 1 shows that the calculated DNSs are mainly distributed in a region with relatively large eccentricities ($ e >0.4 $), so the observed systems with long orbital periods ($ \gtrsim 1 $ day) and low eccentricities ($ \lesssim 0.4 $) can be hardly reproduced. The reason is that with $\Delta M_{\rm ecs} = (1.83-2.25 M_{\odot}) $ the second NSs in most systems are born in a CCS explosion with a high natal kick, which tends to increase the orbital eccentricities and even disrupt the original systems (especially for wide binaries).
Figures 2 and 3 show the similar distributions of Galactic DNSs in the $ P_{\rm orb}-e $ plane with $ \Delta M_{\rm ecs} = (1.83-2.5 M_{\odot} )$ and ($ 1.83-2.75 M_{\odot} $), respectively. In these two cases the value of $ f $ is remarkably increased and can reach as high as 0.85 when $\sigma_{\rm k,ecs} = 20 \rm\, km\, s^{-1} $ and $\sigma_{\rm k,ccs} = 300 \rm\, km\, s^{-1} $ (see Table 2). The second born NSs in most systems are imparted to small natal kicks, and the long-period binaries can effectively avoid disruption. In addition, the predicted DNS binaries tend to have low eccentricities. Generally the observed DNS binaries can be well covered by a high probability region in the $ P_{\rm orb}-e $ diagrams when taking $ \Delta M_{\rm ecs} = (1.83-2.75 M_{\odot} )$, expect the case with $\sigma_{\rm k,ecs} = 80 \rm\, km\, s^{-1} $ (see right panels). Consistent with previous works [e.g. @bp16; @t17], our results suggest that the formation of the second NS be accompanied with a low natal kick.
We then try to quantitatively compare the calculated results with observations to constrain the adopted models with different input parameters. Although the observed Galactic DNS sample is limited by small size and the detection of such systems may be subject to several selection effects [see Section 2.1 of @t17], we employ the Bayesian analysis used by @a15 to compare the goodness of match for every model, which involves two independent orbital parameters of $ P_{\rm orb} $ and $ e $. The Bayes’ theorem gives a probability expression of $$P(M\mid D) = \frac{P(D\mid M)P(M)}{P(D)},$$ where $ P(M\mid D) $ is the posterior probability, i.e. the probability related to our model $ M$ given the observed data $ D$ (i.e., the data set of $ P_{\rm orb} $ and $ e $), $ P(D\mid M) $ is the likelihood of the observed data given that the model is true, which is denoted as $ \Lambda(D) $, $ P(M) $ is the prior probability of our model, and $ P(D) $ a normalizing constant that is independent of the model. As all models have the same prior probability, it can be absorbed into the normalizing constant as $ C = P(M)/P(D)$. So $P(M\mid D)$ can be rewritten as $$P(M\mid D) = C \Lambda(D) .$$ Since $ C $ is independent of the models, the value of $ \Lambda(D) $ gives the relative probability of each model. Because all of the observed DNSs in the sample are independent, the probability of the observed data given the model is equal to the product of the probabilities of each binary system: $$\Lambda(P_{\rm orb},e) = \prod_{i} P(P_{{\rm orb}, i},e_{i}\mid M).$$ Note that the errors of the observed orbital period and eccentricity for each DNS system are very small and can be neglected. Then $ P(P_{{\rm orb}, i},e_{i}\mid M) $ reflects the probability density of a model at a specific point in the $ P_{\rm orb}-e $ parameter space, which corresponds to a specific DNS. Adopting similar treatment of @a15, we calculate $ P(P_{{\rm orb}, i},e_{i}\mid M) $ by binning the simulated systems into two dimensions with a bin size of $ \Delta \log(P_{\rm orb}) = 0.3 $ in orbital period and $ \Delta e = 0.05 $ in eccentricity. Finally the derived value of $ \Lambda(P_{\rm orb},e) $ is used to rank our models (see Table 2). The model that best matches the observations involves the input parameters of $\sigma_{\rm k,ecs} = 40 \rm\, km\, s^{-1} $, $\sigma_{\rm k,ccs} = 300
\rm\, km\, s^{-1} $ and $ \Delta M_{\rm ecs} = (1.83-2.75 M_{\odot}) $.
To display the effect of the mass transfer process during the primordial binary evolution on the DNS population, in Fig. 4 we present the corresponding distributions of DNSs in the $ P_{\rm orb}-e $ plane when the mass transfer efficiency is taken to be a constant value of 0.5 (i.e., Mode II). The left, middle and right panels correspond to $ \Delta M_{\rm ecs} = (1.83-2.25 M_{\odot}) $, $( 1.83-2.5 M_{\odot} )$ and $ (1.83-2.75 M_{\odot}) $, respectively. The kick velocity parameters are all set to be $\sigma_{\rm k,ecs} = 40 \rm\, km\, s^{-1} $ and $\sigma_{\rm k,ccs} = 300 \rm\, km\, s^{-1} $. Compared with the situations in Mode I, the predicted distribution of the DNS systems shifts to a region with shorter orbital periods and seems not to match observations. The main reason is that the birthrate of close DNSs in Mode II is greatly enhanced since a large number of original NS/white dwarf binaries (post-CE products of HMXBs) in Mode I are transformed into DNS systems due to the higher mass transfer efficiency. Further, more efficient mass transfer can significantly increase the secondary mass, so it takes more kinetic energy of the NS to expel the secondary’s envelope during the CE evolution.
The mass distribution of DNS systems
------------------------------------
All of the observed Galactic DNSs except PSR J0453+1559 have similar masses of $ \sim 1.2-1.4 M_{\odot} $ (see Table 1). As expected, recycled pulsars are generally more massive than their non-recycled companions because a certain amount of mass is required to recycle a pulsar. In Figure 5 we present the distributions of the pulsar mass $ M_{\rm p} $ (first formed NS, left), the companion mass $ M_{\rm c} $ (second formed NS, middle) and the total mass $ M_{\rm t} $ (right) for the Galactic DNS systems under different assumed models. The panels form top to bottom correspond to $\Delta M_{\rm ecs}
= (1.83-2.25 M_{\odot})$, $ (1.83-2.5 M_{\odot})$ and $(1.83-2.75 M_{\odot} )$, respectively. In each panel, the black, red, and blue curves correspond to $\sigma_{\rm k,ecs} = $ 20, 40, and $80 \rm\, km\, s^{-1} $, and the dashed and solid curves correspond to $\sigma_{\rm k,ccs} = $ 150 and $300 \rm\, km\, s^{-1} $ respectively. The green curves represent the observed data multiplied by a factor of $ 10^{4} $. We find that the calculated mass distributions of both DNS components generally have two peaks of $ \sim 1.1 M_{\odot} $ [minimum NS mass in the rapid SN model of @f12] and $ \sim 1.3 M_{\odot} $ (the mass of NSs formed from ECS), in agreement with the results of @gm18b and @mg18.
The calculated mass distribution at the peaks of $ \sim 1.1 M_{\odot} $ and $ \sim 1.3 M_{\odot} $ could be used to distinguish the NSs formed from CCS and ECS, respectively. It is unsurprising that the percentage at each peak changes with $ \Delta M_{\rm ecs} $, the former decreases and the latter increases with increasing $ \Delta M_{\rm ecs} $ since more ECS NSs are produced. Even in the case of $\Delta M_{\rm ecs} = (1.83-2.25 M_{\odot})$, quite a fraction of DNS binaries undergo at least one ECS explosion. This is consistent with the result of @gm18a for merging DNSs, indicating that the ECS is a fundamental process for the DNS formation. We can see that the overall distributions of calculated DNS masses, although in agreement with other studies, are actually not compatible with observations. The reason may be that the adopted SN mechanisms are still too simplistic, and there may not be one to one relation between the compact star mass and the core mass of their progenitors. Furthermore a NS can increase its mass due to the possible accretion during the recycling phase.
In spite of the mass distribution for DNS binaries, comparison between the calculated and measured orbital parameters gives the best choice of the input parameters of $\sigma_{\rm k,ecs} = 40 \rm\, km\, s^{-1} $, $\sigma_{\rm k,ccs} = 300
\rm\, km\, s^{-1} $ and $ \Delta M_{\rm ecs} = (1.83-2.75 M_{\odot} )$ (see the rank of different models in Table 2), which are therefore set to be default input parameters in the following.
Formation channels of DNS systems
---------------------------------
The formation channels of DNS systems have been explored by many authors. Besides the canonical channel in which the binary experiences a HMXB stage [@bv91; @tv06], there is an alternative channel which involves a double helium star phase. In the double core CE scenario [@b95] a DNS system is the descendant of a double helium star binary, which is the result of a CE phase in a binary with both components of very similar masses. Both components of the binary have developed a compact core at the beginning of mass transfer. Such a binary will enter the contact phase that is followed by the CE evolution. The post-CE product is a double helium star binary in a relative close orbit. After experiencing two SN explosions, the survived binary is a DNS system [see also @dps06].
By analysing the calculated results, we find that the double helium star binary can be produced in an another way. Here both of the binary components also initially have almost equal mass. The secondary star is still in the main sequence at the onset of mass transfer. The mass transfer proceeds stably between the binary components, after which the primary leaves a helium star and the secondary approaches the end of main sequence. The secondary then evolves to be a supergiant star and transfers mass to the helium star, resulting in CE evolution. The outcome of CE evolution is a double helium star binary. A similar scenario was suggested by @py98, in which the mass exchange during the primordial binary evolution was assumed to be almost conservative, but the process of rejuvenation was not considered in their calculations. In the case of conservative mass transfer, this can cause the secondary to be younger, and stay in the main sequence when the primary explodes to be a NS. A highly non-conservative mass transfer like in Mode I can effectively reduce the rejuvenation of the secondary. The primary helium star firstly explodes to be a NS, which is subsequently recycled by the mass transfer from its helium star companion. When the secondary evolves to be a NS, the binary finally turns into a DNS system.
Adopting the default model with the input parameters of $\sigma_{\rm k,ecs} = 40 \rm\, km\, s^{-1} $, $\sigma_{\rm k,ccs} = 300
\rm\, km\, s^{-1} $ and $ \Delta M_{\rm ecs} = (1.83-2.75 M_{\odot}) $, we obtain the total birthrate of Galactic DNSs to be $ \sim 10\,\rm Myr^{-1} $ (see Table 2). The DNS binaries formed from the canonical channel have a birthrate of $ \sim 4\,\rm Myr^{-1} $, and the two scenarios involving a double helium star binary contribute a similar birthrate of $ \sim 3\,\rm Myr^{-1} $. This means that the double helium star channel is likely to dominate the formation of Galactic DNS systems.
Figure 6 shows the calculated distribution of DNS systems in the $ P_{\rm orb}-e $ plane when taking into account different formation channels. The left and right panels correspond to the results from the canonical channel and the double helium star channel, respectively. In the canonical channel, the DNS systems are dominated by binaries with short orbital periods of $ \lesssim 1 $ day and low eccentricities of $ \lesssim 0.4 $, their distributions show similar features to that in Mode II (see the right panel of Figure 4). The double helium star channel tends to produce the DNS systems with long periods of $ \gtrsim 1 $ day and high eccentricities $ \gtrsim 0.1 $. The difference originates from the fact that the binary orbital energy that is used to expel the CE in the former is considerably smaller than in the latter. Here the fraction of systems that formed from the canonical channel among all DNSs is $ \sim 0.3 $.
In Figure 7 we present the calculated distributions of the DNS systems in the $ P_{\rm orb}-e $ plane, in which the second born NS originates from ECS (left) or CCS (right) explosion. The NSs formed through CCS have received relatively large natal kicks, which tend to produce the DNS systems with large eccentricities. On the contrary, the DNS systems generally have low eccentricities for ECS explosions. A large number of relatively wide systems can survive the SN explosions, and evolve to DNS systems with long orbital periods and low eccentricities. The fraction of the DNS systems that experiences a second ECS explosion can be as high as 0.74 (see Table 2). We note that the PSRs J1757-1854 and B1913+16 have relative short periods and large eccentricities, and their orbital decay timescales are of the order 100 Myr. These two binaries are not easily accounted for by the current scenarios.
Comparison with other works
===========================
In this section, we compare our results with recent population synthesis works of @a15, @k18 and @v18.
The most important parameters in @a15 include the combined CE parameter $ \alpha_{\rm CE} \lambda$, the NS kick velocity distribution and the ECS mass range (same as our $\Delta M_{\rm ecs}$). A significant difference between their work and ours is the treatment of the $ \alpha_{\rm CE} \lambda$, which was set to be constant by @a15. Note also that the double helium star channel is not considered by @a15, who focus on the canonical channel. To match the orbital parameters of the known eight DNSs, their models are required to produce wider binaries than our canonical channel. This indicates that their adopted $ \alpha_{\rm CE} \lambda$ is generally larger than ours. By testing different models to constrain the input parameters, they found that $ \alpha_{\rm CE} \lambda \lesssim 0.25$ is effectively ruled out, the allowed range is $ 0.3-1.0 $ and the best choice is 0.5. Considering the fact that the value of $ \lambda $ for massive supergiants are less than 0.1 [@dt00; @p03; @w16], this implies that the value of $ \alpha_{\rm CE} $ in @a15 is significantly larger than unity. After the discovery of PSR J1930-1852 with the longest known orbital period of 45 days, @s15 argued that many models of @a15 cannot produce such wide DNS systems, probably requiring reconsideration of the binary evolutionary processes.
The investigation of @k18 involves the amplitudes of the NS natal kicks and the treatment of mass and angular momentum transfer. They proposed that a low kick velocity is applied when the envelope of the progenitor star is stripped. This can help produce the observed DNS systems with low eccentricities. Following the method of @s97 to model the binary orbital evolution with Roche lobe overflow, @k18 assumed that the fractions of the mass escaped from the donor and ejected from the accretor are respectively 0.2 and $ \geqslant 0.75 $, and the rest material (with the fraction of $ \leqslant 0.05 $) is accreted by the accretor. In their default model the mass accretion efficiency is also very low, but the channel involving a double helium star binary was not included. For their modelled DNS systems, the probability distribution of the orbital periods is peaked at $ \sim 1-3 $ days, slightly longer than that ($ \lesssim 1 $ day) predicted in our canonical channel. The reason is perhaps that a large fraction of the lost matter from the donor through a fast wind tend to increase the orbital separations of the progenitor systems (this can lead to more expansion of the donor and decrease the binding energy of the envelope during CE evolution, broadening the subsequent DNS systems). From their orbital parameter distribution (Figure 11), there is a shortage of wide systems with orbital period larger than several days with respect to the observed DNSs. Meanwhile close systems with orbital period less than $ \sim 1$ day are also deficient.
@v18 suggested that a bimodal natal kick distribution (with $ \sigma_{\rm k} = 30 \rm\,km\,s^{-1} $ and $ 265 \rm\,km\,s^{-1} $) is preferred over a high-kick component alone. This is in accord with our assumption about the NS kick velocity distribution. They adopted the same treatments of the NS mass and kick velocities for both ultra-stripped SNe [see @t15] and ECSs, which can boost the formation of DNS binaries with low eccentricities. In their fiducial model, the canonical channel is responsible for formation of about 70% of all DNSs and the channel with a double core CE phase approximately contributes about 21%. It can be seen that their double helium star channel contributes a significant fraction of DNS population, but still much smaller than that in our results. Since the probability distribution of the orbital parameters for DNS systems is not clearly shown by @v18, we cannot make further comparisons.
Summary
=======
We have performed a population synthesis study of the formation of DNS systems in the Galaxy. Based on previous investigations on Wolf-Rayet/O binaries, Be/black hole systems and runaway Be stars, we argue that mass transfer in the primordial binaries is likely to be highly inefficient. So we adopt the rotation-dependent mass transfer mode (Mode I) in our calculations. We then investigate the formation channels of the DNSs and the effect of the SN kicks at the NS formation on the distribution of DNS systems. By comparing with observed DNS binaries, we can set important constraints on the model parameters. We summarise our main results as follows.
1\. The second born NSs in most DNS systems have low kick velocities, with the dispersion velocity of less than $ 80 \,\rm km\,s^{-1} $ in a Maxwellian distribution. This also requires that the helium core mass range for ECS is $ 1.83-2.75 M_{\odot} $ if low kicks for ultra-stripped SNe are not taken into account.
2\. The DNS population can be divided into two groups by the orbital period of $ \sim 1 $ day, the systems with short periods mainly form through the canonical channel involving a HMXB phase and the wide binaries experience a double helium star binary stage during the progenitor evolution.
3\. Our best model shows that the birthrate and merger rate of Galactic DNS systems in the canonical channel are $ \sim 4\,\rm Myr^{-1}$ and $ \sim 2\,\rm Myr^{-1}$, and the corresponding rates in the double helium star channel are $ \sim 6\,\rm Myr^{-1}$ and $ \sim 2\,\rm Myr^{-1}$, respectively.
At last we briefly discuss the applicability of the $ \Delta M_{\rm ecs} $ that is ascribed to the process of ECS explosions. Although @a15 extended the upper mass limit to be $ 3.24M_{\odot} $ in their study on DNS formation, our best range of $ \Delta M_{\rm ecs} = (1.83-2.75 M_{\odot}) $ seems to be still wider than previously expected. Investigations indicate that, besides ECS, both ultra-stripped SNe and low-mass Fe-core SNe are expected to have considerably lower NS kicks than CCS of massive stellar cores [@t15; @t17; @j17; @gj18]. This implies that the wide core mass range may include the contributions of ultra-stripped SNe and low-mass Fe-core SNe.
We thank the referee for constructive comments that helped improve this paper. This work was supported by the Natural Science Foundation of China (Nos. 11603010, 11133001, 11333004, 11773015, and 11563003), the Project U1838201 supported by NSFC and CAS, the National Program on Key Research and Development Project (Grant No. 2016YFA0400803), and the Natural Science Foundation for the Youth of Jiangsu Province (No. BK20160611).
[28]{} natexlab\#1[\#1]{}
Abbott B. P. et al., 2016, Phys. Rev. Lett., 116, 061102
Abbott, B. P. et al. 2017, Phys. Rev. Lett., 119, 161101
Andrews, J. J., Farr, W. M., Kalogera, V., & Willems, B. 2015, ApJ, 801, 32
Belczynski, K., & Ziolkowski, J. 2009, ApJ, 707, 870
Belczynski, K., Askar, A., Arca-Sedda, M. et al. 2018, A&A, 615, 91
Beniamini, P., & Piran, T. 2016, MNRAS, 456, 4089
Berger D. H., & Gies D. R., 2001, ApJ, 555, 364
Bhattacharya, D., & van den Heuvel, E. P. J. 1991, PhRv, 203, 1
Boubert, D., & Evans, N. W. 2018, arXiv: 1804.05849
Brown, G. E., 1995, ApJ, 440, 270
Cameron, A. D., Champion, D. J., Kramer, M., et al. 2018, MNRAS, 475, L57
Champion, D. J., Lorimer, D. R., McLaughlin, M. A., et al. 2004, MNRAS, 350, L61
Casares, J., Negueruela, I., Ribó, M., et al. 2014, Nature, 505, 378
Chruslinska, M., Belczynski, K., Klencki, J., & Benacquista, M. 2018, MNRAS, 474, 2937
Corongiu, A., Kramer, M., Stappers, B. W., et al. 2007, A&A, 462, 703
de Mink, S. E., Pols, O. R., & Hilditch, R. W. 2007, A&A, 467, 1181
de Mink, S. E., Langer, N., Izzard, R. G., Sana, H., & de Koter, A. 2013, ApJ, 764, 166
de Mink, S. E., & Belczynski, K., 2015, ApJ, 814, 58
Dewi, J., & Tauris, T. M. 2000, A&A, 360, 1043
Dewi, J. D. M., & Pols, O. R. 2003, MNRAS, 344, 629
Dewi, J. D. M., Podsiadlowski, P., Sena, A. 2006, MNRAS, 368, 1742
Diehl, R. et al., 2006, Nature, 439, 45
Ferdman, R. D., Stairs, I. H., Kramer, M., et al. 2014, MNRAS, 443, 2183
Flannery, B. P., & van den Heuvel, E. P. J. 1975, A&A, 39, 61
Fonseca, E., Stairs, I. H., & Thorsett, S. E. 2014, ApJ, 787, 82
Fryer, C., Belczynski, K., Wiktorowicz, G., et al. 2012, ApJ, 749, 91
Gessner, A. & Janka, H.-T. 2018, arXiv: 1802.05274
Giacobbo, N., & Mapelli, M. 2018a, arXiv: 1805.11100
Giacobbo, N., & Mapelli, M. 2018b, MNRAS, 480, 2011
Grudzinska, M., Belczynski, K., Casares, J., et al. 2015, MNRAS, 452, 2773
Hobbs, G., Lorimer, D. R., Lyne, A. G., & Kramer, M. 2005, MNRAS, 360, 974
Hulse, R. A., & Taylor, J. H., 1975, ApJ, 195, L51
Hurley, J. R., Tout, C. A., & Pols, O. R. 2002, MNRAS, 329, 897
Ivanova, N., Belczynski, K., Kalogera, V., Rasio, F. A., & Taam, R. E. 2003, ApJ, 592, 475
Ivanova, N., Justham, S., Chen, X., et al. 2013, A&AR, 21, 59
Jacoby, B. A., Cameron, P. B., Jenet, F. A., et al. 2006, ApJ, 644, L113
Janka, H.-T., 2017, ApJ, 837, 84
Janssen, G. H., Stappers, B. W., Kramer, M., et al. 2008, A&A, 490, 753
Keith, M. J., Kramer, M., Lyne, A. G., et al. 2009, MNRAS, 393, 623
Khokhlov, S. A., Miroshnichenko, A. S., Zharikov, S. V., et al. 2018, ApJ, 856, 158
Kramer, M., Stairs, I. H., Manchester, R. N., et al. 2006, Sci, 314, 97
Kroupa, P., Tout, C. A., & Gilmore, G. 1993, MNRAS, 262, 545
Kruckow, M. U., Tauris, T. M., Langer, N., Kramer, M., & Izzard, R. G. 2018, arXiv:1801.05433
Langer, N. 1998, A&A, 329, 551
Lazarus, P., Freire, P. C. C., Allen, B., et al. 2016, ApJ, 831, 150
Lipunov, V.M., Postnov, K.A., & Prokhorov, M.E., 1997, MNRAS, 288, 245
Lynch, R. S., Freire, P. C. C., Ransom, S. M., & Jacoby, B. A. 2012, ApJ, 745, 109
Martinez, J. G., Stovall, K., Freire, P. C. C., et al. 2015, ApJ, 812, 143
Martinez, J. G., Stovall, K., Freire, P. C. C., et al. 2017, ApJ, 851, L29
Mapelli, M., & Giacobbo, N. 2018, MNRAS, 479, 4391
Nelson, C. A. & Eggleton, P. P. 2001, ApJ, 552, 664
Ng, C., Kruckow, M. U., Tauris, T. M., et al 2018, MNRAS, 476, 4315
Nieuwenhuijzen, H., & de Jager, C. 1990, A&A, 231, 134
Nomoto, K. 1984, ApJ, 277, 791
Packet, W., 1981, A&A, 102, 17
Peters, P. C. 1964, PhRv, 136, 1224
Petrovic, J., Langer, N., & van der Hucht, K. A. 2005, A&A, 435, 1013
Pfahl, E., Rappaport, S., Podsiadlowski, P., & Spruit, H. 2002, ApJ, 574, 364
Poelarends, A. J. T., Wurtz, S., Tarka, J., Cole Adams, L., & Hills, S. T., 2017, ApJ, 850, 197
Podsiadlowski, Ph., Rappaport, S., & Pfahl, E. D. 2002, ApJ, 565, 1107
Podsiadlowski P., Rappaport S., & Han Z., 2003, MNRAS, 341, 385
Podsiadlowski, P., Langer, N., Poelarends, A. J. T., et al. 2004, ApJ, 612, 1044
Portegies Zwart, S. F., & Yungelson, L. R. 1998, A&A, 332, 173
Robitaille, T. P., & Whitney, B. A. 2010, ApJ, 710, L11
Rocha-Pinto, H. J., Scalo, J., Maciel, W. J., & Flynn, C. 2000, A&A, 358, 869
Shao, Y., & Li, X.-D. 2012, ApJ, 756, 85
Shao, Y., & Li, X.-D. 2014, ApJ, 796, 37
Shao, Y., & Li, X.-D. 2016, ApJ, 833, 108
Smith, L. F., Biermann, P., & Mezger, P. G. 1978, A&A, 66, 65
Soberman G. E., Phinney E. S., & van den Heuvel E. P. J., 1997, A&A, 327, 620
Stancliffe, R. & Eldridge, J. 2009, MNRAS, 396, 1699
Stovall, K., Freire, P. C. C., Chatterjee, S., et al., 2018, ApJ, 854, L22
Swiggum, J., Rosen, R., McLaughlin, M., et al. 2015, ApJ, 805, 156
Tauris, T. M., van den Heuvel, E. P. J., & Savonije, G. J. 2000, ApJ, 530, L93
Tauris, T. M., & van den Heuvel, E. P. J. 2006, in Compact Stellar X-Ray Sources, ed. W. H. G. Lewin & M. van der Klis (Cambridge: Cambridge Univ. Press), 623
Tauris, T., Langer, N., & Podsiadlowski, P. 2015, MNRAS, 451, 2123
Tauris., T. M., Kramer, M., Freire, P., et al., 2017, ApJ, 846, 170
Timmes, F. X., Woosley, S. E., & Weaver, T. A. 1996, ApJ, 457, 834
Tutukov, A. V., & Yungelson, L. R. 1993, ARep, 37, 411
van den Heuvel, E. P. J. 2004, in Proc. 5th INTEGRAL Workshop, ESA SP- 552, ed. V. Schoenfelder, G. Lichti, & C. Winkler. (Noordwijk: ESA Publ. Div. ESTEC), 185
van Leeuwen, J., Kasian, L., Stairs, I. H., et al. 2015, ApJ, 798, 118
Verbunt, F., Igoshev, A., & Cator, E. 2017, A&A, 608,57
Vigna-G[ó]{}mez, A., Neijssel, C., Stevenson, S. et al. 2018, arXiv:1805.07974
Vink, J. S., de Koter, A., & Lamers, H. J. G. L. M. 2001, A&A, 369, 574
Vink, J. S. 2017, A&A, 607, L8
Wang, C., Jia, K., & Li, X.-D. 2016, RAA, 16, 126
Webbink, R. F. 1984, ApJ, 277, 355
Weisberg, J. M., Nice, D. J., & Taylor, J. H. 2010, ApJ, 722, 1030
Woosley, S. E., & Heger, A. 2015, ApJ, 810, 34
Xu, X.-J., & Li, X.-D. 2010, ApJ, 716, 114
![Distributions of Galactic DNS binaries in the orbital period $ P_{\rm orb} $ vs. eccentricity $ e $ plane when taking $ \Delta M_{\rm ecs} = (1.83-2.25 M_{\odot} )$. Each diagram represents the snapshot of the current DNS population with a star formation rate of $ 3 M_{\odot}\,\rm yr^{-1} $ over 10 Gyr. The left, middle and right panels correspond to $\sigma_{\rm k,ecs} = $ 20, 40 and $80 \rm\, km\, s^{-1} $, and the top and bottom panels correspond to $\sigma_{\rm k,ccs} = $ 150 and $300 \rm\, km\, s^{-1} $, respectively. The blue triangles demonstrate the positions of the observed DNSs. \[figure1\]](f1.pdf)
![Same as Fig. 1 but with $ \Delta M_{\rm ecs} = (1.83-2.5 M_{\odot}) $. \[figure2\]](f2.pdf)
![Same as Fig. 1 but with $ \Delta M_{\rm ecs} = (1.83-2.75 M_{\odot}) $. \[figure3\]](f3.pdf)
![Distributions of Galactic DNS binaries in the $ P_{\rm orb} - e $ plane under Mode II. The left, middle and right panels correspond to $ \Delta M_{\rm ecs} = (1.83-2.25 M_{\odot}) $, $ (1.83-2.5 M_{\odot} )$ and $ (1.83-2.75 M_{\odot}) $, respectively. The blue triangles show the positions of the observed DNSs. \[figure4\]](f4.pdf)
![The calculated distributions of pulsar mass $ M_{\rm p} $ (first formed NS, left panels), companion mass $ M_{\rm c} $ (second formed NS, middle panels) and total mass $ M_{\rm t} $ (right panels) for Galactic DNS binaries. The panels from top to bottom correspond to $ \Delta M_{\rm ecs} = (1.83-2.25 M_{\odot} )$, $ (1.83-2.5 M_{\odot} )$ and $ (1.83-2.75 M_{\odot} )$, respectively. In each panel, the black, red and blue curves correspond to $\sigma_{\rm k,ecs} = $ 20, 40 and $80 \rm\, km\, s^{-1} $, and the dashed and solid curves correspond to $\sigma_{\rm k,ccs} = $ 150 and $300 \rm\, km\, s^{-1} $ respectively. The green curves denote the observed data multiplied by a factor of $ 10^{4} $. There are 7 (13) DNS binaries with component (total) masses precisely measured (see Table 1). \[figure6\]](f5.pdf)
![Distributions of Galactic DNS systems formed through the canonical channel (left) and the double helium star channel (right). Here we adopt the default model with the input parameters of $\sigma_{\rm k,ecs} = 40 \rm\, km\, s^{-1} $, $\sigma_{\rm k,ccs} = 300
\rm\, km\, s^{-1} $ and $ \Delta M_{\rm ecs} = (1.83-2.75 M_{\odot}) $. The observational data are plotted with the blue triangles. \[figure6\]](f6.pdf)
![Similar to Figure 6, but with the second born NSs originating from ECS (left) or CCS (right). \[figure6\]](f7.pdf)
------------------------ ---------------- ------- --------------- --------------- --------------- ------------------ ----------------------------- ------------------ -- -- -- -- -- -- -- -- --
Pulsar name $P_{\rm orb} $ $e$ $ M_{\rm p} $ $ M_{\rm c} $ $ M_{\rm t} $ $ P_{\rm spin} $ $ \dot{P}_{\rm spin} $ $ \tau_{\rm c} $
(days) ($M_{\odot}$) ($M_{\odot}$) ($M_{\odot}$) (ms) ($ 10^{-18}\rm s\,s^{-1} $) (Gyr)
J1946+2052 (1) 0.078 0.064 $< 1.31$ $> 1.18$ 2.50 17.0 0.92 0.29
J0737-3039A (2) 0.102 0.088 1.338 1.249 2.587 22.7 1.76 0.204
J1757-1854 (3) 0.184 0.606 1.338 1.395 2.733 21.5 2.63 0.13
J1913+1102 (4) 0.206 0.09 $ <1.84 $ $ >1.04 $ 2.88 27.3 0.161 2.7
J1756-2251 (5) 0.320 0.181 1.341 1.230 2.570 28.5 1.02 0.44
B1913+16 (6) 0.323 0.617 1.440 1.389 2.828 59.0 8.63 0.109
B1534+12 (7) 0.421 0.274 1.333 1.346 2.678 37.9 2.42 0.248
J1829+2456 (8) 1.176 0.139 $ <1.38 $ $ >1.22 $ 2.59 41.0 0.0525 12.4
J1411+2551 (9) 2.616 0.17 $ <1.62 $ $ >0.92 $ 2.54 62.5 0.096 10
J0453+1559 (10) 4.072 0.113 1.559 1.174 2.734 45.8 0.186 3.9
J1518+4904 (11) 8.634 0.249 1.41 1.31 2.72 40.9 0.0272 23.9
J1753-2240 (12) 13.638 0.304 $ - $ $ - $ $ - $ 95.1 0.97 1.55
J1811-1736 (13) 18.779 0.828 $ <1.64 $ $ >0.93 $ 2.57 104.2 0.901 1.83
J1930-1852 (14) 45.060 0.399 $ < 1.32 $ $ >1.30 $ 2.59 185.5 18.0 0.163
J1906+0746$^{a}$ (15) 0.166 0.085 1.291 1.322 2.613 144.1 20300 0.0001
B2127+11C$^{a}$ (16) 0.335 0.681 1.358 1.354 2.713 30.5 4.99 0.1
J1755-2550$^{a}$ (17) 9.696 0.089 $ - $ $ >0.40 $ $ - $ 315.2 2434 0.002
J1807-2500B$^{a}$ (18) 9.957 0.747 1.366 1.206 2.572 4.2 0.0823 0.81
------------------------ ---------------- ------- --------------- --------------- --------------- ------------------ ----------------------------- ------------------ -- -- -- -- -- -- -- -- --
: Basic parameters of DNS systems known in our Galaxy, including orbital period $ P_{\rm orb} $, eccentricity $e$, pulsar mass $ M_{\rm p} $, companion mass $ M_{\rm c} $, total mass $ M_{\rm t} $, spin period $ P_{\rm spin} $ of recycled pulsars, spin period derivative $ \dot{P}_{\rm spin} $ and characteristic age $ \tau_{\rm c} $. \[tbl-1\]
Note. $ \tau_{\rm c} $ values are calculated with the relation of $ P_{\rm spin} / (2\dot{P}_{\rm spin}) $.\
$^{a}$ These binaries are formed in global clusters (possibly through dynamical encounters) or not confirmed to be DNSs (maybe contain a white dwarf companion), which are not included in our analyses.\
Recent references. (1) @s18. (2) @k06. (3) @c18. (4) @l16. (5) @f14. (6) @w10. (7) @fst14. (8) @c04. (9) @m17. (10) @m15. (11) @j08. (12) @k09. (13) @c07. (14) @s15. (15) @v15. (16) @j06. (17) @n18. (18) @l12.
------------------------ ------------------------ ------------------------ -------------------- ------------------- -------------------- ------- ----------------------------------- ------ -- -- --
$\Delta M_{\rm ecs} $ $ \sigma_{\rm k,ecs} $ $ \sigma_{\rm k,ccs} $ $ R_{\rm merger} $ $ R_{\rm birth} $ Number $ f $ $ \log(\Lambda(P_{\rm orb}, e)) $ Rank
($ M_{\odot} $) ($ \rm km\,s^{-1} $) ($ \rm km\,s^{-1} $) ($\rm Myr^{-1} $) ($\rm Myr^{-1} $) ($ \times10^{4} $)
1.83$-$2.25 20 150 8 15 9 0.18 $-35.7$ 7
300 3 5 2 0.48 $-36.1$ 13
40 150 7 13 7 0.1 $-36.1$ 11
300 2 4 2 0.28 $-36.1$ 12
80 150 7 12 7 0.06 $-36.2$ 14
300 2 3 1 0.15 $-37.6$ 18
1.83$-$2.5 20 150 8 20 13 0.45 $-35.3$ 4
300 3 9 7 0.76 $-37.0$ 17
40 150 7 16 10 0.31 $-35.6$ 6
300 3 6 4 0.61 $-35.7$ 8
80 150 7 13 8 0.19 $-36.3$ 15
300 3 4 2 0.41 $-36.8$ 16
1.83$-$2.75 20 150 10 27 19 0.64 $-34.9$ 3
300 5 17 12 0.85 $-35.7$ 9
40 150 9 20 13 0.5 $-34.4$ 2
300 4 10 7 0.74 $-34.1$ 1
80 150 8 15 9 0.36 $-35.9$ 10
300 4 7 4 0.58 $-35.4$ 5
------------------------ ------------------------ ------------------------ -------------------- ------------------- -------------------- ------- ----------------------------------- ------ -- -- --
: Under different assumed models for Galactic DNS population, the obtained merger rate $ R_{\rm merger} $ and birthrate $ R_{\rm birth} $, the estimated total number, and the fraction $ f $ of systems that the second NS formed through ECS among all DNSs. The $ \Lambda(P_{\rm orb}, e) $ denotes the likelihood of the observed data for a given model, and the rank of every model is ordered by $ \Lambda(P_{\rm orb}, e)$. \[tbl-1\]
[^1]: Note that the wind mass loss rates of @v17 are recommended only for stripped stars through the interaction with a companion, and may not be suitable for classical WR stars.
[^2]: The DNS formation timescale is $ \sim 10^{7} $ yr and negligible compared with the timespan of 10 Gyr.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The local atomic and magnetic structures of the compounds $A$MnO$_2$ ($A$ = Na, Cu), which realize a geometrically frustrated, spatially anisotropic triangular lattice of Mn spins, have been investigated by atomic and magnetic pair distribution function analysis of neutron total scattering data. Relief of frustration in CuMnO$_2$ is accompanied by a conventional cooperative symmetry-lowering lattice distortion driven by Néel order. In NaMnO$_2$, however, the distortion has a short-range nature. A cooperative interaction between the locally broken symmetry and short-range magnetic correlations lifts the magnetic degeneracy on a nanometer length scale, enabling long-range magnetic order in the Na-derivative. The degree of frustration, mediated by residual disorder, contributes to the rather differing pathways to a single, stable magnetic ground state in these two related compounds. This study demonstrates how nanoscale structural distortions that cause local-scale perturbations can lift the ground state degeneracy and trigger macroscopic magnetic order. [^1]'
author:
- 'Benjamin A. Frandsen'
- 'Emil S. Bozin'
- Eleni Aza
- Antonio Fernández Martínez
- Mikhail Feygenson
- Katharine Page
- Alexandros Lappas
title: Nanoscale Degeneracy Lifting in a Geometrically Frustrated Antiferromagnet
---
Introduction
============
In geometrically frustrated materials, the spatial arrangement of magnetic moments prevents the simultaneous satisfaction of competing magnetic interactions [@greed;jmc01]. This hinders the formation of magnetic order and may occasionally lead to exotic, macroscopically degenerate spin-liquid ground states for which magnetic order is absent even at zero temperature [@balen;n10; @lee;jpsj10]. In the more common case where order develops at finite temperature, the transition mechanism itself can be quite unusual, often originating from extreme sensitivity to small perturbations resulting from thermal fluctuations, single-ion anisotropy, Dzyaloshinskii-Moriya and dipolar interactions, disorder, magnetoelastic coupling, and more [@moess;pt06]. Coupling between magnetic and other degrees of freedom can lift the intrinsic degeneracies and even support novel phenomena beyond magnetism, such as magnetoelectric states established by uniform lattice distortions [@lee;n08] and domain-wall-driven competing metastable microphases [@kamiy;prl12].
An added layer of complexity may arise when frustrated geometries occur in strongly correlated transition metal oxide (TMO) systems. The possibility of stabilizing unusual, spatially inhomogeneous ground states through competition among simultaneously active electronic degrees of freedom (i.e. spin, charge, and orbital) has been well documented in TMOs [@dagot;s05; @sheno;cpc06]. Combining this affinity for spatial inhomogeneity with geometrical frustration may produce novel behavior.
The insulating sister compounds $\alpha$- (NMO) [@paran;jssc71] and (CMO) [@topfe;zk95] provide a valuable testbed for exploring geometrically frustrated magnetism in strongly correlated TMOs. The $A$MnO$_2$ crystal structures \[Fig. \[fig:structure\](a, b)\] entail spacer layers of primarily monovalent $A$ = Na or Cu cations that separate layers of edge-sharing MnO$_6$ octahedra.
{width="140mm"}
The Jahn-Teller active, antiferromagnetically (AF) coupled Mn$^{3+}$ ions ($S=2$) form an anisotropic triangular network, embodying a simple example of a frustrated topology imposed on the two-dimensional Heisenberg model. At ambient conditions, the triangles are isosceles with $r_1 < r_2 = r_3$, as shown in Fig. \[fig:structure\](c). Consequently, the exchange interactions $J$ (along $r_1$) and $J'$ (along $r_2$ and $r_3$) become nondegenerate, with $J'/J < 1$. The frustration of the triangular lattice is therefore partially relieved. Nevertheless, the magnetic interactions along the two isosceles legs remain degenerate, and the resulting partial frustration of the AF interactions remains important even though these systems would not be considered “highly frustrated” magnets. Indeed, the magnetic frustration has been implicated in numerous unusual properties displayed by these systems, such as quasi-one-dimensional magnetic interactions [@stock;prl09] and rare Higgs-type excitations [@dally;nc18] in NMO. In general, systems with imperfect frustration, e.g. the anisotropic triangular lattice considered here, may still exhibit rich behavior such as ordering by disorder and Peierls-type transitions [@becca;prl02].
In CMO, magnetic long-range order (LRO) [@damay;prb09] develops below = 65 K. This is accomplished through a concomitant monoclinic-to-triclinic structural phase transition that further distorts the triangles such that $r_2 \ne r_3$, fully lifting the degeneracy of the triangular lattice and selecting a unique magnetic ground state [@vecch;prb10]. The ordered state exhibits AF alignment along the shorter formerly isosceles leg and ferromagnetic alignment along the other formerly isosceles leg \[Fig. \[fig:structure\](a)\], with antiferromagnetic coupling between layers [@damay;prb09]. The energy gain from relieving the magnetic frustration outweighs the elastic energy cost of the triclinic distortion [@zorko;sr15]; hence, this effect is magnetically driven. Fig. \[fig:structure\](c) displays the projection of the monoclinic and triclinic unit cells on the Mn layer.
Although NMO is more frustrated than CMO ($J'/J = 0.44$ [@zorko;prb08] compared to 0.27 for CMO [@zorko;sr15], resulting in a greater degree of frustration [@zheng;prb05]), it also exhibits AF LRO below = 45 K \[see Fig. \[fig:structure\](b)\]. It differs from the magnetic structure in CMO only in that the coupling of adjacent manganese layers stacked along the *c* axis in NMO is ferromagnetic, in contrast to the antiferromagnetic coupling between layers in CMO [@giot;prl07]. Moreover, significant short-range magnetic correlations survive well into the paramagnetic state in NMO [@giot;prl07]. Surprisingly, no long-range structural phase transition accompanies the magnetic transition, implying that some other, unknown means must relieve the magnetic degeneracy. Recent work further suggests that the magnetic ground state is intrinsically inhomogeneous [@zorko;nc14], adding to the unusual nature of the magnetism in this compound. Finally, this system exhibits polymorphism [@abaku;cm14] and significant structural disorder [@orlan;prm18] that may influence the magnetostructural properties, although how this may happen remains an open question. Clarifying these issues in NMO continues to be an important objective in the study of $A$MnO$_2$ systems, while also presenting a valuable opportunity for a more general study of magnetic frustration, disorder, and their interplay in TMOs. This complements earlier work focused on magnetic pyrochlores and related materials [@gardn;rmp10].
Here, we discover that the magnetic phase transition in NMO occurs via an unconventional mechanism in which a nanoscale structural distortion supports macroscale magnetic order. Using atomic pair distribution function (PDF) analysis of neutron total scattering data [@egami;b;utbp12], we observe a short-range triclinic distortion in NMO correlated over a 2-nm length scale that exists well above and grows significantly as the temperature decreases. Complementary magnetic PDF analysis [@frand;aca14; @frand;aca15; @frand;prl16] reveals magnetic short-range order (SRO) persisting well above , which likewise grows in magnitude and correlation length upon cooling until finally achieving the LRO state below . Crucially, quantitative refinements confirm that the short-range triclinic distortion lifts the degeneracy of the isosceles triangular network on the nanoscale, thereby enabling the LRO state and enhancing the magnetic correlations above the transition.
While this short-range triclinic distortion is the primary actor, we find that other types of quenched structural disorder also play a role in determining the eventual ground state—specifically, Na vacancies in NMO, and antisite defects involving Cu atoms on the Mn site in CMO. Beyond a certain threshold, this type of disorder suppresses the long-range triclinic structural phase transition, resulting in the preservation of monoclinic symmetry in the average structure of NMO. Taken together, these results describe an unusual mechanism for establishing magnetic LRO in the presence of frustration and a spatially inhomogeneous lattice, through which short-range correlations in the lattice sector promote long-range correlations in the magnetic sector.
Materials and Methods
=====================
Powder samples of and were prepared as described in Ref. [@vecch;prb10; @abaku;cm14]. Temperature-dependent time-of-flight neutron total scattering measurements were conducted using the Nanoscale-Ordered Materials Diffractometer (NOMAD) instrument [@neufe;nimb12] at the Spallation Neutron Source (SNS) of Oak Ridge National Laboratory (ORNL) and the General Materials Diffractometer (GEM) [@willi;pb97] at the ISIS neutron source of Rutherford Appleton Laboratory (RAL). A maximum momentum transfer of 35 Å$^{-1}$ was used when generating the real-space PDF patterns. Analysis and modeling of the atomic PDF were performed with the PDFgui program [@farro;jpcm07] and the DiffPy suite of programs for diffraction analysis [@juhas;aca15]. The diffpy.mpdf package was used for the magnetic PDF analysis. Further details about the atomic and magnetic PDF refinements are provided in the Supplementary Information [^2]. Rietveld refinements were performed within the General Structure Analysis System (GSAS) software [@toby;jac01] using data collected from the high-resolution backscattering banks of the GEM diffractometer.
Results
=======
***Atomic PDF Analysis.*** We first present the atomic PDF analysis of NMO and CMO. The PDF, obtained from the sine Fourier transform of the reduced total scattering structure function $F(Q) = Q(S(Q)-1)$, provides a detailed account of the local atomic structure in real space, irrespective of the character of the long-range structure. Representative PDF patterns for both compounds at room temperature are shown in Fig. \[fig:structure\](d). The expected monoclinic structural models explain the data well, vouching for the quality of the samples. Rietveld refinements of the Bragg intensities verified these results and allowed us to refine the average site occupancies. From this analysis, the actual compositions of our samples are Na$_{0.92}$MnO$_2$ and Cu(Mn$_{0.98}$Cu$_{0.02}$)O$_2$, in line with compositions reported in the literature [@dally;nc18; @dally;prb18]. For simplicity, we continue to label these samples as and .
Intriguingly, the low-temperature PDF fits to the NMO data using the monoclinic model are worse at low $r$ ($<20$ Å) than at high $r$, despite the fact that the average structure is known to be monoclinic at all temperatures. This implies that the local structure on such a short length scale deviates from the average monoclinic structure. Indeed, a triclinic structural model equivalent to that of CMO provides a significantly improved fit to the low-$r$ region at low temperature, as seen in Fig. \[fig:strucFitResults\](a).
{width="140mm"}
Inspections of the low-$r$ fit residuals (lower green curves) for the monoclinic and triclinic models of NMO at 5 K reveals that high-frequency features are consistently smaller for the triclinic model, resulting in a relative improvement in $\chi^2$ of 40% over the monoclinic model. We therefore conclude that the ground state of NMO consists of regions of locally triclinic symmetry existing within the average monoclinic structure; hence, the structure is spatially inhomogeneous. We note that the PDF measurements cannot distinguish between static and dynamic distortions, so the temporal nature of this inhomogeneity is unclear. Importantly, this short-range triclinic distortion lifts the degeneracy of the triangular lattice locally, irrespective of the fact that the average crystal structure retains its monoclinic symmetry to the lowest temperature. The broad, lower-frequency signal seen in the fit residuals in Fig. \[fig:strucFitResults\](a) arises from the magnetic contribution to the PDF, which we will discuss later.
Establishing the temperature dependence of this local triclinic distortion is crucial for understanding its connection to the magnetism in NMO. To this end, we refined both the monoclinic and triclinic structural models against the PDF data, which were collected on a dense temperature grid between 5 and 300 K. For each temperature $T$, fits were performed on a sliding data window ranging from \[1.8 Å - 15 Å\] to \[31.8 Å - 45 Å\] in 2-Å steps. We then computed the normalized difference in $\chi^2$ between the monoclinic and triclinic models as $\Delta \chi^2 / \langle \chi^2 \rangle = \left(\chi^2_{\mathrm{mono}}-\chi^2_{\mathrm{tri}}\right)/\left(0.5\left[\chi^2_{\mathrm{mono}}+\chi^2_{\mathrm{tri}}\right]\right)$ for every refinement, quantifying the improvement offered by the triclinic model: the larger this normalized difference in $\chi^2$, the better the triclinic model relative to the monoclinic model.
In Fig. \[fig:strucFitResults\](b), we plot $\Delta \chi^2 / \langle \chi^2 \rangle$ as a function of $T$ for the shortest and longest fitting ranges. The triclinic model performs significantly better than the monoclinic model for the short fitting range at low $T$. As $T$ increases, $\Delta \chi^2 / \langle \chi^2 \rangle$ steadily decreases until it plateaus at a nonzero value for temperatures above $\sim$150-200 K. For the longer fitting ranges, the triclinic model provides a markedly smaller relative improvement, with no pronounced changes as $T$ varies. These results confirm that the triclinic distortion in NMO has a short-ranged nature and is strongest at low temperature but may persist up to room temperature.
A more comprehensive view of the $T$- and $r$-dependence of this local distortion is given by the color map in Fig. \[fig:strucFitResults\](c). The intensity scales with the triclinic splitting $d = r_2 - r_3$, which is responsible for lifting the local degeneracy of the isosceles triangles. Temperature appears on the vertical axis, and the horizontal axis displays the midpoint of the fitting range. To remove any spurious results arising from possible overfitting with the triclinic model, we scaled the $d$ values by the difference in $\chi^2$ between the monoclinic and triclinic models (see Supplementary Information for further details). The figure indicates that the greatest distortion ($d \sim 0.024$ Å) occurs below $\sim$7 K and at low $r$. This distortion is roughly three times as large as that observed in the average structure of CMO ($d \sim 0.008 $ Å). Averaging over longer distances reduces the distortion, evidenced by the weaker intensity on the right side of the color map. At low temperature, we estimate the correlation length of the local triclinic distortion to be about 20 Å. With increasing temperature, the magnitude of the distortion decreases, although it remains nonzero even at room temperature.
We display the equivalent results for CMO in Fig. \[fig:strucFitResults\](d). Because of the long-range structural phase transition at 65 K, the triclinic splitting remains nonzero over the entire real-space range below this temperature, albeit with somewhat enhanced splitting for short distances, suggesting incomplete ordering. Above 65 K, short-range triclinicity remains, but with a smaller magnitude and shorter real-space extent than in NMO. Overall, we observe that NMO displays significantly more spatial inhomogeneity in the form of local triclinic distortions than does CMO.
Analysis of the local octahedral environment reveals that no dramatic changes in the MnO$_6$ octahedra occur in either compound as the temperature is varied. The splitting of the equatorial Mn-O bonds is commensurate with the overall triclinic splitting of the $a$ and $b$ lattice parameters ($r_2$ and $r_3$ in Fig. \[fig:structure\]). The PDF data at room temperature reveal that the magnitude of the Jahn-Teller (JT) distortion, defined as the ratio of the apical to equatorial Mn-O bonds ($r_{ap}/r_{eq}$) in the MnO$_6$ octahedra, is smaller in (1.173; $r_{ap}=2.258(1) \mathrm{\AA}$, $r_{eq}=1.925(1) \mathrm{\AA}$) than in (1.240; $r_{ap}=2.390(1) \mathrm{\AA}$, $r_{eq}=1.928(1) \mathrm{\AA}$). The JT distortion in shows a very slight decrease of 0.1% as the temperature is lowered from room temperature to 5 K. In , the JT distortion decreases by 0.6% over that temperature range. In the triclinic state, the equatorial bonds become slightly unequal, so we take their average when computing the distortion magnitude.
***Magnetic PDF Analysis.*** Having investigated the local atomic structures of NMO and CMO, we now turn to their magnetic structures. Fig. \[fig:SQcmap\](a) displays a color map of the total scattering structure function $S(Q)$ for NMO, with temperature on the vertical axis and $Q$ on the horizontal axis.
{width="160mm"}
The nuclear Bragg peaks are clearly visible as the intense, temperature-independent features. Below 45 K, several well-defined magnetic Bragg peaks exist between the nuclear peaks, consistent with the expected long-range AF order. Some diffuse scattering remains visible below 45 K, likely due to the presence of mixed paramagnetic and AF phases in the vicinity of the transition. As the temperature increases, the magnetic LRO melts, causing the sharp magnetic Bragg peaks to dissipate into diffuse scattering that persists up to the highest temperatures measured and extends over a broad $Q$-range. Importantly, careful inspection of this diffuse scattering reveals that its $Q$-dependence remains non-uniform well above the transition temperature, indicative of persistent short-range magnetic correlations. To illustrate this further, we display in Fig. \[fig:SQcmap\](b) cuts of $S(Q)$ at 5 K, 50 K, and 260 K. The strong and sharp magnetic Bragg peaks at 5 K give way to a weaker and more diffuse signal at 50 K, yet distinct features are still observable in the scattering pattern. At 260 K, any structure in the diffuse scattering is much less well defined.
Additional insight can be gained by analyzing the magnetic correlations directly in real space using the magnetic PDF (mPDF) method, which is sensitive to local magnetic structure even in the absence of long-range magnetic order [@frand;prl16; @frand;prm17]. Experimentally, the mPDF signal is contained in the fit residual produced by subtracting the best-fit atomic PDF from the total PDF signal; additional details appear in the Supplementary Information. Representative mPDF patterns for NMO at 5 K and 50 K are displayed in Fig. \[fig:SQcmap\](c). These patterns correspond to the long-wavelength oscillatory signal in the atomic PDF fit residuals shown in Fig. \[fig:strucFitResults\](a). The thin gray curves show the unfiltered mPDF signal mixed in with high-frequency noise, which is propagated through the Fourier transform due to the large value of $Q_{\mathrm{max}}$ (chosen to optimize the atomic PDF data). Applying a Fourier filter to remove all frequencies above 6.9 Å$^{-1}$, where the squared magnetic form factor drops below 1% its maximal value, results in the thick black curves in Fig. \[fig:SQcmap\](c). This provides a cleaner signal while preserving the contribution from the magnetic scattering. The overlaid colored curves are fits that will be described subsequently. Simple inspection of the mPDF data at 5 K reveals a well-defined signal over the full $r$-range displayed, indicative of magnetic LRO. The alternating negative and positive peaks reflect the AF spin configuration [@frand;aca14]. At 50 K, however, the signal is reduced in amplitude and fully suppressed in real space beyond about 20 Å, pointing to short-range magnetic correlations with a length scale of $\sim$2 nm.
To investigate the magnetic correlations more quantitatively, we modeled the mPDF using the known LRO AF structure published in Ref. . Our model included as free parameters a scale factor (proportional to the square of the locally ordered moment), a real-space correlation length $\xi$, and two polar angles defining the spin direction. The mPDF fits shown in Fig. \[fig:SQcmap\](c) for 5 and 50 K agree well with the data. At 5 K, the correlation length is limited by the instrumental resolution, and the refined spin direction matches previously published results [@giot;prl07]. Earlier work reported intrinsic inhomogeneities in the magnetic ground state in the form of nanometer-sized bubbles possessing reversed sublattice magnetization relative to the surrounding matrix of Mn spins [@zorko;nc14]. The presence of these bubbles would cause an additional damping in the mPDF signal, which we do not observe. Therefore, any such low-temperature magnetic inhomogeneity, if present, must exist on a length scale of $\sim$5 nm or greater, beyond the sensitivity of the mPDF data. We note that the incommensurability of the magnetic order reported in Ref. would result in a real-space modulation outside the accessible data window of these measurements. At 50 K, the fit yields a correlation length of $\sim$15 Å with a nearly unchanged spin direction. Interestingly, the fit can be improved by about 5% if only a single two-dimensional (2D) layer is considered and by 8% if just a one-dimensional (1D) chain of nearest-neighbor spins is used, suggesting that the short-range magnetic correlations develop a lower-dimensional character above . This agrees with earlier inelastic neutron scattering analysis [@stock;prl09; @dally;nc18; @dally;prb18].
In Fig. \[fig:mPDF-both\](a), we display the refined magnetic correlation length $\xi$ plotted against the dimensionless temperature $T/T_{\mathrm{N}}$. Beginning slightly above $3T_{\mathrm{N}}$, $\xi$ is approximately equal to the Mn-Mn nearest-neighbor (NN) distance (illustrated by the horizontal dashed line). Large error bars above $4T_{\mathrm{N}}$ limit our ability to specify $\xi$ at these temperatures.
![\[fig:mPDF-both\] **Temperature evolution of magnetic correlations in and .** (a) Correlation length of the short-range antiferromagnetic order in determined from mPDF refinements using a three-dimensional model, displayed as a function of dimensionless temperature $T/T_{\mathrm{N}}$. Below the transition (indicated by the vertical dashed line), the correlation length is limited by the instrumental resolution. The horizontal dashed line indicates the nearest-neighbor Mn-Mn distance. Inset: Refined mPDF scale factor as a function of dimensionless temperature for a fitting range of 1.8 to 15 Å (blue circles) and 30 to 45 Å (orange squares). (b) Equivalent results for .](fig_mPDF-both_rescaled_vert){width="70mm"}
Below $3T_{\mathrm{N}}$, the correlation length steadily increases as the temperature decreases, until it reaches its resolution-limited value below the ordering temperature of $\sim$45 K, marked by the vertical dashed line. The inset of Fig. \[fig:mPDF-both\](a) illustrates the temperature dependence of the mPDF scale factor for refinements over 1.8 - 15 Å and 31.8 - 45 Å. For the longer fitting range, the scale factor is zero at high temperature and shows a sharp increase for $T < T_{\mathrm{N}}$, indicating a rapid onset of long-range magnetic order. In contrast, the short-range scale factor shows a much more gradual decrease as the temperature increases, remaining nonzero to 300 K, the highest measured temperature. This quantifies the persistence of significant short-range magnetic correlations at temperatures far above the long-range transition, as expected from our previous qualitative analysis of the scattering data (Fig. \[fig:SQcmap\]). We note that the correlation length and short-range scale factor track proportionally with the amplitude of the local lattice distortion, as seen by comparing the evolution of the local atomic structure in Fig. \[fig:strucFitResults\](b,c) with that of the magnetic correlations in Fig. \[fig:mPDF-both\]a, suggesting that they are closely connected.
The corresponding results for CMO are presented in Fig. \[fig:SQcmap\](d-f) and Fig. \[fig:mPDF-both\](b). Qualitatively, CMO shows similar magnetic behavior as NMO, with short-range correlations surviving into the paramagnetic phase above $T_{\mathrm{N}} = 65$ K. Quantitatively, however, we find that these short-range correlations in CMO are weaker and more limited in spatial extent than for NMO. As seen in Fig. \[fig:mPDF-both\], the correlation length and short-fitting-range scale factor decrease above significantly more rapidly in CMO than in NMO.
Discussion
==========
The key experimental findings obtained from the PDF and mPDF analyses are: (1) NMO exhibits a short-range triclinic distortion correlated over $\sim$2 nm, lifting the magnetic degeneracy locally; (2) this distortion grows in amplitude upon cooling; (3) short-range AF correlations are present in NMO at least up to 300 K and grow in strength and correlation length upon cooling, tracking with the structural distortion; and (4) CMO exhibits a smaller local triclinic distortion and weaker AF correlations than does NMO.
Several insights can be gleaned from these results. First, they clarify the structural ground state of NMO, demonstrating that the average structure remains monoclinic, while the instantaneous local structure is triclinically distorted. Second, the observation that the triclinic distortion and magnetic SRO both grow upon cooling with a very similar temperature dependence indicates that these two effects are closely related. The relatively weaker local structural distortion and AF correlations in the paramagnetic phase of CMO compared to NMO further supports this. Third, the fact that the short-range structural distortion relieves the local magnetic degeneracies naturally suggests that it promotes the growth of the AF magnetic correlations. Intriguingly, the AF correlation length increases upon cooling until matching the $\sim$2 nm length scale of the local triclinic distortion around , at which point magnetic LRO occurs. This may indicate that magnetic coherence across neighboring triclinic regions triggers the transition. Together, these findings provide a cohesive picture to understand the magnetic transition in NMO, in which short-range structural distortions lift the magnetic degeneracy on the nanoscale and eventually enable long-range magnetic order. In a more general context, NMO appears to be a rare example of a geometrically frustrated magnet exhibiting long-range magnetic order coupled to a short-range lattice distortion. Separately, the local triclinic distortion observed in NMO by PDF may explain various unusual experimental results in the literature, including the subtle anisotropic broadening of the low-temperature Bragg peaks seen in previous diffraction work [@giot;prl07], the nanoscale inhomogeneities of the ground state suggested by muon spin relaxation and nuclear magnetic resonance experiments [@zorko;nc14; @zorko;sr15], and the emergent magnetoelectric response [@orlan;prm18].
Given the fundamentally similar structural and magnetic ingredients of NMO and CMO, a natural question to ask is why NMO accomplishes magnetic LRO via a short-range structural distortion, while CMO does the same with a more conventional long-range phase transition. One possibility arises from earlier ab-initio calculations, which predicted a larger energy gain resulting from the triclinic structure in the AF state for CMO than NMO [@zorko;sr15]. However, somewhat surprisingly, the PDF analysis demonstrates that NMO is actually more triclinic on a nanometer length scale than the average CMO triclinic distortion, counter to the predictions. Another possibility would be the presence of stronger magnetoelastic coupling in CMO than in NMO, yet this is also unlikely to be the explanation, as calculations have shown that the magnetoelastic energy scale in both materials is 2-3 orders of magnitude weaker than the magnetic exchange interactions [@zorko;sr15].
Instead, we suggest that structural disorder plays a decisive role. In these complex systems, multiple types of disorder are present. First, the local triclinic distortion discussed previously is present in both systems and plays an active role in promoting AF correlations by lifting the magnetic degeneracy locally. Second, $A$-site occupancy disorder in the form of Na vacancies and excess Cu on the Mn site exist in NMO and CMO, respectively. Third, planar defects due to oxygen-layer gliding exist in both materials. The experimental evidence here and elsewhere indicates that NMO exhibits significantly more structural disorder than CMO. This is evidenced by the more pronounced distortion of the local structure in NMO than CMO, the greater percentage of site occupancy disorder (8% Na vacancies versus 2% excess Cu from antisite defects), and a higher density of planar defects (see Supplementary Information) [@abaku;cm14; @cleme;jes15; @orlan;prm18]. Note that the Na vacancies and Cu impurities allow for conversion of Mn$^{3+}$ ($S=2$) to Mn$^{4+}$ ($S=3/2$) [@li;nm14], which could also create magnetic disorder. It is tempting to relate such mixed-valency mediated disorder to Zener-type carriers [@zener;pr51a; @zener;pr51b], which may cause local inhomogeneities in the spin structure when entering bound states [@degen;pr60]. Verifying this scenario is beyond the resolving capabilities of our experimental probes.
These afore-mentioned factors of quenched disorder [@burgy;prl01] contribute to significant exchange disorder, i.e. randomness in the magnetic exchange interactions between nearby spins. Theoretically, the NN Heisenberg model in the presence of arbitrarily weak magnetoelastic coupling predicts a critical level of exchange disorder. Below this level, a frustrated system enters the LRO Néel state through a concomitant lowering of its average lattice symmetry; above the critical level, the Néel state arises without any long-range structural distortion [@saund;prl07; @saund;prb08]. Our results suggest that CMO rests below the critical level of disorder and NMO above it, although differing degrees of sensitivity to perturbations between the two systems may also contribute [@moess;cjp01]. Considering also that NMO is more frustrated than CMO and likely exhibits stronger electron-lattice coupling via an enhanced Jahn-Teller (JT) distortion [@giot;prl07; @vecch;prb10], the larger exchange disorder may lock the spin order within the local triclinicity and thus help NMO exit degeneracy without a long-range, symmetry-lowering phase transition as in the less frustrated and weaker JT-distorted CMO.
These arguments are in line with the observation that CuMn$_{1-x}$Cu$_x$O$_2$ displays a long-range triclinic distortion for $x < 7\%$ (which includes the current CMO sample with $x = 2\%$), but retains the high-symmetry monoclinic average structure for more disorderd compounds with $x \ge 7\%$ [@garle;prb11; @poien;cm11]. In this picture, compounds with larger values of $x$ possess a level of exchange disorder that exceeds the critical level, preventing the long-range structural transition to the triclinic phase, as in the case of . Additionally, the magnetic structure in CuMn$_{1-x}$Cu$_x$O$_2$ exhibits antiferromagnetic coupling between layers for low values of $x$ in which the triclinic phase is observed, but changes to ferromagnetic coupling between layers for higher $x$ where the long-range triclinic transition is suppressed [@terad;prb11; @garle;prb11]. also shows ferromagnetic coupling between the layers, further supporting the analogy between and heavily Cu-doped . Whether the CuMn$_{1-x}$Cu$_x$O$_2$ system with $x \ge 7\%$ also shows a local triclinic distortion and significant short-range magnetic correlations at high temperature like NMO would be an informative follow-up study. More generally speaking, we note that the importance of local structural disorder revealed here supports similar recent findings in the canonical pyrochlore Yb$_2$Ti$_2$O$_7$ [@bowma;nc19].
In summary, the combined atomic and magnetic PDF analysis of NMO reveals an unusual mechanism of relieving geometrical frustration in which a short-range lattice distortion results in long-range magnetic order. Comparison with CMO in the paramagnetic phase further establishes the close relationship between the short-range structural and antiferromagnetic correlations, while also highlighting the role that structural disorder such as Na vacancies and Cu antisite defects plays in ground state selection for frustrated systems. More generally, this work illuminates the cooperative intertwining of the local atomic and magnetic structures that can occur when spatial inhomogeneity meets geometrical frustration. We suggest that similar physics may be at play in numerous other TMO systems hosting geometrically frustrated magnetism. Finally, the nanoscale magnetic degeneracy lifting that occurs above shares similarities with the nanoscale *orbital* degeneracy lifting observed in other systems, such as the manganites [@qiu;prl05], iron-chalcogenides [@frand;prb19; @koch;prb19], and iridates [@bozin;sr14; @bozin;nc19]. Together, these findings contribute to an emerging understanding of the importance of spontaneous local degeneracy lifting that appears above long-range, symmetry-broken ground states in complex materials.
*Acknowledgments.* This research used resources at the Spallation Neutron Source, a DOE Office of Science User Facility operated by the Oak Ridge National Laboratory. ORNL is managed by UT-Battelle, LLC, under Contract No. DE-AC0500OR22725 for the U.S. Department of Energy. B.A.F. acknowledges support from Brigham Young University. Work at Brookhaven National Laboratory was supported by US DOE, Office of Science, Office of Basic Energy Sciences (DOE-BES) under Contract No. DE-SC0012704. This work has been partially supported by U.S. DOE Grant No. DE-FG02-13ER41967. We thank the Science and Technology Facilities Council (STFC) for the provision of neutron beam time at ISIS Facility. We acknowledge Dr. Jörg Neuefeind for assistance with the measurements at the NOMAD beamline and Dr. Dave Keen for assistance on the GEM beamline. Access to the facilities of the Integrated Infrastructure Initiative ESTEEM2 (EU 7th FP program with reference number 312483) is gratefully acknowledged. We are indebted to Prof. Artem M. Abakumov for the TEM-based characterizations at EMAT. AL thanks the Fulbright Foundation - Greece.
[53]{} ifxundefined \[1\][ ifx[\#1]{} ]{} ifnum \[1\][ \#1firstoftwo secondoftwo ]{} ifx \[1\][ \#1firstoftwo secondoftwo ]{} ““\#1”” @noop \[0\][secondoftwo]{} sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{} @startlink\[1\] @endlink\[0\] @bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1143/JPSJ.79.011004) @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase
10.1103/PhysRevLett.109.067204) [****, ()](\doibase
10.1126/science.1107559) @noop [****, ()]{} [****, ()](\doibase https://doi.org/10.1016/0022-4596(71)90001-6) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevB.82.094404) [****, ()](\doibase 10.1103/PhysRevLett.99.247211) [****, ()](\doibase 10.1103/PhysRevLett.103.077202) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.89.037204) [****, ()](\doibase
10.1103/PhysRevB.80.094410) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevB.77.024412) [****, ()](\doibase 10.1103/PhysRevB.71.134422) [****, ()](\doibase
10.1038/ncomms4222) [****, ()](\doibase 10.1021/cm5011696) [****, ()](\doibase 10.1103/PhysRevMaterials.2.074407) [****, ()](\doibase 10.1103/RevModPhys.82.53) [**](http://store.elsevier.com/product.jsp?lid=0&iid=73&sid=0&isbn=9780080971414), ed. (, , ) [****, ()](\doibase 10.1107/S2053273313033081) [****, ()](\doibase
10.1107/S205327331500306X) [****, ()](\doibase 10.1103/PhysRevLett.116.197204) [****, ()](\doibase 10.1016/j.nimb.2012.05.037) [****, ()](\doibase
https://doi.org/10.1016/S0921-4526(97)00561-9), [****, ()](\doibase
10.1088/0953-8984/19/33/335219) [****, ()](\doibase
10.1107/S2053273315014473) [****, ()](\doibase
10.1107/S0021889801002242) [****, ()](\doibase 10.1103/PhysRevB.98.144444) [****, ()](\doibase 10.1103/PhysRevMaterials.1.074412) @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevLett.87.277202) [****, ()](\doibase 10.1103/PhysRevLett.98.157201) [****, ()](\doibase 10.1103/PhysRevB.77.214438) [****, ()](\doibase 10.1139/p01-123) [****, ()](\doibase
10.1103/PhysRevB.83.172407) [****, ()](\doibase 10.1021/cm102483m) [****, ()](\doibase
10.1103/PhysRevB.84.064432) @noop [****, ()]{} [****, ()](http://journals.aps.org/prl/abstract/10.1103/PhysRevLett.94.177203) @noop [****, ()]{} [****, ()](\doibase 10.1103/PhysRevB.100.020501) [****, ()](\doibase 10.1038/srep04081) @noop [****, ()]{}
[^1]: Notice: This manuscript has been co-authored by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE). The US government retains and the publisher, by accepting the article for publication, acknowledges that the US government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this manuscript, or allow others to do so, for US government purposes. DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan).
[^2]: The Supplementary Information contains additional details about the atomic and magnetic pair distribution function analysis and the transmission electron microscopy characterization of and .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Regge calculus is a powerful method to approximate a continuous manifold by a simplicial lattice, keeping the connectivities of the underlying lattice fixed and taking the edge lengths as degrees of freedom. The discrete Regge model employed in this work limits the choice of the link lengths to a finite number. To get more precise insight into the behavior of the four-dimensional discrete Regge model, we coupled spins to the fluctuating manifolds. We examined the phase transition of the spin system and the associated critical exponents. The results are obtained from finite-size scaling analyses of Monte Carlo simulations. We find consistency with the mean-field theory of the Ising model on a static four-dimensional lattice.'
address:
- |
Atominstitut der Österreichischen Universitäten, TU Wien, A-1040 Vienna, Austria\
Institut für Theoretische Physik, Universität Leipzig, D-04109 Leipzig, Germany
- ' Institut für Theoretische Physik, Universität Leipzig, D-04109 Leipzig, Germany'
- 'Atominstitut der Österreichischen Universitäten, TU Wien, A-1040 Vienna, Austria'
author:
- 'E. Bittner'
- 'W. Janke'
- 'H. Markum'
title: 'Ising spins coupled to a four-dimensional discrete Regge skeleton'
---
Introduction {#intro}
============
Spin systems coupled to fluctuating manifolds are analyzed as a simple example for matter fields coupled to Euclidean quantum gravity. The gravitational action is unbounded from below due to conformal fluctuations. But that does not necessarily render its quantum theory useless or the path integral undefined. Indeed, the existence of a well defined phase was the first and probably most important result of the numerical simulations in four dimensions [@hamber84; @berg]. Its existence and stability were explored in some detail using the standard Regge calculus with continuous link lengths. It turned out that the well defined phase is stable against variations of the measure and the lattice size [@beirl94].
In the discrete Regge model the problem of an unbounded action is not present as in standard Regge calculus. Because of the restriction of possible quadratic link lengths to two values [@beirlz2; @fleming] in the discrete Regge model the action can only reach an extreme but finite value. The expectation values do not diverge if the well defined phase is left. What happens is that the lattice “freezes” at large positive and negative values of the gravitational coupling, as expected for a spin system [@homo]. To get more precise ideas about the behavior of the four-dimensional discrete Regge model, we coupled Ising spins to the fluctuating manifolds. We examined the phase transition of the spin system and the associated critical exponents.
The rest of the paper is organized as follows. In Sec. \[sect2\] we introduce the discrete Regge model and give some details of the analyzed observables. The results of the Monte Carlo simulations are presented in Sec. \[sect3\], and concluding remarks can be found in Sec. \[sect4\].
Models and observables {#sect2}
======================
The situation for the discrete Regge model is both structurally and computationally much simpler than the standard Regge calculus with continuous link lengths. The restriction of the edge lengths to just two values was carefully examined in 2D where an interpolation from $Z_2$ to $Z_\infty$ was performed [@bittnerzn]. It turned out that the phase transition with respect to the cosmological constant is universal. This was tested for pure gravity in 2D and is expected to be the case also in 4D. Compared with standard Regge calculus, numerical simulations of the discrete Regge model can be done more efficiently by implementing look-up tables and using the heat-bath algorithm. In the actual computations we took the squared link lengths as $q_{ij} \equiv q_l = b_l ( 1+\epsilon \sigma_l)$ with $\sigma_l \pm 1$. The Euclidean triangle inequalities are satisfied automatically as long as $\epsilon < \epsilon_{\rm max}$. Because a four-dimensional Regge skeleton with equilateral simplices cannot be embedded in flat space, $b_l$ takes different values depending on the type of the edge $l$. In particular $b_l=1,2,3, 4$ for edges, face diagonals, body diagonals, and the hyperbody diagonal of a hypercube.
We investigated the partition function $$\label{z}
Z=\sum_{\{s\}}\int D[q] \exp\left[-I(q) - K E(q,s)\right],$$ where $I(q)$ is the gravitational action, $$I(q) = - \beta_g \sum_t A_t \delta_t + \lambda \sum_i V_i.$$ The first sum runs over all products of triangle areas $A_t$ times corresponding deficit angles $\delta_t$ weighted by the gravitational coupling $\beta_g$. The second sum extends over the volumes $V_i$ of the 4-simplices of the lattice and allows one together with the cosmological constant $\lambda$ to set an overall scale in the action. The energy of Ising spins $s_i \in Z_2$, $$E(q,s) = \frac{1}{2} \sum_{\langle ij \rangle} A_{ij}\frac{(s_i-s_j)^2}{q_{ij}},$$ is defined as in two dimensions [@gross; @holm; @physA], with the barycentric area $A_{ij}$ associated with a link $l_{ij}$, $$A_{ij} = \sum _{\mbox{\scriptsize $t \supset l_{ij}$}}
\frac{1}{3} A_t.$$ We chose the simple uniform measure as in the pure gravity simulations [@homo]: $$D[q]=\prod_l dq_l {\cal F}(q_l).$$ The function ${\cal F}$ ensures that only Euclidean link configurations are taken into account, i.e., ${\cal F}=1$ if the Euclidean triangle inequalities are fulfilled and ${\cal F}=0$ otherwise. This is always guaranteed for the discrete Regge model by construction.
For every Monte Carlo simulation run we recorded the time series of the energy density $e=E/N_0$ and the magnetization density $m= \sum_i s_i /N_0$, with the lattice size $N_0=L^4$. To obtain results for the various observables ${\cal O}$ at values of the spin coupling $K$ in an interval around the simulation point $K_0$, we applied the reweighting method [@ferrenberg]. Since we recorded the time series, this amounts to computing $$\langle {\cal O} \rangle |_K =\frac{\langle {\cal O}e^{-\Delta K E}\rangle|_{K_0}}
{\langle e^{-\Delta K E}\rangle|_{K_0}}~,$$ with $\Delta K = K - K_0$.
With the help of the time series we can compute the specific heat, $$C(K)=K^2 N_0 (\langle e^2 \rangle- \langle e\rangle^2)~,$$ the (finite lattice) susceptibility, $$\chi(K)=N_0(\langle m^2 \rangle -\langle |m| \rangle^2)~,$$ and various derivatives of the magnetization, $d \langle |m| \rangle/dK$, $d$ln$\langle |m| \rangle
/dK$, and $d$ln$\langle m^2 \rangle/dK$. All these quantities exhibit in the infinite-volume limit singularities at $K_c$ which are shifted and rounded in finite systems. We further analyzed the Binder parameter, $$U_L(K)=1-\frac{1}{3}\frac{\langle m^4 \rangle}{\langle m^2\rangle^2}~.$$ It is well known that the $U_L(K)$ curves for different lattice sizes $L$ cross around $(K_c,U^*)$. This allows an almost unbiased estimate of the critical spin coupling $K_c$.
Simulation results {#sect3}
==================
In four dimensions, after initial discussions [@gaunt79; @gaunt80; @velasco] it is generally accepted that the critical properties of the Ising model on a static lattice are given by mean-field theory with logarithmic corrections. The finite-size scaling (FSS) formulas can be written as [@ralph; @parisi] $$\begin{aligned}
\xi &\propto& L(\log L)^{\frac{1}{4}}, \\
\chi &\propto& L^2 (\log L)^{\frac{1}{2}} \quad = (L (\log L)^{\frac{1}{4}})^{\gamma/\nu},\\
C &\propto& (\log L)^{\frac{1}{3}}, \\
K_c(\infty)-K_c(L) &\propto& L^{-2} (\log L)^{-\frac{1}{6}} = (L (\log L)^{\frac{1}{12}})^{-1/
\nu}, \label{e:4d_kc}\end{aligned}$$ where the critical exponents of mean-field theory are $\alpha=0$, $\beta=1/2$, $\gamma=1$, and $\nu=1/2$. To get more precise ideas about these logarithmic corrections, we first simulated the four-dimensional Ising model on a regular lattice. After this comparative study we turned to the four-dimensional discrete Regge model [@homo] with spin fields.
Ising spins on a regular lattice
--------------------------------
We studied the four-dimensional Ising model on a hypercubic lattice with linear size $L= 3 - 16$, 18, 20, 24, 28, 32, 36, 40, using the single-cluster update algorithm (Wolff) [@wolff]. The simulations were performed with the knowledge of the value of the critical temperature obtained in previous Monte Carlo simulations and high-temperature series analyses [@staufad]: $$K_c=\frac{J}{k_B T_c}=0.149\,694\pm0.000\,002.$$
We performed $n(L) \propto N_0/\langle C(L)\rangle$ cluster updates between measurements for lattices $L\le 24$, with the averaged cluster size $\langle C(L)\rangle$. After an initial equilibration time of about $1\,000 \times n(L)$ cluster updates we took about $50\,000$ measurements for each of the small lattices. For the larger lattices we measured after each cluster update, therefore, we took about $500\,000$ measurements after an initial equilibration of $100\,000$ cluster updates. Analyzing the time series we found integrated autocorrelation times for the energy and the magnetization in the range of unity for the small lattices $L\le 24$ and in the range of $(4-8) \times L$ for the larger lattice sizes. The statistical errors were obtained by the standard jack-knife method using 50 blocks.
Applying the reweighting technique we first determined the maxima of $C$, $\chi$, $d\langle |m| \rangle /dK$, $d$ln$\langle |m| \rangle /dK$, and $d$ln$\langle m^2 \rangle/dK$. The locations of the maxima provide us with five sequences of pseudo-transition points $K_{\rm max}(L)$ for which the scaling variable $x=( K_c - K_{\rm max}(L)) (L (\log L)^\frac{1}{12})^{1/\nu}$ should be constant. Using this fact we then have several possibilities to extract the critical exponent $\nu$ from (linear) least-square fits of the FSS ansatz with multiplicative logarithmic corrections considering Eq. (\[e:4d\_kc\]), $$\begin{aligned}
dU_L/dK &\cong& (L (\log L)^\frac{1}{12})^{1/ \nu} f_0(x) \quad {\rm or}\label{e:4d_duldk}\\
d{\rm ln}\langle |m|^p\rangle /dK &\cong& (L (\log L)^\frac{1}{12})^{1/\nu} f_p(x),\end{aligned}$$ to the data at the various $K_{\rm max}(L)$ sequences. For comparison we also performed fits of a naive power-law FSS ansatz $$\begin{aligned}
dU_L/dK &\cong& L^{1/ \nu} f_0(x) \quad {\rm or}\\
d{\rm ln}\langle |m|^p\rangle /dK &\cong& L^{1/\nu} f_p(x).\label{e:4d_dlnmdk}\end{aligned}$$ The exponents $1/\nu$ resulting from fits using the data for the $N$ largest lattice sizes are collected in Tables I and II. $Q$ denotes the standard goodness-of-fit parameter. For all exponent estimates the FSS ansatz with the logarithmic corrections leads to the weighted average $1/\nu = 1.993(3)$, which is in perfect agreement with the mean-field value $1/\nu=2$, see . With the naive power-law ansatz one also gets an estimate for $1/\nu$ in the vicinity of the mean-field value, but this is clearly separated from the mean-field result, verifying the significance of the multiplicative logarithmic correction, cf. Table II.
Assuming therefore $\nu=0.5$ we can obtain estimates for $K_c$ from linear least-square fits to the scaling behavior of the various $K_{\rm max}$ sequences, as shown in . Using the fits with $L \ge 12$, the combined estimate from the five sequences leads to $K_c=0.149\,697(2)$, which is in agreement with the results using Monte Carlo computer simulations [@staufad] and series expansions [@gaunt; @adstauf].
Knowing the critical coupling we may reconfirm our estimates of $1/\nu$ by evaluating the above quantities at $K_c$. As can be seen in Tables I and II, the statistical errors of the FSS fits at $K_c$ are similar to those using the $K_{\rm max}$ sequences. However, the uncertainty in the estimate of $K_c$ has also to be taken into account. This error is computed by repeating the fits at $K_c \pm \Delta K_c$ and indicated in Tables I and II by the numbers in square brackets. In the computation of the weighted average we assume the two types of errors to be independent. As a result of this combined analysis we obtain strong evidence that the exponent $\nu$ agrees with the mean-field value of $\nu=0.5$.
To extract the critical exponent ratio $\gamma/\nu$ we use the scaling $$\chi_{\rm max} \cong (L (\log L)^{\frac{1}{4}})^{\gamma/\nu}
\label{4d:sus_fit}$$ as well as the scaling of $\chi$ at $K_c$, yielding in the range $L=10 - 40$ estimates of $\gamma/\nu=2.037(9)$ with $Q=0.95$ and $\gamma/\nu=2.008(5)[5]$ ($Q=0.46$), respectively. These estimates for $\gamma/\nu$ are consistent with the mean-field value of $\gamma/\nu=2$. In this is demonstrated graphically by comparing the scaling of $\chi_{\rm max}$ with a constrained one-parameter fit of the form $\chi_{\rm max} = c (L (\log L)^{\frac{1}{4}})^{2}$ with $c=0.526(2)$ ($Q=0.38$, $L \ge 6$).
Concerning the specific heat we expect in the case of the mean-field exponent $\alpha=0$ a logarithmic divergence of the form $$C(x,L)=A(x)+B(x)(\log L)^{\frac{1}{3}}.
\label{eq:C_log_4d}$$ Indeed, the data at the different fixed values of $x$ can all be fitted nicely with this ansatz. In particular, for the fit of $C_{\rm max}$ with 17 points ($L \ge 6$) we obtain $A=-0.324(32)$, $B=1.038(23)$, with a total $\chi^2=11.7 (Q=0.70)$. We also tried an unbiased three-parameter fit using the ansatz $$C(x,L)=A'(x)+B'(x)(\log L)^{\kappa(x)},
\label{eq:C_3para_4d}$$ which in the case of the fit of $C_{\rm max}$ and 16 data points gives $A' \approx -0.36$, $B' \approx 1.75$, and $\kappa=0.33(40)$, with a slightly improved total $\chi^2=10.8 (Q=0.62)$. In we compare these two linear least-square fits. It should be noted, however, that the three-parameter fit is highly unstable and exhibits strong correlations between the three parameters. To illustrate this instability we plot in the exponent $\kappa$ as a function of the smallest lattice size $L_{\rm{min}}$, being the lower bound of the fit range $[L_{\rm{min}},40]$. For comparison we show in the behavior of $\gamma/\nu$ results of the fit corresponding to Eq. (\[4d:sus\_fit\]).
------------------------------------------------------------------------ ---- ---------------- ------
$dU/dK$ at $K^C_{\rm max}$ 22 1.990(5) 0.68
$d$ln$\langle |m|\rangle /dK$ at $K^{\ln\langle |m|\rangle}_{\rm inf}$ 18 1.989(4) 0.94
$d$ln$\langle m^2\rangle/dK$ at $K^{\ln\langle m^2\rangle}_{\rm inf}$ 18 1.998(5) 0.93
weighted average 1.993(3)
$dU/dK$ at $K_c$ 18 1.991(5)\[1\] 0.73
$d$ln$\langle |m|\rangle /dK$ at $K_c$ 18 1.992(5)\[2\] 0.94
$d$ln$\langle m^2\rangle/dK$ at $K_c$ 18 2.000(5)\[2\] 0.94
weighted average 1.995(3)
overall average 1.994(2)
------------------------------------------------------------------------ ---- ---------------- ------
Table I. Fit results for $1/\nu$ with a power-law ansatz with logarithmic corrections, using the data for the $N$ largest lattices. The average is computed by weighting each entry with its inverse squared error. For the fits at our best estimate of $K_c=0.149\,697(2)$ the uncertainty due to the error in $K_c$ is indicated by the numbers in square brackets.
------------------------------------------------------------------------ ---- ---------------- ------
$dU/dK$ at $K^C_{\rm max}$ 18 2.041(9) 0.81
$d$ln$\langle |m|\rangle /dK$ at $K^{\ln\langle |m|\rangle}_{\rm inf}$ 18 2.056(5) 0.75
$d$ln$\langle m^2\rangle/dK$ at $K^{\ln\langle m^2\rangle}_{\rm inf}$ 18 2.066(5) 0.63
weighted average 2.059(3)
$dU/dK$ at $K_c$ 18 2.043(9)\[2\] 0.81
$d$ln$\langle |m|\rangle /dK$ at $K_c$ 18 2.059(5)\[2\] 0.81
$d$ln$\langle m^2\rangle/dK$ at $K_c$ 18 2.067(5)\[2\] 0.70
weighted average 2.061(4)
overall average 2.060(2)
------------------------------------------------------------------------ ---- ---------------- ------
Table II. Fit results for $1/\nu$ with a pure power-law ansatz using $K_c=0.149\,697(2)$. The averages and statistical errors are computed as in Table I.
\[4ddU\]
\[fig\_sus4di\]
\[fig\_c\_E4di\]
Ising spins on a discrete Regge model
-------------------------------------
We simulated the gravitational degrees of freedom (the squared link lengths) of the partition function (\[z\]) using the heat-bath algorithm. For the Ising spins we employed again the single-cluster algorithm. Between measurements we performed $n=10$ Monte Carlo steps consisting of one lattice sweep to update the squared link lengths $q_{ij}$ followed by two single-cluster flips to update the spins $s_i$.
The simulations were done for $\epsilon=0.0875$, cosmological constant $\lambda = 0$ and two different gravitational couplings, $\beta_g=-4.665$ and $\beta_g=22.3$. These two $\beta_g$-values correspond to the two phase transitions of the pure discrete Regge model [@homo], as shown in Fig. \[4dz2\_q\]. The transition at positive gravitational coupling of the standard Regge calculus was described in great detail in Ref. [@hamber00] and shown to be of first order whereas the transition at negative coupling turned out to be of second order for the discrete Regge model [@homo]. Together with an eventual second-order transition of the Ising part, the latter one is a candidate for a possible continuum limit. The lattice topology is given by triangulated tori of size $N_0=L^4$ with $L = 3$ up to 10. From short test runs we estimated the location of the phase transition of the spin model and set the spin coupling $K_0 = 0.024 \approx K_c$ in the long runs for both values of $\beta_g$, see Fig. \[4dU\_short\].
After an initial equilibration time we took about $100\,000$ measurements for each lattice size. Analyzing the time series we found integrated autocorrelation times for the energy and the magnetization in the range of unity for all considered lattice sizes. As in the simulations of the regular lattices the statistical errors were obtained by the standard jack-knife method using 50 blocks.
Completely analogously to the Ising system on a regular lattice we applied reweighting to locate the maxima and used the FSS formulas (\[e:4d\_duldk\])–(\[e:4d\_dlnmdk\]). The exponents $1/\nu$ resulting from fits using the data for the $N$ largest lattice sizes are collected in Tables III and IV for $\beta_g=-4.665$, and in Tables V and VI for $\beta_g=22.3$, respectively. For the simulations at $\beta_g=-4.665$ all exponent estimates with the logarithmic corrections and consequently also their weighted average $1/\nu = 2.025(6)$ are in agreement with the mean-field value $1/\nu=2$, see . For $\beta_g=22.3$ the scatter in the estimates is similar and the weighted average $1/\nu = 2.028(6)$ is again compatible with $1/\nu = 2$. With the naive power-law ansatz one also gets an estimate for $1/\nu$ in the vicinity of the mean-field value, but this is clearly separated from the mean-field result, cf. Tables IV and VI.
Assuming therefore $\nu=0.5$ we can obtain estimates for $K_c$ from linear least-square fits to the scaling behavior of the various $K_{\rm max}$ sequences, as shown in hbox[Fig. \[kc4diz2\] (b)]{} for $\beta_g=-4.665$. Using the fits with $L \ge 4$, the combined estimate from the five sequences leads to $K_c=0.02464(4)$ for $\beta_g=-4.665$, and for $\beta_g=22.3$ we find $K_c=0.02339(4)$, again with $L \ge 4$.
With the knowledge of the critical couplings we may reconfirm our estimates of $1/\nu$ by evaluating the above quantities at $K_c$. As can be inspected in Tables III and V, we obtain from this combined analysis strong evidence that the exponent $\nu$ agrees with the mean-field value of $\nu=0.5$.
To extract the critical exponent ratio $\gamma/\nu$ we use the FSS formula (\[4d:sus\_fit\]) for $\chi_{\rm max}$ as well as the scaling of $\chi$ at $K_c$, yielding for $\beta_g=-4.665$ in the range $L=4 - 10$ estimates of $\gamma/\nu=2.039(9)$ with $Q=0.42$ and $\gamma/\nu=2.036(7)[4]$ ($Q=0.85$), respectively. The corresponding values for $\beta_g=22.3$, using the same fit range, are $\gamma/\nu=2.052(8)$ ($Q=0.57$) and $\gamma/\nu= 2.052(6)[4]$ ($Q=0.01$). These estimates for $\gamma/\nu$ are compatible with the mean-field value of $\gamma/\nu=2$. In this is demonstrated by comparing the scaling of $\chi_{\rm max}$ with a constrained one-parameter fit of the form $\chi_{\rm max} = c (L (\log L)^{\frac{1}{4}})^{2}$ with $c=4.006(10)$ ($Q=0.17$, $L \ge 6$) for $\beta_g=-4.665$ and $c= 4.244(10)$ ($Q=0.001$, $L \ge 6$) for $\beta_g=22.3$, respectively.
The data for the specific heat $C$ at the critical spin coupling $K_c$ are presented in . The fact that $C$ increases very slowly with the size of the lattice means that one will need data from bigger lattices and more statistical accuracy to get an estimate or bound for the critical exponent $\alpha$ from a direct measurement of $C$. Especially, if we assume a logarithmic divergence of $C$ as in the four-dimensional Ising model on regular lattices, we need lattices of comparable size, cf. .
\[4ddUz2\]
------------------------------------------------------------------------ --- ------------------ ------
$dU/dK$ at $K^C_{\rm max}$ 8 2.003(10) 0.47
$d$ln$\langle |m|\rangle /dK$ at $K^{\ln\langle |m|\rangle}_{\rm inf}$ 7 2.032(10) 0.59
$d$ln$\langle m^2\rangle/dK$ at $K^{\ln\langle m^2\rangle}_{\rm inf}$ 7 2.038(10) 0.55
weighted average 2.025(6)
$dU/dK$ at $K_c$ 7 1.981(17)\[13\] 0.70
$d$ln$\langle |m|\rangle /dK$ at $K_c$ 7 2.027(9)\[2\] 0.95
$d$ln$\langle m^2\rangle/dK$ at $K_c$ 7 2.034(9)\[2\] 0.85
weighted average 2.025(6)
overall average 2.025(4)
------------------------------------------------------------------------ --- ------------------ ------
Table III. Fit results for $1/\nu$ with a power-law ansatz with logarithmic corrections for $\beta_g=-4.665$, using the data for the $N$ largest lattices. The average is computed by weighting each entry with its inverse squared error. For the fits at our best estimate of $K_c=0.02464(4)$ the uncertainty due to the error in $K_c$ is indicated by the numbers in square brackets.
------------------------------------------------------------------------ --- ------------------ ------
$dU/dK$ at $K^C_{\rm max}$ 7 2.068(18) 0.60
$d$ln$\langle |m|\rangle /dK$ at $K^{\ln\langle |m|\rangle}_{\rm inf}$ 7 2.122(10) 0.37
$d$ln$\langle m^2\rangle/dK$ at $K^{\ln\langle m^2\rangle}_{\rm inf}$ 7 2.128(10) 0.35
weighted average 2.118(7)
$dU/dK$ at $K_c$ 7 2.068(18)\[12\] 0.59
$d$ln$\langle |m|\rangle /dK$ at $K_c$ 7 2.116(9)\[2\] 0.83
$d$ln$\langle m^2\rangle/dK$ at $K_c$ 7 2.124(9)\[2\] 0.64
weighted average 2.116(7)
overall average 2.117(5)
------------------------------------------------------------------------ --- ------------------ ------
Table IV. Fit results for $1/\nu$ with a pure power-law ansatz for $\beta_g=-4.665$. The averages and statistical errors are computed as in Table III.
------------------------------------------------------------------------ --- ----------------- ------
$dU/dK$ at $K^C_{\rm max}$ 8 1.981(10) 0.64
$d$ln$\langle |m|\rangle /dK$ at $K^{\ln\langle |m|\rangle}_{\rm inf}$ 7 2.043(9) 0.61
$d$ln$\langle m^2\rangle/dK$ at $K^{\ln\langle m^2\rangle}_{\rm inf}$ 7 2.049(9) 0.67
weighted average 2.028(6)
$dU/dK$ at $K_c$ 8 1.993(10)\[1\] 0.76
$d$ln$\langle |m|\rangle /dK$ at $K_c$ 7 2.039(9)\[2\] 0.32
$d$ln$\langle m^2\rangle/dK$ at $K_c$ 7 2.045(9)\[2\] 0.49
weighted average 2.027(6)
overall average 2.028(4)
------------------------------------------------------------------------ --- ----------------- ------
Table V. Fit results for $1/\nu$ with a power-law ansatz with logarithmic corrections for $\beta_g=22.3$. The average is computed by weighting each entry with its inverse squared error. For the fits at our best estimate of $K_c=0.02339(4)$ the uncertainty due to the error in $K_c$ is indicated by the numbers in square brackets.
------------------------------------------------------------------------ --- ----------------- ------
$dU/dK$ at $K^C_{\rm max}$ 7 2.086(11) 0.72
$d$ln$\langle |m|\rangle /dK$ at $K^{\ln\langle |m|\rangle}_{\rm inf}$ 7 2.134(10) 0.54
$d$ln$\langle m^2\rangle/dK$ at $K^{\ln\langle m^2\rangle}_{\rm inf}$ 7 2.139(9) 0.59
weighted average 2.122(6)
$dU/dK$ at $K_c$ 8 2.098(10)\[1\] 0.57
$d$ln$\langle |m|\rangle /dK$ at $K_c$ 7 2.130(9)\[2\] 0.35
$d$ln$\langle m^2\rangle/dK$ at $K_c$ 7 2.135(9)\[2\] 0.48
weighted average 2.122(6)
overall average 2.122(4)
------------------------------------------------------------------------ --- ----------------- ------
Table VI. Fit results for $1/\nu$ with a pure power-law ansatz for $\beta_g=22.3$. The averages and statistical errors are computed as in Table V.
\[fig\_sus4diz2\]
Conclusions {#sect4}
===========
We have performed a study of the Ising model coupled to fluctuating manifolds via Regge calculus. Analyzing the discrete Regge model with two permissible edge lengths it turns out that the Ising transition shows the predicted logarithmic corrections to the mean-field theory. The critical exponents of the phase transition of the Ising spins on a static lattice as well as on a discrete Regge skeleton [@berlin] are consistent with the exponents of the mean-field theory, $\alpha=0$, $\beta=\frac{1}{2}$, $\gamma=1$, and $\nu=\frac{1}{2}$. In summary, our consistent analysis with uniform computer codes yields that the phase transition of a spin system coupled to a discrete Regge skeleton exhibits the same critical exponents and the same logarithmic corrections [@parisi] as on a static lattice.
Acknowledgments {#acknowledgments .unnumbered}
===============
E.B. was supported by Fonds zur Förderung der wissenschaftlichen Forschung under project P14435-TPH and thanks the Graduiertenkolleg “Quantenfeldtheorie: Mathematische Struktur und Anwendungen in der Elementarteilchen- und Festkörperphysik” for hospitality during his extended stay in Leipzig. W.J. acknowledges partial support by the EC IHP Network grant HPRN-CT-1999-00161: “EUROGRID”.
1.2ex
[99]{} H.W. Hamber, in Proceedings of the 1984 Les Houches Summer School, edited by K. Osterwalder and R. Stora, Session XLIII (North-Holland, Amsterdam, 1986); H.W. Hamber and R.M. Williams, Phys. Lett. B [**157**]{}, 368 (1985).
B.A. Berg, in [*Particle Physics and Astrophysics*]{}, Proceedings of the XXVII Int. Universitätswochen für Kernphysik, edited by H. Mitter and F. Widder (Springer, Berlin, 1989); Phys. Rev. Lett. [**55**]{}, 904 (1985); Phys. Lett. B [**176**]{}, 39 (1986). W. Beirl, E. Gerstenmayer, H. Markum, and J. Riedler, Phys. Rev. D [**49**]{}, 5231 (1994). W. Beirl, H. Markum, and J. Riedler, Int. J. Mod. Phys. C [**5**]{}, 359 (1994) and hep-lat/9312054. T. Fleming, M. Gross, and R. Renken, Phys. Rev. D [**50**]{}, 7363 (1994). W. Beirl, A. Hauke, P. Homolka, B. Krishnan, H. Kröger, H. Markum, and J. Riedler, Nucl. Phys. B (Proc. Suppl.) [**47**]{}, 625 (1996); W. Beirl, A. Hauke, P. Homolka, H. Markum, and J. Riedler, Nucl. Phys. B (Proc. Suppl.) [**53**]{}, 735 (1997); J. Riedler, W. Beirl, E. Bittner, A. Hauke, P. Homolka, and H. Markum, Class. Quantum Grav. [**16**]{}, 1163 (1999). E. Bittner, H. Markum, and J. Riedler, Nucl. Phys. B (Proc. Suppl.) [**73**]{}, 789 (1999). M. Gross and H.W. Hamber, Nucl. Phys. B [**364**]{}, 703 (1991). C. Holm and W. Janke, Phys. Lett. B [**335**]{}, 143 (1994). E. Bittner, W. Janke, H. Markum, and J. Riedler, Physica A [**277**]{}, 204 (2000). A.M. Ferrenberg and R.H. Swendsen, Phys. Rev. Lett. [**61**]{}, 2635 (1988). D.S. Gaunt, M.F. Sykes, and S. McKenzie, J. Phys. A [**12**]{}, 871 (1979). S. McKenzie and D.S. Gaunt, J. Phys. A [**13**]{}, 1015 (1980). E. Sánchez-Velasco, J. Phys. A [**20**]{}, 5033 (1987). R. Kenna and C.B. Lang, Phys. Lett. B [**264**]{}, 396 (1991); Nucl. Phys. B [**393**]{}, 461 (1993). H.G. Ballesteros, L.A. Fernández, V. Martín-Mayor, A. Muñoz Sudupe, G. Parisi, and J.J. Ruiz-Lorenzo, Nucl. Phys. B [**512**]{}, 681 (1998). U. Wolff, Phys. Rev. Lett. [**62**]{}, 361 (1989); Nucl. Phys. B [**322**]{}, 759 (1989). D. Stauffer and J. Adler, Int. J. Mod. Phys. C [**8**]{}, 263 (1997). D.S. Gaunt, M.F. Sykes, and S. McKenzie, J. Phys. A [**12**]{}, L339 (1976). J. Adler and D. Stauffer, Int. J. Mod. Phys. C [**6**]{}, 807 (1995). H.W. Hamber, Phys. Rev. D [**61**]{}, 124008 (2000). E. Bittner, W. Janke, and H. Markum, Nucl. Phys. B (Proc. Suppl.) [**106-107**]{}, 989 (2002).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a chiral random matrix model which properly incorporates the flavor-number dependence of the phase transition owing to the [U$_{\rm A}$]{}(1) anomaly term. At finite temperature, the model shows the second-order phase transition with mean-field critical exponents for two massless flavors, while in the case of three massless flavors the transition turns out to be of the first order. The topological susceptibility satisfies the anomalous [U$_{\rm A}$]{}(1) Ward identity and decreases gradually with the temperature increased.'
address: |
$^a$Institute of Physics, The University of Tokyo, Tokyo 153-8902, Japan\
$^b$Department of Physics, The University of Tokyo, Tokyo 113-0033, Japan\
$^c$Physics Department, School of Medicine, Kyorin University, Tokyo 181-8611, Japan
author:
- 'T. Sano$^{a,b}$, H. Fujii$^a$ and M. Ohtani$^c$'
title: ' [U$_{\rm A}$]{}(1) breaking and phase transition in chiral random matrix model'
---
Introduction
============
Chiral symmetry breaking manifests itself in the accumulation of the Dirac modes with zero eigenvalues through the Banks-Casher relation[@BanksC1979]. In the chiral random matrix (ChRM) theory, the Dirac operator is restricted to the space of the constant modes and replaced with a matrix of random entities, retaining the global symmetries of QCD. The ChRM theory has been successful for providing a universal framework to investigate the correlation properties of the low-lying Dirac eigenvalues in the so-called epsilon regime in QCD[@review; @ChRMT]. The ChRM theory can be also regarded as the simplest schematic model for qualitative study of the QCD-like phase diagram from the viewpoint of chiral symmetry[@Halasz:1998qr]. Effects of the external fields such as the temperature $T$[@Jackson:1995nf; @Wettig:1995fg], the quark chemical potential $\mu$ [@Stephanov:1996ki; @Halasz:1997he; @Halasz:1998qr; @Stephanov:2006dn; @Han:2008xj] and the Polyakov loop[@Stephanov:1996he] have been investigated within the ChRM models.
The conventional ChRM model predicts a second-order phase transition at finite temperature $T$[@Jackson:1995nf; @Wettig:1995fg], irrespective of the number of flavors $N_f$. This is a shortcoming as a model of QCD. Based on the universality argument[@Pisarski:1983ms], the chiral transition for $N_f=2$ with the [U$_{\rm A}$]{}(1) anomaly is expected to be of the second order with the O(4) critical exponents, while the transition becomes of the first order for $N_f \ge 3$. Even in a mean-field description, the Ginzburg-Landau effective potential for $N_f =3$ involves the [U$_{\rm A}$]{}(1) breaking determinant term[@KM; @tHooft], which gives rise to a first-order transition. However, it is unknown so far how to incorporate the [U$_{\rm A}$]{}(1)-breaking determinant term into the ChRM models.
The explicit [U$_{\rm A}$]{}(1) breaking, or the anomaly, is included in the ChRM models by treating $\nu$ exact zero modes with definite chirality, which should be interpreted as the zero modes associated with the gauge field configurations of the topological charge $\nu$. The fluctuation of $\nu$ is effectively modeled with the instanton ensemble. The ChRM model supplemented with a Gaussian distribution in $\nu$ has been shown to provide a screening of the topological charge fluctuation at zero temperature, and thus a resolution of the [U$_{\rm A}$]{}(1) problem[@JanikNPZ97]. At finite temperatures, however, this ChRM model results in an unphysical suppression of the topological susceptibility[@OhtaniLWH08; @LehnerOVW09]. The resolution of the [U$_{\rm A}$]{}(1) problem in the ChRM model is then limited only at zero temperature. A modification of the model has been proposed in[@OhtaniLWH08; @LehnerOVW09] to remove the unphysical suppression of the topological susceptibility at finite temperature.
The ground state is fixed exactly with the saddle point condition in the thermodynamic limit. Because the topological charge $\nu$ fluctuates around zero with the variance of ${\mathcal O}(V)$, as dictated by the central limit theorem, the nonzero $\nu$ configurations are irrelevant to the ground state condition in the ChRM model, and therefore the [U$_{\rm A}$]{}(1) breaking seemingly cannot alter the order of the phase transition at finite temperature.
In this paper we propose a ChRM model involving the [U$_{\rm A}$]{}(1) breaking effect, which describes a first-order phase transition for $N_f=3$ and at the same time reproduces the physical behavior for the topological susceptibility. The proposed model consists of the “near-zero” modes and the “topological zero” modes following Refs. [@JanikNZ97; @Jan2]. The latters are interpreted as the zero modes with the right and left chiralities, respectively, associated with instantons and anti-instantons. We introduce number distributions of the topological zero modes based on the instanton gas picture. Summing over the number of the topological zero modes in the partition function, we can derive the [U$_{\rm A}$]{}(1) breaking interaction. Furthermore, this model satisfies the anomalous [U$_{\rm A}$]{}(1) Ward identity[@Crewther77][@JanikNPZ97] and predicts physical temperature dependence for the topological susceptibility[@OhtaniLWH08; @LehnerOVW09].
Interplay between the topological zero modes and the near-zero modes is studied in Ref. [@Nowak:1989gw] to derive the instanton-induced interactions in the coarse-grained instanton-liquid model. The ’topological’ zero modes are also introduced in Ref. [@JanikNZ97] from the determinant interaction in the Nambu–Jona-Lasinio (NJL) model in 0+1 dimensions, but the potential there is unbound in the $N_f=3$ case.
This paper is organized as follows. After reviewing the conventional ChRM model in the next section, we introduce an extended ChRM model which involves two kinds of zero modes, the “near-zero” modes and the “topological zero” modes, motivated by the instanton gas picture in subsec. \[sec:IIa\]. Unlike in the conventional ChRM model, the total number of the topological zero modes is assumed to vary according to the instanton distribution. In subsec. \[sec:IIb\] we propose the binomial distribution for the instanton numbers in a finite space-time volume, and show that after summing up the topological zero modes the resulting effective potential has the [U$_{\rm A}$]{}(1) breaking term and bounded from the below. General properties of the ground state and fluctuations of the model are presented for a certain set of parameters and equal quark masses in sec. \[sec:III\]. Section \[sec:IV\] is devoted to discussions and summary.
Chiral Random Matrix Model {#sec:II}
==========================
The QCD partition function with $N_f$ quark flavors of mass $m_f$ is written as $$\begin{aligned}
Z_\theta^{\rm QCD}=
\sum_{\nu =-\infty}^{\infty} e^{i\nu \theta} Z_\nu^{\rm QCD}
=
\sum_{\nu=-\infty}^{\infty}
e^{i\nu \theta} \left\langle \prod_{f=1}^{N_f}\det(D+m_f) \right\rangle_\nu
\; ,\end{aligned}$$ where $\langle \cdots \rangle_\nu$ denotes the average over the gauge field configurations of fixed topological charge $\nu$, and the $\theta$ parameter has been introduced.
The ChRM model is defined for fixed $\nu$ conventionally as[@Jackson:1995nf] $$\begin{aligned}
Z_\nu = \int dW e^{-N\Sigma^2 {{\rm tr}}W^\dagger W} \prod_{f=1}^{N_f} \det(D+m_f)
\; ,\end{aligned}$$ where the Dirac operator has been replaced with an anti-hermite matrix of constant modes $$\begin{aligned}
D=\left ( \begin{array}{cc} 0 & iW+it{\mathbf{1}_{N-|\nu|/2}}
\\ iW^\dagger+it{\mathbf{1}_{N-|\nu|/2}} & 0 \end{array}
\right )\end{aligned}$$ with an $(N+\nu/2) \times (N-\nu/2)$ random complex matrix $W$ in the chiral representation $\gamma_5 ={\rm diag}(1_{N+\nu/2},-1_{N-\nu/2})$. Here we have introduced the effective temperature $t$ as a deterministic part in the Dirac operator. It may be interpreted as the lowest Matsubara frequency $t = \pi T $. It is readily shown that the matrix $D$ has $|\nu|$ exact zero eigenvalues with definite chirality, which are interpreted as the exact zero modes associated with the topological number $\nu$. The total number of modes $2N$ is finite and proportional to the space-time volume $V$; $N/V={\cal O}(1)$.
After bosonization of this model we obtain the effective potential $\Omega_\nu(S)$ on the chiral manifold; $$\begin{aligned}
Z_\nu
=&\int dS e^{-N\Sigma^2\text{tr}(SS^\dagger)}
\det \left (
(S+m_f)(S^\dagger+m_f )+t^2
\right )^{N-\tfrac{|\nu|}{2}}
\times
\begin{cases}
\det(S+m_f)^{\nu} & (\nu \ge 0)\\
\det(S^\dagger+m_f)^{-\nu} & (\nu<0) \\
\end{cases}
{\nonumber}\\
\equiv
& \int dS e^{-2N\Omega_\nu(S)}
\; ,\end{aligned}$$ where $S$ is an $N_f \times N_f$ complex matrix of the chiral order parameter (see next subsection). The complete partition function is obtained after summing over $\nu$ $$\begin{aligned}
Z_\theta = \sum_{\nu=-2N}^{2N}
e^{-\frac{\nu^2}{2(2N)\tau}}e^{i\nu \theta}
Z_\nu
.
\label{eq:conv}\end{aligned}$$ Here we need to supplement the distribution of $\nu$ characterized by the quenched topological susceptibility $\tau$, which is determined by the pure gluonic dynamics. Note that the range of the topological charge $\nu$ is limited within $\pm 2N$.
In the thermodynamic limit $N\to \infty$, the ground state is exactly determined by the saddle point equation. Since the fluctuation of $\nu$ scales as $\nu^2 \sim N$ and nonzero $\nu$ contribution to the ground state becomes negligible, the ground state is given by the saddle point condition for $\Omega_\nu(S)$ with $\nu=0$, which is symmetric under U($N_f$)$\times$U($N_f$) for $m_f=0$. Furthermore, the $N_f$ dependence is factored out in the potential $\Omega_{0}(S)$ for $S \propto \mathbf{1}_{N_f}$ and $m_f=0$,$^{\rm \footnotemark[1]}$ yielding a second-order phase transition irrespective of $N_f$. Concerning the fluctuation properties, the model (\[eq:conv\]) at zero temperature gives a nonzero singlet pseudo-scalar susceptibility $\chi_{\rm ps0}$ in the chiral limit, resolving the [U$_{\rm A}$]{}(1) problem[@JanikNPZ97]. At finite temperature, however, this leads to an unphysical suppression of the topological susceptibility as mentioned in Introduction.
Model with near- and topological zero modes {#sec:IIa}
-------------------------------------------
Let us consider a variation of the conventional ChRM model with the instanton gas picture in mind. An isolated instanton is a localized object accompanying a right-handed fermion zero mode. In a dilute system of $N_+$ instantons and $N_-$ anti-instantons, we expect $N_+$ right-handed and $N_-$ left-handed zero modes (even at finite temperature). In an effective theory at long distances, effects of the instantons should be integrated out, which will result in [U$_{\rm A}$]{}(1)-breaking effective interactions[@tHooft]. The fundamental assumption in our modeling is the classification of the constant modes into the “near-zero” modes and “topological zero” modes[@JanikNPZ97]. We deal with the $2N$ near-zero modes appearing in the conventional models and include additionally the $N_+ + N_-$ topological zero modes which we regard as the modes accompanied by the instantons. Distributions of $N_+$ and $N_-$ lead to fluctuations of the topological charge $\nu=N_+ - N_-$ as well as the total number $N_+ + N_-$. Eventually we shall sum over $N_+$ and $N_-$ with the mean value of ${\mathcal O}(N)$ and take the thermodynamic limit $N \to \infty$.
For definite numbers of zero modes, we write a Gaussian ChRM model as[@JanikNZ97] $$\begin{aligned}
Z^N_{N_+,N_-}=&
\int dA dB dX dY
e^{-N \Sigma^2 {\rm tr}(AA^\dagger +BB^\dagger +XX^\dagger +YY^\dagger)}
\prod_{f=1}^{N_f} \det(D+m_f)
\label{eq:fixNpf}\end{aligned}$$ with $$\begin{aligned}
D=
\left(
\begin{array}{cccc}
0 &iA+it{\mathbf{1}_N} & 0 & iX \\
iA^\dagger+it{\mathbf{1}_N} & 0 & iY & 0 \\
0 & iY^\dagger & 0 & iB \\
iX^\dagger & 0 & iB^\dagger & 0
\end{array}
\right)
\; .\end{aligned}$$ Here we use the chiral representation in which $\gamma_5={\rm diag} (1_N, -1_N, 1_{N_+}, -1_{N_-})$. Then the matrix $D$, which satisfies $\{ D,\gamma_5 \}=0$, has a block structure with complex matrices $A$, $B$, $X$ and $Y$ as shown above. The $N \times N$ complex matrix $A$ acts on the near-zero modes, while the rectangular complex matrix $B$ of size $N_+ \times N_-$ acts on the topological zero modes. The matrices $X \in \mathbb{C}^{N \times N_-}$ and $Y \in \mathbb{C}^{N \times N_+}$ represent the interactions between the near-zero modes and the topological zero modes. In the following we take the quark mass $m_f$ of flavor $f$ to be diagonal $\propto
\mathbf{1}_{2N+N_+ + N_-}$ in the space of the zero modes. Notice that the temperature term $t$ is introduced only for the near-zero modes, while the topological zero modes are assumed to be insensitive to the temperature $t$. This discrimination is physically legitimate in the instanton gas picture where each topological zero mode is localized around an (anti-)instanton and its eigenvalue is not much affected by the anti-periodic boundary condition in the temporal direction unless the temporal size $1/T$ becomes the same order of the typical temporal extent of the instanton or the topological zero mode.$^{\rm \footnotemark[6]}$
The effective potential on the chiral manifold can be derived from eq. (\[eq:fixNpf\]) in a standard manner. First, we recast the partition function in the form of the integration over Grassmann variables of the near-zero modes $\psi$ and of the topological zero modes $\chi$: $$\begin{aligned}
\det(D+m_f)=
\int d\psi^\dagger d\psi d\chi^\dagger d\chi
\exp \left[ -
(\psi^\dagger_L, \psi^\dagger_R, \chi^\dagger_L, \chi^\dagger_R)
(D+m_f)
\left (
\begin{array}{c}
\psi_R\\
\psi_L\\
\chi_R\\
\chi_L
\end{array}
\right ) \right] ,\end{aligned}$$ where, for each flavor, $\psi^\dagger_{R,L}$ and $\psi_{R,L}$ have $N$ components while $\chi^\dagger_{L{(R)}}$ and $\chi_{R(L)}$ have $N_{+(-)}$ components. Then integration over the random matrices $A$, $B$, $X$ and $Y$ with the Gaussian distributions gives rise to the four-fermion interaction term. We unfold this term with the aid of the Hubbard-Stratonovich transformation as $$\begin{aligned}
&\exp \left[ \frac{1}{N\Sigma^2}
(\psi_L^{\dagger f} \psi_R^g + \chi_L^{\dagger f}\chi_R^g)
(\psi_R^{\dagger g}\psi_L^f + \chi_R^{\dagger g} \chi_L^f)
\right]
{\nonumber}\\
=& \int dS \exp \left[
-(\psi_L^{\dagger f} \psi_R^g + \chi_L^{\dagger f}\chi_R^g)S_{fg}
-(\psi_R^{\dagger f}\psi_L^g + \chi_R^{\dagger f} \chi_L^g)S^\dagger_{fg}
\right]
\exp \left [ -N\Sigma^2 {\rm tr} S^\dagger S \right ]\end{aligned}$$ with an ${N_f \times N_f}$ complex matrix $S$. Here summation over the flavor indices $f$ and $g=1,\cdots ,N_f$ should be understood, and the zero-mode sum is also implicit in (e.g.) $\psi_L^{\dagger f} \psi_R^g \equiv
\sum\limits_{i}\psi^{\dagger f}_{Li} \psi_{Ri}^g$. This transformation allows us to perform the Grassmann integration. We thus obtain the desired form for the partition function: $$\begin{aligned}
Z^N_{N_+,N_-}=&\int dS \; e^{-N\Sigma^2 \text{tr} S^\dagger S}
\; \det {}^N \left [ (S+{ \mathcal{M} })(S^\dagger+{ \mathcal{M} }^\dagger) +t^2 \right ]
\; \det {}^{N_+} (S+{ \mathcal{M} }) \; \det{}^{N_-} (S^\dagger + { \mathcal{M} }^\dagger)
.
\label{eq:fixedia}\end{aligned}$$ The mass matrix ${ \mathcal{M} }=\text{diag}(m_1,\cdots, m_{N_f})$ will be taken as a general $N_f \times N_f$ matrix of the source field later in sec. \[sec:III\].
In the chiral limit ${ \mathcal{M} }=0$, the integrand of (\[eq:fixedia\]) is invariant under $\text{SU}(N_f)\times \text{SU}(N_f)$ transformation, $S\to USV^{-1}$ with $U, V \in \text{SU}(N_f)$, in addition to U$_{\rm V}$(1). On the other hand, under [U$_{\rm A}$]{}(1) transformation, $S\to e^{2i\theta}S$, it acquires a phase factor $e^{2 i N_f \theta \nu}$ due to the difference $\nu \equiv N_+ - N_-$ in the powers of the determinants. In other words, the nonzero $\nu$ breaks the [U$_{\rm A}$]{}(1) symmetry explicitly down to $\Bbb Z_{2\! N_f}$.
Distributions of topological zero modes {#sec:IIb}
---------------------------------------
The complete partition function is obtained after summing $Z_{N_+,N_-}^N$ over the ‘instanton’ numbers $N_+$ and $N_-$ with an appropriate weight reflecting the pure gluon dynamics.$^{\rm \footnotemark[2]}$ Although the distributions of $N_+$ and $N_-$ will be correlated in general, here we assume independent distributions $P(N_\pm)$ for $N_+$ and $N_-$, [*i.e.*]{}, $$\begin{aligned}
Z_\theta =&
\sum_{N_+,N_-}
e^{i \nu \theta} P(N_+)P(N_-)Z_{N_+,N_-}^N
=\int dS e^{-2N \Omega(S;t,m,\theta)}
\; .
\label{eq:ia_sum}\end{aligned}$$
First we consider $P(N_\pm)$ in a dilute instanton gas picture. For one-instanton configuration, one may assign a weight $\kappa$ compared with a no-instanton configuration, and multiply a factor $N \propto V$ taking into account the integration over the instanton location. For a configuration with $N_+$ instantons, we have a weight $$\begin{aligned}
P_{\rm Po}(N_+)=\frac{1}{N_+ !} (\kappa N)^{N_+}\end{aligned}$$ where $N_+ !$ is the symmetry factor [@tHooft]. The same distribution is assumed for $N_-$ as well. This is nothing but the Poisson distribution up to a normalization. The summation with $P_{\rm Po}(N_\pm)$ in eq. (\[eq:ia\_sum\]) results in the exponentiation of the determinant term[@JanikNZ97]: $$\begin{aligned}
\Omega_{\text{Po}}=&
{\tfrac{1}{2}}\Sigma^2 {\rm tr}SS^\dagger
-{\tfrac{1}{2}}\ln \det \left [ (S+{ \mathcal{M} })(S^\dagger+{ \mathcal{M} }^\dagger) +t^2 \right ]
-{\tfrac{1}{2}}\kappa
[e^{i\theta} \det (S+{ \mathcal{M} }) + e^{-i\theta}\det (S^\dagger +{ \mathcal{M} }^\dagger)]
.
\label{eq:njlpf}\end{aligned}$$ The same determinant term as in $\Omega_{\text{Po}}$ is commonly incorporated in other effective models as the “anomaly term” to break the [U$_{\rm A}$]{}(1) symmetry.$^{\rm \footnotemark[3]}$ In the ChRM model, however, this potential is unbound for $N_f=3$. Indeed, for large $S=\phi {\bf 1}_{N_f}$ the term $\det(S+{ \mathcal{M} }) \sim \phi^3$ dominates over the other terms in $\Omega_{\text{Po}}$.
It should be noticed here that the fermion coupling distorts the $N_\pm$ distribution itself. With including the determinant term of the topological zero modes in eq. (\[eq:fixedia\]), the effective distribution for $N_+$ reads $$\begin{aligned}
\widetilde P_{\rm Po}(N_+)=\frac{1}{N_+ !}
\left (\kappa N d\right )^{N_+}\end{aligned}$$ with $d=|\det(S+{ \mathcal{M} })|$, and similarly for $N_-$. We find that the average value of $N_\pm$ increases indefinitely with increasing $d$ as $\left < N_\pm \right > = \kappa N d$. However, the possibility of infinitely many topological zero modes $N_\pm$ should be avoided in the ChRM model as a low-energy effective theory within a finite volume $V\propto N$.
Here we regularize the distribution by setting explicitly a maximum value of ${\cal O}(N)$ for $N_\pm$. We split the finite space-time volume into $\gamma N$ cells with $\gamma$ being a constant of ${\cal O}(1)$, and assign a probability $p$ for a cell to be occupied by a single ‘(anti-)instanton’ and $(1-p)$ for a cell unoccupied. We exclude the possibility of the double occupation of a cell, which is justified by a repulsion between the instantons. Just like in the lattice gas model in statistical mechanics, this assumption results in the binomial distributions for $N_\pm$: $$\begin{aligned}
P(N_\pm)=
\left (\begin{array}{c}
\gamma N \\ N_{\pm}
\end{array}
\right )
\; p^{N_\pm} (1-p)^{\gamma N-N_\pm}
.
\label{eq:bino}\end{aligned}$$ For a small $p$ and a large $\gamma N$, the binomial distribution $P(N_\pm)$ is accurately approximated with the Poisson distribution with the mean $\gamma N p$. But it cannot for a large $p$. The binomial distribution provides a stringent upper bound $\gamma N$ for the number of topological zero modes $N_\pm$, in contrast to the Poisson distribution. The corresponding effective potential for $S$ is found to be $$\begin{aligned}
\Omega(S;t,m,\theta)=&
{\tfrac{1}{2}}\Sigma^2 {\rm tr}SS^\dagger
-{\tfrac{1}{2}}\ln \det \left [ (S+{ \mathcal{M} })(S^\dagger+{ \mathcal{M} }^\dagger) +t^2 \right ]
{\nonumber}\\
&-{\tfrac{1}{2}}\gamma
\left [\ln \left(e^{i\theta} \alpha \det (S+{ \mathcal{M} })+1 \right)
+ \ln \left(e^{-i\theta} \alpha \det (S^\dagger +{ \mathcal{M} }^\dagger)+1\right)
\right ]
\label{eq:thepf}\end{aligned}$$ with $\alpha=p/(1-p)$. This is the ChRM model that we propose and analyze in this paper.
The effective potential (\[eq:thepf\]) is bounded from the below by the $\text{tr}SS^\dagger$ term in contrast to $\Omega_{\rm Po}$ in eq. (\[eq:njlpf\]). The anomaly terms are accommodated under the logarithms in the square bracket. Moreover, for small value of $|\alpha \det (S+m)|$, it reduces to the potential (\[eq:njlpf\]) with $\kappa=\gamma \alpha$. On the other hand, the Poisson approximation fails when $\left < N_\pm \right >$ becomes ${\cal O}(\gamma N)$. We stress here again the fact that the distributions $P(N_\pm)$ is deformed in the presence of the coupling with the topological zero modes as $$\begin{aligned}
\widetilde P(N_+) =
\left (\begin{array}{c}
\gamma N \\ N_{+}
\end{array}
\right )
(p d)^{N_+} (1-p)^{\gamma N-N_+}
,\end{aligned}$$ which changes the probability $p$ to an effective one $\tilde p \equiv p d/(p d+1-p)$. Similarly for $N_-$. Accordingly, the mean number of the zero modes is modified to $\langle \widetilde{N}_ \pm \rangle=\gamma N \tilde p$.
Two remarks are here in order. In the conventional models, the total number of the modes is kept fixed, while the fluctuation of $\nu$ is allowed for resolving the [U$_{\rm A}$]{}(1) problem. On the other hand, in the instanton gas picture, both of $N_+$ and $N_-$ are expected to fluctuate naturally. As stressed in this subsection, the mean number of the modes depends on the magnitude of $S$, which gives an ${\mathcal O}(N)$ effect. The anomaly term appearing in the effective potential in turn affects the saddle point value of $S$. This point is essential to the first-order transition for $N_f=3$, which is overlooked in the earlier works.
Secondly, in eq. (\[eq:thepf\]) one needs to introduce a dimensionful scale to compensate the dimension of determinant if the dimension of mass is assigned to ${ \mathcal{M} }$ as well as $S$. Here we leave all the quantities dimensionless for demonstration of general feature of the model.
$\theta$ dependence
-------------------
The variance of the topological charge $\nu = N_+ - N_-$ for the binomial distribution is computed as $2N \tau = 2N \gamma p (1-p)$, where $\tau$ is the quenched topological susceptibility. In the presence of the fermion coupling, this susceptibility will be replaced with $$\begin{aligned}
\tilde \tau=\gamma \tilde p (1-\tilde p)
=
\gamma \frac{\alpha d}
{(\alpha d+1 )^2}
\; .
\label{eq:tildetau}\end{aligned}$$ We can confirm this fact also by rewriting the anomaly term in eq. (\[eq:thepf\]) as $$-\frac{1}{2}\gamma\ln\left[1+|\alpha\det(S+{\cal M})|^2
+2|\alpha\det(S+{\cal M})|\cos\left(
\theta-\frac{i}{2} \ln
\frac{\text{det}(S+{ \mathcal{M} })}{\text{det}(S^\dagger + { \mathcal{M} }^\dagger)}
\right)\right]\ ,
\label{eq:largeNcform}$$ from which we find again the replaced topological susceptibility $\tilde\tau$ as a coefficient of $\theta^2$. Such a series in $\theta$ of the anomaly term gives a connection to the general arguments on the $\eta'$ in the $1/N_{\rm c}$ expansion [@Witten80].
To make a clear connection with the conventional model, let us ignore for a moment the fluctuation of the total number $N_+ + N_-$ and apply the Gaussian approximation for the $\nu$ distribution. We then obtain the partition function as $$\begin{aligned}
Z_\theta =& \int_{-\infty}^{\infty}d\nu e^{-\tfrac{\nu^2}{2(2N)\tau}}
\; e^{i\nu\theta} \; Z_{N_+,N_-}^N
{\nonumber}\\
=&
\text{const.}\int dS \; e^{-N\Sigma^2 \text{tr} S^\dagger S}
\; \det{}^N \left[ (S+{ \mathcal{M} })(S^\dagger+{ \mathcal{M} }^\dagger)+t^2 \right ]
{\nonumber}\\
& \qquad \times
\; \det{}^{(N_+ +N_-)/2}\left [(S+{ \mathcal{M} })(S^\dagger + { \mathcal{M} }^\dagger)\right ]
\exp \left [ -N\tau
\left (
\theta-\frac{i}{2} \ln
\frac{\text{det}(S+{ \mathcal{M} })}{\text{det}(S^\dagger + { \mathcal{M} }^\dagger)}
\right )^2
\right ]
\; .
\label{eq:GaussSum}\end{aligned}$$ This is almost identical with the model discussed in Ref. [@OhtaniLWH08; @LehnerOVW09], which reproduces the screening of the (unquenched) topological susceptibility (see eq. (\[eq:topsus2\])) as measured on a lattice at finite temperatures. However, this model (\[eq:GaussSum\]) fails to describe a first-order phase transition at finite temperature for $N_f$=3. In fact, the anomaly term appears only as a phase in (\[eq:GaussSum\]) in contrast to (\[eq:largeNcform\]) and drops out when we determine the magnitude of $S=S^\dagger$ in the ground state (with ${ \mathcal{M} }= { \mathcal{M} }^\dagger$). The variation of $N_+ + N_-$ is essential for the anomaly term to affect the saddle point condition, and thus the order of the phase transition.
Ground state and fluctuations {#sec:III}
=============================
In this section we shall study the ground state properties of the system with equal mass, ${ \mathcal{M} }=m \mathbf{1}_{N_f}$, for simplicity. Setting $S = \phi {\bf 1}_{N_f}$ with real $\phi$ and with $\theta=0$, we obtain a simple form of the grand potential: $$\begin{aligned}
\label{eq:gtpot}
\Omega (\phi; t,m)=&
{\tfrac{1}{2}}N_f\Sigma^2 \phi^2
- {\tfrac{1}{2}}N_f \ln \left[ (\phi + m)^2 + t^2 \right]
- \gamma \ln \left | \alpha (\phi + m)^{N_f} +1 \right |
\; .\end{aligned}$$ The factor $N_f$ cannot be factored out in the potential $\Omega$ because of the anomaly term here.
Chiral phase transition: ground state
-------------------------------------
In the thermodynamic limit, $N\to \infty$, the ground state can be analyzed with the solution of the saddle point equation $$\begin{aligned}
\Sigma^2 \phi -\frac{\phi+m}{(\phi+m)^2+t^2}
-\gamma \frac{\alpha (\phi+m)^{N_f-1}}{\alpha(\phi+m)^{N_f}+1} =0
\; .
\label{eq:saddleq}\end{aligned}$$ The scalar quark condensate is related to the solution $\phi$ as $$\begin{aligned}
\left<\bar{\psi}\psi \right>
= -\frac{1}{2N N_f}\frac{\partial}{\partial m} \ln Z(m)
= - \Sigma^2 \phi
.\end{aligned}$$ Without the anomaly term $\alpha=0$, eq. (\[eq:saddleq\]) recovers the flavor-independent gap equation of the conventional model, which has the solution $\phi_0^2=\Sigma^{-2}-t^2$ in the chiral limit. One can estimate the effect of small $m$ and $\alpha$ on $\phi^2$ at the leading order as $$\begin{aligned}
\phi^2 & \sim
\phi_0^2 -\frac{m}{\phi_0} (\Sigma^{-2}-2 t^2)
+\frac{\alpha \gamma}{\Sigma^4}\phi_0^{N_f-2}\; . \end{aligned}$$ At lower temperatures $2t^2< \Sigma^{-2}$ the value of the condensate is decreased by the quark mass term, on the contrary to our intuition. This is because the leading order term of $m$ appears in a combination $-2m\phi/(\phi^2 +t^2)$ in the potential, which simplifies to $-2m/\phi$ for $t=0$, favoring smaller $\phi$ for the potential to be more stabilized.
There is no chiral symmetry for $N_f=1$ because of the anomaly term. Let us discuss phenomenologically interesting cases, $N_f=$ 2 and 3.
### $N_f=2$
Expanding the potential $\Omega$ around $\phi=0$, we find $$\begin{aligned}
\Omega=c_0+c_2\phi^2 +c_4 \phi^4 -h \phi + {\cal O}(\phi^6,\phi^3 m,m^2)\end{aligned}$$ with $c_0=-\ln t^2$, $c_2=\Sigma^2-\alpha \gamma - t^{-2}$, $c_4=(\gamma\alpha^2+t^{-4})/2$ and $h=2(\alpha \gamma +t^{-2})m$. This is the standard form of the Landau-Ginzburg potential for a second order phase transition ($m=0$) with the critical temperature $t_c$ $$\begin{aligned}
t_c=\frac{1}{\sqrt{\Sigma^2-\alpha \gamma}}
\; .\end{aligned}$$ Inclusion of the anomaly term increases the value of $t_c$. The behavior in the vicinity of the critical point is characterized by the mean-field exponents; we find $\beta=1/2$ and $\delta=3$, respectively, from the solutions $\phi^2 = \epsilon
t_c^{-2}/c_4 \propto \epsilon^{2\beta}$ for $\epsilon=(t_c-t)/t_c$ with $0<\epsilon \ll 1$, and $\phi=[m\Sigma^2/(2c_4)]^{1/3}
\propto m^{1/\delta}$ with $m\ne 0$ at $t=t_c$.
### $N_f=3$
Expanding the potential $\Omega$ around $\phi=0$, we find $$\begin{aligned}
\Omega=c_0+c_2\phi^2 +c_3 \phi^3 +c_4 \phi^4 -h \phi +{\cal O}(\phi^6,\phi^2 m, m^2)\end{aligned}$$ with $c_0= - (3/2)\ln t^2$, $c_2=(3/2)(\Sigma^2 - t^{-2})$, $c_3= - \alpha \gamma$, $c_4=(3/4)t^{-4}$ and $h=3t^{-2} m$. The [U$_{\rm A}$]{}(1) symmetry is explicitly broken even for $m=0$ due to the anomalous $\phi^3$ term, which leads to a first-order phase transition.
In Fig. \[fig:cond\], we display the chiral condensate $\phi$ as a function of the temperature $t$ and the quark mass $m$. For numerical demonstration, we have chosen the parameters as $\Sigma=1$, $\gamma=2$ and $\alpha=0.3$ in this paper. We clearly see the second order transition for $N_f=2$, while the first order transition for $N_f=3$ in the chiral limit $m=0$. In $N_f=3$ case, as we increase the current quark mass $m$, we find a terminating point of the first order line at $m_c=0.0265$.
For $N_f \ge 4$, the anomaly term only affects the coefficients of $\phi^n$ ($n \ge N_f$) in the series expansion of the potential.
![Chiral condensate $\phi$ as a function of $t$ and $m$ for $N_f$=2 (left) and 3 (right). \[fig:cond\]](nf2_cc.eps "fig:"){height="8cm"} ![Chiral condensate $\phi$ as a function of $t$ and $m$ for $N_f$=2 (left) and 3 (right). \[fig:cond\]](nf3_cc.eps "fig:"){height="8cm"}
Mesonic masses
--------------
Once the ground state acquires the nonzero expectation value $ S = \phi \mathbf{1}_{N_f} \ne 0$, the symmetry is spontaneously broken down to U$_{\rm V}(N_f)$; $S \to U S U^{-1}$. Degeneracy of the vacua in the chiral limit dictates the massless fluctuations $S=\phi e^{i \pi^a \lambda^a/(\sqrt{2}\phi)} \sim
\phi + i \pi^a \lambda^a/\sqrt{2}$ corresponding to the generators of the broken symmetry.
To find the mesonic masses, we use the parametrization $S=\phi+\lambda^a (\sigma^a+i\pi^a)/\sqrt{2}$ with $\sigma^a, \pi^a \in \mathbb{R}$ and the U($N$) generators $\lambda^a$ normalized as $\text{tr}[\lambda^a \lambda^b]=2\delta^{ab}$ ($a=0, \cdots, N_f^2-1$). Using a formula for a matrix $X$ with a small parameter $\epsilon$, $\det (1 + \epsilon X)=
1 + \epsilon\text{tr}X
+\tfrac{1}{2}\epsilon^2 [(\text{tr}X)^2 -\text{tr}X^2]+{\cal O}(\epsilon^3)$, we can easily expand $\Omega$ (\[eq:thepf\]) around the saddle point solution to define the massses with $\Omega = \Omega_0 + \tfrac{1}{2}M_{{\rm s}a}^2 \sigma^{a2}
+ \frac{1}{2}M_{{\rm ps}a}^2 \pi^{a2} + \cdots$.
The flavor non-singlet masses in the scalar and pseudo-scalar channels, respectively, are found to be $$\begin{aligned}
M_{\rm s}^2 =&
\Sigma^2 - \frac{1}{(\phi+m)^2 +t ^2}
+ \frac{2(\phi+m)^2}{[(\phi+m)^2 +t ^2]^2}
+\gamma \frac{\alpha(\phi+m)^{N_f-2}}{\alpha(\phi+m)^{N_f}+1}
\; ,
\\
M_{\rm ps}^2 =&
\Sigma^2 - \frac{1}{(\phi+m)^2 +t ^2}
-\gamma \frac{\alpha(\phi+m)^{N_f-2}}{\alpha(\phi+m)^{N_f}+1}
\; .\end{aligned}$$ With the saddle point equation, we can re-express the $\pi$ mass as $$\begin{aligned}
M_{\rm ps}^2=\frac{\Sigma^2 m}{\phi+m}
.\end{aligned}$$ In the chiral limit, we have the massless $\pi$. With the explicit breaking of $m$, we find a relation reminiscent of the Gell-Mann-Oakes-Renner, $$\begin{aligned}
(\phi+m)^2 M_{\rm ps}^2 =m\Sigma^2 (\phi+m) \sim
-m \langle \bar \psi \psi \rangle
,\end{aligned}$$ if we identify $\phi+m$ as the pion decay constant $f_\pi$. Interestingly, there is a mass hierarchy, $M_{\rm s}^2:\Sigma^2:M_{\rm ps}^2=2:1:0$, in the chiral limit at zero temperature $t=0$.
On the other hand, the flavor singlet masses for the scalar and pseudo-scalar singlet channels have an additional contribution from the anomaly term as $$\begin{aligned}
M_{\rm s0}^2 =&
M_{\rm s}^2 -\Delta M_0^2
\; ,
\label{eq:s0mass}
\\
M_{\rm ps0}^2 =&
M_{\rm ps}^2 +\Delta M_0^2
\label{eq:ps0mass}\end{aligned}$$ with $$\begin{aligned}
\Delta M_0^2
&\equiv
N_f \gamma\frac{\alpha(\phi+m)^{N_f-2}}{ [ \alpha(\phi+m)^{N_f}+1 ]^2}
\nonumber
\\
&= N_f \frac{\tilde \tau }{(\phi+m)^2}
\; .\end{aligned}$$ The would-be Nambu-Goldstone (NG) mode in the singlet channel becomes massive due to the coupling to the [U$_{\rm A}$]{}(1) interaction $\Delta M_0^2$. This same effect appears as a reduction in the scalar singlet channel. This mass gap is related to the (replaced) quenched susceptibility $\tilde \tau$, similarly to the Witten-Veneziano formula[@Witten80]. It is interesting to note that for $N_f \ge 3$ the [U$_{\rm A}$]{}(1) breaking effect disappears $\Delta M_0^2=0$ from the meson masses when the condensate vanishes in the symmetric phase in the chiral limit[@Lee:1996zy]. In contrast, $N_f$-point functions are affected by the [U$_{\rm A}$]{}(1) breaking term.
In Fig. \[fig:mass\] we show the mesonic masses in the flavor-singlet scalar and pseudo-scalar channels and in the flavor-nonsinglet scalar and pseudo-scalar channels. Both for $N_f$=2 and 3, the flavor-singlet pseudo-scalar meson acquires the nonzero mass via (\[eq:ps0mass\]).
![Temperature dependence of the mesonic masses in the flavor-singlet scalar (s0) and pseudo-scalar (ps0) channels and in the flavor-nonsinglet scalar (s) and pseudo-scalar (ps) channels in the chiral limit ($m=0$; thick lines) and with explicit breaking ($m \ne 0$; thin lines) for $N_f=2$ (left) and $N_f=3$ (right). As nonzero quark mass, we set $m=0.1$ for $N_f$=2 and $m=m_c=0.0265$ for $N_f=3$. []{data-label="fig:mass"}](nf2_masses.eps "fig:"){width="8cm"} ![Temperature dependence of the mesonic masses in the flavor-singlet scalar (s0) and pseudo-scalar (ps0) channels and in the flavor-nonsinglet scalar (s) and pseudo-scalar (ps) channels in the chiral limit ($m=0$; thick lines) and with explicit breaking ($m \ne 0$; thin lines) for $N_f=2$ (left) and $N_f=3$ (right). As nonzero quark mass, we set $m=0.1$ for $N_f$=2 and $m=m_c=0.0265$ for $N_f=3$. []{data-label="fig:mass"}](nf3_masses.eps "fig:"){width="8cm"}
Susceptibilities
----------------
The scalar and pseudo-scalar susceptibilities are defined as the responses to the external fields ${ \mathcal{M} }= (s^a + ip^a)\lambda^a /\sqrt{2}$ in eq. ([\[eq:thepf\]]{}): $$\begin{aligned}
\chi_{\rm s}^{ab}=&
\left. -\frac{\partial^2}{\partial s^a \partial s^b}
\Omega (S({ \mathcal{M} });{ \mathcal{M} })\right |_{{ \mathcal{M} }=m}
\; ,
\nonumber \\
\chi_{\rm ps}^{ab}=&
\left. -\frac{\partial^2}{\partial p^a \partial p^b}
\Omega(S({ \mathcal{M} });{ \mathcal{M} }) \right |_{{ \mathcal{M} }=m}
\; .\end{aligned}$$ In our basis, those susceptibilities are found to be diagonal $\chi_{\rm s, ps}^{ab} \propto \delta^{ab}$ with a simple form: $$\begin{aligned}
\chi =&
\chi^{(0)} \frac{1}{1-\chi^{(0)}/\Sigma^2}
=
\chi^{(0)} \frac{\Sigma^2}{M^2}\end{aligned}$$ with $\chi^{(0)}= \Sigma^2 - M^2$ for each channel. A susceptibility diverges when the corresponding $M^2$ vanishes (e.g.) at a critical point where the effective potential $\Omega(S)$ has a flat direction. Especially, in the broken phase, the susceptibilities of the non-singlet NG modes are indefinite.
The topological susceptibility, $\chi_{\text{top}}$ is defined as the response to the angle $\theta$ $$\begin{aligned}
\chi_{\text{top}} \equiv
\left .
-\frac{1}{2N}\frac{\partial^2}{\partial \theta^2} \ln Z_\theta
\right |_{\theta=0}\;
.
\label{eq:topsus}\end{aligned}$$ Using the saddle point solution with small $\theta$, parametrized as $S(\theta)=\phi+i\eta_0(\theta)\lambda^0/\sqrt{2}$, the susceptibility in the mean field approximation is obtained by differentiating $\Omega(S(\theta);\theta)$ around $\theta=0$ as $$\begin{aligned}
\chi_{\rm top} =&
\frac{\partial^2 \Omega}{\partial \theta^2}
-
\left (
\frac{\partial^2 \Omega}{\partial \theta \partial \eta_0}
\right)^2
\left (\frac{\partial^2 \Omega}{\partial \eta_0^2} \right) ^{-1}
{\nonumber}\\
=&
\tilde \tau
-\frac{N_f\; \tilde \tau^2}{(\phi+m)^2}
\frac{1}{M_{\rm ps 0}^2}
{\nonumber}\\
=&
\left [
\frac{1}{ \tilde \tau } +\frac{N_f}{m\Sigma^2 (\phi+m)}
\right]^{-1}
\; ,
\label{eq:topsus2}\end{aligned}$$ or $$\begin{aligned}
\frac{1}{ \chi_{\text{top}} }
=
\frac{1}{\tilde \tau } + \frac{1}{\tau_m}
\; .
\label{eq:chitop}\end{aligned}$$ Here $\tilde \tau$ is the modified susceptibility defined in eq. (\[eq:tildetau\]) in the previous section. As is well known[@Leutwyler:1992yt], the most prominent effect of the fermion coupling is the screening of the topological susceptibility via the contribution $$\begin{aligned}
\tau_m= \frac{\Sigma^2 m (\phi+m)}{N_f}
=\frac{M_{\rm ps}^2 (\phi+m)^2}{N_f}
\; .\end{aligned}$$ In the quenched limit, $N_f \to 0$, the $\chi_{\rm top}$ recovers the quenched susceptibility $\tau = \gamma p (1-p)$, while in the massless limit the susceptibility $\chi_{\text{top}}$ is screened completely to zero.
By changing the variable $S\to \tilde S=Se^{i\theta/N_f}$ under the integral (\[eq:thepf\]) and then applying the saddle point approximation, we find $\Omega(t,m,\theta)=\Omega(t, m e^{i\theta/N_f},0)$. Here the $\theta$ appears only in the combination with the mass $m$. In this latter form, the source term becomes a linear combination of the scalar and pseudo-scalar sources in the flavor-singlet channel: $\mathcal{M}=m e^{i\theta/N_f}=(s_0 +i p_0)\lambda^0/\sqrt{2}$, and the [U$_{\rm A}$]{}(1) relation for the topological susceptibility[@Crewther77] is immediately derived as $$\begin{aligned}
\chi_{\text{top}} &
=
- \frac{1}{2N}
\left(
\frac{\partial^2\ln Z_\theta}{\partial p_0^2}
\left .
\left (
\frac{\partial p_0} {\partial \theta}
\right )^2
+
\frac{\partial \ln Z_\theta}{\partial s_0}
\frac{\partial^2 s_0}{\partial \theta^2}
\right)\right |_{\theta=0}
{\nonumber}\\
&=
-\frac{m^2}{N_f}\chi_{\text{ps}0}
-\frac{m}{N_f}\langle \bar \psi \psi \rangle
\; .\end{aligned}$$ We have seen that the pseudo-scalar meson in the flavor singlet channel has nonzero mass (\[eq:ps0mass\]) because of the [U$_{\rm A}$]{}(1) breaking term, and accordingly the pseudo-scalar singlet susceptibility remains finite in the broken phase in the chiral limit. Thus for the small but nonzero quark mass $m$, the topological susceptibility $\chi_{\rm top}$ decreases following the chiral condensate $\langle \bar \psi \psi \rangle \sim \phi$ with increasing temperature $t$.
![Temperature dependence of topological susceptibilities for $N_f$=2 (left) and 3 (right). \[fig:top\] ](nf2_rel_top.eps "fig:"){width="8cm"} ![Temperature dependence of topological susceptibilities for $N_f$=2 (left) and 3 (right). \[fig:top\] ](nf3_rel_top.eps "fig:"){width="8cm"}
Discussions and summary {#sec:IV}
=======================
In this work we have considered the ChRM model with the near-zero and topological zero modes. It was known before that summation over the topological zero modes with the Poisson distribution results in the determinant interaction, but unfortunately it gives a pathologic unbound potential for the ChRM model with $N_f=3$[@JanikNZ97]. On the other hand, the Gaussian distribution for topological charge $\nu$ with the total number of the modes fixed, leads to the log-determinant type interaction (\[eq:GaussSum\]). This term in the potential resolves the [U$_{\rm A}$]{}(1) problem at zero and finite temperatures as suggested in Ref. [@OhtaniLWH08; @LehnerOVW09], but still yields a second-order phase transition at finite temperature, irrespective of $N_f$.
We have proposed that (i) the numbers of the topological zero modes with right and left chiralities, respectively, have the distributions related to instanton dynamics in a finite volume, and thus (ii) these distributions have an upper bound of ${\cal O}(N)$. We have adopted the binomial distribution as such a distribution. This gives rise to the stable potential which describes the chiral phase transitions of the second and the first order depending on the number of flavors $N_f$=2 and 3, respectively, and resolves the [U$_{\rm A}$]{}(1) problem as well. We have confirmed through numerical evaluations that the proposed model also reasonably reproduces temperature dependence of meson masses, (pseudo-) scalar susceptibilities and topological susceptibility. Notably, the topological susceptibility is consistent with the universal quark mass dependence, satisfying the anomalous Ward identity.
In the conventional ChRM models, the near-zero modes themselves are customarily assumed to emerge from the instanton dynamics, in contrast to the explicit separation of the near-zero modes and the topological zero modes in our modeling. To assess the foundations and relations of these models, a rigorous analysis of these low-lying modes based on the microscopic dynamics would be needed, which is beyond the scope of our current study.
We have adopted the independent binomial distribution for $N_\pm$ as a simple improvement from the Poisson distribution in order to obtain a stable effective potential. However, the distributions of $N_\pm$ can be correlated non-trivially in general. It will be interesting to examine other possibilities for number distributions of the topological zero modes. Incidentally, we note here that the determinant interactions of the from (\[eq:njlpf\]) in the NJL models in 3+1 dimensions also lead to the unbound effective potentials for a large quark condensate beyond the cutoff although the local minimum is usually chosen as the physical ground state.
Concerning phenomenological applications, first we need to tune the model parameters $\Sigma$, $\alpha$ and $\gamma$ as well as the light and strange quark masses $m_{\rm ud}$ and $m_{\rm s}$ so as to reproduce the empirical properties in the vacuum. Extension to the case at finite baryo-chemical potential is straightforward[@Stephanov:1996ki; @Halasz:1998qr], which allows us to study the phase diagram of the ChRM model with the flavor dependence in the space of the temperature, the chemical potential and the quark masses. Especially the existence of the analog of the QCD critical point(s)[@QCDCP] will be an important subject to be studied. Application of this model at finite isospin and strangeness chemical potentials are also planned. Progress in this direction will be reported elsewhere.
The authors are grateful to T. Wettig for fruitful discussions on this work. M.O. thanks T. Hatsuda, C. Lehner and J.J.M. Verbaarschot for collaboration on a related subject. H.F. and T.S. acknowledge useful discussions in the workshop on “Non-equilibrium quantum field theories and dynamic critical phenomena” at Yukawa Institute of Theoretical Physics in March 2009. They are also very grateful to the members of Komaba theory group for their interests and encouragements. This work is supported in part by Grants-in-Aid (\# 19540273) of MEXT, Japan.
Trinomial Distribution
======================
Here we deal with the trinomial distribution for $N_+$ and $N_-$, as a simplest example of the non-factorizable distribution: $$\begin{aligned}
P(N_+, N_-)
=
\frac{(\gamma N) !}{
N_+ ! N_- ! (\gamma N-N_+ -N_-)!}
\;
p_+^{N_+} p_-^{N_-}
(1-p_+-p_-)^{\gamma N -N_+ -N_-}
\; ,
\label{eq:tri_dist}\end{aligned}$$ where $p_{+(-)}$ is the probability for a single cell to be occupied by an (anti-)instanton. Note that the distribution is symmetric under the exchange of $+$ with $-$. Replacing $P(N_+)P(N_-)$ in eq. (\[eq:ia\_sum\]) with $P(N_+,N_-)$ and setting $p_+ = p_- = p$, we obtain the effective potential $$\begin{aligned}
\Omega_{\rm Tri}(S;m,\theta,t)
=
{\tfrac{1}{2}}\Sigma^2 {\rm tr} S^\dagger S
-
{\tfrac{1}{2}}\ln \det
\left[(S+{\cal M})(S^\dagger +{\cal M}^\dagger )+t^2 \right]
-{\tfrac{1}{2}}\gamma \ln \left[\alpha e^{i\theta}\det(S+{\cal M})
+\alpha e^{-i\theta}\det(S^\dagger +{\cal M}^\dagger)+1\right]
\;
\label{eq:tri_pot}\end{aligned}$$ with $\alpha = p/(1-2p)$. This is quite similar to eq. (\[eq:thepf\]), and qualitative features of the model are unchanged. The quenched topological susceptibility, or the variance of $\nu=N_+ -N_-$, for eq. (\[eq:tri\_dist\]) is computed as $$\begin{aligned}
\tau
=
\gamma p
\; .\end{aligned}$$
[\*]{}
T. Banks and A. Casher, Nucl. Phys. B [**169**]{} (1980), 103. J.J.M. Verbaarschot and T. Wettig, Ann. Rev. Nucl. Sci. [**50**]{} (2000), 343.
E.V. Shuryak and J.J.M. Verbarrschot, Nucl. Phys. A [**560**]{} (1993), 306; J.J.M. Verbaarschot and I. Zahed, Phys. Rev. Lett. [**25**]{} (1993), 3852.
A. M. Halasz, A. D. Jackson, R. E. Shrock, M. A. Stephanov and J. J. M. Verbaarschot, Phys. Rev. D [**58**]{} (1998), 096007 \[[arXiv:hep-ph/9804290]{}\].
A. D. Jackson and J. J. M. Verbaarschot, Phys. Rev. D [**53**]{} (1996), 7223 \[[arXiv:hep-ph/9509324]{}\].
T. Wettig, A. Schafer and H. A. Weidenmuller, Phys. Lett. B [**367**]{} (1996), 28 \[Erratum [*ibid*]{}. B [**374**]{} (1996), 362\] \[[arXiv:hep-ph/9510258]{}\].
A. M. Halasz, A. D. Jackson and J. J. M. Verbaarschot, Phys. Rev. D [**56**]{} (1997), 5140 \[[arXiv:hep-lat/9703006]{}\].
M. A. Stephanov, Phys. Rev. Lett. [**76**]{} (1996), 4472 \[[arXiv:hep-lat/9604003]{}\]. M. A. Stephanov, Phys. Rev. D [**73**]{} (2006), 094508 \[[arXiv:hep-lat/0603014]{}\].
J. Han and M. A. Stephanov, [arXiv:0805.1939 \[hep-lat\]]{}.
M. A. Stephanov, Phys. Lett. B [**375**]{} (1996), 249 \[[arXiv:hep-lat/9601001]{}\]. R. D. Pisarski and F. Wilczek, Phys. Rev. D [**29**]{} (1984), 338. M. Kobayashi and T. Maskawa, Prog. Theor. Phys. [**44**]{} (1970), 1422;
M. Kobayashi, H. Kondo and T. Maskawa, Prog. Theor. Phys. [**45**]{} (1971), 1955.
G. ’t Hooft, Phys. Rev. Lett. [**37**]{} (1976), 8; Phys. Rev. D [**14**]{} (1976), 3432 \[Erratum [*ibid*]{}. D [**18**]{} (1978), 2199\].
R.A. Janik, M.A. Nowak, G. Papp and I. Zahed, Nucl. Phys. B [**498**]{} (1997), 313.
M. Ohtani, C. Lehner, T. Wettig and T. Hatsuda, Mod. Phys. Lett. A [**23**]{} (2008), 2465.
C. Lehner, M. Ohtani, J.J.M. Verbaarschot and T. Wettig, Phys. Rev. D [**79**]{} (2009), 074016 \[[arXiv:0902.2640 \[hep-th\]]{}\].
R.A. Janik, M.A. Nowak and I. Zahed, Phys. Lett. B [**392**]{} (1997), 155.
R.A. Janik, M.A. Nowak, G. Papp and I. Zahed, [*AIP Conf. Proc.*]{} [**494**]{} (1999), 408 \[[arXiv:hep-lat/9911024]{}\].
M. A. Nowak and I. Zahed, Phys. Lett. B [**230**]{} (1989), 108. R.J. Crewther, Phys. Lett. B [**70**]{} (1977), 349.
M.A. Shifman, A.I. Vainshtein and V.I. Zakharov, Nucl. Phys. [**B163**]{} (1980), 46.
E. Witten, Nucl. Phys. B [**156**]{} (1979), 269; E. Witten, Ann. Phys. [**128**]{} (1980), 363. G. Veneziano, Nucl. Phys. B [**159**]{} (1979), 213.
S. H. Lee and T. Hatsuda, Phys. Rev. D [**54**]{} (1996), 1871 \[[arXiv:hep-ph/9601373]{}\].
H. Leutwyler and A. V. Smilga, Phys. Rev. D [**46**]{} (1992), 5607. For recent review, M. Stephanov, PoS LAT2006, (2007) 024.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We address the problem of identifying a graph structure from the observation of signals defined on its nodes. Fundamentally, the unknown graph encodes direct relationships between signal elements, which we aim to recover from observable indirect relationships generated by a diffusion process on the graph. The fresh look advocated here permeates benefits from convex optimization and stationarity of graph signals, in order to identify the graph shift operator (a matrix representation of the graph) given only its *eigenvectors*. These *spectral templates* can be obtained, e.g., from the sample covariance of independent graph signals diffused on the sought network. The novel idea is to find a graph shift that, while being consistent with the provided spectral information, endows the network with certain desired properties such as sparsity. To that end we develop efficient inference algorithms stemming from provably-tight convex relaxations of natural nonconvex criteria, particularizing the results for two shifts: the adjacency matrix and the normalized Laplacian. Algorithms and theoretical recovery conditions are developed not only when the templates are perfectly known, but also when the eigenvectors are noisy or when only a subset of them are given. Numerical tests showcase the effectiveness of the proposed algorithms in recovering social, brain, and amino-acid networks.'
author:
- '[^1]'
bibliography:
- 'citations.bib'
title: Network Topology Inference from Spectral Templates
---
Network topology inference, graph signal processing, network deconvolution, graph sparsification.
Introduction {#S:Introduction}
============
Advancing a holistic theory of networks necessitates fundamental breakthroughs in modeling, identification, and controllability of distributed network processes – often conceptualized as *signals defined on the vertices of a graph* [@barrat2012book; @kolaczyk2009book]. Under the assumption that the signal properties are related to the topology of the graph where they are supported, the goal of graph signal processing (GSP) is to develop algorithms that fruitfully leverage this relational structure [@EmergingFieldGSP; @SandryMouraSPG_TSP13]. Instrumental to that end is the so-termed graph-shift operator (GSO) [@SandryMouraSPG_TSP13], a matrix capturing the graph’s local topology and whose eigenbasis is central to defining graph Fourier transforms [@SandryMouraSPG_TSP14Freq]. Most GSP works assume that the GSO (hence the graph) is known, and then analyze how the algebraic and spectral characteristics of the GSO impact the properties of the signals and filters defined on such a graph. Here instead we take the reverse path and investigate how to use information available from graph signals to infer the underlying graph topology; see also [@DongLaplacianLearning; @MeiGraphStructure; @SSAMGMAR_ssp16; @pasdeloup2016inferenceTSIPN16; @Kalofolias2016inference_smoothAISTATS16].
Our focus in this paper is on identifying graphs that explain the structure of a random signal, meaning that there exists a diffusion process in the GSO that can generate the observed signal. Alternatively, we can say that the goal is to recover the GSO which encodes direct relationships between the elements of the signal from observable indirect relationships generated by a diffusion process. Such a problem is shown to be underdetermined and related to the concept of stationarity of graph signals [@marques2016stationaryTSP16; @perraudinstationary2016]. More precisely, it is established that the sought GSO must have the same eigenvectors as the signal’s covariance matrix. This motivates a two-step network topology inference approach whereby we: i) leverage results from GSP theory to identify the GSO’s eigenbasis from realizations of the diffused signal; and ii) rely on these (possibly imperfect and incomplete) *spectral templates* to recover the GSO by estimating its eigenvalues.
Network topology inference from a set of (graph-signal) observations is a prominent problem in Network Science [@kolaczyk2009book; @sporns2012book]. Since networks encode similarities between nodes, several approaches infer the so-termed *association networks* by constructing graphs whose edge weights correspond to correlations or coherence measures indicating a nontrivial level of association between signal profiles at incident nodes [@kolaczyk2009book Ch. 7.3.1]. This approach is not without merit and widely used in practice, but it exhibits several drawbacks, the main one being that links are formed taking into account only pairwise interactions, ignoring that the observed correlations can be due to latent network effects. Acknowledging these limitations, alternative methods rely on partial correlations [@GLasso2008; @kolaczyk2009book], Gaussian graphical models [@Lake10discoveringstructure; @slawski2015estimation; @meinshausen06; @pavez_laplacian_inference_icassp16], structural equation models [@BazerqueGeneNetworks; @BainganaInfoNetworks], Granger causality [@Brovelli04Granger; @sporns2012book], or their nonlinear (kernelized) variants [@Karanikolas_icassp16; @shen2016kernelsTSP16]. Differently, recent GSP-based network inference frameworks postulate that the network exists as a latent underlying structure, and that observations are generated as a result of a network process defined in such graph. For instance, network structure is estimated in [@MeiGraphStructure] to unveil unknown relations among nodal time series adhering to an autoregressive model involving graph-filter dynamics. A factor analysis-based approach is put forth in [@DongLaplacianLearning] to infer graph Laplacians, seeking that input graph signals are smooth over the learned topologies; see also [@Kalofolias2016inference_smoothAISTATS16]. Different from [@DongLaplacianLearning; @MeiGraphStructure; @Kalofolias2016inference_smoothAISTATS16] that operate on the graph domain, the goal here is to identify graphs that endow the given observations with desired spectral (frequency-domain) characteristics. Two works have recently explored this approach and addressed the problem of identifying a GSO based on its eigenvectors. One is [@SSAMGMAR_ssp16], which assumes perfect knowledge of the spectral templates. The other is [@pasdeloup2016inferenceTSIPN16], which only focuses on a Laplacian GSO. After surveying the required GSP background, in Section \[S:prelim\_problem\] we formulate the problem of identifying a GSO that explains the fundamental structure of a random signal diffused on a graph. The novel idea is to search among all feasible networks for the one that endows the resulting graph-signal transforms with prescribed spectral properties (those guaranteeing graph stationarity [@marques2016stationaryTSP16]), while the inferred graph also exhibits desirable structural characteristics such as sparsity or minimum-energy edge weights. It is argued that the required spectral templates can be pragmatically obtained, e.g., via principal component analysis (PCA) of an ensemble of graph signals resulting from network diffusion dynamics. Additional sources for the spectral templates are provided in Section \[Ss:remark\_templates\]. Using the templates as input, a fairly general optimization problem is then formulated to identify the network structure. For concreteness, emphasis is laid on the recovery of two particular GSOs; namely the adjacency matrix and the normalized graph Laplacian, but our methodology can be applied to other matrix representations of graphs. In Section \[Ss:size\_feasible\_set\] we derive conditions under which the feasible set of the optimization problem reduces to a singleton, a situation in which pursuit of additional network structure is rendered vacuous. When multiple solutions exist, provably-tight convex relaxations – leading to computationally-efficient algorithms – are developed to identify the sparsest GSO consistent with the given eigenspace (Section \[Ss:relaxation\]). We then introduce an inference method for the pragmatic case where knowledge of the spectral templates is imperfect (Section \[Ss:ShiftRecImperfEig\]), and establish that the proposed algorithm can identify the underlying network topology robustly. Last but not least, in Section \[Ss:ShiftRecIncompEig\] we investigate the case where only a subset of the GSO’s eigenvectors are known. Such incomplete spectral templates arise, for example, when the observed graph signals are bandlimited. Comprehensive numerical tests corroborate our theoretical findings and confirm that the novel approach compares favorably with respect to: (i) established methods based on (partial) correlations; and (ii) recent graph signal processing-based topology inference algorithms (Section \[S:Simulations\]). Test cases include the recovery of social and structural brain networks, as well as the identification of the structural properties of proteins from a mutual information graph of the co-variation between the constitutional amino-acids [@Marks2011proteins].
*Notation:* The entries of a matrix $\mathbf{X}$ and a (column) vector $\mathbf{x}$ are denoted by $X_{ij}$ and $x_i$, respectively. Sets are represented by calligraphic capital letters and $\bbX_{\ccalI}$ denotes a submatrix of $\bbX$ formed by selecting the rows of $\bbX$ indexed by $\ccalI$. The notation $^T$ and $^\dag$ stands for transpose and pseudo-inverse, respectively; $\mathbf{0}$ and $\mathbf{1}$ refer to the all-zero and all-one vectors. For a vector $\bbx$, $\diag(\mathbf{x})$ is a diagonal matrix whose $i$th diagonal entry is $x_i$; when applied to a matrix, $\diag(\bbX)$ is a vector with the diagonal elements of $\bbX$. The operators $\circ$, $\otimes$, and $\odot$ stand for the Hadamard (elementwise), Kronecker, and Khatri-Rao (columnwise Kronecker) matrix products. $\| \bbX \|_p$ denotes the $\ell_p$ norm of the *vectorized* form of $\bbX$, whereas $\| \bbX \|_{M(p)}$ is the matrix norm induced by the vector $\ell_p$ norm.
Problem Statement {#S:prelim_problem}
=================
A weighted and undirected graph $\ccalG$ consists of a node set $\ccalN$ of known cardinality $N$, an edge set $\ccalE$ of unordered pairs of elements in $\ccalN$, and edge weights $A_{ij}\in\reals$ such that $A_{ij}=A_{ji}\neq 0$ for all $(i,j)\in\ccalE$. The edge weights $A_{ij}$ are collected as entries of the symmetric adjacency matrix $\bbA$ and the node degrees in the diagonal matrix $\bbD:=\diag(\bbA\bbone)$. These are used to form the combinatorial Laplacian matrix $\bbL_c:=\bbD-\bbA$ and the normalized Laplacian $\bbL:=\bbI - \bbD^{-1/2} \bbA\bbD^{-1/2}$. More broadly, one can define a generic GSO $\bbS\in\reals^{N\times N}$ as any matrix having the same sparsity pattern of $\ccalG$ [@SandryMouraSPG_TSP13]. Although the choice of $\bbS$ can be adapted to the problem at hand, most existing works set it to either $\bbA$, $\bbL_c$, or $\bbL$.
The main focus in this paper is on identifying graphs that explain the structure of a random signal. Formally, let $\bbx=[x_1,...,x_N]^T \in\mbR^N$ be a graph signal in which the $i$th element $x_i$ denotes the signal value at node $i$ of an unknown graph $\ccalG$ with shift operator $\bbS$. Further suppose that we are given a zero-mean white signal $\bbw$ with covariance matrix $\E{\bbw\bbw^T} = \bbI$. We say that the graph $\bbS$ represents the structure of the signal $\bbx$ if there exists a diffusion process in the GSO $\bbS$ that produces the signal $\bbx$ from the white signal $\bbw$, that is $$\begin{aligned}
\label{eqn_diffusion}
\bbx\ =\ \alpha_0 \prod_{l=1}^{\infty} (\bbI-\alpha_l \bbS) \bbw
\ =\ \sum _{l=0}^{\infty} \beta_l \bbS^l \bbw .
$$ While $\bbS$ encodes only one-hop interactions, each successive application of the shift percolates (correlates) the original information across an iteratively increasing neighborhood; see e.g. [@segarra2015graphfilteringTSP15]. The product and sum representations in are common – and equivalent – models for the generation of random signals. Indeed, any process that can be understood as the linear propagation of a white input through a static graph can be written in the form in . These include processes generated by graph filters with time-varying coefficients or those generated by the so-called *diffusion* Laplacian *kernels* [@smola2003kernels], to name a few.
The justification to say that $\bbS$ is the structure of $\bbx$ is that we can think of the edges of $\bbS$ as direct (one-hop) relationships between the elements of the signal. The diffusion described by generates indirect relationships. Our goal is to recover the fundamental relationships described by $\bbS$ from a set $\ccalX:=\{\bbx_p\}_{p=1}^P$ of $P$ independent samples of the random signal $\bbx$.
We show next that this is an underdetermined problem closely related to the notion of stationary signals on graphs [@marques2016stationaryTSP16]. Begin by assuming that the shift operator $\bbS$ is symmetric. Define then the eigenvector matrix $\bbV:=[\bbv_1,\ldots,\bbv_N]$ and the eigenvalue matrix $\bbLam:=\diag(\lam_1,\ldots,\lam_N)$ to write $$\begin{aligned}
\label{eqn_shift_operator_normal}
\bbS =\bbV\bbLam\bbV^T.\end{aligned}$$ Further observe that while the diffusion expressions in are polynomials on the GSO of possibly infinite degree, the Cayley-Hamilton theorem implies they are equivalent to polynomials of degree smaller than $N$. Upon defining the vector of coefficients $\bbh:=[h_0,\ldots,h_{L-1}]^T$ and the graph filter $\bbH\in\reals^{N\times N}$ as $\bbH:=\sum_{l=0}^{L-1} h_l \bbS^l$, the generative model in can be rewritten as $$\begin{aligned}
\label{E:Filter_input_output_time}
\bbx = \bigg(\sum_{l=0}^{L-1}h_l \bbS^l\bigg)\,\bbw
= \bbH \bbw\end{aligned}$$ for some particular $\bbh$ and $L$. Since a graph filter $\bbH$ is a polynomial on $\bbS$ [@SandryMouraSPG_TSP13], graph filters are linear graph-signal operators that have the *same eigenvectors* as the shift (i.e., the operators $\bbH$ and $\bbS$ commute). More important for the present paper, the filter representation in can be used to show that *the eigenvectors of $\bbS$ are also eigenvectors of the covariance matrix $\bbC_x:=\E{\bbx\bbx^T}$.* To that end, substitute into the covariance matrix definition and use the fact that $\E{\bbw\bbw^T} = \bbI$ to write $$\begin{aligned}
\label{E:cov_output_filter}
\bbC_x = \E{\bbH\bbw\big(\bbH\bbw\big)^T}
= \bbH \E{\bbw\bbw^T}\bbH^T
= \bbH\bbH^T .\end{aligned}$$ If we further use the spectral decomposition of the shift in to express the filter as $\bbH = \sum_{l=0}^{L-1} h_l (\bbV\bbLam\bbV^T)^l =\bbV(\sum_{l=0}^{L-1} h_l \bbLam^l)\bbV^T$, we can write the covariance matrix as $$\begin{aligned}
\label{eqn_diagonalize_covariance}
\bbC_x\ =\ \bbV\,\bigg|\sum_{l=0}^{L-1}h_l\bbLam^l\bigg|^2\,\bbV^T
\ :=\ \bbV\diag(\bbp)\bbV^T,\end{aligned}$$ where the matrix squared-modulus operator $|\cdot|^2$ should be understood entrywise, and we have defined the vector $\bbp:=\diag(|\sum_{l=0}^{L-1}h_l\bbLam^l|^2)$ in the second equality.
The expression in is precisely the requirement for a graph signal to be stationary [@marques2016stationaryTSP16 Def. 3]; hence, the problem of identifying a GSO that explains the fundamental structure of $\bbx$ is equivalent to identifying a shift on which the signal $\bbx$ is stationary. In this context, $\bbp$ is termed the power spectral density of the signal $\bbx$ with respect to $\bbS$. A consequence of this fact, which also follows directly from , is that the *eigenvectors* of the shift $\bbS$ and the covariance $\bbC_x$ are the same. Alternatively, one can say that the difference between $\bbC_x$, which includes indirect relationships between components, and $\bbS$, which includes exclusively direct relationships, is only on their *eigenvalues*. While the diffusion in obscures the eigenvalues of $\bbS$, the eigenvectors $\bbV$ remain present in $\bbC_x$ as templates of the original spectrum.
Identity also shows that the problem of finding a GSO that generates $\bbx$ from a white input $\bbw$ with unknown coefficients \[cf. \] is *underdetermined*. As long as the matrices $\bbS$ and $\bbC_x$ have the same eigenvectors, filter coefficients that generate $\bbx$ through a diffusion process on $\bbS$ exist.[^2] In fact, the covariance matrix $\bbC_x$ itself is a GSO that can generate $\bbx$ through a diffusion process and so is the precision matrix $\bbC_x^{-1}$. To sort out this ambiguity, which amounts to selecting the eigenvalues of the shift, we assume that the GSO of interest is optimal in some sense. To be more precise, let $\ccalS$ be a convex set that specifies the type of shift operator we want to identify (details on $\ccalS$ are provided in Section \[Ss:AprioriInfoShift\]) and let $\|\bbS\|_0$ count the number of nonzero entries in the GSO. We then want to identify $\bbS_0^*\in\ccalS$ with the smallest number of nonzero entries $$\begin{aligned}
{2}\label{eqn_zero_norm}
\bbS_0^* := &\argmin_{\{\bbS, \bblambda\}} \
&&\|\bbS\|_0 , \nonumber\\
&\text{s. to } && \bbS = \bbV\bbLam\bbV^T
= \sum_{k=1}^N \lambda_k\bbv_k \bbv_k^T, \qquad
\bbS \in \ccalS,\end{aligned}$$ where $\bblambda=[\lambda_1,\ldots,\lambda_N]^T$. To simplify notation we have purposely ignored the optimal eigenvalues $\bblambda^*_0$ that belong to the argument of the minimum. Also, we have written $\bbV\bbLam\bbV^T \! =\! \sum_{k=1}^N \lambda_k\bbv_k \bbv_k^T$ to emphasize that if the eigenvectors $\bbv_k$ are known, the constraints in are linear on the unknown eigenvalues $\lam_k$. Alternatively, we can introduce criteria in the form of generic convex functions $f ( \bbS, \bblambda)$ and define the shift operator that is optimal with respect to these criteria $$\begin{aligned}
{2}\label{E:general_problem}
\bbS^* := &\argmin_{\{\bbS, \bblambda\}} \
&& f ( \bbS, \bblambda), \nonumber\\
&\text{s. to } && \bbS = \bbV\bbLam\bbV^T
= \sum_{k=1}^N \lambda_k\bbv_k \bbv_k^T, \qquad
\bbS \in \ccalS.\end{aligned}$$ Possible convex choices for the criteria in are to: (i) Adopt $f(\bbS, \bblambda) = f(\bbS) = \| \bbS \|_\mathrm{F}$ which finds a GSO that minimizes the total energy stored in the weights of the edges. (ii) Make $f ( \bbS, \bblambda) = f(\bbS) = \| \bbS \|_\infty$ which yields shifts $\bbS$ associated with graphs of uniformly low edge weights. This can be meaningful, e.g., when identifying graphs subject to capacity constraints. (iii) Minimize $f ( \bbS, \bblambda) = f(\bblambda) = -\lambda_2$, where $\lambda_2$ is the second smallest eigenvalue of $\bbS$. If the GSO is further assumed to be a Laplacian matrix, this yields a shift operator that promotes solutions with fast mixing times [@chung1997spectral].
Independently of the criteria, the definitions in and provide a formal description of a GSO $\bbS$ that is considered to be the best possible description of the structure of the signal $\bbx$. Our goal is to find estimators of these operators as described in the following two formal problem statements.
\[problem\_known\_matrix\] Given a covariance matrix $\bbC_x$ identify the optimal description of the structure of $\bbx$ in the form of the graph-shift operator $\bbS_0^*$ defined in or $\bbS^*$ defined in .
\[problem\_data\_driven\] Given a set $\ccalX:=\{\bbx_p\}_{p=1}^P$ of $P$ independent samples of the random signal $\bbx$ estimate the optimal description of the structure of $\bbx$ in the form of the graph-shift operator $\bbS_0^*$ defined in or $\bbS^*$ defined in .
Problem \[problem\_known\_matrix\] is a simple convex optimization problem in the case of the convex objectives in but necessitates relaxations in the case of the minimum zero-norm formulations in . To solve Problem \[problem\_known\_matrix\] we use ensemble covariance matrices to obtain the eigenvectors and show that the estimation of the eigenvalues yields consistent estimators of sparse network structures. Problem \[problem\_known\_matrix\] is addressed in Section \[S:ShiftInfFullEigen\]. To solve Problem \[problem\_data\_driven\] we first use independent samples of the random signal to estimate the covariance eigenvectors. Then we estimate the eigenvalues using reformulations of and which are robust to errors stemming from the aforementioned eigenvector estimation step; see Section \[S:ShiftInfSubsetEigen\] for a detailed treatment of Problem \[problem\_data\_driven\]. Although Problem \[problem\_known\_matrix\] can be thought as a prerequisite to study Problem \[problem\_data\_driven\], Section \[Ss:remark\_templates\] illustrates other situations in which the estimation of a GSO with prescribed eigenvectors is of practical interest.
\[rmk\_precision\] As already mentioned, the precision matrix $\bbC_x^{-1} \! = \! \bbV\bbLambda^{-1}\bbV^T$ is a possible solution to the problem of finding a GSO that explains the structure of $\bbx$. This establishes a clear connection between and the problem of finding sparse estimates of precision matrices [@kolaczyk2009book Ch. 7]. If the precision matrix $\bbC_x^{-1}$ is the [*sparsest*]{} matrix that explains the structure of $\bbx$, this matrix is also the solution to and we have $\bbS_0^* = \bbC_x^{-1}$. In general, however, $\bbC_x^{-1}$ may not be sparse and, even if it is, there may be [*sparser*]{} graphs that explain $\bbx$. In these cases the solution to is a more parsimonious GSO. We can then think of as a [*generalization*]{} of the problem of finding a sparse precision matrix.
A priori knowledge about the GSO {#Ss:AprioriInfoShift}
--------------------------------
The constraint $\bbS \in \ccalS$ in and incorporates a priori knowledge about $\bbS$. If we let $\bbS = \bbA$ represent the adjacency matrix of an undirected graph with non-negative weights and no self-loops, we can explicitly write $\ccalS$ as follows $$\begin{aligned}
\label{E:SparseAdj_def_S}
\ccalS_{\mathrm{A}} \!:= \! \{ \bbS \, | \, S_{ij} \geq 0, \;\, \bbS\!\in\!\ccalM^N\!\!,\;\, S_{ii} = 0, \;\, \textstyle\sum_j S_{j1} \! = \! 1 \}.\end{aligned}$$ The first condition in $\ccalS_{\mathrm{A}}$ encodes the non-negativity of the weights whereas the second condition incorporates the fact that the unknown graph is undirected, hence, $\bbS$ must belong to the set $\ccalM^N$ of real and symmetric $N \! \times \! N$ matrices. The third condition encodes the absence of self-loops, thus, each diagonal entry of $\bbS$ must be null. Finally, the last condition fixes the scale of the admissible graphs by setting the weighted degree of the first node to $1$, and also rules out the trivial solution $\bbS\!=\!\bbzero$. Naturally, the choice of the first node is (almost) arbitrary; any node with at least one neighbor in the sought graph suffices. Although not considered here, additional sources of information such as knowing the existence (or not) of particular edges can be incorporated into $\ccalS$ as well. Alternatively, when $\bbS = \bbL$ represents a normalized Laplacian [@EmergingFieldGSP], the associated $\ccalS_{\mathrm{L}}$ is $$\begin{aligned}
\label{E:def_S_normalized_laplacian}
\ccalS_{\rm L} \!\! := \! \{ \bbS \, | \, S_{ij}\! \in\! [-\!1, 0] \,\, \text{for} \,\,& i\!\neq\! j, \;\;\bbS\! \in \! \ccalM_{+}^N, \nonumber \\
&S_{ii}\!=\!1 \,\, \forall \,\, i, \,\,\, \lambda_1 = 0 \} .\end{aligned}$$ In $\ccalS_{\rm L}$ we impose that $\bbS$ is symmetric and positive semi-definite, its diagonal entries are $1$ and its off-diagonal entries are non-positive. Moreover, since $\bbS$ is a normalized Laplacian we know that the vector $\sqrt{\bbd}$ containing as entries the square roots of the node degrees is an eigenvector whose associated eigenvalue is zero, and this is incorporated into the last constraint. Notice that for this last constraint to be implementable we should be able to identify $\sqrt{\bbd}$ among all the spectral templates in $\bbV$. This can be done since $\sqrt{\bbd}$ is the only eigenvector whose entries have all the same sign [@biyikougu2007laplacian]. In the same way that fixing a scale discards the solution $\bbS = \mathbf{0}$ for adjacency matrices, the constraint $\lambda_1 = 0$ rules out the uninformative solution $\bbS = \bbI$ from the feasible set $\ccalS_{\rm L}$.
Naturally, the identification of other GSOs can be of interest as well, including for instance the combinatorial Laplacian $\bbL_c$ and the random walk Laplacian [@chung1997spectral]. These can be accommodated in our proposed framework via minor modifications to the set $\ccalS$. For concreteness, we henceforth focus exclusively on adjacency and normalized Laplacian matrices.
Additional sources for the spectral templates {#Ss:remark_templates}
---------------------------------------------
The central focus of this paper is to solve the problems in and when eigenvectors $\bbv_k$ are estimated from a sample set $\ccalX$ (cf. Problem \[problem\_data\_driven\]). Notwithstanding, the network topology inference problems in and are applicable as long as eigenvectors or eigenvector estimates are available. Four examples are outlined next.
\[ssec\_AddSourcesGFT\] Expressing signals $\bbx$ in an alternative domain $\tbx$ by using an orthonormal transform $\tilde{\bbx} := \bbU^T\bbx$, such as Fourier, wavelets, or discrete-cosine, is a cornerstone operation in signal processing. If we make $\bbV=\bbU$ in and we formulate the problem of identifying a graph shift $\bbS = \bbV \bbLambda \bbV^T = \bbU \bbLambda \bbU^T$ whose graph Fourier transform [@SandryMouraSPG_TSP14Freq] $\tilde{\bbx} := \bbV^T\bbx = \bbU^T\bbx$ is the given orthonormal transform of interest. This is important because it reveals the proximity structure between signal components that is implicitly assumed and exploited by the transform $\bbU$.
\[sssec\_linearoperators\] In addition to describing linear diffusion dynamics \[cf. \], graph filters represent linear transformations that can be implemented in a distributed manner [@segarra2015reconstruction; @EUSIPCO_our_interp_2015; @ssamar_distfilters_allerton15]. In the context of distributed algorithms, consider implementing a prescribed linear network operator $\bbB$ using a graph filter $\bbH=\sum_{l=0}^{N-1}h_l\bbS^l$ [@segarra2015graphfilteringTSP15]. A necessary condition to accomplish this goal is that the eigenvectors of the shift $\bbS = \bbV \bbLambda \bbV^T$ and those of the linear transformation $\bbB = \bbV_\bbB \bbLambda_\bbB \bbV_\bbB^T$ must coincide; [@segarra2015graphfilteringTSP15 Prop. 1]. Since $\bbV_\bbB$ can be obtained from the prescribed $\bbB$, the problems in and can be solved using $\bbV=\bbV_\bbB$ as input. Problem , for example, enable us to find the sparsest $\bbS$ which facilitates implementation of a given network operator $\bbB$ via distributed graph filtering.
\[sssec\_sparsify\] Given a GSO $\bbT$, we can use our framework to obtain a different shift $\bbS$ with the same eigenvectors as $\bbT$, but with desirable properties encoded in $\ccalS$ and $f(\bbS,\bblambda)$ \[cf. \]. If we set $f(\bbS,\bblambda)=\|\bbS\|_0$, this *graph sparsification* problem can be addressed by solving using as inputs the eigenvectors of $\bbT$. Note that, different from the setup in Problem \[problem\_known\_matrix\], the matrix $\bbT$ is not necessarily a covariance matrix.
\[sssec\_network\_deonvolution\] The network deconvolution problem is the identification of an adjacency matrix $\bbS$ that encodes direct dependencies when given an adjacency $\bbT$ that includes indirect relationships. The problem is a generalization of channel deconvolution and can be solved by making $\bbS = \bbT \, (\bbI + \bbT)^{-1}$ [@FeiziNetworkDeconvolution]. This solution assumes a diffusion as in that results in a single-pole-single-zero graph filter. A more general approach is to assume that $\bbT$ can be written as a polynomial of $\bbS$ but be agnostic to the form of the filter. This leads to problem formulations and with $\bbV$ given by the eigenvectors of $\bbT$. As in the graph sparsification problem and different from Problem \[problem\_known\_matrix\], the matrix $\bbT$ is not necessarily a covariance matrix.
Topology inference from spectral templates {#S:ShiftInfFullEigen}
==========================================
As discussed in the previous section, the goal is to find a graph shift $\bbS$ that is diagonalized by the given spectral templates $\bbV=[\bbv_1,\ldots,\bbv_N]$. In the absence of additional constraints the problem is ill-posed, so we further impose conditions on $\bbS$ via the set $\ccalS$ and search for the shift that minimizes a pre-specified cost $f$ \[cf. \]. The structure of the feasible set in plays a critical role towards solving our network topology inference problem. The reason is twofold. First, notice that both sets $\ccalS_{\rm A}$ and $\ccalS_{\rm L}$ are convex. Hence, convexity of problem depends exclusively on the choice of the objective $f(\bbS,\bblambda)$, a key property to facilitate the solution of in practice. If $f(\bbS,\bblambda)$ is chosen to be convex – e.g., equal to $\|\bbS\|_p$ with $p\geq 1$ – the overall optimization will be convex too. Second, the dimension of the feasible set is generally small. In fact, it can be shown that in a number of setups the feasible set reduces to a singleton, or otherwise to a low-dimensional subspace. This is important because even if the objective is non-convex, searching over a small space need not be necessarily difficult.
We first investigate the size of the feasible set and provide conditions under which it reduces to a singleton thus rendering $f$ inconsequential to the optimization. Then, for the cases where there are multiple feasible solutions, we focus on the sparsity-promoting formulation, i.e., $f(\bbS,\bblambda)=\|\bbS\|_0$. The resultant optimization is non-convex and in fact NP-hard, so we propose computationally-efficient convex relaxations which are provably tight under some technical conditions.
Size of the feasibility set {#Ss:size_feasible_set}
---------------------------
The feasible set of problem for both $\ccalS_{\mathrm{L}}$ and $\ccalS_{\mathrm{A}}$ is in general small. To be more precise, some notation must be introduced. Define $\bbW \! := \! \bbV \odot \bbV \! \in \! \reals^{N^2 \! \times \! N}$, where $\odot$ denotes the Khatri-Rao product. Notice that from the definition of $\bbS$ we can write $\bbs := \mathrm{vec}(\bbS)$ as $\bbs = \bbW \bblambda$. Hence, each row of $\bbW$ represents the $N$ weighting coefficients that map $\bblambda$ to the corresponding entry of $\bbS$. Further, define the set $\ccalD$ containing the indices of $\bbs$ corresponding to the diagonal entries of $\bbS$ and select the corresponding rows of $\bbW$ to form $\bbW_\ccalD\!\in\!\reals^{N\times N}$. Also, define the matrix $\bbU := \bbV^{\mathbf{1}} \circ \bbV^{\mathbf{1}}\in\! \reals^{N\times N}$, where $\circ$ denotes the elementwise product and $\bbV^{\mathbf{1}} := [\bbone, \bbv_2, \bbv_3, \ldots, \bbv_N ]$. Using these conventions, the following result holds.
\[P:Feasibility\_singleton\] Assume that is feasible, then it holds that:\
a) If $\ccalS = \ccalS_{\mathrm{A}}$, then $\mathrm{rank}({\bbW}_{\ccalD}) \leq N - 1$. Similarly, if $\ccalS = \ccalS_{\mathrm{L}}$, then $\mathrm{rank}(\bbU) \leq N - 1$.\
b) If $\mathrm{rank}({\bbW}_{\ccalD}) = N - 1$ when $\ccalS = \ccalS_{\mathrm{A}}$ or $\mathrm{rank}(\bbU) = N - 1$ when $\ccalS = \ccalS_{\mathrm{L}}$, then the feasible set of is a singleton.
We show statements a) and b) for the case $\ccalS = \ccalS_{\mathrm{A}}$. The proofs for $\ccalS = \ccalS_{\mathrm{L}}$ are analogous and thus omitted. The key of the proof is to note that we may write $\bbW_\ccalD \bblambda = \diag( \bbS ) = \mathbf{0}$ for all feasible $\bblambda$. Hence, feasibility implies that ${\bbW}_{\ccalD}$ is rank-deficient as stated in a). To show b), assume that $\mathrm{rank}({\bbW}_{\ccalD}) = N - 1$ so that $\bblambda$ in $\mathrm{null}({\bbW}_{\ccalD})$ is unique up to a scaling factor. However, since one of the conditions in $\ccalS_A$ forces the first row of $\bbS$ to sum up to $1$ \[cf. \], this scaling ambiguity is resolved and the unique feasible $\bblambda$ (and hence $\bbS$) is obtained.
Proposition \[P:Feasibility\_singleton\] offers sufficient conditions under which reduces to a feasibility problem. More specifically, when condition b) is met, the objective in is inconsequential since there exists only one feasible $\bbS$. For more general cases, however, the GSO that minimizes the particular cost function $f$ chosen is recovered. Among the potential cost functions, the sparsity-inducing $\ell_0$ norm $f(\bbS, \bblambda) = \| \bbS \|_0$ is non-convex, thus, challenging to solve in practice. Due to the widespread interest in identifying sparse graphs (e.g., of direct relationships among signal elements), we devote the ensuing subsection to study this latter case separately.
Relaxation for the sparse formulation {#Ss:relaxation}
-------------------------------------
Many large-scale, real-world networks are sparse [@kolaczyk2009book], so it is often meaningful to infer a sparse GSO where most of the entries in $\bbS$ are zero \[cf. \]. In practice, the usual approach to handle the non-convex $\ell_0$ (pseudo) norm objective in is to relax it to an iteratively re-weighted $\ell_1$ norm. Specifically, with $p$ denoting an iteration index, we aim to solve a sequence $p=1,...,P$ of weighted $\ell_1$-norm minimization problems $$\begin{aligned}
\label{E:SparseAdj_wl11_obj}
\bbS_{\omega}^* \! := \! \argmin_{ \{\bbS, \bblambda\}}\sum_{i,j} \!\omega_{ij}(p)|S_{ij}| \quad \text{s. to } \,\bbS=\sum_{k =1}^N \! \lambda_k\bbv_k \bbv_k^T, \,\bbS \!\in\! \ccalS,\end{aligned}$$ with weights $\omega_{ij}(p):=\tau/\left(|S_{ij}(p-1)|+\delta\right)$, for appropriately chosen positive constants $\tau$ and $\delta$. Intuitively, the goal of the re-weighted scheme in is that if $|S_{ij}(p-1)|$ is small, in the next iteration the penalization $\omega_{ij}(p)$ is large, promoting further shrinkage of $S_{ij}$ towards zero [@candes_l0_surrogate].
Naturally, under condition b) in Proposition \[P:Feasibility\_singleton\] the solutions $\bbS_0^*$ of and $\bbS_{\omega}^*$ of are guaranteed to coincide given that the feasible set is reduced to a singleton. Moreover, even when condition b) is not satisfied, there exist weights $\omega_{ij}$ that guarantee the equivalence of both solutions. To state this formally, define the set $\ccalJ$ containing the indices identifying the support of $\bbS^*_0$ and denote by $\ccalJ^c$ its complement. Whenever $\bbS_0^*$ is the unique solution to , it is not hard to establish that by setting weights in as $\omega_{ij}=1$ for $(i,j)\in\ccalJ^c$ and $\omega_{ij}=0$ otherwise, then $\bbS_{\omega}^*$ is unique and equal to $\bbS_0^*$. The upshot of this simple observation is that there exist optimal weights so that the sparsest solution $\bbS^*_0$ can be recovered by solving a convex optimization problem. This result confers validity to the re-weighted formulation in , nonetheless, we can neither choose these weights without knowing $\bbS_0^*$ a priori nor there is a guarantee that the succession of weights $\omega_{ij}(p)$ converges to these optimal weights. Hence, we now focus on the derivation of theoretical guarantees for a [particular]{} set of weights that can be set a priori, namely, we consider the formulation in which each entry of the GSO is equally weighted. This boils down to solving the convex optimization problem $$\begin{aligned}
\label{E:SparseAdj_l01_obj}
\bbS^*_1 \! := \! \argmin_{ \{\bbS, \bblambda\}} \,\| \bbS\|_{1} \quad \text{s. to } \bbS \! = \! \textstyle\sum_{k =1}^N \lambda_k\bbv_k \bbv_k^T, \,\,\, \bbS \in \ccalS.\end{aligned}$$ Interestingly, under certain conditions we can ensure that the solution $\bbS^*_1$ to the relaxed problem coincides with $\bbS^*_0$. To be more specific, define $\bbs_0^*:= \mathrm{vec}(\bbS_0^*)$, denote by $\ccalD^c$ the complement of $\ccalD$ and partition $\ccalD^c$ into $\ccalK$ and $\ccalK^c$, with the former indicating the positions of the nonzero entries of $\bbs_{0\ccalD^c}^*:=(\bbs_0^*)_{\ccalD^c}$, where we recall that matrix *calligraphic subscripts* select rows. Denoting by $^\dag$ the matrix pseudo-inverse, we define $$\label{E:def_matrix_M}
\bbM := (\bbI - \bbW \bbW^\dag)_{\ccalD^c} \,\in\reals^{N^2-N\times N^2},$$ i.e., the orthogonal projector onto the kernel of $\bbW^T$ constrained to the off-diagonal elements in $\ccalD^c$. With $\bbe_1$ denoting the first canonical basis vector, we construct the matrix $$\label{E:def_matrix_R}
\bbR := [ \bbM, \, \bbe_1 \otimes \bbone_{N-1}]\,\in\reals^{N^2-N\times N^2+1},$$ by horizontally concatenating $\bbM$ and a column vector of size $|\ccalD^c|$ with ones in the first $N-1$ positions and zeros elsewhere. With this notation in place, the following recovery result holds.
\[T:Recovery\] Whenever $\ccalS = \ccalS_{\rm A}$ and assuming problem is feasible, $\bbS^*_1 = \bbS^*_0$ if the two following conditions are satisfied:\
A-1) $\rank(\bbR_{\ccalK}) = |\ccalK|$; and\
A-2) There exists a constant $\delta > 0$ such that $$\label{E:condition_recovery_noiseless}
{\psi_{\bbR}} := \| \bbI_{\ccalK^c}(\delta^{-2}\bbR \bbR^T+\bbI_{\ccalK^c}^T \bbI_{\ccalK^c})^{-1}\bbI_{\ccalK}^T \|_{M(\infty)}<1.$$
Recalling that $\bbs = \mathrm{vec}(\bbS)$, problem for the case where $\ccalS = \ccalS_{\rm A}$ can be reformulated as $$\label{E:proof_noiseless_recovery_010}
\min_{ \{\bbs, \bblambda\}} \;\; \|\bbs\|_1 \;\;\text{s. to } \;\;\bbs= \bbW \bblambda,\;\;\bbs_{\ccalD}=\bbzero,\;\;(\bbe_1 \otimes \bbone_N)^T \bbs = 1,$$ where the last equality imposes that the first column of $\bbS$ must sum up to $1$ \[cf. \]. Notice that the non-negativity constraint in $\ccalS_{\rm A}$ is ignored in . However, if we show that can recover the sparse solution $\bbs_0^*$, then the same solution would be recovered by the more constrained problem . Notice that we may solve for $\bblambda$ in closed form as $\bblambda^* = \bbW^{\dag} \bbs$. Consequently, becomes $$\label{E:proof_noiseless_recovery_020}
\min_{\bbs} \|\bbs\|_1 \;\;\: \text{s. to } (\bbI - \bbW \bbW^\dag) \bbs = \bbzero,\;\bbs_{\ccalD}=\bbzero,\;(\bbe_1 \otimes \bbone_N)^T \bbs = 1.$$ Leveraging the fact that $\bbI - \bbW \bbW^\dag$ is symmetric, the first equality in can be rewritten as \[cf. \] $$\label{E:proof_noiseless_recovery_030}
(\bbI - \bbW \bbW^\dag)^T_\ccalD \bbs_{\ccalD} + \bbM^T \bbs_{\ccalD^c} = \bbzero,$$ and the second equality in forces the first term of to be zero. With these considerations, we may restate as $$\label{E:proof_noiseless_recovery_040}
\min_{ \bbs_{\ccalD^c}} \;\; \|\bbs_{\ccalD^c}\|_1 \;\;\text{s. to } \;\; \bbR^T \bbs_{\ccalD^c} = \bbb,$$ where $\bbb$ is a binary vector of length $N^2+1$ with all its entries equal to $0$ except for the last one that is a $1$. Problem takes the form of classical basis pursuit [@ChenDonohoBP]. Notice that the system of linear equations in is overdetermined since $\bbR^T \in \reals^{N^2+1 \times | \ccalD^c|}$, however, feasibility of guarantees that the mentioned system of equations is compatible. The following two conditions are required for the solution of to coincide with the sparse solution ${\bbs_0^*}_{\ccalD^c}$ (cf. [@zhang2013one]):
- $\mathrm{ker}(\bbI_{\ccalK^c}) \cap \mathrm{ker}(\bbR^T) = \{ \mathbf{0} \}$; and
- There exists a vector $\bby \in \reals^{|\ccalD^c|}$ such that $\bby \in \mathrm{Im}(\bbR)$, $\bby_\ccalK = \mathrm{sign}(({\bbs_0^*}_{\ccalD^c})_\ccalK)$, and $\| \bby_{\ccalK^c}\|_\infty < 1$.
The remainder of the proof is devoted to showing that if conditions *A-1)* and *A-2)* in the statement of the theorem hold true, then a) and b) are satisfied.
To see that *A-1)* implies a) notice that the nullspace of $\bbI_{\ccalK^c}$ is spanned by the columns of $\bbI^T_{\ccalK}$. Hence, for a) to hold we need the $|\ccalK|$ columns of $\bbR^T$ in positions $\ccalK$ to form a full column rank matrix. In condition *A-1)* we require $\bbR_\ccalK$ to be full row rank, which is an equivalent property.
The next step is to show that condition *A-2)* implies b). For this, consider the following $\ell_2$-norm minimization problem $$\begin{aligned}
\label{E:l_2_minimization_dual_certificate}
\min_{ \{\bby, \bbz\}} \; \delta^2\|\bbz\|_2^2 + \|\bby\|_2^2 \;\;\text{s. to } \; \bby= \bbR\bbz,\;\; \bby_\ccalK = \mathrm{sign}(({\bbs_0^*}_{\ccalD^c})_\ccalK),
\end{aligned}$$ where $\delta$ is a positive tuning constant. The inclusion of the term $\delta^2\|\bbz\|_2^2$ in the objective guarantees the existence of a closed-form expression for the optimal solution, while preventing numerical instability when solving the optimization. We will show that the solution $\bby^*$ to problem satisfies the requirements imposed in condition b). The two constraints in enforce the fulfillment of the first two requirements in b), hence, we are left to show that $\| \bby^*_{\ccalK^c}\|_\infty < 1$. Since the values of $\bby_{\ccalK}$ are fixed, the constraint $\bby= \bbR \bbz$ can be rewritten as $\bbI_{\ccalK}^T \mathrm{sign}(({\bbs_0^*}_{\ccalD^c})_\ccalK) = -\bbI_{\ccalK^c}^T \bby_{\ccalK^c} + \bbR \delta^{-1} \delta\bbz$. Then, by defining the vector $\bbt := [\delta\bbz^T, -\bby_{\ccalK^c}^T]^T$ and the matrix $\bbPhi:= [\delta^{-1} \bbR^T, \bbI_{\ccalK^c}]$, can be rewritten as $$\begin{aligned}
\label{E:l_2_minimization_dual_certificate_rewritten}
\min_{\bbt} \;\; \|\bbt\|_2^2 \;\;\text{s. to } \; \bbI_{\ccalK}^T \mathrm{sign}(({\bbs_0^*}_{\ccalD^c})_\ccalK) = \bbPhi^T \bbt.
\end{aligned}$$ The minimum-norm solution to is given by $\bbt^*=(\bbPhi^T)^\dag \bbI_{\ccalK}^T \mathrm{sign}(({\bbs_0^*}_{\ccalD^c})_\ccalK)$ from where it follows that $$\begin{aligned}
\label{E:dual_certificate_solution_norm2_delta}
\bby^*_{\ccalK^c} \!=\! - \bbI_{\ccalK^c}(\delta^{-2}\bbR \bbR^T+\bbI_{\ccalK^c}^T \bbI_{\ccalK^c})^{-1}\bbI_{\ccalK}^T \, \mathrm{sign}(({\bbs_0^*}_{\ccalD^c})_\ccalK).
\end{aligned}$$ Condition a) guarantees the existence of the inverse in . Since $\| \mathrm{sign}(({\bbs_0^*}_{\ccalD^c})_\ccalK) \|_{\infty} \! = \! 1$, we may bound the $\ell_\infty$ norm of $\bby^*_{\ccalK^c}$ as $\| \bby^*_{\ccalK^c} \|_{\infty} \leq \| \bbI_{\ccalK^c}(\delta^{-2}\bbR \bbR^T+\bbI_{\ccalK^c}^T \bbI_{\ccalK^c})^{-1}\bbI_{\ccalK}^T \|_{M(\infty)} = \psi_{\bbR}$. Hence, condition *A-2)* in the theorem guarantees $\| \bby^*_{\ccalK^c} \|_{\infty}<1$ as wanted, concluding the proof.
Theorem \[T:Recovery\] offers *sufficient* conditions under which the relaxation guarantees sparse recovery for adjacency matrices. Simulations in Section \[S:Simulations\] reveal that the bound imposed on $\psi_{\bbR}$ is tight by providing examples where $\psi_{\bbR}$ is equal to 1 and for which recovery fails. In Theorem \[T:Recovery\], condition *A-1)* ensures that the solution to is unique, a necessary requirement to guarantee sparse recovery. Condition *A-2)* is derived from the construction of a dual certificate specially designed to ensure that the unique solution to also has minimum $\ell_0$ norm [@zhang2013one].
Recall that the $\ell_\infty$ norm in is the maximum $\ell_1$ norm across the rows of the argument matrix, which has $|\ccalK_c|$ rows each containing $|\ccalK|$ elements. It is thus expected that sparser graphs (small $|\ccalK|$) might have smaller values of $\psi_{\bbR}$. Furthermore, to have an intuitive understanding of $\psi_{\bbR}$ it is helpful to see that condition *A-2)* is always satisfied whenever $\bbR \bbR^T$ is non-singular. More specifically, for small values of $\delta$ we have that ${\psi_{\bbR}} \approx \delta^2 \| \bbI_{\ccalK^c}(\bbR \bbR^T)^{-1}\bbI_{\ccalK}^T \|_{M(\infty)}$, which can be made arbitrarily small and, in particular, strictly smaller than 1. Matrix $\bbR \bbR^T$ can be shown to be invertible whenever $\mathrm{rank}({\bbW}_{\ccalD}) = N - 1$ (cf. Proposition \[P:Feasibility\_singleton\]). Thus, in the extreme case where the feasible set is a singleton, Theorem \[T:Recovery\] guarantees recovery, as expected. A more general characterization of the ensembles of random graphs that tend to satisfy with high probability is of interest, but left as future research.
The recovery result of Theorem \[T:Recovery\] can be replicated for the case where the shift of interest is a normalized Laplacian, i.e., when $\ccalS = \ccalS_{\rm L}$. To state this formally, if we define $\bbQ := (\bbI - \tilde{\bbU} \tilde{\bbU}^\dag)_{\ccalD^c}$, where $\tilde{\bbU} := \tilde{\bbV} \odot \tilde{\bbV}$ for $ \tilde{\bbV} := [\bbv_2, \bbv_3, \ldots, \bbv_N]$ the following result holds. The proof follows the same steps as those in Theorem \[T:Recovery\] and, thus, is omitted.
\[T:Recovery\_Laplacian\] Whenever $\ccalS = \ccalS_{\rm L}$ and assuming problem is feasible, $\bbS^*_1 = \bbS^*_0$ if the two following conditions are satisfied:\
L-1) $\rank(\bbQ_\ccalK) = | \ccalK |$; and\
L-2) There exists a constant $\delta > 0$ such that $$\label{E:condition_recovery_noiseless_laplacian}
{\psi_{\bbQ}} := \| \bbI_{\ccalK^c}(\delta^{-2}\bbQ \bbQ^T+\bbI_{\ccalK^c}^T \bbI_{\ccalK^c})^{-1}\bbI_{\ccalK}^T \|_{M(\infty)}<1.$$
Imperfect spectral templates {#S:ShiftInfSubsetEigen}
============================
Whenever the number of observed graph signals is limited or the observations are noisy, assuming perfect knowledge of the spectral templates $\bbV$ may be unrealistic. This section broadens the scope of the network inference problems dealt with so far, to accommodate imperfect spectral templates that can either be noisy or incomplete. Specifically, we investigate pragmatic scenarios where: i) only an approximate version of $\bbV$ can be obtained (e.g., from the eigenvectors of a *sample* covariance matrix); and ii) where only a subset of $\bbV$ is available (e.g., when the observed signals are bandlimited and one can only estimate the non-zero frequencies that are present).
Noisy spectral templates {#Ss:ShiftRecImperfEig}
------------------------
We first address the case where knowledge of an approximate version of the spectral templates $\hat{\bbV} = [\hat{\bbv}_1,\ldots,\hat{\bbv}_N]$ is available. The question here is how to update the general formulation in to account for the discrepancies between the estimated spectral templates $\hat{\bbV}$ and the actual eigenvectors of $\bbS$. An instructive reformulation is to include $\bbV= [\bbv_1, \ldots, \bbv_N]$ as decision variables and formulate the following problem $$\begin{aligned}
\label{E:SparseAdj_l1_obj_noisy_vectors}
&\min_{\{\bbS, \bblambda, \bbV\}} \;f( \bbS,\bblambda) \quad \\
&\text{s. to } \,\, \bbS = \textstyle\sum_{k =1}^N \lambda_k\bbv_k \bbv_k^T, \,\,\, \bbS \in \ccalS, \,\,\, d(\bbv_k, \hat{\bbv}_k) \leq \epsilon_k \,\,\forall \, k, \nonumber\end{aligned}$$ where $d(\cdot, \cdot)$ is a *convex* vector distance function, such as the $\ell_p$ norm of the vector difference for $p \! \geq \! 1$. The idea in is to find a sparse $\bbS$ that satisfies the desired properties in $\ccalS$, while its eigenvectors $\bbv_k$ are each of them close to the observed ones $\hat{\bbv}_k$. The value of $\epsilon_k$ must be chosen based on a priori information on the imperfections, such as the number of signals used to estimate the sample covariance, or the statistics of the observation noise. While conceptually simple, problem is more challenging than its noiseless counterpart, since the first constraint is non-convex given that both $\lambda_k$ and $\bbv_k$ are optimization variables.
A more tractable alternative is to form $\bbS':=\sum_{k =1}^{N} \lambda_k \hat{\bbv}_k \hat{\bbv}_k^T$ and search for a shift $\bbS$ that possesses the desired properties while being close to $\bbS'$. Formally, one can solve $$\begin{aligned}
\label{E:SparseAdj_l1_obj_noisy_matrix}
\hat{\bbS}^*&:=\argmin_{\{\bbS, \bblambda, \bbS'\}} \; f( \bbS,\bblambda) \,\,\,\, \\
&\text{s. to } \,\, \bbS' = \textstyle\sum_{k =1}^N \lambda_k \hat{\bbv}_k \hat{\bbv}_k^T, \,\,\,\, \bbS \in \ccalS, \,\,\,d(\bbS, \bbS') \leq \epsilon, \nonumber\end{aligned}$$ where $d(\cdot, \cdot)$ is a *convex* matrix distance whose form depends on the particular application. E.g., if $\|\bbS-\bbS'\|_\mathrm{F}$ is chosen, the focus is more on the similarities across the entries of the shifts, while $\|\bbS-\bbS'\|_{M(2)}$ focuses on their spectrum. Additional conic constraints of the form $\|(\bbS-\bbS')\hbv_k\|_2\leq \lambda_k\epsilon_k$ enforcing that particular eigenvectors are well approximated can also be incorporated. From an application point of view, the formulation in is also relevant to setups where the templates $\hbV$ are not necessarily noisy but the goal is to enlarge the set of feasible GSOs. This can be of interest if, for example, finding an $\bbS$ that is both sparse and with the exact templates collected in $\hbV$ is impossible (cf. Section \[Ss:remark\_templates\]).
The difficulty in solving is determined by $f$ and $\ccalS$. Hence, the challenges and approaches are similar to those in Section \[S:ShiftInfFullEigen\]. For the particular case of sparse shifts, the $\ell_1$ norm relaxation of yields \[cf. \] $$\begin{aligned}
\label{E:SparseAdj_l1_obj_noisy_matrix_v2}
\hat{\bbS}^*_1&:=\argmin_{\{\bbS, \bblambda, \bbS'\}} \; \|\bbS\|_1 \,\,\,\, \\
&\text{s. to } \,\, \bbS' = \textstyle\sum_{k =1}^N \lambda_k \hat{\bbv}_k \hat{\bbv}_k^T, \,\,\,\, \bbS \in \ccalS, \,\,\,d(\bbS, \bbS') \leq \epsilon, \nonumber\end{aligned}$$ where iteratively re-weighted schemes are also possible. Moreover, further uncertainties can be introduced in the definition of the feasible set $\ccalS$, e.g. in the scale of the admissible graphs for the case of $\ccalS = \ccalS_{\rm A}$ (cf. Proposition \[P:noisy\_recovery\] and for additional details).
When the interest is in recovering a normalized Laplacian \[cf. \], a possible implementation is to enforce the constraint $\lambda_1 = 0$ *talis qualis* on entailing that one of the eigenvalues of $\bbS'$ (and not $\bbS$) is equal to zero. However, the smallest eigenvalue of $\bbS$ must be close to zero due to the constraint on the distance between $\bbS$ and $\bbS'$. Alternatively, the objective can be augmented by also considering the nuclear norm $\| \bbS \|_*$ to further promote rank-deficiency on $\bbS$.
To assess the effect of the noise in recovering the sparsest $\bbS$, we define matrices $\hbW$, $\hat{\bbR}$ and $\hat{\bbQ}$ which are counterparts of $\bbW$, $\bbR$ and $\bbQ$ defined prior to Theorem \[T:Recovery\], but based on the noisy templates $\hat{\bbV}$ instead of $\bbV$. Further, we drop the non-negativity constraint in $\ccalS_{\rm A}$ – to obtain $\tilde{\ccalS}_{\rm A}$ – and incorporate the scale ambiguity by augmenting $d(\bbS, \bbS')$ as $\tilde{d}(\bbS, \bbS') = (d(\bbS, \bbS')^2 + (\textstyle\sum_j S_{j1} - 1)^2)^{1/2}$. With this notation, the following result on robust recovery of network topologies holds.
\[P:noisy\_recovery\] When $d(\bbS, {\bbS}') = \| \bbS - {\bbS}' \|_\mathrm{F}$, and assuming that there exists at least one ${\bbS}'$ such that $\tilde{d}(\bbS_0^*, {\bbS}')\leq \epsilon$, the solution $\hbs_1^* := \mathrm{vec}(\hbS_1^*)$ to for $\ccalS = \tilde{\ccalS}_{\rm A}$ with scale ambiguity satisfies $$\label{E:recov_noise}
\| \hbs_1^* - \bbs_0^* \|_1 \leq C \epsilon, \quad \text{with} \,\,\, C = 2 C_1 + 2 C_2 C_3,$$ if the same conditions stated in Theorem \[T:Recovery\] hold but for $\hat{\bbR}$ instead of ${\bbR}$. Constants $C_1$, $C_2$, and $C_3$ are given by $$\label{E:recov_noise_constants}
C_1 \!=\! \frac{\sqrt{| \ccalK |}}{\sigma_{\min}(\hat{\bbR}^T_\ccalK)}, \,\, C_2 \!=\! \frac{1 + \| \hat{\bbR}^T \|_{M(2)} C_1}{1 - \psi_{\hat{\bbR}}}, \,\, C_3 \!=\! \| \hat{\bbR}^{\dag} \|_{M(2)} N,$$ where $\sigma_{\min}( \cdot)$ denotes the minimum singular value of the argument matrix. An analogous result can be derived for the case $\ccalS = \tilde{\ccalS}_\mathrm{L}$ (where the non-positivity constraint is dropped) whenever $\hat{\bbQ}$ satisfies the conditions in Theorem \[T:Recovery\_Laplacian\].
We reformulate in vector form for the case $\ccalS = \tilde{\ccalS}_{\rm A}$ with scale ambiguity to obtain $$\begin{aligned}
\label{E:proof_noisy_recovery_005}
\min_{ \{\bbs, \bblambda, \bbs'\}} \;\; \|\bbs\|_1 \;\;\text{s. to }& \;\;\bbs'= \hbW \bblambda,\;\;\bbs_{\ccalD}=\bbzero, \\
&\|\bbs - \bbs'\|_2^2 + ((\bbe_1 \otimes \bbone_N)^T \bbs - 1)^2 \leq \epsilon^2. \nonumber
\end{aligned}$$ Substituting the first equality constraint in into the inequality constraint, then solving for $\bblambda$ as $\bblambda^* = \hbW^\dag \bbs$, and finally using the second equality constraint to reduce the optimization variables to $\bbs_{\ccalD^c}$, we may restate as \[cf. \] $$\label{E:proof_noisy_recovery_010}
\min_{ \bbs_{\ccalD^c}} \;\; \|\bbs_{\ccalD^c}\|_1 \quad\; \text{s. to } \; \| \hat{\bbR}^T \bbs_{\ccalD^c} - \bbb \|_2 \leq \epsilon,$$ where $\bbb$ is, as in the proof of Theorem \[T:Recovery\], a binary vector with all its entries equal to 0 except for the last one that is equal to 1. Notice that takes the form of a basis pursuit problem with noisy observations [@ChenDonohoBP]. Expressions and can be derived by applying the second claim in [@zhang2013one Theorem 2] to problem . In order to do so, a few factors must be taken into consideration. First, since $\hat{\bbR}^T$ is not full row rank (since it is a tall matrix), constant $C_3$ depends on the $\ell_2$ norm of $\hat{\bbR}^{\dag}$. Moreover, in order to make constants $C_1$, $C_2$, and $C_3$ independent of the dual certificate $\bby \in \reals^{|\ccalD^c|}$ – see condition b) within the proof of Theorem \[T:Recovery\] – we have used that $\| \bby \|_2 \leq N $ and $\| \bby_{\ccalK^c} \|_\infty \leq \psi_{\hat{\bbR}}$, where the first one follows from the fact that the magnitude of every element in $\bby$ is at most $1$ and the second one was shown after .
A similar procedure can be used to show the result pertaining the case where $\ccalS \!= \! \tilde{\ccalS}_{\rm L}$.
When given noisy versions $\hat{\bbV}$ of the spectral templates of our target GSO, Proposition \[P:noisy\_recovery\] quantifies the effect that the noise has on the recovery. More precisely, the recovered shift is guaranteed to be at a maximum distance from the desired shift bounded by the tolerance $\epsilon$ times a constant, which depends on $\hbR$ and the support $\ccalK$. In particular, this implies that as the number of observed signals increases we recover the true graph shift as stated in the following remark.
\[R:consistent\_estimator\] As the number of observed signals increases the sample covariance $\hbC_x$ tends to the covariance $\bbC_x$ and, for the cases where the latter has no repeated eigenvalues, the noisy eigenvectors $\hbV$ tend to the eigenvectors $\bbV$ of the desired shift; see, e.g., [@Ortega90 Theo. 3.3.7]. In particular, with better estimates $\hat{\bbV}$ the tolerance $\epsilon$ in needed to guarantee feasibility can be made smaller, entailing a smaller discrepancy between the recovered $\bbS_1^*$ and the sparsest shift $\bbS_0^*$. In the limit when $\hat{\bbV} = \bbV$ and under no additional uncertainties, the tolerance $\epsilon$ can be made zero and guarantees perfect recovery under conditions *A-1)* and *A-2)* in Theorem \[T:Recovery\] or *L-1)* and *L-2)* in Theorem \[T:Recovery\_Laplacian\].
Incomplete spectral templates {#Ss:ShiftRecIncompEig}
-----------------------------
Thus far we have assumed that the entire set of eigenvectors $\bbV= [\bbv_1, \ldots, \bbv_N]$ is known, either perfectly or corrupted by noise. However, it is conceivable that in a number of scenarios only some of the eigenvectors (say $K$ out of $N$) are available. This would be the case when e.g., $\bbV$ is found as the eigenbasis of $\bbC_x$ and the given signal ensemble is bandlimited. More generally, whenever $\bbC_x$ contains repeated eigenvalues there is a rotation ambiguity in the definition of the associated eigenvectors. Hence, in this case, we keep the eigenvectors that can be unambiguously characterized and, for the eigenvectors with repeated eigenvalues, we include the rotation ambiguity as an additional constraint in our optimization problem.
Formally, assume that the $K$ first eigenvectors $\bbV_K=[\bbv_1,...,\bbv_K]$ are those which are known. Then, the network topology inference problem with incomplete spectral templates can be formulated as \[cf. \] $$\begin{aligned}
\label{E:SparseAdj_l00_onlysomeeig}
\bar{\bbS}^*_1& := \argmin_{ \{\bbS, \bbS_{\bar{K}}, \bblambda\}} \;\; \| \bbS \|_1 \\
& \text{s. to } \;\; \bbS= \bbS_{\bar{K}} + {\textstyle\sum_{k=1}^{K}} \lambda_k \bbv_k \bbv_k^T, \;\; \bbS\in\ccalS, \;\; \bbS_{\bar{K}}\bbV_K=\bb0, \nonumber\end{aligned}$$ where we already particularized the objective to the $\ell_1$ convex relaxation. The formulation in enforces $\bbS$ to be partially diagonalized by the known spectral templates $\bbV_K$, while its remaining component $\bbS_{\bar{K}}$ is forced to belong to the orthogonal complement of $\text{range}(\bbV_K)$. Notice that, as a consequence, the rank of $\bbS_{\bar{K}}$ is at most $N-K$. As in the previous cases, $\ccalS$ incorporates a priori information about the GSO. Notice that the constraint in $\ccalS$ imposing symmetry on $\bbS$ combined with the first constraint in automatically enforce symmetry on $\bbS_{\bar{K}}$, as wanted. An advantage of using only partial information of the eigenbasis as opposed to the whole $\bbV$ is that the set of feasible solutions in is larger than that in . This is particularly important when the templates do not come from a preexisting shift but, rather, one has the freedom to choose $\bbS$ provided it satisfies certain spectral properties. A practical example is the selection of the topology of a sensor network aimed at implementing estimation tasks such as consensus averaging, which can be oftentimes written as rank-one transformations of the sensor observations (cf. Section \[Ss:remark\_templates\] and [@segarra2015graphfilteringTSP15]).
Theoretical guarantees of recovery analogous to those presented in Section \[Ss:relaxation\] can be derived for . To formally state these, the following notation must be introduced. Define $\bbW_K := \bbV_K \odot \bbV_K$ and $\bbUpsilon := [\bbI_{N^2}, \bbzero_{N^2 \times N^2}]$. Also, define matrices $\bbB^{(i,j)} \in \reals^{N \times N}$ for $i < j$ such that $B^{(i,j)}_{ij} = 1$, $B^{(i,j)}_{ji} = -1$, and all other entries are zero. Based on this, we denote by $\bbB \in \reals^{{N \choose 2} \times N^2}$ a matrix whose rows are the vectorized forms of $\bbB^{(i,j)}$ for all $i,j \in \{1, 2, \ldots, N\}$ where $i < j$. In this way, $\bbB \bbs = \bbzero$ when $\bbs$ is the vectorized form of a symmetric matrix. Further, we define the following matrices $$\label{E:def_P_matrices}
\bbP_1 \!:=\!
\begin{bmatrix}
\bbI - \bbW_K \bbW_K^\dag \\
\bbI_\ccalD \\
\bbB \\
\bbzero_{NK \times N^2} \\
(\bbe_1 \otimes \bbone_N)^T
\end{bmatrix}^T, \quad
\bbP_2 \!:=\!
\begin{bmatrix}
\bbW_K \bbW_K^\dag - \bbI \\
\bbzero_{N \times N^2} \\
\bbzero_{{N \choose 2} \times N^2} \\
\bbI \otimes V_K^T \\
\bbzero_{1 \times N^2}
\end{bmatrix}^T,$$ and $\bbP := [\bbP_1^T , \bbP_2^T]^T$. With this notation in place, and denoting by $\ccalJ$ the support of $\bbs_0^* = \mathrm{vec}(\bbS^*_0)$, the following result holds.
\[T:Recovery\_incomplete\] Whenever $\ccalS = \ccalS_{\rm A}$ and assuming problem is feasible, $\bar{\bbS}^*_1 = \bbS^*_0$ if the two following conditions are satisfied:\
A-1) $\rank([{\bbP_1}_\ccalJ^T, \bbP_2^T]) = |\ccalJ| + N^2$; and\
A-2) There exists a constant $\delta > 0$ such that $$\label{E:condition_recovery_incomplete}
\eta_{\bbP} := \| \bbUpsilon_{\ccalJ^c}(\delta^{-2} \bbP \bbP^T + \bbUpsilon_{\ccalJ^c}^T \bbUpsilon_{\ccalJ^c})^{-1}\bbUpsilon_{\ccalJ}^T \|_{M(\infty)}<1.$$
With $\bbs = \mathrm{vec}(\bbS)$ and $\bbs_{\bar{K}} = \mathrm{vec}(\bbS_{\bar{K}})$, we reformulate for $\ccalS = \ccalS_{\rm A}$ as $$\begin{aligned}
\label{E:proof_incomplete_recovery_010}
\min_{ \{\bbs, \bbs_{\bar{K}}, \bblambda\}} & \;\; \| \bbs \|_1 \\
\text{s. to } \;\; &\bbs= \bbs_{\bar{K}} + \bbW_K \bblambda, \; \bbs_{\ccalD} = \bbzero, \; \bbB \bbs = \bbzero, \nonumber \\
& (\bbe_1 \otimes \bbone_N)^T \bbs= 1, \; (\bbI \otimes \bbV_K^T)\bbs_{\bar{K}} = \bbzero. \nonumber
\end{aligned}$$ The first and last constraints in correspond to the first and last constraints in written in vector form. The second constraint in imposes that $\bbS$ has no self-loops, the third one imposes symmetry on $\bbS$, and the fourth one normalizes the first column of $\bbS$ to sum up to 1 \[cf. \]. Notice that the non-negativity constraint in $\ccalS_{\rm A}$ is ignored in ; however, if we show that can recover the sparse solution $\bbs_0^*$, then the same solution would be recovered by the more constrained problem . Using the first constraint to solve for $\bblambda$, we obtain $\bblambda = \bbW_K^\dag (\bbs - \bbs_{\bar{K}})$. Moreover, defining the concatenated variable $\bbt:= [ \bbs^T, \bbs_{\bar{K}}^T]^T$, it follows from the definitions of $\bbUpsilon$ and $\bbP$ that can be reformulated as $$\begin{aligned}
\label{E:proof_incomplete_recovery_020}
\min_{ \bbt } \;\; \| \bbUpsilon \bbt \|_1 \quad \; \text{s. to } \; \bbP^T \bbt = \bbb,
\end{aligned}$$ where $\bbb$ is a vector with every entry equal to $0$ except for the last one which is equal to $1$. We utilize existing results on $\ell_1$-analysis [@zhang2013one] to state that the solution to coincides with the sparsest solution if:
- $\mathrm{ker}(\bbUpsilon_{\ccalJ^c}) \cap \mathrm{ker}(\bbP^T) = \{ \mathbf{0} \}$; and
- There exists a vector $\bby \in \reals^{N^2}$ such that $\bbUpsilon^T \bby \in \mathrm{Im}(\bbP)$, $\bby_\ccalJ = \mathrm{sign}({\bbs_0^*}_{\ccalJ})$, and $\|\bby_{\ccalJ^c}\|_\infty < 1$.
\[F:num\_exp\]
As was the case for Theorem \[T:Recovery\], the proof now reduces to showing that conditions *A-1)* and *A-2)* in the statement of the theorem imply the above conditions a) and b).
From the specific form of $\bbUpsilon$, the kernel of $\bbUpsilon_{\ccalJ^c}$ is a space of dimension $|\ccalJ| + N^2$ spanned by the set of canonical basis vectors $\bbe_i$ of length $2 N^2$ where $i \in \ccalJ \cup \{N^2+1, N^2+2, \ldots, 2 N^2\}$. Thus for a) to hold we need the matrix formed by the columns of $\bbP^T$ indexed by $\ccalJ \cup \{N^2+1, \ldots, 2 N^2\}$ to be full column rank, as stated in condition *A-1)*.
Finally, the procedure to show that *A-2)* implies b) follows the same steps as those detailed in the proof of Theorem \[T:Recovery\] – from onwards – and, thus, is omitted here.
Theorem \[T:Recovery\_incomplete\] provides sufficient conditions for the relaxed problem in to recover the sparsest graph, even when incomplete information about the eigenvectors is available. In practice it is observed that for smaller number $K$ of known spectral templates the value of $\eta_{\bbP}$ in tends to be larger, indicating a less favorable setting for recovery. This observation is aligned with the results obtained in practice; see Fig. \[F:num\_exp\_2\](c).
To state results similar to those in Theorem \[T:Recovery\_incomplete\] but for the recovery of normalized Laplacians, we define $\tilde{\bbU}_K := \tilde{\bbV}_K \odot \tilde{\bbV}_K$ where $ \tilde{\bbV}_K := [\bbv_2, \bbv_3, \ldots, \bbv_K]$ and define the matrices $$\label{E:def_T_matrices}
\bbT_1 :=
\begin{bmatrix}
\bbI - \tilde{\bbU}_K \tilde{\bbU}_K^\dag \\
\bbI_\ccalD \\
\bbB \\
\bbzero_{NK \times N^2} \\
\end{bmatrix}^T, \quad
\bbT_2 :=
\begin{bmatrix}
\tilde{\bbU}_K \tilde{\bbU}_K^\dag - \bbI \\
\bbzero_{N \times N^2} \\
\bbzero_{{N \choose 2} \times N^2} \\
\bbI \otimes V_K^T \\
\end{bmatrix}^T,$$ and $\bbT := [\bbT_1^T , \bbT_2^T]^T$. Under the assumption that the first eigenvector (i.e., the one whose associated eigenvalue is zero) is among the $K$ eigenvectors known, the following result holds. The proof – here omitted – follows the same steps as those in Theorem \[T:Recovery\_incomplete\].
\[T:Recovery\_incomplete\_Laplacian\] Whenever $\ccalS = \ccalS_{\rm L}$ and assuming problem is feasible, $\bar{\bbS}^*_1 = \bbS^*_0$ if the two following conditions are satisfied:\
L-1) $\rank([{\bbT_1}_\ccalJ^T, \bbT_2^T]) = |\ccalJ| + N^2$ ; and\
L-2) There exists a constant $\delta > 0$ such that $$\label{E:condition_recovery_incomplete_laplacian}
\eta_{\bbT} := \| \bbUpsilon_{\ccalJ^c}(\delta^{-2} \bbT \bbT^T + \bbUpsilon_{\ccalJ^c}^T \bbUpsilon_{\ccalJ^c})^{-1}\bbUpsilon_{\ccalJ}^T \|_{M(\infty)}<1.$$
Notice that scenarios that combine the settings in Sections \[Ss:ShiftRecImperfEig\] and \[Ss:ShiftRecIncompEig\], i.e. where the knowledge of the $K$ templates is imperfect, can be handled by combining the formulations in and . This can be achieved upon implementing the following modifications to : considering the shift $\bbS'$ as a new optimization variable, replacing the first constraint in with $\bbS'=\bbS_{\bar{K}}+{\textstyle \sum_{k=1}^{K}} \lambda_k \bbv_k \bbv_k^T$, and adding $d(\bbS,\bbS') \leq \epsilon$ as a new constraint \[cf. \].
Numerical experiments {#S:Simulations}
=====================
We test the proposed topology inference methods on different synthetic and real-world graphs. A comprehensive performance evaluation is carried out whereby we: (i) investigate the recovery of both adjacency and normalized Laplacian matrices; (ii) corroborate our main theoretical findings; (iii) assess the impact of imperfect information in the recovery; (iv) carry out comparisons with state-of-the-art methods; and (v) illustrate how our framework can be used to promote sparsity on a given network.
Topology inference from noiseless templates {#Ss:Sim_Top_inf_noiseless_templates}
-------------------------------------------
Consider Erdős-Rényi (ER) graphs [@bollobas1998random] of varying size $N \in \{10, 20, \dots, 50\}$ and different edge-formation probabilities $p \in \{0.1, 0.2, \ldots, 0.9\}$. For each combination of $N$ and $p$ we generate 100 graphs and try to recover their adjacency $\bbA$ and normalized Laplacian $\bbL$ matrices from the corresponding spectral templates $\bbV$. In Fig. \[F:num\_exp\](a) we plot the proportion of instances where the corresponding optimization problems – problem for $\ccalS = \ccalS_\mathrm{A}$ and $\ccalS = \ccalS_\mathrm{L}$ – have singleton feasibility sets. Notice that multiple solutions are more frequent when the expected number of neighbors of a given node is close to either $1$ or $N$. For intermediate values of $p$, the rank of both ${\bbW}_{\ccalD}$ and $\bbU$ is typically $N-1$, guaranteeing a single feasible point (cf. Proposition \[P:Feasibility\_singleton\]). Using the same set of graphs that those in Fig. \[F:num\_exp\](a), Fig. \[F:num\_exp\](b) shows the recovery rate when solving the iteratively re-weighted problem for both the adjacency (top) and the normalized Laplacian (bottom). As expected, the rates in Fig. \[F:num\_exp\](b) dominate those in Fig. \[F:num\_exp\](a) since every instance with a unique feasible point is recovered successfully. Moreover, the improved rates observed in Fig. \[F:num\_exp\](b) are indicative of the beneficial effect that the weighted $\ell_1$ norm objective has in the recovery.
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
\[F:num\_exp\_2\]
As indicated by Proposition \[P:Feasibility\_singleton\], the rate of recovery is intimately related to the ranks of ${\bbW}_{\ccalD}$ and $\bbU$ for the adjacency and normalized Laplacian cases, respectively. Fig. \[F:num\_exp\](c) further illustrates this relation via a histogram of the rank of $\bbU$ for the 100 graphs with $N=10$ and $p=0.2$. For more than half of the instances, the rank of $\bbU$ is equal to 9 (blue bar) and, as stated in Proposition \[P:Feasibility\_singleton\], for all these graphs there is a unique feasible point (yellow bar) that is successfully recovered (cyan bar). We see that, as the rank of $\bbU$ degrades, uniqueness is no longer guaranteed but for most cases the true graph can still be recovered following the iteratively re-weighted scheme proposed. Only in 8 of the cases where $\mathrm{rank}(\bbU)<9$ the recovery was not successful, entailing a recovery rate of 0.92, as reported in the corresponding entry ($N=10$, $p=0.2$) of the bottom plot in Fig. \[F:num\_exp\](b).
Finally, in order to corroborate the conditions for noiseless recovery stated in Theorem \[T:Recovery\], we draw ER random graphs of size $N \!=\! 20$ and edge-formation probability $p \! = \! 0.25$. For each graph, we make sure that the associated $\bbW_{\ccalD}$ matrix has rank strictly smaller than $N-1$ (to rule out the cases where the feasible set is a singleton), and that condition *A-1)* in Theorem \[T:Recovery\] is satisfied. In Fig. \[F:num\_exp\](d) we plot the number of successes and failures in recovering the adjacency as a function of $\psi_\bbR$ in . We consider 1000 realizations and for each of them the constant $\delta$ in is chosen to minimize $\psi_\bbR$. Fig. \[F:num\_exp\](d) clearly depicts the result of Theorem \[T:Recovery\] in that, for all cases where $\psi_\bbR<1$, relaxation achieves perfect recovery. Equally important, it is clear that the bound stated in is tight since a large proportion of the realizations with $\psi_\bbR$ equal to 1 or just above this value lead to failed recoveries.
Topology inference from noisy and incomplete templates {#Ss:Sim_Top_inf_noisy_templates}
------------------------------------------------------
We consider the identification of unweighted and undirected graphs corresponding to human brains [@hagmann2008mapping]. Each graph consists of $N=66$ nodes, which represent brain regions of interest (ROIs). An edge between two ROIs exists if the density of anatomical connections is greater than a threshold, which is chosen as the largest one that entails a connected graph [@hagmann2008mapping]. We test the recovery from noisy spectral templates $\hat{\bbV}$ \[cf. \] obtained from sample covariances of synthetic signals generated through diffusion processes (cf. Section \[S:prelim\_problem\]). These processes are modeled by graph filters of random degree between $3$ and $7$, and with independent and normally distributed coefficients. Denoting by $\hat{\bbV}_i$ the noisy spectral templates of patient $i \! \in \! \{1, 2, 3\}$ and by $\hat{\bbA}_i$ the adjacency matrices recovered, Fig. \[F:num\_exp\_2\](a) plots the recovery error as a function of the number of signals observed in the computation of the sample covariance. The error is quantified as the proportion of edges misidentified, i.e., $\|\bbA_i \!-\! \hat{\bbA}_i\|_0 / \| \bbA_i \|_0$, and each point in Fig. \[F:num\_exp\_2\](a) is the average across 50 realizations. Notice that for an increasing number of observed signals we see a monotonous decrease in the recovery error. For example, when going from $10^4$ to $10^5$ observations the error is (approximately) divided by seven, when averaged across patients. This is reasonable since a larger number of observations gives rise to a more reliable estimate of the covariance matrix entailing less noisy spectral templates. Traditional methods like graphical lasso [@GLasso2008] fail to recover $\bbS$ from the sample covariance of filtered white signals. For example, when signals are generated using a filter of the form $\bbH = h_0 \bbI + h_1 \bbS$, graphical lasso performs significantly worse than the method based on spectral templates. More precisely, when $10^5$ signals are observed, the recovery error of graphical lasso averaged over 50 realizations and with optimal tuning parameters is 0.303, 0.350, and 0.270 for patients 1, 2, and 3, respectively. Such errors are between 5 and 50 times larger than those reported in Fig. \[F:num\_exp\_2\](a). Further comparisons of our method with graphical lasso and other existing alternatives are provided in Section \[Ss:Sim\_performance\_comparison\].
{width="44.00000%"}
\[F:comparison\_traditional\]
We repeat the previous experiment on four social networks defined on a common set of $N=32$ nodes, which represent students from the University of Ljubljana[^3]. Links for each of the networks capture different types of interactions among the students, and were built by asking each student to select a group of preferred college mates for different situations, e.g., to discuss a personal issue or to invite to a birthday party (see footnote 2 for further details). The considered graphs are unweighted and symmetric, and the edge between $i$ and $j$ exists if either student $i$ picked $j$ in the questionnaire or vice versa. As done for the brain graphs, we test the recovery performance for noisy spectral templates $\hat{\bbV}$ obtained from sample covariances. Fig. \[F:num\_exp\_2\](b) plots the reconstruction error as a function of the number of observed signals for the different networks studied. As was observed in Fig. \[F:num\_exp\_2\](a), we see a monotonous decrease in recovery error for all the analyzed networks.
------------------------- -- --------------- ------------ ----------- -- --------------- --------------- ----------- -- --------------- ------------ -----------
\[1pt\]
\[-6pt\] SpecTemp Kalofolias Dong etal SpecTemp Kalofolias Dong etal SpecTemp Kalofolias Dong etal
\[1pt\]
F-measure [**0.896**]{} 0.791 0.818 [**0.924**]{} 0.868 0.828 [**0.703**]{} 0.651 0.667
edge error [**0.108**]{} 0.152 0.168 [**0.071**]{} 0.149 0.177 [**0.276**]{} 0.318 0.332
degree error [**0.058**]{} 0.071 0.105 [**0.040**]{} 0.055 0.111 [**0.162**]{} 0.201 0.222
[**Barabási-Albert**]{}
F-measure [**0.926**]{} 0.855 0.873 [**0.945**]{} 0.845 0.894 [**0.814**]{} 0.732 0.798
edge error [**0.143**]{} 0.173 0.209 [**0.135**]{} 0.154 0.235 [**0.310**]{} 0.314 0.393
degree error [**0.108**]{} 0.124 0.169 0.109 [**0.092**]{} 0.188 [**0.240**]{} 0.244 0.282
------------------------- -- --------------- ------------ ----------- -- --------------- --------------- ----------- -- --------------- ------------ -----------
\[Tab:comparison\_methods\]
Finally, we illustrate the recovery performance in the presence of incomplete spectral templates by solving for the four networks in Fig. \[F:num\_exp\_2\](b). More specifically, in Fig. \[F:num\_exp\_2\](c) we plot the recovery error as a function of the number $K$ of eigenvectors available. Each point in the plot is the average across 50 realizations in which different $K$ eigenvectors were selected from the $N\!=\!32$ possible ones. As expected, the performance for all four networks improves with the number of spectral templates known. The performance improvement is sharp and precipitous going from a large error of over $0.85$ for three of the networks when 17 spectral templates are known to a perfect recovery for all the networks when 24 eigenvectors are given. Moreover, notice that network $4$ is consistently the easiest to identify both for noisy \[cf. Fig. \[F:num\_exp\_2\](b)\] and incomplete \[cf. Fig. \[F:num\_exp\_2\](c)\] spectral templates. For example, when given 19 spectral templates the error associated with network $4$ is $0.224$ whereas the average across the other three networks is $0.584$. This hints towards the fact that some graphs are inherently more robust for identification when given imperfect spectral templates. A formal analysis of this phenomenon is left as future work.
Performance comparison {#Ss:Sim_performance_comparison}
----------------------
We compare the performance of the presented method based on spectral templates (we refer to it as SpecTemp for conciseness) with established statistical approaches as well as recent GSP-based algorithms.
**Comparison with established methods.** We analyze the performance of SpecTemp in comparison with two widely used methods, namely, (thresholded) correlation [@kolaczyk2009book Ch. 7.3.1] and graphical lasso [@GLasso2008]. Our goal is to recover the adjacency matrix of an undirected and unweighted graph with no self-loops from the observation of filtered graph signals. For the implementation of SpecTemp, we use the eigendecomposition of the sample covariance of the observed signals in order to extract noisy spectral templates $\hat{\bbV}$. We then solve problem for $\ccalS = \ccalS_{\mathrm{A}}$, where $\epsilon$ is selected as the smallest value that admits a feasible solution. We include as a priori knowledge that each node has at least one neighbor. For the correlation-based method, we keep the absolute value of the sample correlation of the observed signals, force zeros on the diagonal and set all values below a certain threshold to zero. This threshold is determined during a training phase, as explained in more detail in the next paragraph. Lastly, for graphical lasso we follow the implementation in [@GLasso2008] based on the sample covariance and select the tuning parameter $\rho$ (see [@GLasso2008]) during the training phase. We then force zeros on the diagonal and keep the absolute values of each entry. Leveraging that the sought graphs are unweighted, for SpecTemp and graphical lasso a fixed threshold of 0.3 is used so that, after recovery, every edge with weight smaller than the threshold is set to zero.
![Comparison of edge recovery error as a function of the number of signals observed for SpecTemp, Kalofolias [@Kalofolias2016inference_smoothAISTATS16], and Dong etal[@DongLaplacianLearning].](figures/compare_Kalo_Dong_observations.pdf){width="41.00000%"}
\[F:comparison\_kalo\_dong\]
We test the recovery of adjacency matrices $\bbS \! = \! \bbA$ of ER graphs with $N \! = \! 20$ nodes and edge probability $p \! = \! 0.2$. We vary the number of observed signals from $10^1$ to $10^6$ in powers of $10$. Each of these signals is generated by passing white Gaussian noise through a graph filter $\bbH$. Two different types of filters are considered. As a first type we consider a *general* filter $\bbH_1 = \bbV \diag(\widehat{\bbh}_1) \bbV^T$, where the entries of $\widehat{\bbh}_1$ are independent and chosen randomly between $0.5$ and $1.5$. The second type is a *specific* filter of the form $\bbH_2 = (\delta \bbI + \bbS)^{-1/2}$, where the constant $\delta$ is chosen so that $\delta \bbI + \bbS$ is positive definite to ensure that $\bbH_2$ is real and well-defined. Following the discussion in Section \[S:prelim\_problem\], this implies that the precision matrix of the filtered signals is given by $\bbC_x^{-1} = \bbH_2^{-2} = \delta \bbI + \bbS$, which coincides with the desired GSO $\bbS$ in the off-diagonal elements. For each combination of filter type and number of observed signals, we generate 10 ER graphs that are used for training and 20 ER graphs that are used for testing. Based on the 10 training graphs, the optimal threshold for the correlation method and parameter $\rho$ for graphical lasso are determined and then used for the recovery of the 20 testing graphs. Given that for SpecTemp we are fixing $\epsilon$ beforehand, no training is required.
\[F:network\_sparsification\]
As figure of merit we use the F-measure [@manning2008introduction], i.e. the harmonic mean of edge precision and edge recall, that solely takes into account the support of the recovered graph while ignoring the weights. In Fig. \[F:comparison\_traditional\] we plot the performance of the three methods as a function of the number of filtered graph signals observed for filters $\bbH_1$ and $\bbH_2$, where each point is the mean F-measure over the 20 testing graphs.
When considering a general graph filter $\bbH_1$ SpecTemp clearly outperforms the other two. For instance, when $10^5$ signals are observed, our average F-measure is $0.81$ while the measures for correlation and graphical lasso are $0.29$ and $0.25$, respectively. Moreover, of the three methods, our approach is the only consistent one, i.e., achieving perfect recovery with increasing number of observed signals. Although striking at a first glance, the deficient performance of graphical lasso was expected. For general filters $\bbH_1$, the precision matrix is given by $\bbC_x^{-1} = \bbH_1^{-2}$ which in general is neither sparse nor shares the support of $\bbS$, the GSO to be recovered. When analyzing the specific case of graph filters $\bbH_2$, where the precision matrix exactly coincides with the desired graph-shift operator, graphical lasso outperforms both our method and the correlation-based method. This is not surprising since graphical lasso was designed for the recovery of sparse precision matrices. Notice however that for large number of observations SpecTemp, without assuming any specific filter model, also achieves perfect recovery and yields an F-measure equal to 1.
**Comparison with GSP methods.** We compare the recovery using SpecTemp with the algorithms in [@DongLaplacianLearning] and [@Kalofolias2016inference_smoothAISTATS16], both methods designed to identify the (combinatorial) Laplacian of a graph when given a set of smooth graph signals. Small modifications can be made to our framework to accommodate this setting, thus permitting a fair comparison. More precisely, in solving we use the set of admissible shifts given by $$\begin{aligned}
\label{E:def_S_combinatorial_laplacian}
\ccalS_{\rm L_c} := \{ \bbS \, | \, S_{ij} \leq 0 \,\, \text{for} \,\,& i\!\neq\! j, \;\;\bbS\! \in \! \ccalM_{+}^N, \;\; \bbS \bbone = \bbzero \}.\end{aligned}$$ Moreover, in order to account for the smoothness of the observed signals in the unknown graph we sort the eigenvectors $\hbv_k$ of the sample covariance in increasing order of their corresponding eigenvalues, and we require the recovered eigenvalues $\bblambda$ to satisfy $\lambda_{i} \geq \lambda_{i+k} + \delta$ for all $i$, and fixed $k$ and $\delta$. In this way, we assign the frequencies with larger presence in the observed signals to low eigenvalues in the recovered Laplacian. Unless otherwise noted, we set $\delta = 0.1$ and $k = 3$.
We compare the three methods of interest on two different types of graphs and three different signal generation models. We consider the recovery of the Laplacian $\bbS = \bbL$ of ER graphs with $N \! = \! 20$ nodes and edge probability $p \! = \! 0.3$ as well as Barabási-Albert preferential attachment graphs [@bollobas1998random] with $N \! = \! 20$ generated from $m_0 = 4$ initially placed nodes, where each new node is connected to $m=3$ existing ones. Following [@Kalofolias2016inference_smoothAISTATS16] we consider three models for smooth graph signals: i) multivariate normal signals with covariance given by the pseudo-inverse of $\bbL$, i.e., $\bbx_1 \sim \ccalN(\bbzero, \bbL^{\dag})$; ii) white signals filtered through an autoregressive (diffusion) process, that is $\bbx_2 = (\bbI + \bbL)^{-1} \bbw$, where $\bbw\sim \ccalN(\bbzero, \bbI)$; and iii) white signals passed through an exponential filter, $\bbx_3 = \exp(-\bbL) \bbw$. For each of the six settings considered (two graphs combined with three signal types) we generate 10 training graphs, 100 testing graphs, and for every graph we generate 1000 graph signals. The training set is used to set the parameters in [@DongLaplacianLearning] and [@Kalofolias2016inference_smoothAISTATS16], and in our case it serves the purpose of selecting the best $\epsilon$ \[cf. \]. To increase the difficulty of the recovery task, every signal $\bbx$ is perturbed as $\hbx = \bbx + \sigma \, \bbx \circ \bbz$, for $\sigma = 0.1$ and where each entry of $\bbz$ is an independent standard normal random variable. We focus on three performance measures, namely, the F-measure as explained in the previous experiment, the $\ell_2$ relative error of recovery of the edges, and the $\ell_2$ relative error of recovery of the degrees. The performance achieved by each method in the testing sets is summarized in Table \[Tab:comparison\_methods\]. In all but one case, our method attains the largest F-measures and the smallest errors for all the graphs and signal types considered.
Finally, for the particular cases of ER graphs and signals $\bbx_1$ (inverse Laplacian), we replicate the above procedure varying the number of observed signals $P$ from $100$ to $1000$. For SpecTemp, we increase $k$ when the number of observations decreases to account for the noisier ordering of the eigenvectors in the sample covariance. In this experiment we use $k=5$ for $P \leq 400$, $k=4$ for $400<P<800$ and $k=3$ for $P \geq 800$. In Fig. \[F:comparison\_kalo\_dong\] we plot the associated $\ell_2$ edge recovery errors. Notice that for small number of observations, the method in [@Kalofolias2016inference_smoothAISTATS16] outperforms SpecTemp whereas the opposite is true when more signals are observed. This can be attributed to the fact that SpecTemp assumes no specific model on the smoothness of the signal, thus, when enough signals are observed our more agnostic, data-driven approach exhibits a clear performance advantage.
Network sparsification {#Ss:Sim_network_sparsification}
----------------------
With reference to the network sparsification problem outlined in Section \[sssec\_sparsify\], our goal here is to identify the structural properties of proteins from a mutual information graph of the co-variation between the constitutional amino-acids [@Marks2011proteins]; see [@FeiziNetworkDeconvolution] for details. For example, for a particular protein, we want to recover the structural graph in the top left of Fig. \[F:network\_sparsification\](a) when given the graph of mutual information in the top right corner. Notice that the structural contacts along the first four sub-diagonals of the graphs were intentionally removed to assess the capability of the methods in detecting the contacts between distant amino-acids. The graph recovered by network deconvolution [@FeiziNetworkDeconvolution] is illustrated in the bottom left corner of Fig. \[F:network\_sparsification\](a) whereas the one recovered using SpecTemp is depicted in the bottom right corner of the figure. Comparing both recovered graphs, SpecTemp leads to a sparser graph that follows more closely the desired structure to be recovered. To quantify this latter assertion, in Fig. \[F:network\_sparsification\](b) we plot the fraction of the real contact edges recovered for each method as a function of the number of edges considered, as done in [@FeiziNetworkDeconvolution]. For example, if for a given method the 100 edges with largest weight in the recovered graph contain $40\%$ of the edges in the ground truth graph we say that the 100 top edge predictions achieve a fraction of recovered edges equal to $0.4$. As claimed in [@FeiziNetworkDeconvolution], network deconvolution improves the estimation when compared to raw mutual information data. Nevertheless, from Fig. \[F:network\_sparsification\](b) it follows that SpecTemp outperforms network deconvolution. Notice that when $\epsilon = 0$ \[cf. \] we are forcing the eigenvectors of $\bbS$ to coincide exactly with those of the matrix of mutual information $\bbS'$. However, since $\bbS'$ is already a valid adjacency matrix, we end up recovering $\bbS = \bbS'$. By contrast, for larger values of $\epsilon$ the additional flexibility in the choice of the eigenvectors allows us to recover shifts $\bbS$ that more closely resemble the ground truth. For example, when considering the top $200$ edges, the mutual information and the network deconvolution methods recover $36\%$ and $43\%$ of the desired edges, respectively, while our method for $\epsilon \! = \! 1$ achieves a recovery of $53\%$. In Fig. \[F:network\_sparsification\](c) we present this same analysis for a different protein and similar results can be appreciated.
Conclusions {#S:Conclusions}
===========
With $\bbS=\bbV\bbLambda\bbV^T$ being the shift operator associated with the graph $\ccalG$, we studied the problem of identifying $\bbS$ (hence the topology of $\ccalG$) using a two-step approach under which we first obtain $\bbV$, and then use $\bbV$ as input to find $\bbLambda$. The problem of finding $\bbLambda$ given $\bbV$ was formulated as a sparse recovery optimization. Efficient algorithms based on convex relaxations were developed, and theoretical conditions under which exact and robust recovery is guaranteed were derived for the cases where $\bbS$ represents the adjacency or the normalized Laplacian of $\ccalG$. In identifying $\bbV$, our main focus was on using as input a set of graph signal realizations. Under the assumption that such signals resulted from diffusion dynamics on the graph or, equivalently, that they were stationary in $\bbS$, it was shown that $\bbV$ could be estimated from the eigenvectors of the sample covariance of the available set. As a consequence, several well-established methods for topology identification based on local and partial correlations can be viewed as particular instances of the approach here presented. The practical relevance of the proposed schemes and the gains relative to existing alternatives were highlighted carrying out numerical tests with synthetic and real-world graphs.
[^1]: Work in this paper is supported by Spanish MINECO grant No TEC2013- 41604-R and USA NSF CCF-1217963. S. Segarra and A. Ribeiro are with the Dept. of Electrical and Systems Eng., Univ. of Pennsylvania. A. G. Marques is with the Dept. of Signal Theory and Comms., King Juan Carlos Univ. G. Mateos is with the Dept. of Electrical and Computer Eng., Univ. of Rochester. Emails: [email protected], [email protected], [email protected] and [email protected]. Part of the results in this paper were submitted to the *2016 IEEE SSP Workshop* [@SSAMGMAR_ssp16] and the *2016 Asilomar Conference on Signals, Systems and Computers* [@SSAMGMAR_Asilomar16].
[^2]: To simplify exposition, the general description of the recovery problem in this section assumes that neither $\bbS$ nor $\bbC_x$ have repeated eigenvalues. Technical modifications in the formulation to accommodate setups where the eigenvalues are not all distinct are discussed in Section \[Ss:ShiftRecIncompEig\].
[^3]: Access to the data and additional details are available at <http://vladowiki.fmf.uni-lj.si/doku.php?id=pajek:data:pajek:students>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Heterogeneous face recognition (HFR) refers to matching face images acquired from different domains with wide applications in security scenarios. This paper presents a deep neural network approach namely Multi-Margin based Decorrelation Learning (MMDL) to extract decorrelation representations in a hyperspherical space for cross-domain face images. The proposed framework can be divided into two components: heterogeneous representation network and decorrelation representation learning. First, we employ a large scale of accessible visual face images to train heterogeneous representation network. The decorrelation layer projects the output of the first component into decorrelation latent subspace and obtains decorrelation representation. In addition, we design a multi-margin loss (MML), which consists of quadruplet margin loss (QML) and heterogeneous angular margin loss (HAML), to constrain the proposed framework. Experimental results on two challenging heterogeneous face databases show that our approach achieves superior performance on both verification and recognition tasks, comparing with state-of-the-art methods.'
author:
- |
Bing Cao\
School of Electronic Engineering\
Xidian University\
Nannan Wang\
School of Telecommunications Engineering\
Xidian University\
Xinbo Gao\
School of Electronic Engineering\
Xidian University\
Jie Li\
School of Electronic Engineering\
Xidian University\
Zhifeng Li\
Tencent AI Lab\
bibliography:
- 'references.bib'
title: 'Multi-Margin based Decorrelation Learning for Heterogeneous Face Recognition'
---
Introduction
============
Heterogeneous face images refer to facial images acquired from different domains, such as visual (VIS) photo, near infrared (NIR) image, thermal infrared image, sketch and images with different resolutions, etc. In recent years, a great deal of efforts [@Tecent3] have been taken to heterogeneous face recognition, i.e. matching VIS face photos with cross-domain face images [@Tecent5]. However, different from impressive progress made in traditional face recognition, it is still a challenging problem for HFR.
Among these HFR scenarios [@cao2018asymmetric], matching VIS face photos with NIR face images is the most straightforward and efficient solution to handle the extreme lighting conditions, which is of wide applications ranging from personal authorization to law enforcement. The lacking of sufficient training data in different domains and the significant cross-domain discrepancy are two most obstacles to train a robust model for NIR-VIS face recognition.
During the last decade, many large-scale VIS face datasets are available, which provide sufficient data to train convolution neural networks (CNN) for traditional face recognition and enormously improve the recognition performance. However, all the NIR face datasets are in small scale, which is not sufficient to train an effective CNN model without overfitting.
To address the other obstacle of significant cross-domain discrepancy, existing methods can be grouped into three categories. Synthesis-based methods transform the cross-domain face images to the same domain. Feature-based methods learn the invariant feature representation for the same identity in different domains. Subspace-based methods project the cross-domain face images to the same domain for HFR. However, these methods can not remove the discrepancy completely and the accuracy is not satisfactory.
In this paper, we propose a novel Multi-Margin based Decorrelation Learning (MMDL) framework to tackle the two aforementioned obstacles. The proposed framework contains two components: heterogeneous representation network and decorrelation representation learning. For the first obstacle, we employ a large scale of accessible visual face images to train heterogeneous representation network, which consists of input layers, output layers and four residual groups. We utilize this network to extract feature representation in a hyperspherical space that is robust to intra-class invariance and inter-class variance in VIS domain. Then, transfer learning is employed to improve the adaptation of this network to cross-domain face images. We impose a decorrelation layer to this network and design a multi-margin loss (MML) to fine-tune it. MML is utilized to minimize the cross-domain intra-class distance and further maximize cross-domain inter-class distance in the hyperspherical space. MML consists of the quadruplet margin loss (QML) and the heterogeneous angular margin loss (HAML).
Our main contributions are summarized as follows:
- We design an effective end-to-end framework to extract invariant representation for both NIR and VIS face images. A decorrelation layer is imposed to heterogeneous representation networks to estimate the decorrelation latent subspace, which results in the two networks share the same parameters.
- The multi-margin loss is designed to constrain the proposed framework. MML contains two components: QML and HAML, which are effective to minimize cross-domain intra-class distance and further maximize cross-domain inter-class distance.
- We propose an alternative optimization to fine-tune the heterogeneous representation networks and decorrelation representation learning, which improves the performance of the proposed framework.
- Experimental results on two challenging HFR databases illustrate that the proposed framework achieves superior performance, comparing with state-of-the-art methods. In addition, we conduct ablation study to demonstrate the effectiveness of various parts of the proposed approach.
Related Work
============
Matching NIR-VIS face images has become an important challenge in biometrics identification and great efforts have been made by researchers to solve this problem in the past decade. Existing HFR methods can be mainly grouped into three categories: synthesis-based methods, feature-based methods and subspace-based methods.
Synthesis-based methods try to synthesize heterogeneous face images from source domain to target domain and compare them in the same domain [@Tecent6]. These methods are designed to reduce the discrepancy between heterogeneous images in pixel-level[@cao2020; @Tecent7]. [@Cao2018TNNLS] took multiple synthesized pseudo face images to improve the recognition accuracy. Generative adversarial network (GAN) was employed to synthesize heterogeneous face images in ADFL [@Song2018Adversarial]. Though, synthesized methods can reduce the discrepancy between heterogeneous face images in pixel-level, the discriminative details are lost seriously, which affects the final recognition performance.
Feature-based methods aim at extracting invariant feature representation for heterogeneous face images [@Tecent4]. These methods are designed to reduce the cross-domain discrepancy in feature-level. To alleviate overfitting, a Coupled Deep Learning (CDL) [@wu2017coupled] introduced nuclear norm constraint on fully connected layer and proposed a cross-modal ranking to reduce cross-domain discrepancy. [@he2018wasserstein] utilized Wasserstein distance to decrease the domain gap and acquire cross-domain invariant representation. Due to the great discrepancy between heterogeneous face images, it is hard to extract cross-domain invariant feature representation.
Subspace-based methods attempt to minimize the cross-domain discrepancy by project cross-domain face features onto a common subspace. In this subspace, heterogeneous face images can be measured directly. [@yi2015shared] employed Restricted Boltzmann Machines (RBMs) to learn the cross-domain representation for NIR and VIS face images. The relationship of cross-domain face images was employed in [@kan2016multi] to develop a multi-view discriminant analysis (MvDA) for HFR. However, it is inevitable to loss valid information in the process of projection, which seriously affects the final performance. Different from existing methods, our MMDL framework takes advantages of both feature-based methods and subspace-based methods.
Proposed Methods
================
Our framework contains two key components: heterogeneous representation networks and decorrelation representation learning as shown in Figure. \[framework\]. The first component extracts the low-dimensional feature representation of NIR and VIS face images. Then, the second component reduces the correlation of these low-dimensional representations and obtain the decorrelation representations that can be measured by cosine distance. In addition, we design the multi-margin loss (MML), which contains quadruplet margin loss (QML) and heterogeneous angular margin loss (HAML), to optimize the proposed framework. In this section, we will detail the proposed multi-margin based decorrelation learning (MMDL) framework and the corresponding optimization scheme.
Decorrelation Representation Learning
-------------------------------------
Let $\Phi$ denotes heterogeneous representation network. For heterogeneous face images, different samples of the same identity share the same invariant feature representation. The proposed network aims at extracting the invariant feature representation. Thus, the parameters $W^{H}$ of $\Phi$ are learned from both NIR and VIS face images. For a NIR image $x^N$ and VIS image $x^V$, the feature representations $y^i\in\mathbb{R}^n\ (i\in\{N,V\})$ that are extracted from $\Phi$ can be denoted as $$\label{eq:1}
\begin{aligned}
y^i = \Phi(x^i,W^{H})\quad (i\in\{N,V\}), \\
\end{aligned}$$ where $\Phi(\cdot)$ is the forward computation process of heterogeneous representation network. $H$ denotes the heterogeneous representation network. $N$ and $V$ represent the NIR domain and VIS domain respectively.
### Decorrelation Representation
![An illustration of the proposed MMDL framework. The heterogeneous representation network is used to extract the feature representation of the input cross-domain face images, the decorrelation layer is utilized to reduce the correlation of variations between the feature representation extracted by heterogenous representation network. With the online quadruplet samples selection, the quadruplet margin loss (QML) is computed by the decorrelation representation. The heterogeneous angular margin loss is introduced to joint optimize the proposed framework with QML. We can directly measure the similarity by decorrelation representation.[]{data-label="framework"}](framework){width="1\columnwidth"}
As demonstrated in previous work [@chen2017efficient], a face image $x$ can be represented by identity information and variations that contains lighting, pose and expression. For HFR, spectrum information is also a kind of variations. As these variations of different samples are correlated [@he2018wasserstein], it is hard to learn a discriminative model and achieve satisfactory performance on matching NIR and VIS face images. Therefore, we impose a decorrelation layer $D$ to heterogeneous representation network $\Phi$, in order to project feature representations $y^i\ (i\in\{N,V\})$ to the decorrelation latent subspace. Then, the outputs of decorrelation layer are the decorrelation representations of heterogeneous face images, which can be denoted as follows: $$\label{eq:2}
\begin{aligned}
z^i = (W^{D})^Ty^i\quad (i\in\{N,V\}),
\end{aligned}$$ where $z^i\in\mathbb{R}^q\ (i\in\{N,V\})$ denotes the decorrelation representations of NIR and VIS face images respectively. $W^{D}\in\mathbb{R}^{n\times q}$ represents the parameters of decorrelation layer and $D$ denotes the decorrelation layer. Therefore, we turn the decorrelation representation issue into estimating the parameter $W^{D}$ of decorrelation layer. For training data $X^i=\{x^i_1,x^i_2,\cdots,x^i_m\}\ (i\in\{N,V\})$, the feature representations $Y^i=\{\mathbf{\emph{y}}^i_1,\mathbf{\emph{y}}^i_2,\cdots,\mathbf{\emph{y}}^i_m\}\ (i\in\{N,V\})$ and decorrelation representations $Z^i=\{\mathbf{\emph{z}}^i_1,\mathbf{\emph{z}}^i_2,\cdots,\mathbf{\emph{z}}^i_m\}\ (i\in\{N,V\})$ can be obtained by feedforward computation.
Suppose that, the correlation of decorrelation representations is minimized, $Z^i_{j}=\{\mathbf{\emph{z}}^i_{j,1},\mathbf{\emph{z}}^i_{j,2}, \cdots, \mathbf{\emph{z}}^i_{j,q}\}$ should be in the coordinates with standard orthogonal basis $\{\mathbf{\emph{w}}_k,\mathbf{\emph{w}}_2,\cdots,\mathbf{\emph{w}}_q\}$, and $z^i_{j,k} = \mathbf{\emph{w}}_k^T\mathbf{\emph{y}}^i_j$. If we reconstruct $\mathbf{\emph{y}}^i_j$ by $\mathbf{\emph{z}}^i_j$, we can obtain $\hat{{\mathbf{\emph{y}}}}=\sum_{k=1}^{q}{z^i_{j,k}\mathbf{\emph{w}}_k}$. Meanwhile, as we utilize cosine distance to measure the similarity of two input cross-domain face images, the angle $\theta$ between $\hat{\mathbf{\emph{y}}}^i$ and $\hat{\mathbf{\emph{y}}}^i$ should be minimize in order to avoid discriminative information loss. Considering the properties of Cosine function, our objective function is as follows: $$\begin{aligned}\label{eq:3}
\max\sum_{j=1}^m\cos\theta_j & =\sum_{j=1}^m\frac{\mathbf{\emph{y}}^i_j\cdot \hat{\mathbf{\emph{y}}}^i_j}{\parallel \mathbf{\emph{y}}^i_j\parallel \parallel \hat{\mathbf{\emph{y}}}^i_j\parallel}\quad (i\in\{N,V\}) \\
& \propto \max\sum_{j=1}^m(\cos\theta_j)^2
\end{aligned}$$ Therefore, our objective function turns into $$\label{eq:4}
\begin{aligned}
\min\sum_{j=1}^m -(\cos\theta_j)^2 & = \sum_{j=1}^m-(\frac{\mathbf{\emph{y}}^i_j\cdot \hat{\mathbf{\emph{y}}}^i_j}{\parallel \mathbf{\emph{y}}^i_j\parallel \parallel \hat{\mathbf{\emph{y}}}^i_j\parallel})^2 \\
& = \sum_{j=1}^m-\frac{(\mathbf{(\emph{y}}^i_j)^T\cdot \sum_{l=1}^{q}{z^i_{j,l}\mathbf{\emph{w}}_k})^2}{(\mathbf{\emph{y}}^i_j)^T\mathbf{\emph{y}}^i_j (\sum_{l=1}^{q}{z^i_{j,l}\mathbf{\emph{w}}_k})^2}\\
& = \sum_{j=1}^m -(\mathbf{\emph{w}}_k)^T(\sum_{l=1}^{q}\frac{\emph{y}^i_j(\emph{y}^i_j)^T}{(\emph{y}^i_j)^T\emph{y}^i_j})\mathbf{\emph{w}}_k \\
& \propto -tr(W^T\frac{YY^T}{Y^TY}W) \\
& s.t. \quad W^TW = I
\end{aligned}$$ We can obtain $W$ by the eigenvalue and eigenvector of $\frac{YY^T}{Y^TY}$ via the lagrangian multiplier, singular value decomposition and $z^i = W^Ty^i$. Therefore, the parameters $W^D$ of decorrelation layer can be estimated by the $W$.
### Multi-Margin Loss
To constrain the proposed framework, we proposed a multi-margin loss, which contains the quadruplet margin loss (QML) and the heterogeneous angular margin loss (HAML).
As demonstrated in [@schroff2015facenet], triplet loss is effective to improve the accuracy of traditional face recognition. However, different from traditional face recognition[@Tecent1; @Tecent2], HFR matches face images from different modalities[@Tecent2]. It is meaningless to constrain the distance of face images from the same domain. Therefore, the contribution of triplet loss to improve the performance on HFR is minimal. Considering the limitation of triplet loss, we propose quadruplet margin loss.
Quadruplet Margin Loss (QML) is designed to increase the distance between inter-class cross-domain images and decrease the distance between intra-class cross-domain images. In order to accelerate convergence, we also take within-domain negative pairs into consideration. The designed online quadruplet samples selection strategy is employed to select groups of four heterogeneous decorrelation representations $\{z^N_j, z^V_j, z^N_k, z^V_l\}$ in a mini-batch as quadruplet tuples, where $\{z^N_j, z^V_j\}$ shares the same identity. $z^N_k$ denotes the closest NIR representations to $z^V_j$ from another identity and $z^V_l$ denotes the closest VIS representations to $z^N_j$ from another identity. Due to we utilize cosine distance to measure the distance between different images. As shown in Figure. \[framework\]. The proposed QML can be computed as follows:
$$\label{eq:5}
\begin{aligned}
\mathcal{L}_{QML}(z^N_j, z^V_j, z^V_l) =& \sum_i^b(\frac{z^N_j\cdot z^V_j}{\|z^N_j\|\|z^V_j\|}-\frac{z^N_j\cdot z^V_l}{\|z^N_j\|\|z^V_l\|}+\alpha_1)\\
& + \sum_i^b(\frac{z^N_j\cdot z^V_j}{\|z^N_j\|\|z^V_j\|}-\frac{z^V_j\cdot z^V_l}{\|z^V_j\|\|z^V_l\|}+\alpha_2),
\end{aligned}$$
$$\begin{aligned}
\mathcal{L}_{QML}(z^N_j, z^V_j, z^N_k, z^V_l) = & \mathcal{L}_{QML}(z^N_j, z^V_j, z^V_l) \\&+ \mathcal{L}_{QML}(z^V_j, z^N_j, z^N_k),
\end{aligned}$$
where $b$ denotes the number of quadruplet tuples in a mini-batch. $\alpha_1$ and $\alpha_2$ are the quadruplet margin. As shown in Figure. \[framework\], QML is designed to decrease the cosine distance between intra-class cross-domain face images and increase the cosine distance between inter-class face images.
Heterogeneous angular margin loss (HAML) is inspired by [@liu2017sphereface] and [@deng2018arcface], which is developed from Softmax loss. Softmax loss is widely used in classification tasks, which is presented as follows: $$\begin{aligned}
\mathcal{L}_s = -\frac{1}{b}\sum^b_{j=1}\log\frac{e^{(W^F_{c_j})^Tz^i_j}}{\sum^n_{v=1}e^{(W^F_v)^Tz^i_j}}\quad (i\in\{N,V\}),
\end{aligned}$$ where $z_j$ belongs to the $c_j$-th class. $W^F_v$ denotes the $v$-th column vector of the weights $W^F\in \mathbb{R}^{q\times c}$ in the last fully connected layer. The number of class is $c$. The target logit [@pereyra2017regularizing] can be transformed to $$\begin{aligned}\label{eq:8}
(W^F_v)^Tz_j = \|W^F_v\|\|z^i_j\|cos\theta_v^i \quad (i\in\{N,V\}) ,
\end{aligned}$$ We fix $\|W^F_v\|=1$ and $\|z^i_j\|=s$ by L2 normalisation, where $s$ is a constant. As we omit these constants, Eq.\[eq:8\] can be reformulated as $(W^F_v)^Tz^i_j = cos\theta_v^i$. Then, all the feature representations are distributed in a hypersphere. The similarity of two face images is determined by the angle between the corresponding feature representations. According to [@deng2018arcface], an angular margin $m$ is added within $\cos \theta$, the HAML can be defined as follows:
$$\begin{aligned}
\mathcal{L}_{HAML} & = -\frac{\lambda_N}{b}\sum^b_{j=1}\log\frac{e^{s(\cos(\theta^N_{c_i}+m_1))}}{e^{s(\cos(\theta^N_{c_i}+m_1))} + \sum^n_{v=1,v\neq c_i}e^{s\cos\theta^N_v}} \\
& -\frac{\lambda_V}{b}\sum^b_{j=1}\log\frac{e^{s(\cos(\theta^V_{c_i}+m_2))}}{e^{s(\cos(\theta^V_{c_i}+m_2))} + \sum^n_{v=1,v\neq c_i}e^{s\cos\theta^V_v}} ,\\
& \qquad \qquad \qquad s.t. \quad \lambda_N + \lambda_V = 1,
\end{aligned}$$
where $\lambda_N$ and $\lambda_V$ represent the trade-off parameters of loss learned from NIR domain VIS domain. As the basic network is pre-trained by VIS face images, we assign greater weight to loss learned from NIR data. The multi-margin loss can be denoted as follows:
$$\label{eq:10}
\begin{aligned}
\mathcal{L}_{MML} = \lambda_1\mathcal{L}_{QML} + \lambda_2\mathcal{L}_{HAML}
\end{aligned}$$
where $\lambda_1$ and $\lambda_2$ are the trade-off parameters of quadruplet margin loss and heterogeneous angular margin loss.
Optimization
------------
An alternative optimization strategy for the proposed MMDL framework is introduced in this subsection. The parameter $W^H$ of heterogeneous representation network is pre-trained by large-scale VIS face images. First, we fix the parameter $W^H$ of heterogeneous representation network and extract the feature representation $Y$ of training data $X$ by Eq. \[eq:1\]. The parameter $W^D$ of decorrelation layer is estimated by feature representation $Y$ according to Eq. \[eq:3\] and Eq. \[eq:4\]. Second, we fix the parameter $W^D$ in the process of learning decorrelation representations $Z$ by Eq. \[eq:2\]. Then, we utilize $Z$ to compute multi-margin loss (MML) by Eq. \[eq:10\] to optimize the parameter $W^H$ of heterogeneous representation network. Finally, the parameter $W^H$ and $W^D$ are fixed, we can obtain the final feature representations of input cross-domain face images by the proposed framework and measure the similarity of them by cosine distance. We summarize the optimization details in Algorithm 1.
**Require**: Training NIR face images $x^N$, training VIS face images $x^V$, learning rate $r$, batch size $b$, the trade-off parameter $\lambda$.\
**Ensure**: The parameter $W^H$ of heterogeneous representation network and the parameter $W^D$ of decorrelation layer.
Pre-train the parameter $W^H$ of heterogeneous representation network. Fix $W^H$ and extract the feature representation $Y$ by Eq. \[eq:1\]. Estimate the parameter $W^D$ of decorrelation layer by $Y$. Fix the $W^D$, compute the decorrelation representation $Z$. Compute loss $\mathcal{L}_{MML}$ by Eq. \[eq:10\]. Update $W^H$ via back-propagation. Fix $W^H$, update $W^D$ by Eq. \[eq:3\] and Eq. \[eq:4\]. **return** $W^H$ and $W^D$.
Experiments
===========
In this section, we evaluate the proposed framework against some state-of-the-art methods, systemically. We conduct experiments on two popular heterogeneous face databases: CASIA NIR-VIS 2.0 Database [@li2013casia], Oulu-CASIA NIR-VIS Database [@MPL3]. Some cropped samples are shown in Figure. \[database\].
--------- --------- --------- ---------- ---------- -------- ----------
QML HAML Rank-1 FAR=0.1% Rank-1 FAR=0.1%
- - - 99.0 98.1 100 94.2
- - $\surd$ 99.4 98.5 100 95.6
- $\surd$ $\surd$ 99.5 98.9 100 95.7
$\surd$ $\surd$ $\surd$ **99.9** **99.4** 100 **97.2**
--------- --------- --------- ---------- ---------- -------- ----------
Databases and Protocols
-----------------------
The CASIA NIR-VIS 2.0 Face Database is the most challenging and the largest NIR-VIS database with large intra-class cross-domain variations, *i.e.* lighting, expression, pose. There are totally 725 subjects, each has 22 VIS face images and 50 NIR face images at most. We follow the partition protocols in [@he2018wasserstein] and evaluate the proposed method on this database with 10-fold experiments. In the training phase, there are about 6100 NIR face images and 2500 VIS face images share 360 identities in each protocol. In the testing phase, there are about 6100 NIR face images in the probe set and 358 VIS face images in the gallery set. The similarity matrix of probe set and gallery set is $6100\times358$, computed by cosine distance. We compare the proposed method on this database at rank-1 recognition accuracy and verification rate (VR)@false accept rate (FAR) $= 0.1\%$.
The Oulu-CASIA NIR-VIS Database consists of 80 subjects with 6 expressions (*i.e.* anger, disgust, fear, happiness, sadness and surprise). We follow the protocols in [@he2018wasserstein] and select 20 identities as the training set. For each expression, we randomly select 8 pairs of NIR-VIS face images, resulting in 96 cross-domain face images (48 pairs of NIR-VIS face images) for one identity. 20 identities are randomly selected from the remaining 60 identities as the testing set. The VIS face images in testing set are used as gallery and the corresponding NIR face images are used as probe. The similarities of all the NIR face images in the probe set and all the VIS face images in the gallery set are computed by cosine distance, which is a $960\times960$ similarity matrix. The rank-1 recognition accuracy and VR @ FAR $=0.1\%$ are reported to evaluate the performance of the proposed method on this database.
![Some NIR-VIS face images. (a) the CASIA NIR-VIS 2.0 face database; (b) the Oulu-CASIA NIR-VIS face database. These samples are all aligned by five landmarks and cropped to the size of $112\times112$. The first row contains NIR face images and the second row contains the corresponding VIS face images.[]{data-label="database"}](database){width="1\columnwidth"}
Experiments Setting
-------------------
The SE-LResNet50E-IR network [@deng2018arcface] is employed as the heterogeneous representation network in the proposed framework. As shown in Figure. \[framework\], the heterogeneous representation network contains one input layer, four residual groups and one output layers, which is pre-trained on the MS-Celeb-1M database [@guo2016ms]. All the images are aligned and cropped to $112\times112$ by five facial landmarks.
We implement all the experiments in this paper by PyTorch under the environment of Python 3.7 on Ubuntu 16.04 system with i7-6700K CPU and NVIDIA TITAN Xp GPU. The dimensions of feature representation $y^i$ and decorrelation representation $z^i$ are set to 512. Stochastic gradient descent (SGD) is utilized for back-propagation. We set the initial learning rate to $1e^{-4}$, which is gradually reduced to $1e^{-6}$. The batch size to 16 and the trade-off parameter $\lambda_N$, $\lambda_V$, $\lambda_1$ and $\lambda_2$ are set to 0.6, 0.4, 10 and 1 respectively. The angular margin $m$ is set to 0.9 in this paper.
Comparsions
-----------
In this subsection, we first explore the improvements benefiting from the three parts of the proposed framework: decorrelation layer part, quadruplet margin loss constraint and heterogeneous angular margin loss constraint. Then, we compare the performance of the proposed framework with state-of-the-art methods.
As shown in Table \[ablation\], we present the performance of rank-1 recognition accuracy and VR@FAR=0.1%. The baseline method utilizes the heterogeneous representation network, without decorrelation layer and multi-margin loss. For CASIA NIR-VIS 2.0 face database, the baseline method achieves rank-1 recognition accuracy of 99.0% and verification rate@FAR=0.1% of 98.1%. With the constraint of heterogeneous angular margin loss, we fine-tune the baseline network and improve the performance to 99.4% recognition accuracy at rank-1 and 98.5% VR@FAR=0.1%. The proposed quadruplet margin loss contributes less to rank-1 recognition accuracy, but improves the verification rate from 98.5% to 98.9%. The proposed Decorrelation layer improves the proposed framework most, from 99.5% to 99.9% of rank-1 recognition accuracy and from 98.9% to 99.4% of VR@FAR=0.1%, which is very close to 100%. It demonstrates the effectiveness of the proposed framework.
As Oulu-CASIA NIR-VIS face database is a many-to-many database, each probe NIR face image has 8 corresponding VIS face images. Therefore, it is relatively easy to achieve satisfactory performance. As shown in Table \[ablation\], the baseline method achieves 100% recognition accuracy. However, our framework still improves the verification rate. By the constraint of heterogeneous angular margin loss, the verification rate is improved from 94.2% to 95.6%. The quadruplet margin loss improves the verification rate marginally. The proposed decorrelation layer improves the verification rate much from 95.7% to 97.2%.
The experimental results demonstrate that all the three components in the proposed framework improve the performance of the baseline method on recognition accuracy and verification rate for HFR.
Method Rank-1 VR@FAR=0.1%
------------------------------------ ---------- -------------
KCSR [@KCSR] 33.8 7.6
KPS [@KPS] 28.2 3.7
KDSR [@KDSR] 37.5 9.3
LCFS [@wang2013learning] 35.4 16.7
H2(LBP3) [@H2LBP3] 43.8 10.1
C-DFD [@CDFD] 65.8 46.2
CDFL [@CDFL] 71.5 55.1
Gabor+RBM [@yi2015shared] 86.2 81.3
VGG [@parkhi2015deep] 62.1 39.7
HFR-CNN [@saxena2016heterogeneous] 85.9 78.0
TRIVET [@TRIVET] 95.7 91.0
IDR [@IDR] 97.3 95.7
CDL [@wu2017coupled] 98.6 98.3
ADFL [@Song2018Adversarial] 98.2 97.2
WCNN [@he2018wasserstein] 98.4 97.6
WCNN + low-rank 98.7 98.4
**MMDL** **99.9** 99.4
: Comparisons of rank-1 recognition accuracy and VR@FAR=0.1% with state-of-the-art HFR methods on the CASIA NIR-VIS 2.0 database.[]{data-label="tab:casia"}
The rank-1 recognition accuracy and verification rates on CASIA NIR-VIS 2.0 Database is shown in Table \[tab:casia\]. We compare the proposed approach with state-of-the-art HFR methods, including traditional methods(*i.e.* KCSR [@KCSR], KPS [@KPS], KDSR [@KDSR], LCFS [@wang2013learning], Gabor+RBM [@yi2015shared], C-DFD [@CDFD], CDFL [@CDFL], H2(LBP3) [@H2LBP3] and CNN-based methods(*i.e.* VGG [@parkhi2015deep], HFR-CNN [@saxena2016heterogeneous], TRIVET [@TRIVET], IDR [@IDR], ADFL [@Song2018Adversarial], CDL [@wu2017coupled], WCNN [@he2018wasserstein]), DVR [@wu2018disentangled]. For traditional state-of-the-art methods, most of these methods try to learn a common subspace or invariant hand-designed feature representation. However, the representational ability of hand-designed feature is limited and it is hard to reduce the correlation of variations for the great gap between heterogeneous face images. Therefore, the performance of traditional HFR methods is not satisfactory. Gabor+RBM achieves the best performance on rank-1 recognition accuracy and VR@FAR=0.1% in comparing with traditional methods, which are only 86.2% and 85.8%. Owing to the strong representational ability of CNN-based method, these methods achieved better performance than traditional methods. The proposed approach also achieves comparable performance of 99.9% recognition accuracy and 99.4% VR@FAR=0.1%.
For the Oulu-CASIA NIR-VIS Database, we also compare the proposed approach with state-of-the-art methods, including traditional methods (*i.e.* MPL3 [@MPL3], KCSR, KPS, KDSR, H2(LBP3)) and CNN-based methods (*i.e.* TRIVET, IDR, ADFL, CDL, WCNN, DVR). For the same reason in the experiments on CASIA NIR-VIS 2.0 database, CNN-based methods achieve much better performance of rank-1 recognition accuracy on this database. However, the performance of VR@FAR=0.1% is not satisfactory for both traditional methods and CNN-based methods. The best VR@FAR=0.1% of state-of-the-art methods is achieved by DVR with 84.9%. The proposed approach achieves superior performance on both rank-1 recognition accuracy of 100% and VR@FAR=0.1% of 97.2%. It demonstrates the effectiveness of the proposed framework. The experimental results are presented in Table \[tab:oulu\_casia\].
We also evaluate the proposed method on CUHK VIS-NIR database [@gong2017heterogeneous] and improve the rank-1 recognition accuracy from 83.9% to 99.7%.
Method Rank-1 VR@FAR=0.1%
----------------------------- --------- -------------
MPL3 [@MPL3] 48.9 11.4
KCSR [@KCSR] 66.0 26.1
KPS [@KPS] 62.2 22.2
KDSR [@KDSR] 66.9 31.9
H2(LBP3) [@H2LBP3] 70.8 33.6
TRIVET [@TRIVET] 92.2 33.6
IDR [@IDR] 94.3 46.2
IDR + low-rank 95.0 50.3
CDL [@wu2017coupled] 94.3 53.9
ADFL [@Song2018Adversarial] 95.5 60.7
WCNN [@he2018wasserstein] 96.4 50.9
WCNN + low-rank 98.0 54.6
**MMDL** **100** **97.2**
: Comparisons of rank-1 recognition accuracy and VR@FAR=0.1% with state-of-the-art HFR methods on the Oulu-CASIA NIR-VIS database.[]{data-label="tab:oulu_casia"}
Conclusion
==========
Considering that the correlation of cross-domain variations, this paper develops a multi-margin based decorrelation learning (MMDL) method, which employs a decorrelation layer to address this problem. The heterogeneous representation network is pre-trained by large-scale VIS face images. Then, the decorrelation layer is imposed on this network. The parameters of heterogeneous representation network and decorrelation layer are optimized by an alternative optimization strategy. The multi-margin loss is proposed to constrain the network in the fine-tune process, which contains two main components: quadruplet margin loss and heterogeneous angular margin loss. Finally, the similarity of decorrelation representations in the hyperspherical space can be measured by Cosine distance. Experimental results on two popular heterogeneous face recognition databases demonstrate that the proposed MMDL framework significantly leads to superior performance in comparing with state-of-the-art methods. In addition, we explore an ablation study to show the improvements acquired by different components of the proposed MMDL framework.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the following fractional elliptic equations of the type, $$(-\Delta)^{\frac12}_A u = \lambda u+f(|u|)u ,\;\textrm{in } \;(-1, 1),\; u=0\;\textrm{in } \;\mathbb R\setminus (-1, 1),$$ where $\lambda$ is a positive real parameter and $(-\Delta)^{\frac12}_A$ is the fractional magnetic operator with $A:\mathbb R\to \mathbb R$ being a smooth magnetic field. Using a classical critical point theorems, we prove the existence of multiple solutions in the non-resonant case when the nonlinear term $f(t)$ has a critical exponential growth in the sense of Trudinger-Moser inequality.'
address: |
Department of Mathematics, Federal University of Paraíba,\
João Pessoa, PB, 58051–900, Brazil
author:
- Pawan Kumar Mishra
- João Marcos do Ó
- Manassés de Souza
title: Multiplicity results for fractional magnetic problems involving exponential growth
---
fractional magnetic operator ,multiplicity ,critical exponential growth ,critical point theorems.
35A15,35R11,35Q60,35B33.\
Introduction
============
We study the following fractional elliptic equations of the type, $$(P)_\lambda \left\{
\begin{array}{rl}
(-\Delta)^{\frac12}_A u &= \lambda u+f(|u|)u ,\;\textrm{in } \;\Omega,\\
u&=0\;\quad \quad \quad \; \quad\textrm{in } \;\mathbb R\setminus \Omega,
\end{array}
\right.$$ where $\Omega = (-1,1)$ and $\lambda$ is a positive real parameter. For a magnetic field $A: \mathbb R\to \mathbb R$ the operator $(-\Delta)^{\frac12}_A$ is known as the fractional magnetic operator. This operator, recently introduced in [@AS], has been defined (upto a normalization constant) as follows $$\label{FMO}
-(-\Delta)_A^{\frac12}u(x)= \frac{1}{\pi}\lim_{\epsilon \to 0}\int_{\mathbb{R}\setminus B(x, \epsilon)} \frac{u(x)-e^{i(x-y)\cdot A(\frac{x+y}{2})}u(y)}{|x-y|^{2}}{\mathrm{d}}y,\;\;\;\;\; \; x\in \mathbb{R},$$ where $B(x, \epsilon)$ denotes the real interval of size $\epsilon$ around $x$. It is clear that, when $A=0$, the above operator is consistent with the usual fractional Laplacian operator (square root of Laplacian) which has seized a lot of attention in the recent past, see [@FL1; @FL2; @FL3; @FL4] and references therein. This operator arises in the description of various phenomena in the several branches of applied sciences, for example, [@App1] uses the fractional Laplacian for linear and nonlinear lossy media, [@App2; @Applbm] use the fractional Laplacian for option pricing in jump diffusion and exponential Lévy models, [@App3] provides the first ever derivation of the fractional Laplacian operator as a means to represent the mean friction in the turbulence modeling and many more.
On the other hand, we can interpret $(-\Delta)_{A}^{\frac12}$ as a fractional analog of the magnetic Laplacian $(\nabla - \mathrm{i} A)^2$, with $A$ being a [bounded]{} potential. In particular, for the physical interest, the study of is apparent in the case $s={1}/{2}$. Indeed the operator in takes inspiration from the definition of a quantized operator corresponding to the classical relativistic Hamiltonian symbol for a relativistic particle of mass $m\geq 0$, that is $$\sqrt{(\xi-A(x))^{2}+m^{2}}+V(x), \quad (\xi, x)\in \mathbb R^{N}\times \mathbb R^{N},$$ which is the sum of the kinetic energy term involving $A(x)$ (magnetic vector potential) and $V(x)$ (potential energy term of electric scalar potential). For the sake of completeness, we emphasized that in the literature there are three kinds of quantum relativistic Hamiltonians depending on how to quantize the kinetic energy term $\sqrt{(\xi-A(x))^{2}+m^{2}}$. As explained in [@T2], these three nonlocal operators are in general different from each other but coincide when the vector potential $A$ is assumed to be linear, so in particular, in the case of constant magnetic fields. For a more detailed description of the operator $(-\Delta)^{s}_{A}$ and relaed problems, we refer the interested readers to [@AS; @FiscellaV1; @T1; @T2; @AME1; @AME; @BMX] and the references therein.
In a latest work, authors in [@FiscellaVecchi] studied a multiplicity result for the following problem in higher dimensions involving fractional magnetic operator $$\label{FPr}
(-\Delta )_A^su=\lambda u+|u
|^{2^*_s-2}u, \;\; \textrm{in}\;\;\Omega,\;\; u=0\;\;\textrm{on}\;\;\mathbb R^N\setminus\Omega,$$ where $\Omega\subset \mathbb R^N$ is an open and bounded set with Lipschitz boundary, $N>2s$, $s\in(0, 1)$ and $2^*_s=2N/(N-2s)$ is the fractional critical Sobolev exponent. The result summarizes as the existence of $m$ pairs of solution of for $\lambda$ lying in the suitable left neighborhood of any eigenvalue with multiplicity $m$ of the magnetic fractional Laplace operator with Dirichlet boundary data.
Note that when $A=0$, the problem transforms into the following problem involving the celebrated fractional Laplace operator $$(-\Delta )^su=\lambda u+u^{2^*_s-2}u, \;\; \textrm{in}\;\;\Omega,\;\; u=0\;\;\textrm{on}\;\;\mathbb R^N\setminus \Omega,$$ which has been studied in fair share by the authors in [@fbm]. Using the abstract critical point theorem, authors in [@fbm] have generalized the results of Cerami, Fortuno and Struwe [@sfs] for the nonlocal setting.
We know from classical fractional Sobolev embedding that $H^{s}(\Omega)$ is continuously embedded in $L^q(\Omega)$ for all $q\in [1, 2^*_s]$, where $2^*_s=2N/(N-2s)$. Note that formally, $2^*_s=\infty$ if $N=2s$. Since $s\in (0, 1)$, the only choice for this fact to be true is $N=1$ and $s=1/2$. At this point a natural question arises to look for an optimal space where $H^{1/2}(\Omega)$ can be embedded. This answer was first given by Ozawa [@Ozawa] and later improved by Iula, Maalaoui and Martinazzi [@matrinazziint] in the form of fractional Trudinger-Moser inequlity (see Lemma \[Trudinger-Moser1\]). This result has motivated many researcher to consider the critical exponent problem in limiting case of fractional Sobolev embedding in dimension 1 such as [@MR3399183; @JPS; @IzSq], with no attempt to give a complete list. (see also in the local case [@Adimurthi; @FMR] and reference therein)
The result obtained in [@FiscellaVecchi] covers all the dimensions except the dimensions $N=2s$ which corresponds to the only dimension, $N=1$ when $s\in (0, 1)$ for $s=1/2$. Up to the best of our knowledge, there is no work dealing with fractional magnetic operator with critical exponential growth except the work of Ambrosio [@Ambrossio] which is related with concentration behavior of solutions for nonlinear Schrödinger equations. This is the main motivation for studying the problem under consideration. In the case of $A=0$, a non-magnetic counter part of $(P)_\lambda$ was partially considered in [@EJDE2]. Inspired from a suitable variant of Trudinger-Moser inequality and by proving the required Moser sequence estimates to study the min-max level under Adimurthi type assumption (see assumption $(H_4)$ below), we have complemented the work of [@FiscellaVecchi] in the dimension one. Our results complete the partial result obtained in [@EJDE2] as well.
Assumptions
-----------
We will consider $f: [0,+\infty) \rightarrow [0,+\infty)$ a continuous function having the critical exponential growth in the following sense: there exists $\beta_0 >0$ such that $$\label{CG}
\tag{CG}
\lim_{t \to +\infty} \frac{f(t) t}{e^{\beta t^2}} = \left\{
\begin{array}{ll}
+\infty, & \hbox{if}\,\, \beta < \beta_0 \\
0, & \hbox{if}\,\, \beta > \beta_0.
\end{array}
\right.$$ Moreover, $f$ satisfies
$(H_1)$
: \[H1\] There exist $t_0 >0$ and $M>0$ such that $$F(t) \leq M f(t) t,\,\,\mbox{for all}\,\, t \geq t_0;$$ where $$F(t) := \int_0^t f(\tau) \tau {\mathrm{d}}\tau,\,\, \mbox{for any}\,\, t>0.$$
$(H_2)$
: $0< F(t) \leq f(t) t^2$ for all $t>0$.
$(H_3)$
: For each $k\geq 1$, $\lambda$ $f$ satisfies $\ell:=\limsup_{t\to 0^+} \frac{2F(t)}{t^2}<\lambda_k-\lambda$, where $\lambda\in (0, \lambda_1)$ if $k=1$ otherwise $\lambda \in (\lambda_{k-1}, \lambda_k)$.
Here $(\lambda_k)$ denotes the sequence of eigenvalues associated to the problem $$\begin{aligned}
\label{EVP}
(-\Delta)_A^\frac{1}{2}u&=\lambda u\;\;\textrm{in}\;\;\Omega,\; u=0\;\;\;\; \textrm{in}\;\;\mathbb R\setminus \Omega.\end{aligned}$$
Spectral properties
-------------------
It is known that there exists a infinite sequence of eigenvalues $\lambda_1\leq\lambda_2\leq .....\leq\lambda_k\leq ....$ with $\lambda_k\rightarrow+\infty$ as $k\rightarrow \infty$. The eigenfunctions $\{\varphi_k\}$ corresponding to each eigenvalue $\lambda_k$ form an orthonormal basis for $L^2(\Omega)$ and an orthogonal basis for $X_{0, A}$, where the space $X_{0, A}$ and corresponding norm $\|\cdot\|_{X_{0, A}}$ is defined in Section 2. Hence $X_{0,A} = H_{k} \oplus H_{k}^{\perp}$, where $H_{k} = \mbox{span}_{\mathbb{R}}\{\phi_1, \phi_2,\cdot\cdot\cdot,\phi_{k}\}$. The following characterization is shown in Proposition 3.3 [@FiscellaVecchi]. $$\label{EVL1}
\lambda_1=\min_{u\in X_{0, A}\setminus \{0\}}\frac{\|u\|^2_{X_{0, A}}}{\|u\|_2^2}$$ Moreover, inductively, for any $k\geq 2$ $$\label{EVL2}
\lambda_{k}=\min_{u\in H^\perp_k\setminus \{0\}}\frac{\|u\|^2_{X_{0, A}}}{\|u\|_2^2}$$
Main results and remarks
------------------------
The objective of this paper is multi-fold. Depending on the location of the parameter $\lambda$ with respect to the spectrum of $(-\Delta)_A^\frac{1}{2}$ with Dirichlet data, we categories the result of this paper in the form of following four main Theorems. The first result deals with the case when the parameter $0<\lambda<\lambda_1$ and the nonlinearity has critical exponential growth. Note that the problem under consideration is no more coercive which is a natural hindrance to study via usual minimization argument. In this case the classical mountain pass theorem gives the existence of a critical point of the corresponding energy functional which results into a nontrivial weak solution of the problem by a one to one correspondence between critical points of the associated energy functional and weak solutions of the problem. Our first result is stated as follows:
\[thm1\] Assume $(H_1)-(H_2)$ and $(H_3)$ with $k=1$. Let $f$ has exponential critical growth together with
$(H_4)$
: $\displaystyle{\liminf_{t \to +\infty} f(t)t^2 e^{\beta_0 t^2}=+\infty}$, [where]{} $\beta_0$ is introduced in .
Then the problem $(P)_\lambda$ has a nontrivial solution.
We point out that the assumption $(H_4)$ was introduced by Adimurthi in [@Adimurthi] in the first instance..
The problem $(P)_\lambda$ exhibits interesting feature when the parameter $\lambda$ lies in between the eigen values $\lambda\in ( \lambda_k, \lambda_{k+1})$ for $k\geq 1$. The second result of the paper highlights this delicate point. The proof of this result invokes the celebrated idea of Linking geomtery.
\[thm2\] Assume $\lambda\in (\lambda_k, \lambda_{k+1})$, $(H_1)-(H_4)$ and that $f$ has exponential critical growth. Then problem $(P)_\lambda$ has a nontrivial solution.
The third theorem of the paper also deals with the critical growth nonlinearity but involves a little stronger assumption due to D. M. Cao (see assumption $(H_5)$ below) instead of Adimurthi assumption $(H_4)$. But with this compromise, we could prove the least bound of critical points of the associated functional by applying another abstract critical point theorem due to [@crlit]. The result says that
\[thm3\] Assume $(H_1)-(H_3)$ and that $f$ has exponential critical growth. Furthermore assume
$(H_5)$
: there exist $p>2$ and a constant $C_p>0$ possibly large such that $f(t)t \geq C_p t^{p-1}$ for all $t\geq 0$.
Define $\lambda_k$ be the $k^{th}$ eigenvalue of the problem with multiplicity $m$. Let $\lambda\in \mathbb R$ and define $\lambda^*=\min\{\lambda_k: \lambda<\lambda_k\}$. If $\lambda<\lambda^*$ and $$C_p>\left( \frac{\beta_0(p-2)}{\pi}\right)^\frac{p-2}{2}\left((\lambda^*-\lambda)2^{\frac{p}{p-2}}\right)^\frac{p}{2},$$ where $\beta_0$ is introduced in , then problem $(P)_\lambda$ admits $m$ pairs of non-trivial weak solutions $\{-u_{\lambda,j}, u_{\lambda, j}\}$, for every $j=1,2,\cdot\cdot\cdot,m$.
The assumption $(H_5)$ was firstly introduced by D. M. Cao in [@Cao].
Before stating the last result of the paper, we introduce what we mean by subcritical growth. We say that $f$ has subcritical growth at $+\infty$ if $$\label{SG}
\tag{SG}
\lim_{t\to+\infty}\frac{|f(t)t|}{e^{\beta t^2}}=0\;\; \text{for all} \;\;\beta>0.$$ Under the light of , it is clear that for some constant $C(\beta)>0$ the nonlinearity satisfies $$\label{subgrowth}
f(t)t\leq C(\beta)e^{\beta t^2}\;\;\text{for all}\;\;\beta>0\;\;\text{ and for all}\;\; t\in \mathbb R.$$ In the last Theorem of the paper, we show that the problem $(P)_\lambda$ exhibits two non-trivial weak solutions under subcritical growth assumption in the sence of . In this case by allowing the nonlinearity to be subcritical, we could prove our result without assuming Cao condition $(H_5)$ or Adimurthi type assumption $(H_4)$.
We conclude the buildup of the last result by introducing the following notations. Since the space $X_{0, A}\hookrightarrow L^{p}(\Omega)$ for all $p\in [2,\; \infty) $, the following supremum is well defined
$$\label{Holder}
S_p=\sup_{\{v\in X_{0, A}\;:\; \|v\|_{X_{0, A}}\leq 1\}} \frac{\|v\|_p}{\|v\|_{X_{0, A}}}.$$
\[thm4\] Assume $f$ satisfies $(H_1)-(H_2)$ together with . Then for every $\rho>(\sqrt{2}S_pC(\beta))^2$ there exists $$\Lambda(\rho):=\frac{1}{S_2^2}\left(1-\frac{\sqrt{2}S_pC(\beta)}{\sqrt{\rho}}\right),$$ where $S_2$, $S_p$ and $C(\beta)$ are defined in and , respectively, such that problem $(P)_\lambda$ has at least two nontrivial weak solutions for every $\lambda \in(0, \Lambda(\rho))$, one of which has norm strictly less than $\rho$.
The proof of the above Theroem is variational and is based on a abstract critical point theorem due to Recceri [@Ricceri] (see Theorem 6).
We point out that these results are true even in the absence of magnetic field, that is, the case when $A=0$.
Functional framework
====================
In this section we give a more general variational set up rather than considering $\Omega=(-1, 1), N=1, s=1/2$ as in our case in this paper. We indicate with $|\Omega|$ the $N$-dimensional Lebesgue measure of a measurable set $\Omega \subset \mathbb{R}^N$. Moreover, for every $z \in \mathbb{C}$ we denote by $\Re z$ its real part, and by $\overline{z}$ its complex conjugate. Let $\Omega \subset \mathbb{R}^N$ be an open set. We denote by $L^2(\Omega,\mathbb{C})$ the space of measurable functions $u:\Omega\to\mathbb{C}$ such that $$\|u\|_{L^2(\Omega)}=\left(\int_{\Omega}|u(x)|^2 \, {\mathrm{d}}x\right)^{1/2}<\infty,$$ where $|\cdot|$ is the Euclidean norm in $\mathbb{C}$.
For $s\in (0,1)$, we define the magnetic Gagliardo semi-norm as $$[u]_{H^{s}_A(\Omega)}:=\left(\frac{1}{2\pi}\iint_{\Omega\times\Omega}\frac{|u(x)-e^{\mathrm{i} (x-y)\cdot
A(\frac{x+y}{2})}u(y)|^2}{|x-y|^{N+2s}}\,{\mathrm{d}}x\,{\mathrm{d}}y\right)^{1/2}.$$ We denote by $H^{s}_A(\Omega)$ the space of functions $u\in L^2(\Omega,\mathbb{C})$ such that $[u]_{H^{s}_A(\Omega)}<\infty$, normed with $$\|u\|_{H^{s}_A(\Omega)}:=\left(\|u\|_{L^2(\Omega)}^2+[u]_{H^{s}_A(\Omega)}^2\right)^{1/2}.$$
However, to encode the boundary condition $u=0$ in $\mathbb{R}^N\setminus\Omega$, the natural functional space to deal with weak solutions of problem $(P)_\lambda$ is $$X_{0,A} := \left\{ u \in H^{s}_{A}(\mathbb{R}^N): u = 0\ \text{in }
\mathbb{R}^N \setminus \Omega \right\}.$$
We define the following real scalar product on $X_{0,A}$ $$\langle u,v\rangle_{X_{0,A}}:=\frac{1}{2\pi} \Re \iint_{\mathbb{R}^{2N}}\frac{\big(u(x)-e^{\mathrm{i} (x-y)
\cdot A(\frac{x+y}{2})}u(y)\big)
\overline{\big(v(x)-e^{\mathrm{i} (x-y)\cdot A(\frac{x+y}{2})}v(y)\big)}}
{|x-y|^{N+2s}}\, {\mathrm{d}}x {\mathrm{d}}y,$$ which induces the norm $$\|u\|_{X_{0,A}} := \left( \frac{1}{2\pi}\iint_{\mathbb{R}^{2N}}\frac{|u(x)- e^{\mathrm{i} (x-y)\cdot
A(\frac{x+y}{2})}u(y)|^2}{|x-y|^{N+2s}}\, {\mathrm{d}}x {\mathrm{d}}y \right)^{1/2}.$$ Under the scalar product defined above, the space $(X_{0,A}, \langle \cdot,\cdot\rangle_{X_{0,A}})$ is a Hilbert space and hence reflexive.
Arguing similar to [@AA] and [@AS], we have the following result.
1. The space $H^{1/2}_A(\mathbb R, \mathbb C)$ is continuously embedded into $L^r(\mathbb R, \mathbb C)$ for any $r\in [2,\infty)$ and compactly embedded into $L^r_{\mathrm {loc}}(\mathbb R, \mathbb C)$ for any $r\in [1,\infty)$.
2. For any $u\in H^{1/2}_A(\mathbb R, \mathbb C)$, we get $|u|\in H^{1/2}(\mathbb R, \mathbb R)$ and $[|u|]\leq [u]_A$. Moreover we also have the following pointwise diamagnetic inequality $$\big||u(x)|-|u(y)|\big|\leq \left|u(x)-e^{i(x-y)\cdot A(\frac{x+y}{2})}u(y)\right| \;\; a. ~e. \;\;x, y \in\mathbb R.$$
3. If $u\in H^{1/2}(\mathbb R, \mathbb R)$ and has compact support then $v=e^{iA(0)\cdot x} u\in H^{1/2}_A(\mathbb R, \mathbb C)$.
As discussed in the introduction, the problems of the type $(P)_\lambda$ are motivated by the following version of the Trudinger-Moser inequality, which is a consequence of the results proved by Ozawa [@Ozawa], Kozono, Sato and Wadade [@Wadade], Martinazzi [@Martinazzi] and Takahashi [@Takahasi].
\[Trudinger-Moser1\] If $\alpha >0$ and $u\in X_{0,0}$, it holds $$\label{TM11}
e^{\alpha u^2} \in L^1(\Omega).$$ Moreover, $$\sup_{\{u \in X_{0,0}\,:\, \|u\|_{1/2,2} \leq 1\}}\int_\Omega e^{\alpha u^{2}} \, {\mathrm{d}}x < \infty,
\label{TM21}$$ for all $0 \leq \alpha \leq \pi$, where $$\|u\|_{1/2,2} := \left(\dfrac{1}{2\pi}\displaystyle\int_{\mathbb R^2}
\dfrac{(u(x)-u(y))^{2}}{|x-y|^{2}} \, {\mathrm{d}}x\, {\mathrm{d}}y \right)^{1/2}.$$
\[lem\] If $0\leq \alpha \leq \pi$, it holds $$\label{Corolario1}
\sup_{\{u \in X_{0,A}\,:\, \|u\|_{X_{0,A}} \leq 1\}}\int_\Omega e^{\alpha |u|^{2}} \, {\mathrm{d}}x < \infty.$$ Moreover, for any $\alpha >0$ and $u \in X_{0,A}$, $$\label{lema2.3}
e^{\alpha |u|^{2}} \in L^{1}(\Omega).$$
The estimating follows from $\||u|\|_{1/2,2} \leq \|u\|_{0,A}$ and Lemma \[Trudinger-Moser1\]. Now we prove the second part of the lemma. Given $u \in X_{0,A}$ and $\varepsilon>0$, there exists $\varphi \in C_{0}^{\infty}(\Omega)$ such that $\||u|-\varphi\|_{1/2,2}<\varepsilon$. Since $$e^{\alpha |u|^{2}} \leq
e^{\alpha(2(|u|-\varphi)^{2}+2\varphi^{2})} \leq
\dfrac{1}{2}e^{4\alpha
(|u|-\varphi)^{2}}+\dfrac{1}{2}e^{4\alpha
\varphi^{2}},$$ it follows that $$\label{jjjj1}
\int_\Omega e^{\alpha |u|^{2}} \, {\mathrm{d}}x \leq
\dfrac{1}{2}\int_\Omega e^{4\alpha\||u|-\varphi\|_{1/2,2}^{2}\left(\frac{|u|-\varphi}{\||u|-\varphi\|_{1/2,2}}\right)^{2}}
\, {\mathrm{d}}x + \dfrac{1}{2}\int_\Omega e^{4\alpha\varphi^{2}} \,
{\mathrm{d}}x.$$ Choosing $\varepsilon >0$ such that ${4\alpha\varepsilon^{2}} <
\pi$, we have ${4\alpha\||u|-\varphi\|_{1/2,2}^{2}}<
\pi$. Then, from Lemma \[Trudinger-Moser1\] and , we obtain $$\int_\Omega e^{\alpha |u|^{2}} \, {\mathrm{d}}x \leq \dfrac{C}{2} +
\dfrac{1}{2} \int_{\textrm{supp}(\varphi)}e^{4\alpha \varphi^{2}} \,
{\mathrm{d}}x< \infty.$$ Thus, the proof is complete.
We say that a function $u \in X_{0,A}$ is a weak solution of $(P)_\lambda$ if $$\langle u,\varphi \rangle_{X_{0,A}}= \lambda \Re \int_{\Omega}u(x)~ \overline{\varphi(x)}\, {\mathrm{d}}x
+ \Re \int_{\Omega}f(|u|)~u~ \overline{\varphi(x)}\, {\mathrm{d}}x,$$ for every $\varphi \in X_{0,A}$.
Clearly, the weak solutions of $(P)_\lambda$ are the critical points of the Euler–Lagrange functional $\mathcal{I}_{A,\lambda}:X_{0,A} \to \mathbb{R}$, associated with $(P)_\lambda$, that is $$\label{Jlam}
\mathcal{I}_{A,\lambda}(u):= \frac12\|u\|_{X_{0,A}}^2
- \frac{\lambda}{2}\|u\|_{2}^2
- \int_\Omega F(|u|) {\mathrm{d}}x,$$ where $F(t)=\int_0^t f(s)s{\mathrm{d}}s$. By using our assumptions and Lemma \[lem\], it is easy to see that $\mathcal{I}_{A,\lambda}$ is well-defined and of class $C^1(X_{0,A}, \mathbb{R})$.
The next lemma will be used to ensure the geometry of the functional $\mathcal{I}_{A,\lambda}$.
If $v \in X_{0,A}$, $\alpha >0$, $q>2$ and $\|v\|_{X_{0,A}}\leq M$ with $\alpha
M^{2}<\pi$, then there exists $C=C(\alpha,M,q)>0$ such that $$\int_\Omega e^{\alpha |v|^{2}}|v|^{q}{\mathrm{d}}x \leq C\|v\|_{X_{0,A}}^{q}.$$
Taking $r > 1$ close to $1$ such that $\alpha r M^{2}<\pi $ and $r'q \geq 1$, where $r'={r}/{(r-1)}$. By Hölder’s inequality, we have $$\label{exponential11}
\int_\Omega e^{\alpha |v|^{2}}|v|^{q}{\mathrm{d}}x \leq
\left(\int_\Omega e^{\alpha r |v|^{2}}{\mathrm{d}}x\right)^{1/r}\|v\|_{r'q}^{q} = \left(\int_\Omega e^{\alpha r \|v\|_{X_{0,A}}^{2}
\left(\frac{|v|}{\|v\|_{X_{0,A}}}\right)^{2}}\,{\mathrm{d}}x \right)^{1/r}
\|v\|_{r'q}^q.$$ Since $\alpha r
M^{2}<\pi$, it follows from and the continuous embedding $X_{0,A} \hookrightarrow L^{r'q}(\Omega)$, that $$\int_\Omega e^{\alpha |v|^{2}}|v|^{q}\,{\mathrm{d}}x \leq C\|v\|_{X_{0,A}}^q.$$ Thus, the proof is complete.
We will show a refinement of . This result will be crucial to show that the functional $\mathcal{I}_{A,\lambda}$ satisfies the Palais-Smale condition.
\[tipolions\](P. L. Lions’ concentration compactness result) If $(v_{n})$ is a sequence in $X_{0,A}$ with $\|v_{n}\|_{X_{0,A}}=1$ for all $n
\in \mathbb{N}$ and $v_{n}\rightharpoonup v$ weakly in $X_{0,A}$, $0<\|v\|_{X_{0,A}}<1$, then for all $0<t<\pi(1-\|v\|_{X_{0,A}}^{2})^{-1}$, we have $$\sup_{n} \int_\Omega e^{t|v_{n}|^{2}}\,{\mathrm{d}}x<\infty.$$
Since $v_{n}\rightharpoonup v$ weakly in $X_{0,A}$ and $\|v_{n}\|_{X_{0,A}}=1$, we get $$\|v_{n}-v\|_{X_{0,A}}^{2}=1-2\langle v_{n},v \rangle_{X_{0,A}}+\|v\|_{X_{0,A}}^{2}\rightarrow
1-\|v\|_{X_{0,A}}^{2}<\dfrac{\pi}{t}.$$ Thus, for $n \in \mathbb{N}$ enough large, we have $t\|v_{n}-v\|_{X_{0,A}}^{2}<\pi$. Thus, we may choose $q>1$ close to 1 and $\varepsilon>0$ satisfying $$\label{ddddjj12}
qt(1+\varepsilon^{2})\|v_{n}-v\|_{X_{0,A}}^{2}<\pi,$$ for $n \in \mathbb{N}$ enough large. By and , there exists $C>0$ such that $$\label{26}
\int_\Omega e^{qt(1+\varepsilon^{2})|v_{n}-v|^{2}} \, {\mathrm{d}}x=
\int_\Omega e^{qt(1+\varepsilon)^{2}\|v_{n}-v\|_{X_{0,A}}^{2}\left(\frac{|v_{n}-v|}{\|v_{n}-v\|_{X_{0,A}}}
\right)^{2}}
\, {\mathrm{d}}x \leq C.$$ Moreover, since $$t|v_{n}|^{2}\leq
t(1+\varepsilon^{2})|v_{n}-v|^{2}+t\left(1+\dfrac{1}{\varepsilon^{2}}\right)|v|^{2},$$ it follows by convexity of the exponential function with $q^{-1}+r^{-1}=1$ that $$\begin{aligned}
e^{t|v_{n}|^{2}}\leq
\dfrac{1}{q}e^{qt(1+\varepsilon^{2})|v_{n}-v|^{2}}+\dfrac{1}{r}e^{rt(1+1/\varepsilon^{2})|v|^{2}}.
\end{aligned}$$ Therefore, by and , we get $$\int_\Omega e^{t|v_{n}|^{2}}\, {\mathrm{d}}x \leq
\dfrac{1}{q}\int_\Omega e^{qt(1+\varepsilon^{2})|v_{n}-v|^{2}} \,
{\mathrm{d}}x + \int_\Omega e^{rt(1+1/\varepsilon^{2})|v|^{2}}\, {\mathrm{d}}x \leq
C,$$ and the result is proved.
Palais-Smale sequence analysis
==============================
In this section, we will study the definition and properties of Palais-Smale sequence and its precompactness. We begin by recalling the following definition of Palais-Smale sequence.
$\{u_n\}\subset X_{0, A}$ is called a Palais-Smale sequence for $\mathcal I_{A, \lambda}$ at a level $c$ (in short $(PS)_c$ sequence) if $$\mathcal I_{A,\lambda}(u_n)\to c\;\;\text{ and } \mathcal I_{A, \lambda}^\prime (u_n)\to 0\;\;\text{ as}\;\; n\to \infty.$$ We say that $\mathcal I_{A,\lambda}$ satisfies Palais-Smale condition at level $c$ if any $(PS)_c$ sequence admits a convergent subsequence in $X_{0, A}$.
\[PS11\] Assume $(H_1)$. Let $(u_n) \subset X_{0,A}$ be a $(PS)_c$ sequence of $I_\lambda$. Then $(u_n)$ is a bounded in $X_{0,A}$.
Let $(u_n) \subset X_{0,A}$ be a $(PS)_c$ sequence of $I_\lambda$, that is, $$\label{EJ23}
\frac{1}{2}\|u_n\|_{X_{0,A}}^2 - \frac{\lambda}{2}\|u_n\|_2^2 - \int_{\Omega} F(|u_n|) {\mathrm{d}}x \to c,\,\, \mbox{as}\,\, n \to +\infty,$$ and $$\label{EJ24}
\left|\Re\langle u_n, v \rangle_{X_{0,A}} - \lambda \Re\langle u_n , v\rangle_{L^2} - \Re\int_{\Omega} f(|u_n|) u_n \overline{v} {\mathrm{d}}x \right| \leq \varepsilon_n \|v\|_{X_{0,A}},\,\, \mbox{for all}\,\,
v \in X_{0,A},$$ where $\varepsilon_n \to 0$ as $n \to +\infty$. It follows from $(H_1)$, that there exists $t_1 >0$ such that $$\label{EJ231}
F(t) \leq \frac{1}{4}f(t) t^2,\,\,\mbox{for all}\,\, t \geq t_1.$$ Using - and $v=u_n$ as test function, we can find $C>0$ such that $$\label{EJ26}
\int_{\Omega} f(|u_n|) |u_n|^2 {\mathrm{d}}x \leq C + 2 \varepsilon_n \|u_n\|_{X_{0,A}}.$$ To complete the proof, we consider two cases.
**Case 1:** $0< \lambda < \lambda_1$
From , and the variational characterization of $\lambda_1$, we have the following estimate for $\|u_n\|_{X_{0,A}}$, $$\left(\frac{\lambda_1- \lambda}{\lambda_1} \right)\|u_n\|_{X_{0,A}}^2 \leq C + 3 \varepsilon_n \|u_n\|_{X_{0,A}}.$$ Consequently, $(u_n)$ is a bounded sequence in $X_{0,A}$ in this case.
**Case 2:** $\lambda_k < \lambda < \lambda_{k+1}$
Given $u \in X_{0,A}$, we write $u = u^{k} + u^{\perp}$, where $u^{k} \in H_{k}$ and $u^{\perp} \in H_{k}^{\perp}$. Notice that $$\label{EJ210}
\Re\langle u, u^{k}\rangle_{X_{0,A}} - \lambda \Re\langle u, u^{k}\rangle_{L^2} = \|u^{k}\|_{X_{0,A}}^2 - \lambda \|u^{k}\|_2^2$$ and $$\label{EJ211}
\Re\langle u, u^{\perp}\rangle_{X_{0,A}} - \lambda \Re\langle u, u^{\perp}\rangle_{L^2} = \|u^{\perp}\|_{X_{0,A}}^2 - \lambda \|u^{\perp}\|_2^2.$$ By , and the variational characterization of $\lambda_{k}$, we obtain $$\begin{aligned}
-\varepsilon_n \|u_n^{k}\|_{X_{0,A}} & \leq \Re\langle u_n, u_n^{k}\rangle_{X_{0,A}} - \lambda \Re\langle u_n, u_n^{k}\rangle_{L^2} -
\Re\int_{\Omega} f(|u_n|) u_n \overline{u_n^{k}} {\mathrm{d}}x \\
& \leq \left(\frac{\lambda_{k}- \lambda}{\lambda_{k}} \right)\|u_n^{k}\|_{X_{0,A}}^2 - \Re\int_{\Omega} f(|u_n|) u_n \overline{u_n^{k}}\, {\mathrm{d}}x.
\end{aligned}$$ Therefore, we can find $C>0$ such that $$\label{EJ212}
\left(\frac{\lambda- \lambda_k}{\lambda_{k}} \right)\|u_n^{k}\|_{X_{0,A}}^2 \leq \varepsilon_n \|u_n^{k}\|_{X_{0,A}} + \|u_n^{k}\|_{\infty} \int_{\Omega} f(|u_n|) |u_n| \,{\mathrm{d}}x.$$ By applying , there exist $C_1, C_2>0$ such that $$\label{EJ27}
\int_{\Omega} f(|u_n|) |u_n| \,{\mathrm{d}}x \leq C_1 + C_2\varepsilon_n \|u_n\|_{X_{0,A}}.$$ Since $H_{k}$ is a finite dimensional subspace, we can find $C>0$ such that $\|u_n^{k}\|_{\infty} \leq C \|u_n^{k}\|_{X_{0,A}}$. Thus, from and , we get $$\label{EJ213}
\|u_n^{k}\|^2_{X_{0,A}} \leq C (\|u_n^{k}\|_{X_{0,A}} + \varepsilon_n \|u_n^{k}\|_{X_{0,A}} + \varepsilon_n \|u_n^{k}\| _{X_{0,A}}\|u_n\|_{X_{0,A}}).$$ Again, by , and the variational characterization of $\lambda_{k+1}$, it follows that $$\begin{aligned}
\varepsilon_n \|u_n^{\perp}\|_{X_{0,A}} & \geq \Re\langle u_n, u_n^{\perp}\rangle_{X_{0,A}} - \lambda \Re\langle u_n, u_n^{\perp}\rangle_{L^2} -
\Re\int_{\Omega} f(|u_n|) u_n \overline{u_n^{\perp}} \,{\mathrm{d}}x \\
& \geq \left(\frac{\lambda_{k+1}- \lambda}{\lambda_{k+1}} \right)\|u_n^{\perp}\|^2_{X_{0,A}} - \Re\int_{\Omega} f(|u_n|) u_n \overline{u_n^{\perp}}\, {\mathrm{d}}x\\
& \geq \left(\frac{\lambda_{k+1}- \lambda}{\lambda_{k+1}} \right)\|u_n^{\perp}\|^2_{X_{0,A}} - \int_{\Omega} f(|u_n|) |u_n|^2\, {\mathrm{d}}x - \|u_n^{k}\|_{\infty} \int_{\Omega} f(|u_n|) |u_n| \,{\mathrm{d}}x.
\end{aligned}$$ This together with and , implies that there exists $C > 0$ such that $$\label{EJ214}
\|u_n^{\perp}\|^2_{X_{0,A}} \leq C (\varepsilon_n \|u_n^{\perp}\|_{X_{0,A}}+ C + \varepsilon_n \|u_n\| _{X_{0,A}}+ \|u_n^{k}\| _{X_{0,A}}+ \varepsilon_n \|u_n^{k}\|_{X_{0,A}} \|u_n\|_{X_{0,A}}).$$ Combining and , we obtain that $$\label{EJ215}
\|u_n\|^2_{X_{0,A}}\leq C (1+ \|u_n\| _{X_{0,A}}+ \varepsilon_n \|u_n\|^2_{X_{0,A}}),$$ and consequently the sequence $\{u_n\}$ is bounded. Thus, we finished the proof.
\[PS\] Assume that $(H_1)-(H_2)$ are satisfied. Then $\mathcal{I}_{A,\lambda}$ satisfies the $(PS)_c$ condition for $c< \frac{\pi}{2 \beta_0}$.
Let $(u_n)$ be satisfying and . By Lemma \[PS11\], we obtain a subsequence denoted again by $(u_n)$ such that, for some $u \in X_{0,A}$, we have $u_n \rightharpoonup u$ in $X_{0,A}$, $u_n \rightarrow u$ in $L^q(\Omega)$ for all $q \in [1, +\infty)$ and $u_n(x) \rightarrow u(x)$ a.e in $\Omega$. It follows from and [@FMR Lemma 2.1], that $\int_{\Omega} f(|u_n|)|u_n|{\mathrm{d}}x \rightarrow \int_{\Omega} f(|u|)|u|{\mathrm{d}}x$, as $n \to +\infty$. Thus, by applying $(H_1)$ and the Generalized Lebesgue Dominated Convergence Theorem we have $$\int_{\Omega} F(u_n){\mathrm{d}}x \rightarrow \int_{\Omega} F(u){\mathrm{d}}x,\,\, \mbox{as}\,\, n \to +\infty.$$ This convergence together with imply $$\label{EJ221}
\lim_{n \to +\infty}\|u_n\|^2_{X_{0,A}} = 2c + \lambda\|u\|_2^2 +2 \int_{\Omega} F(|u|) {\mathrm{d}}x.$$ Consequently, from it follows $$\label{EJ222}
\lim_{n \to +\infty} \int_{\Omega} f(|u_n|)|u_n|^2 {\mathrm{d}}x= 2c +2 \int_{\Omega} F(|u|) {\mathrm{d}}x.$$ From $(H_2)$ and we reach $c \geq 0$. It follows by $(H_2)$ and , that $$\|u\|^2_{X_{0,A}} - \lambda \|u\|_2^2 = \int_{\Omega} f(|u|)|u|^2 {\mathrm{d}}x \geq 2 \int_{\Omega} F(|u|) {\mathrm{d}}x,$$ and consequently, we get that $I_\lambda(u) \geq 0$.
Now we will prove that $u_n \rightarrow u$ in $X_{0,A}$.
In order to achieve this goal we will consider three cases.
**Case 1:** $c=0$
In this case, using we have $$0 \leq\mathcal{I}_{A,\lambda}(u) \leq \liminf_{n \to +\infty} \mathcal{I}_{A,\lambda}(u_n) = \liminf_{n \to +\infty} \frac{1}{2}\|u_n\|^2_{X_{0,A}} - \left(\frac{\lambda}{2}\|u\|_2^2 + \int_{\Omega} F(|u|) {\mathrm{d}}x \right)=0.$$ Consequently, $u_n \rightarrow u$ in $X_{0,A}$, as $n \to +\infty$, as we wanted to demonstrate.
**Case 2:** $c\not=0$ and $u=0$
We will show that this case cannot happen for a $(PS)_c$ sequence. Indeed, since $u=0$, it follows from that, given $\varepsilon >0$, for $n$ large enough, we have $$\label{EJ224}
\|u_n\|^2_{X_{0,A}}\leq 2 c + \varepsilon.$$ Now we notice that, using that $f$ has critical growth, it holds $$\label{EJ225x}
\int_{\Omega} (f(|u_n|)|u_n|)^q {\mathrm{d}}x \leq C \int_{\Omega} e^{q \beta \|u_n\|^2_{X_{0,A}} \left(\frac{|u_n|}{\|u_n\|_{X_{0,A}}}\right)^2} {\mathrm{d}}x.$$ Since $c < \frac{\pi}{2 \beta_0}$, by using , we can choose $q>1$ sufficiently close to $1$, $\beta > \beta_0$ sufficiently close to $\beta_0$, and $\varepsilon$ sufficiently small such that $q \beta \|u_n\|^2 _{X_{0,A}}< \pi$, for $n$ large enough. Thus, by the Trudinger-Moser inequality and we have $$\label{ETC}
\int_{\Omega} (f(|u_n|)|u_n|)^q {\mathrm{d}}x \leq C.$$ From this estimate and the Hölder’s inequality, up to a subsequence, we get $$\int_{\Omega} f(|u_n|)|u_n|^2 {\mathrm{d}}x \rightarrow 0,\,\, \mbox{as}\,\, n \to +\infty.$$ By using , we obtain $\|u_n\|_{X_{0,A}} \rightarrow 0$, as $n \to +\infty$. This contradicts .
**Case 3:** $c\not=0$ and $u\not=0$
Consider $v_n = \dfrac{u_n}{\|u_n\|_{X_{0,A}}}$ and $v= \dfrac{u}{\lim\|u_n\|_{X_{0,A}}}$.
It is clear that $v_n \rightharpoonup v$ weakly in $X_{0,A}$. If $\|v\|_{X_{0,A}} =1$ we conclude the proof. Then, we assume that $\|v\| _{X_{0,A}}<1$.
**Claim:** There exist $q>1$ sufficiently close to $1$ and $\beta > \beta_0$ sufficiently close to $\beta_0$, such that $$q \beta \|u_n \|^2_{X_{0,A}} < \frac{\pi}{1-\|v\|_{X_{0,A}}^2},$$ for $n$ large enough. As a consequence this claim and Lemma \[tipolions\] we have that holds, and we can see as in the **Case 2** that $u_n \rightarrow u$ strongly in $X_{0,A}$. So to complete the proof is enough to prove this statement.
Notice that, up to a subsequence, $$\label{EJ221x}
\lim_{n \to +\infty}\frac{1}{2}\|u_n\|^2 _{X_{0,A}}= c + \frac{\lambda}{2}\|u\|_2^2 + \int_{\Omega} F(|u|) {\mathrm{d}}x.$$ Denote by $$B := \left(c + \frac{\lambda}{2}\|u\|_2^2 + \int_{\Omega} F(|u|) {\mathrm{d}}x\right) (1-\|v\|^2_{X_{0,A}}).$$ Then $$B = c - \mathcal{I}_{A,\lambda}(u)$$ and consequently, $$\lim_{n \to +\infty}\frac{1}{2}\|u_n\|^2_{X_{0,A}} = \frac{B}{1-\|v\|^2_{X_{0,A}}} =\frac{c- \mathcal{I}_{A,\lambda}(u)}{1-\|v\|^2_{X_{0,A}}} < \frac{\pi}{2 \beta_0(1-\|v\|^2_{X_{0,A}})}.$$ This implies the claim.
Mountain pass case when $0<\lambda<\lambda_1$
=============================================
In this case we will use the Mountain Pass Theorem due to Ambrosetti and Rabinowitz [@AR1973].
Let $J: H\to \mathbb R$ be a $C^1$ functional on a Banach space $(H, \; \|\cdot \|)$ satisfying
- there exists some $\beta>0$ such that $J$ satisfies the Palais-Smale condition, $(PS)_c$ in short, for all $c\in (0, \beta)$,
- there exist constants $\rho, \delta>0$ such that $J(u)\geq \delta$ for all $u\in H$ satisfying $\|u\|=\rho$.
- $J(0)<\delta$ and $J(v)<\delta$ for some $v\in H\setminus \{0\}$ with $\|v\|\neq 0$.
Consider $\Gamma:=\{\eta\in C([0, 1],H): \eta(0)=0 \;\text{ and }\; \eta(1)=v\}$ and set $
c_M=\inf_{\eta\in \Gamma}\max_{t\in [0, 1]} J(\eta(t))\geq \delta.
$ Then $c_M\in (0, \beta)$ and it is a critical point of the functional $J$.
In the following propositions we will show the above geometry.
\[MPBelow1\] Assume that $f$ satisfies $(H_1)$. Then there exist $\tilde{M}>0$ and $u\in X_{0, A}$ such that $\mathcal I_{A, \lambda}(u)<-\tilde{M}$ for all $\lambda>0$.
Let us fix some $u_0\in H_k$ with $u_0 \not=0$. Let us introduce the scalar map $\psi: \mathbb R\to \mathbb R$ defined as $\psi(t)=\mathcal I_{A, \lambda}(tu_0)$. Now from assumption $(H_1)$, there are $\mu>2$ and constants $C_1, C_2>0$ such that $$\label{implicationH1}
F(|u|)\geq C_1|u|^\mu-C_2 .$$ Using , and equivalence of $ X_{0, A}$ and $L^\mu$ norms, we get $$\psi(t)\leq \frac{t^2}{2}\|u_0\|_{X_{0,A}}^2-C_1 t^\mu \|u_0\|_{X_{0,A}}^\mu +2C_2$$ which implies that $\psi(t)\to -\infty$ as $t\to \infty$. Hence the result follows.
\[MPBelow\] Assume that $f$ satisfies $(H_3)$. Then there exist $\delta, \rho>0$ such that $\mathcal I_{A,\lambda}(u)\geq \delta$ for $u\in X_{0, A}$ satisfying $\|u\|_{X_{0,A}}=\rho.$
From assumption $(H_3)$, given $\epsilon>0$, there exists $\delta=\delta(\epsilon)>0$ such that $$\frac{2F(t)}{t^2}\leq \ell+\epsilon\;\;\text{for all}\;\; |t|\leq \delta.$$ By the exponential critical growth assumption on the nonlinearity $f$, there exist $C=C(\epsilon)>0$ and $\beta>\beta_0$ such that $$F(|u|)\leq \frac{\ell+\epsilon}{2}|u|^2+C |u|^3 e^{\beta|u|^2},$$ which implies $$\mathcal I_{A, \lambda} (u)\geq \frac{1}{2}\|u\|_{X_{0,A}}^2-\frac{\lambda}{2}\|u\|_2^2-\frac{\ell+\epsilon}{2}\|u\|_2^2-C\int_\Omega |u|^2e^{\beta_0|u|^2} \, {\mathrm{d}}x.$$ Therefore, for $\|u\|_{X_{0,A}}=\rho>0$ sufficiently small such that $\beta \rho^2 < \pi$, by Trudinger-Moser inequality and Hölder’s inequality, we reach $$\mathcal I_{A, \lambda} (u)\geq \frac{1}{2}\|u\|_{X_{0,A}}^2-\frac{\lambda}{2}\|u\|_2^2-\frac{\ell+\epsilon}{2}\|u\|_2^2-C \|u\|^3_{X_{0,A}}=\frac{1}{2}\|u\|^2_{X_{0,A}}-\frac{\lambda+\ell+\epsilon}{2}\|u\|_2^2-C \|u\|^3_{X_{0,A}}$$ possibly for different constant $C>0$. Now using the characterization of $\lambda_1$, we get $$\mathcal I_{A, \lambda} (u)\geq \frac{1}{2}\left(1-\frac{\lambda+\ell+\epsilon}{\lambda_1}\right)\|u\|^2_{X_{0,A}}-C \|u\|^3_{X_{0,A}}.$$ Observe that for a given $\epsilon>0$, sufficiently small, $\lambda+\ell+\epsilon<\lambda_1$. Hence $$\mathcal I_{A, \lambda} (u)\geq C_1\|u\|^2_{X_{0,A}}-C \|u\|^3_{X_{0,A}},$$ where $C_1=\frac{1}{2}\left(1-\frac{\lambda+\ell+\epsilon}{\lambda_1}\right)>0$. Next we denote $g(t)=C_1\rho^2-C\rho^3$ and observe that $g(\rho)\to 0$ as $\rho\to 0$. Hence for sufficiently small $\rho>0$ there exists $\delta>0$ such that $g(\rho)\geq \delta>0$. This completes the proof of the result.
The minimax level
-----------------
To show some estimates on the minimax level we require some facts on the Moser’s functions defined by Moser [@moser] (see also [@Takahasi] for the fractional case). The Moser’s functions are defined as follows $$M_n(x)=\left\{\begin{array}{ll}
\displaystyle \sqrt{\log n}, & |x|< \frac{1}{n}, \\
\displaystyle \frac{\log(1/|x|)}{\sqrt{\log n}}, & \frac{1}n\leq |x|< {1}, \\
0,& |x|\geq {1}. \\
\end{array}\right.$$ The following proposition deals with the asymptotic estimates on Moser’s sequence.
The following estimates are satisfied by $M_n$
(a)
: $\|M_n\|_2^2=\frac{4}{\log n} + o_n(1)$;
(b)
: $\|M_n\|^2_{X_{0,A}}\leq \pi+ O\left(\frac{1}{\log n}\right)$.
The proof of item (a) follows from Takahasi [@Takahasi] (see estimate in Equation (2.5)). To prove the item (b), we will use Euler’s formula $(e^{i\theta}=\cos \theta+i \sin \theta)$ with the notation $\xi_A= \mathrm{i} (x-y)\cdot
A(\frac{x+y}{2})$ as below $$\begin{aligned}
\|M_n\|^2_{X_{0,A}}&= \iint_{\mathbb{R}^{2}}\frac{|M_n(x)- e^{\mathrm{i} (x-y)\cdot
A(\frac{x+y}{2})}M_n(y)|^2}{|x-y|^{2}}\, {\mathrm{d}}x {\mathrm{d}}y \\
&=\iint_{\mathbb{R}^{2}}\frac{|M_n(x)- \cos \xi_A M_n(y)-i \sin \xi_A M_n(y) |^2}{|x-y|^{2}}\, {\mathrm{d}}x {\mathrm{d}}y \\
&=\iint_{\mathbb{R}^{2}}\frac{|M_n(x)- \cos \xi_A M_n(y)|^2+|\sin \xi_A M_n(y) |^2}{|x-y|^{2}}\, {\mathrm{d}}x {\mathrm{d}}y\\
&=\iint_{\mathbb{R}^{2}}\frac{|M_n(x)- M_n(y)|^2+2(1-\cos \xi_A) M_n(x)M_n(y)}{|x-y|^{2}}\, {\mathrm{d}}x {\mathrm{d}}y.
\end{aligned}$$ Hence $$\label{1223}
\|M_n\|^2_{X_{0,A}} = [M_n]^2_{H^s(\mathbb R)}+2\iint_{\mathbb{R}^{2}}\frac{(1-\cos \xi_A) M_n(x)M_n(y)}{|x-y|^{2}}\, {\mathrm{d}}x {\mathrm{d}}y.$$ Let us estimate the second integral in the above equation as follows. Denote $$I = \int_{\mathbb{R}^2} \frac{2(1-\cos(\xi_A(x,y))M_n(x) M_n(y)}{|x-y|^2} {\mathrm{d}}x {\mathrm{d}}y =2( I_1 + I_2),$$ where $$I_1 = \int_{|x-y| < \delta} \frac{(1-\cos(\xi_A(x,y))M_n(x) M_n(y)}{|x-y|^2} {\mathrm{d}}x {\mathrm{d}}y$$ and $$I_2 = \int_{|x-y| \geq \delta} \frac{(1-\cos(\xi_A(x,y))M_n(x) M_n(y)}{|x-y|^2} {\mathrm{d}}x {\mathrm{d}}y,$$ with $\delta >0$ to be chosen later.
In order to estimate $I_1$, notice that since $A \in L^\infty(\mathbb{R})$, we have $$\lim_{|x-y|\rightarrow 0} \left|\frac{(1-\cos(\xi_A(x,y))}{|x-y|^2} \right|=
\lim_{|x-y|\rightarrow 0} \left|\frac{(1-\cos(\xi_A(x,y))}{\xi^2_A(x,y)} A^2\left(\frac{x+y}{2}\right) \right|=
C_A.$$ Then, given $\varepsilon >0$, there exists $\delta=\delta(\varepsilon)>0$ such that $$\left|\frac{(1-\cos(\xi_A(x,y))}{|x-y|^2} \right| \leq C_A + \varepsilon,\,\, \mbox{for all}\,\, |x-y| < \delta.$$ From this we get that $$\begin{aligned}
I_1 &\leq & (C_A + \varepsilon )\int_{|x-y| < \delta} M_n(x) M_n(y) {\mathrm{d}}x {\mathrm{d}}y\\
& \leq & \frac{1}{2}(C_A + \varepsilon )\int_{|x-y| < \delta} (M^2_n(x) + M^2_n(y)) {\mathrm{d}}x {\mathrm{d}}y\\
& = & (C_A + \varepsilon )\int_{|x-y| < \delta} M^2_n(x) {\mathrm{d}}x {\mathrm{d}}y\\
& = & 2 \delta (C_A + \varepsilon ) \|M_n\|_2^2.
\end{aligned}$$ For the integral $I_2$, notice that $$\begin{aligned}
I_2 &\leq & 2 \int_{|x-y| \geq \delta} \frac{M_n(x) M_n(y)}{|x-y|^2} {\mathrm{d}}x {\mathrm{d}}y\\
& \leq & \int_{|x-y| \geq \delta} \frac{M^2_n(x) + M^2_n(y)}{|x-y|^2} {\mathrm{d}}x {\mathrm{d}}y\\
& = & 2\int_{|x-y| \geq \delta} \frac{M^2_n(x) }{|x-y|^2} {\mathrm{d}}x {\mathrm{d}}y\\
& = & \frac{4}{\delta} \|M_n\|_2^2.
\end{aligned}$$ Combining $I_1$ and $I_2$, we obtain $$I \leq \left(4 \delta (C_A + \varepsilon )+ \frac{8}{\delta} \right) \|M_n\|_2^2.$$ Using this last estimate, and (a), we reach (b). This completes the proof of the lemma.
Now, we define the minimax level of $\mathcal I_{A, \lambda}$ by $$\label{EL28}
c(n) : = \inf_{\gamma \in \Gamma} \max_{t \in [0,1]} \mathcal I_{A, \lambda} (\gamma(t)),$$ where $$\Gamma = \{\gamma \in C([0,1], X_{0,A})\,:\, \gamma(0)=0\,\, \mbox{and}\,\, \gamma(1) = R_n z_n\},$$ $R_n$ being such that $\mathcal I_{A, \lambda} (R_n z_n) \leq 0$ and $z_n = \dfrac{M_n}{\|M_n\|_{X_{0,A}}}$.
\[levelminimax\] Assume that $f$ satisfies $(H_4)$. Then there exists $n$ large enough such that $$c(n)< \dfrac{\pi}{2 \beta_0}.$$
It is sufficient to find $n\in \mathbb{N}$ such that $$\label{EL49}
\max_{t \geq 0} \mathcal I_{A, \lambda}(t z_n) < \frac{\pi}{2 \beta_0}.$$ Suppose by contradiction that does not hold. So, for all $n$, this maximum is larger than or equal to $\frac{\pi}{2 \beta_0}$ (it is indeed a maximum, in view of Proposition \[MPBelow1\] and Proposition \[MPBelow\]). Let $t_n >0$ be such that $$\label{EL410}
\mathcal I_{A, \lambda}(t_n z_n) = \max_{t \geq 0} \mathcal I_{A, \lambda}(t z_n).$$ Then, for all $n \in \mathbb{N}$, $$\mathcal I_{A, \lambda}(t_n z_n) \geq \frac{\pi}{2 \beta_0},$$ and, consequently, for all $n \in \mathbb{N}$, we have $$\label{EL411}
t_n^2 \geq \frac{\pi}{ \beta_0}.$$ Let us prove that $t_n^2 \rightarrow \dfrac{\pi}{ \beta_0}$ as $n \to +\infty$. From , we know that $$\frac{d}{dt}(\mathcal I_{A, \lambda}(t z_n))=0\,\,\mbox{when}\,\, t=t_n.$$ Multiplying this last equation by $t_n$ and observing that $\|z_n\|_{X_{0,A}}=1$, we have, for $n$ large enough, that $$\label{EL413}
t_n^2 \geq \int_{B_1(0)} f(t_n z_n) t_n^2 z_n^2 {\mathrm{d}}x.$$ By $(H_4)$, it follows that given $\zeta >0$, there exists $t_{\zeta} >0$ such that $$f(t) t^2 \geq \zeta e^{\beta_0 t^2}\,\,\mbox{for all}\,\, t \geq t_\zeta.$$ From , for $n$ large enough, we obtain $$\label{EL414}
t_n^2 \geq \zeta \int_{B_{1/n}(0)} e^{\beta_0 t_n^2 z_n^2} {\mathrm{d}}x \geq 2\zeta e^{\log n \left(\frac{\beta_0 t_n^2}{\pi + C\log n)^{-1}} -1 \right)},$$ which implies that $\{t_n\}$ is bounded sequence. Moreover, together with gives us that $t_n^2 \searrow \dfrac{\pi}{ \beta_0}$ as $n \to +\infty$.
In order to conclude the proof, observe that from and , we obtain $$t_n^2 \geq 2\zeta e^{\log n \left(\frac{\beta_0 t_n^2 - \pi - C(\log n)^{-1}}{\pi + C(\log n)^{-1}} \right)} \geq 2 \zeta e^{ \frac{ - C}{\pi + C(\log n)^{-1}}},$$ which implies $$\zeta \leq \dfrac{\pi e^{\frac{C}{\pi}}}{2 \beta_0}.$$ Since $\zeta$ is arbitrarily large, we get a contradiction. This completes the proof of the proposition.
Proof of Theorem \[thm1\]
-------------------------
To conclude Theorem \[thm1\] we use Propositions \[MPBelow1\], \[MPBelow\] and \[levelminimax\], and apply the Theorem A.
Linking Case when $\lambda_k<\lambda<\lambda_{k+1}$ for
========================================================
In this case we use the following critical point theorem known as Linking theorem due to Ambrosetti and Rabinowitz [@AR1973].
Let $J: H\to \mathbb R$ be a $C^1$ functional on a Banach space $(H, \; \|\cdot \|)$ such that $H=H_1\oplus H_2$ with $\dim H_1<\infty$. If $J$ satisfies the following
- there exists some $\beta>0$ such that $J$ satisfies the Palais-Smale condition, $(PS)_c$ in short, for all $c\in (0, \beta)$,
- there exist constants $\rho, \delta>0$ such that $J(u)\geq \delta$ for all $u\in H_2$ satisfying $\|u\|=\rho$.
- there exists a $z\not \in H_1$ with $\|z\|=1$ and $R>\rho$ such that $J(u)\leq 0$ for all $u\in \partial Q$, where $Q=\{v+sz:
v\in H_1, \|v\|\leq R \;\;\text{and}\;\; 0\leq s\leq R\}$.
Then $c\in(0, \beta)$ defined as $c=\inf_{\eta\in \Gamma}\max_{u\in Q} J(\eta(u)) \geq \delta$, where $\Gamma=\{\eta \in C(\overline Q, H): \eta(u)=u,\;\;\text{if}\;\; u\in \partial Q\}$, is a critical value of $J$.
In the following propositions we will show the above geometry.
\[LTBelow\] Let $\lambda_k<\lambda<\lambda_{k+1}$ and $f$ satisfies $(H_3)$. Then there exists $a, \rho>0$ such that $$\mathcal I_{A, \lambda}(u)\geq a,\;\;\text{for}\;\; \|u\|_{X_{0, A}}=\rho\;\;\text{and}\;\; u\in H_k^\perp$$
Proof follows the same lines as in Proposition \[MPBelow\] using the characterization of $\lambda_{k+1}$. We remark that we do not require $(AR)$ condition in the proof.
\[lg2\] Let $\lambda_k<\lambda<\lambda_{k+1}$ and $(H_1)$ holds. Define $Q=\{v+sz: v\in H_k, \|v\|\leq R \;\;\text{and}\;\; 0\leq s\leq R\;\;\text{for some }\;R>\rho\}$, where $\rho$ is given in Proposition \[LTBelow\] and $z \in W$ with $\|z\|_{X_{0,A}}=1$. Then $\mathcal I_{A, \lambda}(u)\leq 0$ for all $u\in \partial Q$.
For some given $R>0$, let us split $\partial Q$ into following three parts $$\begin{aligned}
Q_1&=\{u\in H_k: \|u\|_{X_{0, A}}\leq R\};\\
Q_2&=\{ u+sz: u\in H_k, \|u\|_{X_{0, A}}=R\;\;\text{and}\;0\leq s\leq R\};\\
Q_3&=\{ u+Rz: u\in H_k, \|u\|_{X_{0, A}}\leq R\}.
\end{aligned}$$ Now let us compute the energy functional in each of the above splitted boundary components of $Q$. If $u \in Q_1$, then using the characterization of $\lambda_k$ as in , we get $$\mathcal{I}_{A, \lambda}(u)\leq \frac{1}{2}\left(1-\frac{\lambda}{\lambda_k}\right)\|u\|_{X_{0, A}}^2\leq \frac{1}{2}\left(1-\frac{\lambda}{\lambda_k}\right)R^2<0$$ for any choice of $R>0$. Before verifying the claim on $Q_2$ and $Q_3$, let us observe the following. Let us fix some $u_0\in H_k$. and introduce the scalar map $\psi: \mathbb R\to \mathbb R$ defined as $\psi(t)=\mathcal I_{A, \lambda}(tu_0)$. Now from a direct implication of assumption $(H_1)$, there exists $\mu>2$ and $C_1, C_2>0$ such that $$\label{implicationH1Re}
F(|u|)\geq C_1|u|^\mu-C_2 .$$ Using and using equivalence of $ X_{0, A}$ and $L^\mu$ norms, as $Q_2$ is finite dimensional $(\mathrm{dim}=k+1)$, we get $$\psi(t)\leq \frac{t^2}{2}\|u_0\|_{X_{0,A}}^2-\lambda \frac{t^2}{2}\|u_0\|_{X_{0,A}}^2-C_1 t^\mu \|u_0\|_{X_{0,A}}^\mu +2C_2$$ which implies that $\psi(t)\to -\infty$ as $t\to \infty$. Now for any $u\in Q_2$, there exist $v\in H_k$ and $0\leq s\leq R$ such that $u=v+s\omega$. Moreover, $$\|u\|_{X_{0, A}}^2=\|v+sz\|_{X_{0, A}}^2=\|v\|_{X_{0, A}}^2+s^2\|z\|_{X_{0, A}}^2\geq \|v\|^2_{X_{0, A}}=R^2.$$ Therefore if we choose $R>0$ sufficiently large, we have $\mathcal I_{A, \lambda}(u)<0$. Now if $u \in Q_3$ then there exists some $v\in H_k$ such that $u=v+Rz$. Moreover, $$\|u\|_{X_{0, A}}^2=\|v+Rz\|_{X_{0, A}}^2=\|v\|_{X_{0, A}}^2+R^2\|z\|_{X_{0, A}}^2\geq R^2.$$ Now following the similar argument as above one can prove the conclusion of the Proposition for choosing $R>0$ large enough.
The minimax level
-----------------
For the matter we have to select a $z \in W$ such that $\|z\|_{X_{0,A}}=1$ and $\mathcal{I}_{A, \lambda}(u)<\frac{\pi}{2\beta_0}$ for all $u \in Q$.
Let $P_k : X_{0,A} \rightarrow H_k^{\bot}$ be the orthogonal projection. Define $$\label{BEL32}
W_n(x) = P_k(M_n(x)).$$ We need some estimates for $W_n$, which are shown in the next lemma. Before that, knowing that $H_k$ has finite dimension, consider $A_k > 0$ and $B_k > 0$ such that $$\label{BEL33}
\|u\|_{X_{0,A}} \leq A_k \|u\|_2\,\, \textrm{and}\,\, \|u\|_\infty \leq \frac{B_k}{B} \|u\|_2,\,\, \textrm{for all}\,\, u \in H_k,$$ where $B > 0$ is such that $\|M_n\|_2 \leq (\frac{B}{\log n})^{1/2}$ for all $n\in \mathbb N$.
\[lema511\] Let $W_n$ be as defined in . Then the following estimates hold:
(i)
: $1 - \frac{A_k}{\log n} \leq \|W_n\|^2_{X_{0,A}} \leq 1 + O((\log n)^{-1})$;
(ii)
: $W_n(x) \geq \left\{
\begin{array}{ll}
\frac{-B_k}{\sqrt{\log n}}, & \hbox{for all}\,\, x \in (-1,1); \\
\sqrt{\log n} - \frac{B_k}{\sqrt{\log n}}, & \hbox{for all}\,\, x \in (\frac{-1}{n},\frac{1}{n}).
\end{array}
\right.
$
To prove (i) one needs only to notice that $$\begin{aligned}
\|W_n\|^2_{X_{0,A}} &= \|M_n\|^2_{X_{0,A}}
- \|(I - P_k)M_n\|^2_{X_{0,A}} \;\;\textrm{and}\;\;(I-P_k)M_n\in H_k.
\end{aligned}$$ The estimate will follow because of . On the other hand, to verify $(ii)$, as $M_n \geq 0$ in $(-1,1)$ and $M_n = \sqrt{\log n}$ in $(\frac{-1}{n}, \frac{1}{n})$, we have $$W_n(x) = M_n(x) -(I-P_k)(M_n(x)) \geq \left\{
\begin{array}{ll}
-\|(I-P_k)M_n\|_\infty, & \hbox{if}\,\, x \in (-1,1); \\
\sqrt{\log n} - \|(I-P_k)M_n\|_\infty, & \hbox{if}\,\, (\frac{-1}{n}, \frac{1}{n}),
\end{array}
\right.$$ where the inequality follows by observing the definition of $B_k$ in .
In the following, we define $z_n(x)=\frac{W_n(x)}{\|W_n\|_{X_{0, A}}}$, $Q_n=\{v+sz_n: v\in H_k, \|v\|\leq R \;\;\text{and}\;\; 0\leq s\leq R\;\;\text{for some }\;R>\rho\}$ and the minimax level of $\mathcal I_{A, \lambda}$ as follows $$\label{LinLevel}
c(n)=\inf_{\nu\in \Gamma} \sup_{w\in \nu(Q_n)}\mathcal I_{A, \lambda}(w),$$ where $$\Gamma=\{ \nu\in C(Q_n, H):\nu(w)=w \;\text{if} \; w\in \partial Q_n\}.$$ We have the following estimate for above minimax level.
\[minimaxlev2\] Let $c(n)$ be given as in and assumption $(H_1)-(H_5)$ hold. Then for large $n$, $c(n)<\frac{\pi}{2\beta_0}$.
From the definition of $c(n)$, it is enough to show that $$\label{de54}
max\{\mathcal I_{A, \lambda}(v+sz_{n}): v\in H_k, \|v\|_{X_{0, A}}\leq R, 0\leq s\leq R\}<\frac{\pi}{2\beta_0}.$$ Let us proceed by contradiction. Suppose that does not hold, then $$max\{\mathcal I_{A, \lambda}(v+sz_{n}): v\in H_k, \|v\|_{X_{0, A}}\leq R, 0\leq s\leq R\}\geq\frac{\pi}{2\beta_0}.$$ Let $u_n=v_n+s_nz_n$ be the point of maximum in the above expression with $v_n \in H_k$. Then
$$\label{contrassump}
\mathcal I_{A, \lambda}(v_n+s_nz_{n})\geq \frac{\pi}{2\beta_0}.$$
Moreover, since $\mathcal I_{A, \lambda}^\prime (u_n)=0$, we have
$$\label{final1}
\|u_n\|_{X_{0, A}}^2-\lambda \int_\Omega |u_n|^2 dx-\int_\Omega f(|u_n|)|u_n|^2dx=0.$$
Now we finish the proof of the Proposition in following few steps.\
**Step 1:** We claim that $\{v_n\}$ and $\{s_n\}$ are bounded sequences in respective topologies.
There are either of the following two posibilities
- $\dfrac{s_n}{\|v_n\|_{X_{0, A}}}\geq C_0$ for some $C_0>0$ uniformly in $n$.
- $\dfrac{s_n}{\|v_n\|_{X_{0, A}}}\to 0$ in $\mathbb R$, up to a subsequence, as $n\to \infty$.
Suppose $(i)$ holds true. Note that the boundedness of $\{s_n\}$ implies the sequence $\{v_n\}$ is also bounded as $\|v_n\|_{X_{0, A}}\leq s_n/C_0$. Hence, we aim to prove the boundedness of $s_n$ in light of item $(i)$ as above. For, there exists a constant $C$ such that $$\|u_n\|_{X_{0, A}}=\|v_n+s_nz_n\|_{X_{0, A}}\leq \|v_n\|_{X_{0, A}}+s_n\|z_n\|_{X_{0, A}}\leq \frac{s_n}{C_0}+s_n\leq Cs_n.$$ Now from and assumption $(H_4)$, given $\zeta>0$, there exists $t_\zeta>0$ large enough such that $f(t)t^2\geq \zeta e^{\beta_0t^2}$ for all $t>t_\zeta$, we get $$\label{bdd2}
Cs_n^2\geq \int_{B_{1/n}\cap \{|u_n|\geq t_\zeta\}}f(|u_n|)|u_n|^2 {\mathrm{d}}x\geq \zeta\int_{B_{1/n}\cap \{|u_n|\geq t_\zeta\}} e^{\beta_0 |u_n|^2} {\mathrm{d}}x.$$ Now to estimate the integral in the above inequality in right hand side, from Lemma \[lema511\], we have in $B_{1/n}$ for large $n$ and $\epsilon \in (0,1)$, that $$u_n(x)=v_n(x)+s_nz_n(x)=\frac{s_n(\sqrt\log n-\frac{B_k}{\log n})}{\|M_n\|_{X_{0, A}}}\left(\frac{v_n(x)\|M_n\|_{X_{0, A}}}{s_n (\sqrt\log n-\frac{B_k}{\log n})}+1\right)\geq (1-\epsilon)\frac{s_n\sqrt\log n}{\|M_n\|_{X_{0, A}}}.$$ Hence from and using $\|M_n\|_{X_{0,A}}^2 \leq \pi + C(\log n)^{-1}$, we get $$Cs_n^2\geq \zeta\int_{B_{1/n}} e^{\frac{\beta_0(1-\epsilon)^2 s_n^2\log n}{\|M_n\|^2_{X_{0, A}}} } {\mathrm{d}}x= 2\zeta e^{\left(\frac{\beta_0(1-\epsilon)^2 s_n^2}{\pi + C(\log n)^{-1}}-1\right)\log n},$$ which implies $$\label{conclusive}
C s_n^2\geq 2 \zeta e^{\left(\frac{\beta_0(1-\epsilon)^2 s_n^2}{\pi+ C(\log n)^{-1}}-1\right)\log n}.$$ Therefore if $s_n\to \infty$, it contradicts the above inequality. Hence $\{s_n\}$ is a bounded sequence so is $\{v_n\}$.
Next we assume that $(ii)$ occurs. Then $s_n\leq \|v_n\|_{X_{0, A}}$ which implies $\|u_n\|_{X_{0, A}}=\|v_n+s_nz_n\|_{X_{0, A}}\leq 2\|v_n\|_{X_{0, A}}$. Note that if the sequence $\{\|v_n\|\}$ is bounded in $X_{0, A}$ then the sequence $\{s_n\}$ is bounded in $\mathbb R$. Thus we aim to show that $\{\|v_n\|\}$ is bounded in $X_{0, A}$. Assume by contradiction that $\|v_n\|_{X_{0, A}} \to \infty$. From , we have $$\label{claim1sec}
1\geq \int_{\{|u_n|>t_\zeta\}}\frac{f(|u_n|)|u_n|^2}{\|u_n\|_{X_{0, A}}^2}{\mathrm{d}}x\geq \frac{\zeta}{4}\int_{\{|u_n|>t_\zeta\}}\frac{e^{\beta_0 |u_n|^2}}{\|v_n\|_{X_{0, A}}^2} {\mathrm{d}}x.$$ Observe that $$\label{pointwise}
\frac{u_n}{\|v_n\|_{X_{0, A}}}\chi_{\{u_n\geq t_\zeta\}}=\frac{v_n}{\|v_n\|_{X_{0, A}}}+\frac{s_n}{\|v_n\|_{X_{0, A}}}z_n-\frac{u_n}{\|v_n\|_{X_{0, A}}}\chi_{\{u_n<t_\zeta\}}.$$ Since, $\dfrac{s_n}{\|v_n\|_{X_{0, A}}}\to 0$ in $\mathbb R$ and $z_n\to 0$ pointwise almost everywhere in $(-1, 1)$, there exists $v_0\in H_k$ such that $$\label{pointwise2}
\frac{u_n(x)}{\|v_n\|_{X_{0, A}}}\chi_{\{u_n\geq t_\zeta\}}\to v_0\;\; a. e. \; \text{in}\;\; (-1, 1)$$ with $\dfrac{v_n}{\|v_n\|_{X_{0, A}}}\to v_0$ and $\|v_0\|_{X_{0, A}}=1$. Then using $\|v_n\|_{X_{0, A}} \to +\infty$, , and Fatou’s Lemma, we get $$1\geq\frac{\zeta}{4}\int_\Omega\frac{e^{\beta_0 \|v_n\|_{X_{0, A}}^2 (\frac{u_n}{\|v_n\|_{X_{0, A}}}\chi_{\{u_n\geq t_\epsilon\}})^2}}{\|v_n\|_{X_{0, A}}^2} {\mathrm{d}}x \to +\infty\,\,\mbox{as}\,\, n \to +\infty,$$ which is a contradiction. Hence the proof of the claim.
**Step 2:** From step 1, we can assume that there exists $v_0\in X_{0, A}$ and $s_0\in \mathbb R$ such that $v_n\to v_0$ and $s_n\to s_0$, up to a subsequence. Now we claim that $v_0=0$ and $s_0^2=\frac{\pi}{\beta_0}$.
First we show that $s_0^2\geq \frac{\pi}{\beta_0}$. By the definition $$\|u_n\|_{X_{0,A}}^2\to \|v_0\|_{X_{0,A}}^2+s_0^2.$$ Moreover, using $v_n\to v_0$ in $X_{0, A}$, $\|z_n\|_2\to 0$ and Cauchy Schwartz inequality together with embeddings of $X_{0, A}\hookrightarrow L^2((-1, 1))$, we get $u_n\to v_0$ in $L^1((-1, 1)).$ From , we obtain $$\int_\Omega f(|u_n|)|u_n|^2\, {\mathrm{d}}x \leq C.$$ Consequently, from [@FMR Lemma 2.1], we get $\int_{\Omega} f(|u_n|)|u_n|{\mathrm{d}}x \rightarrow \int_{\Omega} f(|v_0|)|v_0|{\mathrm{d}}x$. Thus, by applying $(H_1)$ and the Generalized Lebesgue Dominated Convergence Theorem we have $$\label{djconv}
\int_\Omega F(|u_n|){\mathrm{d}}x\to \int_\Omega F(|v_0|) {\mathrm{d}}x.$$ In light of , and , we get $$\frac{\pi}{2\beta_0}\leq \lim_{n\to \infty}\mathcal I_{A, \lambda}(u_n)= \frac{1}{2}\|v_0\|_{X_{0, A}}^2-\frac{\lambda}{2}\|v_0\|_2^2-\int_\Omega F(|v_0|){\mathrm{d}}x+\frac{s_0^2}{2}\leq \frac{1}{2}\left(1-\frac{\lambda}{\lambda_k}\right)\|v_0\|^2_{X_{0, A}}+\frac{s_0^2}{2}.$$ Since $\lambda \in (\lambda_k, \lambda_{k+1})$, $s_0^2\geq {\pi}/{\beta_0}$.
Now we follow the idea of alternatives as in the step 1. Note that, since $s_0^2\geq {\pi}/{\beta_0}$ and $\|v_n\|_{X_{0, A}}\leq C$ for some $C>0$, the alternative $(ii)$ is not possible to hold. Hence suppose $(i)$ holds. Then, from , we have $$\frac{\beta_0(1-\epsilon)^2 s_0^2}{\pi}-1\leq 0$$ which implies $s_0^2\leq \pi/\beta_0$. Hence the proof.
Next we show that $v_0\equiv 0$. From , using , $v_n\to v_0$ in $X_{0, A}$, $\|z_n\|_{X_{0, A}}=1$, $\|z_n\|_2\to 0$ and $s_n\to s_0$ in $\mathbb R$, we get the following $$\lim_{n\to \infty}\mathcal I_{A, \lambda}(u_n)=\mathcal I_{A, \lambda}(v_0)+\frac{s_0^2}{2}\geq \frac{\pi}{2\beta_0}$$ which implies that $$\mathcal I_{A, \lambda}(v_0)\geq 0.$$ Moreover, by the definition, $$\mathcal I_{A, \lambda}(v_0)=\frac{1}{2}\|v_0\|_{X_{0, A}}^2-\frac{\lambda}{2}\|v_0\|_2^2-\int_\Omega F(|v_0|) {\mathrm{d}}x\leq \frac{1}{2}\|v_0\|_{X_{0, A}}^2-\frac{\lambda}{2}\|v_0\|_2^2\leq \frac{1}{2}\left(1-\frac{\lambda}{\lambda_k}\right)\|v_0\|^2_{X_{0, A}}\leq 0.$$ Hence from above two inequalities $\mathcal I_{A, \lambda}(v_0)=0$. Since $v_0\in H_k$, we have $$0=\mathcal I_{A, \lambda}(v_0)=\frac{1}{2}\left(1-\frac{\lambda}{\lambda_k}\right)\|v_0\|^2_{X_{0, A}}-\int_\Omega F(|v_0|) {\mathrm{d}}x\leq -\int_\Omega F(|v_0|) {\mathrm{d}}x$$ and by the nature of the nonlinearity $$\int_\Omega F(|v_0|) {\mathrm{d}}x\geq 0.$$ On combining these two estimates, we have $$\int_\Omega F(|v_0|) {\mathrm{d}}x=0$$ which implies, from $\mathcal I_{A, \lambda}(v_0)=0$, that $\|v_0\|_{X_{0, A}}=0$. It completes the proof.
In order to conclude the proof, observe that from step 2, up to a subsequence, we have $v_n\to 0$ strongly in $H_k$ and $s_n\to s_0$. Then holds, that is, $$C s_n^2\geq 2 \zeta e^{\left(\frac{\beta_0(1-\epsilon)^2 s_n^2}{\pi+ C(\log n)^{-1}}-1\right)\log n}.$$ Letting $\epsilon \to 0^+$ and later $n \to +\infty$, we get $$\zeta \leq \dfrac{C \pi e^{\frac{C}{\pi}}}{2 \beta_0}.$$ Since $\zeta$ is arbitrarily large, we get a contradiction. This completes the proof of the proposition.
Proof of Theorem \[thm2\]
-------------------------
To conclude Theorem \[thm2\] we use Propositions \[LTBelow\], \[lg2\] and \[minimaxlev2\], and apply the Theorem B.
Proof of Theorem \[thm3\]
=========================
In order to prove Theorem \[thm3\], we will use the following critical point theorem, see [@crlit Theorem 2.4].
\[GCPT\] Let $H$ be a real Hilbert space with the induced norm $\|\cdot\|$ and $J:H\rightarrow \mathbb R$ be a functional of class $C^1(H, \mathbb R)$ satisfying the following conditions:
$(A_1)$
: $J(0)=0$ and $J(-u)=J(u)$;
$(A_2)$
: $J$ satisfies the Palais-Smale condition, in short $(PS)_c$, for $c\in (0, \beta)$ and for some $\beta>0$;
$(A_3)$
: there exist closed subspaces $V, W$ of $H$ and constants $\rho, \delta, \eta$ with $\delta<\eta<\beta$ such that
$(i)$
: $J(u)\leq \eta$ for all $u\in W$;
$(ii)$
: $J(u)\geq \delta$ for any $u\in V$ with $\|u\|=\rho$;
$(iii)$
: $codim~(V)<\infty$ and $dim~W\geq codim~V$.
Then there exist at least $dim~W-codim~V$ pairs of critical points of the functional $J$ with critical values belonging to the interval $[\delta, \eta]$.
Our next aim is to apply Theorem \[GCPT\] in our variational setup. It is clear that the functional $\mathcal I_{A,\lambda} \in C^1(X_{0, A}, \mathbb R)$ and from the definition, $\mathcal I_{A,\lambda}(0)=0$. Since $|-u|=|u|$ implies $\mathcal I_{A,\lambda}(-u)=\mathcal I_{A,\lambda}(u)$. Hence the assumption $(A_1)$ is satisfied. Lemma \[PS\] implies that $I_{A, \lambda}$ satisfies the $(PS)_c$ condition for all $c\in (0, \frac{\pi}{2\beta_0} )$. Hence the assumption $(A_2)$ holds good with $\beta=\frac{\pi}{2\beta_0}$. Next we verify the assumption $(A_3)$. We consider $W= \displaystyle{\rm{span}_{\mathbb R} \{\varphi_1, \varphi _2,.....\varphi_{k+m-1}\}}$ and $$V= X_{0, A} \;\;\text{if}\;\; k=1, \;\;\text{otherwise}\;\;
V=\{u\in X_{0, A}: \langle u,\varphi_j\rangle=0\;\;\forall\; 1\leq j\leq k-1\}.$$ Then both $W$ and $V$ are closed subspaces of $X_{0, A}$ with $k+m-1=$ dim $W\geq$ codim $V=k-1$. Now take $u\in W$ then $u(x)=\displaystyle \sum_{j=1}^{k+m-1}\alpha_j\varphi_j(x)$ and by the orthogonality of eigenfunctions $$\|u\|^2_{X_{0,A}}=\displaystyle \sum_{j=1}^{k+m-1}\alpha_j^2\|\varphi_j\|^2_{X_{0,A}}=\displaystyle \sum_{j=1}^{k+m-1}\lambda_j\alpha_j^2\leq \lambda_k\displaystyle \sum_{j=1}^{k+m-1}\alpha_j^2=\lambda_k\|u\|_2^2=\lambda^*\|u\|_2^2.$$ Now using $(H_5)$, we have $F(|t|)\geq\frac{C_p}{p}|t|^p$, for all $t \in \mathbb{R}$. Thus for $u\in W$ $$\begin{aligned}
\mathcal I_{A,\lambda}(u)&=\frac{1}{2}\|u\|^2_{X_{0,A}}-\frac{\lambda}{2}\|u\|_2^2-\int_\Omega F(|u|)\, {\mathrm{d}}x\\ &\leq \frac{1}{2}(\lambda^*-\lambda)\|u\|_2^2-\int_\Omega F(|u|)\, {\mathrm{d}}x\\
&\leq \frac{1}{2}(\lambda^*-\lambda)2^{\frac{p}{p-2}}\|u\|_p^2-\frac{C_p}{p}\|u\|_p^p.\end{aligned}$$ Define $h(t)=\frac{1}{2}(\lambda^*-\lambda)2^{\frac{p}{p-2}}t^2-\frac{C_p}{p}t^p$ for $t\geq 0$, then $h(t)$ has a maximum at $t_0=\left(\frac{(\lambda^*-\lambda)}{C_p}2^{\frac{p}{p-2}}\right)^\frac{1}{p-2}$. Hence $$I_{A,\lambda}(u)\leq \eta=\left(\frac{1}{2}-\frac{1}{p}\right)\left(\frac{(\lambda^*-\lambda)}{C_p^{2/p}}2^{\frac{p}{p-2}}\right)^\frac{p}{p-2}.$$ Note that we can make $\eta$ to be arbitrary small positive number either by choosing $\lambda$ suitably close to $\lambda^*$ or by taking $C_p>0$ large enough in $(H_5)$. We will determine this closeness later.\
For the second part, we use Proposition \[MPBelow\] and Proposition \[LTBelow\]. Now only thing remains to show is the relation $$\label{relation}
\delta<\eta<\beta.$$ Note that the first inequality can be justified by choosing $\rho$ sufficiently small to make $\delta>0$ arbitrary small in Proposition \[MPBelow\] or in Proposition \[LTBelow\]. Hence $\delta<\eta$ holds good for $\rho>0$ sufficiently small. The ultimate task is to show that $\eta<\beta$. In other words, $$0<\left(\frac{1}{2}-\frac{1}{p}\right)\left(\frac{(\lambda^*-\lambda)}{C_p^{2/p}}2^{\frac{p}{p-2}}\right)^\frac{p}{p-2}<\frac{\pi}{2\beta_0}$$ which leads to a restriction on $\lambda$ and $C_p$ as $\lambda<\lambda^*$ and $$C_p>\left( \frac{\beta_0(p-2)}{\pi}\right)^\frac{p-2}{2}\left((\lambda^*-\lambda)2^{\frac{p}{p-2}}\right)^\frac{p}{2}>0$$ and therefore justifies the choices of $\lambda$ and $C_p$ as in Theorem \[thm3\]. Hence the proof of Theorem \[thm3\].
Proof of Theorem \[thm4\]
=========================
The proof of Theorem \[thm4\] is mainly based on the application of the following result due to [@Ricceri Theorem 6].
\[Recci\] Let $(H, \|\cdot\|)$ be a real reflexive Banach space and $\Phi, \Psi: H\rightarrow \mathbb R$ be two continuously Gateaux differentiable functionals such that $\Phi$ is sequentially weakly lower semicontinuuous and coercive. Further assume that $\Psi$ is sequentially weakly continuous. In addition, assume that, for each $\gamma>0$ the functional $\mathcal I_\gamma: H\to \mathbb R$ $$\mathcal I_\gamma(z):=\gamma \Phi(z)-\Psi(z), \,\, z\in H,$$ satisfies $(PS)_c$ condition for all $c\in \mathbb R$. Then for any $\rho>\inf_H \Psi$ and every $$\gamma>\inf_{u\in \Phi^{-1}(-\infty, \;\rho)}\frac{\sup_{v\in \Phi^{-1}(-\infty, \;\rho)}\Psi(v)-\Psi(u)}{\rho-\Phi(u)}$$ the following alternative holds: either the functional $\mathcal I_\gamma$ has a strict global minimum in $\Phi^{-1}(-\infty, \;\rho)$, or $\mathcal I_\gamma$ has at least two critical points one of which lies in $\Phi^{-1}(-\infty, \;\rho)$.
Here we consider the functional $E_{A, \lambda}: X_{0, A}\to \mathbb R$ as $$\label{Energy}
E_{A, \lambda}(u)=\frac{1}{\lambda}J(u)-K_\lambda(u),$$ where $$J(u)=\frac12\|u\|^2_{X_{0, A}}\;\;\text{and}\;\; K_\lambda(u)=\frac12 \|u\|_2^2+\frac{1}{\lambda}\int_\Omega F(|u|)\, {\mathrm{d}}x.$$ It is straightforward to see that $J$ is continuously Gateaux differentiable, sequentially weakly lower semicontinuuous and coercive.
If $f$ satisfies $(H_1)-(H_2)$ and , then $K_\lambda$ is continuously Gateaux differentiable, sequentially weakly lower semicontinuuous and coercive.
Since $f$ has subcritical growth the proof is easy and we will omit it. Moreover, $K_\lambda$ is sequentially weakly continuous.
The next result is about the Palais-Smale condition.
If $f$ satisfies $(H_1)$ and , then the functional defined in satisfies $(PS)_c$ for all $c\in \mathbb R$.
Let us consider $\{u_n\}$ be a Palais-Smale sequence for the functional $E_{\lambda, A}$, that is, $$\label{EJ23R}
\frac{1}{2\lambda}\|u_n\|^2_{X_{0,A}} - \frac{1}{2}\|u_n\|_2^2 - \frac{1}{\lambda}\int_{\Omega} F(|u_n|) {\mathrm{d}}x \to c,\,\, \mbox{as}\,\, n \to +\infty,$$ and $$\label{EJ24R}
\left|\frac{1}{\lambda}\Re\langle u_n, v \rangle_{X_{0,A}} - \Re\langle u_n , v\rangle_{L^2} - \frac{1}{\lambda}\Re\int_{\Omega} f(|u_n|) u_n \overline{v} \,{\mathrm{d}}x \right| \leq \varepsilon_n \|v\|_{X_{0,A}},\,\, \mbox{for all}\,\,
v \in X_{0,A}.$$ To prove the claim of the above proposition, we divide the proof into a few steps.\
**Step 1:** The Palais-Smale sequence is bounded.\
The proof of this step follows the same lines as in Lemma \[PS\]. Consequently, there exists $u_o\in X_{0, A}$ such that $u_n\rightharpoonup u_o$ weakly in $X_{0, A}$, $u_k\to u_o$ in $L^q(\Omega)$ for all $q\in [1, \infty)$ and $u_n(x)\to u_o(x)$ a.e. in $(\Omega)$.
**Step 2:** The following convergence holds $$\int_\Omega f(|u_n|)|u_n|^2 {\mathrm{d}}x \to \int_\Omega f(|u_o|)|u_o|^2 {\mathrm{d}}x.$$ To prove the claim of this step, we proceed as follows $$\begin{aligned}
\left|\int_\Omega f(|u_n|)|u_n|^2 {\mathrm{d}}x- \int_\Omega f(|u_o|)|u_o|^2 {\mathrm{d}}x\right|&\leq \left|\int_\Omega f(|u_n|)|u_n|(|u_n|-|u_o|)\, {\mathrm{d}}x- \int_\Omega (f(|u_n|)|u_n|- f(|u_o|)|u_o|)|u_o| \,{\mathrm{d}}x\right|.\end{aligned}$$ Let us denote $$I_1=\int_\Omega |f(|u_n|)|~|u_n|~\Big||u_n|-|u_o|\big| \,{\mathrm{d}}x\;\;\text{and}\;\; I_2=\int_\Omega \Big(f(|u_n|)|u_n|- f(|u_o|)|u_o|\Big)|u_o| \,{\mathrm{d}}x.$$ We estimate these integrals one by one as follows. We begin with $I_1$, by using the estimate , the elementary inequality $\Big||a|-|b|\Big|\leq |a-b|$ for $a, b\in \mathbb C$ and Hölder’s inequality as $$I_1\leq C\int_\Omega e^{\beta |u_n|^2}|u_n-u_o| {\mathrm{d}}x\leq C\left(\int_\Omega e^{q\beta \|u_n\|^2_{X_{0, A}}\left(\frac{|u_n|}{\|u_n\|_{X_{0, A}}}\right)^2} {\mathrm{d}}x
\right)^\frac{1}{q}\left(\int_\Omega|u_n-u_o|^p{\mathrm{d}}x\right)^\frac{1}{p}.$$ Now using the fact that $u_n\to u_o$ in $L^q(\Omega)$ for all $q\in [1, \infty$) and $q~\beta~\|u_n\|_{X_{0, A}}^2<\pi$ for suitable choosen $\beta>0$ we get that $I_1\to 0$ as $n\to \infty$. Next we show the similar convergence for $I_2$. The proof of this convergence follows from the Lemma 2.1 of [@FMR] once $\int_\Omega f(|u_n|)|u_n|^2{\mathrm{d}}x<C_1$ which follows from and boundedness of the sequence $\{u_n\}$. Hence $I_2\to 0$ as $n\to \infty$. Consequently the claim of the Step 2 is proved.\
**Step 3:** Up to a subsequence, $u_n\to u_o$ in $X_{0, A}$.\
Take $v={u}_o$ in , we get $$\label{rufl1}
\frac{1}{\lambda}\|u_0\|^2_{X_{0,A}} - \|u_o\|_2^2 - \frac{1}{\lambda}\int_\Omega f(|u_o|) |u_o|^2 {\mathrm{d}}x =0.$$ On the other hand, if we take $v={u}_n$ in and use Step 2, we get $$\label{rufl2}
\frac{1}{\lambda}\|u_n\|^2_{X_{0,A}} - \|u_o\|_2^2 - \frac{1}{\lambda}\int_\Omega f(|u_o|) |u_o|^2 {\mathrm{d}}x \to 0.$$ From and , we have $\|u_n\|^2 \to \|u_o\|^2$ in $\mathbb R$ and hence $u_n\to u_o$ in $X_{0, A}$.
Justification for the choice of $\lambda$
-----------------------------------------
In the statement of Theorem \[Recci\], it can be noticed that the result holds good for all $\rho>0$ in light of the definition of $J$. Now we define the range of $\lambda$ as follows. $$\frac{1}{\lambda}>\Theta_\lambda,
\,\,\text{where}\,\,\Theta_\lambda=\inf_{u\in J^{-1}(-\infty, \;\rho)}\frac{\sup_{v\in J^{-1}(-\infty, \;\rho)}K_\lambda(v)-K_\lambda(u)}{\rho-J(u)}.$$ Since $J(0)=0=K_\lambda(0)$, $$\Theta_\lambda\leq \frac{1}{\rho}\sup_{v\in J^{-1}(-\infty, \;\rho)}K_\lambda(v)=\frac{1}{2\rho}\sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\|v\|_2^2+\frac{1}{\rho\lambda}\sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\int_\Omega F(|v|)\, {\mathrm{d}}x.$$ On the other hand, using $$\label{gned1}
\sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\|v\|_2^2\leq 2\rho S_2^2.$$ Under the assumption $(H_2)$ and , we can get the following estimate $$\sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\int_\Omega F(|v|)\, {\mathrm{d}}x\leq \sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\int_\Omega f(|v|)|v|^2 \, {\mathrm{d}}x \leq C \sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\int_\Omega e^{\beta |v|^2}|v|\, {\mathrm{d}}x.$$ Now using Hölder’s inequality with conjugate exponents $1/p+1/q=1$ in the last term, we get $$\sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\int_\Omega e^{\beta |v|^2}|v|\, {\mathrm{d}}x\leq
\sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\left(\int_\Omega e^{q\beta |v|^2}\, {\mathrm{d}}x\right)^\frac{1}{q}
\left(\int_\Omega |v|^p {\mathrm{d}}x\right)^\frac{1}{p}.$$ Again recalling and choosing $\beta>0$ sufficiently small such that $\beta q\|v\|_{X_{0, A}}^2<\pi$, by Trudinger-Moser inequality , we get $$\label{gned2}
\sup_{\{v\in X_{0, A}\,:\, \|v\|_{X_{0,A}}\leq (2\rho)^\frac12\}}\int_\Omega F(|v|) {\mathrm{d}}x\leq C_1(\beta)\|v\|_p\leq C_\beta (2\rho)^\frac12 S_p\,.$$ By combining and , we have the following estimate for $\Theta_\lambda$ $$\Theta_\lambda\leq S_2^2+\frac{\sqrt{2}C_\beta S_p}{\lambda\sqrt{\rho}}.$$ Therefore if we choose $$0< \lambda<\Lambda(\rho):=\frac{1}{S_2^2}\left(1-\frac{\sqrt{2}C_\beta S_p}{\sqrt{\rho}}\right)$$ for any $\rho>2C_\beta^2S_p^2$, we can justify the choice of $\lambda$ as in Theorem \[thm4\]. Now only thing remain to show that the possibility of global minima for the functional $E_{\lambda,\;A}$ will not occur. From assumption $(H_1)$, there exists $\mu>2$ and $C_1, C_2>0$ such that $$\label{implicationHF}
F(|u|)\geq C_1|u|^\mu-C_2 .$$ Using , we estimate for arbitrary but fixed $u\in X_{0, A}$ $$E_{A, \lambda}(t u)\leq \frac{t^2}{2\lambda}\|u\|_{X_{0, A}}^2-\frac{t^2}{2}\|u\|_2^2-\frac{C_1 t^\mu}{\lambda}\|u\|_\mu^\mu+\frac{2C_2}{\lambda}$$ which implies that $E_{A, \lambda}(tu)\to -\infty$ as $t\to \infty$. Hence $E_{A, \lambda}$ cannot have a strict global minimum in $J^{-1}(-\infty, \;\rho)$.
Acknowledgement {#acknowledgement .unnumbered}
===============
Research supported in part by INCTmat/MCT/Brazil, CNPq and CAPES/Brazil.
References {#references .unnumbered}
==========
[99]{}
W. Abdelhedi and H. Chtioui, *On a Nirenberg-type problem involving the square root of the Laplacian*, Journal of Functional Analysis, 265 (2013), 2937–2955.
Adimurthi, *Existence of positive solutions of the semilinear Dirichlet problem with critical growth for the n-Laplacian*, Ann. Sc. Norm. Super. Pisa Cl. Sci. 17 (1990), 393–413.
A. Ambrosetti and P. H. Rabinowitz, *Dual variational methods in critical point theory and applications*, J. Funct. Anal. 14 (1973), 349–381.
V. Ambrosio, G. M. Bisci, and D. Repovš, *Nonlinear equations involving the square root of the Laplacian*, Discrete & Continuous Dynamical Systems - S, 12 (2019), 151-170.
V. Ambrosio and P. d’Avenia, *Nonlinear fractional magnetic Schrödinger equation: existence and multiplicity*, J. Differential Equations 264 (2018), 3336–3368.
V. Ambrosio, *On a fractional magnetic Schrödinger equation in $\mathbb R$ withexponential critical growth*, Nonlinear Analysis 183 (2019), 117-148.
D. Applebaum, *Lévy processes-from probability to finance and quantum groups*, Notices Am. Math. Soc. 51 (2004), 1336–1347.
P. Bartolo and V. Benci, *Abstract critical point theorems and applications to some nonlinear problems with strong resonance at infinity*, Nonlinear analysis: Theory, methods & applications 7 (1983), 981-1012.
X. Cabré and J. Tan, *Positive solutions of nonlinear problems involving the square root of the Laplacian*, Advances in Mathematics 224 (2010), 2052–2093. D. M. Cao, *nontrivial solution of semilinear elliptic equation with critical exponent in $\mathbb{R}^2$*, Comm. Partial Diff. Eq. 17 (1992), 407-435.
G. Cerami, D. Fortuno and M. Struwe, *Bifurcation and multiplicity results for nonlinear elliptic problems involving critical Sobolev exponents*, Ann. Inst. Henri poincare 1 (1984), 341-350.
W. Chen and S. Holm, *Fractional Laplacian time-space models for linear and nonlinear lossy media exhibiting arbitrary frequency power-law dependency*, The Journal of the Acoustical Society of America 115 (2004), 1424-–1430. R. Cont and E. Voltchkova, *A finite difference scheme for option pricing in jump diffusion and exponential lévy models*, SIAM Journal on Numerical Analysis 43 (2005), 1596–-1626.
P. d’Avenia and M. Squassina, *Ground states for fractional magnetic operators*, ESAIM Control Optim. Calc. Var. 24 (2018), 1–24.
D. G. de Figueiredo, O. H. Miyagaki and B. Ruf, *Elliptic equations in $\mathbb{R}^2$ with nonlinearities in the critical growth range*, Calc. Var. Partial Differential Equations 3 (1995), 139–153.
J. M. do Ó, O. H. Miyagaki, and M. Squassina. *Nonautonomous fractional problems with exponential growth*, NoDEA Nonlinear Differential Equations Appl. 22 (2015), 1395–1410.
B. P. Epps and B. Cushman-Roisin, *Turbulence modeling via the fractional Laplacian*, arXiv preprint arXiv:1803.05286, 2018.
A. Fiscella , G. M. Bisci and R. Servadei, *Bifurcation and multiplicity results for critical nonlocal fractional Laplacian problems*, Bulletin des Sciences Mathématiques 140 (2016), 14-35.
A. Fiscella and E. Vecchi, *Bifurcation and multiplicity results for critical magnetic fractional problems*, Electronic Journal of Differential Equations 153 (2018), 1-18.
A. Fiscella, A. Pinamonti and E. Vecchi, [*Multiplicity results for magnetic fractional problems*]{}, J. Differential Equations [**263**]{} (2017), 4617–4633.
J. Giacomoni, P. K. Mishra, and K. Sreenadh, *Fractional elliptic equations with critical exponential nonlinearity*, Adv. Nonlinear Anal. 5 (2016), 57-74.
A. Iannizzotto and M. Squassina, *1/2-Laplacian problems with exponential nonlinearity*, J. Math. Anal. Appl. 414 (2014), 372-385. T. Ichinose, *Essential selfadjointness of the Weyl quantized relativistic Hamiltonian*, Ann. Inst. H. Poincaré Phys. Théor., 51 (1989), 265-–297. T. Ichinose and H. Tamura, *Imaginary-time path integral for a relativistic spinless particle in an electromagnetic field*, Commun. Math. Phys., 105 (1986), 239-–257.
S. Iula, A. Maalaoui and L. Martinazzi, *A fractional Moser-Trudinger type inequality in one dimension and its critical points*, Differential Integral Equations 29 (2016), 455-492.
H. Kozono, T. Sato and H. Wadade, *Upper bound of the best constant of a Trudinger-Moser inequality and its application to a Gagliardo-Nirenberg inequality*. Indiana Univ. Math. J. 55 (2006), 1951-1974.
L. Martinazzi, *Fractional Adams-Moser-Trudinger type inequalities*, Nonlinear Anal. 127 (2015), 263–278.
P. K. Mishra and K. Sreenadh, *Bifurcation and multiplicity of solutions for the fractional Laplacian with critical exponential nonlinearity*, Electronic Journal of Differential Equations 203 (2016), 1-9.
J. Moser, *A sharp form of an inequality by N. Trudinger*, Indiana Univ. Math. J. 20 (1970/71), 1077-1092.
T. Ozawa, *On critical cases of Sobolev’s inequalities*, J. Funct. Anal. 127 (1995), 259-269.
A. Pinamonti, M. Squassina and E. Vecchi, [*Magnetic BV functions and the Bourgain-Brezis-Mironescu formula*]{}, to appear on Advances in Calculus of Variations, Preprint. arXiv:1609.09714.
A. Pinamonti, M. Squassina and E. Vecchi, [*The Maz’ya-Shaposhnikova limit in the magnetic setting*]{}, J. Math. Anal. Appl. [**449**]{} (2017), 1152–1159.
B. Ricceri, *On a classical existence theorem for nonlinear elliptic equations, in: M. Thera (Ed.)*, Experimental, Constructive and Nonlinear Analysis, in: Conf. Proc., Can. Math. Soc., 27 (2000), 275-278.
F. Takahashi, *Critical and subcritical fractional Trudinger-Moser-type inequalities on $\mathbb{R}$*, Adv. Nonlinear Anal. 8 (2019), 868-884.
J. Tan, *The Brezis-Nirenberg type problem involving the square root of the Laplacian*, Calc. Var. Partial Differential Equations, 36 (2011), 21-41.
B. Zhang, M. Squassina and X. Zhang, [*Fractional NLS equations with magnetic field, critical frequency and critical growth*]{}, Manuscripta Math. [**155**]{} (2018), no. 1-2, 115–140.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
According to the Langlands functoriality conjecture, broadened to the setting of spherical varieties (of which reductive groups are special cases), a map between $L$-groups of spherical varieties should give rise to a functorial transfer of their local and automorphic spectra. The “Beyond Endoscopy” proposal predicts that this transfer will be realized as a comparison between the (relative) trace formulas of these spaces.
In this paper we establish the local transfer for the identity map between $L$-groups, for spherical affine homogeneous spaces $X=H\backslash G$ whose dual group is ${\operatorname{SL}}_2$ or ${\operatorname{PGL}}_2$ (with $G$ and $H$ split). More precisely, we construct a transfer operator between orbital integrals for the $X\times X/G$-relative trace formula, and orbital integrals for the Kuznetsov formula of ${\operatorname{PGL}}_2$ or ${\operatorname{SL}}_2$. Besides the $L$-group, another invariant attached to $X$ is a certain $L$-value, and the space of test measures for the Kuznetsov formula is enlarged, to accommodate the given $L$-value.
The fundamental lemma for this transfer operator is proven in a forthcoming paper of Johnstone and Krishna. The transfer operator is given explicitly in terms of Fourier convolutions, making it suitable for a global comparison of trace formulas by the Poisson summation formula, hence for a uniform proof, in rank one, of the relations between periods of automorphic forms and special values of $L$-functions.
address: 'Department of Mathematics and Computer Science, Rutgers University at Newark, 101 Warren Street, Smith Hall 216, Newark, NJ 07102, USA.'
author:
- Yiannis Sakellaridis
bibliography:
- 'biblio.bib'
title: 'Functorial transfer between relative trace formulas in rank one.'
---
Introduction
============
Relative functoriality {#ssrelfunctoriality}
----------------------
According to the Relative Langlands Program, the local and automorphic spectra of a spherical $G$-variety $X$ should be determined by its $L$-group ${{^LG}}_X$, which comes equipped with a distinguished morphism $$\label{Lgroups}
{{^LG}}_X\times{\operatorname{SL}}_2\to {{^LG}},$$ cf. [@GN; @SV; @KnSch].
Roughly speaking, this means, locally, that the Plancherel formula for $L^2(X(F))$ (where $F$ is a local field) should read: $$\left<\Phi_1,\overline{\Phi_2}\right>_{L^2(X)} = \int_{\varphi} J_\varphi^{{\operatorname{Planch}}}(\Phi_1\otimes\Phi_2) \nu_X(\varphi),$$ where $\nu_X$ is the standard measure [@SV §17.3] on the space of Langlands parameters into ${{^LG}}_X$, and $J_\varphi^{{\operatorname{Planch}}}$ is a *relative character* $$J_{\varphi} :\mathcal S(X\times X)\to \Pi_\varphi \otimes \widetilde{\Pi_{\varphi}} \to {\mathbb{C}},$$ where $\Pi_{\varphi}$ is the sum of irreducible representations in the Arthur packet associated to the composition of $\varphi$ with the canonical map .
Globally, an analogous decomposition (in terms of “global Langlands parameters”) should hold for the spectral side of the *relative trace formula* of $X$ — more precisely, for its stable part —, a distribution on the adelic points of the quotient $X\times X/G$ (with $G$ acting diagonally), whose spectral decomposition should read, roughly: $${{\operatorname{RTF}}}_X(\Phi_1\otimes \Phi_2) = \int_{\varphi} J_\varphi^{{\operatorname{gl}}}(\Phi_1\otimes\Phi_2) \mu_X(\varphi).$$ Moreover, the global distributions $J_\varphi^{{\operatorname{gl}}}$, which can be expressed in terms of squares of periods of automorphic forms, should (under some assumptions on $X$) be equal to Euler products of the local distributions $J_{\varphi_v}^{{\operatorname{Planch}}}$, establishing a link between periods of automorphic forms and special values of $L$-functions; this is the generalized Ichino–Ikeda conjecture proposed in [@SV §17.4].
We currently have no general tools to address these very general, and uniform, conjectures. In this paper, I will propose a uniform approach which works in the case when ${{^LG}}_X={\operatorname{SL}}_2$ or ${\operatorname{PGL}}_2$ and, hopefully, generalizes to higher rank (although I cannot yet propose such a generalization). The idea is to find a way to compare the relative trace formula for *any* such variety, with the corresponding Kuznetsov formula, i.e., the relative trace formula for the Whittaker model $X=(N,\psi)\backslash G^*$ of the group $G^*={\operatorname{PGL}}_2$ or ${\operatorname{SL}}_2$ (respectively). The Kuznetsov formula, not the Arthur–Selberg trace formula, seems to be the appropriate base case for such a type of functoriality, but it requires some modification, because it does not produce, on the spectral side, the same $L$-functions as the relative trace formula for $X$. Roughly speaking, the spectral side of the Kuznetsov formula is weighted by the factors $$\frac{1}{L(\varphi, {\mathrm{Ad}}, 1)},$$ (where $\varphi$ denotes a global Langlands parameter into ${{^LG}}_X = {{^LG}}^*$), while the relative trace formula for $X$ will have an extra $L$-factor, depending on $X$, in the numerator: $$\frac{L_X(\varphi)}{L(\varphi, {\mathrm{Ad}}, 1)}.$$
For example, in the case of the Arthur–Selberg trace formula (when $X=H$, a reductive group), we have $L_X(\varphi) = L(\varphi, {\mathrm{Ad}}, 1)$, which is why no $L$-values appear in the end, while for $X={\mathbbm{G}_m}\backslash {\operatorname{PGL}}_2$ we have $L_X(\varphi)= L(\varphi,{{\operatorname{Std}}}, \frac{1}{2})^2$, corresponding to the square of the Hecke period. These $L$-functions are obtained by *enlarging* the space of test measures for the Kuznetsov formula. Thus, our comparison is achieved via a *transfer operator*, which is a linear isomorphism $$\mathcal T: \mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi) \xrightarrow\sim \mathcal S(X\times X/G),$$ between the appropriately enlarged space of test measures for the Kuznetsov formula of $G^*$, and the standard space of test measures for the relative trace formula of $X$.
Rank-one spherical varieties
----------------------------
These ideas were explored, in the special cases $X = T\backslash {\operatorname{PGL}}_2$ (where $T$ is a torus) and $X={\operatorname{SL}}_2 = {{\operatorname{SO}}}_3\backslash {{\operatorname{SO}}}_4$, in the papers [@SaBE1; @SaBE2; @SaHanoi; @SaTransfer1; @SaTransfer2]. However, it was not clear at that point if those cases were part of a general pattern, or just reflections of methods already known. In this paper, I demonstrate for the first time that there is a general “operator of functoriality” in rank one, as general and uniform as the aforementioned conjectures.
Spherical varieties of rank one are, in some sense, the building blocks of all spherical varieties, in the same sense as the group ${\operatorname{SL}}_2$ is the building block of all reductive groups: to a general spherical variety $X$, and each simple coroot $\gamma$ of its dual group (better known as the “spherical roots” of $X$), there is an associated rank-one (up to center) spherical variety $X_\gamma$ which is a degeneration of $X$. Thus, the comparisons studied here should be essential in understanding cases of higher rank.
The list of spherical varieties of rank one consists of a finite number of families, classified by Akhiezer in [@Akhiezer] — see also the tables of [@Wasserman; @KnVS]. Up to the action of the “center” $\mathcal Z(X):={{\operatorname{Aut}}}^G(X)$, the *affine homogeneous spherical varieties* $X=H\backslash G$ over an algebraically closed field in characteristic zero whose dual group $\check G_X$ is either ${\operatorname{SL}}_2$ or ${\operatorname{PGL}}_2$ are listed in the following table:
$ \begin{array}{|c|c|c|c|c|c|}
\hline
&X & P(X) & \check G_X & \gamma & L_X \\
& & & & &\\
\hline
\hline
A_1& {\mathbbm{G}_m}\backslash{\operatorname{PGL}}_{2} &B & {\operatorname{SL}}_{2} & \alpha & L({{\operatorname{Std}}},\frac{1}{2})^2 \\
\hline
A_n& {\operatorname{GL}}_n \backslash{\operatorname{PGL}}_{n+1} &P_{1,n-1,1} & {\operatorname{SL}}_{2} & \alpha_1+\dots + \alpha_n& L({{\operatorname{Std}}},\frac{n}{2})^2 \\
\hline
B_n& {{\operatorname{SO}}}_{2n}\backslash{{\operatorname{SO}}}_{2n+1} & P_{{{\operatorname{SO}}}_{2n-1}} & {\operatorname{SL}}_2 & \alpha_1 + \dots + \alpha_n& L({{\operatorname{Std}}}, n-\frac{1}{2}) L({{\operatorname{Std}}},\frac{1}{2})\\
\hline
C_n& {\operatorname{Sp}}_{2n-2}\times {\operatorname{Sp}}_{2}\backslash {\operatorname{Sp}}_{2n} & P_{{\operatorname{SL}}_2 \times {\operatorname{Sp}}_{2(n-2)}} & {\operatorname{SL}}_{2} & \alpha_1+2\alpha_2+\dots + 2\alpha_{n-1}+ \alpha_n & L({{\operatorname{Std}}},n-\frac{1}{2}) L({{\operatorname{Std}}},n-\frac{3}{2}) \\
\hline
F_4& {\operatorname{Spin}}_9\backslash F_4 & P_{{\operatorname{Spin}}_7} & {\operatorname{SL}}_2 & \alpha_1+2\alpha_2+3\alpha_3+2\alpha_4 & L({{\operatorname{Std}}}, \frac{11}{2}) L({{\operatorname{Std}}}, \frac{5}{2}) \\
\hline
G_2& {\operatorname{SL}}_3\backslash G_2& P_{{\operatorname{SL}}_2} & {\operatorname{SL}}_2 & 2\alpha_1+ \alpha_2 & L({{\operatorname{Std}}}, \frac{5}{2}) L({{\operatorname{Std}}}, \frac{1}{2}) \\
\hline
\hline
D_2& {\operatorname{SL}}_2={{\operatorname{SO}}}_{3}\backslash{{\operatorname{SO}}}_{4} & B & {\operatorname{PGL}}_2 & \alpha_1 + \alpha_2 & L({\mathrm{Ad}},1) \\
\hline
D_n& {{\operatorname{SO}}}_{2n-1}\backslash{{\operatorname{SO}}}_{2n} & P_{{{\operatorname{SO}}}_{2n-2}} & {\operatorname{PGL}}_2 & 2\alpha_1+ \dots + 2\alpha_{n-2} + \alpha_{n-1}+\alpha_n & L({\mathrm{Ad}},n-1) \\
\hline
D_4'' & {\operatorname{Spin}}_7\backslash{\operatorname{Spin}}_8 & P_{{\operatorname{Spin}}_6} & {\operatorname{PGL}}_2 & 2\alpha_1+ 2\alpha_{2} + \alpha_{3}+\alpha_4 & L({\mathrm{Ad}},3)\\
\hline
B_3''& G_2\backslash{\operatorname{Spin}}_7 & P_{{\operatorname{SL}}_3} & {\operatorname{PGL}}_2 & \alpha_1 + 2\alpha_2 + 3\alpha_3 & L({\mathrm{Ad}},3)\\
\hline
\end{array}
$ $$\label{thetable}$$
The various columns of this table will be explained below. In this paper, I work over a local field $F$ in characteristic zero, and will only consider the case when both $G$ and $H$ are split over $F$. Under these restrictions, as we will see (Proposition \[uniqueclass?\]), each line in the first group of the table above (from $A_1$ to $G_2$) corresponds to a unique isomorphism class of $G$-varieties, while each line in the second group (from $D_2$ to $B_3''$) corresponds to a set of isomorphism classes parametrized by square classes in $F^\times$.
For almost all of the varieties in the table above, a version of the local relative Langlands conjecture for $L^2(X)$ was established by Wee Teck Gan and Raúl Gomez [@GG], on a case-by-case basis using the usual and exceptional theta correspondences. Similar, and other, methods can be used to study global periods; examples in the literature include [@RS; @GanGur; @Flicker].
In any case, the local and global conjectures should be seen as corollaries of a deeper fact, which is encoded in the comparison of trace formulas that I establish in this paper. Moreover, the approach of the present paper is *classification-free* (except for a minor result in Lemma \[lemmaclosedorbits\]), and relies on a sophisticated theory developed by Friedrich Knop, on the geometry of the moment map $$T^*X \to {\mathfrak{g}}^*.$$
I now explain the various entries in the table: The column $\gamma$ denotes the *normalized spherical root* of the spherical variety, in the language of [@SV], described in the basis of simple roots labelled as in Bourbaki. This is the positive coroot for the canonical embedding . (The $L$-groups can be replaced by their connected components $\check G_X$, $\check G$, here, since we take $G$ to be split.) This spherical root is either a root of $G$ or the sum of two strongly orthogonal roots; I have chosen the representative of the equivalence class up to center to be such that the dual group is ${\operatorname{SL}}_2$ in the former case, and ${\operatorname{PGL}}_2$ in the latter. Thus, we obtain two families of spherical varieties of rank one, whose prototypes are, respectively, the examples labelled $A_1$ and $D_2$ above (which, of course, are special cases of $A_n$ and $D_n$). The case $D_4''$ is obtained from $D_4$ by application of the triality automorphism of ${\operatorname{Spin}}_8$ (which does not descend to ${{\operatorname{SO}}}_8$). Because of the two prototypes, we say (following [@SV]) that the spherical root is “of type $T$” (for “torus”) in the first family and “of type $G$” (for “group”) in the second.
By $P(X)$ we denote the conjugacy class of parabolics stabilizing the open Borel orbit. In the table, I describe the parabolics in a way that should be self-explanatory, by indicating the semisimple part of their Levi quotient $L(X)$ or, in the case of ${\operatorname{GL}}_n$, an ordered partition of $n$. In the case of $G=G_2$, $P_{{\operatorname{SL}}_2}$ is such that its Levi contains a long root. Notice that the roots of the Levi $L(X)$ are always orthogonal to the spherical root $\gamma$. The parabolic $P(X)$ determines the restriction of the map to the “Arthur-${\operatorname{SL}}_2$” factor, which has to map to a principal ${\operatorname{SL}}_2$ in the Levi subgroup of $\check G$ dual to $P(X)$.
Finally, $L_X$ stands for a $\frac{1}{2}{\mathbb{Z}}$-graded representation $r=\bigoplus_{n \in \frac{1}{2}{\mathbb{Z}}} r_n$ of $\check G_X$, which I call the “$L$-value associated to $X$”, thinking of the $L$-value $$\prod_n L(r_n\circ \varphi, n)$$ attached to any Langlands parameter $\varphi$ into $\check G_X$. For this reason, I denote this graded representation by $\prod_n L(r_n, n)$. This is the $L$-value attached to the square of the global $H$-period, according to the generalization of the Ichino–Ikeda conjecture [@II] proposed in [@SV §17.4] and the local unramified calculation, performed for classical groups only, of [@SaSph]. The same calculation can be generalized to all cases, including non-classical groups, but, in any case, the $L$-value here will be determined directly in terms of the geometry of the $G$-space $X\times X$, as follows:
- When the spherical root is of type $T$, the associated $L$-value is always of the form $L({{\operatorname{Std}}}, s_1)L({{\operatorname{Std}}}, s_2)$, for some positive half-integers $s_1\ge s_2$. These are determined by the relations $s_1+s_2=\frac{\dim X}{2}$, and $ s_1 = \frac{\left<\check\gamma, \rho_{P(X)}\right>}{2}$.
- When the spherical root is of type $G$, we have $L_X = L({\mathrm{Ad}}, s_0)$, with $s_0= \left<\check\gamma, \rho_{P(X)}\right> = \frac{\dim X-1}{2}$, always an integer.
Here, $2\rho_{P(X)}$ is the sum of positive roots in the nilpotent radical of the Lie algebra of $P(X)$; it can be considered as a cocharacter into the canonical maximal torus of $\check G_X$, hence the value of the spherical coroot $\check \gamma$ (the positive root of $\check G_X$) makes sense on it.
Notation and the main result {#ssnotation}
----------------------------
All varieties will be defined over a local, locally compact field $F$ in characteristic zero, and we write $X=X(F)$, etc, when no confusion arises. In particular, all measures or functions will be on the $F$-points of the varieties under consideration.
We denote by $\mathcal S(X)$ the space of (${\mathbb{C}}$-valued) Schwartz measures on the $F$-points of a smooth variety $X$; these are smooth measures which, in the non-Archimedean case, are of compact support, and in the Archimedean case are of rapid decay, together with their polynomial derivatives. For uniformity of language, I will often write “smooth of rapid decay” to describe this behavior (of the measure and its derivatives), with the understanding that this means compact support in the non-Archimedean case. (Whenever needed, the space of Schwartz *functions* will be denoted by $\mathcal F(X)$.)
The notation $X\sslash G$ will stand for the affine, invariant-theoretic quotient ${\operatorname{Spec}}F[X]^G$ of a $G$-variety $X$, and if $\pi: X\to X\sslash G$ denotes the canonical quotient map, the image $\pi_! \mathcal S(X)\subset {{\operatorname{Meas}}}(X\sslash G)$ of the pushforward map of measures will be denoted by $\mathcal S(X/G)$. In the Archimedean case, where $\mathcal S(X)$ is a nuclear Fréchet space, the space $\mathcal S(X/G)$ inherits a quotient Fréchet topology; in the non-Archimedean case, any reference to topology should be ignored.
Let $X$ be one of the spaces in Table , with a reductive group $G$ acting on it, and let $G^*$ denote the group ${\operatorname{PGL}}_2$, if $\check G_X={\operatorname{SL}}_2$, or ${\operatorname{SL}}_2$, if $\check G_X={\operatorname{PGL}}_2$. Let $N\subset G^*$ be the upper triangular unipotent subgroup, identified with the additive group ${\mathbbm{G}_a}$ in the obvious way, and let $\psi: F\to {\mathbb{C}}^\times$ be a nontrivial character, considered also as a character of $N$. We fix throughout an additive Haar measure on $F$, which is self-dual with respect to $\psi$. We extend the notation of Schwartz spaces to the quotient that we will denote by $(N,\psi)\backslash G^*/(N,\psi)$: If $A^*\subset G^*$ is the torus of diagonal elements, and $w = \begin{pmatrix} & -1 \\ 1\end{pmatrix}$, we embed $A^*$ in the affine quotient $N\backslash G^*\sslash N$ by $A^*\ni a \mapsto [wa]$, the class of the element $wa$, and let $\mathcal S(N,\psi\backslash G^*/N,\psi)$ denote the space of measures on $A^*$ of the form $$f(a) = \pi^\psi_!(\Phi dg):= \left(\int_{N\times N} \Phi(n_1 wa n_2) \psi^{-1}(n_1n_2) d(n_1,n_2)\right)\cdot \delta(a) da,$$ where $\Phi$ is a Schwartz *function* on $G^*$, $\delta$ is the modular character of the upper triangular Borel subgroup, and $da$ is a Haar measure on $A^*$. This is a twisted version of the pushforward of the measure $\Phi dg$ to $N\backslash G^*\sslash N$, which for suitable Haar measures reads: $$\pi_! (\Phi dg)(a) = \left(\int_{N\times N} \Phi(n_1 wa n_2) d(n_1,n_2)\right)\cdot \delta(a) da.$$
We also fix a coordinate on $A^*$ which we denote by $$\xi(a) = e^\alpha(a), \mbox{ when } G^*={\operatorname{PGL}}_2;$$ $$\zeta (a) = e^\frac{\alpha}{2}(a), \mbox{ when } G^*={\operatorname{SL}}_2,$$ where $\alpha$ is the positive (upper triangular) root of $G^*$, and we use exponential notation to denote the corresponding character, since weights are written additively. Then, $N\backslash G^*\sslash N$ is identified with $\mathbbm A^1$, with coordinate $\xi$, resp. $\zeta$. The elements of $\mathcal S(N,\psi\backslash G^*/N,\psi)$, viewed as measures on $\mathbbm A^1=N\backslash G^*\sslash N$, are smooth of rapid decay away from zero, while in a neighborhood of $0\in \mathbbm A^1$ they have a singularity which is called the “Kloosterman germ”, because in the non-Archimedean case they are smooth multiples of the measures $$\xi\mapsto \left(\int_{|u|^2=|\xi|} \psi^{-1} \left(\frac{u}{\xi} + u^{-1}\right) du \right) d^\times \xi,$$ resp.$$\zeta\mapsto \left(\int_{u\in \pm 1 + \mathfrak p} \psi^{-1} \left(\frac{u+u^{-1}}{\zeta}\right) du \right) d\zeta.$$ (Notice that there are two separate germs in the case of ${\operatorname{SL}}_2$, corresponding to the choice of $\pm 1$.)
We define enlarged spaces of test measures $$\mathcal S^-_{L_X} (N,\psi\backslash G^*/N,\psi) \supset \mathcal S (N,\psi\backslash G^*/N,\psi)$$ for the Kuznetsov formula, associated to the $L$-values $L_X$ that appear in Table , as follows (see also [@SaTransfer1 §2.2]): their elements coincide with elements of $\mathcal S (N,\psi\backslash G^*/N,\psi)$ away from infinity, but in a neighborhood of infinity, instead of being of rapid decay, they are allowed to be of the following form:
- When $G^*={\operatorname{PGL}}_2$ and $L_X=L({{\operatorname{Std}}}, s_1)L({{\operatorname{Std}}},s_2)$ with $s_1\ge s_2$, $$\label{expansionPGL2}
(C_1(\xi^{-1}) |\xi|^{\frac{1}{2}-s_1}+C_2(\xi^{-1}) |\xi|^{\frac{1}{2}-s_2}) d^\times \xi,$$ where $C_1$ and $C_2$ are smooth functions; this should be replaced by $$|\xi|^{\frac{1}{2}-s_1}(C_1(\xi^{-1}) +C_2(\xi^{-1}) |\xi|^{s_1-s_2} \log|\xi|) d^\times \xi$$ when $|\xi|^{s_1-s_2}$ is a smooth function — that is, in the non-Archimedean case, when $s_1=s_2$, in the real case when $s_1-s_2\in 2\mathbb N$, and in the complex case when $s_1-s_2\in \mathbb N$. (We use the arithmetic normalization of absolute values, which is compatible with norms to the base field; this is the square of the usual absolute value in the complex case.)
- When $G^*={\operatorname{SL}}_2$ and $L_X=L({\mathrm{Ad}},s_0)$, $$\label{expansionSL2}
C(\zeta^{-1}) |\zeta|^{1-s_0} d^\times \zeta,$$ where $C$ is a smooth function.
In the Archimedean case, all of these spaces have an obvious Fréchet topology. The basic theorem proven in this paper is the following:
\[maintheorem\] Let ${\mathfrak{C}}_X=(X\times X)\sslash G$. There is an isomorphism ${\mathfrak{C}}_X \simeq \mathbbm A^1$, and the map $X\times X\to \mathbbm A^1$ is smooth away from the preimage of two points of $\mathbbm A^1$, that we will call singular. We fix the isomorphisms as follows:
- When $\check G_X={\operatorname{SL}}_2$, we take the set of singular points to be $\{0, 1\}$, with $X^{{\operatorname{diag}}}\subset X\times X$ mapping to $1 \in {\mathfrak{C}}_X\simeq \mathbbm A^1$.
- When $\check G_X = {\operatorname{PGL}}_2$, we take the set of singular points to be $\{-2,2\}$, with $X^{{\operatorname{diag}}}\subset X\times X$ mapping to $2 \in {\mathfrak{C}}_X\simeq \mathbbm A^1$.
Then, there is a continuous linear isomorphism: $$\mathcal T: \mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi) \xrightarrow\sim \mathcal S(X\times X/G),$$ given by the following formula:
- When $\check G_X={\operatorname{SL}}_2$ with $L_X=L({{\operatorname{Std}}}, s_1) L({{\operatorname{Std}}},s_2)$, $s_1\ge s_2$, $$\mathcal Tf(\xi) = |\xi|^{s_1-\frac{1}{2}} \left( |\bullet|^{\frac{1}{2}-s_1} \psi(\bullet) d\bullet\right) \star \left( |\bullet|^{\frac{1}{2}-s_2} \psi(\bullet) d\bullet\right) \star f(\xi).$$
- When $\check G_X={\operatorname{PGL}}_2$ with $L_X=L({\mathrm{Ad}}, s_0)$, $$\mathcal Tf(\zeta) = |\zeta|^{s_0-1} \left( |\bullet|^{1-s_0} \psi(\bullet) d\bullet\right) \star f(\zeta).$$
By $\left( |\bullet|^s \psi(\bullet) d\bullet\right) \star $ we denote the operator of multiplicative convolution by the measure $\left( |x|^s \psi(x) dx\right)$ (in the variable $y=\xi$ or $\zeta$, respectively): $$\left( |\bullet|^s \psi(\bullet) d\bullet\right) \star f (y) = \int_{F^\times} |x|^s \psi(x) f(x^{-1}y) dx = |y|^{s+1} \int f(u^{-1}) |u|^s \psi(uy) du.$$ The measure $dx$ is the *additive* Haar measure on $F$ that we have fixed. Convolution should be understood as the Fourier transform of the distribution $u\mapsto f(u^{-1}) |u|^s$, followed by multiplication by $|y|^{s+1}$.
The operator $\mathcal T$ is clearly the correct operator of functoriality between the relative trace formula for $X$ and the Kuznetsov formula. Indeed, it was shown in [@SaBE1; @SaBE2; @SaTransfer1] that in the basic cases $A_1, D_2$ it satisfies the appropriate fundamental lemma for the Hecke algebra, and that it pulls back relative characters to relative characters (see [@SaHanoi §6–7] for precise references); these statements can also be confirmed in the general $A_n$-case by “unfolding”. In an upcoming paper [@JK], Daniel Johnstone and Rahul Krishna prove the appropriate fundamental lemma for the transfer operator in all cases. There remains to prove the fundamental lemma for the full Hecke algebra, in order to be able to use this operator globally (together with the “Hankel transforms” for the functional equations of the standard and adjoint $L$-functions, discussed in [@SaHanoi §8] and [@SaTransfer2 §8]), and obtain a uniform proof of functoriality and the relation between $X$-periods of automorphic forms and the $L$-value $L_X$.
It also happens in rank one, as was already observed in [@SaTransfer1] for the two basic cases, that the operator $\mathcal T$ has precisely the same form (in suitable coordinates) as the analogous operator $\mathcal T_\emptyset$ for the boundary degeneration $X_\emptyset$ of $X$ (a horospherical variety). This is explained in Section \[sec:boundary\]. This provides a conceptual reason for the formula of the transfer operator, but I do not know why it is equal to $\mathcal T_\emptyset$ “on the nose”, instead of being just a deformation of it.
Outline of the proof
--------------------
As mentioned, the proof of Theorem \[maintheorem\] is classification-free, and relies on Friedrich Knop’s theory of the moment map. The main issue is to analyze the quotient $X\times X/G$, and to describe, more or less explicitly, the germs of pushforward Schwartz measures for the map $X\times X \to X\times X\sslash G$.
To every spherical variety $X$ one can attach a canonical “universal Cartan”, that is, a torus $A_X$, and a “little Weyl group” $W_X$ acting on it. The dual torus to $A_X$ is the canonical maximal torus of the dual group $\check G_X$, and $W_X$ is its Weyl group. Hence, in the rank-one case that we are considering, $A_X\simeq {\mathbbm{G}_m}$ and $W_X = {\mathbb{Z}}/2$, acting on ${\mathbbm{G}_m}$ by inversion.
The main result about the space $\mathcal S(X\times X/G)$ of pushforward measures, for $X$ as in Table , is the following:
\[Xtheorem\] There is a canonical isomorphism ${\mathfrak{C}}_X:= X\times X\sslash G\simeq A_X\sslash W_X$, and the map $X\times X\to {\mathfrak{C}}_X$ is smooth away from the preimages of $[\pm 1]$, where $[\pm 1]$ denote the images of $\pm 1\in A_X$ in $A_X\sslash W_X$.
In particular, there are two distinguished closed $G$-orbits $X_1 = X^{{\operatorname{diag}}}$ and $X_{-1}$ (over $[\pm 1]$, respectively); if $d_{\pm 1}$ denote their codimensions, then $d_1=\dim X$ and $$d_{-1} = \epsilon \left<2\rho_{P(X)},\check\gamma\right>-d_1+2,$$ where $\check\gamma$ is the spherical coroot, $2\rho_{P(X)}$ is the sum of roots in the unipotent radical of $P(X)$, and $$\epsilon = \begin{cases} 1, \mbox{ when the spherical root is of type $T$ (dual group ${\operatorname{SL}}_2$)};\\
2, \mbox{ when the spherical root is of type $G$ (dual group ${\operatorname{PGL}}_2$).}
\end{cases}$$ In the case of root of type $G$, $d_1=d_{-1}$.
The space $\mathcal S(X\times X/G)$ consists of those measures on ${\mathfrak{C}}_X(\simeq \mathbbm A^1)$ which are smooth and of rapid decay, together with their polynomial derivatives, away from neighborhoods of $[\pm 1]$ (compactly supported in the non-Archimedean case), while in the neighborhood of $[\pm 1]$ their germs coincide with germs for the twisted pushforward maps: $$\mathbbm A^2/({\mathbbm{G}_m}, |\bullet|^\frac{2-d_{\pm 1}}{2}),$$ for spherical root of type $T$, and $$\mathfrak{sl}_2/(B_{{\operatorname{ad}}}, \delta_2^\frac{3-d_{\pm 1}}{2}),$$ for spherical root of type $G$, where $B_{{\operatorname{ad}}}$ denotes the Borel subgroup of ${\operatorname{PGL}}_2$, and $\delta_2$ is its modular character.
For the precise meaning of these “twisted pushforwards”, I point the reader to the precise formulation of Theorem \[Xtheorem2\]. In other words, the germs for the general case are twisted versions of the germs for the “basic cases” $A_1$ and $D_2$. This indirect description of the germs allows us to relate these germs of pushforward measures for $X\times X/G$ with the Kloosterman germs for the Kunzetsov formula of $G^*$, based on results of [@SaTransfer1].
Since $X=H\backslash G$ is homogeneous, we can also write $X\times X\sslash G = H\backslash G\sslash H$; when $X$ is symmetric (as is the case for most of the cases in Table , except for those denoted by $G_2$ and $B_3''$), the identification of this with $A_X\sslash W_X$ is due to Richardson [@Richardson].
In any case, to obtain this and Theorem \[Xtheorem\] in general, I use Knop’s theory of the moment map in a somewhat paradoxical way: While the cotangent bundle together with its moment map $T^*X\to g^*$ is classically used for microlocal analysis on $X$, here I use it to obtain an explicit resolution[^1] of the space $X\times X$ under the $G$-action. The basic idea is, roughly, to study the space $$Z:=T^*X\times_{{\mathfrak{g}}^*} T^*X,$$ which is the union of conormal bundles to the $G$-orbits on $X\times X$. Where the $G$-orbits are of codimension one, their conormal bundles are of dimension one, and the map from the projectivization: $${\mathbbm{P}}Z \to X\times X$$ restricts to an isomorphism. The important issue is to understand the conormal bundles where this map fails to be an isomorphism.
It eventually turns out that ${\mathbbm{P}}Z$ is not quite the correct resolution, because it can be quite singular. A closely related space is a space that I denote by ${\mathbbm{P}}J_X$, and which is obtained in Section \[sec:resolution\] as follows:
Let ${\mathfrak{a}}_X^*$ be the dual Lie algebra to the torus $A_X$, and ${\mathfrak{c}}_X^*={\mathfrak{a}}_X^*\sslash W_X$ — both of these spaces are isomorphic to the affine line. There is a smooth abelian group scheme $J$ over ${\mathfrak{c}}_X^*$ whose general fiber is isomorphic to $A_X$, but the isomorphism is only determined up to the action of $W_X$, and whose special fiber (over $0=$ the image of $0\in {\mathfrak{a}}_X^*$) is isomorphic to $\{\pm 1\} \times {\mathbbm{G}_a}$. This group scheme is known, for example, as the group scheme of regular centralizers over the Kostant section of $\mathfrak{sl}_2$, and it can be abstractly defined as $$J = \left({\mathrm{Res}}_{{\mathfrak{a}}_X^* / {\mathfrak{c}}_X^*} (A_X \times {\mathfrak{a}}_X^*)\right)^{W_X},$$ where ${\mathrm{Res}}_{{\mathfrak{a}}_X^* / {\mathfrak{c}}_X^*}$ denotes Weil restriction of scalars from ${\mathfrak{a}}_X^*$ to ${\mathfrak{c}}_X^*$.
Knop has shown [@KnAut] that, except perhaps for the non-identity component of the special fiber of $J$, this group scheme acts canonically on $T^*X$ over ${\mathfrak{g}}^*$. (This action is canonical in that its differential is the Hamiltonian vector field induced from canonical isomorphisms $T^*X\sslash G\xrightarrow\sim {\mathfrak{c}}_X^*$ and $\operatorname{Lie}(J) \simeq T^* {\mathfrak{c}}_X^*$.) Thus, being a bit imprecise as far as the action of the non-identity component of the special fiber goes, we have a map $$J\times_{{\mathfrak{c}}_X^*} T^*X \to T^*X\times_{{\mathfrak{g}}^*} T^*X,$$ and it is its composition with the map to $X\times X$ (after projectivization) which will give rise to the desired resolution: $${\mathbbm{P}}(J\times_{{\mathfrak{c}}_X^*} T^*X) \to X\times X.$$
On the other hand, we have, by definition, a canonical quotient map $J\to A_X\sslash W_X$, and this can be used to prove the isomorphism $X\times X\sslash G\xrightarrow\sim A_X\sslash W_X$.
Recall that there is a bijection between points of $X\times X\sslash G$ and closed (geometric) orbits of $G$ on $X\times X$. The diagonal $X_1=X^{{\operatorname{diag}}}\hookrightarrow X\times X$ corresponds to the class of $1\in A_X$, and the fiber of $J\times_{{\mathfrak{c}}_X^*} T^*X$ over it is just its conormal bundle $N^*_{X_1} = T^*X$. To correct the imprecision about the non-identity component $\{-1\}\times {\mathbbm{G}_a}$ of the special fiber of $J$, we replace $(\{-1\}\times {\mathbbm{G}_a}) \times_{{\mathfrak{c}}_X^*} T^*X$ by a copy of ${\mathbbm{G}_a}\times_{{\mathfrak{c}}_X^*} N_{X_{-1}}^*$, where $X_{-1}$ denotes the closed $G$-orbit over $[-1]\in
A_X\sslash W_X$, $N^*_{X_{-1}}$ denotes its conormal bundle, and ${\mathbbm{G}_a}$ maps to $0\in {\mathfrak{c}}_X^*$. This replacement leads to a smooth scheme $J_X\rightrightarrows T^*X$, birational to $J\times_{{\mathfrak{c}}_X^*} T^*X$, such that the resulting map from its “projectivization” $${\mathbbm{P}}J_X \to X\times X,$$ is isomorphic to the blowup of $X\times X$ at the closed orbits $X_1$ and $X_{-1}$.
The formula of Theorem \[Xtheorem\] on the codimensions of orbits is obtained in Section \[sec:integration\] by a degeneration argument — developing the analog of the Weyl integration formula for $X\times X$ under the diagonal $G$-action, and deforming $X$ to its horospherical “boundary degeneration” $X_\emptyset$, where this integration formula is very explicit.
The map ${\mathbbm{P}}J_X \to A_X\sslash W_X$ is easy to describe, and a standard analysis of pullbacks of Schwartz measures under resolutions shows, in Section \[sec:Schwartzmeasures\], that the elements of the pushforward space $\mathcal S(X\times X/G)$ are measures on $A_X\sslash W_X\simeq \mathbbm A^1$ whose singularities at $[\pm 1]$ are linear combinations of multiplicative characters of the form $x\mapsto |x|^{\frac{d_{\pm 1}}{2}-1} \eta(x)$, where $\eta$ is quadratic.
The last task, in Section \[sec:germs\] is to understand this linear combination of these characters. By linearization, this is equivalent to understanding the pushforwards of Schwartz measures under a map $$V\xrightarrow{Q} \mathbbm A^1,$$ where $Q$ is a nondegenerate, split quadratic form on a vector space $V$ of dimension $d=d_{\pm 1}$. A key proposition, \[reductiontobasic\], identifies these pushforwards with twisted pushforwards on a two- or three-dimensional quadratic space, as in Theorem \[Xtheorem\].
Acknowledgments
---------------
This project started as a joint project with Daniel Johnstone and Rahul Krishna, who eventually undertook the proof of the fundamental lemma. I am very grateful to them for many helpful conversations, including an observation of R. Krishna which greatly simplified the proof of the key proposition \[reductiontobasic\]. I am very grateful to the University of Chicago and Ngô Bao Châu for their hospitality during the winter and spring quarters of 2017, at the end of which I discovered the existence of a uniform transfer operator, and to the Institute for Advanced Study where most of the writing was done during the academic year 2017–2018. I also thank Wee Teck Gan for providing several references. Finally, I would like to acknowledge my intellectual debt to Hervé Jacquet and Friedrich Knop, two formidable mathematicians who, in the 80s and 90s, from different perspectives, laid the ground for the questions that I am addressing in this paper, not always properly appreciated [@JaNotices] by the mainstream of the field; I believe that we have only seen the tip of the iceberg as far as relations between their work go.
This work was supported by NSF grants DMS-1502270, DMS-1801429, and by a stipend at the IAS from the Charles Simonyi Endowment.
The moment map and the structure of Borel orbits
================================================
Invariant theory of the cotangent bundle and its polarizations {#ssmomentmap}
--------------------------------------------------------------
Throughout the paper, $X$ will denote one of the homogeneous spherical varieties of rank one appearing in Table . However, in this section I revisit (and slightly reformulate) the theory of the cotangent bundle of $X$ due to Friedrich Knop, which holds true for any homogeneous, quasi-affine spherical variety $X$ under the action of a connected reductive group $G$.
To any such $X$, one attaches a conjugacy class of parabolics, denoted by $P(X)$, characterized by the property that, if $B\subset G$ is a Borel subgroup and $\mathring X\subset X$ its open orbit, (a representative of) $P(X)$ is given by $$P(X):= \{g\in G| \mathring Xg =\mathring X\}.$$
Let $A$ denote the reductive quotient of a Borel subgroup — viewed as a unique torus up to unique isomorphism, the so-called *(universal) Cartan of $G$*. Let ${\mathcal{B}}$ denote the full flag variety of $G$, and let ${\mathcal{B}}_X$ denote the flag variety of parabolics in the conjugacy class of $P(X)$.
The unipotent radical of a parabolic $P$ will be denoted by $U_P$. Having fixed a Borel subgroup $B$, we will also use the letter $N$ for $U_B$. The quotient $\mathring X\sslash N$ is a homogeneous space under the action of $A$; its action factors through the faithful action of a quotient $A\twoheadrightarrow A_X$ which we will call *the (universal) Cartan of $X$*. In fact, it is known that $A_X$ is a quotient of $P(X)$, and that $P(X)$ acts on $\mathring X\sslash N$ through this quotient. The rank of $A_X$ is, by definition, the rank of $X$; thus, for all varieties of Table , $A_X\simeq{\mathbbm{G}_m}$. The character group of $A_X$ will be denoted by $\Lambda_X$, and called the weight lattice of $X$. We use similar notation for other $B$-orbits (or $B$-orbit closures) $Y$: $\Lambda_Y$ will denote the set of characters of nonzero rational $B$-eigenfunctions on $Y$, and $A_Y = {\operatorname{Spec}}F[\Lambda_Y]$ the torus quotient by which $A$ acts on $Y\sslash N$. The *rank* of $Y$ is the rank of the group $\Lambda_Y$. It is known that $\mathring X$ has maximal rank among all $B$-orbits on $X$.
We will denote Lie algebras of algebraic groups by the same letter in Gothic lowercase, and linear duals by a star exponent. The cotangent space $T^*X$ of $X$ comes equipped with a moment map $$\mu: T^*X\to {\mathfrak{g}}^*.$$ This gives rise to a $G$-invariant map: $$T^*X \to {\mathfrak{g}}^*\sslash G= {\mathfrak{a}}^*\sslash W.$$ We let $\hat{\mathfrak{g}}^*={\mathfrak{g}}^*\times_{{\mathfrak{a}}^*\sslash W} {\mathfrak{a}}^*$, and $\tilde{\mathfrak{g}}^*=\{(Z,B)| Z\in {\mathfrak{g}}^*, B\in {\mathcal{B}}, Z\in {\mathfrak{u}}_B^\perp\}$; the latter is the Springer–Grothendieck resolution, and we have natural, proper maps $\tilde{\mathfrak{g}}^*\to \hat{\mathfrak{g}}^*\to{\mathfrak{g}}^*$.
We define the following covers of the cotangent bundle:
- $\widehat{T^*X} = T^*X \times_{{\mathfrak{a}}^*\sslash W} {\mathfrak{a}}^* = T^*X \times_{{\mathfrak{g}}^*} \hat{\mathfrak{g}}^*$.
- $\widetilde{T^*X} = \{(v,B)| v \in T^*X, B\in {\mathcal{B}}, \mu(v) \in {\mathfrak{u}}_B^\perp\} = T^*X \times_{{\mathfrak{g}}^*} \tilde{\mathfrak{g}}^*$.
Hence, we have proper maps $\widetilde{T^*X}\to \widehat{T^*X} \to T^*X$.
Following Knop, we construct canonical maps, that we will call *Knop’s sections*, $$\begin{aligned}
\hat\kappa_X: (X\times {\mathcal{B}}_X)^0 \times {\mathfrak{a}}_X^* & \to & \widehat{T^*X}, \\
\tilde\kappa_X: (X\times {\mathcal{B}})^0 \times {\mathfrak{a}}_X^* & \to & \widetilde{T^*X}\end{aligned}$$ over $X$, linear in the ${\mathfrak{a}}_X^*$-argument, where the exponent “$0$” denotes the subset of pairs $(x, P)$ with $x$ in the open $P$-orbit. The maps are given as follows: by linearity, it is enough to define them for the lattice $\Lambda_X={\operatorname{Hom}}(A_X,{\mathbbm{G}_m})\subset {\mathfrak{a}}_X^*$, where we consider characters as elements of ${\mathfrak{a}}_X^*$ by identifying them with their differentials at the identity. Let $\chi\in \Lambda_X$, $P\in {\mathcal{B}}_X$ or ${\mathcal{B}}$, and let $f_\chi$ be a rational, nonzero $P$-eigenfunction on $X$ with eigencharacter $\chi$. If $x$ is in the open $P$-orbit, then $$\hat\kappa_X(x,P,\chi) = (d_x \log f_\chi,\chi)$$ and $$\tilde\kappa_X(x, P,\chi) = (d_x \log f_\chi,P)$$ where $d_x$ denotes the differential evaluated at $x$, and $d_x \log f_\chi= \frac{d_x f_\chi}{f_\chi(x)}.$
The following facts are known, or can easily be inferred, from the work of Knop:
1. All maps $\widetilde{T^*X}\to \widehat{T^*X} \to T^*X$ are proper and dominant.
This is obvious from the definitions.
2. There is, by definition, a natural map $\widetilde{T^*X}\to X\times {\mathcal{B}}$; this associates to every irreducible component of $\widetilde{T^*X}$ a $G$-orbit on $X\times {\mathcal{B}}$, namely, the largest $G$-orbit in its image. If we fix $B\in {\mathcal{B}}$, $G$-orbits on $X\times {\mathcal{B}}$ are in bijection with $B$-orbits on $X$. Under this map, the irreducible components *of maximal dimension* in $\widetilde{T^*X}$ are in bijection with the *Borel orbits of maximal rank* in $X$.
This is [@KnOrbits Proposition 6.3]. We will denote by $\widetilde{T^*X}^\bullet$ the irreducible component corresponding to the open Borel orbit; it is the closure of the image of $\tilde\kappa_X$ in $\widetilde{T^*X}$.
3. Considering only these components of maximal dimension in $\widetilde{T^*X}$, and their images in $\widehat{T^*X}$, we obtain a canonical bijection between *the irreducible components of $\widehat{T^*X}$* and the *Borel orbits of maximal rank in $X$*.
This is [@KnOrbits Theorem 6.4], together with the non-degeneracy statement of [@KnMotion Lemma 3.1]. Thus, this bijection is characterized by the fact that the component corresponding to a $B$-orbit $Y$ contains all pairs $(v\in T^*_Y X, Z\in {\mathfrak{a}}^*)$ with $\mu(v)\in \mathfrak u_B^\perp$ and $Z$ its image under the canonical map $\mathfrak u_B^\perp\to {\mathfrak{a}}^*$. In particular, the closure of the image of Knop’s section $\hat\kappa_X$ is the irreducible component corresponding to the open $B$-orbit, to be denoted by $\widehat{T^*X}^\bullet$.
4. The stabilizer of $\widehat{T^*X}^\bullet$ under the natural action of $W$ on $\widehat{T^*X}$ (induced from its action on ${\mathfrak{a}}^*$) is a certain semidirect product $W_{L(X)}\rtimes W_X$, and its image in ${\mathfrak{a}}^*$ coincides with ${\mathfrak{a}}_X^*$. Here, $W_{L(X)}$ is the Weyl group of the Levi quotient of $P(X)$, which is the largest subgroup of $W$ acting trivially on ${\mathfrak{a}}_X^*$, and $W_X$ is the so-called *little Weyl group* of the spherical variety, which acts faithfully on ${\mathfrak{a}}_X^*$. For the examples of Table , $W_X={\mathbb{Z}}/2$.
The fact that $W_{L(X)}$ is precisely the centralizer of ${\mathfrak{a}}_X^*$ again follows from the non-degeneracy statement of [@KnMotion Lemma 3.1]; I point the reader to the proof of [@KnOrbits Theorem 6.2] for the other statements.
The image of the moment map, followed by the Chevalley quotient: $$T^*X \xrightarrow{\mu} \mathfrak g^* \to \mathfrak a^*\sslash W$$ is equal to the image of the map $$\mathfrak a_X^*\sslash W_X \to \mathfrak a^*\sslash W,$$ induced from the inclusion $\mathfrak a_X^*\hookrightarrow \mathfrak a^*$. Knop has shown that the map to ${\mathfrak{a}}^*\sslash W$ lifts canonically to a map $$\mu_{{\operatorname{inv}}}: T^*X \to {\mathfrak{c}}_X^*:={\mathfrak{a}}_X^*\sslash W_X,$$ descending from the map $\widehat{T^*X}^\bullet\to {\mathfrak{a}}_X^*$. Under this map, ${\mathfrak{c}}_X^*$ is identified with the invariant-theoretic quotient $T^*X\sslash G$ [@KnWeyl Satz 7.1]. The map $\mu_{{\operatorname{inv}}}$ is the *invariant moment map*. Thus, we have a commutative diagram $$\xymatrix{
T^*X \ar[r]\ar[dr]_{\sslash G} & {\mathfrak{g}}_X^* \ar[r]\ar[d]^{\sslash G} & {\mathfrak{g}}^* \ar[d]^{\sslash G}\\
& {\mathfrak{c}}_X^*= {\mathfrak{a}}^*_X\sslash W_X \ar[r] & {\mathfrak{a}}^*\sslash W,
}$$ where ${\mathfrak{g}}_X^*$ is the spectrum of the integral closure of the image of $F[{\mathfrak{g}}^*]$ in $F[T^*X]$ (denoted $M_X$ in [@KnWeyl §6]).
We let $\mathring{\mathfrak{a}}_X^*$ denote the open subset where $W_X$ acts freely and $\mathring{\mathfrak{c}}_X^*$ its image — in our rank-one cases, these are just the complements of zero. Vectors in $T^*X$ (and its various covers) which live over $\mathring{{\mathfrak{c}}}_X^*$ will be called *regular semisimple*, and denoted $T^*X^{{\operatorname{rs}}}$. The reader should not confuse this notion with the property of being regular in ${\mathfrak{g}}^*$; in fact, the centralizer of the image of an element of $T^*X^{{\operatorname{rs}}}$ in ${\mathfrak{g}}^*$ is conjugate to a Levi of $P(X)$ over the algebraic closure. Hence, “regular semisimple” elements in $T^*X$ have an image in ${\mathfrak{g}}^*$ which is semisimple and “as regular as possible”, though not necessarily regular.
Now we restrict to the case when $X$ has rank one and $W_X={\mathbb{Z}}/2$. Thus, ${\mathfrak{a}}_X^*$ is a one-dimensional vector space, and $\mathring{\mathfrak{a}}_X^*$ is the complement of zero. We can also identify ${\mathfrak{c}}_X^*$ with $\mathbbm A^1$, always letting the point $0\in {\mathfrak{c}}_X^*$ (the image of $0\in {\mathfrak{a}}_X^*$) correspond to $0\in \mathbbm A^1$. Then, the invariant moment map $T^*X\to {\mathfrak{c}}_X^*$ can be considered as a quadratic form on the fibers of $T^*X$.
\[nondegenerate\] For $X$ affine homogeneous of rank one with $W_X={\mathbb{Z}}/2$, the map $T^*X\to {\mathfrak{c}}_X^*$ is a nondegenerate quadratic form on every fiber, and ${\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$ is equal to the image of $T^*X^{{\operatorname{rs}}}$ in ${\mathfrak{g}}^*$.
Let $x\in X$ with stabilizer $H$; the fiber of $T^*X$ over $x$ is canonically identified with $\mathfrak h^\perp$. Since $H$ is reductive, the restriction to $\mathfrak h^\perp$ of a nondegenerate invariant symmetric form on ${\mathfrak{g}}^*$ is also nondegenerate, and the corresponding quadratic form $$\mathfrak h^\perp \to \mathbbm A^1$$ is $H$-invariant. It thus has to factor through the invariant-theoretic quotient $$\mathfrak h^\perp\sslash H = T^*X\sslash G = {\mathfrak{c}}_X^*,$$ which therefore has to be nondegenerate.
By [@KnWeyl Satz 5.4], the closure of the image of the moment map is equal to the set of $G$-translates of $(\mathfrak l_1+{\mathfrak{u}}_P)^\perp$, where $P$ is a parabolic in the class of $P(X)$, and $L_1 U_P$ is the kernel of the map $P\to A_X$. Choosing a linear section $\sigma$ of the natural quotient map $(\mathfrak l_1+{\mathfrak{u}}_P)^\perp\to {\mathfrak{a}}_X^*$, the preimage of $\mathring {\mathfrak{a}}_X^*$ in this set is equal to $\sigma(\mathring {\mathfrak{a}}_X^*)\cdot G$, which is smooth. The quotient $\sigma(\mathring {\mathfrak{a}}_X^*)\cdot G\sslash G$ is isomorphic to $\mathring c_X^*$ (because ${\mathfrak{a}}_X^*$ is of dimension one, and $W_X={\mathbb{Z}}/2$, so ${\mathfrak{c}}_X^*\to {\mathfrak{a}}^*\sslash W$ is necessarily an embedding), thus the finite map of normal $G$-varieties: ${\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}\to \sigma(\mathring {\mathfrak{a}}_X^{*,{{\operatorname{rs}}}})\cdot G$ has to be an isomorphism.
\[rsinKnop\] For $X$ of rank one with $W_X={\mathbb{Z}}/2$, the restrictions of Knop’s sections $\hat\kappa_X$ and $\tilde\kappa_X$ to $\mathring{\mathfrak{a}}_X^*$ are isomorphisms onto the subsets of regular semisimple vectors on $\widehat{T^*X}^\bullet$, resp. $\widetilde{T^*X}^\bullet$.
We have a dominant, proper map $\widetilde{T^*X}^\bullet\to \widehat{T^*X}^\bullet$, and the image of $\tilde\kappa_X$ surjects onto the image of $\hat\kappa_X$, so it is enough to prove the proposition for $\widetilde{T^*X}$.
The $B$-orbits of non-maximal rank have rank zero, and therefore any cotangent vector over such an orbit which is perpendicular to $\mathfrak u_B$ maps to $0\in {\mathfrak{a}}^*$. Therefore, the regular semisimple elements $(v, B)\in \widetilde{T^*X}^{{\operatorname{rs}}}$ all live over $B$-orbits of maximal rank, that is, if $x$ is the image of $v$ in $X$, then $x\cdot B$ is a $B$-orbit of maximal rank. It suffices to show that those which belong to $\widetilde{T^*X}^\bullet$ live over the open orbit.
If not, that is, if there is a regular semisimple vector $(v,B)\in \widetilde{T^*X}^\bullet$ which lives over an orbit $Y$ of maximal rank other than the open one, hence also belongs to the irreducible component of $\widetilde{T^*X}^Y\subset \widetilde{T^*X}$ indexed by $Y$, that means that the intersection of two distinct irreducible components $$\widetilde{T^*X}^\bullet \cap \widetilde{T^*X}^Y$$ contains regular semisimple vectors. In particular, the same holds for the intersection of the corresponding irreducible components of $\widehat{T^*X}$, $$\widehat{T^*X}^\bullet \cap \widehat{T^*X}^Y.$$
I claim that, in rank one with $W_X={\mathbb{Z}}/2$, the map $\widehat{T^*X}\to T^*X$ is étale over $T^*X^{{\operatorname{rs}}}$; indeed, the image of $T^*X^{{\operatorname{rs}}}$ is the subset $\mathring{\mathfrak{c}}^*_X$ of ${\mathfrak{a}}^*\sslash W$, and the normalizer of ${\mathfrak{a}}_X^*$ in $W$ has to coincide with $W_{L(X)}\rtimes W_X$, because $W_X={\mathbb{Z}}/2$ is the group of automorphisms finite order of the lattice $\Lambda_X\subset {\mathfrak{a}}_X^*\simeq\mathbbm A^1$, and $W_{L(X)}$ is its centralizer. The distinct $W$-conjugates of $\mathring{\mathfrak{a}}_X^*$ have empty intersections (because these are distinct one-dimensional vector subspaces of ${\mathfrak{a}}^*$ with their origins removed), thus we have $$\widehat{T^*X}^{{\operatorname{rs}}}= \bigsqcup_{w\in W_{L(X)}\rtimes W_X\backslash W} T^*X^{{\operatorname{rs}}}\times_{\mathring{\mathfrak{c}}^*_X} (\mathring{\mathfrak{a}}_X^*)^w$$ as algebraic varieties, where $w$ denotes a representative for the given coset, and $(\mathring{\mathfrak{a}}_X^*)^w$ is the $w$-conjugate of $\mathring{\mathfrak{a}}_X^*$ inside of ${\mathfrak{a}}^*$. Hence, the map $\widehat{T^*X}^{{\operatorname{rs}}}\to T^*X^{{\operatorname{rs}}}$ is the base change of the étale maps $(\mathring{\mathfrak{a}}_X^*)^w\to \mathring{\mathfrak{c}}_X^*$, hence étale.
But this implies that the components of $\widehat{T^*X}^{{\operatorname{rs}}}$ have empty intersections.
\[corollaryGtransitive\] In the setting of the previous proposition, $G$ acts transitively on every fiber of $\widetilde{T^*X}^{\bullet,{{\operatorname{rs}}}}$, $\widehat{T^*X}^{\bullet,{{\operatorname{rs}}}}$ or $T^*X^{{\operatorname{rs}}}$ over, respectively, $\mathring{\mathfrak{a}}^*_X$, $\mathring{\mathfrak{a}}^*_X$ or $\mathring{\mathfrak{c}}^*_X$.
It is enough to prove the statement for $\widetilde{T^*X}^{\bullet,{{\operatorname{rs}}}}$. By Proposition \[rsinKnop\], Knop’s section is an isomorphism onto regular semisimple vectors: $$(X\times {\mathcal{B}})^0 \times \mathring{\mathfrak{a}}_X^* \xrightarrow\sim \widetilde{T^*X}^{\bullet,{{\operatorname{rs}}}}.$$ The group $G$ acts transitively on $(X\times {\mathcal{B}})^0$, hence on the fiber over any point in $\mathring{\mathfrak{a}}_X^*$.
In particular, considering the map $T^*X\to {\mathfrak{g}}_X^*$ and setting $$\hat{\mathfrak{g}}_X^*={\mathfrak{g}}_X^*\times_{{\mathfrak{c}}_X^*} {\mathfrak{a}}_X^*,$$ so that we have a map $$\widehat{T^*X}^\bullet\to \hat{\mathfrak{g}}_X^*,$$ the $G$-stabilizer of any element $\hat Z\in \hat{\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$, resp. $Z\in {\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$, acts transitively on its fiber in $\widehat{T^*X}^\bullet$, resp. $T^*X$. The $G$-stabilizer of such an element is a Levi subgroup of $G$ over the algebraic closure and, in the case of $\hat{\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$, a Levi subgroup $L$ equipped with a choice of parabolic $P$ in the class of $P(X)$. (Indeed, this is the parabolic in the class $P(X)$ containing $B$, for any lift $(Z, B)$ of $\hat Z$ to $\tilde{\mathfrak{g}}_X^*$ — equivalently, the parabolic $P\in {\mathcal{B}}_X$ in the domain of Knop’s section.) Knop has shown [@KnMotion Proposition 2.4] that $L$ acts on the fiber of $\hat Z$ precisely through the quotient $L\to P\to A_X$. This gives rise to a canonical action of $A_X$ on $\widehat{T^*X}^{\bullet,{{\operatorname{rs}}}}$, commuting with the action of $G$:
$$\label{AXaction}
A_X\times \widehat{T^*X}^{\bullet,{{\operatorname{rs}}}}\to \widehat{T^*X}^{\bullet,{{\operatorname{rs}}}}.$$
Corollary \[corollaryGtransitive\] implies:
\[polarizedtorsor\] In the setting of Proposition \[rsinKnop\], $\widehat{T^*X}^{\bullet,{{\operatorname{rs}}}}$ is (canonically) an $A_X$-torsor over $\hat{\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$.
In Section \[sec:resolution\], we will see (following Knop, again) how to formulate the analog of this for the map $T^*X^{{\operatorname{rs}}}\to {\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$, and to extend it to an action of a group scheme over the whole space ${\mathfrak{g}}_X^*$.
Borel orbits over the algebraic closure
---------------------------------------
From now, $X$ is always affine homogeneous spherical $G$-variety of rank one, with $W_X={\mathbb{Z}}/2$. Its weight lattice $\Lambda_X$ (the character group of $A_X$) is thus isomorphic to ${\mathbb{Z}}$.
In the present subsection, we work over $\bar F$, the algebraic closure of $F$.
Recall, again, that the rank of a $B$-orbit on $X$ is the rank of the torus $B_x\backslash B/N$, where $B_x$ is the stabilizer of a point $x$ on the orbit, and $N\subset B$ the unipotent radical. For what follows, for a Borel subgroup $B$ and a simple positive root $\alpha$, we denote by $P_\alpha$ the parabolic generated by $B$ and the simple root space associated to the root $-\alpha$, and by $\mathcal R(P_\alpha)$ its radical (so that $P_\alpha/\mathcal R(P_\alpha)\simeq {\operatorname{PGL}}_2$). (We also use the notation $P_{\alpha\beta}$ for the parabolic generated by $B$ and the negative root spaces associated to two simple roots $\alpha, \beta$, etc.)
Knop has defined in [@KnOrbits] a rank-preserving action of the Weyl group of $G$ on the set of Borel orbits (or, equivalently, the set of $B$-orbit closures), which is transitive on the subset of orbits of maximal rank, which includes the open orbit. For the reflection $w_\alpha$ associated to a simple root $\alpha$, and a $B$-orbit $Y$, one considers the spherical ${\operatorname{PGL}}_2$-variety $YP_\alpha/\mathcal R(P_\alpha)$ which is of one of the following four types:
1. ${\operatorname{PGL}}_2\backslash {\operatorname{PGL}}_2$, i.e., a point;
2. $T\backslash {\operatorname{PGL}}_2$, where $T$ is a torus;
3. $\mathcal N(T)\backslash {\operatorname{PGL}}_2$, where $\mathcal N(T)$ is the normalizer of a torus;
4. $S\backslash {\operatorname{PGL}}_2$, with $N_2\subset S \subset B_2$, where $B_2\supset N_2$ denote the Borel subgroup of ${\operatorname{PGL}}_2$, and its unipotent radical.
In the first three cases, there is a single orbit of largest rank in $YP_\alpha$, and it is fixed by $w_\alpha$. In the last case, there are two such orbits, say $Y$ and $Z$, and $w_\alpha$ interchanges them; moreover, for their character groups $\Lambda_Y={\operatorname{Hom}}(A_Y,{\mathbbm{G}_m})$, $\Lambda_Z={\operatorname{Hom}}(A_Z,{\mathbbm{G}_m})$, we have: $$\label{charsKnopaction}
\Lambda_Z = \Lambda_Y^{w_\alpha}$$ inside of ${\operatorname{Hom}}(A,{\mathbbm{G}_m})$.
Since, in our case, $X$ is of rank one, all $B$-orbits are of rank one or zero.
Following Brion [@BrOrbits], we have by [@SV §3.1]:
\[Briongenerators\] There is a $B$-orbit $Z$ of rank one, and a simple root $\alpha$, or two orthogonal simple roots $\alpha,\beta$, such that, setting $P=P_\alpha$, resp. $P=P_{\alpha\beta}$, the spherical variety $ZP/\mathcal R(P)$ is isomorphic to one of the following:
1. $T\backslash {\operatorname{PGL}}_2$, where $T$ is a torus;
2. $\mathcal N(T)\backslash {\operatorname{PGL}}_2$, where $\mathcal N(T)$ is the normalizer of a torus;
3. ${\operatorname{PGL}}_2$ as a ${\operatorname{PGL}}_2^2$-space (when $P=P_{\alpha\beta}$).
Moreover, these possibilities are mutually exclusive, as in the first case the weight lattice $\Lambda_X$ is spanned by a root $\gamma$ of $G$, in the second case it is spanned by the double $2\gamma$ of a root, and in the third case it contains the sum $\gamma=\gamma_1+\gamma_2$ of two strongly orthogonal roots.
The weight $\gamma$ of the lemma above called the (normalized) *spherical root* of $X$, by [@SV §3.1]. Correspondingly to the three cases, we will say that the spherical root is of type $T$, $N$ or $G$.
I caution the reader that this is not the standard normalization of spherical roots in the theory of spherical varieties (e.g., as in [@Luna]), and it also differs from a different normalization that appears in [@KnAut]; however, I will call $\gamma$ the spherical root, unless there is a danger of confusion, in which case the adjective “normalized” will appear.
If the spherical root is of type $G$, all $B$-orbits are of rank one.
Indeed, otherwise there will be a $B$-orbit $Z$ of rank zero, and a simple root $\alpha$, such that $ZP_\alpha$ contains a $B$-orbit of rank one. But this can happen only if $ZP_\alpha/\mathcal R(P_\alpha)$ is isomorphic to $T\backslash {\operatorname{PGL}}_2$ or $\mathcal N(T)\backslash {\operatorname{PGL}}_2$, which is impossible by the above lemma.
We now study closed $B$-orbits, first over the algebraic closure:
Let $Z\subset X$ be a $B$-orbit, and $H\subset G$ the stabilizer of a point on $Z$. The following are equivalent:
1. $H\cap B$ is a Borel subgroup of $H$;
2. $Z$ is closed.
Here, for non-connected groups, by slight abuse of language we use “Borel” for any solvable subgroup such that the quotient is projective, whether it is connected or not.
This is obvious, by considering $H$-orbits on the flag variety of $G$.
\[lemmaclosedorbits\] Let $Z\subset X$ be a closed $B$-orbit. Then, one of the following two holds:
1. $Z$ is of rank zero, and for every simple root $\alpha$ such that $Y:=ZP_\alpha\ne Z$, we have $Y/\mathcal R(P_\alpha) \simeq T\backslash {\operatorname{PGL}}_2$ or $\mathcal N(T)\backslash {\operatorname{PGL}}_2$ (notation as above);
2. or, $Z$ is of rank one, and for all simple roots $\alpha$ we have $ZP_\alpha=Z$, except for two orthogonal simple roots $\alpha, \beta$ for which, setting $Y:=ZP_{\alpha\beta}$, we have $Y/\mathcal R(P_{\alpha\beta}) \simeq
{\operatorname{PGL}}_2^{{\operatorname{diag}}}\backslash {\operatorname{PGL}}_2^2$.
In the first case, we only need to exclude the possibility that $ZP_\alpha/\mathcal R(P_\alpha)=S\backslash {\operatorname{PGL}}_2$ with $N_2\subset S\subset B_2$. Let $H\subset G$ be the stabilizer of a point on $Z$. Since $Z$ is of rank zero, $H\cap B$ and $G\cap B$ have the same rank and, in particular, contain a common maximal torus $T$. Decomposing the Lie algebras of $G$ and $H$ into $T$-eigenspaces, we see that each root space for $H$ is also a root space for $G$. This means that the statement $ZP_\alpha/\mathcal R(P_\alpha)=S\backslash {\operatorname{PGL}}_2$ lifts to the statement that $\mathfrak h\cap \mathfrak p_\alpha$ contains the root space corresponding to the root $\alpha$ (but not its opposite). Thus, if $L_\alpha'$ denotes the commutator of the standard Levi (with respect to $T$) of $P_\alpha$, then $Y=ZP_\alpha$ contains the space $(T\cap L_\alpha')\backslash L_\alpha'$ as a subvariety. But this is nontrivial projective, a contradiction, since $X$ is assumed affine.
In the second case, observe first that the spherical root is necessarily of type $G$. Indeed, if $H$ is the stabilizer of a point in the closed orbit, it follows from the previous lemma that ${{\operatorname{rk}}}(H)={{\operatorname{rk}}}(G)-1$. Since $X$ is homogeneous, this holds for all stabilizers, and therefore there cannot be a Borel orbit of rank zero.
I will now rely on the classification of Table , since I currently do not have a proof which avoids any kind of classification. Since all $B$-orbits are of rank one, they form a partially ordered set (by dimension) which can be identified, using Knop’s action, with the homogeneous set $(W_{L(X)}\rtimes W_X)\backslash W$ for the Weyl group, with the minimal lentgh of a representative of a coset corresponding to the codimension of the orbit. The little Weyl group $W_X={\mathbb{Z}}/2$, for spherical roots of type $G$ (i.e., the cases $D_n$–$B_3''$ of Table \[thetable\]), is generated by the element $w_\gamma=w_{\gamma_1} w_{\gamma_2}$, where $\gamma_1,\gamma_2$ are the two strongly orthogonal roots such that $\gamma=\gamma_1+\gamma_2$. These partially ordered sets, together with the graph of Knop’s action, have been depicted in [@SaSph 6.19, 6.16], and one observes that there is actually a unique minimal element (Borel orbit) in this partially ordered set, and two unique, mutually orthogonal simple roots raising it to the same Borel orbit.
A pair $(Y,P)$ as in the lemma, where $P=P_\alpha$ in the first case and $P=P_{\alpha\beta}$ in the second, will be called *a basic orbit-parabolic pair*. Notice that $Y\subset X$ is closed, since this is the case for $Z$ and the action map $Z\times^B P\to X$ is proper. Such a pair will play an important role in various arguments in this paper, since by the above lemma it allows us to reduce many arguments to the basic rank-one cases, labelled $A_1$ and $D_2$ in Table . In reality, the precise choice of parabolic will never matter; only its class matters, and if ${\mathcal{B}}_P$ denotes the flag variety of parabolics in this class, $Y$ can be replaced by the $G$-orbit on $X\times {\mathcal{B}}_P$, whose fiber over $P$ is $Y$.
In this paper we do not consider arbitrary homogeneous spherical varieties of rank one; it turns out that only one representative in the equivalence class of a variety modulo $G$-automorphisms is appropriate for the relative trace formula comparison that we are performing. More precisely:
\[correctrepresentative\] If the spherical root $\gamma$ is of type $N$ — equivalently, if $\Lambda_X$ is spanned by $2\gamma$, then there is an equivariant two-fold cover $X'\to X$ with $\Lambda_{X'}$ spanned by $\gamma$.
If the spherical root $\gamma$ is of type $G$, then there is an equivariant finite cover $X'\to X$ (possibly $X'=X$) such that $\Lambda_{X'}$ is spanned by $\frac{\gamma}{2}$. Moreover, ${{\operatorname{Aut}}}^G(X')=\mathbb Z/2$.
Moreover, in both cases, the stabilizers of points on $X'$ are connected.
The first statement is [@Luna Lemme 6.4.1].
For the second, if $X=H\backslash G$, replace $G$ by the simply connected cover of its derived group; it necessarily acts transitively on $X$, because if the connected center of $G$ was not acting trivially, the rank of $X$ would be greater than one. So, we can without loss of generality denote that by $G$. I claim that $\frac{\gamma}{2} \in {\operatorname{Hom}}(A,{\mathbbm{G}_m})$. To show this, it is enough to show that its pairing with every simple coroot is an integer. Without loss of generality, we may replace $\gamma$ by its Weyl group conjugate $\gamma'=\alpha-\beta$ which belongs to the character group of a closed $B$-orbit $Z$, where $\alpha, \beta$ are two orthogonal simple roots, as in Lemma \[lemmaclosedorbits\]. Clearly, the pairing of $\frac{\gamma'}{2}$ with $\check\alpha, \check\beta$ is integral. On the other hand, by the same lemma, for every simple root $\delta\ne \alpha, \beta$, we have $YP_\delta = Y$, hence $\left<\gamma', \check\delta\right>=0$, proving the claim. The result on the existence of $X'$ now follows as a simple special case of [@Luna Théorème 2]; it also follows from the same that there exists an $X$ whose weight lattice is spanned by $\gamma$. The cover $X'\to X$ gives rise to a nontrivial involution $\tau$ of $X'$, hence to an embedding ${\mathbb{Z}}/2\hookrightarrow {{\operatorname{Aut}}}^G(X')$. (This embedding can also be established as a special case of [@KnAut Theorem 1.2].) On the other hand, we also have an embedding $$\iota:{{\operatorname{Aut}}}^G(X') \hookrightarrow {\operatorname{Hom}}(\Lambda_{X'}, \bar F^\times),$$ by considering the action of an automorphism on the lines of $B$-eigenvectors in the function field $\bar F(X')$. To show that ${\mathbb{Z}}/2$ is the whole automorphism group, it suffices to show that $\iota(\tau)$ is trivial on the (normalized) spherical root $\gamma$.
The set of $B$-orbits of maximal rank is acted upon transitively by the Weyl group action of Knop. Any $G$-automorphism $\sigma$ of $X'$ preserves the open Borel orbits and commutes with Knop’s action on the set of Borel orbits, hence preserves all orbits of maximal rank. Moreover, it is easy to see (by the definition of Knop’s action, and reduction to the case of ${\operatorname{SL}}_2$) that the analogs of the map $\iota$ for any $B$-orbit $Z$ of maximal rank: $$\iota_Z: {{\operatorname{Aut}}}^G(X') \to {\operatorname{Hom}}(\Lambda_Z, \bar F^\times)$$ are compatible with the $W$-action: $\iota_{Z^w}(\sigma) = \iota_Z(\sigma)^w$. (Recall that the character groups are related by .)
Let $(Y',P_{\alpha\beta})$, now, be a basic orbit-parabolic pair for $X'$, as in Lemma \[lemmaclosedorbits\]. Since $Y'$ is preserved by $G$-automorphisms, a $G$-automorphism $\sigma$ of $X'$ restricts to a $P_{\alpha\beta}$-automorphism of $Y'$. The quotient $Y'/\mathcal R(P_{\alpha\beta})$ is isomorphic to ${\operatorname{PGL}}_2$, and the weight lattice of its closed $B$-orbit is spanned by the weight $\alpha-\beta$. Since ${\operatorname{PGL}}_2$ has no ${\operatorname{PGL}}_2\times{\operatorname{PGL}}_2$-automorphisms, we see that $\iota_{Y'}(\sigma)$ has to be trivial on $\alpha-\beta$, and therefore $\iota(\sigma)$ has to be trivial on its $W$-conjugate $\gamma$.
If the stabilizers were not connected, then there would exist a further cover $X''\to X'$, giving an embedding of character groups of rank one with nontrivial cokernel: $\Lambda_{X'}\subset \Lambda_{X''}$. But this is impossible in both cases for the spherical root $\gamma$ of $X'$: If it is of type $T$, the same should hold for $X''$, and their character groups are both spanned by $\gamma$; if it is of type $G$, the same should hold for $X''$, and the character group in such a case is spanned by either $\gamma$ or $\frac{\gamma}{2}$ (in this case, by the latter), because no smaller fraction of $\gamma$ can belong to the weight lattice of $A$. Thus, stabilizers are connected.
A variety $X'$ as in the proposition will be called the “correct representative” (of its equivalence class modulo $G$-automorphisms).
Borel orbits and the moment map over $F$ {#ssrationality}
----------------------------------------
We now return to our non-algebraically closed field $F$ in characteristic zero, with $G$ split over $F$. We maintain the assumptions of the previous subsection for $X$, and, moreover, we assume that it is the correct representative given by Proposition \[correctrepresentative\].
Under these assumptions, our main goal for the rest of this section is to prove the following:
\[rationality\] The following are equivalent:
1. The stabilizer $H$ of one, equivalently any, $F$-point on $X$ is a split reductive group.
2. One, equivalently every, closed $B_{\bar F}$-orbit on $X_{\bar F}$ is defined over $F$.
3. The invariant moment map $T^*X\to {\mathfrak{c}}_X^*$, viewed as a quadratic form on the fibers of $T^*X$, is split (maximally isotropic) on one, equivalently every, fiber.
Notice that $H$ is connected, by the fact that we are working with correct representatives, and Proposition \[correctrepresentative\]. The proposition is not true without the assumption on “correct representatives”, e.g., for the variety $\mathcal N(T)\backslash {\operatorname{PGL}}_2$.
We begin with some preliminary lemmas and constructions. If $B\subset G$ is a Borel subgroup, since $B$ is split solvable, by standard Galois cohomology, every $B$-orbit which is defined over $F$ has an $F$-point. This holds, in particular, for the open $B$-orbit, which is unique hence defined over $F$, therefore $X$ has $F$-points.
\[maxrankdefined\] All $B$-orbits of maximal rank (over $\bar F$) are defined over $F$. Moreover, if $Y$ is a $B$-orbit of maximal rank, then $Y(F)$ meets any $G(F)$-orbit on $X(F)$ nontrivially.
The open $B$-orbit is defined over $F$ and has rank one. By the definition of Knop’s action of the Weyl group on the set of Borel orbits, and the fact that $G$ is split, one sees from the definition that the action is defined over $F$, therefore every $F$-orbit of maximal rank is defined over $F$. Any $G(F)$-orbit is open in $X(F)$ (in the Hausdorff topology induced by the topology of $F$), and therefore has to contain $F$-points of the open $B$-orbit. Finally, we use Knop’s action to deduce the same result for any $B$-orbit of maximal rank: if $Y^{w_\alpha} = Z$, with $Z$ open in $YP_\alpha$, and $Z(F)$ contains a point $z$ in a given $G(F)$-orbit, then there is a $g\in P_\alpha(F)$ such that the $P_\alpha(F)$-stabilizer of $zg$ is contained in $B(F)$ (by the fact that the map $P_\alpha(F)\to B\backslash P_\alpha(F)$ is surjective), therefore the same $G(F)$-orbit also contains a point of $Y(F)$.
Now let $(Y,P)$ be a basic orbit-parabolic pair, see Lemma \[lemmaclosedorbits\]. By this lemma, the variety $Y$ contains a $B$-orbit of maximal rank (rank one), which is therefore defined over $F$. Therefore, $Y$ is defined over $F$. Moreover, since $Y$ is closed, the action map $Y\times^P G\to X$ is proper.
The above map is surjective on $F$-points, i.e., $Y(F)\times^{P(F)} G(F)\twoheadrightarrow X(F)$.
Since $Y$ contains a $B$-orbit of maximal rank, this follows from Lemma \[maxrankdefined\].
Now we define a proper cover of the cotangent bundle $T^*X$ which is intermediate between this and a component of the cover $\widetilde{T^*X}$ defined in §\[ssmomentmap\]. Let $(Y,P)$ be a basic orbit-parabolic pair. Similarly to the definition of $\widetilde{T^*X}$, we let $$\widetilde{T^*X}^P = \{(v,P')| v\in T^*X, P'\sim P, \mu(v) \in {\mathfrak{u}}_{P'}^\perp\}.$$ Here, $P'\sim P$ means that $P'$ is conjugate to $P$. (Really, $P$ denotes here a class of parabolics, and $Y$ can be thought of as a $G$-orbit on $P\backslash G\times X$; nothing depends on a choice in this class.)
We let $\widetilde{T^*X}^{P,Y} = T_Y^*X^{{\mathfrak{u}}_P^\perp} \times^P G$, where the exponent $~^{{\mathfrak{u}}_P^\perp}$ means that the image under the moment map belongs to ${\mathfrak{u}}_P^\perp$.
\[partialcover\] $\widetilde{T^*X}^{P,Y}$ is an irreducible component of $\widetilde{T^*X}^P$ and, hence, is proper and dominant over $T^*X$.
It is clearly irreducible, and a closed subset of $\widetilde{T^*X}^P$. Moreover, considering the natural map $\widetilde{T^*X}\to \widetilde{T^*X}^P$, it is clear from the definition that $\widetilde{T^*X}^{P,Y}$ contains the image of the irreducible component of maximal dimension corresponding to the open $B$-orbit in $Y$; therefore, it is an irreducible component of $\widetilde{T^*X}^P$, and dominant over $T^*X$.
Now let $Y_2=Y/U_P$. Since $Y$ is homogeneous under $P$, this is homogeneous under the Levi quotient $L$ of $P$, and its quotient by $\mathcal Z(L)$ is the rank-one spherical variety $Y/\mathcal R(P)$ described in Lemma \[lemmaclosedorbits\], hence isomorphic to $T\backslash {\operatorname{PGL}}_2$ or ${\operatorname{PGL}}_2$. There are natural quotient maps of (the total spaces of) vector bundles $$T_Y^*X^{{\mathfrak{u}}_P^\perp} \to T^*Y^{{\mathfrak{u}}_P^\perp} \to T^* Y_2,$$ with the former an isomorphism on the base and the latter an isomorphism on the fiber.
\[lemmaY2\] The connected center of $L$ acts trivially on $Y_2$, and $Y_2$ is isomorphic either to $T\backslash {\operatorname{PGL}}_2$, where $T$ is a torus, or to ${\operatorname{SL}}_2$. Moreover, we have $T^* Y_2\sslash L \xrightarrow\sim {\mathfrak{c}}_X^*$, fitting into a natural commutative diagram: $$\label{moddiagram}
\xymatrix{
\widetilde{T^*X}^{P,Y} \ar[r] \ar[d] & T^*X\ar[d] \\
T^* Y_2 \times^P G \ar[r] & {\mathfrak{c}}_X^*.
}$$
The variety $Y_{2,{{\operatorname{ad}}}}:= Y_2/\mathcal Z(L) = Y/\mathcal R(P)$ is, by Lemma \[lemmaclosedorbits\], isomorphic to $T\backslash {\operatorname{PGL}}_2$ or ${\operatorname{PGL}}_2$. If the identity component of $\mathcal Z(L)$ was acting nontrivially on $Y_2$, the rank of $Y_2$, hence of $Y$, would be larger than one, a contradiction. Thus, $Y_2 \to Y_{2,{{\operatorname{ad}}}}$ is a finite cover. Since $X$ is taken to be the correct representative in its class, the character group $\Lambda_X$ is generated by the spherical root $\gamma$, if that is of type $T$, or by $\frac{\gamma}{2}$, if it is of type $G$. Correspondingly, by , the character group of $Y$ (equivalently, of $Y_2$) is generated by $\alpha$, in the first case, and by $\frac{\alpha+\beta}{2}$, in the second, in the notation of Lemma \[lemmaclosedorbits\]. But this means that in the first case $Y_2=Y_{2,{{\operatorname{ad}}}}$, while in the second it has to be a two-fold connected cover of it, hence isomorphic to ${\operatorname{SL}}_2$.
We have $A_{Y_2} = A_Y = A_X^w$ for some element $w$ of the Weyl group of $G$, and $W_{Y_2} = W_X = {\mathbb{Z}}/2$. This gives rise to a canonical isomorphism $${\mathfrak{c}}_{Y_2}^* = {\mathfrak{c}}_X^*,$$ and the construction of the invariant moment maps $T^* Y_2\to {\mathfrak{c}}_{Y_2}^*$ and $T^*X\to {\mathfrak{c}}_X^*$ is clearly compatible with this isomorphism, showing the commutativity of the diagram.
\[propVP\] Let $(Y,P)$ be a basic orbit-parabolic pair; then the map $\widetilde{T^*X}^{P,Y} \to T^*X$ is surjective on $F$-points.
Let $y\in Y$ with image $y_2\in Y_2$, and let $V\supset V_P$ and $V_2$ be, respectively, the fibers of $T^*X$, $\widetilde{T^*X}^{P,Y}$, and $T^*Y_2$ over $y$, $(y,P)$, and $y_2$, respectively. The kernel of the map $V_P\to V_2$ is an isotropic subspace of $V$ (with respect to the quadratic map $V\to {\mathfrak{c}}_X^*$) of dimension $$\label{dimensions}
\dim\ker(V_P\to V_2) = \frac{\dim V - \dim V_2}{2}.$$
The quadratic space $V$ is split (maximally isotropic)[^2] if and only if $V_2$ is.
Notice that $V_2$ is isomorphic to $\mathbbm A^2$, for spherical roots of type $T$, and to $\mathfrak{sl}_2$, for spherical roots of type $G$, with $P\cap H$ acting through a $1$-dimensional torus quotient in the first case, and through a quotient isomorphic to ${\operatorname{PGL}}_2$, in the second case.
Since the map is proper, to prove surjectivity it is enough to prove that the image is dense.
Let $L$ denote the Levi quotient of the class of parabolics $P$, considered as an abstract algebraic group depending only on the class of $P$, defined uniquely up to conjugacy. We have a canonical map of coadjoint quotients $$\mathfrak l^*\sslash L \to {\mathfrak{g}}^*\sslash G.$$ Let $$\widehat{{\mathfrak{g}}^*}^P = {\mathfrak{g}}^* \times_{{\mathfrak{g}}^*\sslash G} \mathfrak l^*\sslash L,$$ $$\widetilde{{\mathfrak{g}}^*}^P = \{(Z,P')| P'\sim P, Z \in {\mathfrak{u}}_{P'}^\perp\},$$ so we have natural, proper maps $$\widetilde{{\mathfrak{g}}^*}^P \to \widehat{{\mathfrak{g}}^*}^P \to {\mathfrak{g}}^*.$$ At the level of $F$-points, the second arrow is not surjective, but the first arrow is.
Now, considering ${\mathfrak{c}}_X^*$ as a subset of ${\mathfrak{g}}^*\sslash G$ and, similarly, ${\mathfrak{c}}_{Y_2}^*$ as a subset of $ \mathfrak l^*\sslash L$, the canonical isomorphism ${\mathfrak{c}}_{Y_2}^*\simeq {\mathfrak{c}}_X^*$ that we saw in Lemma \[lemmaY2\] gives rise to a lift: $$\label{liftcX} {\mathfrak{c}}_X^* \to \mathfrak l^*\sslash L.$$ The composition of this with the invariant moment map gives a lift: $$\label{liftTX} T^*X \to \widehat{T^*X}^P:= T^*X \times_{{\mathfrak{g}}^*} \widehat{{\mathfrak{g}}^*}^P,$$ which fits into a commutative diagram $$\xymatrix{ \ar[d]\widetilde{T^*X}^Y \ar[r] & \widehat{T^*X}^Y \ar[d]\ar[r] & {\mathfrak{a}}^*_Y \ar[d] \\
\widetilde{T^*X}^{P,Y} \ar[dr]\ar[r] & \widehat{T^*X}^P \ar[r]\ar@/_/[d] & {\mathfrak{c}}_X^*\\
& T^*X, \ar@/_/[u]\ar[ur] & }$$ where $\widetilde{T^*X}^Y$, $\widehat{T^*X}^Y$ denote the irreducible components corresponding to the open $B$-orbit in $Y$ (see §\[ssmomentmap\]). The scheme-theoretic image of $T^*X$ in $\widehat{T^*X}^P$ is the same as the image of $\widehat{T^*X}^Y$, hence an irreducible component $\widehat{T^*X}^{P,Y}$ of $\widehat{T^*X}^P$. The fact that the map $\widetilde{{\mathfrak{g}}^*}^P \to \widehat{{\mathfrak{g}}^*}^P$, hence its base change $\widetilde{T^*X}^{P,Y} \to \widehat{T^*X}^{P,Y}$, is surjective on $F$-points, proves the surjectivity statement.
Let us now count dimensions. The map $\widetilde{T^*X}^Y\to \widetilde{T^*X}^{P,Y}$ is finite, hence $$\dim \widetilde{T^*X}^{P,Y} = \dim \widetilde{T^*X}^Y = \dim \widehat{T^*X} + \dim {\mathcal{B}}_{L(X)} = \dim T^*X + \dim {\mathcal{B}}_{L(X)},$$ where ${\mathcal{B}}_{L(X)}$ denotes the full flag variety of the Levi quotient of $P(X)$. In terms of the above vector spaces, setting $P_H= P\cap H$, where $H$ is the stabilizer of $y\in Y$, we have $$T^*X = V\times^H G,$$ $$\widetilde{T^*X}^{P,Y} = V_P \times^{P_H} G,$$ hence the relation above translates to $$\label{firstdim}
\dim V_P + \dim (P_H\backslash H) = \dim V + \dim {\mathcal{B}}_{L(X)}.$$
On the other hand, by Corollary \[corollaryGtransitive\], the general fiber of the map $T^*X\to {\mathfrak{c}}_X^*$ is a single $G$-orbit, which is of the form $L_1\backslash G$ over the algebraic closure, where $L_1$ is (non-canonically) isomorphic to $\ker(L(X)\to A_X)$, with $L(X)$ a Levi of $P(X)$; the general $G$-orbit on $\widetilde{T^*X}$ and $\widetilde{T^*X}^{P,Y}$ is isomorphic over the algebraic closure to $B_{L_1}\backslash G$, where $B_{L_1}$ is a Borel subgroup of $L_1$.
Similarly, the general $L$-orbit on $T^*Y_2$ is isomorphic to $T_1\backslash L$, where $T_1$ is (non-canonically) isomorphic to $\ker(A\to A_X)$ over the algebraic closure; notice that $P(Y_2)$ is the class of Borel subgroups of $L$. If we consider $T^*Y_2$ as a $P$-space, we have to include $U_P$, which acts trivially, in the point stabilizers.
Hence, the relative dimension of the map $T_Y^*X^{{\mathfrak{u}}_P^\perp} \to T^*Y_2$, by counting dimensions of stabilizers in $G$, is $$\dim U_P - \dim U_{L(X)} = \dim U_P - \dim {\mathcal{B}}_{L(X)}.$$ The analogous relation for the corresponding vector spaces, by counting dimensions of stabilizers in $H$, is $$\label{seconddim} \dim V_P - \dim V_2 = \dim U_{P_H} - \dim {\mathcal{B}}_{L(X)} = \dim (P_H\backslash H) - \dim {\mathcal{B}}_{L(X)}.$$
Combining and , we obtain .
Since the kernel of the map $V_P\to V_2$ is isotropic (it maps to $0\in {\mathfrak{c}}_X^*$ by the commutativity of ), the quadratic space $V$ is maximally isotropic if and only if $V_2$ is.
We are now ready to prove Proposition \[rationality\].
Suppose that the stabilizer $H$ of a point on $X$ is split. Since $H$ is connected (Proposition \[correctrepresentative\]), a Borel subgroup of $H$ is split, solvable, connected, and is contained in a Borel subgroup $B$ of $G$ (over $F$). Then, the $H$-orbit represented by $1$ on $G/B$ is projective, hence closed, and the corresponding $B$-orbit on $X=H\backslash G$ is defined over $F$.
Vice versa, if a closed $B_{\bar F}$-orbit on $X_{\bar F}$ is defined over $F$ — equivalently, has a point with stabilizer $H$, then $B\cap H$ is a split Borel subgroup of $H$, hence $H$ is split.
Now let $(Y,P)$ be a basic orbit-parabolic pair, as in Lemma \[lemmaclosedorbits\]. By Lemma \[lemmaY2\], the variety $Y_2 = Y/U_P$ is isomorphic to $T\backslash {\operatorname{PGL}}_2$ for some torus $T$ or to ${\operatorname{SL}}_2$, hence the closed $B$-orbit(s) on $Y_2$, and on $Y$, are defined over $F$ if and only if we are not in the case of $T\backslash {\operatorname{PGL}}_2$ with $T$ non-split. This is equivalent to each, equivalently one, fiber of $T^*Y_2\to {\mathfrak{c}}_X^*$ being isotropic, which by Proposition \[propVP\] is equivalent to each, equivalently one, fiber of $T^*X \to {\mathfrak{c}}_X^*$ over a point in $Y$ being maximally isotropic. The surjectivity of the map $\widetilde{T^*X}^{P,Y}\to T^*X$ on $F$-points, by the same proposition, implies that the same is true for every fiber of $T^*X$ hence, applying this argument in the reverse direction, one closed $B_{\bar F}$-orbit being defined over $F$ implies that every closed $B_{\bar F}$-orbit is defined over $F$.
Recall that we have been assuming that $G$ is split, and that $X$ is a “correct representative” of its class modulo $G$-isomorphisms; the statement is not true for varieties such as $\mathcal N(T)\backslash {\operatorname{PGL}}_2$.
\[typeGalwayssplit\] For spherical roots of type $G$, all $B$-orbits are of rank one (maximal), hence defined over $F$ by Lemma \[maxrankdefined\]. Therefore, by the proposition, in that case all stabilizers are split.
Finally, a result on the forms of a given variety $H\backslash G$ with $G$ and $H$ split:
\[uniqueclass?\] If $X_{\bar F}$ is the correct representative of a class of rank one affine homogeneous spherical $G_{\bar F}$-varieties (with $W_X={\mathbb{Z}}/2$) over $\bar F$, then a form of $X$ as a $G$-variety over $F$, where $G$ denotes the split form of $G_{\bar F}$, always exists. Moreover, if the spherical root is of type $T$, there is a unique such form with split stabilizers, and if the spherical root is of type $G$ the isomorphism classes of such forms are naturally a torsor for the group $F^\times/(F^\times)^2$, and stabilizers are always split.
The existence of a model over $F$ follows from the general Theorem 0.2 of [@Borovoi]. To apply it, we first replace the spherical subgroup $H_{\bar F}$ (over the algebraic closure) by its *spherical closure* $\bar H_{\bar F}\subset \mathcal N(H_{\bar F})$; I refer the reader to the aforementioned reference for the definition of spherical closure. By the theorem, the resulting variety $\bar X_{\bar F} = \bar H_{\bar F}\backslash G_{\bar F}$ has a form $\bar X$ as a $G$-variety over $F$. As in the proof of Proposition \[correctrepresentative\], if we replace $G$ by the simply connected cover of its derived group and $\bar H$ by the identity component of its preimage in this simply connected cover, we will obtain a variety $X$ of rank one whose weight lattice is “as large as possible”, that is, spanned by the spherical root $\gamma$ if that is of type $T$, and by $\frac{\gamma}{2}$ if it is of type $G$; thus, $X$ has to be a form of $X_{\bar F}$ (and, in particular, the action of the simply connected cover factors through $G$).
Given such a form $X$, the set of $G$-forms of $X$ is parametrized by the first Galois cohomology group $H^1(F,{{\operatorname{Aut}}}^G(X))$.
If the spherical root is of type $G$, then, by Remark \[typeGalwayssplit\], stabilizers are always split. Moreover, by Proposition \[correctrepresentative\], ${{\operatorname{Aut}}}^G(X)={\mathbb{Z}}/2$, so the forms are a torsor for $H^1(F,{{\operatorname{Aut}}}^G(X)) = F^\times/(F^\times)^2$.
If the spherical root is of type $T$, then there is a Borel orbit (over the algebraic closure) of rank zero, and therefore the rank of $H_{\bar F}$ is equal to the rank of $G$, where $H_{\bar F}$ denotes the stabilizer of any point on $X_{\bar F}$. Let $T_{\bar F}\subset H_{\bar F}$ be a maximal torus. The Lie algebra ${\mathfrak{g}}_{\bar F}$ splits into a direct sum of $T_{\bar F}$-eigenspaces, and the subalgebra ${\mathfrak{h}}_{\bar F}$ is a subsum of that. If, now, $G$ is defined and split over $F$, we may assume that $T_{\bar F}$ is the extension to $\bar F$ of a maximal split torus $T\subset G$, hence the eigenspaces are defined over $F$, and the subalgebra ${\mathfrak{h}}_{\bar F}$ is the extension to $\bar F$ of a subalgebra over $F$. In other words, there is a form of $X$ such the stabilizer of a point (hence every point, by Proposition \[rationality\]) is split.
Assume that $X$ is such a form. Let $(Y,P_\alpha)$ be a basic orbit-parabolic pair, so that $Y$ contains a $B$-orbit of rank one, and two closed $B$-orbits of rank zero. These closed $B$-orbits now are defined over $F$, by Proposition \[rationality\]. Arguing as in the proof of Proposition \[correctrepresentative\], we get an injection $${{\operatorname{Aut}}}^G(X) \hookrightarrow {{\operatorname{Aut}}}^{L_\alpha}(Y/U_{P_\alpha}),$$ where $L_\alpha$ is the Levi quotient of $P_\alpha$. The weight lattice of $Y/U_{P_\alpha}$ is spanned by $\alpha$, hence the $L_\alpha$-variety $Y/U_{P_\alpha}$ is isomorphic to the quotient of $L_\alpha$ by a maximal torus $T$, which now has to be split. The group ${{\operatorname{Aut}}}^{L_\alpha}(Y/U_{P_\alpha})$ is isomorphic to ${\mathbb{Z}}/2$, with the nontrivial automorphism interchanging the two closed Borel orbits. Hence, if $\sigma$ is a nontrivial $G$-automorphism of $X$, it interchanges the two closed $B$-orbits in $Y$. As a result, for any nontrivial element of $H^1(F,{{\operatorname{Aut}}}^G(X))$, defining a form $X'$ of $X$, these $B$-orbits are not defined over $F$ in $X'$, and, by Proposition \[rationality\] again, stabilizers of points on $X'$ are not split. Therefore, $X$ is the unique form with split stabilizers.
Consider the $G={{\operatorname{SO}}}_4=({\operatorname{SL}}_2\times{\operatorname{SL}}_2)/\{\pm 1\}^{{\operatorname{diag}}}$-action on ${\operatorname{GL}}_2$. All orbits are isomorphic to $H\backslash G={{\operatorname{SO}}}_3\backslash{{\operatorname{SO}}}_4={\operatorname{PGL}}_2\backslash {{\operatorname{SO}}}_4$, but the $G(F)$-conjugacy class of the embedding of $H$ in $G$ depends on the square class of the determinant.
Structure and resolution of $X\times X/G$ {#sec:resolution}
=========================================
In this section we study the diagonal action of $G$ on $X\times X$.
In the group case $X=H$, $G=H\times H$, the invariant theoretic quotient $X\times X\sslash G$ coincides with the invariant theoretic quotient of $H$ by $H$-conjugacy, and is naturally identified with $A_H\sslash W_H$, by the Chevalley isomorphism. Moreover, the quotient map $H\to A_H\sslash W_H$ is smooth over the points where the quotient map $A_H\to A_H\sslash W_H$ is smooth.
A Chevalley isomorphism for $X\times X\sslash G = H\backslash G\sslash H$ was proven by Richardson [@Richardson] in the case where $X=H\backslash G$ is a symmetric variety. Almost all spherical varieties of Table are symmetric, but this is not the case for the examples denoted by $G_2$ and $B_3''$. In any case, we need a more general and conceptual way to analyze this quotient, that will be useful in proving various results that we need (and, hopefully, useful in view of higher rank cases, where only a small fraction of spherical varieties are symmetric).
This general approach will be provided to us by Knop’s theory of the cotangent bundle of $X$ and the invariant collective motion. It will turn out, in rank one, that even for the non-symmetric cases the quotient $X\times X\sslash G$ is (canonically) isomorphic to $A_X\sslash W_X$, with singularities of the quotient map $X\times X\to A_X\sslash W_X$ only over the singularities of $A_X\to A_X\sslash W_X$; this fact does not generalize to higher rank, although the constructions of Knop do.
By the end of this section, we will have constructed a resolution of the $G$-space $X\times X$, denoted $${\mathcal{R}}: {\mathbbm{P}}J_X \to X\times X.$$ As mentioned in the introduction, the term “resolution” refers here to the fibers of the map $X\times X\to X\times X\sslash G$, which under the resolution become normal crossings divisors.
Knop’s abelian group scheme; the $[\pm 1]$ and nilpotent divisors {#ssKnopsgroupscheme}
-----------------------------------------------------------------
Consider $A_X\times {\mathfrak{a}}_X^*$ as a constant group scheme over ${\mathfrak{a}}_X^*$, with the simultaneous action of $W_X$ on $A_X$ and on ${\mathfrak{a}}_X^*$. Let $$J = ({\mathrm{Res}}_{{\mathfrak{a}}_X^*/{\mathfrak{c}}_X^*} (A_X\times {\mathfrak{a}}_X^*))^{W_X},$$ where ${\mathrm{Res}}_{{\mathfrak{a}}_X^*/{\mathfrak{c}}_X^*}$ denotes Weil’s restriction of scalars from ${\mathfrak{a}}_X^*$ to ${\mathfrak{c}}_X^*$. It is a group scheme over ${\mathfrak{c}}_X^*$, with a canonical birational morphism $$\label{birational} J\to (A_X\times {\mathfrak{a}}_X^*)\sslash W_X,$$ which is an isomorphism over $\mathring{{\mathfrak{c}}}_X^* = {\mathfrak{c}}_X^*\smallsetminus\{0\}$. Since we are in rank one with $W_X={\mathbb{Z}}/2$, we can identify ${\mathfrak{c}}_X^*=\mathbbm A^1$, with a coordinate $\xi$ such that $\xi=0$ corresponds to $0\in{\mathfrak{a}}_X^*$. Then $$\label{Jcoordinates} J \simeq {\operatorname{Spec}}F[t_0,t_1,\xi]/(t_0^2-\xi t_1^2 - 1).$$ This group scheme is also familiar as the group scheme of regular centralizers over the Kostant section of the Lie algebra $\mathfrak{sl}_2$, under the adjoint action of ${\operatorname{SL}}_2$.
We have a canonical identification $$\label{Jbc} J\times_{{\mathfrak{c}}_X^*} \mathring{{\mathfrak{a}}}_X^* \simeq A_X \times \mathring{{\mathfrak{a}}}_X^*,$$ compatible with pullback of sections from ${\mathfrak{c}}_X^*$ to ${\mathfrak{a}}_X^*$, and the identification of sections of $J$ with $W_X$-equivariant sections of $A_X\times {\mathfrak{a}}_X^*$.
On the other hand, the fiber of $J$ over $0\in {\mathfrak{c}}_X^*$ is isomorphic to ${\mathbbm{G}_a}\times {\mathbb{Z}}/2$. We let $J^0\subset J$ denote the open group subscheme whose fiber over any point of ${\mathfrak{c}}_X^*$ is the connected component of the fiber of $J$. The group scheme $J$ has a canonical action of ${\mathbbm{G}_m}$, induced from the action of ${\mathbbm{G}_m}$ on $A_X \times {\mathfrak{a}}_X^*$ (on the second factor). In the coordinates above: $$a\cdot (\xi, t_0, t_1) = (a^2\xi, t_0, a^{-1} t_1).$$
In what follows, for any variety $Y$ equipped with a map to ${\mathfrak{c}}_X^*$ we will denote $$J\bullet Y:= J\times_{{\mathfrak{c}}_X^*} Y$$ (and similarly for $J^0$).
We can distinguish three divisors on $J\bullet Y$, two of them to be denoted $[\pm 1]_Y$, and another to be denoted $(J\bullet Y)^{{\operatorname{nilp}}}$. The first two are the images of $(\pm 1) \cdot Y$ (where $(\pm 1)$ are understood as $W_X$-invariant sections of $A_X$ over ${\mathfrak{a}}_X^*$, hence as sections of $J$ over ${\mathfrak{c}}_X^*$), and the third is the preimage of $0\in {\mathfrak{c}}_X^*$. In the coordinates used above, the sum of the divisors $[\pm 1]_Y$ is given by the equation $t_1=0$ (with the value of $t_0$ distinguishing the irreducible components), and $(J\bullet Y)^{{\operatorname{nilp}}}$ is given by $\xi =0$. Notice that the union of these three divisors is precisely the preimage of the corresponding points $[\pm 1]\in A_X\sslash W_X$ under the map $J\bullet Y\to A_X\sslash W_X$ descending from ; in the coordinates above, this union corresponds to the equations $t_0=\pm 1$.
Notice that $J\bullet Y$ is smooth over $Y$ (because it is obtained by base change from the smooth group scheme $J\to {\mathfrak{c}}_X^*$), and the divisors $[\pm 1]_Y$ are isomorphic to $Y$. Hence, if $Y$ is smooth, this is also the case for these divisors and the scheme $J\bullet Y$. On the other hand, the nilpotent divisor $(J\bullet Y)^{{\operatorname{nilp}}}$ is smooth on its intersection with the smooth locus of the morphism $Y\to {\mathfrak{c}}_X^*$.
\[lemmatransverse\] Assume that $Y\to {\mathfrak{c}}_X^*$ is smooth. Then, the divisors $[\pm 1]_Y$ and $(J\bullet Y)^{{\operatorname{nilp}}}$ intersect transversely, and the morphism $J\bullet Y \to A_X\sslash W_X$ is smooth away from $[\pm 1]_{Y}$.
All these properties are stable under smooth base change, and since the morphism $Y\to {\mathfrak{c}}_X^*$ is smooth, the problem reduces to the case $Y={\mathfrak{c}}_X^*$, $J\bullet Y=J$.
It is then immediate to check from the equation $t_0^2-\xi t_1^2 = 1$ that the intersections of the divisors $\xi=0$ and $t_1=0$ are transverse.
For the second statement, we note that we can identify $A_X\sslash W_X\simeq \mathbbm A^1$ so that the map $J\to A_X\sslash W_X$ corresponds to the coordinate $t_0$. The cotangent space of $J$ is generated by $dt_0, dt_1, d\xi$ subject to the equation $$2t_0 dt_0 -2\xi t_1 dt_1 -t_1^2 d\xi =0.$$ Thus, $dt_0 = 0$ only when $t_1=0$.
We now specialize to the scheme $Y=T^*X$, endowed with the invariant moment map $\mu_{{\operatorname{inv}}}: T^*X\to {\mathfrak{c}}_X^*$. Lemma \[nondegenerate\] implies that this map is smooth away from the zero section (whose complement we will denote by $T^*X_{\ne 0}$), hence, by \[lemmatransverse\], we get:
\[corollarydivisors\] The divisors $[\pm 1]_{T^*X_{\ne 0}}$ and $(J\bullet T^*X_{\ne 0})^{{\operatorname{nilp}}}$ in $J\bullet T^*X_{\ne 0}$ intersect transversely, and the morphism $J\bullet T^*X_{\ne 0} \to A_X\sslash W_X$ is smooth away from $[\pm 1]_{T^*X_{\ne 0}}$.
Integration of the invariant collective motion
----------------------------------------------
The relative Lie algebra ${\operatorname{Lie}}(J)$ of $J$ over ${\mathfrak{c}}_X^*$ can be canonically identified with the cotangent space of ${\mathfrak{c}}_X^*$. Indeed, sections of both are canonically identified with $W_X$-invariant sections of the cotangent bundle of ${\mathfrak{a}}_X^*$. Thus, a section of the cotangent bundle of ${\mathfrak{c}}_X^*$ can be viewed as a section of ${\operatorname{Lie}}(J)$, and at the same time induces a Hamiltonian vector field on $T^*X$, by pullback and the symplectic structure. Knop has shown [@KnAut Theorem 4.1] that there is an action over ${\mathfrak{g}}_X^*$ $$\label{actionJ0T}\xymatrix{\mathcal C: & J^0 \bullet T^*X \ar[dr] \ar[rr] && T^*X \ar[dl]\\ && {\mathfrak{g}}_X^*}$$ that integrates this vector field. Over $\mathring{\mathfrak{c}}_X^*={\mathfrak{c}}_X^*\smallsetminus\{0\}$, this action lifts through the isomorphism to the canonical action of $A_X$ that we discussed on the regular semisimple part of the polarization. In particular, on the regular semisimple part this action is induced from the action of the centralizers of coadjoint vectors: if $v\in T^*X^{{\operatorname{rs}}}$ with image $\mu(v)=Z\in {\mathfrak{g}}^*$, the centralizer of $Z$ is a twisted Levi $L$ (conjugate over the algebraic closure to a Levi of $P(X)$), and it acts on $v$ through a quotient which, over the algebraic closure, is isomorphic (up to the $W_X$-action) to $A_X$.
This action may, but does not always, extend to $J$, as the following examples show:
Let $X={\operatorname{SL}}_2$ under the $G={{\operatorname{SO}}}_4={\operatorname{SL}}_2\times {\operatorname{SL}}_2/\{\pm 1\}^{{\operatorname{diag}}}$-action. Then ${\mathfrak{c}}_X^*=\mathfrak{sl}_2\sslash {\operatorname{SL}}_2$ (under the adjoint action), and $J$ can be identified with the group scheme of regular centralizers over ${\mathfrak{c}}_X^*$, i.e., the group scheme of centralizers in ${\operatorname{SL}}_2$ over a Kostant section $${\mathfrak{c}}_X^*\hookrightarrow \mathfrak{sl}_2^{\rm reg}.$$ Hence, it acts (faithfully) on $(T^*{\operatorname{SL}}_2)_{\ne 0}$, by either left or right translation; it is easy to see that this extends to the trivial action on the zero section. In particular, the action of $J^0$ extends to $J$.
Let $X={\mathbbm{G}_m}\backslash {\operatorname{PGL}}_2$. The group scheme of regular centralizers for ${\operatorname{PGL}}_2$ again acts faithfully on $T^*X$, but in this case it is isomorphic to $J^0$. However, one can easily see that the entire group scheme $J$ acts, with the action of the $(-1)$-section induced from the nontrivial $G$-automorphism of $X$.
\[pgl3\] Let $X=H\backslash G={\operatorname{GL}}_2\backslash {\operatorname{PGL}}_3$, the variety of direct sum decompositions ${\mathbbm{G}_a}^3=V_2\oplus V_1$ of a based three-dimensional vector space into the sum of a two- and a one-dimensional subspace. The fiber $\mathfrak h^\perp$ of $T^*X$ over the point $x_0$ corresponding to the decomposition $\left<e_1, e_2\right> \oplus \left< e_3\right>$ is isomorphic to ${{\operatorname{Std}}}\oplus {{\operatorname{Std}}}^*$, the direct sum of the standard representation and its dual, as a (right) representation of $H={\operatorname{GL}}_2$. The quadratic form of the invariant moment map is $$(v, v^*)\mapsto \left<v, v^*\right>.$$ Hence, $T_{x_0}^*X^{{\operatorname{nilp}}}=$ the variety of mutually orthogonal pairs $(v, v^*)$.
Represent $H$ as the upper left copy of ${\operatorname{GL}}_2$ in ${\operatorname{PGL}}_3$, identify ${\mathfrak{g}}^*= {\mathfrak{g}}= \mathfrak{sl}_3$ through the trace pairing, and consider representatives $$Z_{\epsilon, y}= \begin{pmatrix}
&& y \\ && \epsilon^2 \\ 0 & 1 & 0
\end{pmatrix},$$ with $\epsilon\ne 0$, for the split regular semisimple $H$-orbits on $\mathfrak h^\perp$ (i.e., those over the image of $\mathring{\mathfrak{a}}_X^*(F)\to {\mathfrak{c}}_X^*(F)$). The variable $y$ is redundant, at the moment, but will play a role as $\epsilon \to 0$. The centralizer of $Z_{\epsilon, y}$ under the right coadjoint representation of $G$ is the torus of matrices (modulo center) of the form $$\begin{pmatrix}
a\\
\frac{(-2a+b+c)y}{2\epsilon^2} & \frac{b+c}{2} & \frac{b-c}{2 \epsilon} \\
\frac{(b-c)y}{2\epsilon} & \frac{(b-c)\epsilon}{2} & \frac{b+c}{2}
\end{pmatrix}.$$ It acts on the point $x_0\in X$ through the quotient $(a,b,c)\mapsto \frac{b}{c}$, which is isomorphic to $A_X\simeq {\mathbbm{G}_m}$ (up to inversion).
Let us examine whether the action of $-1\in A_X$ on $T^*X^{{\operatorname{rs}}}$ extends to the nilpotent limit $\epsilon\to 0$. Representing $-1$ by a matrix as above, corresponding to $(a,b,c)=(1,1,-1)$: $$g_{\epsilon, y}= \begin{pmatrix}
1\\
\frac{-y}{\epsilon^2} & 0 & \frac{1}{ \epsilon} \\
\frac{y}{\epsilon} & \epsilon & 0
\end{pmatrix} \sim \begin{pmatrix}
1\\
-y & 0 & \epsilon \\
\frac{y}{\epsilon^2} & 1 & 0
\end{pmatrix}$$ (where $\sim$ means same left $H$-coset), we easily see that $x_0\cdot g_{\epsilon, y}$ does not have a limit in $X$ as $\epsilon \to 0$, unless $y=0$. On the other hand, if $y$ also tends to zero in such a way that $\frac{y}{\epsilon^2}$ has a limit, the action extends to the limit. Geometrically, this means that if we blow up $\mathfrak h^\perp= {{\operatorname{Std}}}\oplus {{\operatorname{Std}}}^*$ over the divisor ${{\operatorname{Std}}}\cup {{\operatorname{Std}}}^*$, and remove the strict transform of the nilpotent divisor, then the action of $-1$ extends to the blowup — this will be relevant when discussing the “second” singular orbit on $X\times X/G$, below, see Example \[pgl3-second\].
\[torsor\] For $X$ of rank one with $W_X={\mathbb{Z}}/2$, the restriction of $T^*X\to {\mathfrak{g}}_X^*$ over ${\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$ (that is, over $\mathring{\mathfrak{c}}_X^*$ under the invariant moment map) is a $J\bullet {\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$-torsor.
This is a consequence of Corollary \[polarizedtorsor\]: Since $\widehat{T^*X}^{{\operatorname{rs}}}$ is an $A_X$-torsor over $\hat{\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$, and the action is the base change of the action of the group scheme $J\bullet {\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$ on $T^*X^{{\operatorname{rs}}}$, the latter is also a torsor.
Resolution in a neighborhood of the diagonal
--------------------------------------------
We denote by $N^*_A B$ the conormal bundle in a smooth variety $B$ of a subvariety $A$. In particular, for $B=X\times X$, we will just be using the notation $N^*_A$ for $N^*_A B$; we denote by $N_A$ the normal bundle.
The fiber product $J\bullet T^*X$ carries a natural ${\mathbbm{G}_m}$-action, induced from the action on $T^*X$ and on $J$. The quotient ${\mathbbm{G}_m}\backslash (J\bullet T^*X_{\ne 0})$ will be denoted by the symbol of projectivization, ${\mathbbm{P}}(J\bullet T^*X)$. We use analogous notation for all similar spaces with a ${\mathbbm{G}_m}$-action.
Consider the combination of the projection and action maps: $$\label{Jtoconormal}
J^0\bullet T^*X \to T^*X \times_{{\mathfrak{g}}^*} T^*X.$$
An immediate corollary of Lemma \[torsor\] is:
\[cortorsor\] The map is an isomorphism over $\mathring{\mathfrak{c}}_X^*$ (i.e., on the sets of regular semisimple vectors).
By Lemma \[nondegenerate\], over $\mathring{\mathfrak{c}}_X^*$ we have $T^*X^{{\operatorname{rs}}}\times_{{\mathfrak{g}}^*} T^*X^{{\operatorname{rs}}}= T^*X^{{\operatorname{rs}}}\times_{{\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}} T^*X^{{\operatorname{rs}}}$, and by Lemma \[torsor\] the right hand side is isomorphic to $J^0\bullet T^*X^{{\operatorname{rs}}}= J\bullet T^*X^{{\operatorname{rs}}}$.
The space on the right hand side of can be thought of as the union of all conormal bundles to all $G$-orbits on $X\times X$. Strictly speaking, we should take the fiber product of $T^*X$ with itself over the *anti-diagonal* copy $\{(Z,-Z)\}$ of ${\mathfrak{g}}^*$, but it is more convenient, notationally, to multiply the second variable by $-1$ and work over the diagonal copy. In any case, this union of conormal bundles comes with its own map to ${\mathfrak{g}}^*$, which is not the moment map for the $G^{{\operatorname{diag}}}$-action on $X\times X$ (that one is trivial), but “remembers” the fact that $X\times X$ had a $G\times G$-action. This is the microlocal analog of the spectral decomposition of the relative trace formula.
The rough idea behind the resolution of $X\times X$ that we are about to construct is that, generically, the projectivization of this union of conormal bundles is isomorphic to $X\times X$ (because we are in rank one, and generic $G$-orbits will be of codimension one), hence the projectivization of $T^*X \times_{{\mathfrak{g}}^*} T^*X$ is, roughly, a “resolution” of $X\times X$. However, this space is quite singular, and we will use an extension of the space on the left as a smooth replacement.
\[arounddiagonal\] Consider the map $${\mathcal{R}}: {\mathbbm{P}}(J^0\bullet T^*X) \to X\times X,$$ descending from . We regard ${\mathbbm{P}}T^*X$ as a divisor in ${\mathbbm{P}}(J^0\bullet T^*X)$, descending from the divisor $[1]_{T^*X}$ (see §\[ssKnopsgroupscheme\]).
The map ${\mathcal{R}}$ factors through the blowup $\widetilde{X\times X}^1\to X\times X$ at the diagonal $X_1:=X^{{\operatorname{diag}}}$, and is an isomorphism from a $G$-stable neighborhood of ${\mathbbm{P}}T^*X$ to a neighborhood of the exceptional divisor in $\widetilde{X\times X}^1$.
By the universal property of blowups, the map ${\mathcal{R}}$ factors through a morphism to the blowup $$\label{toblowup}
\tilde{\mathcal{R}}: {\mathbbm{P}}J_X\to \widetilde{X\times X}^1,$$ sending the divisor ${\mathbbm{P}}T^*X$ to the exceptional divisor of the blowup, which is isomorphic to ${\mathbbm{P}}TX$.
To show that this map is an isomorphism in a $G$-stable neighborhood of the divisor ${\mathbbm{P}}T^*X$, it is enough to show that the induced map $d\tilde{\mathcal{R}}$ from the normal bundle of the divisor ${\mathbbm{P}}T^*X$ to the normal bundle of the exceptional divisor is an isomorphism.
The normal bundle of ${\mathbbm{P}}T^*X$ can be identified with ${\mathbbm{P}}({\operatorname{Lie}}J \bullet T^*X)$, and the normal bundle of the exceptional divisor can be identified with the blowup of the tangent bundle $TX$ at the zero section, i.e., of the normal bundle to $X_1 = X^{{\operatorname{diag}}}$. The map $d\tilde{\mathcal{R}}$ is lifted from the analogous map $$\label{mapLie} d{\mathcal{R}}: {\mathbbm{P}}({\operatorname{Lie}}J \bullet T^*X_{\ne 0}) \to TX= N_{X_1},$$ the partial differential of the map ${\mathcal{R}}$. We compute this map:
Recall that ${\operatorname{Lie}}J$ is canonically isomorphic to the cotangent space of ${\mathfrak{c}}_X^*$. A section $\sigma$ of ${\operatorname{Lie}}J$ gives, by pullback of differential forms via the invariant moment map, a section $\mu_{{\operatorname{inv}}}^*\sigma$ of the cotangent bundle of $T^*X$, hence a vector field $v_\sigma$ on $T^*X$, by the symplectic structure. If $\pi: T^*X\to X$ denotes the canonical projection, and $\pi^*(TX)$ is the pullback of the tangent bundle, we have a canonical projection of vector bundles on $T^*X$ $${{\operatorname{pr}}}: T(T^*X)\to \pi^*(TX),$$ corresponding to the projection of vector fields on $T^*X$ “to the $X$-direction”.
Let $T^*_X (T^*X) = T^*(T^*X)/\pi^*(T^*X)$ denote the relative cotangent bundle of $T^*X$ over $X$; we similarly have a projection of vector bundles $${{\operatorname{pr}}}': T^*(T^*X) \to T^*_X(T^*X).$$ Moreover, we have canonical identifications $$T^*_X (T^*X) \xrightarrow\sim T^*X\times_X TX \overset{\sim}\leftarrow \pi^*(TX),$$ and the image ${{\operatorname{pr}}}(v_\sigma)$ of the aforementioned vector field, as a section of $\pi^*(TX)$, coincides under this identification with “the restriction of $\mu_{{\operatorname{inv}}}^*\sigma$ to the fiber direction”, that is, with the section ${{\operatorname{pr}}}'(\mu_{{\operatorname{inv}}}^*\sigma)$ of $T^*_X(T^*X)$.
Consider, for example, an identification ${\mathfrak{c}}_X^*\xrightarrow{\xi}\mathbbm A^1$ (with the point “zero” preserved), and let $\sigma=d\xi$. Then, $\xi\circ\mu_{{\operatorname{inv}}}$ can be viewed as a quadratic form on $T^*X$, and it gives rise to a map $$\iota_\xi:T^*X\to TX$$ over $X$. As we have seen in Lemma \[nondegenerate\], the quadratic form is nondegenerate; hence, $\iota_\xi$ is an isomorphism.
The differential of the quadratic form, restricted to each fiber, is the graph of $\iota_\xi$, considered as a subset of $T^*X\times_X TX = \pi^*(TX)$. Therefore, the section of ${\operatorname{Lie}}J$ corresponding to $d\xi$ defines a vector field on $T^*X$, whose projection to the $X$-direction is the graph of $\iota_\xi$. Hence, the map $d{\mathcal{R}}$ descends from $$v^*\in T^*X_{\ne 0} \mapsto (d\xi, v^*) \in {\operatorname{Lie}}J \bullet T^*X_{\ne 0} \mapsto \iota_\xi(v^*) \in TX.$$
In particular, the map $d{\mathcal{R}}$ is fiberwise an isomorphism, hence its lift $d{\mathcal{R}}$ to the blowup is an isomorphism, and $\tilde R$ is an isomorphism from a $G$-stable neighborhood of the divisor ${\mathbbm{P}}J_X$ to a $G$-stable neighborhood of the exceptional divisor.
Notice that, under the isomorphism ${\operatorname{Lie}}J = T^*{\mathfrak{c}}_X^*$, the action of ${\mathbbm{G}_m}$ on ${\operatorname{Lie}}J$ is the following: it acts in the canonical way on the base ($a\in {\mathbbm{G}_m}$ acts by multiplying a coordinate $\xi$ on ${\mathfrak{c}}_X^*$ by $a^2$); this induces an inverse pullback isomorphism: $$(a^*)^{-1}: T^*{\mathfrak{c}}_X^* \to T^*{\mathfrak{c}}_X^*,$$ and we multiply this by $a$, fiberwise. In terms of the coordinates $t_0^2-\xi t_1^2=1$, the action of ${\mathbbm{G}_m}$ on $J$ is given by $a\cdot (\xi, t_0, t_1) = (a^2\xi, t_0, a^{-1} t_1)$. One can now directly see that the map ${\operatorname{Lie}}J \bullet T^*X_{\ne 0} \to TX$ that was described above is, indeed, ${\mathbbm{G}_m}$-equivariant.
The proposition implies:
\[corregulardense\] There is an open, dense, $G$-stable subset $(X\times X)^\circ \subset X\times X$ on which every $G$-orbit is of codimension one, and the nonzero vectors of its conormal bundle are regular semisimple.
Indeed, consider the space $T^*X\times_{{\mathfrak{g}}^*} T^*X$ as the union of the conormal bundles of all $G$-orbits on $X\times X$; the map from $J\bullet T^*X^{{\operatorname{rs}}}$ is an isomorphism onto the regular semisimple subset $T^*X^{{\operatorname{rs}}}\times_{{\mathfrak{g}}^{*,{{\operatorname{rs}}}}} T^*X^{{\operatorname{rs}}}$ by Corollary \[cortorsor\], and the space $J\bullet T^*X^{{\operatorname{rs}}}$ is a union of ${\mathbbm{G}_m}\times G$-orbits of codimension one. By Proposition \[arounddiagonal\], a dense open $G$-stable subset of its projectivization is isomorphic to an open subset of a $G$-stable neighborhood of the diagonal $X_1=X^{{\operatorname{diag}}}$.
Closed orbits and invariant-theoretic quotients
-----------------------------------------------
The following lemma will be very basic in our analysis of the space $X\times X$:
\[lemmaregular\] Every $G$-orbit in the image of the regular semisimple set ${\mathbbm{P}}(J^0\bullet T^*X^{{\operatorname{rs}}})$ under the map ${\mathcal{R}}$ is closed.
In particular, non-closed $G$-orbits on $X\times X$ do not contain regular semisimple vectors in their conormal bundles.
The second statement follows from the first, by Corollary \[cortorsor\].
To prove the first, we need some preliminary results. Choose a basic orbit-parabolic pair $(Y,P)$ as in Lemma \[lemmaclosedorbits\]. The subvariety $Y\times Y\subset X\times X$ is closed, hence the map $(Y\times Y)\times^P G\to X\times X$ is proper.
Consider the space $\widetilde{T^*X}^{P,Y} = T_Y^*X^{{\mathfrak{u}}_P^\perp} \times^P G$, defined in §\[ssrationality\]. By Lemma \[partialcover\], it surjects onto $T^*X$, hence $J^0\bullet \widetilde{T^*X}^{P,Y}$ surjects onto $J^0\bullet T^*X$ (and same for the subsets of regular semisimple vectors). But the latter has dense image in $X\times X$, by Proposition \[arounddiagonal\], hence so does the former. The subspace $J^0\bullet T_Y^*X^{{\mathfrak{u}}_P^\perp}$ maps to $Y\times Y$, and we have a commutative diagram: $$\label{Ydiagram} \xymatrix{
J^0\bullet \widetilde{T^*X}^{P,Y} \ar[r] \ar[d]^{{\mathcal{R}}_Y} & J^0\bullet T^*X \ar[d] \\
\tilde Y\times^P G \ar[r] & X\times X,}$$ where $\tilde Y\subset Y\times Y$ denotes the closure of the image of $J^0\bullet T_Y^*X^{{\mathfrak{u}}_P^\perp}$ in $Y\times Y$ under the map to $Y\times Y$. Hence, the map $$\tilde Y\times^P G\to X\times X$$ is surjective and proper. It is thus enough to show that every $P$-orbit in the image of $J\bullet T_Y^*X^{{\mathfrak{u}}_P^\perp,{{\operatorname{rs}}}}$ in $\tilde Y$ is closed.
Let $H\subset G$ be the stabilizer of a point on $Y$, and $H_P=H\cap P$. Consider the quotient $Y\to Y_2 = Y/U_P$, and remember that $Y_2$ is isomorphic to $T\backslash {\operatorname{PGL}}_2$ or ${\operatorname{SL}}_2$ (Lemma \[lemmaY2\]), with the connected center of the Levi quotient of $P$ acting trivially; in particular, stabilizers of points in $Y_2$ are reductive. Thus, we can fix a Levi decomposition $H_P = H_L \cdot H_U$, and a Levi subgroup $L\subset P$ which contains $H_L$. Notice that $Y_2 = H_L\backslash L$, hence we can also consider $Y_2$ as a *subvariety* of $Y$ (depending on the choices that we have made).
Use an invariant bilinear form on ${\mathfrak{g}}$ to identify ${\mathfrak{u}}_P^\perp = \mathfrak p$. By semisimplicity, the image of every element of $T_Y^*X^{{\mathfrak{u}}_P^\perp,{{\operatorname{rs}}}}$ under the moment map is $U_P$-conjugate to an element of $\mathfrak l$. Thus, it is enough to show that the $P$-orbit of ${\mathcal{R}}_Y(a\cdot v)$ is closed, for any $v\in T_Y^*X^{{\mathfrak{u}}_P^\perp,{{\operatorname{rs}}}}$ with $\mu(v)\in \mathfrak l$, and any $a\in J$ over the image of $v$ in ${\mathfrak{c}}_X^*$.
Recall that the $J$-action on $v$ is induced from the action of the centralizer of $\mu(v)$ in $G$; since this acts by a one-dimensional quotient, the same action is induced from the centralizer of $\mu(v)$ in $L$, which is a torus acting nontrivially. Hence, given $a\in J$ over $\mu_{{\operatorname{inv}}}(v)$, there is an $l\in L$ with $a\cdot v = v\cdot l$. Hence, considering $Y_2$ as a subvariety of $Y$, ${\mathcal{R}}_Y(a\cdot v)\in Y_2\times Y_2$. To avoid confusion, we will be denoting by $y$ a point in $Y_2\times Y_2$ considered as a subset of $Y\times Y$, and by $\bar y$ its image in $Y_2\times Y_2$ considered as a quotient of $Y\times Y$. It is immediate to confirm that the $L^{{\operatorname{diag}}}$-orbit of ${\mathcal{R}}_Y(a\cdot v)$ is closed in $Y_2\times Y_2$; for every point $y$ on that orbit, the preimage of $\bar y$ under the *quotient* $\tilde Y\to Y_2\times Y_2$ is the closure of the $U_P$-orbit of $y$, considered as a point in $\tilde Y$. But $\tilde Y$ is affine, hence the $U_P$-orbit of $y$ is closed. Hence, the $P$-orbit of ${\mathcal{R}}_Y(a\cdot v)$ is closed, completing the proof of the lemma.
Now recall the birational map : $J\to (A_X\times {\mathfrak{a}}_X^*)\sslash W_X$. This induces a map $$J^0\bullet T^*X\to (A_X\times {\mathfrak{a}}_X^*)\sslash W_X,$$ which is a smooth geometric quotient[^3] by the $G$-action when restricted to $J\bullet T^*X^{{\operatorname{rs}}}$ (i.e., over $\mathring{\mathfrak{c}}_X^*$), by Lemma \[torsor\].
\[GITquotients\] There is a commutative diagram $$\xymatrix{
J^0\bullet T^*X_{\ne 0} \ar[r] \ar[d] & T^*X_{\ne 0} \times_{{\mathfrak{g}}_X^*} T^*X_{\ne 0} \ar[r] & X\times X \ar[d] \\
(A_X\times {\mathfrak{a}}_X^*)\sslash W_X \ar[r] & A_X\sslash W_X\ar[r] & {\mathfrak{C}}_X:=X\times X\sslash G }$$ which identifies:
- ${\mathfrak{C}}_X$ with $A_X\sslash W_X$;
- for every $c\in {\mathfrak{C}}_X$ with corresponding closed $G$-orbit $C\subset X\times X$, the fiber of $J\bullet T^*X^{{{\operatorname{rs}}}}$ over $c$ with the set $N_C^{*,{{\operatorname{rs}}}}$ of regular semisimple vectors in the conormal bundle to $C$;
- the quotient $N^*_C\sslash G$ with the fiber of $(A_X\times {\mathfrak{a}}_X)^*\sslash W_X$ over $c$.
Since $T^*X\sslash G={\mathfrak{c}}_X^*$, we have ${\mathbbm{P}}(J^0 \bullet T^*X)\sslash G = A_X\sslash W_X$, so the composition ${\mathbbm{P}}(J^0 \bullet T^*X)\to X\times X\to {\mathfrak{C}}_X$ indeed factors through a map $A_X\sslash G\to {\mathfrak{C}}_X$. On the other hand, with notation as in the proof of Lemma \[lemmaregular\], and by the diagram , the composition $${\mathbbm{P}}(J^0\bullet \widetilde{T^*X}^{P,Y}) \twoheadrightarrow {\mathbbm{P}}(J^0 \bullet T^*X) \to X\times X$$ factors through a proper, surjective map $\tilde Y\times^P G\to X\times X$. I claim that the composition $${\mathbbm{P}}(J^0\bullet \widetilde{T^*X}^{P,Y}) \to {\mathbbm{P}}(J^0 \bullet T^*X) \to A_X\sslash W_X$$ factors through a $G$-invariant map $$\tilde Y \times^P G \to A_X\sslash W_X.$$ Indeed, recall that $\tilde Y$ is the closure of the image of $J^0\bullet T_Y^*X^{{\mathfrak{u}}_P^\perp}$ in $Y\times Y$; the map $J^0\bullet T_Y^*X^{{\mathfrak{u}}_P^\perp}\to \tilde Y\sslash P$ factors through $A_X\sslash W_X$ for the same reasons, but on the other hand we have a quotient map $\tilde Y \to Y_2\times Y_2$ (again, in the notation of the proof of Lemma \[lemmaregular\]), and $Y_2\times Y_2\sslash P = A_X\sslash W_X$, so the map to $A_X\sslash W_X \to \tilde Y\sslash P$ is an isomorphism.
Therefore, the map $A_X\sslash W_X\to {\mathfrak{C}}_X$ is surjective. Proposition \[arounddiagonal\] implies that it is also birational. Since both $A_X\sslash W_X$ and ${\mathfrak{C}}_X$ are normal, the map has to be an isomorphism.
Recall that geometric points of ${\mathfrak{C}}_X$ correspond bijectively to closed geometric orbits of $G$ on $X\times X$. By Lemma \[lemmaregular\], these have to be *precisely* the images of ${\mathbbm{P}}(J^0\bullet T^*X^{{\operatorname{rs}}})$ in $X\times X$. Notice that, by Lemma \[torsor\], the set ${\mathbbm{P}}(J^0\bullet T^*X^{{\operatorname{rs}}})$ contains a unique $G$-orbit over any point of ${\mathfrak{C}}_X$.
Let $c\in {\mathfrak{C}}_X$ be such a point, $C\subset X\times X$ the closed $G$-orbit over $c$, and denote by an index $~_c$ various fibers over $c$. By Corollary \[cortorsor\], the fiber $J^0\bullet T^*X^{{\operatorname{rs}}}$ over $c$ coincides with the regular semisimple part $N^{*,{{\operatorname{rs}}}}_C$ of its cotangent bundle. Hence, $$N^{*,{{\operatorname{rs}}}}_C\sslash G= (J\bullet T^*X^{{\operatorname{rs}}})_c\sslash G = ((A_X\times\mathring{\mathfrak{a}}_X^*)/W_X)_c.$$
On the other hand, the invariant moment map gives rise to a $G$-invariant map $N^*_C\to {\mathfrak{c}}_X^*$ which, by considering the regular semisimple and the zero vectors, has to be surjective. Since $N^*_C\sslash G$ is normal, it has to coincide with the spectrum of the integral closure of $F[{\mathfrak{c}}_X^*]$ in the function field of $((A_X\times\mathring{\mathfrak{a}}_X^*)/W_X)_c$, which coincides with $((A_X\times {\mathfrak{a}}_X^*)/W_X)_c$; that is, with ${\mathfrak{a}}_X^*$ (up to $\pm 1$), if $c\ne [\pm 1]$, and with ${\mathfrak{c}}_X^*$, if $c=[\pm 1]$.
\[containsnormal\] The closure of the image of the map contains ${N^*_C}$, for any *closed* $G$-orbit $C\subset X\times X$. Every closed $G$-orbit contains regular semisimple vectors in its conormal bundle.
The second statement was already explained in the proof of the previous lemma, and the first follows by taking the closure of regular semisimple vectors, and using the fact that $N^*_C$ is irreducible.
The statement is not true for non-closed $G$-orbits, which could contribute smaller irreducible components to $T^*X\times_{{\mathfrak{g}}^*} T^*X$.
Consider the case of $X={\operatorname{GL}}_2\backslash{\operatorname{PGL}}_3$, discussed in Example \[pgl3\], where we identified $T^*_{x_0}X={\mathfrak{h}}^\perp$ as the representation ${{\operatorname{Std}}}\oplus {{\operatorname{Std}}}^*$ of $H={\operatorname{GL}}_2$. Let $v^*\in h^\perp$ be a nonzero *irregular* nilpotent vector, i.e., either in ${{\operatorname{Std}}}$ or in ${{\operatorname{Std}}}^*$. The orbit of $x_0$ under its centralizer $G_{v^*}$ is two-dimensional. Hence, the fiber of $$T^*X\times_{{\mathfrak{g}}^*} T^*X$$ over $v^*\in T_{x_0}^*X$ under the first projection is (at least) two-dimensional, while the fiber of $$J^0\bullet T^*X$$ over it is one-dimensional. Thus, there are conormal vectors to $G$-orbits on $X\times X$ which are not contained in the closure of the image of .
Blow-up of $X\times X$ at the closed orbits
-------------------------------------------
We have already seen (Proposition \[arounddiagonal\]) that the map ${\mathbbm{P}}(J^0 \bullet T^*X)\to X\times X$ is an isomorphism, generically; in particular, generic fibers over ${\mathfrak{C}}_X=A_X\sslash W_X$ (Proposition \[GITquotients\]) are single $G$-orbits. We can now determine which ones:
\[genericisomorphism\]
1. Let $c\in {\mathfrak{C}}_X$ with corresponding closed $G$-orbit $C\subset X\times X$. If $c\ne [\pm 1]$, the linear map $N^*_C \to N^*_C\sslash G= {\mathfrak{a}}_X^*$ (up to $\pm 1$) is an *isomorphism* on each fiber over $C$. If $c=[\pm 1]$, the quadratic map $N^*_C\to N^*_C\sslash G={\mathfrak{c}}_X^*$ is *nondegenerate* on each fiber over $C$.
2. The map ${\mathbbm{P}}(J^0 \bullet T^*X)\to X\times X$ is an isomorphism over $\mathring{\mathfrak{C}}_X := {\mathfrak{C}}_X\smallsetminus\{[\pm 1]\}$.
3. The map $X\times X\to A_X\sslash W_X$ is a smooth geometric quotient by the $G$-action away from $[\pm 1]\in A_X\sslash W_X$.
Let $V$ be the fiber of $N^*_C$ over a point of $C$, and $H_0$ the stabilizer of that point. If the linear map (when $c\ne [\pm 1]$) $$V\to {\mathfrak{a}}_X^*,$$ where ${\mathfrak{a}}_X^*$ is identified, up to $\pm 1$, with the fiber of $A_X\times {\mathfrak{a}}_X^*\sslash W_X$ over $c$, resp. if the quadratic map (when $c=[\pm 1]$) $$V\to {\mathfrak{c}}_X^*$$ were trivial on a nonzero, necessarily $H_0$-stable subspace $V_0\subset V$, that space would have an $H_0$-stable complement $$V= V_0 \oplus V_0',$$ identifying the invariant-theoretic quotients $$N^*_C\sslash G = V\sslash H_0 = V_0'\sslash H_0 = {\mathfrak{c}}_X^*.$$ In particular, the $G$-orbit of a generic point of $N^*_C$ (corresponding to an $H_0$-orbit on $V$ not belonging to $V_0'$) is not closed, a contradiction, since by Proposition \[GITquotients\] and Corollary \[corollaryGtransitive\] the fibers over all points of $\mathring {\mathfrak{c}}_X^*$ are $G$-homogeneous.
This proves the first claim, and it implies that closed $G$-orbits in $X\times X$ over $c\ne [\pm 1]$ are of codimension one, hence coincide with the whole fiber. Hence, by Corollary \[cortorsor\], the map from $J\bullet T^*X^{{\operatorname{rs}}}$ to the union $T^*X_{\ne 0} \times_{{\mathfrak{g}}^*} T^*X_{\ne 0}$ of nonzero vectors in the union of conormal bundles is an isomorphism over $\mathring {\mathfrak{C}}_X$, therefore the projectivization ${\mathbbm{P}}(J\bullet T^*X^{{\operatorname{rs}}})$ (or, equivalently, ${\mathbbm{P}}(J^0\bullet T^*X)$) is isomorphic to $X\times X$ over this subset.
The last claim follows from the analogous claim for ${\mathbbm{P}}(J \bullet T^*X^{{\operatorname{rs}}})$, where it is obvious.
Finally, we are ready to construct a resolution of $X\times X$. We now know that this space contains two closed $G$-orbits, $X_1 = X^{{\operatorname{diag}}}$ and $X_{-1}$, around which the map to $A_X\sslash W_X$ may fail to be a geometric quotient, namely, the ones over the points $[\pm 1]\in{\mathfrak{C}}_X$. The resolution that we will construct will eventually turn out to be, simply, the blowup at those two subsets. However, generalizing Proposition \[arounddiagonal\], we will construct this resolution by a slight modification of the space $J\bullet T^*X$, that we will denote by $J_X$.
The scheme $J_X$ will be glued from two open subsets: the first is $U_1:=J^0\bullet T^*X$. The second is $U_{-1} := J^0\bullet N^*_{X_{-1}}$, where $X_{-1}$ is the closed $G$-orbit corresponding to $[-1]\in {\mathfrak{C}}_X$. We define $J_X= U_1\cup U_{-1}$, glued over their subsets of regular semisimple vectors as follows: Notice that $U_1^{{\operatorname{rs}}}=J\bullet T^*X^{{\operatorname{rs}}}$. Moreover, by Proposition \[GITquotients\], we have an identification of $N^{*,{{\operatorname{rs}}}}_{X_{-1}}$ with the subset $(-1)\cdot T^*X^{{{\operatorname{rs}}}}$ of $J\bullet T^*X^{{\operatorname{rs}}}$, hence $U_{-1}^{{\operatorname{rs}}}= J\bullet (-1)\cdot T^*X^{{{\operatorname{rs}}}}$. This defines the isomorphism $$U_{-1}^{{\operatorname{rs}}}= J\bullet (-1)\cdot T^*X^{{{\operatorname{rs}}}} \ni (j,(-1)\cdot v^*)\mapsto ((-1)\cdot j, v^*)\in J\bullet T^*X^{{\operatorname{rs}}}= U_1^{{\operatorname{rs}}},$$ hence the scheme $J_X$. This scheme retains a map $$\label{birationalX} J_X \to (A_X\times {\mathfrak{a}}_X^*)\sslash W_X,$$ whose restriction to $\mathring{\mathfrak{c}}_X^*$ is equal to $J\bullet T^*X^{{\operatorname{rs}}}$.
The map extends to $J_X$: $$\label{JtoconormalX} J_X \to T^*X\times_{{\mathfrak{g}}^*} T^*X,$$ and we define distinguished divisors $[1]_{J_X}$, $[-1]_{J_X}$ and $J_X^{{\operatorname{nilp}}}$, similarly as in §\[ssKnopsgroupscheme\].
Extending Corollary \[corollarydivisors\],
\[corollarydivisorsJX\] The divisors $[\pm 1]_{J_{X,\ne 0}}$ and $J_{X,\ne 0}^{{\operatorname{nilp}}}$ in $J_{X,\ne 0}$ intersect transversely, and the morphism $J_{X,\ne 0} \to A_X\sslash W_X$ is smooth away from $[\pm 1]_{J_{X,\ne 0}}$.
Here, $J_{X,\ne 0}$ denotes the complement of the zero section $J^0\bullet X\subset J^0\bullet T^*X$ in $U_1$, and of the zero section $J^0\bullet X_{-1}\subset J^0\bullet N^*_{X_{-1}}$ in $U_{-1}$.
On the open subset $U_1$, this is contained in Corollary \[corollarydivisors\]. For $U_{-1}$, the same proof, based on Lemma \[lemmatransverse\], works, because of the non-degeneracy statement of the first part of Proposition \[genericisomorphism\].
Now consider the composition $J_X \to T^*X\times_{{\mathfrak{g}}^*} T^*X \to X\times X$. On $J_{X,\ne 0}$, it clearly factors through the projectivization ${\mathbbm{P}}J_X$.
\[isblowup\] The morphism $${\mathcal{R}}: {\mathbbm{P}}J_X \to X\times X$$ is isomorphic to the blowup of $X\times X$ at the closed $G$-orbits $X_1$ and $X_{-1}$. The preimage of any point of ${\mathfrak{C}}_X=X\times X\sslash G$ under the composition of the maps ${\mathbbm{P}}J_X\to X\times X\to {\mathfrak{C}}_X$ is a normal crossings divisor.
The statement identifying ${\mathbbm{P}}J_X$ as a blowup has already been proven away from $[-1]\in {\mathfrak{C}}_X$, by a combination of Propositions \[arounddiagonal\] and \[genericisomorphism\]. On a $G$-stable neighborhood of $[-1]_{J_X}$, it can be proven by exactly the same arguments as in Proposition \[arounddiagonal\]. Notice that, now, the map $N^*_{X_{-1}}\to {\mathfrak{c}}_X^*$, viewed as a quadratic form on the fibers by fixing a coordinate $\xi$ on ${\mathfrak{c}}_X^*$, gives rise to a map from the conormal to the normal bundle: $$N^*_{X_{-1}}\to N_{X_{-1}},$$ and the non-degeneracy of this quadratic form (Proposition \[genericisomorphism\]) implies that this map is an isomorphism. This fact implies, as in Proposition \[arounddiagonal\], that the map from ${\mathbbm{P}}J_X$ to the blowup of $X\times X$ at $X_{-1}$ is an isomorphism around $[-1]_{J_X}$.
The preimage of any point on ${\mathfrak{C}}_X$ is either a unique (smooth) $G$-orbit of codimension one in $X\times X$, or is contained in the divisors $[\pm 1]_{{\mathbbm{P}}J_X}$ and ${\mathbbm{P}}J_X^{{\operatorname{nilp}}}$; by Corollary \[corollarydivisorsJX\], these have normal crossings.
I finish this section by stating rationality properties of the map $N^*_{X_{-1}}\to {\mathfrak{c}}_X^*$, analogous to those of the invariant moment map that were proven in §\[ssrationality\]. Notice that, up to this point in this section, we have not used the fact that we are working with the “correct representative” of a variety in its class modulo $G$-automorphisms (see Proposition \[correctrepresentative\]), but now we will.
Let $(Y,P)$ be a basic orbit-parabolic pair, as in Proposition \[propVP\]; hence, $Y$ is a closed $P$-orbit, and the quotient $Y/\mathcal R(P)$ is isomorphic to $T\backslash {\operatorname{PGL}}_2$ or to ${\operatorname{PGL}}_2$. We have defined a cover $\widetilde{T^*X}^P\to T^*X$, and an irreducible component $\widetilde{T^*X}^P$ thereof; by base change, we get analogous covers for $J\bullet T^*X$. We let $$\widetilde{J_X}^P = \{(v,P')| v\in J_X, P'\sim P, \mu(v) \in {\mathfrak{u}}_{P'}^\perp\}$$ (where $\mu$ also denotes the moment map for $J_X$), and we let $ \widetilde{J_X}^{P,Y}$ be the closure of $J\bullet \widetilde{T^*X}^{P,Y,{{\operatorname{rs}}}}$ in $\widetilde{J_X}^P$. (Recall that $J_X^{{\operatorname{rs}}}= J\bullet T^*X^{{\operatorname{rs}}}$.) Finally, recalling that $N^*_{X_{-1}}$ is the closure of $(-1)\cdot T^*X^{{\operatorname{rs}}}$ in $J_X$, let $\widetilde{N^*_{X_{-1}}}^{P,Y}$ be the closure of $(-1)\cdot \widetilde{T^*X}^{P,Y,{{\operatorname{rs}}}}$ in $\widetilde{J_X}^{P,Y}$.
Explicitly, fix the pair $(Y,P)$, let $x_1\in Y\subset X$, let $V = T^*_{x_1} X$, and $V_P = V\cap \mu^{-1}({\mathfrak{u}}_P^\perp)$, as in Proposition \[propVP\], $v\in V_P^{{\operatorname{rs}}}$. Then $(-1)\cdot v \in T^*X$, the translate of $v$ under the action of the $(-1)$-section of $J$, lives over a point $x_2$ which also belongs to $Y$, because this action can be induced by the centralizer of $v$ in a Levi subgroup of $P$ (as in the proof of Lemma \[lemmaregular\]). The point $x=(x_1,x_2)\in X\times X$ belongs to the closed orbit $X_{-1}$, by Lemma \[lemmaregular\], and setting $H_i$ for the stabilizer of $x_i$, and $V' = ({\mathfrak{h}}_1+{\mathfrak{h}}_2)^\perp\subset {\mathfrak{g}}^*$, $V'_P=V'\cap {\mathfrak{u}}_P^\perp$, we have $$\widetilde{N^*_{X_{-1}}}^{P,Y} = V'_P\times^{(P\cap H_1\cap H_2)} G \longrightarrow V'\times^{(H_1\cap H_2)} G = N^*_{X_{-1}}.$$
Let $Y_2 = Y/U_P$, as before, and let $Y_{2,-1}$ be analog of $X_{-1}$ for $Y_2$ — that is, the closed $L$-orbit on $Y_2\times Y_2$ which contains the projections of cotangent pairs $(v, (-1)v)$, where $v\in T^*Y_2^{{\operatorname{rs}}}$. If $\bar x$ is the image of $x$ in $Y_{2,-1}$, and $V_2'$ the fiber of $N^*_{Y_{2,-1}}(Y_2\times Y_2)$ over $\bar x$, then we have a quotient map $V'_P\to V'_2$.
\[propVP2\] In the setting above, the map $\widetilde{N^*_{X_{-1}}}^{P,Y} \to N^*_{X_{-1}}$ is surjective on $F$-points.
The kernel of the map $V_P'\to V_2'$ is an isotropic subspace of $V'$ (with respect to the quadratic map $V'\to {\mathfrak{c}}_X^*$) of dimension $$\label{dimensions2}
\dim\ker(V'_P\to V'_2) = \frac{\dim V' - \dim V'_2}{2}.$$
The quadratic space $V'$ is split (maximally isotropic) if and only if $V'_2$ is, which happens if and only if the quadratic space $V$ is.
Hence, by the equivalences of Proposition \[rationality\], the fibers of $N^*_{X_{-1}}$ are split quadratic spaces if and only if the stabilizer of one point on $X$ is split.
The proof is identical to that of Proposition \[propVP2\]; we just need to add how the property of $V'_2$ being split relates to the property of $V_2$ being split. But this is clear from considering the space $Y_2$ or, equivalently, $Y_{2,{{\operatorname{ad}}}}= Y_2/\mathcal Z(L)$, the latter being isomorphic to $T\backslash {\operatorname{PGL}}_2$ or to ${\operatorname{PGL}}_2$. In the both cases, there is an automorphism of rank $2$ of $Y_2$, which, applied to one copy of $Y_2$, interchanges the orbits $Y_2^{{\operatorname{diag}}}$ and $Y_{2,-1}\subset Y_2\times Y_2$; so, one, equivalently all fibers of the conormal bundle of the former are split if and only if one, equivalently all, fibers of the conormal bundle of the latter are.
Finally:
\[surjectiveonpoints\] Assume that one, equivalently all, stabilizers of points on $X$ are split. Then the map $X\times X\to {\mathfrak{C}}_X = A_X\sslash W_X$ is surjective on $F$-points.
By Proposition \[rationality\], the stabilizers being split is equivalent to the map $T^*X\to {\mathfrak{c}}_X^*$ being totally isotropic, which implies that it is surjective. This means that the projection $J^{{\operatorname{rs}}}\bullet T^*X^{{\operatorname{rs}}}\to J^{{\operatorname{rs}}}$ is surjective on $F$-points. The map $J^{{\operatorname{rs}}}\to A_X\sslash W_X$ is also surjective on $F$-points. Hence, the composition ${\mathbbm{P}}J_X^{{\operatorname{rs}}}= {\mathbbm{P}}(J^{{\operatorname{rs}}}\bullet T^*X^{{\operatorname{rs}}})\to X\times X\to {\mathfrak{C}}_X$ is surjective on $F$-points.
\[pgl3-second\] Let us consider the case of the variety $X={\operatorname{GL}}_2\backslash {\operatorname{PGL}}_3 =$ the variety of decompositions ${\mathbbm{G}_a}^3=V_2\oplus V_1$, that we already saw in Example \[pgl3\]. Letting $x_1$ be the decomposition $\left<e_1,e_2\right>\oplus\left<e_3\right>$, and $P=$ the stabilizer of the plane $\left<e_2,e_3\right>$, we will get $x_2=$ a decomposition $\left<e_1 + ce_2, e_3\right> \oplus \left<e_2\right>$, with the scalar $c$ depending on the chosen cotangent vector. Then we see that $H_1\cap H_2 = P\cap H_1\cap H_2 \simeq{\mathbbm{G}_m}$, $V'=V_P'=({\mathfrak{h}}_1+{\mathfrak{h}}_2)^\perp = {{\operatorname{Std}}}_1 \oplus {{\operatorname{Std}}}_1^*$, where ${{\operatorname{Std}}}_1$ denotes the standard one-dimensional representation of ${\mathbbm{G}_m}$, and $V'$ intersects the nilpotent cone in ${\mathfrak{g}}^*$ along *irregular* orbits only. The reader should compare this with Example \[pgl3\], where we saw that the nilpotent limit of the action of the $(-1)$-section of $J$ on regular semisimple vectors does not exist, but it does exist at the exceptional divisor of the blowup along the (irregular nilpotent) divisor ${{\operatorname{Std}}}\oplus {{\operatorname{Std}}}^*$. This blowup, with the strict transform of the nilpotent divisor removed, is isomorphic to the conormal bundle $N^*_{X_{-1}}$.
Integration formula {#sec:integration}
===================
The goal of this section is to prove the theorem below, which uses the following standard fact: if $H_1, H_2\subset G$ are two subgroups that are conjugate over the algebraic closure, and the normalizer of $H_i$ acts trivially on the top exterior power $\bigwedge^{{\operatorname{top}}}\mathfrak h_i$ (in particular, $H_i$ is unimodular), then any choice of $G$-invariant measure $\mu_1$ on $H_1\backslash G$ induces, in a canonical way, a $G$-invariant measure $\mu_2$ on $H_2\backslash G$; simply, write $\mu_1=|\omega_1|$ for some invariant volume form $\omega_1$, and let $\mu_2=|\omega_2|$ for some invariant volume form $\omega_2$, such that, over the algebraic closure, $\omega_2$ is conjugate to $\epsilon\omega_1$, for some $\epsilon \in \bar F$ with $|\epsilon|=1$. Recall that we have fixed a Haar measure on $F$, so that the absolute value of a volume form is a well-defined measure.
\[thmintegrationformula\] Let $\omega$ be a nonzero $G\times G$-invariant volume form on $X\times X$ defining an invariant measure $|\omega|$. The $G$-stabilizers of points on $X\times X$ over any $c\ne [\pm 1]\in A_X\sslash W_X$ are all conjugate, over the algebraic closure, to the kernel $L_1$ of the map $L(X)\twoheadrightarrow A_X$, where $L(X)$ is a Levi subgroup of a parabolic of type $P(X)$. Fix compatible $G$-invariant measures $d\dot{g}$ on all of them (s. comment above).
Then, identifying $A_X\sslash W_X\simeq \mathbbm A^1$ and letting $c_{\pm 1}$ be the coordinates of the points $[\pm 1]$, there is an additive Haar measure $dc$ such that the following integration formula holds: $$\label{integrationformula} \int_{X\times X} \Phi(x) |\omega|(x) = \int_{A_X\sslash W_X} |c-c_1|^{\frac{d_1}{2}-1} |c-c_{-1}|^{\frac{d_{-1}}{2}-1} \left(\int_{(X\times X)_c} \Phi(\dot{g}) d\dot{g} \right) dc.$$
Here, $d_1=\dim X={{\operatorname{codim}}}X_1$, and $$\label{codimensionseq}
d_{-1}={{\operatorname{codim}}}X_{-1} = 2\epsilon \left<\rho_{P(X)},\check\gamma\right>-d_1+2,$$ where $\check\gamma$ is the spherical coroot, and $$\epsilon = \begin{cases} 1, \mbox{ for roots of type $T$ (dual group ${\operatorname{SL}}_2$)};\\
2, \mbox{ for roots of type $G$ (dual group ${\operatorname{PGL}}_2$).}
\end{cases}$$ Moreover, in the case of type $G$ we have $$\label{equalcodim}d_1=d_{-1} = \left<2\rho_{P(X)},\check\gamma\right>+1.$$
Notice that the formula for ${{\operatorname{codim}}}X_{-1}$ is new, and will be proven as a corollary of the integration formula.
There are also analogous integration formulas for the normal/conormal bundles of the orbits $X_1$ and $X_{-1}$. Notice that, by the nondegenerate quadratic forms obtained by the invariant-theoretic quotients $$N_{X_1}^*\to {\mathfrak{c}}_X^* \leftarrow N_{X_{-1}}^*$$ (see Proposition \[genericisomorphism\]), the normal and conormal bundles are $G$-equivariantly isomorphic.
\[thmintegrationformula-linear\] Let $\omega$ be a nonzero $G$-invariant volume form on $N_{X_{\pm 1}}^*$ which restricts to Haar measures on the fibers. The $G$-stabilizers of points on $X\times X$ over any $\xi\ne 0\in {\mathfrak{c}}_X^*$ are all conjugate, over the algebraic closure, to the kernel $L_1$ of the map $L(X)\twoheadrightarrow A_X$, where $L(X)$ is a Levi subgroup of a parabolic of type $P(X)$. Fix compatible $G$-invariant measures $d\dot{g}$ on all of them.
Then, there is an additive Haar measure $d\xi$ on ${\mathfrak{c}}_X^*\simeq \mathbbm A^1$ such that the following integration formula holds: $$\int_{N^*_{X_{\pm 1}}} \Phi(x) |\omega|(x) = \int_{{\mathfrak{c}}_X^*} |\xi|^{\frac{d_{\pm 1}}{2}-1} \left(\int_{(N^*_{X_{\pm 1}})_\xi} \Phi(\dot{g}) d\dot{g} \right) d\xi.$$
The proof of Theorem \[thmintegrationformula-linear\] is completely analogous to that of Theorem \[thmintegrationformula\], and therefore I will only present that of Theorem \[thmintegrationformula\], leaving the reformulation for the other to the reader.
Pullback to the polarization
----------------------------
Consider the map ${\mathbbm{P}}J_X\xrightarrow{{\mathcal{R}}} X\times X \to A_X\sslash W_X$. Recall that $J_X^{{\operatorname{rs}}}= J\bullet T^*X^{{\operatorname{rs}}}$; we let $$\widehat{J_X}^{\bullet,{{\operatorname{rs}}}} = J\bullet \widehat{T^*X}^{\bullet,{{\operatorname{rs}}}} = A_X \times \widehat{T^*X}^{\bullet,{{\operatorname{rs}}}} \subset J_X\times_{{\mathfrak{c}}_X^*} {\mathfrak{a}}_X^*,$$ where $\widehat{T^*X}^\bullet$ is the distinguished irreducible component of the polarized cotangent bundle that was defined in §\[ssmomentmap\].
We have a commutative diagram $$\xymatrix{
{\mathbbm{P}}\widehat{J_X}^{\bullet,{{\operatorname{rs}}}} \ar[r]^{p}\ar[d] & {\mathbbm{P}}J_X \ar[r]^{{\mathcal{R}}}\ar[d] & X\times X \ar[dl]\\
A_X \ar[r] & A_X\sslash W_X.}$$
Recall that we denote by $[\pm 1]_{{\mathbbm{P}}J_X}$ the exceptional divisors of the blowup ${\mathcal{R}}$, and by $d_{\pm 1}$ the codimensions of the orbits $X_{\pm 1}$.
\[canonicaldivisor\] Let $K_{X\times X}$ be the canonical bundle on $X\times X$. Then $${\mathcal{R}}^* K_{X\times X} = K_{{\mathbbm{P}}J_X} ((d_1-1)[1]_{{\mathbbm{P}}J_X} + (d_{-1}-1)[-1]_{{\mathbbm{P}}J_X}).$$
This is immediate from the characterization of ${\mathbbm{P}}J_X$ as the blowup of $X\times X$ at the two divisors $X_1=X^{{\operatorname{diag}}}$ and $X_{-1}$ (Proposition \[isblowup\]).
Hence, if $\omega$ is a nonzero, $G\times G$-invariant volume form on $X\times X$, the divisor of its pullback to ${\mathbbm{P}}J_X$ is $$[{\mathcal{R}}^*\omega]=(d_1-1) \left[ [1]_{{\mathbbm{P}}J_X}\right] + (d_{-1}-1)\left[[-1]_{{\mathbbm{P}}J_X}\right],$$ where $d_1 = {{\operatorname{codim}}}(X_1) = \dim X$, and $d_{-1} = {{\operatorname{codim}}}(X_{-1})$.
Set $Y={\mathbbm{P}}\widehat{J_X}^{\bullet,{{\operatorname{rs}}}}$, for notational simplicity. The map $p:Y \to {\mathbbm{P}}J_X $ is an étale ${\mathbb{Z}}/2$-cover onto its image, and notice that $\widehat{J_X}^{\bullet,{{\operatorname{rs}}}} = A_X \times \widehat{T^*X}^{\bullet,{{\operatorname{rs}}}}$, canonically. Thus, setting $\hat{\mathcal{R}}=p\circ{\mathcal{R}}$, we have $$\label{omegatopolar}[\hat{\mathcal{R}}^*\omega] = (d_1-1) [Y_1] + (d_{-1}-1)[Y_{-1}].$$ where $Y_{\pm 1} = [\pm 1]_Y\subset Y$.
Fix a pair $(x,B)\in (X\times {\mathcal{B}})^\circ$, defining an embedding $\mathring{\mathfrak{a}}_X^*\to \widehat{T^*X}^{\bullet,{{\operatorname{rs}}}}$ by Knop’s section $\hat\kappa_X$ (§\[ssmomentmap\]), and let $L_1$ be the stabilizer of the points in the image. If $L$ denotes the centralizer of the image of such a point under the polarized moment map to $\hat{\mathfrak{g}}_X^*$, identified with the Levi quotient $L(X)$ of $P(X)$, then $L_1\supset [L,L]$, and $L/L_1\simeq A_X$, canonically because of the polarization. The action map identifies $$\widehat{T^*X}^{\bullet,{{\operatorname{rs}}}} \simeq \mathring{\mathfrak{a}}_X^* \times L_1\backslash G.$$ Hence: $$\label{polarisom}{\mathbbm{P}}\widehat{J_X}^{\bullet,{{\operatorname{rs}}}} = A_X \times L_1\backslash G.$$
Fix an invariant volume form $\omega_{L_1\backslash G}$ on $L_1\backslash G$. Then, by and the fact that the only nowhere vanishing regular functions on a torus are characters, there is a Haar volume form $\omega_{A_X}$ on $A_X$ and an $m\in {\mathbb{Z}}$ such that $$\label{formonpolarization} \hat{\mathcal{R}}^*\omega = (a-1)^{d_1-1} (a+1)^{d_{-1}-1} a^m \cdot \omega_{A_X}\wedge \omega_{L_1\backslash G}$$ under , where we have identified $A_X\simeq {\mathbbm{G}_m}$ to fix a coordinate $a$. On the other hand, this has to be invariant under the $W_X$-Galois action $a\mapsto a^{-1}$, hence $$m = 1- \frac{d_1+d_{-1}}{2}.$$
Descent to $X\times X$
----------------------
The integration formula , now, follows from by descending to $A_X\sslash W_X \simeq \mathbbm A^1$: Fix a coordinate $c$ on that space, with $c_{\pm 1}$ the coordinates of the points $[\pm 1]$. In a sufficiently small neighborhood $U$ of any point of $A_X\sslash W_X\smallsetminus\{[\pm 1]\}$ the stabilizers of all points are conjugate to a group $L_1'$ which is conjugate over the algebraic closure to $L_1$, and the preimage of $U$ in $X\times X$ is, in the semialgebraic topology, $G$-equivariantly isomorphic to $U\times L_1'\backslash G$. Thus, fixing the compatible measures $d\dot{g}$ on the $G$-orbits as in the statement of the theorem, there is an integration formula of the form: $$\int_{X\times X} |\omega| = \int_{A_X\sslash W_X} \varphi(c) \int_{(X\times X)_c} d\dot{g} dc$$ for some nonnegative measurable function $\varphi$ on $A_X\sslash W_X\smallsetminus\{[\pm 1]\}$, and some additive Haar measure $dc = |\omega_{\mathbbm A^1}|$ on $\mathbbm A^1$. On the other hand, writing, in such a neighborhood $c$, the measure $d\dot{g}$ on $L_1'\backslash G$ as $|\omega'|$, for some invariant volume form $\omega'$, we see by applying over a suitable algebraic extension of $F$ that the pullback of $\omega_{\mathbbm A^1}\wedge \omega'$ to ${\mathbbm{P}}\widehat{J_X}^{\bullet,{{\operatorname{rs}}}}$ has to be a multiple of $(a-1)^{d_1-1} (a+1)^{d_{-1}-1} a^m \cdot \omega_{A_X}\wedge \omega_{L_1\backslash G}$ by a rational function $f(a)^{-1}$ with $|f(a)| = \varphi(c(a))$. Without loss of generality, the pullback of $\omega'$ is equal to $\omega_{L_1\backslash G}$, and an easy calculation shows that, up to a scalar, $$|\hat{\mathcal{R}}^*\omega_{\mathbbm A^1}| = |c-c_1|^\frac{1}{2}|c-c_{-1}|^\frac{1}{2} |\omega_{A_X}|,$$ and $|a-(\pm 1)|^2\cdot {|a|^{-1}} = |c - c_{\pm 1}|$ (up to a fixed scalar), hence: $$\varphi(c) = |c - c_1|^{\frac{d_1}{2}-1} |c - c_{-1}|^{\frac{d_{-1}}{2}-1}$$ for a suitable $dc$.
Degeneration {#ssdegeneration}
------------
We have proven the integration formula of Theorem \[thmintegrationformula\], except for the determination of the codimension $d_{-1}$ of the orbit $X_{-1}$. In this subsection we will prove the codimension formula : $$(d_1-1) + (d_{-1}-1) = 2\epsilon \left<\rho_{P(X)},\check\gamma\right>,$$ where $\check\gamma$ is the spherical coroot, and $$\epsilon = \begin{cases} 1, \mbox{ for roots of type $T$ (dual group ${\operatorname{SL}}_2$)};\\
2, \mbox{ for roots of type $G$ (dual group ${\operatorname{PGL}}_2$).}
\end{cases}$$ with $d_1=d_{-1} = \left<2\rho_{P(X)},\check\gamma\right>+1$ in the case of type $G$.
To prove this, we degenerate $X$ to its *boundary degeneration* $X_\emptyset$, which is a horospherical variety. The codimension formula will follow by comparing the integration formula to the corresponding formula for $X_\emptyset\times X_\emptyset$, which is very easy to compute.
More precisely, consider the decomposition of the coordinate ring of $X$ as a $G$-module into irreducibles: $$F[X] = \bigoplus_{\lambda \in \Lambda_X^+} F[X]_\lambda.$$ The indices $\lambda$ denote, here, the highest weight of the representation, and $\Lambda_X^+\subset \Lambda_X:={\operatorname{Hom}}(A_X,{\mathbbm{G}_m})\simeq {\mathbb{Z}}$ is the monoid of weights such that the corresponding rational $B$-eigenfunction is regular; from now on we identify it with $\mathbb N$. The above decomposition is not an algebra grading, a fact that is known to be equivalent to the fact that $W_X\ne 1$ (see [@KnAut]). Instead, it corresponds to an algebra filtration $$\mathcal F_\lambda = \bigoplus_{\nu\le \lambda} F[X]_\nu,$$ i.e., $\mathcal F_\lambda \cdot \mathcal F_\mu\subset \mathcal F_{\lambda+\mu}$. The Rees family: $$F[\mathscr X] := \bigoplus_\lambda \mathcal F_\lambda \cdot t^\lambda \subset F[X][t]$$ defines an affine $G$-variety $\mathscr X$ over $\mathbbm A^1={\operatorname{Spec}}F[t]$, together with an action of ${\mathbbm{G}_m}$ that extends to a morphism $\tau: X\times \mathbbm A^1\to \mathscr X$; more canonically, ${\mathbbm{G}_m}= {\operatorname{Spec}}F[\Lambda_X] = A_X$, and the base $\mathbbm A^1$ of this family is the affine embedding $\overline{A_X}\supset A_X$ on which elements of $\Lambda_X^+$ extend to regular functions, so the defining morphism is $$\label{tau}
\xymatrix{ X\times \overline{A_X}\ar[dr]\ar[rr]^\tau && \mathscr X\ar[dl]^\pi\\ & \overline{A_X},}$$ and extends to a canonical action of $A_X\times G$ on $\mathscr X$ over $\overline{A_X}$.
The special fiber $X_\emptyset:=\mathscr X_0$ is an affine horospherical $G$-variety[^4] with $P(X_\emptyset) = P(X)$; its open $G$-orbit $X_\emptyset^\bullet$ is isomorphic to $U^-L_1\backslash G$, where $U^-$ is the unipotent radical of a parabolic opposite to $P(X)$, and $L_1$ is the kernel of the map $L(X)\twoheadrightarrow A_X$, as above. Notice that the image of the defining morphism $\tau$ lies in the complement of the open orbit over $0\in \overline{A_X}$.
It is known that there is a family of $G$-invariant volume forms on the homogeneous parts of the fibers of $\mathscr X\to \mathbbm A^1$ that is everywhere nonvanishing, cf. [@SV §4.2]. More precisely, let us fix a parabolic $P$ in the class of $P(X)$, and let $\mathring X$ be the open $P$- (and Borel) orbit. Restricting to $U$-invariants, the above decomposition becomes a grading $$F[X]^U = \bigoplus_\lambda F[X]^U_\lambda,$$ with $F[X]^U_\lambda=$ the (one-dimensional) highest weight subspace of $F[X]_\lambda$, on which $P(X)$ acts by the character $\lambda$ of the quotient $A_X$. Correspondingly, the family $\mathscr X\sslash U$ becomes constant: $$\label{constantmodU}
\sigma:\mathscr X\sslash U \xrightarrow\sim X\sslash U \times \overline{A_X},$$ but it can be seen from the definitions that this isomorphism is related to the one (call it $\tau_U$) that we obtain by descending the defining map $\tau$ of by: $$\label{comparison} \sigma^{-1}(\bar x, t) = \tau_U(\bar x\cdot t^{-1}, t),$$ where we have used the canonical action of $A_X$ on $X\sslash U$; that is, the action of $A_X$ on $\mathscr X$ descends to the action $$\label{comparison2} a\cdot (\bar x, t) = (\bar x\cdot a, at)$$ on $X\sslash U \times \overline{A_X}$.
If $\mathring{\mathscr X}$ is the union of open $P$-orbits on the various fibers, the restriction of $\sigma$ defines an isomorphism $$\mathring{\mathscr X}/U \xrightarrow\sim \mathring X/U \times \overline{A_X},$$ which, by the local structure theorem [@BLV Théorème 1.4], [@KnMotion Theorem 2.3], can be lifted $P=L\cdot U$-equivariantly: $$\label{tildesigma}\tilde\sigma: \mathring{\mathscr X} \simeq S \times U \times \overline{A_X},$$ for some Levi subgroup $L\subset P$, with $S\simeq \mathring X/U$ an $L$-stable subvariety of $X$, acting by conjugation on $U$ and via the quotient $A_X$, simply transitively, on $S$. More precisely, $L$ is the centralizer of the image of an element $v\in T^*\mathring X^{{\mathfrak{u}}^\perp,{{\operatorname{rs}}}}$ under the moment map, and $S$ is a “flat”, that is, the $L$-orbit of the image of $v$ on $X$.
From now on, by abuse of notation, any reference to $\mathscr X$ should be taken to refer to the smooth locus of the map $\pi:\mathscr X\to \overline{A_X}$. Consider $\Omega:=\Omega_{\mathscr X/ \mathbbm A^1}=$ the relative cotangent sheaf of $\mathscr X\xrightarrow\pi \mathbbm A^1$. Restricted to any fiber, it is canonically identified with the cotangent bundle of that fiber. Its top exterior power, $\bigwedge^{{\operatorname{top}}}\Omega$, restricts to the bundles of volume forms on the fibers (of the smooth locus).
There is a $G$-invariant section $\omega$ of $\bigwedge^{{\operatorname{top}}}\Omega$ which restricts to a nonvanishing volume form on each fiber. Moreover, such a form is an $A_X$-eigenform satisfying, for every $a\in A_X$, $$\label{omegaeigen} a^*\omega_{\mathscr X} = e^{2\rho_{P(X)}}(a) \cdot \omega_{\mathscr X}.$$
Notice that we use exponential notation for the character $2\rho_{P(X)}=$the sum of roots in the unipotent radical of $P(X)$, since we use additive notation for roots.
A nonzero, $P$-invariant volume form on $X/U \times U$ pulls back by to a $P$-invariant section $\omega_{\mathring{\mathscr X}}$ of the line bundle $\bigwedge^{{\operatorname{top}}}\Omega$ on $\mathring{\mathscr X}$. If $X$ admits a $G$-invariant measure (as is the case for affine homogeneous spaces), so does $X_\emptyset$ [@SV §4.2], and the $g$-pullback of $\omega_{\mathring{\mathscr X}}$, for every $g\in G$, is a $gPg^{-1}$-invariant section of $\Omega$ on $\mathring{\mathscr X}g^{-1}$, which coincides with $\omega_{\mathring{\mathscr X}}$ on the intersection with $\mathring{\mathscr X}$; thus, these translates glue to a global section $\omega_{\mathscr X}$ of $\Omega$ over the union $\mathscr X^\bullet$ of all open $G$-orbits on the fibers.
This section restricts, by construction, to a nonvanishing volume form on each fiber. Notice, also, that any other such form should be a multiple of $\omega_{\mathscr X}$ by a nowhere vanishing regular function on $\overline{A_X}\simeq \mathbbm A^1$, hence by a scalar.
Regarding the action of $A_X$, it is enough to prove for the restriction of $\omega_{\mathscr X}$ to $\mathring{\mathscr X}$. In terms of the isomorphism , the $P$-invariant form is given by $$\omega_{\mathring{\mathscr X}}(s,u,t) = e^{2\rho_{P}}(s) \cdot \omega_S(s) \wedge \omega_U(u),$$ where $\omega_S$ is an $A_X$-invariant volume form on $S$, $\omega_U$ is a $U$-invariant volume form on $U$, and we have identified $S\simeq A_X$ by choosing a base point. The action of $A_X$ on $\mathring{\mathscr X}$ is given by on $S\times \overline{A_X}$, and trivial action on $U$, therefore this form is $e^{2\rho_{P(X)}}$-equivariant.
Now we move to the space $\widetilde{{\mathscr X}}:=\mathscr X\times_{\overline{A_X}}\mathscr X$. Again, we only work over the smooth locus of the morphism to $\overline{A_X}$. The tensor product of $\omega_{\mathscr X}$ with itself gives rise to a section $\omega_{\widetilde{\mathscr X}}$ of the top exterior power of the relative cotangent bundle of $\widetilde{\mathscr X}\to \overline{A_X}$, which restricts to an invariant, nonvanishing, $G\times G$-invariant volume form on the open orbit of each fiber.
\[GITfamily\] There is an isomorphism $\widetilde{{\mathscr X}}\sslash G \simeq \mathbbm A^1\times \overline{A_X}$ over $\overline{A_X}$.
Recall the heighest weight decomposition $F[X] = \bigoplus_\lambda F[X]_\lambda$, where $\lambda$ ranges in a monoid $\Lambda_X^+\simeq \mathbb N$ of weights of $A_X$. Notice that the highest-weight modules $F[X]_\lambda$ are necessarily self-dual; indeed, twisting the action of $G$ on $X=H\backslash G$ by a Chevalley involution does not change its isomorphism class as a $G$-variety (because $H$ is reductive), hence preserves the lattice $\Lambda_X^+$. Thus, $(F[X]_\lambda\otimes F[X]_\lambda)^G=F$, where $G$ here acts diagonally.
This gives the structure of a graded vector space to $$F[X\times X]^G = \bigoplus_{\lambda, \mu} (F[X]_\lambda\otimes F[X]_\mu)^G = \bigoplus_{\lambda} (F[X]_\lambda\otimes F[X]_\lambda)^G =\bigoplus_{\lambda} F,$$ which also corresponds to a filtration of rings, with associated graded ${{\operatorname{gr}}}F[X\times X]^G = F[\overline{T_1}]$, where $\overline{T_1}$ is the image, in the grading, of a nonzero element $T_1$ of $F$ in the copy labelled by the first nontrivial element of $\Lambda_X^+$.
But this shows that $T_1\in F[X\times X]^G$ generates the ring freely, thus, $F[X\times X]^G\simeq F[T_1]$.
Moving now to the coordinate ring $F[\mathscr X]\subset F[X][t]$ of the Rees family, this argument shows that $$F[\widetilde{\mathscr X}]^G = F[T_1 t^2, t],$$ hence $\widetilde{\mathscr X}\sslash G \simeq \mathbbm A^2 = \mathbbm A^1 \times \overline{A_X}$.
Notice that at $t=0$, this specializes to an isomorphism $X_\emptyset\times X_\emptyset \sslash G \simeq \mathbbm A^1$ which is $A_X$-equivariant when $A_X$ acts by the *square* of the generator $\lambda_1$ of $\Lambda_X^+$ on $\mathbbm A^1$ (because the action of $A_X$ on $\widetilde{\mathscr X}$ restricts to its diagonal action on the two copies of the special fiber $X_\emptyset\times X_\emptyset$). More generally, the action of $A_X$ on $\mathbbm A^1 \times \overline{A_X}={\operatorname{Spec}}F[y, t]$, where $y=T_1t^2$ as in the proof above, is given by $$a\cdot (y,t) = (\lambda_1^2(a) y, \lambda_1(a) t).$$
As mentioned, the restriction of the form $\omega_{\widetilde{\mathscr X}}$ to the fiber over any $t\in \overline{A_X}$ is a $G\times G$-invariant, nonzero volume form $\omega_t$; on the special fiber, it satisfies the integration formula: $$\label{integration-degen}\int_{X_\emptyset\times X_\emptyset} |\omega_0| = \int_{\mathbbm A^1} |c|^{2\epsilon \left<\rho_{P(X)},\check\gamma\right>} \left(\int_{L_1\backslash G} |\omega_{L_1\backslash G}|\right) dc,$$ where $\epsilon$ is as in Theorem \[thmintegrationformula\]. Indeed, the special fiber contains an open dense subset which is $A_X\times G$-equivariantly isomorphic to $A_X\times L_1\backslash G$, and which corresponds to the open Bruhat cell under the isomorphism $X_\emptyset^\bullet\times X_\emptyset^\bullet/G = H_\emptyset\backslash G/H_\emptyset$, where $H_\emptyset \simeq \ker(P(X)^-\to A_X)$, with $P(X)^-$ opposite to $P(X)$. The parabolic $P(X)^-$ is actually conjugate to $P(X)$: indeed, if $X=H\backslash G$, since $H$ is reductive there is a Chevalley involution of $G$ which fixes $H$, and hence preserves the isomorphism class of $X$ — but this turns the class of $P(X)$ to the class of $P(X)^-$ — see also the discussion in §\[ssvolformproof\]. Thus, the integration formula for the open Bruhat cell with respect to $P(X)^-$ reads: $$\int_G \Phi(g) dg = \int_{A_X} \int_{H_\emptyset\times U_{P(X)}^-} \Phi(u_1 aw u_2) d(u_1, u_2) \cdot |e^{2\rho_{P(X)}}(a)| da,$$ where $w$ is the longest element of the Weyl group, and this easily translates to . Here, we need to take into account that there is an isomorphism $$H_\emptyset\backslash G\sslash H_\emptyset \xrightarrow\sim \mathbbm A^1$$ which pulls back to the character $\frac{\gamma}{\epsilon}$ (a generator for $\Lambda_X={\operatorname{Hom}}(A_X,{\mathbbm{G}_m})$) under the sequence of maps $$A_X\to H_\emptyset A_X w H_\emptyset \hookrightarrow G\to H_\emptyset\backslash G\sslash H_\emptyset \xrightarrow\sim \mathbbm A^1.$$ Hence, the inverse of this sequence of maps (restricted to ${\mathbbm{G}_m}$) is given by the cocharacter $\epsilon\check\gamma$.
On the other hand, consider the integration formula , taking into account that the points $c_1,c_{-1}$ on $X\times X\sslash G$, expressed now in the coordinate $T_1$ as above, when we vary the parameter $t\ne 0$ become $c_{\pm 1} t^2$ in the coordinate $T_1 t^2$. The limit as $t\to 0$ must coincide with the integration formula on $X_\emptyset\times X_\emptyset$, proving the codimension formula .
Finally, for spherical roots of type $G$ we have, by Proposition \[correctrepresentative\], a nontrivial $G$-automorphism of $X$ of order $2$. Applied to the first copy of $X$ in $X\times X$, this automorphism does not preserve the diagonal $X_1$, hence has to interchange it with the unique other semisimple $G$-orbit which can have codimension larger than one, that is, with $X_{-1}$. This completes the proof of Theorem \[thmintegrationformula\].
Proof with volume forms {#ssvolformproof}
-----------------------
I will also outline a proof of the codimension formula that directly uses the volume forms that appeared in the proof of the integration formula , without using any measures or integrals on the sets of $F$-points. The argument may be of independent interest, because it indicates how the resolution $${\mathcal{R}}: {\mathbbm{P}}J_X\to X\times X$$ can be placed in the family $\widetilde{{\mathscr X}}=\mathscr X\times_{\mathbbm A^1}\mathscr X$, as $X$ degenerates. The reader can skip this subsection.
Since the result that we need for the rest of the paper has already been proven, I will only sketch the arguments, employing well-known facts from the theory of spherical varieties without many explanations. One basic fact that we will need, which follows from the local structure theorem of [@KnMotion Theorem 2.3] together with an identification [@SV Proposition 2.5.2] of $\mathscr X^\bullet$ (the union of open $G$-orbits in the fibers of $\mathscr X\to\mathbbm A^1$) with an open subset of the “normal bundle degeneration” of the wonderful compactification of $X$ — or else, follows from the local structure theorem applied to $\mathscr X^\bullet$ as an $A_X\times G$-spherical variety — is the following:
Let $\hat v\in \widehat{T^*X}^{{\operatorname{rs}}}$ be in the image of Knop’s section, $x_0=\hat\pi(\hat v)$ (where $\hat\pi$ is the projection to $X$), $S= \hat\pi(A_X\cdot \hat v)\subset X$, a “flat”, and $\mathscr S=$ the closure of $A_X\cdot S = \tau(S\times A_X)$ in $\mathscr X^\bullet$. Then, the map $$\label{tolimit}A_X\times A_X\ni (l,t)\mapsto \tau(\pi(l t^{-1}\cdot \hat v), t)$$ extends to an $L\times A_X$-equivariant isomorphism (where $L$ is the centralizer of the image $\hat\mu(\hat v)$ of $\hat v$ under the moment map) $$A_X\times \overline{A_X} \simeq \mathscr S.$$ Moreover, if $P$ is the parabolic with Levi $L$ corresponding to the polarization of $\hat v$, and $\mathring{\mathscr X}$ is the union of open $P$-orbits on the fibers of $\mathscr X$, we have $$\mathring{\mathscr X} \simeq \mathscr S \times U_P$$ under the natural action of $P=L\cdot U_P$ on the right hand side.
The map should be compared with the isomorphism ; in fact, the former is a lift of the latter. The proposition has the following corollary: Let $\Omega = \Omega_{\mathscr X/\mathbbm A^1}$ be the relative cotangent sheaf of the family $\mathscr X$, as before, identified, by abuse of notation, with its total space over $\mathscr X^\bullet$. Let $\hat\Omega^\bullet = \Omega\times_{{\mathfrak{c}}_X^*} {\mathfrak{a}}_X^*$; then $\hat\Omega^\bullet$ is an irreducible component of the polarization $\Omega\times_{{\mathfrak{a}}^*\sslash W} {\mathfrak{a}}^*$, whose fiber over $1\in \overline{A_X}$ is the distinguished irreducible component $\widehat{T^*X}^\bullet$. However, the restriction of $\hat\Omega^\bullet$ to the special fiber will contain *two* irreducible components, with only one being the distinguished one, see Remark \[rmkdist\] below.
\[corlimit\] The map $$\widehat{T^*X}^{{\operatorname{rs}}}\times A_X \ni (\hat v, t)\mapsto \hat\tau(t^{-1}\hat v, t),$$ where $\hat\tau$ is the lift of the morphism $\tau$ to polarized cotangent bundles, extends to an open embedding $$\widehat{T^*X}^{{\operatorname{rs}}}\times \overline{A_X} \hookrightarrow \hat\Omega^{\bullet,{{\operatorname{rs}}}},$$ whose restriction to the special fiber has image equal to the distinguished connected/irreducible component $\widehat{T^*X_\emptyset^\bullet}^{\bullet,{{\operatorname{rs}}}}$.
\[rmkdist\] Notice that any $\hat v\in \widehat{T^*X}^{\bullet,{{\operatorname{rs}}}}$ belongs to the image of Knop’s section in our case, by Proposition \[rsinKnop\]. Elements of $\hat\Omega^{\bullet,{{\operatorname{rs}}}}$ over $t\in \overline{A_X}$ can be identified with pairs $(v,P)$ with $v\in T^*(\mathscr X^\bullet_t)^{{\operatorname{rs}}}$, $P\in {\mathcal{B}}_X$, and $\mu(v)\in {\mathfrak{u}}_P^\perp$. Letting $\pi$ denote the projection from the cotangent bundle to the base $\mathscr X^\bullet_t$, for $t\ne 0$ Proposition \[rsinKnop\] implies that $\pi(v)$ is always in the *open* $P$-orbit, while for $t=0$ the point $\pi(v)$ can belong either to the open or to the closed $P$-orbit (Bruhat cell), giving rise to the two components in the fiber of $\hat\Omega^\bullet$. The distinguished component of the corollary above corresponds to the open $P$-orbit.
Now we proceed to putting the resolution ${\mathcal{R}}$ in a family. First, observe that the construction of the map ${\mathcal{R}}$ does not produce a resolution of the special fiber $X_\emptyset\times X_\emptyset$; the reason is that Proposition \[arounddiagonal\] fails in this case, and the integration of the invariant collective motion, which in the case of $X_\emptyset$ coincides with the action of $A_X$, just takes $X_\emptyset^{{\operatorname{diag}}}$ to the closed “Bruhat cell” $$(1\times A_X)\cdot X_\emptyset^{{\operatorname{diag}}}\subset X_\emptyset \times X_\emptyset$$ which, in the rank-one case that we are considering, maps to the “zero” point of $X_\emptyset \times X_\emptyset\sslash G \simeq \mathbbm A^1$. (The proof of Proposition \[arounddiagonal\] fails in this case, because the invariant moment map $T^*X_\emptyset\to {\mathfrak{a}}^*_X$ has nontrivial kernel on every fiber.) Hence, the construction of the resolution ${\mathcal{R}}$ cannot be performed over the whole family.
An idea that suggests what to do is to view one of the two factors in $X\times X$ as “dual” to the other, denoted by $X^\vee$, and, similarly, its horospherical space $X_\emptyset^\vee$ as “dual” to $X_\emptyset$. By the “dual” of a $G$-variety $X$, I mean the $G$-variety that is isomorphic to $X$ as a variety, but with the action of $G$ twisted by a Chevalley involution: $$x^\vee \cdot g := x\cdot g^c,$$ where $x^\vee$ denotes the point on $X^\vee$ corresponding to $x\in X$ under the isomorphism $X\xrightarrow\sim X^\vee$, and $c$ is an involution on $G$ in the conjugacy class of Chevalley involutions.
Of course, for $X=H\backslash G$ affine homogeneous there is a Chevalley involution which preserves $H$, and therefore $X^\vee\simeq X$, $G$-equivariantly. This also proves that $X_\emptyset^\vee \simeq X_\emptyset$; in particular, the opposite parabolic $P(X)^-$ is conjugate to $P(X)$, and the kernel of the quotient of universal Cartans $A\to A_X$ is stable under the action of the longest Weyl group element of $G$.
The drawback of viewing $X$ and $X^\vee$ as “in principle” different $G$-varieties is that the notion of “diagonal” in $X\times X^\vee$, which was the starting point of our resolution ${\mathcal{R}}$, disappears. But this is precisely what we need when passing to the limit, where the diagonal $X_\emptyset^{{\operatorname{diag}}}$ maps to the point $0\in \mathbbm A^1 = X_\emptyset\times X_\emptyset\sslash G$.
Thinking of the second copy as $X^\vee$ suggests that it is more natural, when polarizing cotangent vectors, to choose *opposite* polarizations in the two copies, that is: polarizations which are conjugate by the longest element of the Weyl group. Unless $X$ is a torus quotient of $G$, the action of the longest Weyl group element on $A_X$ is nontrivial (because the kernel of $A\to A_X$ is not contained in the Cartan of the derived group of $G$), hence in the rank-one case that we are considering, it has to coincide with the (inversion) action of the nontrivial element $w_\gamma\in W_X$.
To proceed with the argument, and in order to avoid any confusion between the “equivariant” and the “Chevalley-twisted” identification of $X^\vee$ with $X$, I will not use the notation $X^\vee$ from now on — but the reader should keep the idea of dual spaces in mind. Instead, I will twist the polarization $\widehat{T^*X}$ on the second copy by the action of the nontrivial element $w_\gamma\in W_X$ on ${\mathfrak{a}}_X^*$: $\lambda\mapsto {^{w_\gamma}\lambda}= -\lambda$. Namely, define a “polarized union of conormal bundles” (to the $G$-orbits) $$\widehat{N^*(X\times X)}:= \widehat{T^*X}^\bullet \times_{\hat{\mathfrak{g}}_X^*,w_\gamma} \widehat{T^*X}^\bullet,$$ where the appearance of $w_\gamma$ in the fiber product means that the map $$\widehat{T^*X}^\bullet\to \hat{\mathfrak{g}}_X^* = {\mathfrak{g}}_X^*\times_{{\mathfrak{c}}_X^*} {\mathfrak{a}}_X^*$$ on the second copy is twisted by $w_\gamma$. The regular semisimple points $\widehat{N^*(X\times X)}^{{\operatorname{rs}}}$ can be identified with quadruples $$(v_1, P_1, v_2, P_2),$$ with $v_1, v_2\in T^*X^{{\operatorname{rs}}}$, $\mu(v_1)=\mu(v_2) \in {\mathfrak{u}}_{P_1}^\perp\cap {\mathfrak{u}}_{P_2}^\perp$, and $P_2\sim P(X)$ *opposite* to $P_1$.
We similarly define this polarization for every fiber of the map $\mathscr X^\bullet\to \overline{A_X}$. We also define $$\widehat{N^*\widetilde{\mathscr X^\bullet}} = \hat\Omega^\bullet \times_{\hat{\mathfrak{g}}_X^*,w_\gamma} \hat\Omega^\bullet,$$ the total space of which is a variety living over $\widetilde{\mathscr X^\bullet}= \mathscr X^\bullet\times_{\overline{A_X}} \mathscr X^\bullet$, which restricts to the above polarized union of conormal bundles on each fiber.
Start with the embedding $\widehat{T^*X}^\bullet\overset{\Delta}\hookrightarrow \widehat{N^*(X\times X)}$ which is obtained by applying the nontrivial element $w_\gamma\in W_X$ to the second copy in the diagonal embedding $\widehat{T^*X}^\bullet\hookrightarrow \widehat{T^*X}^\bullet\times_{\hat{\mathfrak{g}}_X^*} \widehat{T^*X}^\bullet$. We will denote the image of $\Delta$ by $\widehat{T^*X}^\Delta$.
We now pass to the limit afforded by Corollary \[corlimit\] in *both* copies. Having twisted the polarized moment map in the second copy by $w_\gamma$ means that the diagonal action of $A_X$ on $\widehat{N^*(X\times X)}^{{\operatorname{rs}}}$ is lifted from the *anti-diagonal* action of Knop’s group scheme $J$ on $T^*X\times_{{\mathfrak{g}}_X^*} T^*X$: $$j\cdot (v_1, v_2)= (jv_1, j^{-1} v_2).$$ We now have:
The map $$\label{infamily}\widehat{T^*X}^{\Delta,{{\operatorname{rs}}}} \times A_X \times A_X \ni (\hat v, t_1, t)\mapsto \hat\tau((t_1, 1)\cdot (t^{-1})^{{\operatorname{diag}}}\hat v, t)$$ extends to an open embedding $$\widehat{T^*X}^{\Delta,{{\operatorname{rs}}}} \times A_X \times \overline{A_X} \hookrightarrow {\widehat{N^*\widetilde{\mathscr X^\bullet}}}^{{\operatorname{rs}}},$$ whose restriction to the fiber over any $t\in \overline{A_X}$ has image equal to the connected/irreducible component of $\widehat{N^*(\mathscr X_t\times \mathscr X_t)}^{{\operatorname{rs}}}$ represented by quadruples $$(v_1, P_1, v_2, P_2)$$ with $v_1, v_2\in T^*X^{{\operatorname{rs}}}$, $\mu(v_1)=\mu(v_2) \in {\mathfrak{u}}_{P_1}^\perp\cap {\mathfrak{u}}_{P_2}^\perp$, $P_2\sim P(X)$ opposite to $P_1$ and $\pi(v_i)$ in the *open* $P_i$-orbit (for $i=1,2$).
The embedding $$\widehat{T^*X}^{\Delta,{{\operatorname{rs}}}} \times \overline{A_X} \hookrightarrow {\widehat{N^*\widetilde{\mathscr X^\bullet}}}^{{\operatorname{rs}}},$$ with image in the component of quadruples $(v_1, P_1, v_2, P_2)$ as in the proposition, follows from Corollary \[corlimit\]. Now, this component (of the fiber over any $t\in \mathbbm A^1$) is an $A_X\times A_X$-torsor over $\hat{\mathfrak{g}}_X^{*,{{\operatorname{rs}}}}$ by Corollary \[polarizedtorsor\] — easily extended to the horospherical case —, thus it is an $A_X$-torsor over the image of $\widehat{T^*X}^{\Delta,{{\operatorname{rs}}}}\times\{t\}$, under the action of $A_X$ on the first (or second) copy. The result follows.
Now consider the composition of the maps $$\hat{\mathscr R} :{\mathbbm{P}}\widehat{T^*X}^{\Delta,{{\operatorname{rs}}}} \times A_X \times \overline{A_X}\to {\mathbbm{P}}{\widehat{N^*\widetilde{\mathscr X^\bullet}}}^{{\operatorname{rs}}}\to \widetilde{\mathscr X}.$$ The pullback of the form $\omega_{\mathscr X}$ to ${\mathbbm{P}}\widehat{T^*X}^{\Delta,{{\operatorname{rs}}}} \times A_X \times \overline{A_X}$ restricts to a volume form like on the fiber over any $t\ne 0$, except that, due to the presence of $(t^{-1})^{{\operatorname{diag}}}$ in the definition of the map , the points $\pm 1\in A_X$ are now shifted by a factor of $t^2$. Identifying $A_X\simeq {\mathbbm{G}_m}$ through the generating character $\frac{\gamma}{\epsilon}$, the restriction of $\hat{\mathscr R}^*\omega_{\mathscr X}$ to the fiber over $t\ne 0$ is equal to $$(a-t^2)^{d_1-1} (a+t^2)^{d_{-1}-1} a^{1-\frac{d_1+d_{-1}}{2}} \cdot \omega_{A_X}\wedge \omega_{L_1\backslash G},$$ in the notation of . (Recall that ${\mathbbm{P}}\widehat{T^*X}^{\Delta,{{\operatorname{rs}}}} \simeq L_1\backslash G$ is a single $G$-orbit.)
Such a form restricts to an $A_X$-eigenform with eigencharacter $|\bullet|^\frac{d_1+d_{-1}-2}{2}$ on the special fiber ${\mathbbm{P}}\widehat{T^*X}^{\Delta,{{\operatorname{rs}}}} \times A_X \times \{0\}$, when $A_X$ acts only on the second factor, and is identified with ${\mathbbm{G}_m}$ through the generating character $\frac{\gamma}{\epsilon}$. On the other hand, the analog of in terms of volume forms states that the restriction of the pullback to the special fiber has to be an $A_X$-eigenform with eigencharacter $e^{2\rho_{P(X)}}$. We conclude that $$\frac{d_1+d_{-1}-2}{2} \cdot \frac{\gamma}{\epsilon} = 2\rho_{P(X)},$$ which is equivalent to the codimension formula .
Schwartz measures {#sec:Schwartzmeasures}
=================
We are ready to consider the pushforward of Schwartz measures: $$\label{pushforward} \mathcal S(X\times X)\to {{\operatorname{Meas}}}({\mathfrak{C}}_X),$$ whose image we have denoted by $\mathcal S(X\times X/G)$. (Recall that ${\mathfrak{C}}_X$ denotes the invariant-theoretic quotient $X\times X\sslash G\simeq A_X\sslash W_X$.) From now on, we assume that $X$ is not only a “correct representative” in its class modulo $G$-automorphisms, but also that stabilizers are split; thus, by Lemma \[surjectiveonpoints\], the map $X\times X\to {\mathfrak{C}}_X$ is surjective on $F$-points.
In this section we will obtain as much information as possible from abstract principles about the space $\mathcal S(X\times X/G)$, using the blowup ${\mathbbm{P}}J_X$. We use the blowup in the way that it is used in Igusa integrals: as a resolution of the map $X\times X\to {\mathfrak{C}}_X$, in the sense that preimages of points are normal crossings divisors, see Proposition \[isblowup\].
We will actually be working mainly with the linearizations of this $G$-space. The main result of the section is Theorem \[pushfthm\], leaving us only a certain linear combination of scalars to compute in the next section. Strictly speaking, the techniques of the next section are sufficient to obtain the main results, Theorems \[Xtheorem\] and \[maintheorem\], but using the resolution puts the results into a conceptual context, up to the computation of a linear combination of coefficients.
Generalities on Schwartz measures {#ssgeneralities-Schwartz}
---------------------------------
Before we proceed, we need some preliminaries on cosheaves of Schwartz measures.
Let $Z$ be a smooth variety, and $D\subset Z$ a divisor. We let $C^\infty(\bullet,D)$ denote the sheaf of functions on the $F$-points of $Z$ which, locally, are of the form $\Phi(z) |\epsilon_D(z)|$, where $\Phi$ is a smooth function and $\epsilon_D$ is a local generator for the divisor $D$. Informally, we consider such functions as “smooth sections of the complex line bundle $|\mathcal L_D|$ associated to $D$”. Consider the restricted topology of semialgebraic sets on the $F$-points of $Z$. (The $F$-points of Zariski open subsets will be enough, for our purposes.) We can define a cosheaf $\mathcal S(\bullet, D)$ of Schwartz measures valued in $|\mathcal L_D|$, namely, those measures which are the product of *a nowhere vanishing smooth measure of polynomial growth (together with its derivatives)* and *a section of $C^\infty(\bullet,D)$ of rapid decay* (together with its derivatives; compactly supported in the non-Archimedean case). These notions make sense: if $U\subset Z$ is such an open subset, a measure of polynomial growth on $U$ (together with its derivatives) is a measure which, over a finite cover of $U$ by open affine subsets, in affine coordinates $(x_1, \dots, x_n)$ can be written as $\Phi(x_1,\dots, x_n) dx_1\cdots dx_n$, where $\Phi$ is a function such that $|T\Phi|$ is bounded by the maximal absolute value of a finite set of polynomials in the $x_i$’s, for every polynomial differential operator $T$. And, a section of $C^\infty(U,D)$ of rapid decay (in the Archimedean case, together with its derivatives) is simply a section which, on the $F$-points of a smooth compactification $\bar Z$ of $Z$, vanishes in the complement of $U$, together with all its derivatives.
For a closed subset $Y\subset Z$ (“closed” means semialgebraic, again, but the reader can restrict their attention to Zariski closed) we define the *stalk* $\mathcal S_Y(\bullet,D)$ as the cosheaf on $Z$, supported on $Y$, whose sections over an open $U\subset Z$ are the quotient $$\mathcal S_Y(U,D)= \mathcal S(U,D)/\mathcal S(U\smallsetminus Y,D).$$ The *fiber* $\overline{\mathcal S_Y(\bullet,D)}$ is the cosheaf whose sections over $U\subset Z$ are the quotient $$\overline{\mathcal S_Y(U,D)} = \mathcal S(U,D)/C^\infty_{{\operatorname{temp}}}(U,[Y])\mathcal S(U,D),$$ where $C^\infty_{{\operatorname{temp}}}(U,[Y])$ denotes the ideal of those tempered (i.e., of polynomial growth together with their polynomial derivatives) smooth functions that vanish on $Y$. In the non-Archimedean case, the natural map $\mathcal S_Y(\bullet,D)\to \overline{\mathcal S_Y(\bullet,D)}$ is, of course, an isomorphism.
Our analysis of the pushforward starts from the following:
Let $Z\to Y$ be a smooth map of smooth varieties which is surjective on $F$-points. Then the pushforward of $\mathcal S(Z)$ is equal to $\mathcal S(Y)$.
This is standard, see, e.g., [@SaStacks Proposition 3.1.2].
Let $U\subset X\times X$ be the preimage of $\mathring{\mathfrak{C}}_X:={\mathfrak{C}}_X\smallsetminus\{[1], [-1]\}$. Then the pushforward of $\mathcal S(U)$ is the space of Schwartz measures $\mathcal S(\mathring{\mathfrak{C}}_X)$.
Indeed, the map $X\times X\to {\mathfrak{C}}_X$ is smooth over $\mathring{\mathfrak{C}}_X$ by Proposition \[genericisomorphism\], and the map is surjective on $F$-points by Lemma \[surjectiveonpoints\].
Thus, our remaining task is to determine the behavior of the elements of $\mathcal S(X\times X/G)$ close to the points $[\pm 1]\in {\mathfrak{C}}_X$. To this end, we can linearize the problem: Let $x\in X_{\pm 1}$, with $H_{\pm 1}$ its stabilizer in $G$ and $V_{\pm 1}$ its fiber in the conormal bundle $N^*_{X_\pm 1}$, so that $V_{\pm 1}\sslash H_{\pm 1} = N^*_{X_{\pm 1}}\sslash G = {\mathfrak{c}}_X^*$, by Proposition \[GITquotients\]. We let $\mathcal S(V_{\pm 1}/H_{\pm 1})$ be the pushforward of $\mathcal S(V_{\pm 1})$ under the map $V_{\pm 1}\to {\mathfrak{c}}_X^*$.
\[linearization\] There is an $F$-analytic isomorphism between a neighborhood $U_1$ of $0\in {\mathfrak{c}}_X^*(F)$ (in the Hausdorff topology on $F$-points) and a neighborhood $U_2$ of $[\pm 1]\in {\mathfrak{C}}_X(F)$, such that the space of restrictions to $U_1$ of elements of $\mathcal S(V_{\pm 1}/H_{\pm 1})$ is equal, under this isomorphism, to the space of restrictions to $U_2$ of the pushforwards of elements of $\mathcal S(X\times X)$ supported in a certain $G(F)$-stable neighborhood of $x$.
The restriction to a $G(F)$-stable neighborhood of $x$ is because the map $G(F)\to X_{\pm 1}(F)=H_{\pm 1}\backslash G(F)$ sending $g$ to $x\cdot g$ may not be surjective on $F$-points. Eventually, as we will see, the normal fibers $V_{\pm 1}$ of all points on $X_{\pm 1}$ contribute the same germs of pushforward measures, so this detail will not matter.
The pushforward map $\mathcal S(V_{\pm 1})\to \mathcal S(V_{\pm 1}/H_{\pm 1})$ factors through the $H_{\pm 1}$-coinvariants of $\mathcal S(V_{\pm 1})$, and similarly the pushforward map $\mathcal S(X\times X)\to \mathcal S(X\times X/G)$ factors through the $G$-coinvariants of $\mathcal S(X\times X)$.
Fix an isomorphism ${\mathfrak{c}}_X^*\simeq \mathbbm A^1$ and use the resulting nondegenerate quadratic form (Proposition \[genericisomorphism\]) $V_{\pm 1}\to {\mathfrak{c}}_X^*\simeq\mathbbm A^1$ to identify $V_{\pm 1}$ with its linear dual, the fiber of the *normal* bundle.
By Luna’s étale slice theorem [@Lunaslice], there is an $H_{\pm 1}$-stable subvariety $W\subset X\times X$ containing $x$, and a Cartesian diagram of pointed spaces with étale diagonal maps: $$\label{Luna}
\xymatrix{
& (W\times^{H_{\pm 1}}G,x) \ar[dl]\ar[dr]\ar[dd]
\\
(V_{\pm 1}\times^{H_{\pm 1}} G,0) \ar[dd] && (X\times X,x) \ar[dd] \\
& W\sslash H_{\pm 1}\ar[dl]\ar[dr]
\\
V\sslash H_{\pm 1} = {\mathfrak{c}}_X^* && {\mathfrak{C}}_X= X\times X\sslash G.}$$
The étale diagonals induce isomorphisms between neighborhoods $U_1$ of $0\in {\mathfrak{c}}_X^*(F)$ and $U_2$ of $[\pm 1]\in {\mathfrak{C}}_X(F)$, and [@SaStacks Corollary 4.2.1] implies that such a diagram induces an isomorphism between the coinvariant spaces over these neighborhoods; more precisely (since we are not treating $V_{\pm 1}/H_{\pm 1}$ as a stack), between the $H_{\pm 1}(F)$-coinvariants of elements of $\mathcal S(V_{\pm 1})$ supported in the preimage of $U_1$, and the $G(F)$-coinvariants of elements of $\mathcal S(X\times X)$ supported in the intersection of the preimage of $U_2$ with the $G(F)$-orbit of the Luna slice $W(F)$. In particular, the pushforwards of those measures to $U_1\simeq U_2$ coincide.
From now on, and until the end of this section, we denote $V_{\pm 1}$ simply by $V$, and $H_{\pm 1}$ simply by $H$. The reader should not confuse that, in the case of $X_{-1}$, with the representation $X=H\backslash G$ used elsewhere in this paper. The dimension $d_{\pm 1}$ of $V$ will be denoted simply by $d$.
Pullback to the blowup
----------------------
Let ${\mathcal{R}}_V:\tilde V\to V$ be the blowup of $V$ at the origin, and $E$ the preimage of $0$ (the exceptional divisor); it is the linear analog of the resolution ${\mathcal{R}}: {\mathbbm{P}}J_X \to X\times X$, and we have an extension of Luna’s Cartesian diagram, still with étale diagonals,
$$\label{Luna2}\xymatrix{
& \tilde W\times^{H} G \ar[dl]\ar[dr]\ar[dd]
\\
\tilde V\times^{H} G \ar[dd] && {\mathbbm{P}}J_X \ar[dd] \\
& W\times^{H} G \ar[dl]\ar[dr]\ar[dd]
\\
V\times^{H} G \ar[dd] && X\times X \ar[dd] \\
& W\sslash H\ar[dl]\ar[dr]
\\
V\sslash H && {\mathfrak{C}}_X= X\times X\sslash G.}$$
However, we will not use this diagram, as we will work only on the linearization, based on Proposition \[linearization\].
Pullback of Schwartz measures under the blowup ${\mathcal{R}}_V$ gives rise to a closed embedding: $$\label{Schwartzembedding}
\mathcal S(V)\overset{{\mathcal{R}}_V^*}{\hookrightarrow} \mathcal S(\tilde V, (d-1)[E]).$$
The space on the right is the space of Schwartz measures valued in the complex line bundle defined by the divisor $(d-1)[E]$, introduced in §\[ssgeneralities-Schwartz\].
This follows from writing any Schwartz measure, locally, as $f=\Phi\cdot |\omega|$, where $\Phi$ is a Schwartz function and $\omega$ a Haar volume form on $V$, and taking into account that the divisor of ${\mathcal{R}}_V^*\omega$ is $(d-1)[E]$.
\[stalks\] The fiber $\overline{\mathcal S_E(\tilde V, (d-1)[E])}$ of the Schwartz cosheaf $\mathcal S(\tilde V,(d-1)[E])$ over the exceptional divisor $E$ is canonically isomorphic to the space ${{\operatorname{Meas}}}^\infty(V\smallsetminus\{0\})^{{\mathbbm{G}_m}, |\bullet|^d}$ of smooth measures on $V\smallsetminus\{0\}$ which are eigenmeasures for the multiplicative group of dilations with eigencharacter $|\bullet|^d$. [^5]
In particular, $\overline{\mathcal S_E(\tilde V, (d-1)[E])}$ contains a canonical line $\overline{\mathcal S_E(\tilde V, (d-1)[E])}_{{\operatorname{Haar}}}$, that will be called the line of “Haar” elements, corresponding under this isomorphism to Haar measures on $V$. This line is the image of the subspace $\mathcal S(V)$ in the fiber. In the non-Archimedean case, where the fiber and the stalk coincide, the subspace $\mathcal S(V)$ coincides with the space of those elements of $\mathcal S(\tilde V, (d-1)[E])$ whose image in the stalk over the exceptional divisor lies in $\overline{\mathcal S_E(\tilde V, (d-1)[E])}_{{\operatorname{Haar}}}$.
In the Archimedean case, there is a distinguished $\mathcal C^\infty_0(V)$-submodule $\mathcal S_E(\tilde V, (d-1)[E])_{{\operatorname{Haar}}}$ of the stalk $\mathcal S_E(\tilde V, (d-1)[E])$ (where $\mathcal C^\infty_0(V)$ denotes the stalk at $0\in V$ of the ring of smooth functions), the one generated by Schwartz measures on $V$ which in the neighborhood of the origin are equal to a Haar measure, and, again, the subspace $\mathcal S(V)$ coincides with the space of those elements of $\mathcal S(\tilde V, (d-1)[E])$ whose image in the stalk over the exceptional divisor lies in $\mathcal S_E(\tilde V, (d-1)[E])_{{\operatorname{Haar}}}$.
The blowup $\tilde V$ is canonically the total space of the tautological line bundle over ${\mathbbm{P}}V=$ the exceptional divisor; let $\pi:\tilde V\to {\mathbbm{P}}V$ be the projection to the zero section.
Any element of $\mathcal S(\tilde V, (d-1)[E])$ can be written as a product $\Phi {\mathcal{R}}_V^* dv$, where $\Phi$ is a Schwartz function on $\tilde V$ and $dv$ is a Haar measure on $V$. The fiber can thus be identified with the space of measures of the form $v\mapsto f(\pi(v)) {\mathcal{R}}_V^* dv$, where $f$ is a Schwartz function on $E={\mathbbm{P}}V$ (the restriction of $\Phi$ to the exceptional divisor). This is canonically isomorphic to the space of measures described in the statement of the lemma.
The rest of the statements are even more tautological.
Pushforward to ${\mathfrak{c}}_X^*$
-----------------------------------
Now we consider pushforwards of Schwartz measures to ${\mathfrak{c}}_X^*$.
\[propbigimage\] The image of the pushforward map $$\mathcal S(\tilde V, (d-1)[E]) \to {{\operatorname{Meas}}}({\mathfrak{c}}_X^*)$$ consists precisely of those measures which are smooth away from $0$, of rapid decay (together with their polynomial derivatives) at infinity (compactly supported, in the non-Archimedean case), and in a neighborhood of $0$ have the form: $$\label{pushforwards-all} C_0(\xi) + |\xi|^{\frac{d}{2}-1} \sum_{\eta\in \widehat{F^\times/(F^\times)^2}} C_\eta(\xi) \cdot \eta(\xi),$$ where $\eta$ runs over all quadratic characters of $F^\times$, $C_0$ and the $C_\eta$’s are smooth measures, and $\xi$ is a coordinate on ${\mathfrak{c}}_X^*\simeq \mathbbm A^1$, except when $|\xi|^{\frac{d}{2}-1}\eta(\xi)$ is smooth for some $\eta$, that is:
- when $\frac{d}{2}-1 = 0$ and $\eta=1$, or
- when $F=\mathbb R$ and $\frac{d}{2}-1$ is an even integer and $\eta$ is trivial, or an odd integer and $\eta$ is the sign character, or
- when $F=\mathbb {\mathbb{C}}$ (so, $\eta=1$) and $d$ is even,[^6]
in which case the term $|\xi|^{\frac{d}{2}-1} \eta(\xi)\cdot C_\eta(\xi)$ should be replaced by $|\xi|^{\frac{d}{2}-1} \eta(\xi) \log|\xi|\cdot C_\eta(\xi)$. In the Archimedean case, this map is continuous with respect to the obvious Fréchet topology on these measures, determined by Schwartz seminorms away from zero, and by absolute values of the derivatives of the functions $\frac{C_0}{d\xi}$, $\frac{C_\eta}{d\xi}$ at zero.
I claim that, locally around any point of the exceptional divisor, there is a coordinate chart $(\epsilon_E, x_1, \dots, x_{d-1})$, where $\epsilon_E=0$ is a local equation for $E$, such that the map $\xi:\tilde V\to {\mathfrak{c}}_X^*\simeq \mathbbm A^1$ is given by $\xi=\epsilon_E^2 x_1$. Indeed, this is seen immediately by writing the split quadratic form $V\to {\mathfrak{c}}_X^*$ in coordinates.[^7]
The asserted form of the pushforward of Schwartz measures under such a map is quite a standard result. One way to prove it is using Mellin transforms: The Mellin transform of a pushforward measure $\xi_! f$ with respect to the variable $\xi$, with $f = |\epsilon_E|^{d-1}\cdot C(\epsilon_E, x_1, \dots, x_{d-1})$, where $C$ is a Schwartz measure in $d$ variables, is $$\widecheck{\xi_! f}(\chi) := \int_F \xi_! f(\xi) \chi^{-1}(\xi) = \int \bar C(\epsilon_E, x_1) |\epsilon_E|^{d-1} \chi^{-1}(\epsilon_E^2 x_1),$$ where $\bar C$ is the pushforward of $C$ with respect to the map $(\epsilon_E, x_1, \dots, x_{d-1})\mapsto (\epsilon_E, x_1)$.
This is the Tate zeta integral of a Schwartz measure in two variables, in one of the variables against the character $\chi^{-1}$ and in the other against the character $|\bullet|^{d-1} \chi^{-2}$. In the non-Archimedean case, it has poles at $\chi=|\bullet|^{-1}$ and at the points $\chi=\eta \cdot |\bullet|^{\frac{d}{2}-1}$ (double if any of these points coincide, simple otherwise), where $\eta$ ranges over all quadratic characters. Such a Mellin transform corresponds to a measure on the line which in a neighborhood of $\xi=0$ is of the form $$C_0(\xi) + |\xi|^{\frac{d}{2}-1} \sum_{\eta\in \widehat{F^\times/(F^\times)^2}} C_\eta(\xi) \cdot \eta(\xi),$$ unless $d=2$, in which case the pole at $\chi=|\bullet|^{-1}$ is double, and the corresponding singular term is of the form $C_1(\xi) \cdot \log|\xi|$.
A similar argument works in the Archimedean case, where double poles appear whenever the product $|\xi|^{\frac{d}{2}-1} \eta(\xi)$ is a smooth function of $\xi$. Here, the above Tate integral maps continuously into the appropriate “Paley–Wiener space” in the language of [@SaTransfer1 Remark 2.1.6], with the location and multiplicity of poles determined by the characters $|\xi|^{\frac{d}{2}-1} \eta(\xi)$, which corresponds to the Fréchet space of measures as in the statement of the proposition.
Our final task will be to determine the image $\mathcal S(V/H)$ of the subspace $\mathcal S(V)\hookrightarrow \mathcal S(\tilde V, (d-1)[E])$. This will be completed in the next section. We start with the following observation:
\[pushfthm\] The space $\mathcal S(V/H)$ contains the space $\mathcal S({\mathfrak{c}}_X^*)$ of Schwartz measures on ${\mathfrak{c}}_X^*$.
Moreover, in the expression for the pushforward of a measure $f\in \mathcal S(\tilde V, (d-1)[E])$, the coefficients $\frac{C_\eta}{d\xi}(0)$ depend only on the image of $f$ in the fiber $$\overline{\mathcal S(\tilde V, (d-1)[E])}={{\operatorname{Meas}}}^\infty(V\smallsetminus\{0\})^{{\mathbbm{G}_m}, |\bullet|^d}$$ (see Lemma \[stalks\]). In particular, by Lemma \[stalks\], for all $f\in \mathcal S(V)$ these coefficients will lie in a one-dimensional subspace of ${\mathbb{C}}^{\widehat{F^\times/(F^\times)^2}}$.
Let $(a_\eta)_\eta$ be a vector spanning this one-dimensional subspace. Then $\mathcal S(V/H)$ is the space of those measures of the form $$\label{pushforwards} C_0(\xi) + |\xi|^{\frac{d}{2}-1} C_{{\operatorname{sing}}}(\xi) \sum_{\eta\in \widehat{F^\times/(F^\times)^2}} a_\eta \cdot \eta(\xi),$$ where $C_0, C_{{\operatorname{sing}}}$ are Schwartz measures, and the same modification as in Proposition \[propbigimage\] applies to the case where $|\xi|^{\frac{d}{2}-1}\eta(\xi)$ is smooth.
As we have seen, the complement of the origin is smooth and surjective over ${\mathfrak{c}}_X^*$, hence the image of $\mathcal S(V\smallsetminus\{0\})$ is equal to $\mathcal S({\mathfrak{c}}_X^*)$.
Hence, the germs of the measures $C_\eta$ at $0$ depend only on the image of $f$ in the stalk $\mathcal S_E(\tilde V, (d-1)[E])$.
In the non-Archimedean case, “germ of $C_\eta$” means simply the value of $\frac{C_\eta}{d\xi}$ at $0$, and, as we saw in Lemma \[stalks\], the stalk $\mathcal S_E(\tilde V, (d-1)[E])$ is identified with ${{\operatorname{Meas}}}^\infty(V\smallsetminus\{0\})^{{\mathbbm{G}_m}, |\bullet|^d}$, so the claim follows.
In the Archimedean case, we will show that the $H$-coinvariants of the stalk $\mathcal S_0(V)$ are generated over the stalk $C_0^\infty({\mathfrak{c}}_X^*)$ by any measure which is nonvanishing at the origin.[^8]
For this, consider the descending filtration of the stalk $\mathcal S_0(V)$ which defines its topology, i.e., $F^n\mathcal S_0(V)=$ the germs of smooth measures $f=\Phi dv$ (where $dv$ is a Haar measure) such that all partial derivatives of $\Phi$ of order $<n$ vanish at the origin. By the ${\mathbbm{G}_m}$-action on $V$, this filtration corresponds to a grading on the dense subspace of ${\mathbbm{G}_m}$-finite germs. Notice that $X\times X$ admits a $G$-invariant measure, and therefore the Haar measure $dv$ is $H$-invariant; thus, we can choose such a measure to identify the $H$-modules of functions and measures. The graded piece $F^n\mathcal S_0(V)/F^{n+1}\mathcal S_0(V)$ is then identified with ${\operatorname{Sym}}_{\mathbb{R}}^n(V^*)\otimes_{\mathbb{R}}{\mathbb{C}}$, and therefore the $H$-coinvariants of the stalk are $$\begin{aligned}
\label{sympowers}\nonumber \mathcal S_0(V)_H = \lim_{\underset{n}\leftarrow} \left({\operatorname{Sym}}_{\mathbb{R}}^n(V^*)\otimes_{\mathbb{R}}{\mathbb{C}}\right)^{H} &=& \lim_{\underset{n}\leftarrow} \left({\operatorname{Sym}}_{\mathbb{C}}^n(V^*\otimes_{\mathbb{R}}{\mathbb{C}})\right)^{H_{\mathbb{C}}} = \\ &=&\begin{cases} {\mathbb{C}}[[\xi]], & \mbox{ if } F={\mathbb{R}},\\ {\mathbb{C}}[[\xi, \bar\xi]],& \mbox{ if } F={\mathbb{C}},\end{cases}\end{aligned}$$ where we have treated $H$ as a real group, so that $H_{\mathbb{C}}$ denotes its complexification.
The space ${\mathbb{C}}[[\xi]]$, if $F={\mathbb{R}}$, and ${\mathbb{C}}[[\xi, \bar\xi]]$, if $F={\mathbb{C}}$, is naturally identified with the stalk $C^\infty_0({\mathfrak{c}}_X^*)$ at zero of the ring of smooth functions. The calculation is equivariant with respect to the action of this stalk. Thus, the above calculation shows that the $H$-coinvariants of $\mathcal S_0(V)$ are freely generated over $C^\infty_0({\mathfrak{c}}_X^*)$ by the germ of any element $\Phi dv$ with $\Phi(0)\ne 0$.
Thus, the germs of pushforwards will also be generated, over $C^\infty_0({\mathfrak{c}}_X^*)$ and up to smooth measures, by the germ of the pushforward of any such measure $\Phi dv$. Consider such a measure with $\Phi(v)$ constant (and $\ne 0$) close to the origin. The pushforward map is ${\mathbbm{G}_m}$-equivariant (with respect to the quadratic action on ${\mathfrak{c}}_X^*$), hence, in terms of the expression , the germ of the pushforward of such a measure $\Phi dv$ will be of the form $$C_0(\xi) + |\xi|^{\frac{d}{2}-1} \sum_{\eta\in \widehat{F^\times/(F^\times)^2}} C_\eta(\xi) \cdot \eta(\xi),$$ where the measures $C_\eta$ are *constant* around $\xi=0$. Thus, the image of the Haar *stalk* $\mathcal S_E(\tilde V, (d-1)[E])_{{\operatorname{Haar}}}$ (in the notation of Lemma \[stalks\]) in the singular quotient of the stalk of $\mathcal S(V/H)$ at zero (i.e., ignoring the term $C_0$) is of the form $$\label{Haarstalk} |\xi|^{\frac{d}{2}-1} C_{{\operatorname{sing}}}(\xi) \sum_{\eta\in \widehat{F^\times/(F^\times)^2}} a_\eta \cdot \eta(\xi),$$ and in particular is completely determined by the coefficients $\alpha_\eta$, which depend only on the image of an element in the *fiber* $\overline{\mathcal S_E(\tilde V, (d-1)[E])}_{{\operatorname{Haar}}}$.
Therefore, we are left with computing the ratio between the coefficients $a_\eta$, which correspond to the singular part of the pushforward of a measure on $V$ which restricts to a Haar measure in a neighborhood of the origin.
Determination of the germs {#sec:germs}
==========================
Reduction to the basic cases
----------------------------
We will actually not compute the ratio of the coefficients $a_\eta$ explicitly in all cases, but rather prove, by reducing to an ${\operatorname{SL}}_2$- or ${\operatorname{PGL}}_2$-example, that they match the contributions of the Kloosterman germs under the transfer operator from the Kuznetsov formula. The cases $d=$ even and $d=$ odd will be quite different, as we will see. We fix throughout the isomorphism $\xi:{\mathfrak{c}}_X^*\xrightarrow\sim \mathbbm A^1$.
The main result of this subsection is Proposition \[reductiontobasic\], which says that pushforwards of Schwartz measures for a $d$-dimensional split quadratic space (under the quadratic map) are equal to twisted pushforwards for a two- or three-dimensional quadratic space (same parity as $d$); this will complete the proof of Theorem \[Xtheorem\].
The two- or three-dimensional quadratic space $V_2$ is obtained from $V$ by choosing a maximal isotropic subspace $M\subset V$ and a hyperplane $M'\subset M$; then $V_2 = M'^\perp/M'$. Let us go through the argument carefully:
Fix such a maximal isotropic subspace $M$. Since $V$ is split, the orthogonal complement $M^\perp$ is either equal to $M$ (when $d$ is even), or contains $M$ as a hyperplane (when $d$ is odd). The quotient $V/M^\perp$ is isomorphic to the linear dual $M^*$ through the quadratic form, and the parabolic $P\subset {{\operatorname{SO}}}(V)$ stabilizing $M$ surjects to ${\operatorname{GL}}(M^*)$.
The integration (pushforward) map $\mathcal S(V)\to {\mathbb{C}}$ factors through surjective pushforward maps: $$\label{sequence} \mathcal S(V)\to \mathcal S(V/M^\perp) \to \mathcal S({\mathbbm{P}}M^*) \to {\mathbb{C}}.$$
Let $M'\in {\mathbbm{P}}M^*$, identified (and denoted by the same letter) with a hyperplane in $M$. Let $P'\subset P$ be the stabilizer of the flag $M'\subset M\subset V$. The space $\mathcal S({\mathbbm{P}}M^*)$, considered as a representation of $P$, can be identified, up to a scalar which we fix, with the (unnormalized) induced representation ${\operatorname{Ind}}_{P'}^P (\delta_{P'/U_P})$, where $\delta_{P'/U_P}$ denotes the modular character of the image of $P'$ in the Levi quotient of $P$. By Frobenius reciprocity, the $P$-equivariant map $\mathcal S(V)\to \mathcal S({\mathbbm{P}}M^*)$ is given by a $(P', \delta_{P'/U_P})$-equivariant functional. The Lemma that follows determines this functional:
Let $\Phi$ be a Schwartz function on $V$, and $dv$ a Haar measure. Then, for suitable Haar measures, $$\label{PV} \int_V \Phi(v) dv = \int_{{\mathbbm{P}}M^*} \left(\int_{{\mathbbm{G}_m}} \int_{M^\perp} \Phi(av + v_1) dv_1 |a|^{\dim M} d^\times a \right) dv.$$
Notice that the expression in brackets, viewed as a function of $v\in M^*\smallsetminus\{0\}$, is $({\mathbbm{G}_m}, |\bullet|^{-d})$-equivariant, hence $dv$ denotes an invariant measure on ${\mathbbm{P}}M^*$, valued in the dual of the line bundle of $({\mathbbm{G}_m}, |\bullet|^{-d})$-equivariant functions on $M^*\smallsetminus\{0\}$. More precisely, under the action of $P'$, the expression in brackets is $\delta_{P'/U_P}$-equivariant, and $dv$ is an invariant measure on ${\mathbbm{P}}M^*$, valued in the line bundle dual to the one induced from this character of $P'$.
This lemma is just a reformulation of the sequence .
Let us reformulate the inner integral of : Fix $M'\in {\mathbbm{P}}M^*$, understood again as a hyperplane in $M$. Its preimage in $V$ under the rational map $V\to {\mathbbm{P}}M^*$ is equal to $M'^\perp \smallsetminus M^\perp$. Fix a nonzero vector $v\in M^*$ in the line corresponding to $M'$; then the functional $L: av\mapsto a$ is a linear functional on the one-dimensional space of multiples of $v$ in $M^*$ or, equivalently, a functional $$L: M'^\perp \to {\mathbbm{G}_a}.$$
The quotient $V_2:=M'^\perp/M'$ is a nondegenerate quadratic space of dimension $2$ or $3$ (same parity as $V$). Fix a Haar measure $dv'$ on $M'$, and let $\Phi\mapsto \Phi_2$ be the corresponding pushforward map (integration over cosets of $M'$ against $dv'$) $$\mathcal F (M'^\perp) \twoheadrightarrow \mathcal F(V_2),$$ where $\mathcal F$ denotes the spaces of Schwartz functions. Then the inner integral of can be written as $$\int_{V_2} \Phi_2(v_2) L(v_2)^{\dim M-1} dv_2,$$ for a suitable Haar measure $dv_2$.
Let us explicate this integral:
- If $V_2$ is a two-dimensional (split) quadratic space, then $\dim M=\frac{d}{2}$, and there are coordinates $(x,y)$ such that the quadratic form is $\xi=xy$ and the functional $L$ is $L = x$, so the integral reads: $$\label{twodim} \int_{V_2} \Phi_2(x,y) |x|^\frac{d-2}{2} dx dy.$$
- If $V_2$ is a three-dimensional (split) quadratic space, then $\dim M=\frac{d-1}{2}$, and there is an isomorphism $V_2\simeq\mathfrak{sl}_2$ with quadratic form $\xi = -\det$ and $L\begin{pmatrix} A & B \\ C & -A \end{pmatrix} = C$, so the integral reads: $$\label{threedim} \int_{V_2} \Phi_2\begin{pmatrix} A & B \\ C & -A \end{pmatrix} |C|^{\frac{d-3}{2}} dA dB dC.$$
In either case, these integrals can be disintegrated against the quadratic form, but we need to choose a section $\sigma: {\mathfrak{c}}_X^* \to V_2$ of the quadratic form, since the integrand is not invariant over the fibers. Choose this section $\sigma$ so that its image is contained in an affine line of the hyperplane $L=1$; then it is necessarily contained in the affine line $\sigma(0)+\bar M$ (where $\bar M=M/M'$, the image of $M$ in $V_2$); explicitly:
- $\sigma(\xi) = (1,\xi)$ in the coordinates above when $V_2$ is two-dimensional;
- $\sigma(\xi) = \begin{pmatrix} & \xi \\ 1 & \end{pmatrix}$ when $V_2$ is three-dimensional.
Let $B_2\subset {{\operatorname{SO}}}(V_2)$ be the stabilizer of the isotropic line $\bar M=M/M'$, and let $\delta_2^{-1}$ be the absolute value of the character by which it acts on $\bar M$. Then:
\[rewrite\] The expressions and can be written: $$\label{onV2} \int_{{\mathfrak{c}}_X^*} \int_{B_2} \Phi_2(\sigma(\xi) b) \delta_2(b)^{\dim M-1} db d\xi,$$ for a suitable right Haar measure on $B_2$.
Immediate, by calculation.
By this lemma, we have a new integration formula for $V$ in terms of the quadratic form, that includes $B_2$-orbital integrals on $V_2$, twisted by the character $\delta_2(b)^{\dim M-1}$. Thus, the integration formula for a $d$-dimensional quadratic space involves a twisted integration formula for a $2$- or $3$-dimensional quadratic space:
Let $\Phi$ be a Schwartz function on $V$, and $dv$ a Haar measure. Let $K\subset P$ be any compact subgroup such that $K\to P'\backslash P={\mathbbm{P}}M^*$ is surjective. Then, for suitable (right) Haar measures, $$\label{PV2} \int_V \Phi(v) dv = \int_{{\mathfrak{c}}_X^*} \left(\int_{K} \int_{B_2} \int_{M'} \Phi((\sigma(\xi) b + v')k ) \delta_2(b)^{\dim M-1} dv' db dk\right) d\xi.$$
Here, we have lifted the section $\sigma$ to $M'^\perp$ and the group $B_2$ to $P'$, by choosing a section $V_2\to M'^\perp$.
The integral over $K$ replaces the integral over ${\mathbbm{P}}M^*$ in ; since it is a compact integral of a smooth function, it can be moved to the interior, and the result follows by applying .
Representation-theoretically, the two inner integrals of represent a $(P', \delta_{P'/U_P})$-equivariant functional, hence a morphism $$\mathcal S(V)\to {\operatorname{Ind}}_{P'}^{{{\operatorname{SO}}}(V)}(\delta_{P'/U_P}) = {\operatorname{Ind}}_P^{{{\operatorname{SO}}}(V)} {\operatorname{Ind}}_{P'}^P(\delta_{P'/U_P}).$$ The integral over $K$ corresponds to the quotient ${\operatorname{Ind}}_{P'}^P(\delta_{P'/U_P})\to {\mathbb{C}}$ (the trivial representation), so the expression in brackets can be seen as a morphism $$\mathcal S(V)\to {\operatorname{Ind}}_P^{{{\operatorname{SO}}}(V)} ({\mathbb{C}}).$$ By the invariance of the left hand side of , this morphism is ${{\operatorname{SO}}}(V)$-invariant, hence has image in the trivial subrepresentation of ${\operatorname{Ind}}_P^{{{\operatorname{SO}}}(V)}({\mathbb{C}})$.
Let $\ell_\xi(\Phi)$ be the ${{\operatorname{SO}}}(V)$-invariant functional represented by the expression in brackets of . Comparing with the integration formula of Theorem \[thmintegrationformula-linear\] (for the special case $H={{\operatorname{SO}}}(V)$), we get: $$\ell_\xi(\Phi) = |\xi|^{\frac{d}{2}-1} O_\xi(\Phi),$$ where the $O_\xi$’s are orbital integrals on the fibers over $\xi\ne 0$, against invariant measures $d\dot{g}_\xi$ obtained by identifying all nondegenerate ${{\operatorname{SO}}}(V)$-orbits over the algebraic closure, and choosing volume forms as in Theorem \[thmintegrationformula-linear\].
Thus, we arrive at the following result about the coefficients $a_\eta$ of the expression :
\[reductiontobasic\] If $d$ is even, we have $a_\eta=0$ except for $\eta=1$, and there is an equality between the space of pushforward measures for $V\xrightarrow{\xi} \mathbbm A^1$ and the measures on $\mathbbm A^1$ of the form $$\label{twistedpfT} \xi\mapsto \left(\int_{{\mathbbm{G}_m}} \Phi_2(a, a^{-1}\xi)|a|^\frac{d-2}{2} d^\times a\right) d\xi,$$ where $\Phi_2$ varies among Schwartz functions on $\mathbbm A^2$.
If $d$ is odd, there is an equality between the space of pushforward measures for $V\xrightarrow{\xi} \mathbbm A^1$ and the measures on $\mathbbm A^1$ of the form $$\label{twistedpfG} \xi\mapsto \left(\int_{B_{{\operatorname{ad}}}} \Phi_2\left({\mathrm{Ad}}(b^{-1}) \begin{pmatrix}
& \xi \\ 1
\end{pmatrix}
\right)\delta_2(b)^\frac{d-3}{2} db\right) d\xi,$$ where $\Phi_2$ varies among Schwartz functions on $\mathfrak{sl}_2$, $B_{{\operatorname{ad}}}$ denotes the upper triangular Borel subgroup of ${\operatorname{PGL}}_2$, $\delta_2$ is its modular character, and $db$ is a right Haar measure.
Here, sticking with standard notation, we have denoted by ${\mathrm{Ad}}$ the *left* adjoint representation of ${\operatorname{PGL}}_2$ on $\mathfrak{sl}_2$; but recall that our convention is that $G$ acts on the right on $X\times X$, hence $H$ acts on the right on $V$, and this convention is extended to the group ${{\operatorname{SO}}}(V)$.
1. In other words, the germs are reduced to *twisted versions* of the infinitesimal versions of the *basic cases* $X={\mathbbm{G}_m}\backslash {\operatorname{PGL}}_2$ and $X={\operatorname{SL}}_2={{\operatorname{SO}}}_3\backslash {{\operatorname{SO}}}_4$. Indeed, the linearizations of those two are, respectively, $\mathbbm A_2/{\mathbbm{G}_m}$ and $\mathfrak{sl}_2/{\operatorname{PGL}}_2$, and the latter can also be replaced by $\mathfrak{sl}_2/B_{{\operatorname{ad}}}$, because the affine quotients $\mathfrak{sl}_2\sslash {\operatorname{PGL}}_2$ and $\mathfrak{sl}_2\sslash B_{{\operatorname{ad}}}$ are the same. Putting an appropriate character on ${\mathbbm{G}_m}$ or $B_{{\operatorname{ad}}}$, we obtain the germs for the general case. This fact will be used to relate those germs to the Kloosterman germs of the Kuznetsov formula, under the transfer operator.
2. As we saw in , , the measures , can be considered as twisted pushforwards of the Haar measures $\Phi_2 dv$ (where $dv$ is a Haar measure on $\mathbbm A^2$, resp. $\mathfrak{sl}_2$), dual to the twisted pullback maps: $$\label{twistedpbT}
\Psi \mapsto \tilde\Psi(x, y) = \Psi(xy) |x|^\frac{d-2}{2},$$ resp. $$\label{twistedpbG}
\Psi \mapsto \tilde\Psi\begin{pmatrix} A & B \\ C & -A \end{pmatrix} = \Psi(A^2+BC) |C|^{\frac{d-3}{2}}.$$
Let $\ell_\xi(\Phi)$ be the ${{\operatorname{SO}}}(V)$-invariant functional represented by the expression in brackets of . By that formula, the pushforward $f$ of $\Phi dv$ can be written $$f(\xi) = \ell_\xi(\Phi) d\xi = \int_{K} \int_{B_2} \Phi_2^K(\sigma(\xi) b) \delta_2(b)^{\dim M-1} db dk,$$ where $\Phi_2^K \in \mathcal F(V_2)$ is the Schwartz function $$\label{PhiK} v_2\to \int_{K} \int_{M'} \Phi((v_2+v')k) dv' dk.$$
As we have seen in Lemma \[rewrite\], this is equal to the expressions , in the two cases, applied to the function $\Phi_2^K$. We just need to argue that these spaces of pushforward measures obtained from a function of the form $\Phi_2^K$ is the same as the space obtained from an arbitrary Schwartz function $\Phi_2\in \mathcal F(V_2)$. The pushforward map $\mathcal S(M'^\perp)\to \mathcal S(V_2)$ is surjective, and starting from an arbitrary Schwartz measure $\Phi_2 dv_2\in \mathcal S(V_2)$ we can choose a preimage $\Phi_1 dv_1\in \mathcal S(M'^\perp)$. Without loss of generality (in terms of the output of , ), we will assume that $\Phi_1$ is $K\cap P'$-invariant.
Notice that we have freedom in choosing $K$, as long as the map $K\to P'\backslash P={\mathbbm{P}}M^*$ is surjective. Identify $M^*$ as a subspace of $V$ through an isotropic splitting of the quotient $V\to M^*$, so that we have a direct sum decomposition $V=M^\perp\oplus M^*$, and choose $K$ in the Levi subgroup ${\operatorname{GL}}(M^*)\subset P$. Any element of $K$ fixing the line in $M^*$ corresponding to $M'$ has to belong to $K\cap P'$. Thus, for any two $v_1, v_2\in M'^\perp$ with nonzero image in $M^\perp$, the relation $v_1\cdot k=v_2$ for some $k\in K$ implies that $v_2\in v_1\cdot (K\cap P')$. Hence, the map of topological quotients $$M'^\perp/K\cap P' \to V/K,$$ surjective by our assumption on $K$, is also injective. Thus, $\Phi_1$, a $K\cap P'$-invariant function on $M'^\perp$, is the restriction of a unique $K$-invariant function $\Phi$ on $V$; in particular, the average $\Phi_2^K$ (defined in terms of $\Phi$, and using probability measure on $K$) is equal to the function $\Phi_2$ that we started from.
In the non-Archimedean case, it is immediate to see that if $\Phi_1$ is smooth, so is $\Phi$. In the Archimedean case, taking $K={{\operatorname{SO}}}_n({\mathbb{R}})$ when $F={\mathbb{R}}$ and $K=\operatorname{U}_n({\mathbb{R}})$ when $F={\mathbb{C}}$ (where $n=\dim M$), the quotient $M^*/K$ can be identified with ${\mathbb{R}}_{\ge 0}$ through the distance function from the origin. For any one-dimensional subspace $Fv$ of $M^*$, any $K\cap {\operatorname{GL}}_1(F)$-invariant smooth function on $Fv$ is the restriction of a smooth radial function on $M^*$.
Finally, in the case $d=$ even, the Mellin transforms of the twisted orbital integrals are Tate integrals in two variables, without any need to pass to a blowup, and from this it is immediate to see that the nontrivial quadratic characters do not appear in the expression ; thus, $a_\eta=0$ except for $\eta=1$.
By Proposition \[linearization\], this completes the proof of Theorem \[Xtheorem\], which we state here more precisely. Notice that the precise local ($F$-analytic) isomorphism between a neighborhood of $0\in {\mathfrak{c}}_X^*$ and a neighborhood of $[\pm 1]\in {\mathfrak{C}}_X$, mentioned in Proposition \[linearization\], is not important, since the germs are invariant under any $F$-analytic automorphism. In particular, fixing isomorphisms ${\mathfrak{c}}_X^*\simeq\mathbbm A^1_\xi$ and ${\mathfrak{C}}_X=\mathbbm A^1_c$ with $c_{\pm 1}$ corresponding to $[\pm 1]$, we can take $\xi = c-c_{\pm 1}$.
\[Xtheorem2\] There is a canonical isomorphism ${\mathfrak{C}}_X:= X\times X\sslash G\simeq A_X\sslash W_X$, and the map $X\times X\to {\mathfrak{C}}_X$ is smooth away from the preimages of $[\pm 1]$, where $[\pm 1]$ denote the images of $\pm 1\in A_X$ in $A_X\sslash W_X$.
In particular, there are two distinguished closed $G$-orbits $X_1 = X^{{\operatorname{diag}}}$ and $X_{-1}$ (over $[\pm 1]$, respectively); if $d_{\pm 1}$ denote their codimensions, then $d_1=\dim X$ and $$d_{-1} = \epsilon \left<2\rho_{P(X)},\check\gamma\right>-d_1+2,$$ where $\check\gamma$ is the spherical coroot, $2\rho_{P(X)}$ is the sum of roots in the unipotent radical of $P(X)$, and $$\epsilon = \begin{cases} 1, \mbox{ when the spherical root is of type $T$ (dual group ${\operatorname{SL}}_2$)};\\
2, \mbox{ when the spherical root is of type $G$ (dual group ${\operatorname{PGL}}_2$).}
\end{cases}$$ In the case of root of type $G$, $d_1=d_{-1}$.
The space $\mathcal S(X\times X/G)$ consists of measures on ${\mathfrak{C}}_X(\simeq \mathbbm A^1_c)$ which are smooth and of rapid decay, together with their polynomial derivatives (compactly supported in the non-Archimedean case) away from neighborhoods of $c_{\pm 1}$ (the coordinates of the points $[\pm 1]$), while in neighborhoods of $c_{\pm 1}$ they are of the form — when the spherical root is of type $T$ — or — when the spherical root is of type $G$—, with $\xi = c-c_{\pm 1}$ and $d=d_{\pm 1}$.
Recall that, up to a linear combination of coefficients, the singularities of the measures of the form , have been described explicitly in Theorem \[pushfthm\]. In the Archimedean case, there is a natural Fréchet topology on the space of these measures, and by Proposition \[propbigimage\] the map from $\mathcal S(X\times X)$ is continuous; hence, the quotient topology on $\mathcal S(X\times X/G)$ coincides with the natural Fréchet topology on the space of such measures.
Completion of the proof of the main theorem {#sspfmainthm}
-------------------------------------------
We are now ready to prove Theorem \[maintheorem\], which I repeat, for the convenience of the reader. Recall that $G^*$ is such that its dual group is $\check G_X$, that is: $G^*={\operatorname{PGL}}_2$ when the spherical root of $X$ is of type $T$, and $G^*={\operatorname{SL}}_2$ when the spherical root is of type $G$.
\[maintheorem2\] Let ${\mathfrak{C}}_X=(X\times X)\sslash G$. There is an isomorphism ${\mathfrak{C}}_X \simeq \mathbbm A^1$, and the map $X\times X\to \mathbbm A^1$ is smooth away from the preimage of two points of $\mathbbm A^1$, that we will call singular. We fix the isomorphisms as follows:
- When $\check G_X={\operatorname{SL}}_2$, we take the set of singular points to be $\{0, 1\}$, with $X^{{\operatorname{diag}}}\subset X\times X$ mapping to $1 \in {\mathfrak{C}}_X\simeq \mathbbm A^1$.
- When $\check G_X = {\operatorname{PGL}}_2$, we take the set of singular points to be $\{-2,2\}$, with $X^{{\operatorname{diag}}}\subset X\times X$ mapping to $2 \in {\mathfrak{C}}_X\simeq \mathbbm A^1$.
Then, there is a continuous linear isomorphism: $$\mathcal T: \mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi) \xrightarrow\sim \mathcal S(X\times X/G),$$ given by the following formula:
- When $\check G_X={\operatorname{SL}}_2$ with $L_X=L({{\operatorname{Std}}}, s_1) L({{\operatorname{Std}}},s_2)$ with $s_1\ge s_2$, $$\mathcal Tf(\xi) = |\xi|^{s_1-\frac{1}{2}} \left( |\bullet|^{\frac{1}{2}-s_1} \psi(\bullet) d\bullet\right) \star \left( |\bullet|^{\frac{1}{2}-s_2} \psi(\bullet) d\bullet\right) \star f(\xi).$$
- When $\check G_X={\operatorname{PGL}}_2$ with $L_X=L({\mathrm{Ad}}, s_0)$, $$\label{transferG}
\mathcal Tf(\zeta) = |\zeta|^{s_0-1} \left( |\bullet|^{1-s_0} \psi(\bullet) d\bullet\right) \star f(\zeta).$$
\[remarkconstants\] The points $s_1, s_2, s_0$ are determined by the geometry, according to the following formulas: $$s_1+s_2=\frac{\dim X}{2};$$ $$s_1 = \frac{\left<\check\gamma, \rho_{P(X)}\right>}{2},$$ and therefore, by , $$s_1-s_2= \frac{\dim X_{-1}}{2} -1;$$ $$\label{s0}
s_0 = \left<\check\gamma, \rho_{P(X)}\right> = \frac{\dim X-1}{2} = \frac{\dim X_{-1}-1}{2},$$ the last one by .
Let us start with the case $G^*={\operatorname{SL}}_2$. The space $\mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi)$ can be thought of as the space of sections of a cosheaf over ${\mathbbm{P}}^1(F)$; in a neighborhood of infinity, its elements have the form $$f(\zeta) = |\zeta|^{1-s_0} \Phi(\zeta^{-1}) d^\times \zeta,$$ where $\Phi$ is a smooth function.
If we consider the subspace $\mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi)^0$ of Schwartz sections in the complement of $0\in {\mathbbm{P}}^1$, it is immediate to see that the transfer operator of defines a continuous isomorphism between this space and the space $\mathcal S(\mathbbm A^1)$ of usual Schwartz measures on the $F$-points of the affine line.
There remains to determine the behavior of the stalk at zero under this transform, i.e., the behavior of the “Kloosterman germs”. Applying [@SaTransfer2 Proposition 8.3.3] with $\chi= \delta_2^{\frac{1}{2}-s_0}$, we get that the operator $\mathcal T$ maps the space $\mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi)$ isomorphically to $\mathcal S(\frac{{\operatorname{SL}}_2}{B_{{\operatorname{ad}}}, \delta_2^{1-s_0}})$, in the notation of that proposition. This is precisely the space of measures on the affine line $\mathbbm A^1$ (with coordinate $c$), which are smooth away from $c=\pm 2$, of rapid decay (together with their derivatives) at infinity, and in a neighborhood of $c=\pm 2$, setting $\xi=c\mp 2$, are of the form , with $\frac{d-3}{2}=s_0-1$. By Theorem \[Xtheorem2\], this is precisely the space $\mathcal S(X\times X/G)$.
In the case $G^*={\operatorname{PGL}}_2$, we start again with the subspace $\mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi)^0$ of Schwartz sections in the complement of zero; those, now, are of the form $$f(\xi)= \left(\Phi_1(\xi^{-1}) |\xi|^{\frac{1}{2}-s_1} + \Phi_2(\xi^{-1}) |\xi|^{\frac{1}{2}-s_2}\right) d^\times \xi$$ in a neighborhood of infinity, with the suitable logarithmic modification when $|\xi|^{s_1-s_2}$ is smooth. (All $\Phi_i$’s, here and below, denote smooth functions.) The following results about the effects of multiplicative convolutions on such measures can be seen directly by considering their Mellin transforms, which belong to the appropriate “Paley–Wiener space”, in the language of [@SaTransfer1 Remark 2.1.6], using the functional equation of multiplicative Fourier convolutions of [@SaTransfer1 §2.1.7]. The asymptotic exponents of the measures determine the location and multiplicity of poles of the Mellin transforms, and all the maps are continuous in the Archimedean case; since the details are similar to the aforementioned reference, they are left to the reader.
The first convolution, by $\left( |\bullet|^{\frac{1}{2}-s_2} \psi(\bullet) d\bullet\right)$, takes the space $\mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi)^0$ to the space of measures which at infinity are of the form $\Phi_3(\xi^{-1}) |\xi|^{\frac{1}{2}-s_1} d^\times \xi$, while at zero are of the form $\Phi_4(\xi) |\xi|^{\frac{1}{2}-s_2} d\xi$, and otherwise smooth. The second convolution, by $\left( |\bullet|^{\frac{1}{2}-s_1} \psi(\bullet) d\bullet\right)$, takes this space to the space of measures which are of rapid decay (compactly supported, in the non-Archimedean case), of the form $\left(\Phi_5(\xi) |\xi|^{\frac{1}{2}-s_1} + \Phi_6(\xi) |\xi|^{\frac{1}{2}-s_2}\right) d\xi$ in a neighborhood of $\xi=0$, and otherwise smooth. Finally, multiplication by the factor $|\xi|^{s_1-\frac{1}{2}}$ (by our convention that $s_1=\max(s_1,s_2)$) turns the germs at zero to $$\left(\Phi_5(\xi) + \Phi_6(\xi) |\xi|^{s_1-s_2}\right) d\xi =\left( \Phi_5(\xi) + \Phi_6(\xi) |\xi|^{\frac{d_{-1}}{2}-1}\right) d\xi.$$
There remains to examine the effect of the transfer operator to the “Kloosterman germs”, i.e., to the stalk of $\mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi)$ at zero; more precisely, we want to show that this stalk contributes an extra summand of $$\Phi_7(\xi) |\xi-1|^{s_1+s_2-1} d\xi = \Phi_7(\xi) |\xi-1|^{\frac{d_1}{2}-1} d\xi$$ in a neighborhood of $\xi=1$.
To that end, we can argue as in the case $G^*={\operatorname{SL}}_2$, appealing to Theorem \[Xtheorem2\] and Proposition \[reductiontobasic\], and using the “unfolding” technique, as in the proof of [@SaTransfer2 Proposition 8.3.3], to study the behavior of the Kloosterman germs, i.e., the germs of the elements of $\mathcal S^-_{L_X}(N,\psi\backslash G^*/N,\psi)$ at zero, under the transfer operator. (See also [@SaBE1 Theorem 5.4], where this technique was used to prove a certain fundamental lemma.)
Alternatively, the transform of this stalk under the transfer operator can be computed directly, as in the proof of [@SaBE1 Theorem 5.1]. Since the details are the same, I only mention that, after application of the first convolution, by $\left( |\bullet|^{\frac{1}{2}-s_2} \psi(\bullet) d\bullet\right)$, this stalk contributes an extra summand of the form $$\Phi_8(\xi) \psi^{-1}(\xi^{-1}) |\xi|^{s_2-\frac{1}{2}} d\xi$$ at zero, which after the second convolution, by $\left( |\bullet|^{\frac{1}{2}-s_1} \psi(\bullet) d\bullet\right)$, gives rise to the extra term $$\Phi_7(\xi) |\xi-1|^{s_1+s_2-1} d\xi$$ in a neighborhood of $\xi=1$; multiplication by the factor $|\xi|^{s_1-\frac{1}{2}}$ does not alter this germ.
Relation to the boundary degeneration {#sec:boundary}
=====================================
Finally, I would like to give a conceptual explanation for the transfer operator $\mathcal T$ of Theorem \[maintheorem\]. The explanation uses boundary degenerations, and is analogous to the one discussed in [@SaTransfer1 §4.3, 5] for the special cases $X={\mathbbm{G}_m}\backslash {\operatorname{PGL}}_2$ and $X={\operatorname{SL}}_2$. Since the arguments are analogous, I only sketch them without many details.
Here we will assume that $F$ is non-Archimedean. Recall the (homogeneous) boundary degeneration $X_\emptyset^\bullet$ introduced in §\[ssdegeneration\]. There is a canonical “asymptotics” morphism $e_\emptyset^*: \mathcal S(X)\to \mathcal S^+(X_\emptyset)$, where $\mathcal S^+(X_\emptyset)$ denotes a certain space of smooth measures on $X_\emptyset^\bullet$, whose support has compact closure in $X_\emptyset$, which contains the space $\mathcal S(X_\emptyset^\bullet)$ of smooth, compactly supported measures on $X_\emptyset^\bullet$. Moreover, the nontrivial element $w_\gamma\in W_X$ acts by an involution called the “scattering operator” $\mathfrak S_{w_\gamma}$ on $\mathcal S^+(X_\emptyset)$. This involution is $(A_X,w_\gamma)$-equivariant, that is, it intertwines the action of $a\in A_X$ with the action of ${^{w_\gamma}a}=a^{-1}$, when this action is normalized to be unitary: if we identify $X_\emptyset^\bullet$ with $S\backslash G$, where $S\subset P(X)^-$ (the opposite of $P(X)$) is the kernel of the map $P(X)^-\to A_X$, and $\delta_{P(X)}$ is the modular character of $P(X)$ (the inverse of the modular character of $P(X)^-$), then this unitary action is given by $$\label{norm-measures} a\cdot f(Sg) = \delta_{P(X)}^{-\frac{1}{2}}(a) f(Sag).$$ (Recall that $f\in \mathcal S^+(X_\emptyset)$ is a measure, not a function.)
Now, the scattering operator can be expressed in terms of standard intertwining operators and “Radon transforms”, as in [@SaTransfer1 Theorem 3.5.1]. This expression involves a slight extension of the results of [@SaSph], including to the cases where $G$ is not a classical group. Since the same methods apply, I will dispense with all the details and just formulate the result. The formulation of the result would be very easy, if it was not for the normalization of “standard intertwining operators”. This requires some careful definitions.
To begin with, notice that we have canonical “twisted pushforward maps”: $$\label{Mellin} \mathcal S(X_\emptyset^\bullet)\to \mathcal S(A_X\backslash X_\emptyset^\bullet,\chi),$$ where the space on the right denotes measures on $A_X\backslash X_\emptyset^\bullet$, valued in the line bundle whose sections are $(A_X,\chi)$-equivariant functions on $X_\emptyset^\bullet$, under the normalized action on functions that is dual to : $$\label{norm-functions} a\cdot \Phi(Sg) = \delta_{P(X)}^{\frac{1}{2}}(a) \Phi(Sag).$$ We will denote the map by $f\mapsto \check f(\chi)$, and think of it as a Mellin transform. The transform extends (meromorphically in $\chi$) to elements of $\mathcal S^+(X_\emptyset)$, and an element of $\mathcal S^+(X_\emptyset)$ can be recovered from its Mellin transform by Mellin inversion; I point to [@SaTransfer1] for details. Thus, we will describe the scattering endomorphism of this space by describing a meromorphic (in $\chi$) family of morphisms: $$\mathscr S_{w_\gamma,\chi} : \mathcal S(A_X\backslash X_\emptyset^\bullet,\chi^{-1}) \to \mathcal S(A_X\backslash X_\emptyset^\bullet,\chi).$$
Notice that, up to a choice of scalar (depending on $\chi$), the space $\mathcal S(A_X\backslash X_\emptyset^\bullet,\chi)$ is isomorphic to the induced principal series representation $$I_{P(X)^-}^G(\chi) = {\operatorname{Ind}}_{P(X)^-}^G (\chi\delta_{P(X)}^{-\frac{1}{2}}),$$ thus the morphism $\mathscr S_{w_\gamma,\chi}$ has to be (for almost all $\chi$) a multiple of the “standard intertwining operator” $$\mathfrak R_\chi: I_{P(X)^-}^G(\chi^{-1}) \to I_{P(X)^-}^G(\chi)$$ given by the integral $$\mathfrak R_\chi f(g) = \int_{U_{P(X)}^-} f(w_\gamma ug) du.$$ The issue is how to fix the isomorphism with $I_{P(X)^-}^G(\chi)$, and the representative $w_\gamma$ in the Weyl group, as well as the measure $u$, that appear in the definition of $\mathfrak R_\chi$. It turns out that there is a canonical way to do define $\mathfrak R_\chi$ (which I will call “spectral Radon transform”), having fixed first a nontrivial additive character $\psi: F\to {\mathbb{C}}^\times$. Let me, however, formulate the main result first, assuming that the character $\psi$ has been fixed such that our fixed Haar measure on $F$ is self-dual, and the appropriate definition of $\mathfrak R_\chi$ has been given:
\[thmscattering\] For the cases of Table , in terms of the canonical spectral Radon transforms $\mathfrak R_\chi$ that will be described below, $$\mathfrak R_\chi: \mathcal S(A_X\backslash X_\emptyset^\bullet, \chi^{-1}) \to \mathcal S(A_X\backslash X_\emptyset^\bullet,\chi),$$ the scattering operator $\mathscr S_{w_\gamma,\chi}$ for the nontrivial element $w_\gamma$ of $W_X$ is given by $$\mathscr S_{w_\gamma,\chi} = \mu_X(\chi) \cdot \mathfrak R_{\chi},$$ where $\mu_X$ is given by the following formulas:
- for spherical roots of type $T$, $$\mu_X(\chi) = \gamma(\chi, \frac{\check\gamma}{2}, 1-s_1, \psi^{-1}) \gamma(\chi, \frac{\check\gamma}{2}, 1-s_2, \psi) \gamma(\chi,-\check\gamma, 0,\psi),$$
- for spherical roots of type $G$, $$\mu_X(\chi) = \gamma(\chi,\check\gamma, 1-s_0 ,\psi^{-1}) \gamma(\chi,-\check\gamma,0,\psi).$$
Here, $\gamma(\chi, \check\lambda, s, \psi)$ denotes the gamma factor of the local functional equation for the abelian $L$-function associated to the composition of $\chi$ with the cocharacter $\check\lambda: {\mathbbm{G}_m}\to A_X$ — see [@SaTransfer1 §2.1] for a recollection of the standard notational conventions. Notice that the factor $\mu_X(\chi)$ does not change if we replace $\psi$ by $\psi_u(x) = \psi(ux)$, for any $u\in F^\times$ with $|u|=1$, or by interchanging $s_1$ and $s_2$; thus, it is independent of $\psi$, the latter having been fixed so that the fixed measure on $F$ is self-dual.
This theorem has the following corollary, as in [@SaTransfer1 Theorem 3.6.3]; it is expressed in terms of a *canonical* open embedding $A_X\hookrightarrow X_\emptyset\times X_\emptyset\sslash G$, which is part of the careful definition of $\mathfrak R_\chi$ (see below):
Consider the meromorphic family of functionals $I_\chi$ obtained as pullbacks of the composition of maps $$\xymatrix{ \mathcal S(X\times X)\ar[rr]^{e_\emptyset^*\otimes e_\emptyset^*} && \mathcal S^+(X_\emptyset\times X_\emptyset) \ar[rr]^{\,\,\,\,\,\,\,\,\,\,\,\, \int \chi^{-1}\delta_{P(X)}^{-\frac{1}{2}}} && {\mathbb{C}},}$$ where the last arrow denotes the integral against the pullback of the character $\chi^{-1}\delta_{P(X)}^{-\frac{1}{2}}$ from $A_X$ to (a dense open subset of) $X_\emptyset\times X_\emptyset$.
For an open dense subset of $\chi$’s (in the complex group $\widehat{A_X}_{\mathbb{C}}$ of characters of $A_X$), the functional $I_\chi$ is a *relative character* for the normalized principal series representation $\pi_\chi=I_{P(X)}(\chi)$, that is, it factors through a morphism $$\mathcal S(X\times X) \to \pi_\chi\otimes \widetilde{\pi_\chi} \xrightarrow{\left<\,\, , \,\, \right>} {\mathbb{C}},$$ the product $I_\chi \mu_X(\chi)$ is invariant under the $W_X$-action $\chi\mapsto \chi^{-1}$, and for suitable choices of invariant measures, the most continuous part of the Plancherel decomposition of $L^2(X)$ (corresponding to a canonical subspace $L^2(X)_\emptyset\subset L^2(X)$) reads: $$\label{PlancherelX} \left<f_1, \overline{f_2}\right>_\emptyset = \int_{\widehat{A_X}/W_X} I_\chi(f_1\otimes f_2) \mu_X(\chi) d\chi.$$
Finally, let $\mathcal S^+(N\backslash G^*)$ be the image of the asymptotics map $e_\emptyset^*$ for the Whittaker model of $G^*$: $$e_\emptyset^*: \mathcal S(N,\psi\backslash G^*) \to \mathcal S^+(N\backslash G^*),$$ where $N$ is the upper triangular unipotent subgroup, identified in the standard way with ${\mathbbm{G}_a}$, and $\psi$ is our fixed additive character, and let $\mathcal S^+(N\backslash G^*/N)$ be the image of the pushforward map from $\mathcal S^+(N\backslash G^*)\otimes \mathcal S^+(N\backslash G^*)$ to $N\backslash G^*\sslash N = (N\backslash G^*) \times (N\backslash G^*)\sslash G^*$. Identifying the Cartan $A^*$ of $G^*$ with the torus of diagonal elements through the upper triangular Borel, the embedding $A^*\to wA^*\subset G^*$ fixed in §\[ssnotation\] descends to an embedding of $A^*$ in $N\backslash G^*\sslash N$. We similarly have a relative character $J_\chi$ on $\mathcal S(N,\psi\backslash G^*)\otimes \mathcal S(N,\psi^{-1}\backslash G^*)$, obtained as the composition $$\xymatrix{ \mathcal S(N,\psi\backslash G^*)\otimes \mathcal S(N,\psi^{-1}\backslash G^*) \ar[rr]^{\,\,\,\,\,\,\,\,\,\, e_\emptyset^*\otimes e_\emptyset^*} && \mathcal S^+(N\backslash G^*/N) \ar[rr]^{\,\,\,\,\,\,\,\,\int \chi^{-1}\delta_{B^*}^{-\frac{1}{2}}} && {\mathbb{C}}}.$$ Notice that here we are integrating here against the character $\chi^{-1}\delta_{B^*}^{-\frac{1}{2}}$ of $A^*\subset N\backslash G^*\sslash N$, where $B^*\subset G^*$ is the Borel subgroup of $G^*$; this makes $J_\chi$ a relative character for the normalized principal series $I_{B^*}^{G^*}(\chi)$.
We identify $A^*$ with the Cartan $A_X$ of $X$, so that the positive root of $G^*$ corresponds to the spherical root $\gamma$. The relative characters $J_\chi$ satisfy a similar Plancherel formula for the most continuous part of $L^2(N,\psi\backslash G^*)$, again by [@SaTransfer1 Theorem 3.6.3]: $$\label{PlancherelW} \left<f_1, \overline{f_2}\right>_\emptyset = \int_{\widehat{A_X}/W_X} J_\chi(f_1\otimes f_2) \gamma(\chi,-\check\gamma, 0,\psi) d\chi.$$
Identify $N\backslash G^*\sslash N\simeq{\mathbbm{G}_a}$ in a way compatible with the coordinates $\xi$ and $\zeta$ that were used in §\[ssnotation\], and identify $X_\emptyset\times X_\emptyset\sslash G\simeq {\mathbbm{G}_a}$ as follows:
- if the spherical root $\gamma$ is of type $G$, through the map $e^{\frac{\gamma}{2}}$ on $A_X \subset X_\emptyset\times X_\emptyset\sslash G$.
- if the spherical root $\gamma$ is of type $T$, through the map $(-e^\gamma)$. The negative sign before the character $e^\gamma$ is due to our choice of coordinates for $X\times X\sslash G$, and was explained in [@SaTransfer1 §5].
Consider a diagram $$\label{diagramdegen}
\xymatrix{\mathcal S(N,\psi\backslash G^*)\otimes \mathcal S(N,\psi^{-1}\backslash G^*)
\ar[rr]^{e_\emptyset^*\otimes e_\emptyset^*} && \mathcal S^+(N\backslash G^*/N) \ar[d]^{\mathcal T_\emptyset} \\
\mathcal S(X\times X) \ar[rr]^{e_\emptyset^*\otimes e_\emptyset^*} && \mathcal S^+(X_\emptyset\times X_\emptyset/G) \ar[rrr]^{~\,\,\,\,\,\,\,\,\,\,\mu_X(\chi) \int \chi^{-1}\delta_{P(X)}^{-\frac{1}{2}}} &&& {\mathbb{C}},}$$ with the map $\mathcal T_\emptyset$ to be determined.
The composition of arrows in the bottom row is the relative character $\mu_X(\chi) I_\chi$ of the Plancherel formula . By the philosophy of relative functoriality, this relative character for $X$ should correspond to the relative character $\gamma(\chi,-\check\gamma, 0,\psi) J_\chi$ of the Plancherel formula .
As in [@SaTransfer1 Theorems 4.3.1 and 5.0.2], we get:
There is a unique $A_X$-equivariant operator $$\mathcal T_\emptyset: \mathcal S^+(N\backslash G^*/N) \to \mathcal S^+(X_\emptyset \times X_\emptyset/G)$$ such that the functional of the diagram pulls back to the relative character $\gamma(\chi,-\check\gamma, 0,\psi) J_\chi$ on $\mathcal S(N,\psi\backslash G^*)\otimes \mathcal S(N,\psi^{-1}\backslash G^*)$.
Moreover, in the coordinates fixed above, the operator is given by the *same* formula as the transfer operator $\mathcal T$ of Theorem \[maintheorem\], that is:
- When $\check G_X={\operatorname{SL}}_2$ with $L_X=L({{\operatorname{Std}}}, s_1) L({{\operatorname{Std}}},s_2)$, $s_1\ge s_2$, $$\mathcal T_\emptyset f(\xi) = |\xi|^{s_1-\frac{1}{2}} \left( |\bullet|^{\frac{1}{2}-s_1} \psi(\bullet) d\bullet\right) \star \left( |\bullet|^{\frac{1}{2}-s_2} \psi(\bullet) d\bullet\right) \star f(\xi).$$
- When $\check G_X={\operatorname{PGL}}_2$ with $L_X=L({\mathrm{Ad}}, s_0)$, $$\mathcal T_\emptyset f(\zeta) = |\zeta|^{s_0-1} \left( |\bullet|^{1-s_0} \psi(\bullet) d\bullet\right) \star f(\zeta).$$
The term “$A_X$-equivariant”, here, refers to the normalized action of $A_X$ on $\mathcal S^+(X_\emptyset \times X_\emptyset/G)$ that descends from , and, similarly, its analogously normalized action (but using the modular character $\delta_{B^*}$ instead of $\delta_{P(X)}$) on $\mathcal S^+(N\backslash G^*/N)$. The factor $|\xi|^{s_1-\frac{1}{2}}$, resp. $|\zeta|^{s_0-1}$, in the formula for $\mathcal T_\emptyset$ is due to the difference between the characters $\delta_{B^*}^{-\frac{1}{2}}$ and $\delta_{P(X)}^{-\frac{1}{2}}$ in the definition of the relative characters $I_\chi$ and $J_\chi$; in terms of the torus $A_X$, this factor can be written $|e^{\rho_{P(X)} - \frac{\gamma}{2}}| = |e^{\rho_{P(X)} - \rho_{B^*}}| = \delta_{P(X)}^{\frac{1}{2}}\delta_{B^*}^{-\frac{1}{2}}$.
I conjecture that descends to a commutative diagram $$\xymatrix{
\mathcal S(N,\psi\backslash G^*/N,\psi) \ar[rr]^{e_\emptyset^*\otimes e_\emptyset^*}\ar[d]^{\mathcal T} && \mathcal S^+(N\backslash G^*/N) \ar[d]^{\mathcal T_\emptyset} \\
\mathcal S(X\times X/G) \ar[rr]^{e_\emptyset^*\otimes e_\emptyset^*} && \mathcal S^+(X_\emptyset\times X_\emptyset/G) }.$$ This would imply that the relative characters under $\mathcal T$ satisfy: $$\mathcal T^*(\mu_X(\chi) I_\chi) = \gamma(\chi,-\check\gamma, 0,\psi) J_\chi.$$ This was proven for the basic cases $A_1$ and $D_2$ in [@SaTransfer1]. For $\chi$ unramified, given the fundamental lemma in the upcoming work of Johnstone and Krishna [@JK], it is directly related to the extension of this fundamental lemma to the whole Hecke algebra.
Finally, on the definition of the spectral Radon transforms $$\mathfrak R_\chi: \mathcal S(A_X\backslash X_\emptyset^\bullet, \chi^{-1}) \to \mathcal S(A_X\backslash X_\emptyset^\bullet,\chi),$$ used in Theorem \[thmscattering\]: The definition depends on determining a distinguished $G$-orbit $X_\emptyset^R$ in the “open Bruhat cell” of $X_\emptyset^\bullet\times X_\emptyset^\bullet$, that is, whose image in ${\mathcal{B}}_X\times {\mathcal{B}}_X$ under the map induced from $X_\emptyset^\bullet\to {\mathcal{B}}_X$ (the map taking a point to the normalizer of its stabilizer) belongs to the open $G$-orbit.
Given the distinguished orbit $X_\emptyset^R$, the operator $\mathfrak R_\chi$ is the one that descends from the following Radon transform on measures on $X_\emptyset^\bullet$: $$\mathcal S(X_\emptyset^\bullet)\ni \Phi dx\mapsto \mathfrak R(\Phi dx) (x) = \left(\int_{(x,y)\in X_\emptyset^R} \Phi(y) dy\right)\cdot dx.$$ Here, $dx$ denotes an invariant measure on $X_\emptyset$. This definition depends on fixing ($G$-equivariantly) measures $dy$ on the fibers of $X_\emptyset^R$ with respect to the first projection, which should be fixed so that the spectral scattering operators $\mathscr S_{w_\gamma,\chi}$ are unitary for $\chi$ unitary. Since this characterizes the measures uniquely, I leave their description to the reader, noting that [@SaTransfer1 §3] contains some basic examples.
We are left with describing the distinguished $G$-orbit $X_\emptyset^R$. What follows is the description of this orbit that is dictated by the methods used in [@SaSph; @SaTransfer1], but I will not explain the arguments behind it — I only include the result for the sake of completeness.
A $G$-orbit on $X_\emptyset^\bullet\times X_\emptyset^\bullet$ is equivalent to an equivariant isomorphism $\iota: X_\emptyset^\bullet \xrightarrow\sim X_\emptyset^h$, where $X_\emptyset^h$ is the space of *generic horocycles* on $X_\emptyset^h$, that is, the space of pairs $(P,Z)$, where $P\in {\mathcal{B}}_X$ and $Z\subset X_\emptyset$ is a $U_P$-orbit *in the open $P$-orbit* (Bruhat cell). Indeed, such an isomorphism defines a distinguished $G$-orbit $$X_\emptyset^R=\{(x,y) \in X_\emptyset^\bullet\times X_\emptyset^\bullet| y\in \iota(x)\},$$ and, vice versa, can be recovered from it. Similarly, this is equivalent to describing a distinguished $G$-orbit $$X_\emptyset^{R,h} \subset X_\emptyset^h\times X_\emptyset^h.$$ This is the orbit that we will describe. Notice that, by [@SV Lemma 2.8.1], there is a canonical isomorphism $X_\emptyset^h\simeq X^h$, where $X^h$ is defined in the same way, except that $Z\subset X$. Thus, we are looking to describe a distinguished $G$-orbit $X^{R,h}\subset X^h\times X^h$.
Let $(Y, P)$ be a basic orbit-parabolic pair, as in Lemma \[lemmaclosedorbits\], and let $Y_0\subset Y$ be the open $B$-orbit; without fixing a Borel subgroup, we can consider $Y_0$ as a $G$-orbit on $X\times {\mathcal{B}}$ — denoted by $\tilde Y_0$ to avoid confusion. Similarly to the definition of $X^h$, we can define $$X^{h,Y} = \{(B, Z)| B\in {\mathcal{B}}, \, Z \mbox{ is a $U_B$-orbit with } (z,B)\in \tilde Y_0\mbox{ for any }z\in Z\}.$$ Hence, these are not generic horocycles, but horocycles corresponding to the $B$-orbit $Y_0$.
Let $\tilde X^h$ be the base change $X^h\times_{{\mathcal{B}}_X}{\mathcal{B}}$ of $X^h$ to the full flag variety. Fixing a Borel subgroup $B$, we have non-canonical isomorphisms: $$\label{noncan1} \tilde X^h = A_X\times^B G,$$ $$\label{noncan2} \tilde X^{Y,h}= A_Y\times^B G.$$
Let $w_{\gamma'}$ be the Weyl group element $w_\alpha$ or $w_\alpha w_\beta$, respectively (where $w_\alpha$, $w_\beta$ denote simple reflections), for each of the two cases of Lemma \[lemmaclosedorbits\]. Let $w_1\in W$ be an element of minimal length such that $Y_0^{w_1}=\mathring X$ under Knop’s action; so, ${{\operatorname{codim}}}Y_0 = \operatorname{length}(w_1)$, and $A_Y^{w_1} = A_X$, see . Moreover, the nontrivial element $w_\gamma \in W_X$ is equal to $w_1^{-1} w_{\gamma'} w_1$.
The $P$-variety $Y_2:=Y/U_P$ is isomorphic either to ${\mathbbm{G}_m}\backslash{\operatorname{PGL}}_2$ or to ${\operatorname{SL}}_2$ under the action of the Levi quotient $L$ of $P$, by Lemma \[lemmaY2\]. In each of the two cases, a distinguished $L$-orbit on $Y_2^h\times Y_2^h$, living over the open Bruhat cell, was described in [@SaTransfer1 §3.4, 3.5]. This corresponds to a $G$-orbit on $X^{h,Y}$, living over the Bruhat cell corresponding to $w_{\gamma'}$. We denote by $X^{R,h,Y}$ this $G$-orbit.
Now, the product $X^{Y,h}\times \tilde X^h$ lives over the product ${\mathcal{B}}\times {\mathcal{B}}$. There is a *distinguished* $G$-orbit $X'^h\subset X^{Y,h}\times \tilde X^h$ that lives over the Bruhat cell corresponding to $w_1$, that is, over the $G$-orbit of the pair $(B,Bw_1)$: it contains all pairs of horocycles $((w_1^{-1} B w_1, y N \tilde w_1 ) , (B, x N))$, where $N$ denotes again the unipotent radical of $B$, and $x\in \mathring X$ (the open $B$-orbit), $y\in Y_0$ and $\tilde w_1$ is an element of $N\backslash B w_1 B \subset N\backslash G$ such that $y N \tilde w_1 N = xN$.
Consider the space of quadruples $$(x_1,y_1,y_2,x_2)$$ with $x_1, x_2\in X^h$, $y_1,y_2\in X^{Y,h}$, $(y_i, x_i)\in X'^h$ (for $i=1,2$) and $(y_1, y_2)\in X^{R,h,Y}$. It lives over the $G$-orbit of the quadruple $$(B, B w_1^{-1}, B w_1^{-1} w_{\gamma'}, B w_1^{-1} w_{\gamma'} w_1)$$ of Borel subgroups.
*The distinguished $G$-orbit $X^{R,h}\subset X^h\times X^h$, now, is the image of this subset under the first and last projections, composed with the projection $\tilde X^h\to X^h$.* The reader can immediately check that this is indeed a $G$-orbit, using the non-canonical isomorphisms , .
[^1]: The space $X\times X$ is smooth, but here we take into account the $G$-action, and use the term “resolution” to refer to the fibers of the quotient map $X\times X\to X\times X\sslash G$: a resolution is a blowup that turns them into normal crossings divisors.
[^2]: We will be using “split” in a slightly non-standard way: for quadratic spaces of odd dimension $d$, “split” will mean maximally isotropic, i.e., containing a $\frac{d-1}{2}$-dimensional isotropic subspace.
[^3]: A smooth geometric quotient $X\to Y$ is a smooth surjective morphism of $G$-varieties, with $G$ acting trivially on $Y$, such that geometric fibers are $G$-orbits. \[geomquotient\]
[^4]: In other references, $X_\emptyset$ denotes just the open $G$-orbit in $X_\emptyset$. Here, I have found it more convenient to use $X_\emptyset$ for the affine degeneration, and $X_\emptyset^\bullet$ for its open $G$-orbit.
[^5]: The action of $F^\times$ on measures is defined in duality with its (unnormalized) action on functions: $\left<a\cdot \mu, \Phi\right>=\left<\mu, a^{-1}\cdot \Phi\right>$, where $a^{-1}\cdot \Phi(x)= \Phi(a^{-1}x)$. In particular, Haar measure is $({\mathbbm{G}_m},|\bullet|^d)$-equivariant.
[^6]: We use the arithmetic normalization of absolute values, which is compatible with norms to the base field; this is the square of the usual absolute value in the complex case.
[^7]: In terms of the map ${\mathbbm{P}}J_X\to {\mathfrak{C}}_X$, of which the map $\tilde V\to {\mathfrak{c}}_X^*$ is the “linearization” by , and given that $J\bullet N^*_{X_{\pm 1},\ne 0}$ is smooth over $J$ by non-degeneracy of the quadratic forms, the first of the maps ${\mathbbm{P}}J_X\to {\mathbbm{P}}J \xrightarrow{Q} {\mathfrak{C}}_X$ is smooth, and the second is given, in coordinates $t_0^2-\xi t_1^2=1$ for $J$, and a suitable identification $A_X\sslash W_X = \mathbbm A^1$, by $Q=(t_0 \pm 1)^{-1} \cdot \xi t_1^2$. In a neighborhood of $t_0=\pm 1$, where the function $(t_0 \pm 1)^{-1}$ is a nonvanishing, smooth semialgebraic function of $(\xi, t_1)$, we can set $\xi'=(t_0 \pm 1)^{-1}\xi$, and we get that the map is given by $Q=\xi' t_1^2$.
[^8]: We define coinvariants of Fréchet spaces by dividing by the *closure* of the space generated by elements of the form $f-h\cdot f$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the distance-calibrated spectral energy distribution (SED) of TRAPPIST-1 using a new medium resolution (R$\sim$6000) near-infrared FIRE spectrum and its *Gaia* parallax. We report an updated bolometric luminosity ($L_\mathrm{bol}$) of $-3.216 \pm 0.016$, along with semi-empirical fundamental parameters: effective temperature $T_\mathrm{eff}=2628 \pm 42$ K, mass=$90 \pm 8~M_\mathrm{Jup}$, radius=$1.16 \pm 0.03~R_\mathrm{Jup}$, and log $g$=$5.21 \pm 0.06$ dex. It’s kinematics point toward an older age while spectral indices indicate youth therefore, we compare the overall SED and near-infrared bands of TRAPPIST-1 to field-age, low-gravity, and low-metallicity dwarfs of similar $T_\mathrm{eff}$ and $L_\mathrm{bol}$. We find field dwarfs of similar $T_\mathrm{eff}$ and $L_\mathrm{bol}$ best fit the overall and band-by-band features of TRAPPIST-1. Additionally, we present new [@Alle13] spectral indices for the SpeX SXD and FIRE spectra of TRAPPIST-1, both classifying it as intermediate gravity. Examining $T_\mathrm{eff}$, $L_\mathrm{bol}$, and absolute $JHKW1W2$ magnitudes versus optical spectral type places TRAPPIST-1 in an ambiguous location containing both field- and intermediate-gravity sources. Kinematics place TRAPPIST-1 within a subpopulation of intermediate-gravity sources lacking bonafide membership in a moving group with higher tangential and $UVW$ velocities. We conclude that TRAPPIST-1 is a field-age source with subtle spectral features reminiscent of a low surface gravity object. To resolve the cause of TRAPPIST-1’s intermediate gravity indicators we speculate two avenues which might be correlated to inflate the radius: (1) magnetic activity or (2) tidal interactions from planets. We find the M8 dwarf LHS 132 is an excellent match to TRAPPIST-1’s spectral peculiarities along with the M9$\beta$ dwarf 2MASS J10220489$+$0200477, the L1$\beta$ 2MASS J10224821$+$5825453, and the L0$\beta$ 2MASS J23224684$-$3133231 which have distinct kinematics making all three intriguing targets for future exoplanet studies.'
author:
- 'Eileen C. Gonzales'
- 'Jacqueline K. Faherty'
- Jonathan Gagné
- Johanna Teske
- Andrew McWilliam
- Kelle Cruz
bibliography:
- 'references.bib'
title: 'A Reanalysis of the Fundamental Parameters and Age of TRAPPIST-1[^1]'
---
Introduction
============
The majority of stars in our galaxy are low mass, M dwarfs being the most numerous with the longest main sequence lifetime [@Boch10]. Their low mass and abundance in the solar neighborhood make M dwarfs favorable targets for exoplanet observations. Their small radii enable easier detection of Earth-sized planets using transit and radial velocity methods, therefore they are prime targets when searching for rocky planets within a star’s habitable zone.
With numerous searches for exoplanets- such as *Kepler* (aimed at detecting planets around Sun-like stars @Boru10) and *TESS* (an all sky survey searching for planets smaller then Neptune around nearby stars, @Rick15)- understanding stellar properties of M dwarfs as exoplanet host stars is extremely pertinent to understanding planet habitability. [@Kane16] found 40% of all *Kepler* planet candidates (1) with radii less than 2 $R_\oplus$ and (2) lying within an optimistically-sized habitable zone, orbit stars cooler than 4000 K. This is despite cool dwarfs being less than 5% of initial *Kepler* targets [@Bata10]. [@Muld15c] and [@Gaid16] determined planets with radii of 1–3 $R_\oplus$ occur 2–4 times higher around M dwarfs than FGK stars. Furthermore, [@Dres15] estimate the frequency of habitable Earths around M dwarfs to be $2.5\pm0.2$ planets. Such objects would have radii of $1-4$ $R_\oplus$ and periods shorter than 200 days. They calculate an occurrence rate of $0.56\substack{+0.06 \\ -0.05}$ Earth-sized planets with periods shorter than 50 days and $0.46\substack{+0.07 \\ -0.05}$ super-Earths ($1-1.5\,R_\oplus$) with periods shorter than 50 days per early-type M dwarf. [@Ball18] predict that *TESS* will detect $900 \pm 350$ planets around $715\pm255$ M dwarfs spectral typed M1V–M4V.
Four nearby mid- to late-type M dwarfs with habitable zone planets are: 2MASS J23062928$-$0502285 (hereafter TRAPPIST-1), Proxima Centauri, LHS 1140, and Teegarden’s Star. TRAPPIST-1, a M7.5 dwarf at a distance of $\sim12.4$ parsecs, hosts a system of seven rocky Earth-sized exoplanets [@Gill16; @Gill17] with four lying in the habitable zone. Proxima Centauri, a moderately active M5.5 dwarf, is our nearest stellar neighbor 1.295 parsecs away. It hosts an earth-sized planet ($1.3\,M_\oplus$) that could have liquid water on its surface [@Angl16]. LHS 1140, a metal-poor M4.5 dwarf older than 5 Gyr and about 15 parsecs away, hosts a super-Earth ($\sim7M_\oplus$) and an Earth-sized planet ($\sim2M_\oplus$) with an Earth-like composition [@Ditt17; @Ment18]. Teegarden’s Star, an old magnetically quiet M6.5 dwarf located 3.8 parsecs away, was classified as intermediate gravity by [@Gagn15b] and hosts two Earth-sized planets both within the conservative habitable zone [@Zech19]. Having the most precise stellar parameters for these systems is critical for understanding planet habitability because stellar size and temperature drive the habitable zone boundaries.
While most M dwarfs are stars, late-type M dwarfs can be either stars or brown dwarfs depending on age. Brown dwarfs are low-mass, low-temperature objects unable to sustain stable hydrogen burning in their cores and thus cool throughout their lifetime. With masses $<75\ M_\mathrm{Jup}$, brown dwarfs lie between the boundary of low mass stars and planets [@Saum96; @Chab97]. They typically fall into three main age subpopulations: field dwarfs, low surface gravity dwarfs, and subdwarfs. Field dwarfs anchor the brown dwarf spectral classification system, while the low-gravity and subdwarfs show differences in their observed spectra deviating them from the field classification. Red infrared colors, enhanced metal oxide in the optical, and weak alkali lines differentiate low-gravity dwarfs from the field sources [@Kirk06; @Kirk10; @Cruz09; @Alle10]. High-likelihood or candidate membership in young nearby moving groups has been seen for many low-gravity sources [@Liu_13; @Liu_16; @Fahe16; @Kell16; @Schn16; @Gagn17]. Low-gravity dwarfs are further classified into two gravity groups- (1) very low gravity, designated by $\gamma$ in the optical or by VL-G in the infrared, and (2) intermediate gravity, designated by $\beta$ in the optical or INT-G in the infrared. Subdwarfs are low-luminosity metal-poor sources that exhibit blue near-infrared colors [@Burg03c; @Burg09a], kinematics consistent with halo-membership [@Burg08a; @Dahn08; @Cush09], enhanced metal-hydride absorption bands along with weak or absent metal oxides, and enhanced collision-induced H$_2$ absorption (@Burg03c and references therein).
In this paper we examine the fundamental parameters of TRAPPIST-1 to determine if it is a typical field M-dwarf host star or if it is more akin to low gravity or subdwarf equivalents. Previously published data on TRAPPIST-1 is presented in Section \[PubTrap\]. New FIRE and SpeX SXD spectra observations are discussed in Section \[Obs\]. Section \[motivation\] presents the motivation of our chosen samples and analysis process. Section \[FundParamTrap\] discusses how we derive the fundamental parameters of TRAPPIST-1 using its distanced-calibrated SED and the [@Fili15] method. Section \[CompTrap\] discusses comparative samples for TRAPPIST-1. Sections \[FieldAgeAssumption\] and \[YoungAgeAssumption\] examine the full SED of TRAPPIST-1 and the $Y$, $J$, $H$ and $K$ near-infrared (NIR) spectra of comparative objects. Section \[subdwarfNaIKIcomps\] examines whether an extreme subsolar metallicity for TRAPPIST-1 might explain anomalous spectral features. Section \[FinalThoughts\] makes concluding remarks on the age of TRAPPIST-1 after examining all samples. Section \[LHS132\] examines LHS 132 and other sources from [@Burg17] comparing their overall SEDs to TRAPPIST-1. Lastly, Section \[Discussion\] presents [@Alle13] indices and gravity classifications for the entire comparative sample, absolute magnitude and fundamental parameters versus spectral type comparisons, and a comparison of the kinematics of TRAPPIST-1 to $\beta$-gravity sources. We also present possible reasons for the gravity classification that TRAPPIST-1 receives.
Published Data on TRAPPIST-1 {#PubTrap}
============================
The discovery of a M-dwarf
--------------------------
TRAPPIST-1 was discovered by [@Gizi00] as part of a search using optical and near-infrared sky survey data from the Second Palomar Sky Survey (POSS-II) and the Two Micron All Sky Survey (2MASS). [@Gizi00] spectral typed TRAPPIST-1 as an M7.5 based on its optical spectrum taken by the RC spectrograph at Kitt Peak on the 4 m telescope. Additional optical spectra of TRAPPIST-1 are presented in [@Cruz07], [@Schm07], [@Rein09] and [@Burg15]. NIR spectra are presented in [@Tann12], [@Bard14], [@Cruz18], and this paper (FIRE). Thus, there are currently 5 optical and 4 NIR spectra of TRAPPIST-1.
*Gaia* DR2 (@GaiaDR1 [@GaiaDR2; @Lind18]) provides the most precise proper motion and parallax values for TRAPPIST-1 although it was also observed by numerous other surveys (@Cost06, @Schm07, @Wein16, and @Boss17, @vanG18). Radial velocity measurements of TRAPPIST-1 were reported in [@Rein09], [@Tann12] (using the NIRSPEC spectrum), and [@Burg15] (using the MagE spectrum). In this paper we present updated $UVW$ velocities using updated position and parallax values from *Gaia* DR2 paired with the [@Tann12] radial velocity.
The equivalent width of the 6563Å H$\alpha$ line, an indicator of activity for M-dwarfs, has been measured in many papers [@Gizi00; @Schm07; @Rein10; @Barn14; @Burg15]. The width measurements vary between $\sim2.3-7.7$Å. Ratios of log($L_\mathrm{\alpha}/L_\mathrm{bol}$) in [@Gizi00], [@Schm07] and [@Rein10] indicate that TRAPPIST-1 is a moderately active M dwarf.
Bolometric luminosity, effective temperature, radius, mass, gravity and age for TRAPPIST-1 were initially determined in [@Fili15] via distance-scaled SED fitting. After the discovery of TRAPPIST-1 as an exoplanet host star with seven rocky planets, [@Gill17] presented fundamental parameters from a Markov chain Monte Carlo (MCMC) analysis with *a priori* knowledge of stellar properties from [@Fili15]. Using their improved parallax measurement, [@vanG18] determined a bolometric luminosity almost two times as precise as the [@Fili15] value and used that to determine more accurate values of $T_\mathrm{eff}$, radius, and mass for TRAPPIST-1. Values for published data from the literature are listed in Table \[tab:Trapdata\].
[l c c ]{} R.A. & $23^h 06^m 29.36^s$ & 1\
Decl. & $-05 ^\circ 02' 29''.2$ & 1\
R.A. (epoch 2015.0) & $346.63 \pm 0.11$ & 2,3\
Decl.(epoch 2015.0) & $-5.043 \pm 0.093$ & 2,3\
Spectral type & M7.5 & 4\
$\pi$ (mas) & $80.45 \pm 0.12$ & 2,3\
$Gaia$ BP (mag) & $18.998 \pm 0.048$ & 2,5\
$Gaia$ RP (mag) & $14.1 \pm 0.01$ & 2,5\
PS $g$ (mag) & $19.35 \pm 0.02$ & 6\
PS $r$ (mag) & $17.87 \pm 0.01$ & 6\
PS $i$ (mag) & $15.13 \pm 0.01$ & 6\
PS $z$ (mag) & $13.73 \pm 0.01$ & 6\
PS $y$ (mag) & $12.97 \pm 0.01$ & 6\
2MASS $J$ (mag) & $11.354 \pm 0.022$ & 1\
2MASS $H$ (mag) & $10.718 \pm 0.021$ & 1\
2MASS $K_{s}$ (mag) & $10.296 \pm 0.023$ & 1\
WISE $W1$ (mag) & $10.042 \pm 0.023$ & 7\
WISE $W2$ (mag) & $9.80 \pm 0.02$ & 7\
WISE $W3$ (mag) & $9.528 \pm 0.041$ & 7\
WISE $W4$ (mag) & $<8.397$ & 7\
$\mu_\alpha$ (mas yr$^{-1}$) & $930.88\pm0.25$ & 2,3\
$\mu_\delta$ (mas yr$^{-1}$) & $-479.40\pm0.17$ & 2,3\
$V_{r}$ (km s$^{-1}$) & $-52.8 \pm 0.16$ & 8\
$V_\mathrm{tan}$ (km s$^{-1}$) & $61.69 \pm 0.10$ & 9\
$U$ (km s$^{-1}$) & $-44.1\pm0.1$ & 9\
$V$ (km s$^{-1}$) & $-67.2\pm0.3$ & 9\
$W$ (km s$^{-1}$) & $11.7\pm0.4$ & 9\
$X$ (pc) & $2.369 \pm 0.004$ & 9\
$Y$ (pc) & $6.41\pm0.01$ & 9\
$Z$ (pc) & $-10.38\pm0.02$ & 9\
$L_\mathrm{bol}$ log($L_*/L_{\odot}$) & $-3.216 \pm 0.016$ & 9\
$T_\mathrm{eff}$ (K) & $2628 \pm 42$ & 9\
Radius ($R_\mathrm{Jup}$) & $1.16 \pm 0.03$ & 9\
Mass ($M_\mathrm{Jup}$ )& $90 \pm 8$ & 9\
log $g$ (dex) & $5.21 \pm 0.06$ & 9\
Age (Gyr) & $0.5-10$ & 9\
Distance (pc) & $12.43 \pm 0.02$ & 9\
$[\mathrm{Fe/H}]$ (dex) & $+0.04 \pm 0.08$ & 10\
H$\alpha$ EW (Å) & $2.34-4.17$ & 11\
log($L_\mathrm{H\alpha}/L_\mathrm{bol}$) & min:$-4.85$, max:$-4.60$ & 11\
Discovery as an exoplanet host star
-----------------------------------
TRAPPIST (the TRansiting Planet and PlanetIsimals Small Telescope) monitored TRAPPIST-1 in the near-infrared (NIR) from mid September 2015 to the end of December 2015. Follow up photometry in the optical on the Himalayan Chandra 2-meter Telescope and in the NIR with the Very Large Telescope and UK Infrared Telescope helped to confirm signatures of three of exoplanets: TRAPPIST-1b, TRAPPIST-1c, and TRAPPIST-1d [@Gill16]. TRAPPIST-1d was later identified by follow up observations to be the signature of four planets: TRAPPIST-1d, TRAPPIST-1e, TRAPPIST-1f, and TRAPPIST-1g, along with the discovery of TRAPPIST-1h [@Gill17]. All seven of the planets have Earth-sized radii, ranging from 0.755 to 1.127 $R_\mathrm{Earth}$, with four of the planets TRAPPIST-1d,e,f, and g, lying in the habitable zone [@Gill17]. These planets are in a zone where temperatures are cool enough to potentially have long-lived liquid water present on the surface.
Observations {#Obs}
============
FIRE
----
We used the Folded-port InfraRed Echellette (FIRE; @Simcoe13) spectrograph on the 6.5m Baade *Magellan* telescope to obtain near-infrared spectra of TRAPPIST-1. Observations were taken on 2017 July 28 using echellette mode and the 0.$\arcsec$45 slit, high gain (1.2 e-/DN), and sample-up-the-ramp (SUTR) readout covering the full 0.8 - 2.5 $\micron$ band. Each exposure was 600s long with ABBA nodding, and are bracketed by quartz lamp, ThAr lamp, and telluric standards. Airmass ranged from 1.7-1.0 and seeing from 0.7-1.1 over the night. The night sky emission lines were significantly larger than the internal Th-Ar comparison lines, even though the system was in best focus; this suggests that FIRE was not properly collimated at the time of observation. The result is that the resolving power we obtained was $\lambda$/$\Delta \lambda \sim$ 6000, not the R$\sim$8,000 that should have been obtained. This causes lines in our data to be shallower but wider with the equivalent widths preserved. All FIRE exposures of TRAPPIST-1 obtained on 2017 UT July 27 outside of transit were combined and reduced with the Interactive Data Language (IDL) FireHose v2 package[^2] [@Boch09; @zenodofirehose], as described in [@Gagn15b].
SpeX
----
The following new SpeX spectra were obtained for objects we use in a comparative analysis to TRAPPIST-1.
### SXD
A SpeX SXD spectrum was obtained for 2MASS J06085283$-$2753583 on 2007 November 13 across the $0.7-2.55\,\upmu$m region. The $0.\arcsec5$ slit was used providing a resolving power of $\lambda$/$\Delta \lambda \sim1200$. We obtained a two 200s exposures and four 300s exposures for a total integration time of 20 minutes using ABBA nodding. The slit was aligned to the parallactic angle and we observed at an airmass of 1.49. The spectrum was telluric corrected and flux calibrated using the spectrum of the A0 V standard HD 52487 taken at a similar airmass. Internal flat-field and Ar arc lamps exposures were taken for pixel response and wavelength calibration. The data was then reduced using standard procedures and the SpeXtool Package [@Cush04].
Is TRAPPIST-1 Young? {#motivation}
====================
[@Burg17] noted TRAPPIST-1 exhibited weaker FeH absorption and a more triangular $H$ band, features that are typically associated with youth (see @Fahe16 for details). Using the [@Alle13] surface gravity indices, they determined the low-resolution SpeX spectrum of TRAPPIST-1 displayed signs of an intermediate gravity object, suggesting a young age. However, from examination of the kinematics and the lack of enhanced VO absorption in the SpeX prism spectrum, [@Burg17] concluded that the $\beta$ gravity classification may have arisen from other physical factors and thus is unrelated to youth. Indeed, all previous studies [@Gizi00; @Fili15; @Fahe09] indicated that TRAPPIST-1 was a field object.
The low gravity indicators in the spectrum of TRAPPIST-1 motivate the remainder of the paper where we create comparison subsamples to examine whether youth or some other physical factor drive the observed parameters. Throughout our comparisons we examine whether the overall SED shape or specific features in the new FIRE medium-resolution NIR spectrum of TRAPPIST-1 show signs of low surface gravity. We also determine gravity indices for all objects in our sample as another aspect of our comparison to TRAPPIST-1, which is discussed in depth in Section \[AL13Indicescomparison\].
Fundamental Parameters of TRAPPIST-1 {#FundParamTrap}
====================================
Fundamental parameters for TRAPPIST-1 were determined from its distance-calibrated SED using the technique of [@Fili15]. Parameter values were determined using SEDkit[^3], which requires spectra, photometry, and a parallax to create the distance-scaled SED and to determine the bolometric luminosity. The spectra, photometry, and parallax used in the generation of the SED of TRAPPIST-1 can be found in Tables \[tab:Trapdata\] and \[tab:opticalphot\]–\[tab:SpectraReferences\].
Using the optical and NIR spectra, we first construct a composite spectrum of TRAPPIST-1. The overlapping region from $0.8305-0.95\, \upmu$m (shown as a fuzzy blue-green line in Figure \[fig:Regimes\]) was combined as an average. The composite spectrum is then scaled to the absolute magnitudes in each filter. We do not create synthetic magnitudes, those determined through empirical relations in the absence of photometric data as in [@Fili15], instead we only use observed photometric data. The SED of TRAPPIST-1 is shown in Figure \[fig:Regimes\], with the various components labeled.
The bolometric luminosity ($L_\mathrm{bol}$) was determined by integrating under the distance-calibrated SED from 0–1000 $\upmu$m, using a distance of $12.43\pm0.02$ pc. To obtain a radius estimate and the effective temperature ($T_\mathrm{eff}$), we used the DMEstar models [@Feid12] to extend the brown dwarf [@Saum08] hybrid cloud evolutionary models into the low-mass stellar range. They were connected via a cubic spline interpolation in the regions with no evolutionary model coverage. The same was done using the @Chab00 evolutionary models. For all three models we use an age range of $0.5-10$ Gyr, corresponding to the field ultracool dwarf age range [@Fili15]. The final radius range was set as the maximum and minimum from all model predictions as done in @Fili15. The effective temperature was calculated using the inferred radius along with the bolometric luminosity using the Stefan-Boltzmann law. The uncertainty on the $T_\mathrm{eff}$ comes primarily from the uncertainty in the age of the system (leading to the range of radii) and the SED flux measurement. However, as noted in [@Dupu13] slight differences in radii do not have a large effect on the calculated $T_\mathrm{eff}$. The range of masses was determined using the @Saum08 and @Chab00 evolutionary models.
Using this approach, we derived the following parameters assuming a field age: $L_\mathrm{bol} = -3.216\pm0.016$, $T_\mathrm{eff} = 2628\pm42$K, $R = 1.16\pm0.03$$R_\mathrm{Jup}$, $M = 90\pm8 \;M_\mathrm{Jup}$, $\mathrm{log}\,g = 5.21\pm0.06$dex. Given the speculation on the age of TRAPPIST-1, we also repeat this process and assume an age range of $0-0.5$ Gyr to address the intermediate gravity classification. The fundamental parameters derived for TRAPPIST-1 for both age assumptions are also listed in Table \[tab:LitFunParams\], which compares our values to literature values.
[l c c c c c c c c c]{} log $L_*/L_{\odot}$ & $-3.216 \pm 0.016$ & $-3.216 \pm 0.016$ & $\cdots$ & $-3.281\pm 0.028$ & $-3.28 \pm 0.03$ & $-3.28 \pm 0.016$\
$T_\mathrm{eff}$ (K) & $2628 \pm 42$& $2310 \pm 230$ & $\cdots$ & $2559 \pm 50$ & $2557 \pm 64$ & $2516 \pm 41$\
Radius ($R_\mathrm{Jup}$) & $1.16 \pm 0.03$ & $1.5 \pm 0.3$ & $1.177 \pm 0.029$& $1.138 \pm 0.035$ & $1.14 \pm 0.04$ & $1.177 \pm 0.029$\
Mass ($M_\mathrm{Jup}$) & $90 \pm 8$ & $49 \pm 34$ & $83.8 \pm 7.3$ & $84 \pm 7.6$ & $86.07 \pm 9.28$ & $93.2 \pm 6.3$\
log$g$ & $5.21 \pm 0.06$ & $4.61 \pm 0.54$ & $\cdots$ & $\cdots$ & $5.22 \pm 0.08$ & $\cdots$\
Age (Gyr) & $0.5-10$ & $<0.5$ & $7.6 \pm 2.2$& $>0.5\tablenotemark{a}$ & $0.5-10$ & $\cdots$\
Parallax (mas) & $80.45\pm0.12$ & $80.45\pm0.12$ & $\cdots$& $82.58\pm2.58$ & $82.58\pm2.58$ & $82.4 \pm 0.8$\
Fundamental parameter comparison to the literature
--------------------------------------------------
Table \[tab:LitFunParams\] contains our calculated fundamental parameters assuming two age ranges (1) $0.5-10$ Gyr, the field age constraint from [@Fili15] and (2) $<0.5$ Gyr, to address the intermediate gravity classification. Also listed in Table \[tab:LitFunParams\] are literature values for derived fundamental parameters of TRAPPIST-1 from [@Fili15], [@Gill17], [@Burg17], and [@vanG18]. Our bolometric luminosity value differs from previous measurements by up to 2 $\sigma$, which can be accounted for by our use of the more precise *Gaia* DR2 parallax measurement as well as replacing the wider Johnson-Cousins band measurements with the narrower band Pan-STARRS values. However, our radius, $T_\mathrm{eff}$, and mass values agree with literature values within 1$\sigma$ with the exception of our $0.5-10$ Gyr $T_\mathrm{eff}$ value which differs by 1.4 $\sigma$ to the [@vanG18] value.
A Comparative Sample for TRAPPIST-1 {#CompTrap}
===================================
Sample Selection and Properties
-------------------------------
[l c c c c c c c c c c c c]{} 23 06 29.36 & $-$5 02 29.2 & 2MASS J23062928$-$0502285 & TRAPPIST-1 & 1 & M7.5 & 1 & $\cdots$ &$\cdots$ & $80.45 \pm 0.12$ & 2\
&&& &&&&\
&&& &&&&\
03 20 59.65 & $+$18 54 23.3 & 2MASS J03205965$+$1854233 & J0320$+$1854 & 3 & M8 & 3 & $\cdots$ & $\cdots$ &$68.28 \pm 0.15$ & 2\
10 13 07.35 & $-$13 56 20.4 & SSSPM J1013$-$1356 & J1013$-$1356 & 4 & sdM9.5 & 3 &$\cdots$ & $\cdots$ & $18.37 \pm 0.39$ & 2\
14 56 38.31 & $-$28 09 47.3 & LHS 3003 & $\cdots$ & 5 & M7 & 3 & M7 & 6 & $141.69\pm 0.11$ & 2\
16 55 35.29 & $-$8 23 40.1 & vB 8 & $\cdots$ & 7 & M7 & 8 & M7 & 9 & $153.81 \pm 0.11 $ & 2\
19 16 57.62 & $+$05 09 02.2 & vB 10 & $\cdots$ & 7 & M8 & 10 & M8 & 9 & $168.96 \pm 0.13$ & 2\
&&& &&&&\
04 36 27.88 & $-$41 14 46.5 & 2MASS J04362788$-$4114465 & J0436$-$4114& 11 & M8$\beta$ & 12 & M9 VL-G & 13 & $25.30 \pm 0.12$ & 2\
06 08 52.83 & $-$27 53 58.3 & 2MASS J06085283$-$2753583 & J0608$-$2753 & 14 & M8.5$\gamma$ & 15 & L0 VL-G & 13 & $22.65 \pm 0.19$ & 2\
12 47 44.28 & $-$38 16 46.4& 2MASS J12474428$-$3816464 & J1247$-$3816 & 16 & $\cdots$ &$\cdots$ & M9 VL-G & 16 & $11.81 \pm 0.48$ & 2\
&&& &&&&\
04 43 37.61 & $+$00 02 05.1 & 2MASS J04433761$+$0002051 & J0443$+$0002 & 17 & M9$\gamma$ & 18 & L0 VL-G & 13 & $47.41 \pm 0.19$ & 2\
05 18 46.16 & $-$27 56 45.7 & 2MASS J05184616$-$2756457 & J0518$-$2756 & 19 & L1$\gamma$ & 20 & L1 VL-G & 13 & $17.27 \pm 0.81$ & 2\
12 07 48.36 & $-$39 00 04.3 & 2MASS J12074836$-$3900043 & J1207$-$3900 & 16 & L0$\gamma$ & 16 & L1 VL-G & 16 & $14.92 \pm 0.90$ & 2\
&&& &&&&\
08 53 36.19 & $-$03 29 32.1 & 2MASS J08533619$-$0329321 & J0853$-$0329 & 21 & M9 & 22 & M9 & 6 & $115.30 \pm 0.11$ & 2\
10 48 14.64 & $-$39 56 06.2 & DENIS–P J1048.0$-$3956 & J1048$-$3956 & 23 & M9 & 23 &$\cdots$ & $\cdots$ & $247.22 \pm 0.12$ & 2\
14 44 20.67 & $-$20 19 22.3 & SSSPM J1444$-$2019 & J1444$-$2019 & 24 & sdM9 & 24 & $\cdots$ &$\cdots$ &$57.80 \pm 0.56$ & 25\
17 12 51.21 & $-$5 07 24.9 & GJ 660.1B & $\cdots$ & 26 & $\cdots$ & $\cdots$ & d/sdM7 & 27 & $50.1 \pm 3.62$ & 28\
18 35 37.90 & $+$32 59 54.5 & 2MASS J18353790+3259545 & J1835$+$3259 & 29 & M8.5 & 30 & $\cdots$ & $\cdots$ & $175.82 \pm 0.09$ & 2\
20 00 48.41& $-$75 23 07.0 & 2MASS J20004841$-$7523070 & J2000$-$7523 & 31 & M9$\gamma$ & 32 & M9$\gamma$ & 32 & $33.95 \pm 0.15$ & 2\
&&& &&&&\
05 32 53.46 & $+$82 46 46.5 & 2MASS J05325346$+$8246465 & J0532$+$8246 & 33 & sdL7 & 34 & $\cdots$ &$\cdots$ & $40.24 \pm 0.64$ & 2\
12 56 37.13 & $-$02 24 52.4 & SDSS J125637.13$-$022452.4 & J1256$-$0224& 35 & sdL3.5 & 36 & $\cdots$ &$\cdots$ & $12.55 \pm 0.72$ & 2\
14 39 00.31 & $+$18 39 38.5 & LHS 377 & $\cdots$ & 37 & sdM7 & 38 & $\cdots$ &$\cdots$ & $25.75 \pm 0.10$ & 2\
16 10 29.00 & $-$00 40 53.0 & LSR J1610$-$0040 & J1610$-$0040 & 39 & sdM7 & 34 & $\cdots$ &$\cdots$ & $30.73 \pm 0.34$ & 40\
20 36 21.65 & $+$51 00 05.2 & LSR J2036$+$5059 & J2036$+$5059 & 41 & sdM7.5 & 42 & $\cdots$ &$\cdots$ & $23.10 \pm 0.29$ & 25\
&&& &&&&\
01 02 51.00 & $-$37 37 43.0 & LHS 132 & $\cdots$ & 5 & M8 & 43 & M8 & 44 & $87.87 \pm 0.12$ & 2\
23 41 28.68 & $-$11 33 35.6 & 2MASS J23412868$-$1133356 & J2341$-$1133 & 19 & M8 & 19 & $\cdots$ &$\cdots$ & $20.33 \pm 0.24$ & 2\
23 52 05.00 & $-$11 00 43.5 & 2MASS J23520507$-$1100435 & J2352$-$1100 & 19 & M7 & 19 & M8$\beta$ & 45 & $24.88 \pm 0.51$ & 2\
[l c c c c c c c c c]{} TRAPPIST-1 & M7.5 & $\cdots$ & $-3.216 \pm 0.016$ & $2628 \pm 42$ & $1.16 \pm 0.03$ & $90 \pm 8$ & $5.21 \pm 0.06$ & $0.5-10$ & $12.43 \pm 0.02$\
TRAPPIST-1 & M7.5 & $\cdots$ & $-3.216 \pm 0.016$ & $2310 \pm 230$ & $1.5 \pm 0.3$ & $49 \pm 34$ & $4.61 \pm 0.54$ & $0-0.5$ & $12.43 \pm 0.02$\
LHS 132 & M8 & M8 & $-3.264 \pm 0.015$ & $2579 \pm 41$ & $1.14 \pm 0.03$ & $87 \pm 8$ & $5.21 \pm 0.07$ & $0.5-10$ & $11.38 \pm 0.02$\
J0320$+$1554 & M8 & $\cdots$ & $-3.226 \pm 0.007$ & $2613 \pm 35$ & $1.16 \pm 0.03$ & $89 \pm 8$ & $5.21 \pm 0.06$ & $0.5-10$ & $14.65 \pm 0.03$\
J0436$-$4114 & M8$\beta$ & M9$\gamma$ & $-2.927 \pm 0.019$ & $2560 \pm 260$ & $1.7 \pm 0.34$ & $60 \pm 43$ & $4.58 \pm 0.56$ & $0-0.5$ & $39.52 \pm 0.19$\
J0443$+$0002 & M9$\gamma$ & L0$\gamma$ & $-3.194\pm 0.009$ & $2232 \pm 29$ & $1.65 \pm 0.04$ & $25 \pm 2$ & $4.34 \pm 0.06$ & $0.021-0.027$ & $21.09 \pm 0.08$\
J0518$-$2756 & L1$\gamma$ & L1$\gamma$ &$-3.273 \pm 0.041$ & $2229 \pm 57$ & $1.51 \pm 0.03$ & $31 \pm 3$ & $4.51 \pm 0.05$ & $0.038-0.048$ & $57.9 \pm 2.7$\
J0532$+$8246 & sdL7 & $\cdots$ &$-4.28 \pm 0.07$ & $1670 \pm 70$ & $0.84 \pm 0.02$ & $72 \pm 5 $ & $5.4 \pm 0.05$ & $5-10$ & $24.85 \pm 0.39$\
J0608$-$2753 & M8.5$\gamma$ & L0$\gamma$ &$-2.996 \pm 0.014$ & $2510 \pm 250$ & $1.64 \pm 0.33$ & $ 57 \pm 40 $ & $4.59 \pm 0.56$ & $0-0.5$ & $44.15 \pm 0.38$\
J0853$-$0329 & M9 & M9 & $-3.485 \pm 0.033$ & $2330 \pm 70$ & $1.08 \pm 0.05$ & $78 \pm 10$ & $5.21 \pm 0.1$ & $0.5-10$ & $8.67 \pm 0.01$\
J1013$-$1356 & sdM9.5 & $\cdots$ &$-3.303 \pm 0.027$ & $2628 \pm 43$ & $1.05 \pm 0.01$ & $92 \pm 1$ & $5.32 \pm 0.05$ & $5-10$ & $54.4 \pm 1.2 $\
J1048$-$3956 & M9 & $\cdots$ & $-3.485 \pm 0.019$ & $2330 \pm 60$ & $1.08 \pm 0.05$ & $78 \pm 10$ & $5.21 \pm 0.1$ & $0.5-10$ & $4.045 \pm 0.002$\
J1207$-$3900 & L0$\gamma$ & L1$\gamma$ &$-3.337 \pm 0.053$ & $2013 \pm 66$ & $1.72 \pm 0.04$ & $15 \pm 1$ & $4.1 \pm 0.03$ & $0.007-0.013$ & $67 \pm 4$\
J1247$-$3816 & $\cdots$ & M9$\gamma$ & $-2.84 \pm 0.039$ & $2630 \pm 290$ & $1.79 \pm 0.39$ & $67 \pm 49$ & $4.56 \pm 0.57$ & $0-0.5$ & $84.6 \pm 3.5$\
J1256$-$0224 & sdL3.5 & $\cdots$ & $-3.63 \pm 0.05$ & $2301\pm71$ & $0.94\pm0.02$ & $83\pm2$ & $5.37\pm0.01$ & $0.5-10$ & $79.7\pm4.6$\
LHS 377 & sdM7 & $\cdots$ & $ -3.041 \pm 0.008$ & $2840 \pm 60$ & $1.212 \pm 0.05$ & $106 \pm 1$ & $5.25 \pm 0.05$ & $5-10$ & $38.84 \pm 0.16$\
J1444$-$2019 & sdM9 & $\cdots$ & $-3.49\pm 0.02$ & $2442\pm68$ & $0.98\pm 0.05$ & $87\pm1$ & $5.35\pm0.05$ & $5-10$ & $17.3\pm0.17$\
LHS 3003 & M7 & M7 & $-3.224 \pm 0.012$ & $2616 \pm 38$ & $1.16 \pm 0.03$ & $89 \pm 8$ & $5.21 \pm 0.06$ & $0.5-10$ & $7.06 \pm 0.01$\
J1610$-$0040 & sdM7 & $\cdots$ & $-2.98 \pm 0.01$ & $2880 \pm 20$ & $1.27 \pm 0.01$ & $111 \pm 1$ & $5.23 \pm 0.05$ & $5-10$ & $32.54 \pm 0.36$\
vB 8 & M7 & M7 & $-3.192 \pm 0.006$ & $2642 \pm 35$ & $1.18 \pm 0.03$ & $91 \pm 7$ & $5.21 \pm 0.06$ & $0.5-10$ & $6.501\pm0.005$\
GJ 660.1B & $\cdots$ & d/sdM7 & $-3.523 \pm 0.063$ & $2409 \pm 91$ & $0.97 \pm 0.02$ & $86 \pm 2$ & $5.36 \pm 0.01$ & $0.5-10$ & $20 \pm 1.4$\
J1835$+$3295 & M9 & M9& $-3.502\pm0.016$ & $2321\pm58$ & $1.07\pm0.05$ & $77\pm10$ & $5.21\pm 0.1$ & $0.5-10$ & $5.687\pm 0.003$\
vB 10 & M8 & M8 & $-3.298 \pm 0.018$ & $2540 \pm 52$ & $1.13 \pm 0.04$ & $86 \pm 8$ & $5.22 \pm 0.07$ & $0.5-10$ & $5.918 \pm 0.005$\
J2000$-$7523 & M9$\gamma$ & M9$\gamma$ & $-3.04\pm 0.02$ & $2389\pm 44$ & $1.72\pm0.05$ & $30\pm3$ & $4\pm0.07$ & $0.021-0.027$ & $29.45\pm0.13$\
J2036$+$5059 & sdM7.5 & $\cdots$ & $-3.057 \pm 0.035$ & $2832 \pm 58$ & $1.2 \pm 0.01$ & $105 \pm 1$ & $5.25 \pm 0.05$ & $0.5-10$ & $43.29 \pm 0.54$\
J2341$-$1133 & M8 & $\cdots$ & $-2.921\pm0.015$ & $2898\pm41$ & $1.34\pm0.03$& $109\pm7$ & $5.18\pm0.04$ & $0.5-10$& $49.18\pm0.58$\
J2352$-$1100 & M7 & M8$\beta$ & $-2.797\pm0.018$ & $2913\pm56$& $1.53\pm0.05$& $100\pm11$& $5.02\pm0.06$& $0.13-0.2$& $40.19\pm0.82$\
To determine if TRAPPIST-1 exhibits similar or different features compared to the average field M dwarf, we constructed three comparative samples.
- Sample \#1: Assumes TRAPPIST-1 is a field age ($0.5-10$ Gyr) source and contains comparative field, very low-gravity ($\gamma$), and old dwarfs with similar effective temperatures (within $1\, \sigma$ of TRAPPIST-1) and/or bolometric luminosity (within up to $2.5\, \sigma$) of TRAPPIST-1
- Sample \#2: Assumes an age of $<0.5$ Gyr for TRAPPIST-1 and contains field, very low-gravity ($\gamma$), and old dwarfs with similar effective temperatures within $1\, \sigma$
- Sample \#3: Assumes TRAPPIST-1 is a field age and contains subdwarfs of varying $T_\mathrm{eff}$ and $L_\mathrm{bol}$ (sdM7 and later) with medium resolution data (R $\sim 5000$).
The effective temperature samples compare the overall SED shape and spectral features of TRAPPIST-1 to objects with similar atmospheric chemistry. The bolometric luminosity sample examines how the flux of TRAPPIST-1 is redistributed across different wavelengths compared to comparable $L_\mathrm{bol}$ sources. The subdwarf sample explores the and doublets in the $J$ band to see if TRAPPIST-1 exhibits similarities with these metal poor old objects.
Objects in each comparative sample were chosen from (1) [@Fili15], which examined a large sample of field and low-gravity objects, (2) [@Fahe16], which examined a large sample of low-gravity sources, or (3) [@Gonz18], which examined subdwarfs later than sdM6 with parallax measurements. The bolometric luminosities in each sample were empirically derived, while their effective temperatures were semi-empirically derived using radii from evolutionary models depending on the age. Ages were determined by updated membership in moving groups (Faherty et al. in prep) for young sources from [@Fahe16]. For field sources from [@Fili15] we use their field dwarf age range of 0.5–10 Gyr, and for subdwarfs from [@Gonz18] we use their subdwarf age range of 5–10 Gyr. Exact evolutionary models used for radii values are listed in [@Fili15], [@Fahe16], and [@Gonz18].
In order to resolve features in a band-by-band comparative analysis of the spectra, we required objects to have medium-resolution NIR data ($\lambda/\Delta\lambda > 1000$ at $J$ band). Field dwarf comparison SEDs chosen from [@Fili15] were reconstructed with the same data used in that work, with the exception of replacing their low-resolution SpeX data with medium-resolution NIR data in this work. Low-gravity dwarf comparison SEDs from [@Fahe16] were constructed with new NIR FIRE data or re-reduced spectra. Subdwarf comparison SEDs were constructed with the same data from [@Gonz18], with the exception of GJ 660.1B which was not part of that sample, SSSPM J1013$-$1356 which only uses the NIR spectrum in this work, and LSR 2036$+$5059 which includes synthetic WISE photometry points for proper appending of the Rayleigh-Jeans tail.
All SEDs constructed for this paper used measured photometric values alone with the exceptions listed above. We did not incorporate synthetic photometry as done by [@Fili15] and [@Fahe16]. We also used updated parallax measurements from *Gaia* DR2 for objects chosen from [@Fili15] and [@Fahe16]. Therefore our values differ slightly but fall within $\sim1\, \sigma$ for all values except bolometric luminosity.
Sample \#1 contains 5 field sources that have similar $T_\mathrm{eff}$ and $L_\mathrm{bol}$, three sources with similar $T_\mathrm{eff}$, and three with similar $L_\mathrm{bol}$ for a total of 8 sources when examining $T_\mathrm{eff}$ or $L_\mathrm{bol}$. Sample \#2 contains 8 sources- 3 field dwarfs, 3 subdwarfs, and 2 low-gravity sources. Sample \#3 contains the 5 subdwarfs from [@Gonz18] which have medium-resolution $J$-band data. Details for all comparative objects are listed in Table \[tab:Sample\]. The corresponding parallaxes, photometry, and spectra used in the creation of the SEDs are listed in Tables \[tab:opticalphot\]–\[tab:SpectraReferences\]. Table \[tab:SampleFunParams\] lists the derived fundamental parameters for TRAPPIST-1 along with the comparative objects.
Spectral Analysis For Sample \#1: Assuming a Field Age for TRAPPIST-1 {#FieldAgeAssumption}
=====================================================================
Full SED Comparisons {#FullSEDFeildAssumption}
--------------------
Under the assumption that TRAPPIST-1 is field age, we present the overall SEDs for objects of equivalent temperature and/or bolometric luminosity to that of TRAPPIST-1 in Figures \[fig:TeffLbolSEDs\] and \[fig:Fieldzoom\]. The sample is composed of three age subsets: field, low-gravity (young), and subdwarfs (old). All objects of similar $T_\mathrm{eff}$ are within $1\,\sigma$ of TRAPPIST-1 and of similar $L_\mathrm{bol}$ are within $2.5\,\sigma$.
Figure \[fig:TeffLbolSEDs\]a compares the overall SED assuming an older age for TRAPPIST-1 with field dwarfs of similar $T_\mathrm{eff}$ and $L_\mathrm{bol}$ across $0.3-31\,\upmu$m. All comparative field dwarfs in the sample are within one spectral type of TRAPPIST-1. From $0.65-1\, \upmu$m there is a spread in the SEDs (zoomed version in Figure \[fig:Fieldzoom\]a), however this spread does not appear to correspond to temperature or bolometric luminosity of the objects. There appears to be a tighter temperature-dependent sequence from $1.28-1.8\, \upmu$m, which can be seen in Figure \[fig:Fieldzoom\]b most clearly from the start of the $H$ band out to $1.70\,\upmu$m. Other than the large spread in the red optical, the overall SED shape of TRAPPIST-1 is similar to all sources in this sub-sample.
Figure \[fig:TeffLbolSEDs\]b compares TRAPPIST-1 to very low-gravity ($\gamma$) dwarfs of similar $T_\mathrm{eff}$ sample across the $0.5-23\,\upmu$m region. The low-gravity comparison objects are not members of known moving groups so we can’t give a definitive age bracket for them however, their spectra show strong indications of youth such as triangular $H$-band and weakened FeH (see @Fahe16, Faherty et al. in prep). From $\sim 0.80-1\,\upmu$m, we see TRAPPIST-1 directly on top of J0608$-$2753, however from $J$ band onward all low-gravity sources are much brighter than TRAPPIST-1.
Figure \[fig:TeffLbolSEDs\]c displays the overall SED of TRAPPIST-1 against an older subdwarf of similar $T_\mathrm{eff}$ and $L_\mathrm{bol}$ across the $0.5-13\,\upmu$m region. From $\sim0.7-1\, \upmu$m SSSPM J1013$-$1356 (hereafter J1013$-$1356) is brighter than TRAPPIST-1, while beyond $1\, \upmu$m TRAPPIST-1 is much brighter. TRAPPIST-1 is clearly poorly fit by the subdwarf in this sample.
SEDs of low-gravity objects with equivalent $L_\mathrm{bol}$ (within $2.5\, \sigma$) of TRAPPIST-1 across the $0.5-14\,\upmu$m region are shown in Figure \[fig:TeffLbolSEDs\]d. 2MASS J04433761$+$0002051 (hereafter J0443$+$0002) is a member of the $21-27$ Myr [@Bell15] $\beta$ Pictoris moving group, 2MASS J05184616$-$2756457 (hereafter J0518$-$2756) is a member of the $38-48$ Myr [@Bell15] Columba moving group, and lastly 2MASS J12074836$-$3900043 (hereafter J1207$-$3900) is a member of the $7-13$ Myr [@Bell15] TW Hydra moving group (@Fahe16, Faherty et al. in prep.). From $0.5\,\upmu$m to the start of the $J$ band, TRAPPIST-1 is brighter than all low-gravity sources, while in the $J$ and $H$ bands J0443$+$0002 and TRAPPIST-1 are of similar brightness. Beyond $W1$ TRAPPIST-1 is fainter than the low-gravity sources.
Overall under the assumption of field age for TRAPPIST-1, comparing the shape to field, low-gravity, and subdwarfs leads to the strong conclusion that this is a field source.
Band-by-Band Analysis {#bandbybandField}
---------------------
While the overall SED comparisons give a general examination of where flux is matched between sources, the subtle but significant feature detail is lost. Detailed line analysis can be used to tease out signatures of gravity or metallicity therefore in this subsection we compare NIR band-by-band features.
### Similar $T_\mathrm{eff}$
Figure \[fig:Teffbandbyband\]a shows the $0.95-1.10\, \upmu$m $Y$-band data with FeH, VO, and H$_2$O features labeled. The spectra for the $Y$ band were normalized by averaging the flux taken across the relatively featureless $0.98-0.988\, \upmu$m region. From first glance, we see that TRAPPIST-1 appears to be most similar in shape in the $Y$ band to the low-gravity dwarf J0436$-$4114 and the field dwarf vb10. The slope from $0.95 - 0.99\, \upmu$m of TRAPPIST-1 is most similar to those of the field dwarfs however, it also matches the slope of the low-gravity dwarf J0436$-$4114. The Wing-Ford FeH band-head of TRAPPIST-1 appears to be of similar depth to J0436$-$4114, but only slightly shallower than vB 8 and vB 10. From $1-1.05\,\upmu$m the spectrum of TRAPPIST-1 overlaps that of J0436$-$4114, but shares the shape of the field dwarfs more than that of the low-gravity dwarfs. There is an indication of slight VO absorption in the spectrum of TRAPPIST-1, similar in depth to the low-gravity dwarf J0436$-$4114. From all of these features we therefore conclude the TRAPPIST-1 $Y$-band spectrum exhibits a blend of both field and low-gravity dwarf features.
Figure \[fig:Teffbandbyband\]b shows the $1.12-1.35\, \upmu$m $J$-band data with FeH and H$_2$O molecular features, as well as the and alkali doublets labeled. The $J$-band spectra were normalized by the average flux over the featureless $1.29-1.31\, \upmu$m region. The depth of the and doublets of TRAPPIST-1 are slightly deeper than the field dwarfs and much deeper than the low-gravity dwarfs. The FeH absorption of TRAPPIST-1 is similar to the field dwarfs, while beyond the $1.25\,\upmu$m doublet the shape is similar for all objects in the sample. TRAPPIST-1 has some $J$-band spectral features similar to field dwarfs, while others differ from both the field and low-gravity dwarfs.
Figure \[fig:Teffbandbyband\]c shows the $1.42-1.80\, \upmu$m $H$-band data with FeH and H$_2$O molecular features labeled. The $H$-band spectra were normalized by the average flux over the featureless $1.5-1.52\, \upmu$m region. The $H$-band shape of TRAPPIST-1 is similar to the field objects, however its overall shape is also similar to J0436$-$4114 but slightly less triangular. There are no features in the $H$-band spectrum of TRAPPIST-1 that match those of the low-gravity dwarf J0608$-$2753. In the $H$ band, TRAPPIST-1 is more like a field object.
Figure \[fig:Teffbandbyband\]d shows the $2.0-2.35\, \upmu$m $K$-band data with H$_2$O, CO, and collision-induced H$_2$ absorption features labeled. The $K$-band spectra were normalized by the average flux over the $2.16-2.20\, \upmu$m region due to the relatively flat spectral region. TRAPPIST-1 has a visible doublet like the field dwarfs, while the low-gravity dwarfs don’t display this feature significantly or at all. The depth of TRAPPIST-1’s CO lines are similar to the field dwarfs, while the low-gravity dwarfs have shallower CO absorption lines. Thus in the $K$ band TRAPPIST-1 best matches the field dwarfs.
### Similar $L_\mathrm{bol}$
In the $Y$ band, Figure \[fig:Lbolbandbyband\]a, the depth of the Wing-Ford FeH band head for TRAPPIST-1 is similar to all comparative sources except J0320$+$1854 and J1207$-$3900 which are deeper than TRAPPIST-1. The shape of the spectrum carved out by the longer FeH band ($\sim 1-1.04\,\upmu$m) for TRAPPIST-1 is most similar first to J0443$+$0002 and second to the field dwarfs.
The and doublets of TRAPPIST-1 appears to be deeper than all comparative objects in Figure \[fig:Lbolbandbyband\]b. Interestingly, the low-gravity dwarfs J0518$-$2756 and J1207$-$3900 have relativity deep alkali lines for low-gravity sources. The shape of the FeH feature, particularly near $1.20\,\upmu$m, for TRAPPIST-1 is shallower than the low-gravity dwarfs, more similar to the field dwarfs. Between the second and the H$_2$O band we see that J0518$-$2756 and J1207$-$3900 slope slightly redward, while TRAPPIST-1 has a flat slope like the other field objects and J0443$+$0002. Thus in the $J$ band we see TRAPPIST-1 showing a hybrid of features both similar to field and low-gravity dwarfs like J0443$+$0002.
Figure \[fig:Lbolbandbyband\]c shows the $H$ band, where the low-gravity dwarfs are more triangular in shape compared to TRAPPIST-1. Again, we see that the $H$ band of TRAPPIST-1 clearly resembles that of a field object. The overall $K$ band shape of TRAPPIST-1, in Figure \[fig:Lbolbandbyband\]d, is similar to the field dwarfs and J0443$+$0002. TRAPPIST-1 shows absorption and CO depths matching the field dwarfs.
When compared to sources of similar $T_\mathrm{eff}$ or $L_\mathrm{bol}$, the band-by-band fits show under the assumption of field age TRAPPIST-1 exhibits a blend of field and low-gravity spectral features.
Spectral Analysis For Sample \#2: Assuming an Age of $<0.5$ Gyr for TRAPPIST-1 {#YoungAgeAssumption}
==============================================================================
Despite the above conclusion that the overall SED of TRAPPIST-1 is well fit as a field source, the SpeX SXD, prism, and FIRE spectra all show subtle signatures deviating from normal leading to an intermediate gravity classification. If we assume that TRAPPIST-1 is not a field-aged source due to its $\beta$ gravity classification, then we should treat it as we have other $\beta$ gravity sources and assume an age range of $<0.5$ Gyr. With this age assumption, the $T_\mathrm{eff}$ comparison sample we presented in Section \[FieldAgeAssumption\] would no longer be suitable. Here we present the new temperature samples of field, young and old age sources compared to TRAPPIST-1 assuming a younger age leading to a cooler temperature and larger radius.
Full SEDs
---------
Figures \[fig:YoungAgeSample\]a, b, and c compare TRAPPIST-1 to field-age, low-gravity, and subdwarfs of similar temperatures over the $0.5-13\,\upmu$m range. Compared to the field-age sources in Figure \[fig:YoungAgeSample\]a, TRAPPIST-1 is brighter than all sources across the entire range, having only a similar brightness from $0.63-0.72\,\upmu$m. In Figure \[fig:YoungAgeSample\]b, TRAPPIST-1 overlaps with the comparative sample from $\sim0.6-1\,\upmu$m, however beyond $1\,\upmu$m TRAPPIST-1 is fainter. This trend is similar to what is seen in Figure \[fig:TeffLbolSEDs\]b, showing no matter the assumed age TRAPPIST-1 does not resemble the very-low-gravity sources. TRAPPIST-1 overlaps with the comparative sample in Figure \[fig:YoungAgeSample\]c from $\sim0.6-1\,\upmu$m, however spectral features in this region are not well fit by the subdwarfs. Beyond $\sim1\,\upmu$m, TRAPPIST-1 is brighter than the comparative sample, however it displays a more triangular $H$ band similar to GJ 660.1B. Again, as seen with the low-gravity sources, no matter the assumed age range TRAPPIST-1’s overall SED is poorly fit by subdwarfs.
Band-by-Band Analysis {#bandbybandYoung}
---------------------
Assuming a younger age for TRAPPIST-1 we take a closer look at the comparative sample described in detail in Tables \[tab:Sample\] and \[tab:SampleFunParams\]. At the younger age, TRAPPIST-1 is 319 K cooler hence the sample changes from an assumed field age. All panels in Figure \[fig:YoungAgeBands\] show how TRAPPIST-1 fits to field and low-gravity equivalent sources in the $Y$, $J$, $H$, and $K$ bands.
In Figure \[fig:YoungAgeBands\]a TRAPPIST-1 is best fit by the field dwarf DENIS–P J1048.0$-$3956 (hereafter J1048$-$3956), from $0.95-0.99\,\upmu$m, while from $0.99-1.06\,\upmu$m TRAPPIST-1 is best fit by low-gravity dwarf 2MASS J20004841$-$7523070 (hereafter J2000$-$7523). The majority of the $Y$ band is poorly fit by both comparative field dwarfs and the low-gravity dwarf J0608$-$2753. While TRAPPIST-1 has regions that are matched well to J1048$-$3956 and J2000$-$7523, the field- and low-gravity dwarfs when assuming an older age more closely match TRAPPIST-1’s spectrum in the $Y$ band (see Section \[bandbybandField\]).
TRAPPIST-1 shares alkali depth and spectral shape features most similar to the field dwarfs in Figure \[fig:YoungAgeBands\]b, similar $H$-band shape as the field dwarfs in \[fig:YoungAgeBands\]c, and similar NaI depth to J0853$-$0356 and overall $K$-band shape to both field dwarfs in \[fig:YoungAgeBands\]d. These band-by-band similarities to the field dwarfs clearly anchor the classification of TRAPPIST-1 as a field aged source.
Sample \#3: Could TRAPPIST-1’s spectral features be low metallicity mimicing low gravity? {#subdwarfNaIKIcomps}
=========================================================================================
The low-metallicity d/sd GJ 660.1B discussed in [@Agan16] also showed signatures of youth similar to TRAPPIST-1 with its triangular $H$ band. Previous work by [@Kirk10] also saw this feature in spectra of late-M dwarfs with high proper motion and with no evidence of youth. [@Agan16] measured the [@Alle13] indices for GJ 660.1B assuming two different spectral types, M7 and M9.5. When typed as an M9.5, two of the indices received intermediate gravity scores, while when typed as M7 (the final decided upon spectral type) all scores were field gravity. Thus [@Agan16] (and references therein) state that additional opacity from the $1.55-1.6\, \upmu$m FeH absorption band and stronger H$_2$O, due to reduced condensate opacity of low-metallicity subdwarfs, may help shape the $H$-band continuum of mild subdwarfs and therefore potentially skew gravity index-based classifications.
Since the $H$-band shape of TRAPPIST-1 is similar to GJ 660.1B, we investigate if the intermediate gravity classification could be due to a low metallicity. However, there is no medium-resolution spectral data of GJ 660.1B or J1013$-$1356 and J1444$-$2019, two of the subdwarfs from the field and younger age assumption samples, therefore we instead compare the $J$-band and doublets of TRAPPIST-1 to subdwarfs with medium resolution spectra from [@Gonz18]. Figure \[fig:SubdwarfJband\] displays these sources in a decreasing effective temperature sequence. Both lines in the doublet of TRAPPIST-1 in Figure \[fig:SubdwarfJband\]a appear to be of similar depth, unlike the subdwarfs, and are deeper than the subdwarf doublets. In Figure \[fig:SubdwarfJband\]b the $1.17\,\upmu$m doublet of TRAPPIST-1 is narrower than the subdwarfs and has a depth in between that of the binary LSR J1610$-$0040 and SDSS J125637.13$-$022452.4 (hereafter J1256$-$0224), while the $1.25\,\upmu$m doublet is deeper and narrower than the subdwarfs. Thus the $J$-band alkali lines of TRAPPIST-1 are not similar to those of subdwarfs. In conclusion while low metallicity may mimic some low-gravity features, we find no evidence for that in the case of TRAPPIST-1.
Final Thoughts on TRAPPIST-1’s Age {#FinalThoughts}
==================================
From comparing the overall SEDs in the various samples, we see that TRAPPIST-1 most resembles those of the field dwarfs no matter what age we assume. When comparing the NIR band-by-band fits, TRAPPIST-1 is most similar to the field sources in some areas, while it has some similarities to aspects of the low-gravity sources. From our examination with the subdwarfs in the $J$ band, we see no evidence for the low-gravity spectral features to be caused by low-metallicity. Thus we conclude that the spectral features of TRAPPIST-1 are a blend of the field dwarfs of equivalent $T_\mathrm{eff}$ and $L_\mathrm{bol}$ and that of low-gravity sources of similar $L_\mathrm{bol}$. TRAPPIST-1 is likely a field age source with these spectral features originating from some cause other than youth.
Examining LHS 132 as an ideal comparative source {#LHS132}
================================================
When determining the age of TRAPPIST-1 [@Burg17] used the M7 dwarf 2MASS J23520507$-$1100435 (hereafter J2352$-$1100) and the M8 dwarfs LHS 132 and 2MASS 23412868$-$1133356 (hereafter J2341$-$1133) to constrain the age based on surface gravity features. J2352$-$1100 is a member of the $130-200$ Myr old [@Bell15] AB Doradus moving group (Faherty et al. in prep.), while both LHS 132 and J2341$-$1133 are field age sources [@Fili15; @Cruz07]. We have created full SEDs of these sources to compare to TRAPPIST-1 in more detail than done in [@Burg17].
In Figure \[fig:RebuttalSED\] we see that LHS 132 fits the overall SED shape of TRAPPIST-1 very well, while J2352$-$1100 and J2341$-$1133 both have significantly more flux over the entire region. LHS 132 has a similar $T_\mathrm{eff}$ and $L_\mathrm{bol}$ as TRAPPIST-1, while J2352$-$1100 and J2341$-$1133 are much hotter, brighter, and have larger radii. Looking at the SEDs alone, J2352$-$1100 and J2341$-$1133 are both poor comparison sources to TRAPPIST-1 since these objects are fundamentally different, while LHS 132 is still a fair comparative source. LHS 132 also receives an intermediate gravity classification which is discussed further in Section \[AL13Indicescomparison\]. Therefore, LHS 132 is a target to explore in more detail to see if other aspects match those of TRAPPIST-1.
Discussion {#Discussion}
==========
Comparison of the [@Alle13] Gravity Indices {#AL13Indicescomparison}
-------------------------------------------
Because of the intermediate gravity (INT-G or $\beta$) classification that [@Burg17] found for TRAPPIST-1 when using the [@Alle13] gravity-sensitive indices, we calculated the indices for our entire comparative sample using our modified version of the `ALLERS13_INDEX` IDL code on both low- and medium-resolution spectra when available. The index values and final gravity scores are listed in Tables \[tab:gravityindiceslow\] and \[tab:gravityindicesmed\]. The equivalent width measurements for the medium resolution gravity score are listed in Table \[tab:eqw\]. All objects receive the same gravity class using both the low- and medium-resolution indices, with the exception of vB 10 which receives a FLD-G classification with low-resolution, but a $\beta$ classification with medium resolution due to the FeH$_J$ score. We note that while our FeH$_J$ score is a 2, the FeH$_J$ feature does not appear to look different from the other field sources in our sample. The FeH$_J$ score from [@Martin17] was a 0, therefore we may be getting a spurious measurement.
The best fit source from the [@Burg17] sample, LHS 132, also receives an INT-G gravity classification. LHS 132 and TRAPPIST-1 have similar radii, mass, log $g$, $T_\mathrm{eff}$, and $L_\mathrm{bol}$ as well as the same scores for each of the gravity indices. Therefore, whatever physical factor is causing TRAPPIST-1 to receive an $\beta$ classification may also be the same for LHS 132.
All subdwarfs receive a $\beta$ gravity class, with the exception of the spectroscopic binary J1610$-$0400, which received a FLD-G classification, and LHS377 which did not receive a gravity class due to a lack of measurement for the FEH$_z$ index. Therefore as stated in [@Alle13ConfP; @Agan16; @Burg17; @Martin17] there is some aspect of the spectrum that fools the indices into classifying objects with older ages as $\beta$, which could be due to metallicity. However, as stated in [@Burg17] and from our comparison to subdwarfs, TRAPPIST-1 is not low metallicity nor does it show $J$-band features similar to subdwarfs and thus this is unlikely to be the cause of the $\beta$ classification for TRAPPIST-1.
Figures \[fig:Indices\]a–c show the low-resolution spectrum index scores for the comparative sample and TRAPPIST-1 with the field dwarf polynomial and the dividing line between a score of 1 or 2 from [@Alle13]. A score of 1 indicates intermediate gravity, while a score of 2 indicates low-gravity for that index. Looking at the low-resolution gravity scores, TRAPPIST-1 received scores of 1, indicating intermediate gravity, in all indices using the prism and SXD data. The FIRE data however, received a score of 0 in the $H$-cont index, indicating field gravity, and a score of 1 in the other indices. All FeH$_z$ and $_J$ index measurements (see Figure \[fig:Indices\]a and b) for TRAPPIST-1, lie in the intermediate gravity region as defined by [@Alle13]. For the $H$-cont index (Figure \[fig:Indices\]c) we see that the FIRE measurement lies in the field dwarf region, while both SpeX measurements lie in the intermediate gravity region.
Medium resolution gravity scores of TRAPPIST-1, reveal a difference in the line scores between the SXD and FIRE spectra. The 1.169, 1.17, and 1.253 $\upmu$m equivalent widths are plotted for all sources in our sample with medium resolution in Figure \[fig:EquivalentWidths\] along with the equivalent widths of all sources in the [@Martin17] sample. We do not show the 1.224 $\upmu$m equivalent width plot, since as shown by [@Alle13] and [@Martin17] there is no visible trend. All three lines for the SXD spectrum of TRAPPIST-1 received a score of 0, while the FIRE spectrum 1.169 $\upmu$m and 1.253 $\upmu$m lines received scores of 1 and the 1.177 $\upmu$m line received score of 0. The 1.169 $\upmu$m and 1.253 $\upmu$m equivalent width measurements for the FIRE spectrum lies just below the field sources from [@Martin17] and not far from the corresponding SXD measurement. As supported by our band-by-band analysis in Sections \[bandbybandField\] and\[bandbybandYoung\], TRAPPIST-1 appears to have low surface gravity features despite our conclusion that its overall SED is best fit by a field age.
Comparison to Trends with Spectral Type
----------------------------------------
To further examine the signatures of youth seen in the NIR spectrum of TRAPPIST-1 we place it in context with fundamnetal parameters of field sources from [@Fili15], low-gravity sources from [@Fahe16], and subdwarfs from [@Gonz18]. Figure \[fig:SptvLbol\] shows the comparison of $L_\mathrm{bol}$ versus spectral type, Figure \[fig:SptvTeff\] shows the comparison of $T_\mathrm{eff}$ versus spectral type, and Figures \[fig:AbsMagsJHK\] and \[fig:AbsMagsW1W2\] compare absolute magnitudes versus spectral type for the $J$,$H$,$K$,$W1$,and $W2$ bands.
As shown in [@Gonz18], all sources are mixed when comparing $L_\mathrm{bol}$ versus optical spectral type in Figure \[fig:SptvLbol\]. TRAPPIST-1 lands in an area where there are other field sources, however there is no visible trend for where M7.5 $\beta$ sources should be located on the diagram due to only one M7.5 $\beta$ source other than TRAPPIST-1 plotted. [@Fahe16] found that $\beta$ gravity sources in their sample that were not members of known moving groups fell along the field sequence, and thus not all late M dwarf $\beta$ sources are young.
Figure \[fig:SptvTeff\] compares $T_\mathrm{eff}$ versus optical spectral type with TRAPPIST-1 again landing in an ambiguous location. TRAPPIST-1 lies near the field dwarfs, but is surrounded by low-gravity sources as well. As seen in [@Fahe16] the low-gravity and field dwarf polynomials overlap in the M-dwarf region, right where TRAPPIST-1 is located.
Figures \[fig:AbsMagsJHK\] and \[fig:AbsMagsW1W2\] compare the absolute magnitudes in the $J$,$H$,$K$,$W1$, and $W2$ bands of field, low-gravity and subdwarfs. The M-dwarf $\beta$ gravity sources lie in the same location as the field sources or just slightly above in the $J$ band and begin to move slightly higher than the field sequence by $K$ band. By $W1$ and $W2$ the $\beta$ sources are further above the field sequence. TRAPPIST-1 remains within the field sequence from $J$ through $W2$ and does not appear to move in brightness like the other $\beta$ sources. This further supports the idea that TRAPPIST-1 is a field source requiring a different physical explanation for its low surface gravity features.
Comparison of kinematics to other $\beta$ gravity sources
---------------------------------------------------------
### UVW
[l c c c c c c c c ]{} TRAPPIST-1 & M7.5 & $\cdots$ & $-44 \pm 0.1$ & $-67.2 \pm 0.3$ & $11.7 \pm 0.4$ & $61.69 \pm 0.10$ & SpT: 1, Rest: 2\
Teegarden’s Star & M6.5 & M7.5$\beta$ & $-69.46 \pm 0.31$ & $-71.17 \pm 0.15$ & $-58.68 \pm 0.25$ & $93.03 \pm 0.10$ & SpT: 3, 4, UVW: 5, Vel: 2\
LHS 132 & M8 & M8 & $-22.4 \pm 1.2$ & $37 \pm 1.2$ & $-68.08 \pm 0.95$ & $80.4 \pm 0.14$ & SpT: 6, 7, Rest: 2\
2MASS J10220489$+$0200477 & M9$\beta$ & M9 &$14.87 \pm 4.49$ & $-53.28 \pm 19.96$ & $-49.14 \pm 14.81$ & $65.06 \pm 1.35$ & SpT:8, 9, UVW: 10, Vel: 11\
2MASS J10224821$+$5825453 & L1$\beta$ & L1 &$-69.35 \pm 2.74$ & $-67.62 \pm 3.48$ & $0.1 \pm 0.87$ & $95.61 \pm 0.57$ & SpT: 12, 8, UVW: 10, Vel: 11\
2MASS J23224684$-$3133231 & L0$\beta$ & L2$\beta$ & $40.26 \pm 2.74$ & $-30.87 \pm 3.18$ & $-24.72 \pm 1.27$ & $54.39 \pm 0.76$ & SpT: 13, 9, UVW:10 , Vel: 11\
2MASS J00332386$-$1521309 & L4$\beta$ & L1 & $-52.85 \pm 5.68$ & $-26.91 \pm 3.93$ & $3.23 \pm 0.86$ & $34.1\pm1.5$ & SpT: 12, 9, UVW:10 , Vel: 11\
To determine if $\beta$ sources are kinematically distinct, we examined $UVW$ velocities of all $\beta$ sources from [@Fahe16]. The $UVW$ velocities are displayed in a Toomre diagram (Figure \[fig:Toomre\]) along with $\gamma$ sources from [@Fahe16], Teegarden’s Star, and LHS 132. For objects in this work labeled as “not in group” or “non-members”, these are sources from [@Fahe16] that could be ambiguous members, candidate members, or true non-members of any currently known group, thus objects that are not bonafide members.
In Figure \[fig:Toomre\] we see all $\gamma$ sources, whether in known moving groups or not, clustered in parameter space that corresponds to the thin disk, while the $\beta$ sources are found across the thin and well into the thick disk region. Thus the $\beta$ sources with total velocities greater than 50 km s$^{-1}$ may not truly be young, but display signatures of youth for some unaccounted reason. TRAPPIST-1, LHS 132, and Teegarden’s Star lie in the thick disk region, along with two intermediate gravity sources from [@Fahe16]- 2MASS J10220489$+$0200477 (hereafter J1022$+$0200) and 2MASS J10224821$+$5825453 (hereafter J1022$+$5825). The two intermediate gravity sources in the thin/thick disk region are 2MASS J23224684$-$3133231 (hereafter J2322$-$3133) and 2MASS J003323.86-1521309 (hereafter J0033$-$1521). Kinematics of these seven sources are listed in Table \[tab:compkinematics\]. Since Teegarden’s star also lies in the same region as TRAPPIST-1 and has recently been found to host at least two planets [@Zech19], we suggest one idea for the low-gravity features may be the tug of planets on their host star. Consequently, LHS 132, J1022$+$0200, J1022$+$5825 may be ideal targets for M dwarf planet searches.
### Tangential Velocity
Figure \[fig:Vtan\] shows the distribution of tangential velocities for low-gravity sources from [@Fahe16], with their updated membership and $V_\mathrm{tan}$ values from Faherty et al. (in prep). In this sample, any source classified as $\beta$ gravity in either optical or near-infrared receives a classification of $\beta$ (i.e. M8 in optical, but M8 $\beta$ in NIR, is designated as a $\beta$ in this sample). The same follows for a source with a $\gamma$ classification. For sources that received a $\beta$ classification in one regime, but a $\gamma$ in another, we choose the more extreme gravity classification.
Figure \[fig:Vtan\]a shows the distribution of tangential velocities for $\gamma$ gravity sources which are members (on the right) or non-members (on the left) of known moving groups. We see that members and non-members have similar distributions of $V_\mathrm{tan}$, both peaking in the range. The non-member distribution is not significantly different from the member distribution.
Figure \[fig:Vtan\]b shows the distribution for $\beta$ gravity sources that are members and non-members of known moving groups. Unlike the $\gamma$ gravity sources we see that $\beta$ sources in moving groups have $V_\mathrm{tan}$ ranging from , whereas non-member sources have a larger range of $V_\mathrm{tan}$. However, the bulk of $\beta$ gravity non-members fall in the range seen for member sources. Our calculated $V_\mathrm{tan}$ for TRAPPIST-1 places it outside of the bulk velocity region for $\beta$ gravity sources. We also calculated $V_\mathrm{tan}$ for LHS 132, which places it in the same region as TRAPPIST-1 and the four sources plotted which are Teegarden’s Star, J1022$+$0200, J1022$+$5825, and J2322$-$3133. However, J0033$-$1521 lies in the bulk region. Therefore, TRAPPIST-1 along with Teegarden’s Star, J1022$+$0200, J1022$+$5825, J2322$-$3133, and LHS 132 are kinematically distinct from the other suspected young M dwarfs.
Speculation on the $\beta$ gravity class for TRAPPIST-1
-------------------------------------------------------
The $\beta$ classification of TRAPPIST-1 could be due to radius inflation. Two possible causes of this could be: (1) magnetic activity and/or (2) tidal interactions of the planets with the star. In the case of the former, [@Chab07] state that theoretical models for sources with M$=0.08$ M$_\odot$, show that radii can range from $0.10-0.14$ R$_\odot$ for black spot coverage of up to 50% thus offering one pathway towards a radius variation that would mimic a young M dwarf that had not contracted yet. [@Luge17] and [@Vida17] examined the $K$2 light curve and found evidence of cool, stable magnetic spots on TRAPPIST-1. However, most recently [@Morr18] examined the Spitzer 3.5 $\upmu$m and 4.5 $\upmu$m light curves and found no signature of cool spots leaving this line of evidence inconclusive for magnetic influence. For the second possible cause, tidal interactions of the planets with the star, could be the cause of the classification for both TRAPPIST-1 and Teegarden’s Star. While it is beyond the scope of this work to examine the full influence of the planets on their host star, we suggest that an excellent test for this theory would be to look for planets around LHS 132, the other source that matches many of TRAPPIST-1’s features, as well as J1022$+$0200 and J1022$+$5825 which are kinemtically distinct from other $\beta$ sources.
Conclusions
===========
In this work we present a distance-calibrated SED of TRAPPIST-1 using a new NIR FIRE spectrum and a new parallax from the *Gaia* DR2 data release. With our new distance-calibrated SED we compare TRAPPIST-1 to objects of similar effective temperature and/or bolometric luminosity for young, old, and field-aged sources considering an age for TRAPPIST-1 of either $0.5-10$ Gyr or $<0.5$ Gyr. The $J$-band and lines of TRAPPIST-1 were compared to those of the subdwarfs with medium-resolution data from [@Gonz18]. We also looked at TRAPPIST-1 related to objects from [@Burg17]. We present updated or new fundamental parameters for our comparative sample using Gaia parallaxes and Pan-Starrs photometry when available.
Using our derived fundamental parameters we find field dwarfs of similar $T_\mathrm{eff}$ and $L_\mathrm{bol}$, and LHS 132 a M8 dwarf classified as intermediate gravity best fit the SED shape of TRAPPIST-1. From our band-by-band comparisons, TRAPPIST-1 exhibits a blend of field and young spectral features.
We measure the [@Alle13] indices for TRAPPIST-1, along with our entire comparative sample. TRAPPIST-1 receives a $\beta$ gravity classification when using three different spectra indicating it might be young. Examining spectral indices versus spectral type, TRAPPIST-1 lies in the $\beta$ gravity space, while when looking at equivalent width versus spectral type, TRAPPIST-1 falls with the field sources.
In an effort to better understand the $\beta$ gravity population and TRAPPIST-1, we plot $L_\mathrm{bol}$ and $T_\mathrm{eff}$ versus optical spectral type as well as $J$, $H$, $K$, $W1$, and $W2$ absolute magnitudes versus spectral type. We find TRAPPIST-1 lies in an area that has both field and $\beta$ sources when examining $L_\mathrm{bol}$ and $T_\mathrm{eff}$ versus optical spectral type and absolute magnitudes versus optical spectral type.
We present updated $UVW$ velocities for TRAPPIST-1 using the new *Gaia* astrometry and compare its kinematics to $\beta$ and $\gamma$ sources which are confirmed members or not bonafide members of known moving groups, Teegarden’s Star, and LHS 132. TRAPPIST-1 along with Teegarden’s Star, LHS 132, J1022$+$0200, J1022$+$5825, and J2322$-$3133 fall within a subpopulation of $\beta$-gravity sources that are not bonafide members of known moving groups and have higher $UVW$ and tangential velocities.
Lastly, we present two possible causes for the $\beta$ classification of TRAPPIST-1. First, TRAPPIST-1 may have significant magnetic influence as observationally examined by observing cool stable spots. At present there are contradictory information in the literature on this topic. [@Luge17] and [@Vida17] found evidence for spots while [@Morr18] did not leaving this line of explanation inconclusive. Our second proposed explanation could be related to tidal interactions with planets and we suggest LHS 132, J1022$+$0200, J1022$+$5825, and J2322$-$3133 might be an excellent targets for exoplanet campaigns given their similarities to both TRAPPIST-1 and Teegarden’s Star.
We thank the Magellan telescope operators for their help in collecting FIRE spectra. We thank D. Bardalez Gagliuffi and C. Galindo for their help in obtaining the SpeX spectra of J0608$-$2753. This research was supported by the NSF under Grant No. AST-1614527 and Grant No. AST-1313278, as well as by NASA under *Kepler* Grant No. 80NSSC19K0106. E.G. thanks the LSSTC Data Science Fellowship Program, which is funded by LSSTC, NSF Cybertraining Grant No. 1829740, the Brinson Foundation, and the Moore Foundation; her participation in the program has benefited this work. J.T. acknowledges support for this work provided by NASA through Hubble Fellowship grant HST-HF2-51399.001 awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS5-26555. This research has made use of the BDNYC Data Archive, an open access repository of M, L, T and Y dwarf astrometry, photometry and spectra. This paper includes data gathered with the 6.5 meter Magellan Telescopes located at Las Campanas Observatory, Chile. This publication makes use of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration. This work has made use of data from the European Space Agency (ESA) mission [*Gaia*]{} (<https://www.cosmos.esa.int/gaia>), processed by the [*Gaia*]{} Data Processing and Analysis Consortium (DPAC, <https://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the [*Gaia*]{} Multilateral Agreement.
Photometry and Spectra Tables for SEDs
======================================
[l c c c c c c c c c c c c c c c]{} TRAPPIST-1 & $19.35\pm0.02$ & $17.87\pm0.01 $ & $15.13\pm0.01$ & $13.73\pm0.01$ & $12.97\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &$\cdots$ & 1\
J0320$+$1854 & $19.8\pm0.03$ & $18.2\pm0.01$ & $15.6\pm0.01 $ & $14.24\pm0.01$ &$13.43\pm0.01$ &$20.411\pm 0.093$ &$20.021\pm 0.016$ & $17.996\pm 0.007$ & $15.303 \pm 0.004$ & $13.448 \pm 0.004$ & 1, 2\
J0443$+$0002 & $21.51\pm0.07$ & $19.66\pm0.02$ & $16.96\pm0.01 $& $15.38\pm0.01 $& $14.4\pm0.01$ & $\cdots$ & $21.99\pm 0.07$ & $19.76\pm0.02$ & $17.0048\pm 0.0055$ & $15.0886 \pm 0.0053$ & 1, 2\
J0518$-$2756 & $\cdots$ & $\cdots$ & $19.99\pm0.02$ & $18.49\pm0.01$ & $17.5\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J0532+8246 & $\cdots$ & $\cdots$ & $\cdots$ &$18.07\pm0.01$ & $16.92\pm0.03 $& $\cdots$ & $\cdots$ & $22.59\pm0.26$ & $20.366\pm0.054 $ & $ 17.581\pm0.02$ & 1, 3\
J0608$-$2753 & $\cdots$ & $20.94\pm0.02$ & $18.14\pm0.01$ & $16.55\pm0.01$ & $15.54\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J0853$-$0329 & $19.8\pm0.04$ & $17.85\pm0.01$ & $15.5\pm0.01$ & $14.06\pm0.01$ & $13.1\pm0.01 $ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J1013$-$1356 & $20.99\pm0.07$ & $19.25\pm0.01$ & $17.18\pm0.01$ & $ 16.25\pm0.01$ & $15.9\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$& $\cdots$ & 1\
J1256$-$0224 & $\cdots$ & $\cdots$ & $19.47\pm0.04$ & $18.0\pm0.03$ & $17.54\pm0.01$& $\cdots$ & $\cdots$ & $\cdots$ &$19.41 \pm 0.02$ & $17.71 \pm 0.02$ & 1, 4\
LHS 377 & $18.99\pm0.02$ & $17.69\pm0.01$ &$15.68\pm0.01$ &$14.84\pm0.01$ &$ 14.48\pm0.01$ &$21.89 \pm 0.13$ & $19.384 \pm 0.011$ & $17.681 \pm 0.006$ & $15.696 \pm 0.005$ & $14.707 \pm 0.006$ & 1, 2\
J1444$-$2019 & $\cdots$ & $19.31\pm0.03$ & $16.09\pm0.05$ & $\cdots$ & $14.06\pm0.02$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 5\
LHS 3003 & $17.65\pm0.01$ & $16.21\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 5\
J1610$-$0040 & $20.12\pm0.1$ & $18.08\pm0.01$ & $15.9\pm0.01$ & $14.86\pm.01$ & $ 14.41\pm0.02$ & $\cdots$ & $\cdots$ & $17.976 \pm 0.007$ & $15.903 \pm 0.004$ & $14.663 \pm 0.005$ & 1,6\
vB 8 & $17.43\pm0.01$ & $16.02\pm0.01$ & $\cdots$ & $\cdots$& $\cdots$ & $19.201 \pm 0.035$ &$17.738 \pm 0.006 $ & $16.046 \pm 0.003$ & $13.221 \pm 0.002$ & $11.766 \pm 0.003 $ & 1, 2\
GJ 660.1B & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $14.4\pm0.18$& $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 5\
J1835$+$3295 & $19.02\pm0.01$ & $17.18\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
vB 10 & $\cdots$ & $16.59\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J2036+5059 & $19.61\pm0.02$ & $18.14\pm0.01$ & $16.13\pm0.01$ & $15.24\pm0.01$ & $14.89\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J2341$-$1133 & $21.1\pm0.03$ & $19.84\pm0.02$ & $17.13\pm0.01$ & $15.86\pm0.01$ & $15.13\pm0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J2352$-$1100 & $20.57\pm0.03$& $19.18\pm0.01$ & $16.49\pm0.01$ & $15.23\pm0.01$ & $14.49\pm0.01$ & $\cdots$ & $20.930\pm 0.042$ & $19.250 \pm 0.015$ & $16.4830 \pm 0.0044$ &$14.9416 \pm 0.0047$ & 1,2\
[l c c c c c c c]{}\[h\] TRAPPIST-1 & $18.998 \pm 0.048$ & $14.1 \pm 0.01$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
LHS 132 & $\cdots$ & $\cdots$ & $ 18.53 \pm 0.021$ & $16.3 \pm 0.008$ & $13.88 \pm 0.012$ & $13.83 \pm 0.03$ & 2, 3\
J0853$-$0329 & $\cdots$ & $\cdots$ & $18.94\pm0.032$ & $\cdots$ & $\cdots$ & $\cdots$ & 2\
J1048$-$3956 & $\cdots$ & $\cdots$ & $17.532\pm0.057$ & $15.051\pm0.014$ & $\cdots$ & $\cdots$ & 4\
J1247$-$3816 & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $17.85 \pm 0.16$ & 3\
LHS 3003 & $\cdots$ & $\cdots$ & $16.95 \pm 0.014$ & $14.9 \pm 0.006$ & $12.53\pm0.008$ & $\cdots$ & 2\
vB 8 & $\cdots$ & $\cdots$ & $16.85 \pm 0.059$ & $\cdots$ & $12.25\pm0.015$ & $\cdots$& 2\
J2000$-$7523 & $\cdots$ & $\cdots$ & $21.157\pm0.008$ & $18.379\pm0.001$ & $16.119\pm0.024$ & $\cdots$ & 5\
[l c c c c c c c c c c c c]{} TRAPPIST-1 & $11.354 \pm 0.022$ & $10.718 \pm 0.021$ & $10.296 \pm 0.023$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
LHS 132 & $11.13 \pm 0.023$ & $10.479 \pm 0.024$ & $10.069 \pm 0.021 $ & $11.165 \pm 0.07 $ & $10.03 \pm 0.07$ & $\cdots$ & $\cdots$ & $\cdots$ &1, 2\
J0320$+$1854 & $11.759 \pm 0.021$ & $11.066 \pm 0.022$ & $10.639 \pm 0.018$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J0436$-$4114 & $13.097 \pm 0.026$ & $12.43 \pm 0.022$ & $12.05 \pm 0.024$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J0443$+$0002 & $ 12.507 \pm 0.026 $ & $11.804 \pm 0.024$ & $11.216 \pm 0.021$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J0518$-$2756 & $15.262 \pm 0.043$ & $ 14.295 \pm 0.046$ & $ 13.615 \pm 0.04$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J0532+8246 & $15.179 \pm 0.058$ & $14.904 \pm 0.091$ & $14.92 \pm 0.15$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J0608$-$2753 & $13.595 \pm 0.028$ & $12.897 \pm 0.026$ & $12.37 \pm 0.024$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J0853$-$0329 & $11.212\pm0.026$ & $10.469\pm0.026$ & $9.942\pm0.024$ & $\cdots$ & $\cdots$ & $11.18\pm0.05$ & $10.48\pm0.05$ & $9.91\pm0.05$ & 1, 3\
J1013$-$1356 & $14.621 \pm 0.032$ & $14.382 \pm 0.049$ & $14.398 \pm 0.078$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J1048$-$3956 & $9.538\pm0.022$ & $8.905\pm0.044$ & $8.447\pm0.023$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J1207$-$3900 & $15.494 \pm 0.058$ & $14.608 \pm 0.04$ & $ 14.04 \pm 0.059$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J1247$-$3816 & $14.785 \pm 0.031$ & $14.096 \pm 0.035$ & $13.573 \pm 0.038$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J1256$-$0224 & $16.10 \pm 0.11$ & $15.79\pm0.15$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $16.078 \pm 0.016$ & $16.605 \pm 0.099$ & 1, 4, 5\
LHS 377 & $13.194 \pm 0.029$ & $12.73 \pm 0.03$ & $12.479 \pm 0.025$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J1444$-$2019 & $12.546 \pm 0.026$ & $12.142 \pm 0.026$ & $11.933 \pm 0.026$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
LHS 3003 & $9.965 \pm 0.026$ & $9.315 \pm 0.022$& $8.928 \pm 0.027$ & $\cdots$ & $\cdots$ & $ 9.94 \pm 0.05$ & $9.43 \pm 0.05$ & $8.93 \pm 0.05$ & 1,6\
J1610$-$0040 & $12.911 \pm 0.022$ & $12.302 \pm 0.022$ & $12.302 \pm 0.022$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
vB 8 & $9.776 \pm 0.029$ & $9.201 \pm 0.024$ & $8.816 \pm 0.023$ & $9.737 \pm 0.04$ & $8.819 \pm 0.06$ & $\cdots$ & $\cdots$ & $\cdots$ &1, 2\
GJ 660.1B & $ 13.052 \pm 0.045$ & $12.565 \pm 0.023$ & $ 12.227 \pm 0.027$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J1835$+$3295 & $10.27\pm0.022$ & $9.617\pm0.021$ & $9.171\pm0.018$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
vB 10 & $9.908 \pm 0.025$ & $9.226 \pm 0.026$ & $8.765 \pm 0.022$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J2000$-$7523 & $15.07 \pm 0.048$ &$14.003 \pm 0.036$& $13.42 \pm 0.042$ &$\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J2036+5059 & $13.611 \pm 0.029$ & $13.160 \pm 0.036$ & $12.936 \pm 0.033$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J2341$-$1133 & $13.546 \pm 0.023$ & $12.939 \pm 0.03$ & $12.546 \pm 0.033$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J2352$-$1100 & $12.84 \pm 0.022$ & $12.166\pm0.021$ & $11.742 \pm 0.02$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
[l c c c c c c c c c c c]{} TRAPPIST-1 &$\cdots$ & $\cdots$ & $10.042 \pm 0.023$ & $9.80 \pm 0.02$ & $9.528 \pm 0.041$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
LHS 132 &$\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $9.64 \pm 0.02 $ & $9.62 \pm 0.02$ & $9.52 \pm 0.02$ & $9.48 \pm 0.01$ & 2\
J0320$+$1854 & $\cdots$ & $\cdots$ & $10.347 \pm 0.023$ & $10.148 \pm 0.02 $ & $9.874 \pm 0.048$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J0436$-$4114 & $\cdots$ & $\cdots$ & $11.74 \pm 0.023$ & $11.46 \pm 0.021$ & $11.111 \pm 0.082$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ &1\
J0443$+$0002 & $\cdots$ & $\cdots$ & $10.826 \pm 0.024$ & $ 10.476 \pm 0.021$ & $ 10.031 \pm 0.054$ & $\cdots$ & $ 10.55 \pm 0.02$ & $ 10.45 \pm 0.02$ & $ 10.35 \pm 0.03 $ & $ 10.22 \pm 0.03$ & 1, 3\
J0518$-$2756 & $\cdots$ & $\cdots$ & $ 13.045 \pm 0.024$ & $12.661 \pm 0.026$& $12.581 \pm 0.349$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J0532$+$8246 & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $13.37 \pm 0.03$ & $13.22 \pm 0.02$ & $13.23 \pm 0.1 $ & $13.03 \pm 0.1$ & 1\
J0608$-$2753 & $\cdots$ & $\cdots$ & $11.976 \pm 0.024$ & $11.623 \pm 0.021$ & $11.31 \pm 0.11$ & $\cdots$ & $11.75 \pm 0.02$ & $11.62 \pm 0.02$ & $11.52 \pm 0.03$ & $11.44 \pm 0.03$ & 1, 3\
J0853$-$0329 & $9.39\pm0.07$ & $9.62\pm0.1$ & $9.624\pm0.023$ & $9.381\pm0.02$ & $8.967\pm0.028$ & $8.756\pm0.428$ & $9.41\pm0.02$ & $9.39\pm0.03 $ & $9.22\pm0.01$ & $9.13\pm0.01$ & 2, 4\
J1013$-$1356 & $\cdots$ & $\cdots$ & $13.782 \pm 0.028$ & $13.545 \pm 0.035$ & $12.68 \pm 0.51$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J1048$-$3956 & $\cdots$ & $\cdots$ & $ 8.103\pm0.024$ & $7.814\pm0.021$ & $7.462\pm0.018$ & $7.226\pm0.087$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 5\
J1207$-$3900 & $\cdots$ & $\cdots$ & $13.64 \pm 0.024$ & $13.204 \pm 0.027$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 5\
J1247$-$3816 & $\cdots$ & $\cdots$ & $13.119 \pm 0.024$ & $12.532 \pm 0.024$ & $10.953 \pm 0.077$ & $8.84 \pm 0.29$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1,5\
J1256$-$0224 & $\cdots$ & $\cdots$ & $15.214 \pm 0.038$ & $15.106 \pm 0.098$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
LHS 377 & $\cdots$ & $\cdots$ & $12.298 \pm 0.027$ & $12.051 \pm 0.025$ & $11.67 \pm 0.11$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
J1444$-$2019 & $\cdots$ & $\cdots$ & $11.464 \pm 0.024$ & $11.211 \pm 0.022$ & $10.967 \pm 0.09$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 5\
LHS 3003 & $8.43 \pm 0.03$ &$\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $8.47\pm0.02$ & $8.49\pm0.01$ & $8.39\pm0.02$ & $8.36\pm0.01$ & 2,6\
J1610$-$0040 & $\cdots$ & $\cdots$ & $11.639 \pm 0.025$ & $11.639 \pm 0.025$ & $11.639 \pm 0.025$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 1\
vB 8 & $\cdots$ & $\cdots$ & $8.588 \pm 0.023$ & $8.365 \pm 0.021$ & $8.132 \pm 0.022$ & $7.857 \pm 0.181$ & $8.37 \pm 0.02$ & $8.38 \pm 0.01$ & $8.28 \pm 0.02$ & $ 8.24 \pm 0.02$ & 1, 2\
GJ 660.1B & $\cdots$ & $\cdots$ & $11.689\pm0.24$ & $11.496\pm0.318$ &$\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 7\
J1835$+$3295 & $\cdots$ & $\cdots$ &$8.803\pm0.022$ & $8.539\pm0.019$ & $8.16\pm0.019$ & $7.886\pm0.132$ & $8.55\pm0.02$ &$8.55\pm0.01$ &$8.39\pm0.01$ & $8.29\pm0.01$ & 1, 2\
vB 10 & $\cdots$ & $\cdots$ & $8.465 \pm 0.023$ & $8.249 \pm 0.02$ & $8.08 \pm 0.022$ & $\cdots$ & $8.29 \pm 0.02$ & $8.3 \pm 0.03$ & $8.15 \pm 0.01$ & $8.14 \pm 0.01$ & 1, 2\
J2000$-$7523 & $\cdots$ & $\cdots$ &$12.819 \pm 0.024$ &$12.431 \pm 0.024$ &$11.64 \pm 0.15$ &$\cdots$ &$\cdots$ &$\cdots$ &$\cdots$ & $\cdots$& 1\
J2036$+$5059 & $\cdots$ & $\cdots$& $12.667\pm0.24$ & $12.436\pm0.318$ &$\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 7\
J2341$-$1133 & $\cdots$ & $\cdots$ & $12.224 \pm 0.025$ & $12.001 \pm 0.024$ & $11.721 \pm 0.218$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 5\
J2352$-$1100 & $\cdots$ & $\cdots$ & $11.44 \pm 0.025$ & $11.146 \pm 0.022$ & $10.849 \pm 0.109$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & 5\
[l c c c c c c c c c]{} TRAPPIST-1 & KPNO 4m: R–C Spec & 2003–07–10 & 1 & FIRE & 2017–07–28 & 2 & $\cdots$ & $\cdots$ & $\cdots$\
LHS 132 & CTIO 1.5m: R–C Spec & 2003–11–09 & 1 & SpeX Prism & 2008–09–07 & 3 & $\cdots$ & $\cdots$ & $\cdots$\
J0320$+$1854 & GoldCam & 2000–10–01 & 4 & SpeX SXD, LXD1.9 & 2009–12–01 & 5 & $\cdots$ & $\cdots$ & $\cdots$\
J0436$-$4114 & LRIS & 2009–10–11 & 6 & SpeX SXD & 2012–09–20 & 7 & $\cdots$ & $\cdots$ & $\cdots$\
J0443$+$0002 & CTIO 1.5m: R–C Spec & 2002–01–28 & 4 & SpeX SXD & 2012–02–05 & 8 & $\cdots$ & $\cdots$ & $\cdots$\
J0518$-$2756 & LRIS & 2009–02–17 & 9 & FIRE & 2015–12–20 & 11 & $\cdots$ & $\cdots$ & $\cdots$\
J0532+8246 & LRIS & 2003–01–03 & 11 & NIRSPEC & 2002–12–24 & 11 & IRS & 2005–03–23& 12\
J0608$-$2753 & CTIO 4m: R–C Spec & 2002–01–25 & 4 & SpeX SXD & 2007–11–13 & 13 & $\cdots$ & $\cdots$ & $\cdots$\
J0853$-$0329 & CTIO 4m: R–C Spec & 2003–04–21& 14 & SpeX SXD & 2009–12–01 & 5 & $\cdots$ & $\cdots$ & $\cdots$\
J1013$-$1356 & GMOS-N & 2004–11–21 & 15 & SpeX Prism & 2004–03–12& 16 & IRS & 2005–06–07& 17\
J1048$-$3956 & CTIO 1.5m: R–C Spec& 2003–05–15 & 14 & SpeX SXD & 2009–12–01 & 5 & $\cdots$ & $\cdots$ & $\cdots$\
J1207$-$3900 & MagE & 2013–05–14 & 18 & FIRE & 2015–12–22 & 10 & $\cdots$ & $\cdots$ & $\cdots$\
J1247$-$3816 & $\cdots$ & $\cdots$ & $\cdots$ & SpeX Prism & 2013–05–10 & 18 &$\cdots$ & $\cdots$ & $\cdots$\
J1256$-$0224 & LDSS3 & 2006–05–07 & 19 & FIRE & 2016–08–13 & 20 & $\cdots$ & $\cdots$ & $\cdots$\
LHS 377 & X-Shooter (UVB,VIS) & 2014–02–20 & 21 & X-Shooter & 2014–02–20 & 21 & IRS & 2005–07–01 & 12\
J1444$-$2019 & $\cdots$ & $\cdots$ & $\cdots$ & SpeX Prism & 2005–03–23 & 3 & $\cdots$ & $\cdots$ & $\cdots$\
LHS 3003 & GoldCam & 2003–03–13 & 14 & SpeX Prism & 2008–07–29 & 3 & IRS & 2006–01–14 & 22\
J1610$-$0040 & MkIII & 2003-02-19 & 23 & SpeX Prism & 2003–07–06 & 24 & $\cdots$ & $\cdots$ & $\cdots$\
vB 8 & KPNO 4m: R–C Spec & 2002–09–25 & 14 & SpeX SXD, LXD1.9 & 2001–07–12 & 5 & $\cdots$ & $\cdots$ & $\cdots$\
GJ 660.1B & $\cdots$ & $\cdots$ & $\cdots$ & SpeX Prism & 2011–03–09 & 25 & $\cdots$ & $\cdots$ & $\cdots$\
J1835$+$3295 & KPNO 4m: R–C Spec & 2001–07–22 & 1 & SpeX Prism & 2003–09–05 & 9 & $\cdots$ & $\cdots$ & $\cdots$\
vB 10 & KPNO 4m: R–C Spec & 2002–09–26 & 14 & SpeX SXD, LXD1.9 & 2001–06–13 & 5, 26 & IRS & 2005-10-11 & 27\
J2000$-$7523 & CTIO 4m: R–C Spec & 2003–04–23 & 14 & FIRE & 2013–07–28 & 28 & $\cdots$ & $\cdots$ & $\cdots$\
J2036+5059 & KAST &2001-12-09 & 29 & SpeX Prism & 2003–10–06 & 24 & $\cdots$ & $\cdots$ & $\cdots$\
J2341$-$1133 & GoldCam & 2002–07–06 & 1 & SpeX Prism & 2010–07–07 & 3 & $\cdots$ & $\cdots$ & $\cdots$\
J2352$-$1100 & GoldCam & 2002–07–06 & 1 & SpeX Prism & 2010-07–07 & 3 & $\cdots$ & $\cdots$ & $\cdots$\
[l c c c c c c c c c c]{} TRAPPIST-1 & prism & M7.5 & $\cdots$ & $1.078\pm0.008$ & $1.054\pm0.003$ & $1.059\pm0.010$ & $0.981 \pm 0.008$ & 1n11 & INT-G & 1\
& SXD & M7.5 & $\cdots$ & $1.119 \pm 0.001$ & $1.070 \pm 0.002$ & $1.070 \pm 0.001$ & $0.971 \pm 0.001$ & 1n11 & INT-G & 2\
& FIRE & M7.5 & $\cdots$ & $1.105 \pm 0.001$ & $1.084 \pm 0.001$ & $1.062 \pm 0.001$ & $0.951 \pm 0.001$ & 1n10 & INT-G & 3\
LHS 132 & prism & M8 & M8 & $1.141 \pm 0.005$ & $1.055 \pm 0.010$ & $1.071 \pm 0.009$ & $0.980 \pm 0.006$ & 1n11 & INT-G & 4\
J0320$+$1854 & prism & M8 & $\cdots$ & $1.244 \pm 0.009$ & $1.056 \pm 0.009$ & $1.076 \pm 0.015$ & $0.952 \pm 0.007$ & 0n10 & FLD-G & 1\
& SXD & M8 & $\cdots$ & $1.228 \pm 0.003$ & $1.049 \pm 0.003$ & $1.095 \pm 0.002$ & $0.941 \pm 0.001$ & 0n00 & FLD-G & 6\
J0436$-$4114 & prism & M8$\beta$ & M9$\gamma$ & $1.076 \pm 0.015$ & $1.104 \pm 0.015$ & $1.065 \pm 0.013$ & $0.982 \pm 0.011$ & 2n12 & VL-G & 1\
J0443$+$0002 & prism & M9$\gamma$ & L1$\gamma$ & $ 1.109 \pm 0.011 $ & $ 1.163 \pm 0.016 $ & $ 1.056 \pm 0.014 $ & $ 0.971 \pm 0.009$ & 2122 & VL-G & 7\
& FIRE & M9$\gamma$ & L1$\gamma$ & $ 1.117 \pm 0.001 $ & $ 1.204 \pm 0.002 $ & $ 1.075 \pm 0.001 $ & $ 0.979 \pm 0.001$ & 2112 & VL-G & 3\
J0518$-$2756 & prism & 11 & 11 & $ 1.141 \pm 0.019 $ & $ 1.278 \pm 0.043 $ & $ 1.060 \pm 0.015 $ & $ 0.945 \pm 0.006$ & 1221 & VL-G & 1\
& FIRE & 11 & 11 & $ 1.097 \pm 0.001 $ & $ 1.300 \pm 0.003 $ & $ 1.073 \pm 0.001 $ & $ 0.948 \pm 0.001$ & 2211 & VL-G & 3\
J0532$+$8246 & NIRSPEC & sdL7 & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $ 1.134 \pm 0.002 $ & $\cdots$ & $\cdots$ & 8\
J0608$-$2753 & prism & M8.5$\gamma$ & L0$\gamma$ & $ 1.066 \pm 0.005 $ & $ 1.195 \pm 0.011 $ & $ 1.062 \pm 0.018 $ & $ 0.996 \pm 0.010 $ & 2112 & VL-G & 1\
& SXD & M8.5$\gamma$ & L0$\gamma$ & $ 1.047 \pm 0.003 $ & $ 1.173 \pm 0.002 $ & $ 1.071 \pm 0.003 $ & $ 0.985 \pm 0.002 $ & 2112 & VL-G & 9\
J0853$-$0329 & SXD & M9 & M9 & $ 1.201 \pm 0.003 $ & $ 1.119 \pm 0.004 $ & $ 1.101 \pm 0.003 $ & $ 0.941 \pm 0.001 $& 0n00 & FLD-G &6\
J1013$-$1356 & prism & sdM9.5 & $\cdots$ &$ 1.133 \pm 0.014 $ & $ 0.987 \pm 0.008 $ & $ 1.024 \pm 0.003 $ & $ 0.957 \pm 0.013 $ & 1021 & INT-G & 10\
J1048$-$3956 & SXD & M9 & $\cdots$ & $ 1.250 \pm 0.003 $ & $ 1.054 \pm 0.003 $ & $ 1.114 \pm 0.002 $ & $ 0.936 \pm 0.002 $ &0n00 & FLD-G & 6\
J1207$-$3900 & prism & L0$\gamma$ & L1$\gamma$ & $ 1.016 \pm 0.016 $ & $ 1.264 \pm 0.019 $ & $ 1.049 \pm 0.016 $ & $ 0.999 \pm 0.015 $ & 2222 & VL-G & 11\
& SXD & L0$\gamma$ & L1$\gamma$ & $ 1.019 \pm 0.005 $ & $ 1.517 \pm 0.018 $ & $ 1.095 \pm 0.002 $ & $ 1.021 \pm 0.002 $ & 2212 & VL-G & 11\
& FIRE & L0$\gamma$ & L1$\gamma$ & $ 1.052 \pm 0.001 $ & $ 1.388 \pm 0.003 $ & $ 1.052 \pm 0.002 $ & $ 0.994 \pm 0.001 $ & 2222 & VL-G & 3\
J1247$-$3816 & prism & $\cdots$ & M9$\gamma$ &$ 1.047 \pm 0.010 $ & $ 1.102 \pm 0.007 $ & $ 1.048 \pm 0.016 $ & $ 0.987 \pm 0.011 $ & 2n22 & VL-G & 11\
J1256$-$0224 & prism & sdL3.5 & $\cdots$ & $ 1.341 \pm 0.052 $ & $ 0.977 \pm 0.017 $ & $ 1.049 \pm 0.008 $ & $ 0.919 \pm 0.015 $ & 1021 & INT-G &12\
& FIRE & sdL3.5 & $\cdots$ & $ 1.336 \pm 0.005 $ & $ 0.982 \pm 0.001 $ & $ 1.090 \pm 0.001 $ & $ 1.082 \pm 0.004 $ & 1012 & INT-G & 13\
LHS 377 & X-Shooter & sdM7 & $\cdots$ & $\cdots$ & $ 0.991 \pm 0.001 $ & $ 0.982 \pm 0.002 $ & $ 0.824 \pm 0.001 $ & $\cdots$ & $\cdots$ & 14\
J1444$-$2019 & prism & sdM9 & $\cdots$ &$ 1.256 \pm 0.019 $ & $ 0.985 \pm 0.013 $ & $ 1.067 \pm 0.007 $ & $ 0.922 \pm 0.006 $ & 1n10 & INT-G & 4\
LHS3003 & prism & M7 & M7 & $ 1.121 \pm 0.006 $ & $ 1.027 \pm 0.007 $ & $ 1.069 \pm 0.008 $ & $ 0.978 \pm 0.007 $ & 0n00 & FLD-G & 4\
J1610$-$0040 & SXD & sdM7 & $\cdots$ & $ 1.102 \pm 0.002 $ & $ 0.984 \pm 0.001 $ & $ 1.058 \pm 0.001 $ & $ 0.958 \pm 0.001 $ & 1n00 & FLD-G & 15\
vB 8 & prism & M7 & M7 & $ 1.155 \pm 0.008 $ & $ 1.015 \pm 0.007 $ & $ 1.064 \pm 0.008 $ & $ 0.978 \pm 0.009 $ & 0n00 & FLD-G &5\
& SXD & M7 & M7 & $ 1.155 \pm 0.002 $ & $ 1.014 \pm 0.002 $ & $ 1.074 \pm 0.001 $ & $ 0.972 \pm 0.001 $ & 0n00 & FLD-G & 6\
GJ 660.1B & prism & $\cdots$ & d/sdM7 &$ 1.200 \pm 0.017 $ & $ 1.012 \pm 0.011 $ & $ 1.087 \pm 0.006 $ & $ 0.958 \pm 0.006 $& 0n00 & FLD-G& 16\
J1835$+$3295 & prism & M8.5 & $\cdots$ &$ 1.142 \pm 0.049 $ & $ 1.057 \pm 0.012 $ & $ 1.094 \pm 0.006 $ & $ 0.927 \pm 0.003 $ &1n00& FLD-G & 1\
vB 10 & SXD & M8 & M8 &$ 1.148 \pm 0.012 $ & $ 1.052 \pm 0.002 $ & $ 1.071 \pm 0.002 $ & $ 0.950 \pm 0.006 $ & 0n10 & FLD-G &6, 17\
J2000$-$7523 & FIRE & M9$\gamma$ & M9$\gamma$ & $ 1.094 \pm 0.001 $ & $ 1.195 \pm 0.001 $ & $ 1.076 \pm 0.002 $ & $ 0.969 \pm 0.001 $& 1n11 & INT-G &18\
J2036$+$5059 & prism & sdM7.5 & $\cdots$ &$ 1.090 \pm 0.011 $ & $ 0.977 \pm 0.009 $ & $ 1.029 \pm 0.003 $ & $ 0.976 \pm 0.007 $ & 1n21 & INT-G & 4\
& SXD & sdM7.5 & $\cdots$ & $ 1.095 \pm 0.002 $ & $ 0.977 \pm 0.001 $ & $ 1.032 \pm 0.001 $ & $ 0.969 \pm 0.001 $ & 1n21 & INT-G & 15\
J2341$-$1133 & prism & M8 & $\cdots$ & $ 1.105 \pm 0.010 $ & $ 1.059 \pm 0.012 $ & $ 1.069 \pm 0.012 $ & $ 0.973 \pm 0.006$ & 1n11 & INT-G & 4\
J2352$-$1100 & prism & M7 & M8$\beta$ & $1.110 \pm 0.004$ & $1.054 \pm 0.008$ & $1.066 \pm 0.009$ & $0.979 \pm 0.006$ & 1n11 & INT-G & 4\
[l c c c c c c c c c c]{} TRAPPIST-1 & SXD & M7.5 & $\cdots$ & $ 1.119 \pm 0.001 $ & $ 1.093 \pm 0.010 $ & $ 1.070 \pm 0.002 $ & $ 1.070 \pm 0.001 $ & $ 0.971 \pm 0.001 $ & 1n01 & INT-G\
& FIRE & M7.5 & $\cdots$ & $ 1.105 \pm 0.001 $ & $ 1.110 \pm 0.009 $ & $ 1.084 \pm 0.001 $ & $ 1.062 \pm 0.001 $ & $ 0.951 \pm 0.001 $ & 1n10 & INT-G\
J0320$+$1854 & SXD & M8 & $\cdots$ & $ 1.228 \pm 0.003 $ & $ 1.168 \pm 0.015 $ & $ 1.049 \pm 0.003 $ & $ 1.095 \pm 0.002 $ & $ 0.941 \pm 0.001 $ & 0n00 & FLD-G\
J0443$+$0002 & FIRE & M9$\gamma$ & L0$\gamma$ & $ 1.117 \pm 0.001 $ & $ 1.113 \pm 0.010 $ & $ 1.204 \pm 0.002 $ & $ 1.075 \pm 0.001 $ & $ 0.979 \pm 0.001 $ & 2122 & VL-G\
J0518$-$2756 & FIRE & L1$\gamma$ & L1$\gamma$ & $ 1.097 \pm 0.001 $ & $ 1.137 \pm 0.011 $ & $ 1.300 \pm 0.003 $ & $ 1.073 \pm 0.001 $ & $ 0.948 \pm 0.001 $ & 2221 & VL-G\
J0532$+$8246 & NIRSPEC & sdL7 & $\cdots$ & $\cdots$ & $1.239 \pm 0.026$ & $\cdots$ & $1.134 \pm 0.002$ & $\cdots$ & $\cdots$ & $\cdots$\
J0608$-$2753 & SXD & M8.5$\gamma$ & L0$\gamma$ & $ 1.047 \pm 0.003 $ & $ 1.081 \pm 0.012 $ & $ 1.173 \pm 0.002 $ & $ 1.071 \pm 0.003 $ & $ 0.985 \pm 0.002 $ & 2122 & VL-G\
J0853$-$0329 & SXD & M9 & M9 &$ 1.201 \pm 0.003 $ & $ 1.119 \pm 0.004 $ & $ 1.192 \pm 0.016 $ & $ 1.101 \pm 0.003 $ & $ 0.941 \pm 0.001 $ & 0n00 & FLD-G\
J1048$-$3956 & SXD & M9 & $\cdots$ &$ 1.250 \pm 0.003 $ & $ 1.190 \pm 0.022 $ & $ 1.054 \pm 0.003 $ & $ 1.114 \pm 0.002 $ & $ 0.936 \pm 0.002 $ & 0n00 &FLD-G\
J1207$-$3900 & SXD & L0$\gamma$ & L1$\gamma$ & $ 1.019 \pm 0.005 $ & $ 1.178 \pm 0.040 $ & $ 1.517 \pm 0.018 $ & $ 1.095 \pm 0.002 $ & $ 1.021 \pm 0.002 $ & 2222 & VL-G\
& FIRE & L0$\gamma$ & L1$\gamma$ & $ 1.052 \pm 0.001 $ & $ 1.119 \pm 0.011 $ & $ 1.388 \pm 0.003 $ & $ 1.052 \pm 0.002 $ & $ 0.994 \pm 0.001 $ & 2222 & VL-G\
J1256$-$0224 & FIRE & sdL3.5 & $\cdots$ & $ 1.336 \pm 0.005 $ & $ 1.161 \pm 0.020 $ & $ 0.982 \pm 0.001 $ & $ 1.090 \pm 0.001 $ & $ 1.082 \pm 0.004 $ & 1012 & INT-G\
LHS 377 & X-Shooter & sdM7 & $\cdots$ & $\cdots$ & $ 1.047 \pm 0.003 $ & $ 0.991 \pm 0.001 $ & $ 0.982 \pm 0.002 $ & $ 0.824 \pm 0.001 $ & $\cdots$ & $\cdots$\
J1610$-$0040 & SXD & sdM7 & $\cdots$ & $ 1.102 \pm 0.002 $ & $ 1.073 \pm 0.010 $ & $ 0.984 \pm 0.001 $ & $ 1.058 \pm 0.001 $ & $ 0.958 \pm 0.001 $ & 1n00 & FLD-G\
vB 8 & SXD & M7 & M7 & $ 1.155 \pm 0.002 $ & $ 1.111 \pm 0.011 $ & $ 1.014 \pm 0.002 $ & $ 1.074 \pm 0.001 $ & $ 0.972 \pm 0.001 $ & 0n00 & FLD-G\
vB 10 & SXD & M8 & M8 & $ 1.148 \pm 0.012 $ & $ 1.036 \pm 0.002 $ & $ 1.052 \pm 0.002 $ & $ 1.071 \pm 0.002 $ & $ 0.950 \pm 0.006 $ & 2n00 & INT-G\
J2000$-$7523 & FIRE & M9$\gamma$ & M9$\gamma$ &$ 1.094 \pm 0.001 $ & $ 1.195 \pm 0.001 $ & $ 1.102 \pm 0.007 $ & $ 1.076 \pm 0.002 $ & $ 0.969 \pm 0.001 $& 1n21& INT-G\
J2036$+$5059 & SXD & sdM7.5 & $\cdots$ & $ 1.095 \pm 0.002 $ & $ 1.055 \pm 0.009 $ & $ 0.977 \pm 0.001 $ & $ 1.032 \pm 0.001 $ & $ 0.969 \pm 0.001 $ & 2n21 & INT-G\
[l c c c c c c c c]{} TRAPPIST-1 & SXD & M7.5 & $\cdots$ & $11.762 \pm 0.095$ & $4.566 \pm 0.084$ & $6.891 \pm 0.073$ & $4.124 \pm 0.12$ & $4.618 \pm 0.067$\
& FIRE & M7.5 & $\cdots$ & $12.049 \pm 0.015$ & $3.913 \pm 0.019$ & $6.690 \pm 0.014$ & $4.043 \pm 0.024$ & $4.027 \pm 0.014 $\
J0320$+$1854 & SXD & M8 & $\cdots$ & $14.426 \pm 0.074$ & $5.685 \pm 0.064 $ & $8.38 \pm 0.068$ & $5.04 \pm 0.10$ & $5.749 \pm 0.063$\
J0443$+$0002 & FIRE & M9$\gamma$ & L0$\gamma$ & $ 9.223 \pm 0.065 $ & $ 2.941 \pm 0.059$ & $ 4.509 \pm 0.054$ & $ 3.629 \pm 0.091$ & $ 2.974 \pm 0.052$\
J0518$-$2756 & FIRE & L1$\gamma$ & L1$\gamma$ & $ 6.25 \pm 0.18$ & $ 4.65 \pm 0.16$ & $ 5.31 \pm 0.14$ & $ 4.56\pm 0.22$ & $3.99\pm 0.12$\
J0532$+$8246 & NIRSPEC & sdL7 & $\cdots$ & $\cdots$ & $ 12.8 \pm 1.8$ & $ 16.0 \pm 1.7$ & $ 0.6 \pm 1.2$ & $ 6.66 \pm 0.88$\
J0608$-$2753 & SXD & M8.5$\gamma$ & L0$\gamma$ & $ 7.12 \pm 0.45$ & $ 2.97 \pm 0.36$ & $ 4.78 \pm 0.33$ & $ 1.42\pm 0.52$ & $ 1.46 \pm 0.27$\
J0853$-$0329 & SXD & M9 & M9 & $13.645\pm0.063$ & $5.956\pm0.057$ & $8.389\pm0.057$ &$4.640\pm 0.092$ &$5.682\pm0.051$\
J1048$-$3956 & SXD & M9 & $\cdots$ &$13.739\pm 0.044$& $7.073\pm 0.043$& $9.620\pm 0.038$& $5.205\pm 0.074$ & $6.604\pm 0.036$\
J1207$-$3900 & SXD & L0$\gamma$ & L1$\gamma$ & $ 12.0 \pm 3.4 $ & $ 3.6\pm 2.4$ & $ 2.5 \pm 1.8$ & $ -10.0 \pm 6.8$ & $ 6.6 \pm 3.3$\
& FIRE & L0$\gamma$ & L1$\gamma$ & $ 6.83 \pm 0.14$ & $ 2.62 \pm 0.13$ & $ 3.10 \pm 0.12$ & $ 3.39 \pm 0.18$ & $ 1.73\pm 0.11$\
J1256$-$0224 & FIRE & sdL3.5 & $\cdots$ & $ 12.56\pm 0.62$ & $ 6.72 \pm 0.42$ & $ 9.09 \pm 0.33$ & $ 2.45 \pm 0.61$ & $ 4.30 \pm 0.34$\
LHS 377 & X-Shooter & sdM7 & $\cdots$ & $ 3.0 \pm 1.0$ & $ 1.11 \pm 0.20$ & $ 3.53 \pm 0.33$ & $ 1.33 \pm 0.12$ & $ 1.31 \pm 0.15$\
J1610$-$0040 & SXD & sdM7 & $\cdots$ & $ 13.08 \pm 0.18$ & $ 4.93 \pm 0.16$ & $ 7.15 \pm 0.13$ & $ 3.43\pm 0.23$ & $ 5.01 \pm 0.13$\
vB 8 & SXD & M7 & M7 & $ 12.169 \pm 0.073$ & $ 4.543 \pm 0.065$ & $ 6.947 \pm 0.056$ & $ 3.87\pm 0.10$ & $ 4.608\pm 0.052$\
vB 10 & SXD & M8 & M8 & $\cdots$ & $4.71 \pm 0.27$ & $6.99 \pm 0.26$ & $5.43 \pm 0.25$ & $4.98 \pm 0.30$\
J2000$-$7523 & FIRE & M9$\gamma$ & M9$\gamma$ &$7.60\pm0.11$& $2.428\pm0.088$& $3.822\pm 0.071$&$2.59\pm0.12$ & $2.053\pm 0.064$\
J2036$+$5059 & SXD & sdM7.5 & $\cdots$ & $ 7.86 \pm 0.32$ & $ 2.18\pm 0.29$ & $ 4.36 \pm 0.27$ & $ 0.71 \pm 0.39$ & $2.20 \pm 0.24$\
[^1]: This paper includes data gathered with the 6.5 meter Magellan Telescopes located at Las Campanas Observatory, Chile.
[^2]: Available at <https://github.com/jgagneastro/FireHose_v2/>
[^3]: SEDkit is available on GitHub at <https://github.com/hover2pi/SEDkit>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Refining some results of S. S. Dragomir, several new reverses of the triangle inequality in inner product spaces are obtained.'
author:
- |
[**Arsalan Hojjat Ansari and Mohammad Sal Moslehian**]{}\
Dept. of Math., Ferdowsi Univ.\
P. O. Box 1159, Mashhad 91775\
Iran\
E-mail: [email protected]\
URL: http://www.um.ac.ir/$\sim$moslehian/
title: 'Refinements of Reverse Triangle Inequalities in Inner Product Spaces[^1]'
---
Introduction.
=============
It is interesting to know under which conditions the triangle inequality went the other way in a normed space $X$; in other words, we would like to know if there is a constant $c$ with the property that $c\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|$ for any finite set $x_1,\cdots,x_n\in X$. M. Nakai and T. Tada $\cite{N-T}$ proved that the normed spaces with this property are precisely those of finite dimensional.
The first authors investigating reverse of the triangle inequality in inner product spaces were J. B. Diaz and F. T. Metcalf $\cite{D-M}$ by establishing the following result as an extension of an inequality given by M. Petrovich $\cite{PET}$ for complex numbers:
[**Diaz-Metcalf Theorem.**]{} Let $a$ be a unit vector in the inner product space $(H;\langle .,.\rangle)$. Suppose the vectors $x_k \in H, k\in\{1,\cdots,n\}$ satisfy $$\begin{aligned}
0\leq r\leq\frac{Re\langle x_k,a\rangle}{\|x_k\|}, k\in\{1,\cdots,n\}\end{aligned}$$ Then $$\begin{aligned}
r\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|.\end{aligned}$$ where equality holds if and only if $$\begin{aligned}
\sum_{k=1}^n x_k =r\sum_{k=1}^n \|x_k\|a.\end{aligned}$$
Inequalities related to the triangle inequality are of special interest; cf. Chapter XVII of $\cite{M-P-F}$ and may be applied to get nice inequalities in complex numbers or to study vector-valued integral inequalities $\cite{DRA1}, \cite{DRA2}$.
Using several ideas and notation of $\cite{DRA1}, \cite{DRA2}$ we modify or refine results of S. S. Dragomir and get some new reverses of triangle inequality.
We use repeatedly the Cauchy-Schwarz inequality without mentioning it. The reader is refered to $\cite{RAS}, \cite{DRA3}$ for the terminology on inner product spaces.
Main Results.
=============
The following theorem is an strengthen of theorem 1 of $\cite{DRA2}$ in which the real numbers $r_1, r_2$ are not neccesarily nonnegative. The proof seems to be different as well.
[**Theorem 1.**]{} Let $a$ be a unit vector in the complex inner product space $(H;\langle .,.\rangle)$. Suppose the vectors $x_k \in H, k\in\{1,\cdots,n\}$ satisfy $$0\leq r_1 ^2 \|x_k\|\leq Re\langle x_k,r_1 a\rangle, 0\leq r_2 ^2 \|x_k\| \leq Im\langle x_k,r_2 a \rangle$$ for some $r_1,r_2\in[-1,1].$ Then we have the inequality $$(r_1 ^2 +r_2 ^2)^\frac{1}{2}\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|.$$ The equality holds in (2) if and only if $$\sum_{k=1}^n x_k =(r_1 +ir_2)\sum_{k=1}^n \|x_k\|a.$$
[**Proof.**]{} If $r_1 ^2 +r_2 ^2 =0$,theorem is trivial. Assume that $r_1 ^2 +r_2 ^2 \neq 0$. Summing inequalities (1) over $k$ from $1$ to $m$, we have $$\begin{aligned}
(r_1 ^2 +r_2 ^2)\sum_{k=1}^n \|x_k\|\leq Re\langle \sum_{k=1}^n x_k ,r_1 a\rangle+Im\langle \sum_{k=1}^n x_k ,r_2a\rangle\\
=Re\langle \sum_{k=1}^n x_k ,(r_1+ir_2)a\rangle\\
\leq |\langle \sum_{k=1}^n x_k ,(r_1+ir_2)a\rangle|\\
\leq\|\sum_{k=1}^n x_k \| \|(r_1 +ir_2)a\|\\
=(r_1^2 +r_2^2)^\frac{1}{2}\|\sum_{k=1}^n x_k \|.\end{aligned}$$ Hence $2$ holds.
If (3) holds, then $$\begin{aligned}
\| \sum_{k=1}^n x_k\|=\|(r_1 +ir_2)\sum_{k=1}^n \|x_k\|a\|=(r_1^2 +r_2^2)^\frac{1}{2}\sum_{k=1}^n \|x_k\|.\end{aligned}$$
Conversely, if the equality holds in (2), we have\
$$\begin{aligned}
(r_1^2 +r_2^2)^\frac{1}{2}\|\displaystyle{\sum_{k=1}^n x_k}\|&=&(r_1^2+r_2^2)\displaystyle{\sum_{k=1}^n \|x_k\|}\leq Re\langle\displaystyle{\sum_{k=1}^n x_k},(r_1+ir_2)a\rangle\\
&\leq&|\langle\displaystyle{\sum_{k=1}^n x_k},(r_1 +ir_2)a\rangle|\leq(r_1^2 +r_2^2)^\frac{1}{2}\|\displaystyle{\sum_{k=1}^n x_k\|}.\end{aligned}$$ From this we deduce $$|\langle\sum_{k=1}^n x_k ,(r_1 +ir_2)a\rangle|=\|\sum_{k=1}^n x_k\|\|(r_1 +ir_2)a\|.$$ Consequently there exists $\eta\geq 0$ such that $\displaystyle{\sum_{k=1}^n} x_k =\eta (r_1 +ir_2)a$
From this we have $$(r_1^2 +r_2^2)^\frac{1}{2}\eta=\|\eta(r_1 +ir_2)a\|=\|\sum_{k=1}^n x_k\|=(r_1^2 +r_2^2)^\frac{1}{2}\sum_{k=1}^n \| x_k\|.$$ Hence $\eta=\displaystyle{\sum_{k=1}^n \|x_k\|}.\Box$
The next theorem is a refinement of Corollary 1 of $\cite{DRA2}$ since, in the notation of the theorem, $\sqrt{2-p_1^2-p_2^2}\leq \sqrt{\alpha_1^2+\alpha_2^2}$.
[**Theorem 2.**]{} Let $a$ be a unit vector in the complex inner product space $(H;\langle .,.\rangle)$. Suppose the vectors $x_k \in H-\{0\}, k\in\{1,\cdots,n\}$, such that $$\|x_k-a\|\leq p_1, \|x_k-ia\|\leq p_2, p_1, p_2\in(0,\sqrt{\alpha^2+1})$$ where $\alpha=\displaystyle{\min_{1\leq k\leq n}}\|x_k\|$. Let $$\begin{aligned}
\alpha_1=\min\{\frac{\|x_k\|^2-p_1^2+1}{2\|x_k\|}:1\leq k\leq n\}, \alpha_2=\min\{\frac{\|x_k\|^2-p_2^2+1}{2\|x_k\|}:1\leq k\leq n\},\end{aligned}$$ Then we have the inequality $$\begin{aligned}
(\alpha_1^2+\alpha_2^2)^\frac{1}{2}\sum_{k=1}^n \| x_k\|\leq\|\sum_{k=1}^n x_k\|\end{aligned}$$ where the equality holds if and only if $$\begin{aligned}
\sum_{k=1}^n x_k=(\alpha_1+i\alpha_2)\sum_{k=1}^n \|x_k\|a\end{aligned}$$
[**Proof.**]{} From the first inequality in $(4)$ we have $$\langle x_k-a,x_k-a\rangle\leq p_1^2$$ $$\| x_k\|^2+1-p_1^2\leq2Re\langle x_k, a\rangle, k=1,\cdots,n$$ $$\frac{\|x_k\|^2-p_1^2+1}{2\|x_k\|}\|x_k\|\leq Re\langle x_k,a\rangle$$ consequently $$\alpha_1\|x_k\|\leq Re\langle x_k,a\rangle.$$ Similarly from the second inequality we obtain $$\alpha_2\|x_k\|\leq Re\langle x_k,ia\rangle=Im\langle x_k,a\rangle.$$
Now apply Theorem 1 for $r_1=\alpha_1,r_2=\alpha_2.\Box$
[**Corollary 3.**]{} Let $a$ be a unit vector in the complex inner product space $(H;\langle .,.\rangle)$. Suppose the vectors $x_k \in H-\{0\}, k\in\{1,\cdots,n\}$ such that $$\|x_k-a\|\leq 1 , \|x_k-ia\|\leq 1.$$Then $$\begin{aligned}
\frac{\alpha}{\sqrt 2}\sum_{k=1}^n \| x_k\|\leq\|\sum_{k=1}^n x_k\|\end{aligned}$$ in which $ \alpha=\displaystyle{\min_{1\leq k\leq n}}\|x_k\|.$ The equality holds if and only if $$\begin{aligned}
\sum_{k=1}^n x_k=\alpha\frac{(1+i)}{2}\sum_{k=1}^n \|x_k\|a\end{aligned}$$
[**Proof.**]{} Apply Theorem 2 for $\alpha_1=\frac{\alpha}{2}=\alpha_2.\Box$
[**Theorem 4.**]{} Let $a$ be a unit vector in the inner product space $(H;\langle .,.\rangle)$ over the real or complex number field. Suppose that the vectors $x_k \in H-\{0\}, k\in\{1,\cdots,n\}$ satisfy $$\|x_k-a\|\leq p , p\in(0,\sqrt{\alpha^2+1}),\alpha=\displaystyle{\min_{1\leq k\leq n}}\|x_k\|.$$
Then we have the inequality $$\alpha_1\sum_{k=1}^n \|x_k\|\leq\|\sum_{k=1}^n x_k\|$$ where $$\alpha_1=\min\{\frac{\|x_k\|^2-p^2+1}{2\|x_k\|}:1\leq k\leq n\}$$.
The equality holds if and only if $$\sum_{k=1}^n x_k=\alpha_1\sum_{k=1}^n \|x_k\|a$$
[**Proof.**]{} The proof is similar to Theorem 2 in which we use Theorem 1 with $r_2=0.\Box$
The next Theorem is a generalization of Theorem 1. It is a modification of Theorem 3 of $\cite{DRA2}$, however our proof is apparently different.
[**Theorem 5.**]{} Let $a_{1},\ldots ,a_{m}$ be orthonormal vectors in the complex inner product space $(H;\langle .,.\rangle)$. Suppose that for $1\leq t\leq m ,r_t,\rho_t\in R$ and that the vectors $x_k \in H, k\in\{1,\cdots,n\}$ satisfy\
$$0\leq r_t ^2 \|x_k\|\leq Re\langle x_k,r_t a_t\rangle, 0\leq \rho_t ^2 \|x_k\| \leq Im\langle x_k,\rho_t a_t \rangle, t\in\{1,\cdots,m\}$$ Then we have the inequality $$(\sum_{t=1}^m r_t ^2+\rho_t ^2)^\frac{1}{2}\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|$$ The equality holds in (7) if and only if $$\sum_{k=1}^n x_k =\sum_{k=1}^n \|x_k\|\sum_{t=1}^m (r_t+i\rho_t)a_t.$$
[**Proof.**]{} If $\displaystyle{\sum_{t=1}^m}(r_t ^2 +\rho_t ^2)=0$, theorem is trivial. Assume that $\displaystyle{\sum_{t=1}^m}(r_t ^2 +\rho_t ^2)\neq 0.$ Summing inequalities (6) over $k$ from $1$ to $n$ and again over $t$ from $1$ to $m$ we get $$\sum_{t=1}^m(r_t ^2 +\rho_t ^2)\sum_{k=1}^n\|x_k\|\leq Re\langle \sum_{k=1}^n x_k ,\sum_{t=1}^m r_t a_t \rangle+Im \langle \sum_{k=1}^n x_k ,\sum_{t=1}^m \rho_t a_t \rangle$$ $$=Re\langle \sum_{k=1}^n x_k ,\sum_{t=1}^m r_t a_t \rangle+Re \langle \sum_{k=1}^n x_k ,i\sum_{t=1}^m \rho_t a_t \rangle$$ $$= Re\langle \sum_{k=1}^n x_k ,\sum_{t=1}^m (r_t+i\rho_t)a_t \rangle$$ $$\leq |\langle \sum_{k=1}^n x_k ,\sum_{t=1}^m (r_t+i\rho_t)a_t \rangle|$$ $$\leq\|\sum_{k=1}^n x_k\|\|\sum_{t=1}^m (r_t+i\rho_t)a_t\|$$ $$=\|\sum_{k=1}^n x_k\|(\sum_{t=1}^m(r_t ^2 +\rho_t ^2))^\frac{1}{2}.$$ Then $$(\sum_{t=1}^m(r_t ^2 +\rho_t ^2))^\frac{1}{2}\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|.$$ If $(8)$ holds, then $$\|\sum_{k=1}^n x_k\|=\|\sum_{k=1}^n \|x_k\|\|\sum_{t=1}^m (r_t+i\rho_t)a_t\|=
\sum_{k=1}^n \|x_k\|(\sum_{t=1}^m(r_t ^2 +\rho_t ^2))^{1/2}.$$ Conversely, if the equality holds in $(7)$ we obtain from (6) that $$(\sum_{t=1}^m(r_t ^2 +\rho_t ^2))^{1/2}\|\sum_{k=1}^n x_k\|=$$ $$=\sum_{t=1}^m(r_t ^2 +\rho_t ^2)\sum_{k=1}^n\|x_k\|\leq Re\langle \sum_{k=1}^n x_k ,\sum_{t=1}^m (r_t+i\rho_t)a_t \rangle\leq$$ $$\leq |\langle \sum_{k=1}^n x_k ,\sum_{t=1}^m (r_t+i\rho_t)a_t \rangle|\leq$$ $$\leq\| \sum_{k=1}^n x_k \|\|\sum_{t=1}^m (r_t+i\rho_t)a_t\|=\| \sum_{k=1}^n x_k \|(\sum_{t=1}^m(r_t ^2 +\rho_t ^2))^{1/2}$$ Thus we have $$|\langle \sum_{k=1}^n x_k ,\sum_{t=1}^m (r_t+i\rho_t)a_t \rangle|=\| \sum_{k=1}^n x_k \|\|\sum_{t=1}^m (r_t+i\rho_t)a_t\|.$$ Consequently there exists $\eta\geq 0$ such that $$\sum_{k=1}^n x_k =\eta\sum_{t=1}^m (r_t+i\rho_t)a_t$$ from which we have $$\eta(\sum_{t=1}^m (r_t^2+\rho_t^2))^\frac{1}{2}=\|\eta\sum_{t=1}^m (r_t+i\rho_t)a_t\|=\|\sum_{k=1}^n x_k\|=\sum_{k=1}^n\| x_k\|(\sum_{t=1}^m (r_t^2+\rho_t^2))^\frac{1}{2}$$Hence $$\eta=\sum_{k=1}^n\| x_k\|.\Box$$
[**Corollary 6.**]{} Let $a_{1},\ldots ,a_{m}$ be orthornormal vectors in the inner product space $(H;\langle .,.\rangle)$ over the real or complex number field. Suppose for $1\leq t\leq m $ that the vectors $x_k \in H, k\in\{1,\cdots,n\}$ satisfy $$0\leq r_t ^2 \|x_k\|\leq Re\langle x_k,r_t a_t\rangle.$$ Then we have the inequality $$(\sum_{t=1}^m r_t ^2 )^\frac{1}{2}\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|.$$ The equality holds if and only if $$\sum_{k=1}^n x_k =\sum_{k=1}^n \|x_k\|\sum_{t=1}^m r_t a_t.$$
[**Proof.**]{} Apply Theorem 5 for $\rho_t=0.\Box$
[**Theorem 7.**]{} Let $a_{1},\ldots ,a_{m}$ be orthornormal vectors in the complex inner product space $(H;\langle .,.\rangle)$. Suppose that the vectors $x_k \in H-\{0\}, k\in\{1,\cdots,n\}$ satisfy $$\|x_k-a_t\|\leq p_t ,\|x_k-ia_t\|\leq q_t, p_t, q_t\in(0,\sqrt{\alpha^2+1}),1\leq t\leq m$$ where $ \alpha=\displaystyle{\min_{1\leq k\leq n}}\|x_k\|.$ Let $$\alpha_t=\min\{\frac{\|x_k\|^2-p_t^2+1}{2\|x_k\|}:1\leq k\leq n\},\beta_t=\min\{\frac{\|x_k\|^2-q_t^2+1}{2\|x_k\|}:1\leq k\leq n\}.$$ Then we have the inequality $$(\sum_{t=1}^m\alpha_t^2+\beta_t^2)^\frac{1}{2}\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|$$ where equality holds if and only if $$\sum_{k=1}^n x_k=\sum_{k=1}^n\|x_k\|\sum_{t=1}^m(\alpha_t+i\beta_t)a_t.$$
[**Proof.**]{} For $1\leq t\leq m, 1\leq k\leq n$ it follows from $\|x_k-a_t\|\leq p_t$ that $$\langle x_k-a_t\rangle,x_k-a_t\rangle\leq p_t^2$$ $$\frac{\|x_k\|^2-p_t^2+1}{2\|x_k\|}\|x_k\|\leq Re\langle x_k,a_t\rangle0$$ $$\alpha_t\|x_k\|\leq Re\langle x_k,a_t\rangle$$ and similarly $$\beta_t\|x_k\|\leq Re\langle x_k,ia_t\rangle=Im\langle x_k,a_t\rangle,$$ Now applying Theorem 4 with $r_t=\alpha_t, \rho_t=\beta_t$ we deduce the desired inequality.$\Box$
[**Corollary 8.**]{} Let $a_{1},\ldots ,a_{m}$ be orthornormal vectors in the complex inner product space $(H;\langle .,.\rangle)$. Suppose that the vectors $x_k \in H, k\in\{1,\cdots,n\}$ satisfy $$\|x_k-a_t\|\leq 1 ,\|x_k-ia_t\|\leq 1, 1\leq t\leq m$$ Then $$\frac{\alpha}{\sqrt{2}}\sqrt{m}\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|.$$ The equality holds if and only if $$\sum_{k=1}^n x_k=\alpha\frac{(1+i)}{2}\sum_{k=1}^n\|x_k\|\sum_{t=1}^m a_t$$
[**Proof.**]{} Applying Theorem 7 for $\alpha_t=\frac{\alpha}{2}=\beta_t.\Box$
[**Remark.**]{} It is interesting to note that
$$\frac{\alpha}{\sqrt{2}}\sqrt{m}\leq\frac{\|\sum_{k=1}^n x_k\|}{\sum_{k=1}^n\|x_k\|}\leq 1.$$ $$\alpha\leq\sqrt{\frac{2}{m}}$$
[**Corollary 9.**]{} Let $a$ be a unit vector in the complex inner product space $(H;\langle .,.\rangle)$. Suppose that the vectors $x_k \in H-\{0\}, k\in\{1,\cdots,n\}$ satisfy\
$$\|x_k-a\|\leq p_1, \|x_k-ia\|\leq p_2, p_1, p_2\in(0,1].$$ Let $$\alpha_1=\min\{\frac{\|x_k\|^2-p_1^2+1}{2\|x_k\|}:1\leq k\leq n\},\alpha_2=\min\{\frac{\|x_k\|^2-p_2^2+1}{2\|x_k\|}:1\leq k\leq n\}.$$ If $\alpha_1\neq(1-p_1^2)^\frac{1}{2}$, or $\alpha_2\neq(1-p_2^2)^\frac{1}{2}$, then we have the following strictly inequality $$(2-p_1^2-p_2^2)^\frac{1}{2}\sum_{k=1}^n \| x_k\|<\|\sum_{k=1}^n x_k\|$$
[**Proof.**]{} If equality holds, then by Theorem 2 we have $$(\alpha_1^2+\alpha_2^2)^\frac{1}{2}\sum_{k=1}^n \| x_k\|\leq\|\sum_{k=1}^n x_k\|=(2-p_1^2-p_2^2)^\frac{1}{2}\sum_{k=1}^n \| x_k\|$$ and so $$(\alpha_1^2+\alpha_2^2)^\frac{1}{2}\leq(2-p_1^2-p_2^2)^\frac{1}{2}.$$ On the other hand for $1\leq k\leq n$, $$\frac{\|x_k\|^2-p_1^2+1}{2\|x_k\|}\geq(1-p_1^2)^\frac{1}{2}$$ and so $$\alpha_1 \geq (1-p_1^2)^\frac{1}{2}.$$ Similarly $$\alpha_2\geq(1-p_2^2)^\frac{1}{2}.$$ Hence $$(2-p_1^2-p_2^2)^\frac{1}{2}\leq(\alpha_1^2+\alpha_2^2)^\frac{1}{2}$$ Thus $$\sqrt{\alpha_1^2+\alpha_2^2}=(2-p_1^2-p_2^2)^\frac{1}{2}.$$ Therefore $$\alpha_1=(1-p_1^2)^\frac{1}{2} {\rm and} \alpha_2=(1-p_2^2)^\frac{1}{2}$$ a contradiction.$\Box$
The following result looks like Corollary 2 of $\cite{DRA2}$.
[**Theorem 10**]{}. Let $a$ be a unit vector in the complex inner product space $(H;\langle .,.\rangle), M\geq m >0, L\geq \ell>0$ and $x_k \in H-\{0\}, k\in\{1,\cdots,n\}$ such that $$Re\langle Ma-x_k,x_k-ma\rangle\geq0,Re\langle L ia-x_k,x_k-\ell ia\rangle\geq0$$ or equivalently, $$\|x_k-\frac{m+M}{2}a\|\leq\frac{M-m}{2},\|x_k-\frac{L+\ell}{2}ia\|\leq\frac{L-\ell}{2}.$$ Let $$\alpha_{m,M}=\min\{\frac{\|x_k\|^2+mM}{(m+M)\|x_k\|}:1\leq k\leq n\}$$ and $$\alpha_{\ell,L}=\min\{\frac{\|x_k\|^2+\ell L}{(\ell+L)\|x_k\|}:1\leq k\leq n\}$$ Then we have the inequlity $$(\alpha_{m,M}^2+\alpha_{\ell,L}^2)^\frac{1}{2}\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|.$$ The equality holds if and only if $$\sum_{k=1}^n x_k=(\alpha_{m,M}+i\alpha_{\ell,L})\sum_{k=1}^n\|x_k\|a.$$
[**Proof**]{}. For each $1\leq k\leq n$, it follows from $$\|x_k-\frac{m+M}{2}a\|\leq\frac{M-m}{2}$$ that $$\langle x_k-\frac{m+M}{2}a, x_k-\frac{m+M}{2}\rangle\leq(\frac{M-m}{2})^2.$$ Hence $$\|x_k\|^2+mM\leq(m+M)Re\langle x_k, a\rangle.$$ Then $$\frac{\|x_k\|^2+mM}{(m+M)\|x_k\|}\|x_k\|\leq Re\langle x_k, a\rangle$$ consequently $$\alpha_{m,M}\|x_k\|\leq Re\langle x_k, a\rangle.$$ Similarly from the second inequlity we deduce $$\alpha_{\ell,L}\|x_k\|\leq Im\langle x_k, a\rangle.$$ Applying Theorem 1 for $ r_1=\alpha_{m,M},r_2=\alpha_{\ell,L}$, we infer the desired inequality.$\Box$
[**Theorem 11**]{}. Let $a$ be a unit vector in the complex inner product space $(H;\langle .,.\rangle), M \geq m > 0, L\geq \ell > 0$ and $x_k \in H-\{0\}, k\in\{1,\cdots,n\}$ such that $$Re\langle Ma-x_k,x_k-ma\rangle\geq 0,Re\langle Lia-x_k,x_k-\ell ia\rangle\geq 0$$ or equivalently $$\|x_k-\frac{m+M}{2}a\|\leq\frac{M-m}{2},\|x_k-\frac{L+\ell}{2}ia\|\leq\frac{L-\ell}{2}.$$Let $$\alpha_{m,M}=\min\{\frac{\|x_k\|^2+mM}{(m+M)\|x_k\|}:1\leq k\leq n\}$$ and $$\alpha_{\ell,L}=\min\{\frac{\|x_k\|^2+\ell L}{(\ell+L)\|x_k\|}:1\leq k\leq n\}.$$ If $\alpha_{m,M}\neq 2\frac{\sqrt{mM}}{m+M}$, or $\alpha_{\ell,L}\neq 2\frac{\sqrt{\ell L}}{\ell+L}$, then we have $$2(\frac{mM}{(m+M)^2}+\frac{\ell L}{(\ell+L)^2})^\frac{1}{2}\sum_{k=1}^n\|x_k\|<\|\sum_{k=1}^n x_k\|.$$
[**Proof**]{}. If $2(\frac{mM}{(m+M)^2}+\frac{\ell L}{(\ell+L)^2})^\frac{1}{2}\displaystyle{\sum_{k=1}^n}\|x_k\|=\|\displaystyle{\sum_{k=1}^n} x_k\|$ then by theorem 10 we have $$(\alpha_{m,M}^2+\alpha_{\ell,L}^2)^\frac{1}{2}\sum_{k=1}^n\|x_k\|\leq\|\sum_{k=1}^n x_k\|\\
=2(\frac{mM}{(m+M)^2}+\frac{\ell L}{(\ell+L)^2})^\frac{1}{2}\sum_{k=1}^n\|x_k\|.$$ Consequently $$(\alpha_{m,M}^2+\alpha_{\ell,L}^2)^\frac{1}{2}\leq 2(\frac{mM}{(m+M)^2}+\frac{\ell L}{(\ell+L)^2})^\frac{1}{2}.$$ On the other hand for $ 1\leq k\leq n,\frac{\|x_k\|^2+mM}{(m+M)\|x_k\|}\geq 2\frac{\sqrt{mM}}{m+M}$, and $\frac{\|x_k\|^2+\ell L}{(\ell+L)\|x_k\|}\geq 2\frac{\sqrt{\ell L}}{\ell+L}$, so $$(\alpha_{m,M}^2+\alpha_{\ell,L}^2)^\frac{1}{2}\geq 2(\frac{mM}{(m+M)^2}+\frac{\ell L}{(\ell+L)^2})^\frac{1}{2}.$$ Then $$(\alpha_{m,M}^2+\alpha_{\ell,L}^2)^\frac{1}{2}=2(\frac{mM}{(m+M)^2}+\frac{\ell L}{(\ell+L)^2})^\frac{1}{2}.$$ Hence $$\alpha_{m,M}= 2\frac{\sqrt{mM}}{m+M}$$ and $$\alpha_{\ell,L}= 2\frac{\sqrt{\ell L}}{\ell+L}$$ a contradection.$\Box$
Finally we mention two applications of our results to the complex numbers.
[**Corollary 12.**]{} Let $a\in C$ with $|a|=1$. Suppose that $z_k \in C, k\in\{1,\cdots,n\}$ such that $$|z_k-a|\leq p_1 , |z_k-ia|\leq p_2,p_1, p_2\in(0,\sqrt{\alpha^2+1})$$ where $$\alpha=\min\{|z_k|:1\leq k\leq n\}.$$ Let $$\alpha_1=\min\{\frac{|z_k|^2-p_1^2+1}{2|z_k|}:1\leq k\leq n\},\alpha_2=\min\{\frac{|z_k|^2-p_2^2+1}{2|z_k|}:1\leq k\leq n\}.$$ Then we have the inequality $$\sqrt{\alpha_1^2+\alpha_2^2}\sum_{k=1}^n | z_k|\leq|\sum_{k=1}^n z_k|.$$ The equality holds if and only if $$\sum_{k=1}^n z_k=(\alpha_1+i\alpha_2)(\sum_{k=1}^n |z_k|)a.$$
[**Proof.**]{} Apply Theorem 2 for $H=C.\Box$
[**Corollary 13.**]{} Let $a\in C$ with $|a|=1$. Suppose that $z_k \in C, k\in\{1,\cdots,n\}$ such that $$|z_k-a|\leq 1 , |z_k-ia|\leq 1.$$ If $\alpha=\min\{|z_k|:1\leq k\leq n\}$. Then we have the inequality $$\frac{\alpha}{\sqrt 2}\sum_{k=1}^n |z_k|\leq|\sum_{k=1}^n z_k|$$ the equality holds if and only if $$\sum_{k=1}^n z_k=\alpha\frac{(1+i)}{2}(\sum_{k=1}^n |z_k|)a.$$
[**Proof.**]{} Apply Corollary 3 for $H=C.\Box$
[99]{} J.B. Diaz and F.T. Metcalf, A complementary triangle inequality in Hilbert and Banach spaces, Proc. Amer. Math. Soc. 17(1) (1966), 88-97. S.S. Dragomir, Reverses of the triangle inequality in inner product spaces, arXiv:math.FA/0405495. S.S. Dragomir, Some reverses of the generalized triangle inequality in complex inner product spaces, arXiv:math.FA/0405497. S.S. Dragomir, [*Discrete inequalities of the Cauchy-Bunyakovsky-Schwarz type*]{}. Nova Science Publishers, Inc., Hauppauge, NY, 2004. D.S. Mitrinovi' c, J.E. Pečari' c and A.M. Fink, Classical and New Inequalities in Analysis, Kluwer Academic Publishers, Dordrecht/Boston/London, 1993. M. Nakai and T. Tada, The reverse triangle inequality in normed spaces, New Zealand J. Math. 25 (1996), no. 2, 181–193. M. Petrovich, Module dune somme, L Ensignement Math' ematique, 19 (1917), 53-56. Th.M. Rassias, [*Inner Product Spaces and Applications*]{}, Chapman-Hall,1997.
[^1]: [*2000 Mathematics Subject Classification*]{}. Primary 46C05; Secondary 26D15.\
[*Key words and phrases*]{}. Triangle inequality, reverse inequality, Diaz-Metkalf inequality, inner product space.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A flux of cosmic rays (CRs) propagating through a diffuse ionized gas can excite MHD waves, thus generating magnetic disturbances. We propose a generic model of CR penetration into molecular clouds through their diffuse envelopes, and identify the leading physical processes controlling their transport on the way from a highly ionized interstellar medium to a dense interior of the cloud. The model allows us to describe a transition between a free streaming of CRs and their diffusive propagation, determined by the scattering on the self-generated disturbances. A self-consistent set of equations, governing the diffusive transport regime in an envelope and the MHD turbulence generated by the modulated CR flux, is essentially characterized by two dimensionless numbers. We demonstrate a remarkable mutual complementarity of different mechanisms leading to the onset of the diffusive regime, which results in a universal energy spectrum of the modulated CRs. In conclusion, we briefly discuss implications of our results for several fundamental astrophysical problems, such as the spatial distribution of CRs in the Galaxy as well as the ionization, heating, and chemistry in dense molecular clouds.'
author:
- 'A. V. Ivlev$^1$, V. A. Dogiel$^2$, D. O. Chernyshov$^{2,3,4,5}$, P. Caselli$^1$, C.-M. Ko$^5$ & K. S. Cheng$^3$'
title: 'Penetration of Cosmic Rays into Dense Molecular Clouds: Role of Diffuse Envelopes[^1]'
---
Introduction {#intro}
============
Cosmic rays (CRs) represent a crucial ingredient for the dynamical and chemical evolution of interstellar clouds. Interaction of CRs with molecular clouds is accompanied by various processes generating observable radiation signatures, such as ionization of molecular hydrogen [see, e.g., @oka05; @dalg06; @indri12] and iron [e.g., @dog98; @dog11; @tati12; @yus; @nobu15; @kriv17], as well as production of neutral pions whose decay generates gamma rays in the GeV [e.g., @yang1; @yang2; @tib15] and TeV [e.g., @ahar06; @abram16; @abdal17] energy ranges. Being a unique source of ionization in dark clouds, where the interstellar radiation cannot penetrate, CRs provide a partial coupling of the gas to magnetic field lines, which could slow down or prevent further contraction of the cloud [e.g., @shu87]. CRs are fundamental for the starting of astrochemistry, as they promote the formation of H$_3^+$ ions, which can easily donate a proton to elements such as C and O, and thus eventually form molecules containing elements heavier than H [e.g., @yamamoto17]. Through the ionization of H$_2$ molecules and the consequent production of secondary electrons, CRs are an important heating source of dark regions [e.g., @goldsmith01]. Their interaction with H$_2$ can also result in molecular excitation, followed by fluorescence producing a tenuous UV field within dark clouds and dense cores [@cecchi92; @shen04; @ivlev15a]; this UV field can photodesorb molecules from the icy dust mantles and help maintaining a non-negligible amount of heavy molecules (such as water) in the gas phase [e.g., @caselli12]. Furthermore, CRs can directly impinge on dust grains and heat up the icy mantles, causing catastrophic explosions of these mantles [@leger85; @ivlev15b] and activating the chemistry in solids [@shingledecker17]. Finally, CRs play a fundamental role in the charging of dust grains and the consequent dust coagulation [@okuzumi09; @ivlev15a; @ivlev16], particularly important for the formation of circumstellar disks [e.g., @zhao16] and planet formation in more evolved protoplanetary disks [e.g., @testi14].
One of the fundamental questions is how interstellar (IS) CRs penetrate into molecular clouds, i.e., what are the governing mechanisms of this process and how does this affect the CR spectrum inside the clouds. The crucial point here is that the IS spectrum may be significantly modified while traversing the outer diffuse envelope of a cloud, before reaching the cloud interior.
There are, at least, three important factors which may critically affect the CR spectra inside the clouds:
1. The cloud structure is strongly nonuniform. Dense cloud cores with the gas density $n_{\rm g}= 10^4-
10^7$ cm$^{-3}$ are surrounded by low-density envelopes with $n_{\rm g}=10-10^3$ cm$^{-3}$ [see @lis90; @protheroe]. In the central molecular zone these envelopes occupy up to 30% of the space [see @oka05; @indri12].
2. It is known since a long time [see @lerche67; @kuls69] that a CR flux propagating through a plasma can excite MHD waves and, thus, create magnetic disturbances. A linear analysis [e.g., @dog85] suggests that the waves are expected to be excited near most of the molecular clouds. However, it is still an open question as to whether the resulting disturbances are essential [see @skill76; @cesar78] or not [see @gabici15] for the CR penetration into the clouds.
3. CR energy losses in the envelope are determined by ionization, proton-proton collisions, and MHD-wave excitation [see @skill76; @padovani; @padovani1; @ivlev15a; @schlick16]. A relative importance of these processes also needs to be carefully analyzed.
Attempts to analyze a system of nonlinear equations describing the CR-wave interaction in molecular clouds were undertaken in several publications [see, e.g., @skill76; @cesar78; @gabici15]. We notice however, that in all these cases the analysis was based on relatively simple estimates rather than on the exact solution of the equations. Nevertheless, @skill76 showed that interactions of CRs with waves should lead to depletion of their density inside the clouds at energies below $\sim100$ MeV. Later, @cesar78 demonstrated that the depletion can be even stronger if the effect of magnetic field compression is taken into account. In a recent paper, @gabici15 estimated the flux velocity of CRs penetrating into a cloud to be about the Alfven speed for all energies. (Below we will see that this estimate is correct only for relatively low energies.) For the sake of completeness, one should also mention analysis of the CR-wave interaction undertaken by @dog94 for processes of CR escape from the Galaxy, and by @recc16a [@recc16b] to describe the spatial distribution of Galactic CRs and the CR-driven Galactic winds. These problems are, however, clearly beyond the scope of our paper.
The principal goal of the present paper is an attempt to formulate a self-consistent generic model of CR penetration into molecular clouds through their diffuse envelopes. We identify the leading physical processes controlling the CR propagation on the way from a highly ionized interstellar medium to a dense interior of the cloud. In our analysis we do not presume a regime of CR propagation in the envelope, but instead derive it from the model. This allows us to reveal the mutual interplay of the factors mentioned above, and thus to address a number of important specific questions, such as:
1. What is the regime of CR propagation in molecular cloud envelopes – do CRs freely cross the envelope, or do they experience significant scattering by the self-generated MHD turbulence?
2. What characteristics of the interstellar CR spectrum and parameters of a diffuse envelope determine the propagation regime?
3. Do CRs lose a significant part of their energy by MHD-wave excitation in the envelope, or do regular losses due to interaction with gas dominate?
4. Can (some of) the above processes cause a strong self-modulation of the CR flux penetrating into a dense core?
The paper is organized as follows: In Section \[Eqs\] we present a self-consistent set of equations, governing the diffusive regime of CR transport in a molecular cloud envelope and the MHD turbulence generated by the modulated CR flux. In Section \[normal\] we write the governing equations in the dimensionless form and show that the diffusive regime is described by a single dimensionless number $\nu$ (wave damping rate), while a transition to the free-streaming regime is characterized by the small parameter $\epsilon$ (ratio of the Alfven velocity to the speed of light). In Section \[model\] we consider an idealized problem setup, where CRs propagate toward an “absorbing wall” and the energy losses due to their interaction with gas are negligible. This allows us to determine basic conditions of the onset of the diffusion zone in the cloud envelope, and to identify generic properties of the nonlinear CR diffusion. In Section \[realistic\] we study the effect of gas losses on the diffusion and, in particular, on the magnitude of the modulated CR flux penetrating into the cloud. Finally, in Section \[discussion\] we point out a remarkable mutual complementarity of different mechanisms leading to the onset of the diffusive regime, which results in a universal energy spectrum of the modulated CRs. Implications of our results for several fundamental astrophysical problems are briefly discussed.
Governing equations {#Eqs}
===================
In weakly ionized cloud envelopes, where the gas density $n_{\rm g}$ typically does not exceed $\sim10^3$ cm$^{-3}$, the strength of the magnetic field $B$ is practically independent of $n_{\rm g}$ [and is of the order of $10~\mu$G, see @crutch12]. For this reason, we do not consider effects of large-scale variations of $B$, which may be essential for CR propagation in dense cloud cores [e.g., @cesar78; @schlick08]. Also, since the Larmor radius of CRs with energies relevant to our problem is much smaller than the spatial extent of a typical envelope, a stream of such rapidly gyrating CRs is parallel to the magnetic field. Hence, the problem can be considered as one-dimensional, with the coordinate $z$ measured along the field line.
A CR flux can effectively excite Alfven and fast magnetosonic waves in a cold magnetized plasma. Low-frequency disturbances of the magnetic field associated with these waves can, in turn, effectively scatter CRs. The maximum growth rate is achieved for the waves propagating along the magnetic field in the direction of the CR flux. The growth rate is then the same for both wave modes [@kuls69], propagating with the Alfven phase velocity, $$v_{\rm A}=\frac{B}{\sqrt{4\pi m_{\rm i}n_{\rm i}}}\,,$$ where $n_{\rm i}$ and $m_{\rm i}$ are the ion density and mean ion mass, respectively.
Let us introduce steady-state local distribution functions of CRs in the momentum and energy space, averaged over pitch angle and denoted as $F(p,z)$ and $N(E,z)$, respectively. They are related to each other via $$4\pi p^2F(p,z)=vN(E,z)\equiv4\pi j(E,z),$$ where $j(E,z)$ is the so-called [*CR energy spectrum*]{}. The particle momentum as a function of the kinetic energy is $$\label{p(E)}
p(E)=c^{-1}\sqrt{E(E+2m_{\rm p}c^2)}\,,$$ the physical velocity is $v(E)=p(E)c^2/(E+m_{\rm p}c^2)$. The local flux of CRs through a unit area and per unit energy interval is defined as[^2] $$S(E,z)\simeq-\min\left\{D\frac{\partial N}{\partial z}+v_{\rm A}N,~S_{\rm free}\right\}.
\label{smin}$$ In such a definition, the flux continuously changes between the [*diffusive regime*]{} (first term; in what follows it is referred to as the [*modulated flux*]{}), where the mean free path of CRs due to pitch-angle scattering on MHD turbulence is sufficiently small, and the [*free-streaming regime*]{} (second term), where the scattering is negligible. For the former regime, where the pitch-angle distribution is quasi-isotropic, the flux consists of the diffusion and advection parts [see, e.g., @went74], with $D(E,z)$ being the spatial diffusion coefficient of CRs. In turn, the magnitude of the free-streaming flux, $$\label{sfree}
S_{\rm free}(E,z)=\langle\mu\rangle vN,$$ is determined by average pitch angle of CRs in this regime, $\langle\mu\rangle$, which is generally not small. A discussion of different free-streaming zones and estimates for the corresponding $\langle\mu\rangle$ is presented in Appendix \[<mu>\].
The steady-state CR flux is governed by the transport equation [see, e.g., @skill76; @ber90] $$\frac{\partial S}{\partial z}=-\frac{\partial }{dE}\left(\dot E_{\rm g}N\right),
\label{Diff_Eq}$$ where $\dot E_{\rm g}(E)$ describes energy losses due to collisions with gas (“gas losses”). Here, we omit on purpose “wave losses”, i.e., the term due to the adiabatic expansion of the magnetic disturbances associated with MHD waves. The role of this term is discussed in Section \[WL\], where we show that the wave losses are generally unimportant for our problem. Furthermore, for waves propagating in one direction the mechanism of momentum diffusion (Fermi acceleration) does not operate [see, e.g., @ber90], and therefore the corresponding term is also not included in Equation (\[Diff\_Eq\]).
The diffusion coefficient of CRs [@kuls69; @ber90], $$\label{Diff}
D(E,z)=\frac{v^2}2\int_0^1d\mu\frac{1-\mu^2}{\nu_{\rm w}}\,,$$ is determined by diffusion of their pitch angle $\mu$. The latter is characterized by the effective frequency of CR scattering by MHD waves, $$\nu_{\rm w}(E,z,\mu)=2\pi^2\Omega_B(E)\frac{k_{\rm res}W(k_{\rm res},z)}{B^2}\,,$$ where $W(k,z)$ is the total spectral energy density of MHD waves, as discussed below, and $\Omega_B=(m_{\rm p}v/p)\Omega$ is the gyrofrequency of a proton, expressed via gyrofrequency scale $$\Omega=\frac{eB}{m_{\rm p}c}\,.$$ Wavenumber $k_{\rm res}$ at a given energy is related to $\mu$ by a condition of the first-harmonic cyclotron resonance, $$\label{k_res}
|\mu| vk_{\rm res}=\Omega_B\,,$$ or, equivalently, $|\mu| pk_{\rm res}=m_{\rm p}\Omega$. This condition assumes that $v$ is much larger than $v_{\rm A}$, which sets a lower bound of $\sim\frac12m_{\rm p}v_{\rm A}^2$ for the kinetic energy of CRs in our consideration.
To identify generic effects of self-generated turbulence in weakly ionized envelopes, we assume no other sources of turbulence and therefore no pre-existing MHD waves. The latter assumption is reasonable since, in the absence of internal sources, such waves in a typical envelope experience relatively strong damping and therefore can be neglected compared to the self-excited waves. The spectral energy density $W(k,z)$ for each wave mode is governed by a wave equation, including dominant processes of excitation, damping, transport, as well as of nonlinear wave interaction. We employ the following steady-state equation [@laga83; @norman96; @ptus06]: $$v_{\rm A}\frac{\partial W}{\partial z}+\frac{\partial }{\partial k}\left(\frac{kW}{T_{\rm NL}}\right)
=2(\gamma_{\rm CR}-\nu_{\rm damp})W.
\label{Wave_Eq}$$
A nonlinear interaction of waves, leading to their cascading to larger $k$, is described in Equation (\[Wave\_Eq\]) with the simplest phenomenological model characterized by the cascade timescale $T_{\rm NL}$ [@ptus06]. For the Iroshnikov-Kraichnan cascade[^3] [@irosh64; @kraich65] of acoustic MHD waves in an incompressible plasma, the timescale can be evaluated as the characteristic time of “collisions” between oppositely traveling wave packets, $\sim(kv_{\rm
A})^{-1}$, multiplied by the number of collisions needed to accumulate a large distortion of the packets, $\sim m_{\rm
i}n_{\rm i} v_{\rm A}^2/(kW)$ [@gold97]. This yields $$T_{\rm NL}^{-1}(k)=C_{\rm NL}\frac{k^2W(k)}{m_{\rm i}n_{\rm i} v_{\rm A}}\,,$$ where $C_{\rm NL}\sim 1$ is an unknown constant. We assume $T_{\rm NL}$ to be the same for the excited MHD modes [@gold97], and then Equation (\[Wave\_Eq\]) can be employed to describe the total spectral density of MHD waves.
The wave damping rate $\nu_{\rm damp}$ due to ion collisions with gas is proportional to the ratio $m_{\rm g}/m_{\rm i}$ of the mean mass of a gas particle to the mean ion mass, $$\nu_{\rm damp}\simeq\frac12\frac{m_{\rm g}}{m_{\rm i}}\nu_{\rm g}\,.$$ It is determined by the momentum-transfer cross section of ion-gas collisions (averaged over velocities), $\nu_{\rm
g}=\langle\sigma v\rangle_{\rm ig} n_{\rm g}$. We recall that waves can only be sustained when their frequency exceeds the damping rate, so for MHD waves the wavenumber should exceed the value of $\sim\nu_{\rm damp}/v_{\rm A}$ [@kuls69]. With the resonance condition (\[k\_res\]), this implies the upper limit on the energy of CRs that can contribute to the wave excitation, $E\lesssim eBv_{\rm A}/\nu_{\rm damp}$. For typical conditions in diffuse envelopes ($n_{\rm
g}\sim100$ cm$^{-3}$, $B\sim10-100~\mu$G) we obtain the energy limit $\sim1-100$ TeV. This limitation does not affect the results presented below, as the relevant energies turn out to be much smaller.
Finally, $\gamma_{\rm CR}$ is the (amplitude) growth rate of MHD waves excited by streaming CRs. These waves propagate along the magnetic field in the same direction as the CR flux (to the left in Figure \[fig1\]), and their growth rate is given by the following general formula, both for clockwise and counter-clockwise polarization [@went74; @skilb; @ber90]: $$\begin{aligned}
\gamma_{\rm CR}(k,z)=-\pi^3\frac{e^2v_{\rm A}}{c^2}\int_{-1}^{1}d\mu\:(1-\mu^2)\hspace{2.5cm}\label{gamma}\\
\times\int_0^{\infty}dp\:p^2v\:\delta(|\mu|pk-m_{\rm p}\Omega) \left(\frac{\partial f}{\partial \mu}-\frac{v_{\rm A}}{v}
p\frac{\partial f}{\partial p}\right),\nonumber\end{aligned}$$ where $v\gg v_{\rm A}$ is assumed. Here, $f(p,z,\mu)\equiv F(p,z)+\delta f(p,z,\mu)$ is the anisotropic distribution of CR in the momentum space, with $\langle\delta f\rangle_{\rm \mu}=0$, and $\delta(x)$ is the Dirac delta function. In the diffusive regime [*and*]{} for a weak anisotropy, $|\delta f|\ll F$, the combination of derivatives in Equation (\[gamma\]) is approximately equal to $-(v/\nu_{\rm w})\partial F/\partial z$ (the contribution of the gas losses is normally negligible here). Taking onto account Equation (\[Diff\]), we see that in this case $\gamma_{\rm CR}$ is determined by the diffusion part of the modulated CR flux. In Sections \[model\] and \[realistic\] we discuss mechanisms leading to the occurrence of gradients in the CR density.
Following @skilb, we introduce an effective cosine of the pitch angle, $\mu=\mu_*~(>0)$, in resonance condition (\[k\_res\]). This provides one-to-one relation between $k_{\rm res}$ and $E$, reducing Equation (\[k\_res\]) to $$\label{k_res1}
k_{\rm res}(E)=\frac{m_{\rm p}\Omega}{\mu_*p(E)}\,.$$ With this approximation, elemental integration in Equation (\[Diff\]) yields a simple expression for the diffusion coefficient, $$\label{Diff1}
D(E,z) \simeq \frac{1}{6\pi^2\mu_*}\frac{vB^2}{k^2W}\,,$$ with $k^2W$ evaluated for $k(E)$ from Equation (\[k\_res1\]). Similarly, by substituting $|\mu|=\mu_*$ in the delta-function in Equation (\[gamma\]) and performing the integration, we derive $$\label{gamma1}
\gamma_{\rm CR}(k,z) \simeq\pi^2 \frac{e^2v_{\rm A}}{m_{\rm p}c^2\Omega} pvD \frac{\partial N}{\partial z}\,,$$ where the (energy-dependent) rhs is evaluated for $E(k)$ from Equation (\[k\_res1\]). Thus, with approximation (\[k\_res1\]) the growth rate is exactly proportional to the diffusion part of the modulated flux. Equation (\[gamma1\]) remains applicable also in the free-streaming regime, after replacing $D\partial N/\partial z$ with difference $S_{\rm free}-v_{\rm A}N$.
It is noteworthy that, generally, from Equations (\[Diff\]) and (\[gamma\]) it follows that $D$ is a functional of $W^{-1}$, and $\gamma_{\rm CR}$ is a functional of $W^{-1}\partial N/\partial z$. Effectively, this implies dependence of $\mu_*$ on $k$, which can only be deduced by solving the resulting set of integral equations (\[Diff\_Eq\]) and (\[Wave\_Eq\]). However, this fact may only slightly change energy scalings of the results derived below with approximation (\[k\_res1\]), and therefore should not affect our principal conclusions.
Role of wave losses {#WL}
-------------------
In Equation (\[Diff\_Eq\]) we omitted wave losses – a term representing the conventional adiabatic contribution, proportional to the velocity gradient of MHD disturbances [see, e.g., @ber90]. After simple algebra, this term (to be added under the energy derivative on the rhs) can be written as $$\dot E_{\rm w}N= -\frac13\frac{du}{dz}pvN,$$ where $u=-v_{\rm A}$ is the velocity of the disturbances in the diffusive regime. We see that for our problem the adiabatic losses only operate at the border between the diffusion and the free-streaming zones, changing the CR flux by a value of $\sim v_{\rm A}N$, i.e., of the order of the advection part in Equation (\[smin\]). Thus, the wave losses merely lead to a renormalization of the advection.
In Sections \[model\] and \[realistic\] we demonstrate that the advection part of the modulated flux can usually be neglected for realistic conditions. Therefore, the wave losses are not expected to noticeably modify our results.
Dimensionless units and dependence on physical parameters {#normal}
=========================================================
To write governing equations (\[Diff\_Eq\]) and (\[Wave\_Eq\]) in a dimensionless form, we use the following normalization of $E$, $k$, and $p$: $$\label{Ekp}
\tilde E=\frac{E}{m_{\rm p}c^2}\,,\quad \tilde k=\mu_*\frac{ck}{\Omega}=\frac1{\tilde p}=\frac1{\sqrt{\tilde E(\tilde E+2)}}\,,$$ which naturally follows from Equations (\[p(E)\]) and (\[k\_res1\]). In some cases it is also practical to utilize the normalized physical velocity, $$\tilde v=\frac{\sqrt{\tilde E(\tilde E+2)}}{\tilde E+1}\,.$$ For brevity, we may use either of these variables to present results below.
Next, we introduce dimensionless CR spectrum, $$\tilde j=\frac{vN}{4\pi j_*}\,,$$ normalized by the characteristic value of the IS spectrum, $j_*=j_{\rm IS}(E=m_{\rm p}c^2)$. Now, in order to eliminate coefficients in CR flux (\[smin\]) for the diffusive regime and, simultaneously, in wave equation (\[Wave\_Eq\]), we introduce dimensionless wave energy density $\tilde W=W/W_*$ and coordinate $\tilde z=z/z_*$, normalized by $$W_*=\frac{2\pi^2\mu_*^2}{C_{\rm NL}}\frac{m_{\rm p}^2c^3v_{\rm A}j_*}{\Omega}$$ and $$\label{z*}
z_*=\frac{C_{\rm NL}}{3\pi^3\mu_*}\frac{m_{\rm i}n_{\rm i}}{m_{\rm p}^2\Omega j_*}\,.$$ Then Equations (\[Diff\_Eq\]) and (\[Wave\_Eq\]) are reduced to $$\label{Diff_norm}
\frac{\partial \tilde S}{\partial\tilde z}=-\frac{\partial}{\partial \tilde p}\left(\tilde L_{\rm g}\tilde j\right),$$ $$\label{Wave_norm}
\tilde k^{3/2}\frac{\partial}{\partial \tilde k}\left(\tilde k^{3/2}\tilde W\right)
=\frac{\tilde D}{2\tilde k}\frac{\partial \tilde j}{\partial \tilde z}-\nu,$$ where $\tilde L_{\rm g}$ and $\nu$ are dimensionless gas loss function and gas damping rate, respectively (both defined later in this Section), while $$\label{D_norm}
\tilde D=\frac{\tilde v\tilde p^2}{\tilde W}\,,$$ is the normalized diffusion coefficient. Dimensionless CR flux, $\tilde S=-\tilde v S/(4\pi j_*\epsilon)$, becomes $$\tilde S=\min\left\{\tilde D\frac{\partial \tilde j}{\partial \tilde z}+\tilde j,~\tilde S_{\rm free}\right\},
\label{smin_norm}$$ where the free-streaming term is $$\label{sfree_norm}
\tilde S_{\rm free}=\frac{\langle\mu\rangle}{\epsilon}\tilde v\tilde j.$$ With the used normalization, the flux of free-streaming CRs is inversely proportional to the small parameter $$\label{epsilon}
\epsilon=\frac{v_{\rm A}}{c}\,,$$ which is a measure of the contrast between the characteristic flux velocities in the two regimes (typically, $\epsilon\sim
10^{-3}-10^{-4}$). Note that in the transport equation (\[Wave\_norm\]) we dropped the term $\sim\epsilon \tilde
W^{-1}\partial \tilde W/\partial \tilde z$ representing advection: Based on results of Section \[approximate\], it is of the order of $\epsilon\nu$ and therefore is negligible compared to the rhs.
The gas losses can be conveniently expressed in terms of the loss function $L_{\rm g}(E)=-\dot E_{\rm g}/n_{\rm g}v$, which is a universal function of energy only (for a given gas composition). In the normalized form, it is $$\label{tilde_L}
\tilde L_{\rm g}=\frac1{\epsilon}\frac{n_{\rm g}z_*L_{\rm g}}{ m_{\rm p}c^2}\,.$$ In the free-streaming regime, where $W\simeq0$, the small parameter $\epsilon$ cancels out in Equation (\[Diff\_norm\]) and CR transport naturally becomes independent of $v_{\rm A}$. Upon transition to the diffusive regime, the effective loss rate is increased by a factor of $\epsilon^{-1}$, reflecting the corresponding increase of the distance traversed by self-trapped CRs.
Thus, with the used normalization, the [*only*]{} dimensionless number entering governing Equations (\[Diff\_norm\]) and (\[Wave\_norm\]) (for a given loss function $L_{\rm g}$) is the damping rate $$\label{nu}
\nu=\frac{3\pi\mu_*}{4C_{\rm NL}}\frac{m_{\rm g}z_*\nu_{\rm g}}{m_{\rm i}c}\,,$$ while the small parameter $\epsilon$ characterizes a transition between the diffusive and free-streaming regimes.[^4]
The scaling dependence of $\nu$ and $\epsilon$ on the physical parameters is given by the following general expressions: $$\begin{aligned}
\nu& = 8.7 \left(\frac{m_{\rm g}/m_{\rm p}}{2.3}\right)
\left(\frac{j_* m_{\rm p} c^2}{1.3~\mbox{cm}^{-2}\mbox{s}^{-1}\mbox{sr}^{-1}}\right)^{-1} \label{scale_nu}\\
&\times\left(\frac{n_{\rm i}/n_{\rm g}}{3\times 10^{-4}}\right)\left(\frac{n_{\rm g}}{100~\mbox{cm}^{-3}}\right)^2
\left(\frac{B}{0.1~\mbox{mG}}\right)^{-1},\nonumber \\
\epsilon&= 1.2\times10^{-3} \left(\frac{m_{\rm i}/m_{\rm p}}{12}\right)^{-1/2}\label{scale_epsilon}\\
&\times \left(\frac{n_{\rm i}/n_{\rm g}}{3\times 10^{-4}}\right)^{-1/2}\left(\frac{n_{\rm g}}{100~\mbox{cm}^{-3}}\right)^{-1/2}
\left(\frac{B}{0.1~\mbox{mG}}\right).\nonumber\end{aligned}$$ To give results in absolute units, we also use the normalization length, $$\begin{aligned}
z_*&= 2.8 \times 10^{18}\frac{C_{\rm NL}}{\mu_*}\left(\frac{m_{\rm i}/m_{\rm p}}{12}\right)
\left(\frac{j_* m_{\rm p} c^2}{1.3~\mbox{cm}^{-2}\mbox{s}^{-1}\mbox{sr}^{-1}}\right)^{-1} \nonumber \\
&\times\left(\frac{n_{\rm i}/n_{\rm g}}{3\times 10^{-4}}\right)\left(\frac{n_{\rm g}}{100~\mbox{cm}^{-3}}\right)
\left(\frac{B}{0.1~\mbox{mG}}\right)^{-1}~\mbox{cm}. \nonumber\end{aligned}$$
The illustrative numerical results presented in Sections \[model\] and \[realistic\] are obtained by varying density of gas $n_{\rm g}$. For simplicity, it is assumed that hydrogen is in molecular form and carbon photoionization by IS radiation field is the main source of charged species [see, e.g., @oka06]. Hence, $m_{\rm g}/m_{\rm p}\simeq2.3$, $m_{\rm
i}/m_{\rm p}=12$, and $n_{\rm i}/n_{\rm g}\simeq 4\times 10^{-4}$, adopting the solar chemical composition with ionized carbon. The magnetic field is set to $B = 100~\mu$G, in order to increase the magnitude of $\epsilon$ (which improves convergence of the numerical scheme). For the ion-gas collisions we use $\langle\sigma v\rangle_{\rm ig} \simeq 2.1 \times
10^{-9}$ cm$^3$/s, corresponding to molecular hydrogen at a temperature of 100 K [see, e.g., @kuls69]. Finally, we set $C_{\rm NL}=\mu_*=1$ and employ the following model spectrum for interstellar CRs [@ivlev15a]: $$\label{IS_spectrum}
j_{\rm IS}(E)=\frac{1.4\times10^{-9}\tilde E^{-0.8}}{(0.55+\tilde E)^{1.9}}~~{\rm eV^{-1}cm^{-2}s^{-1}sr^{-1}}.$$ With these physical parameters, $\nu$ and $\epsilon$ are related via $$\epsilon\nu^{1/4} = 1.7\times 10^{-3},$$ and below we only indicate the value of $\nu$.
In Appendix \[numerical\] we describe the algorithm to solve Equations (\[Diff\_norm\]) and (\[Wave\_norm\]) numerically, and also give the gas loss function $L_{\rm g}(E)$ used to obtain numerical results presented in Section \[realistic\].
A model problem: Absorbing wall {#model}
===============================
We start with an idealized problem setup sketched in Figure \[fig1\], and consider propagation of CRs toward an “absorbing wall” (which mimics a dense interior of a molecular cloud). The CR flux generates MHD turbulence upstream from the wall (located at $z=0$), implying diffusive regime for CR propagation. Therefore, one can set $N(E,0)=0$ as the standard boundary condition for the diffusion equation at an absorbing wall.[^5] At the outer envelope boundary (located at $z=H$) the CR density is given by the interstellar value, $N(E,H)=N_{\rm IS}(E)$. The principal aim of this simplified consideration is to identify generic properties of nonlinear CR propagation, self-consistently described by the transport and wave equations discussed above.
We start with a case where the gas losses are unimportant, so the rhs of Equation (\[Diff\_norm\]) can be set equal to zero. Then the transport equation in the diffusive regime has a straightforward solution, $$\label{j_diff}
\frac{j(E,z)}{j_{\rm IS}(E)}=\frac{N(E,z)}{N_{\rm IS}(E)}=\frac{1-e^{-\eta(E,z)}}{1-e^{-\eta(E,H)}}\,,$$ determined by “diffusion depth” $$\label{zeta}
\eta(E,z)= \int_0^{\tilde z}\frac{dx}{\tilde D(E,x)}\equiv v_{\rm A}\int_0^z\frac{dx}{D(E,x)}\,.$$ The magnitude of the resulting modulated flux (\[smin\]) is $$\label{flux}
S(E)=\frac{v_{\rm A}N_{\rm IS}(E)}{1-e^{-\eta(E,H)}}\,,$$ (hereafter, we omit the minus sign in front of $S$). By virtue of Equation (\[Ekp\]) the solution can also be presented as a function of $k$. One can see that $\eta$ is a measure of the relative importance of diffusion and advection in the modulated CR flux: For $\eta\ll1$ Equation (\[j\_diff\]) is reduced to the solution of the standard diffusion equation ($v_{\rm A}$ cancels out), for $\eta\gg1$ the CR density becomes constant and flux (\[flux\]) saturates at $v_{\rm
A}N_{\rm IS}$.
Below we show that the diffusive regime for given $E$ does not necessarily extend up to the outer envelope boundary, but may terminate at the [*outer border*]{} of the diffusion zone $z_0(E)<H$, where $W\to0$. In this case, the free-streaming regime with $N(E,z)=N_{\rm IS}(E)$ operates at $z>z_0$, and the solution does not depend on $H$.
By substituting Equation (\[j\_diff\]) in Equation (\[Wave\_norm\]) we derive the following wave equation for self-consistent turbulent field in the diffusive regime: $$\label{W_SC}
\tilde k^{3/2}\frac{\partial}{\partial \tilde k}\left(\tilde k^{3/2} \tilde W\right)
=\frac{\tilde j_{\rm IS}(k)}{2\tilde k}\frac{e^{-\eta(k,z)}}{1-e^{-\eta_0(k)}}- \nu,$$ where $\eta(k,z)$ is given by Equation (\[zeta\]) with $E(k)$ from Equation (\[Ekp\]), $$\eta(k,z)=\tilde k^2\sqrt{1+\tilde k^2}\int_0^{\tilde z}dx\:\tilde W(k,x),$$ and $\eta_0(k)=\eta(k,z_0)$. We recall that the excitation term in Equation (\[W\_SC\]) is proportional to the diffusion part of the modulated flux which, in turn, cannot exceed the flux of free streaming CRs. Then from Equations (\[smin\_norm\]) and (\[sfree\_norm\]) it follows that in the diffusive regime, with $j(E,z)$ from Equation (\[j\_diff\]), condition $\eta_0\gtrsim v_{\rm A}/v$ must always be fulfilled. This lower bound of $\eta_0$ (which is a small number, since $v\gg v_{\rm A}$ is assumed) represents the necessary condition of applicability for the diffusion approximation.
We notice that requirement $$\label{zeta_min}
\eta\gtrsim \frac{v_{\rm A}}{v}$$ coincides with the condition that the mean free path of CRs, $\sim D/v$, is smaller than the inhomogeneity scale length, $\sim N/|\partial N/\partial z|$, as one can easily derive from Equations (\[j\_diff\]) and (\[zeta\]); simultaneously, this ensures that the velocity of the CR flux does not exceed the physical velocity. Therefore, we shall consider inequality (\[zeta\_min\]) as the sufficient condition of applicability of the diffusion approach. The resulting [*inner border*]{} of the diffusion zone $z_{\rm min}(E)$ is determined from condition $\eta(E,z_{\rm min})\sim v_{\rm A}/v$.
The threshold energy $E_{\rm ex}$, below which CRs excite waves, can be readily derived from the balance of the growth rate in the [*free-streaming*]{} regime and the damping rate. By replacing the diffusion flux on the rhs of Equation (\[Wave\_norm\]) with the free-streaming expression from Equation (\[sfree\_norm\]), we obtain the following equation: $$\label{threshold}
\frac{\tilde E_{\rm ex}+2}{\tilde E_{\rm ex}+1}\tilde E_{\rm ex}\tilde j_{\rm IS}(E_{\rm ex})
=\frac{2\epsilon \nu}{\langle\mu\rangle}\,,$$ where $\langle\mu\rangle$ is the average pitch angle in the free-streaming zone I (see Appendix \[<mu>\] and Figure \[figA1\] therein). For sufficiently steep, monotonic energy spectra, e.g., $\tilde j_{\rm IS}= \tilde E^{-\alpha}$ with $\alpha>1$, waves are excited if $E<E_{\rm ex}$; the threshold energy scales as $$E_{\rm ex}\propto\left(\frac{m_{\rm g}n_{\rm g}}{j_*}\sqrt{\frac{n_{\rm i}}{m_{\rm i}}}\:\right)^{-\frac1{\alpha-1}}.$$ Equation (\[threshold\]) also shows that CRs with $j_{\rm IS}\propto E^{-1}$ represent a [*critical case*]{}, where the excitation occurs when the flux magnitude matches the damping threshold.
Numerical analysis shows that the magnitude of $W$ in the turbulent zone is typically high enough for the condition of the diffusion approximation to be well fulfilled. Thus, it is reasonable to solve wave equation (\[W\_SC\]) for $k>k_{\rm
ex}\equiv k(E_{\rm ex})$ with condition $W(k_{\rm ex},z)=0$. The solution in $(k,z)$ space is applicable for $\eta(k,z)\gtrsim v_{\rm A}/v$, while outer turbulent border $z_0(k)$ is obtained from $W(k,z_0)=0$.
Approximate solution {#approximate}
--------------------
One can obtain a simple approximate solution of Equation (\[W\_SC\]), providing a fairly accurate and general description of the turbulent regime. From the numerical integration performed for different values of $\nu$ we found that, as long as $\eta_0\lesssim1$ and $\nu$ is not too small, the turbulent field can be reasonably approximated by a decreasing linear function of coordinate (see Appendix \[linear\_W\] and the figure therein), $$\label{field}
\tilde W(k,z)\simeq w(k)+w'(k)\tilde z,$$ with $w'<0$, so the outer border of the diffusion zone is $\tilde z_0(k)=-w(k)/w'(k)$. Equation (\[field\]) breaks down close to $k_{\rm ex}$, but this does not affect properties of the whole diffusion zone.
We first study the case of small diffusion depth, $\eta_0\lesssim1$, which allows us to expand the exponentials on the rhs of Equation (\[W\_SC\]). We retain only linear terms in the resulting $z$-polynomial and equate to zero the corresponding coefficients, which gives us two equations for $w(k)$ and $w'(k)$. One equation yields $$\label{zeta_0}
\eta_0(k)=\frac{\tilde j_{\rm IS}(k)}{2\tilde k\nu}\,,$$ which is simply a balance of the excitation and damping on the rhs of Equation (\[W\_SC\]), written for small $\eta$; the lhs, i.e., the cascade term for $w(k)$, is neglected here compared to $\nu$ – this assumption is confirmed [*a posteriori*]{}. The other equation leads to $$\frac{d}{d\tilde k}\left(\tilde k^{3/2}w'\right)=\nu \sqrt{\tilde k(1+\tilde k^2)}\:w'\tilde z_0\,,$$ showing that the cascade is essential for $w'(k)$. By combining Equation (\[zeta\_0\]) with relation $\eta_0(k)=
-\frac12\tilde k^2 \sqrt{1+\tilde k^2}\:w'\tilde z_0^2$ and setting $w'(k_{\rm ex})=0$, we get the solution which can be conveniently written as $$\label{w'}
2\sqrt{\frac{-\tilde k^{3/2}w'(k)}{\nu}}=\int_{\tilde k_{\rm ex}}^{\tilde k}dx\sqrt{\frac{\sqrt{1+x^2}\:
\tilde j_{\rm IS}(x)}{x^{7/2}}}\,.$$ Then $\tilde z_0(k)$ is readily obtained by employing the above relation for $\eta_0(k)$, and $w(k)=-\tilde z_0(k)w'(k)$. We note that a realistic IS spectrum, such as Equation (\[IS\_spectrum\]), is a rather steeply increasing (decreasing) function at small $\tilde k$ (large $\tilde E$). Therefore, if $\tilde k_{\rm ex}\lesssim1$, the integral in Equation (\[w’\]) is dominated by larger $k$, i.e., the contribution of $k\simeq k_{\rm ex}$ vanishes asymptotically.
With this solution we can verify the simplifications/assumptions made to obtain it: First, we recall that the advection term $\sim\epsilon \tilde W^{-1}\partial \tilde W/\partial \tilde z$ was dropped in Equation (\[Wave\_norm\]). For $k\gg k_{\rm
ex}$ we get $\epsilon\tilde W^{-1}|\partial \tilde W/\partial \tilde z|\simeq\epsilon|w'|/w\sim\epsilon\nu\sqrt{1+\tilde
k^2}$, which is indeed small compared to $\nu$. Second, by substituting solution $w(k)\sim\tilde j_{\rm IS}(k)/\tilde k^3$ to the cascade term in the lhs of Equation (\[W\_SC\]) we conclude that the latter is small compared to $\nu$ too, as long as $\eta_0\lesssim1$.
Condition $\eta_0\lesssim1$ implies a certain upper limit on $k$, since $\eta_0(k)$ is an increasing function (for realistic IS spectra). For larger $\eta_0$ (and $k$), numerical results indicate that spatial nonlinearity of the turbulent field becomes significant (see Appendix \[linear\_W\]). Nevertheless, Equation (\[field\]) still provides useful qualitative description of the diffusion zone. For $\eta_0\gg1$, term $e^{-\eta_0}$ in Equation (\[W\_SC\]) can be neglected. In this case, to determine $w(k)$ and $w'(k)$ we write the resulting wave equation for $z=0$ and $z=z_0$. The former gives $$\label{w}
\tilde k^{3/2}\frac{d}{d\tilde k}\left(\tilde k^{3/2}w\right)=\frac{\tilde j_{\rm IS}(k)}{2\tilde k}-\nu,$$ showing that excitation exceeds damping at larger $k$, so that now the cascade plays a crucial role. In the latter equation, we neglect the term $\propto e^{-\eta_0}$ and, after simple transformation, obtain the following equation for $z_0(k)$: $$\label{z_0}
\frac{d\ln z_0}{d\tilde k}=-\frac{\nu}{\tilde k^3w(k)}\,.$$ Equation (\[w\]) allows straightforward integration for given $j_{\rm IS}(k)$, and the derived $w(k)$ has to be matched with that obtained from Equations (\[w’\]). By substituting the result in Equation (\[z\_0\]) and integrating it, we get $z_0(k)$ for large $\eta_0$.
Diffusion zone {#zone}
--------------
Figure \[fig2\] illustrates the characteristic form of the diffusion zone in $(E,z)$ plane. The numerically calculated diffusion border is plotted for several values of $\nu$ (solid lines). The right branch of each contour is the outer border of the zone $z_0(E)$, approximately derived in Section \[approximate\], while the left branch corresponds to inner border $z_{\rm min}(E)$, determined by condition (\[zeta\_min\]). The branches cross at the highest “critical” point $E\simeq
E_{\rm ex}(\nu)$, determined by Equation (\[threshold\]). The analytical curves $z_0(E)$ and $z_{\rm min}(E)$, obtained from solution (\[w’\]) (dotted lines), demonstrate a good overall agreement with the numerical results. A stronger deviation is observed toward the critical point, where the approximate solution breaks down. Also, at lower energies analytical $z_0(E)$ deviates increasingly from the numerical curve when $\nu$ is small.
Using solution (\[w’\]), one can deduce how the shape of the diffusion zone depends on the form of the IS spectrum and the main physical parameters. For $\tilde j_{\rm IS}(E)=\tilde E^{-\alpha}$ with $\alpha(E)$ determined by a model spectrum, Equation (\[IS\_spectrum\]) or analogous [@ivlev15a], it is practical to consider two limiting cases – the ultra-relativistic limit, where $\tilde k=1/\tilde E\ll1$, and the non-relativistic case, where $\tilde k=1/\sqrt{2\tilde
E}\gg1$. Equation (\[w’\]) yields the outer border, $\tilde z_0(E)\sim1/\nu$ for $\tilde E\gg1$ and $\tilde
z_0(E)\sim\sqrt{\tilde E}/\nu$ for $\tilde E\ll1$. Substituting a solution for $w(k)$ in condition $\eta(E,z_{\rm
min})\sim\epsilon/\tilde v$, we obtain the inner border, $\tilde z_{\rm min}(E)\sim\epsilon \tilde E^{\alpha-1}$ for $\tilde
E\gg1$ and $\tilde z_{\rm min}(E)\sim\epsilon \tilde E^{\alpha-1/2}$ for $\tilde E\ll1$. In absolute units, this gives the following dependence on the physical parameters: $$\label{scaling_z1}
z_{\rm min}\propto \frac{\sqrt{m_{\rm i}n_{\rm i}}}{j_*}\,,\quad z_0\propto \frac{m_{\rm i}}{n_{\rm g}}\,.$$
If $\eta_0\gtrsim1$, which corresponds to large $k$ and/or small $\nu$, solution (\[w’\]) is no longer applicable and the turbulent field is qualitatively described by Equations (\[w\]) and (\[z\_0\]). The former yields $\tilde k^3
w(k)\sim\tilde j_{\rm IS}(k)$ for large $k$, and then from the latter equation we invoke that $z_0(k)$ asymptotically tends to a constant value. This explains the behavior of numerically calculated $z_0(E)$ at lower $E$ and small $\nu$, seen in Figure \[fig2\] for $\nu=0.3$ and 3.5.
The diffusion zone is formed when $z_{\rm min}(E)\lesssim z_0(E)$. Using the above estimates for the inner and outer borders, we then arrive to a simple criterion of the diffusive regime, valid for all energies where $\eta_0(E)\lesssim1$: $$\label{zone_eq}
\epsilon\nu \tilde E^{\alpha-1}\lesssim1.$$ Expectedly, this criterion is essentially equivalent to the excitation criterion (\[threshold\]) in the free-streaming regime. Equation (\[zone\_eq\]) shows that if $\alpha>1$ for any $E$, the diffusion zone shrinks monotonically with $\nu$ toward lower energies, until the basic resonance condition (\[k\_res\]) becomes inapplicable at $v\lesssim v_{\rm A}$. Current models of the IS spectra, such as Equation (\[IS\_spectrum\]), suggest $\alpha<1$ for non-relativistic CRs. Then the diffusion zone for sufficiently large $\nu$ becomes an isolated “island”, and eventually disappears when product $\epsilon\nu$ exceeds a certain maximum value $(\epsilon\nu)_{\rm max}\sim1$. The exact value of $(\epsilon\nu)_{\rm max}$ is derived from Equation (\[threshold\]) and corresponds to the maximum of its lhs; e.g., for IS spectrum (\[IS\_spectrum\]) the maximum is at $E\simeq60$ MeV, and $(\epsilon\nu)_{\rm max}\sim1$. Then from Equations (\[scale\_nu\]) and (\[scale\_epsilon\]) we obtain the maximum gas density $n_{\rm g}\sim3\times10^4$ cm$^{-3}$, above which no turbulence can be excited by CRs with such energy spectrum.[^6] In Figure \[fig2\], the diffusion zone completely disappears at $\nu\sim3\times10^3$.
Figure \[fig2\] also indicates that, for very small $\nu$, the derived outer border $z_0(E)$ at higher energies may be larger than the envelope size $H$. Then the diffusion zone is bound between $z_{\rm min}(E)$ and $H$, and the solution obtained in Section \[approximate\] for $W(k,z)$ is modified. Nevertheless, as long as the resulting $\eta(k,H)\equiv\eta_H$ is small, its value is determined from the same excitation-damping balance that leads to Equation (\[zeta\_0\]), and therefore $\eta_H$ is equal to the derived $\eta_0$. In this case, the condition of the diffusion regime to operate is simply $z_{\rm min}(E)\lesssim H$.
CR flux {#CR_flux}
-------
>From Equation (\[flux\]) it follows that the value of diffusion depth $\eta_0$ (or $\eta_H$) completely determines the CR flux penetrating the cloud. Figure \[fig3\] illustrates dependence $\eta_0(E;\nu)$. For $\eta_0\lesssim1$ it is well described by Equation (\[zeta\_0\]) with subtracted “inner border” value $\epsilon/\tilde v$, as determined by condition (\[zeta\_min\]). For large $\eta_0$, the exact dependence becomes unimportant for calculating $S(E)$, since the exponential in Equation (\[flux\]) can be safely neglected.
Let us summarize the behavior of $S(E)$. At sufficiently high energies, the CR flux is not affected by turbulence and equal to the free-streaming value, $$\label{flux0}
E>E_{\rm ex}:\qquad S_{\rm free}(E)=4\pi\langle\mu\rangle j_{\rm IS}(E).$$ A continuous transition to the modulated flux occurs at $E=E_{\rm ex}(\nu)$, determined by Equation (\[threshold\]). For smaller $E$, from Equations (\[flux\]) and (\[zeta\_0\]) we obtain the following general formula: $$\label{flux_final}
E<E_{\rm ex}:\qquad \frac{S(E)}{4\pi j_*}=\frac{\tilde E+1}{\sqrt{\tilde E(\tilde E+2)}}
\left(\frac{\epsilon \tilde j_{\rm IS}(E)}{1-e^{-\eta_0(E)}}\right),$$ with diffusion depth $$\label{zeta_00}
\eta_0(E)=\sqrt{\tilde E(\tilde E+2)}\:\frac{\tilde j_{\rm IS}(E)}{2\nu}\,.$$
For $\eta_0\lesssim1$, where the exponential in the denominator of Equation (\[flux\_final\]) can be expanded, the resulting leading term does not depend on $j_{\rm IS}(E)$. In this case we obtain “diffusion-dominated” flux, $$\label{flux1}
%\eta_0\lesssim1:\qquad
S_{\rm DD}(E)=\frac{\tilde E+1}{\tilde E+2}\left(\frac{8\pi\epsilon\nu j_*}{\tilde E}\right),$$ where advection is unimportant and therefore its magnitude is governed by a balance of the excitation and damping in wave equation (\[W\_SC\]). This is the reason why it obeys a [*universal*]{} energy dependence, scaling as $\propto E^{-1}$ both in the non-relativistic and ultra-relativistic limits (or, equivalently, as $\propto(pv)^{-1}$). Furthermore, from Equations (\[scale\_nu\]) and (\[scale\_epsilon\]) it follows that $$\label{flux_par}
S_{\rm DD}\propto m_{\rm g}n_{\rm g}\sqrt{\frac{n_{\rm i}}{m_{\rm i}}}\,,$$ i.e., the flux does not depend on $j_*$ and thus is [*solely*]{} determined by the physical parameters of the envelope. We want to emphasize that this expression can be deduced from a theoretical analysis by @skill76, by substituting their Equation (6) into the second term of their Equation (8).
At even lower energies, $\eta_0$ exceeds unity for smaller $\nu$, as evident from Figure \[fig3\]. Then advection dominates and the flux tends to $v_{\rm A}N_{\rm IS}(E)$, which is $$\label{flux2}
%\eta_0\gtrsim1:\qquad
S_{\rm AD}(E)= \frac{\tilde E+1}{\sqrt{\tilde E(\tilde E+2)}}\:4\pi\epsilon j_{\rm IS}(E).$$ The analysis performed by @gabici15 corresponds to our case $\eta_0\sim1$, and therefore their conclusion that the velocity of the CR flux penetrating into a cloud is of the order of $v_{\rm A}$ represents a crossover to the advection-dominated flux.
Figure \[fig4\] shows the modulated CR flux obtained analytically, from Equation (\[flux\_final\]) for IS spectrum (\[IS\_spectrum\]), and compared with the numerically calculated flux. One can see that the analytical results provide a fairly accurate description of $S(E)$ in the whole energy range; only for very small $\nu$, a slight deviation (about 50%) is observed at intermediate energies, where $\eta_0(E)\sim1$ (as one can see from Figure \[fig3\]).
Both panels of the figure clearly demonstrate a transition from free streaming to the diffusive regime, occurring at $E=E_{\rm ex}(\nu)$ and manifested by a kink at each curve. In the left panel the curves are normalized by $S_{\rm DD}(E)$ and, hence, at $E<E_{\rm ex}$ they collapse into the horizontal line at the unity level as long as $\eta_0(E)\lesssim1$ (for $E>E_{\rm ex}$ they approximately scale as $\propto
E\tilde j_{\rm IS}(E)/\nu^{3/4}$). In the right panel $S(E)$ is normalized by $S_{\rm AD}(E)$, and thus a crossover to the advection-dominated flux occurs if the curves approach the unity level (for $E>E_{\rm ex}$ the curves tend to $\epsilon^{-1}\propto\nu^{1/4}$). The crossover takes place only for small $\nu$, otherwise the flux remains diffusion-dominated at all energies shown.
We point out that Equation (\[flux\_final\]) is insensitive to the particular model of nonlinear wave cascade. As shown in Section \[approximate\], the cascade term in Equation (\[W\_SC\]) is negligible for small $\eta_0$ (where $S\simeq S_{\rm
DD}$), whereas for large $\eta_0$ the CR flux tends to the advection asymptote $v_{\rm A}N_{\rm IS}$, i.e., the cascade term may affect the flux only near the crossover point $\eta_0(E)\sim1$. This has been verified with numerical calculations performed for the Kolmogorov cascade [with $T_{\rm NL}$ taken from @ptus06], indeed showing minor deviations from the presented results in the crossover energy range.
Effect of energy losses {#realistic}
=======================
In the previous Section we derived intrinsic properties of the turbulent diffusion zone generated under idealized conditions, where CRs propagate toward an absorbing wall, and the energy losses due to interaction with gas are unimportant. This approach presumes the intrinsic spatial scale of the diffusion zone, $z_0(E)$, to be much smaller than the CR loss length at a given energy. For realistic parameters of diffuse envelopes, the latter assumption is not always justified, especially in the non-relativistic case.
For this reason, let us now move away from the initial assumption that CRs propagate freely through the envelope until they reach the turbulent zone near the absorbing wall, to see what impact the gas losses may have on the diffusion and, most importantly, how the flux self-modulation is affected by the losses.
The principal difference introduced to the problem by the gas losses is that the CR flux is no longer conserved, as follows from Equation (\[Diff\_norm\]). Therefore, the losses naturally generate a CR density gradient and, hence, stimulate wave excitation across the whole envelope, starting from its outer boundary (whereas before the gradient was only present near the absorbing wall). For this reason it is more convenient to analyze results in the frame of reference where $z=0$ is located at the outer boundary, as shown in Figure \[fig5\]. Thus, now $|z_0|$ is referred to as the inner (“downstream”) border of the diffusion zone and $|z_{\rm min}|~(<|z_0|)$ is the outer (“upstream”) border.
Solution for the excitation-damping balance {#inhomogeneous}
-------------------------------------------
The general excitation criterion (\[threshold\]) does not depend on a particular problem setup and, hence, can also be used when the losses are present. Turbulence sets in (and, as pointed out in Section \[model\], the diffusive approximation is thereby justified) when the excitation term on the rhs of Equation (\[Wave\_norm\]) becomes equal to damping. Furthermore, the role of the cascade term on the lhs remains largely negligible at $k\gtrsim k_{\rm ex}$: as we demonstrate below in this Section, the condition of applicability of the excitation-damping balance is relaxed compared to the loss-free case (where the cascade term can be neglected for $\eta_0\lesssim1$). Therefore, from Equation (\[Wave\_norm\]) we obtain $$\label{Wave_norm2}
\frac{\tilde D}{2\tilde k}\frac{\partial \tilde j}{\partial \tilde z}\simeq\nu.$$ We see that $\tilde D\partial \tilde j/\partial \tilde z$, the diffusion part of flux (\[smin\_norm\]), does not depend on coordinates (for given $\nu$) and therefore does not contribute to transport equation (\[Diff\_norm\]). The latter is then reduced to $$\label{Diff_norm2}
\frac{\partial \tilde j}{\partial\tilde z}=-\frac{\partial}{\partial \tilde p}\left(\tilde L_{\rm g}\tilde j\right),$$ giving the local CR spectrum, i.e., the advection part of flux (\[smin\_norm\]).
Equation (\[Diff\_norm2\]) has a general solution in $(p,z)$ space, $$\label{solution_gen}
\tilde L_{\rm g}(p)\tilde j(p,z)=\Phi\left(\tilde z-\int \frac{d\tilde p}{\tilde L_{\rm g}(\tilde p)}\right),$$ where function $\Phi(x)$ is determined by the boundary condition $\tilde j(p,0)=\tilde j_{\rm IS}(p)$. To illustrate the overall behavior and obtain useful closed-form expressions, let us again consider a power-law IS energy spectrum, $\tilde
j_{\rm IS}(E)=\tilde E^{-\alpha}$, and treat separately the non-relativistic and ultra-relativistic cases.
For $\tilde E\lesssim1$ the gas losses are dominated by ionization [@hayakawa]. The loss function can be approximated by $\tilde L_{\rm g}(E)\simeq A_{\rm ion}\tilde E^{-b}$, with the exponent in the range of $0\lesssim
b\lesssim1$. The solution resulting from Equation (\[solution\_gen\]) is $$\label{solution1}
\tilde j(E,z)=\tilde j_{\rm IS}(E)\left(1+(2b+1)\frac{\tilde L_{\rm g}(E)}{\sqrt{2\tilde E}}|\tilde z|\right)
^{-\frac{2(b+\alpha)}{2b+1}}.$$ The standard expression for non-relativistic ionization losses with $b=1$ is determined by [@ginz79] $$A_{\rm ion}=\frac38\frac{m_{\rm e}}{m_{\rm p}}\frac{n_{\rm g}z_*\sigma_{\rm T}\ln\Lambda}{\epsilon}\,,$$ where $\Lambda$ is the argument of the Coulomb logarithm for the ionization losses (for hydrogen, $\Lambda\simeq20$), $\sigma_{\rm T}=6.6\times10^{-25}$ cm$^{-2}$ is the Thomson cross section of electron, and $m_{\rm e}/m_{\rm p}=1/1836$ is the electron-to-proton mass ratio.
In the relativistic case, the pion production occurring in proton-proton collisions above the threshold energy of $\simeq280$ MeV is the main mechanism for the energy losses [@hayakawa]. The loss function can be approximated by $\tilde L_{\rm g}(E)\simeq A_{\pi}\tilde E$, where [@mann94] $$A_{\pi}=0.65\frac{n_{\rm g}z_*\sigma_{\pi}}{\epsilon}$$ is proportional to the effective cross section $\sigma_{\pi}\simeq3\times10^{-26}$ cm$^{-2}$ (neglecting a weak logarithmic energy dependence). Then Equation (\[solution\_gen\]) yields $$\label{solution2}
\tilde j(E,z)=\tilde j_{\rm IS}(E)e^{-(\alpha-1)A_{\pi}|\tilde z|}\,.$$ The derived results also allow us to verify the (initially assumed) excitation-damping balance, Equation (\[Wave\_norm2\]), i.e., to identify conditions when the cascade term in Equation (\[Wave\_norm\]) is negligible: Since the relative contribution of the cascade term increases with $k$ (i.e., with decreasing $E$), it is sufficient to consider the non-relativistic case. Substituting Equations (\[solution1\]) in Equation (\[Wave\_norm2\]) and taking into account Equation (\[D\_norm\]) gives an estimate for $W(k)$, to be inserted in the lhs of Equation (\[Wave\_norm\]). We obtain that the latter is small compared to $\nu$ when $(A_{\rm ion}/\nu^2)\tilde E^{-(\alpha+1/2)}\lesssim1$, which can be equivalently rewritten as $\eta_0\lesssim\tilde E\nu/A_{\rm ion}$ with $\eta_0$ from Equation (\[zeta\_0\]). Comparing this with condition $\eta_0\lesssim1$ for the loss-free case, we conclude that for $\tilde E\gtrsim A_{\rm ion}/\nu$ ($\simeq10^{-4}\nu^{1/4}$ for the presented results, i.e., for all energies shown) the excitation-damping balance is indeed more easily satisfied in the presence of losses.
Onset of diffusion zone {#zone_onset}
-----------------------
A condition of applicability of the diffusive regime is that the CR mean free path, $\sim D/v$, is smaller than the characteristic spatial scale. In dimensionless form, the mean free path $\sim\epsilon\tilde D/ \tilde v$ should be smaller than the relevant scale of the present problem, $\sim|\tilde z|$. By employing Equation (\[Wave\_norm2\]), the condition is reduced to $$\label{condition}
\frac{\tilde E+2}{\tilde E+1}\tilde E|\tilde z|\frac{\partial \tilde j}{\partial \tilde z}\gtrsim2\epsilon\nu,$$ where $\tilde j(E,z)$ is a solution of Equation (\[Diff\_norm2\]).
Equation (\[condition\]) is the [*necessary*]{} condition of applicability of the diffusive regime in the presence of losses. For given $E$, its lhs is a function of $z$, whose maximum is of the order of $\sim\tilde E\tilde j_{\rm IS}(\tilde
E)$. Hence, for $\tilde j_{\rm IS}(E)=\tilde E^{-\alpha}$ condition (\[condition\]) essentially coincides with criterion (\[zone\_eq\]) of the diffusive regime, derived for the absorbing wall problem.
The sufficient applicability condition requires that the diffusion zone is formed within the envelope, i.e., that the outer border $|z_{\rm min}(E)|$ at which inequality (\[condition\]) is first fulfilled is smaller than the envelope size $H$. For the loss mechanisms discussed in Section \[inhomogeneous\], we have $$\begin{aligned}
\tilde E\lesssim1:&\qquad |\tilde z_{\rm min}|\sim\frac{\epsilon\nu}{(\alpha+1)A_{\rm ion}}\tilde E^{\alpha+1/2}\label{z_min1}\\
\intertext{and}
\tilde E\gtrsim1:&\qquad |\tilde z_{\rm min}|\sim\frac{\epsilon\nu}{(\alpha-1)A_{\pi}}\tilde E^{\alpha-1}.\label{z_min2}\end{aligned}$$ Since $A_{\rm ion}/A_{\pi}\simeq7\times10^{-3}\ln\Lambda$ is practically a constant $\sim0.1$, a smooth crossover between the two cases occurs at energy about a few tenths of GeV. With Equation (\[z\*\]) we notice that in absolute units, $$\label{scaling_z2}
|z_{\rm min}|\propto \frac{B}{j_*m_{\rm i}}\,,$$ the coordinate of the diffusion onset is proportional to $B$ and does not depend on $n_{\rm g}$ or $n_{\rm i}$. As regards the dependence on $E$, it is determined by a particular IS energy spectrum. In Figure \[fig6\] (discussed in the next Section), $|z_{\rm min}(E)|$ is the left border of the plotted diffusion zone, calculated for IS spectrum (\[IS\_spectrum\]); it scales approximately as $\propto E^{1.3}$ in the non-relativistic case.
Once requirement $|z_{\rm min}|\lesssim H$ is fulfilled and the diffusive regime operates, the dimensionless CR flux is given by the corresponding expression in Equation (\[smin\_norm\]), with $\tilde D\partial \tilde j/\partial \tilde
z=2\tilde k\nu$ and $\tilde j(E,z)$ from Equation (\[Diff\_norm2\]). We see that the diffusion part of the modulated flux dominates over the advection part when $2\tilde k\nu \gtrsim \tilde j$. This remarkably coincides with condition $\eta_0\lesssim1$ of the diffusion-dominated flux for the loss-free case – with the only difference that now $\eta_0$ should be evaluated not for $\tilde j_{\rm IS}(E)$ but for derived $\tilde j(E,z)$. Then the modulated flux (in absolute units) is still given by Equation (\[flux1\]) obtained for the loss-free case; moreover, in the presence of losses, $S_{\rm DD}(E)$ dominates over a broader range of parameters, since $\eta_0$ should be additionally multiplied by a factor of $j/j_{\rm IS}\leq1$.
If advection dominates over diffusion, transport equation (\[Diff\_norm2\]) still describes the advection part of flux (\[smin\_norm\]). In this case, the modulated flux is given by Equation (\[flux2\]) with $\tilde j_{\rm IS}(E)$ replaced by $\tilde j(E,z)$.
CR flux {#zone_onset2}
-------
Summing up the above results, we conclude that the modulated CR flux in the presence of losses can be written as a simple superposition of the diffusion and advection asymptotes: The diffusion flux is given by Equations (\[flux1\]), and the advection flux is described by a modified Equation (\[flux2\]), with $\tilde j_{\rm IS}(E)$ replaced by solution $\tilde
j(E,z)$ of Equation (\[Diff\_norm2\]). This yields $$\label{flux_final2}
S(E,z)\simeq S_{\rm DD}(E)\left(1+\sqrt{\tilde E(\tilde E+2)}\:\frac{\tilde j(E,z)}{2\nu}\right),$$ where the relative magnitude of the advection flux is equal to the modified diffusion depth (\[zeta\_00\]).
It is noteworthy that the sum of $S_{\rm DD}$ and $S_{\rm AD}$ not only provides the correct asymptotic behavior – as demonstrated below, Equation (\[flux\_final2\]) also allows us to accurately describe a crossover between them. This can be understood by bearing in mind a remark we made in the end of Section \[inhomogeneous\]: At higher energies, the losses tend to extend the range of applicability of the excitation-damping balance, Equation (\[Wave\_norm2\]), which directly determines $S_{\rm DD}(E)$. Therefore, Equations (\[flux1\]) remains accurate where the crossover to advection occurs.[^7] Moreover, the losses generally reduce the relative magnitude of the advection flux, so that the crossover may not take place at all.
>From Equations (\[smin\]) and (\[sfree\]) it follows that the diffusive regime operates as long as the modulated flux, approximately equal to $S_{\rm DD}(E)$, is smaller than the [*local*]{} free-streaming flux, which is proportional to $\tilde j(E,z)$. Equation (\[Diff\_norm2\]) suggests that this condition is violated at sufficiently large $|z|$, where $\tilde j(E,z)$ becomes too small due to the losses. The corresponding inner border of the diffusion zone, $|z_0(E)|$, can be directly obtained from excitation criterion (\[threshold\]) (written for given $E$) where, again, $\tilde j_{\rm
IS}(E)$ is replaced by $\tilde j(E,z)$, $$\label{threshold2}
\frac{\tilde E+2}{\tilde E+1}\tilde E\tilde j(E,z_0)=\frac{2\epsilon \nu}{\langle\mu\rangle}\,.$$ Here, $\langle\mu\rangle$ is the average pitch angle of CRs for $|z|>|z_0(E)|$, which corresponds to a “downstream” free-streaming zone (see Appendix \[<mu>\]). Since the exact value of $\langle\mu\rangle\sim1$ is unimportant for the presented analysis, for simplicity we keep the same notation as for the CR flux in the free-streaming zone I.
The diffusion zone in the presence of losses is shown in Figure \[fig6\], where the left border $|z_{\rm min}(E)|$ is determined from condition (\[condition\]) and the right border $|z_0(E)|$ is derived from Equation (\[threshold2\]). The overall shape of the zone and its qualitative change with $\nu$ are quite similar to what we see in Figure \[fig2\] for the absorbing-wall case (we remind that distance $|z|$ in Figure \[fig6\] is measured in the negative direction). However, $|z_{\rm min}|$ and $|z_0|$ are much larger than the respective spatial scales ($z_{\rm min}$ and $z_0$) in Figure \[fig2\]. Also, Equation (\[scaling\_z2\]) shows that $|z_{\rm min}|$ does not depend on $\nu$, i.e., the diffusion zone shrinks due to a rapid decrease of $|z_0|$ with $\nu$,[^8] while for the absorbing-wall case both borders move toward each other as $\nu$ increases (see Equation (\[scaling\_z1\])).
The free-streaming flux $S_{\rm free}(E,z)= 4\pi \langle\mu\rangle j(E,z)$ at $|z|>|z_0(E)|$ (as well as for $E>E_{\rm ex}$) is determined by $j(E,z)$ which is a solution of transport equation (\[Diff\_norm\]). A general form of the solution in $(p,z)$ space is $$\label{solution_gen2}
\tilde L_{\rm g}(p)\tilde j(p,z)=\Phi\left(\tilde z-\langle\mu\rangle\int \frac{d\tilde p\:\tilde v}{\epsilon
\tilde L_{\rm g}(\tilde p)}\right),$$ and the resulting $S_{\rm free}(E,z)$ has to be matched at $z=z_0(E)$ with Equation (\[flux\_final2\]). Of course, the free streaming regime is only realized when $|z_0(E)|<H$, otherwise the CR flux penetrating the cloud is directly given by Equation (\[flux\_final2\]).
The characteristic behavior of the modulated CR flux in the presence of losses is illustrated in Figure \[fig7\] for $\nu=3.5$, again calculated for IS spectrum (\[IS\_spectrum\]). One can see that the analytical curves obtained from Equation (\[flux\_final2\]) are in excellent agreement with the numerical results. The way how the losses modify the flux is evident by comparing these curves with the corresponding loss-free curve plotted in the left panel of Figure \[fig4\]: The flux is attenuated with the distance at lower energies, thus suppressing a crossover to the advection-dominated flux, clearly seen in Figure \[fig4\] for $\nu=3.5$ (where the curve in the left panel steadily increases toward smaller $E$). Furthermore, at $|z|>|z_0(E)|$ the losses induce a “backward” transition to the free-streaming regime, seen as the kink for $|z|=10^{20}$ cm. For larger $\nu$ (not shown here), where the advection contribution is practically negligible, the curves become almost horizontal in the diffusive regime and, hence, undistinguishable from those in Figure \[fig4\]. This striking similarity is a manifestation of the universal behavior characterizing the diffusion-dominated flux $S_{\rm
DD}(E)$.
Discussion and conclusions {#discussion}
==========================
A comparison of results obtained in Sections \[model\] and \[realistic\] demonstrates that, when calculating the magnitude of the modulated CR flux, it is largely unimportant what leading mechanism – absorbing wall or gas losses – causes the self-modulation: Figure \[fig6\] suggests that in the presence of losses the condition of diffusion onset, $|z_{\rm min}(E)|\lesssim H$, is usually fulfilled for non-relativistic CRs (assuming typical envelope size of 3–10 pc), and hence they are modulated due to turbulence induced near the outer envelope boundary. For relativistic CRs losses are typically unimportant at a scale of the envelope, and their self-modulation occurs near the absorbing cloud wall; according to Figure \[fig2\], the respective condition $z_{\rm min}(E)<H$ is well satisfied. Nevertheless, the resulting CR flux remains [*universal at all energies*]{} below $E_{\rm ex}$ – it is described by the diffusion-dominated asymptote $S_{\rm DD}(E)$, Equation (\[flux1\]). Figures \[fig4\] and \[fig7\] indicate that the effect of advection, causing a deviation from this dependence, only becomes significant if $\nu\lesssim10$ (according to Equation (\[scale\_nu\]), the corresponding gas density in the envelope typically must be well below $\sim100$ cm$^{-3}$).
Of course, the gas losses can destroy universality of the energy spectrum for low-energy CRs penetrating into the cloud: Figure \[fig6\] shows that, at lower energies and for sufficiently large $\nu$ ($\gtrsim100$), the right border of the diffusion zone $|z_0(E)|$ becomes smaller than typical $H$. As discussed in Section \[realistic\], the further free-streaming propagation of such CRs in the envelope is described by Equation (\[solution\_gen2\]), and their flux is proportional to the local spectrum $j(E,z)$. If the remaining distance $H-|z_0(E)|$ exceeds the integral term in the parentheses (multiplied by $z_*$), the attenuation modifies the universal spectrum of $S_{\rm DD}(E)$ before CRs reach the cloud.
The presented results allow us to address several important questions regarding interaction of CRs with molecular clouds, and draw the following major conclusions:
1. [*Dimensionless numbers.*]{} Generic features of CR propagation in low-density envelopes are completely determined by two dimensionless numbers: gas damping rate $\nu$, Equation (\[nu\]), which governs the diffusive transport regime (due to the self-generated MHD turbulence), and small parameter $\epsilon$, Equation (\[epsilon\]), which controls a transition between the diffusive regime and a free streaming of CRs (where the turbulence is unimportant).
2. [*Diffusive propagation.*]{} The turbulence generated by CRs in the envelope affects their transport at energies below the excitation threshold $E_{\rm ex}$, Equation (\[threshold\]), which is a function of the product $\epsilon\nu$. As a result, the CR flux becomes self-modulated before penetrating into the cloud – it changes from a free-streaming flux, determined by given IS energy spectrum $j_{\rm IS}(E)$, to the universal diffusion-dominated flux $S_{\rm DD}(E)$, scaling as $\propto E^{-1}$ both in the non-relativistic and ultra-relativistic limits. The locations of the diffusion zones (regions of the diffusive propagation) in the envelope are determined by the leading mechanism of self-modulation for given $E<E_{\rm ex}$: The zone can either be formed near the inner boundary (for higher-energy CRs, whose propagation is unaffected by the gas losses) or near the outer boundary (for lower energies, where the losses are essential).
3. [*Wave losses.*]{} In Section \[WL\] we showed that taking into account the wave losses basically leads to a renormalization of the advection flux $S_{\rm AD}(E)$, Equation (\[flux2\]). Since a contribution of $S_{\rm AD}$ to the modulated CR flux is significant only for relatively small $\nu$, the effect of wave losses can be practically always neglected.
4. [*Important physical parameters.*]{} The excitation threshold $E_{\rm ex}(\epsilon\nu)$ does not depend on the magnetic field $B$; it is a function of the physical parameters of the envelope as well as of the magnitude and the form of $j_{\rm IS}(E)$. One of our key findings is that the universal flux $S_{\rm DD}(E)$ is insensitive to the particular model of nonlinear wave cascade, depends neither on $B$ nor on $j_{\rm IS}(E)$, and thus is only determined by densities and masses of the neutral and ionized species in the envelope, Equation (\[flux\_par\]).
5. [*Magnitude of the self-modulation.*]{} The CR modulation due to self-generated turbulence is conveniently characterized by the flux ratio $$\frac{S_{\rm DD}(E)}{S_{\rm free}(E)}\simeq\frac{\epsilon\nu}{\tilde E\tilde j_{\rm IS}(E)}\,,$$ determined by Equations (\[flux0\]) and (\[flux1\]). For IS spectra analogous to that of Equation (\[IS\_spectrum\]), the product $\tilde E\tilde j_{\rm IS}(E)$ achieves a broad maximum ($\sim1$) at $E\sim100$ MeV. Therefore, the strongest modulation occurs at these energies, where the reduction is $\sim\epsilon\nu$; for typical envelopes, the flux can decrease by up to two orders of magnitude.
The conclusion that the CR flux penetrating into denser cloud regions has a universal energy dependence at $E<E_{\rm ex}$, solely determined by the physical parameters of the envelope, is of substantial general interest and importance. One of the reasons is that gamma-ray emission, measured from molecular clouds at different distances from the Galactic Center [see, e.g., @digel; @yang1; @tib15], is considered to provide information about the global distribution of CRs in the Galaxy [see e.g. @ahar01; @casa10]. The derived spatial distribution of Galactic CRs is then interpreted as a result of global-scale CR propagation and used as an input for models of the CR origin [see, e.g., @Bloe2; @Breit; @galprop; @recc16a]. Thus, the fact that the modulated flux is independent of the spectrum of Galactic CRs may have profound implications for such analysis.
Also, observations indicate that the central regions of the Galactic Disk are enhanced by molecular hydrogen in the form of very dense molecular clouds and diffuse gas [@oka05]. The latter occupies about 30% of the volume of the central molecular zone, and therefore the overall effect of the local self-modulation, which we predict to occur in these diffuse regions, can be significant. For example, the spectrum of CR protons deduced by @acero16 and @yang3 from the Fermi data for the inner Galaxy is harder than that in the outer Galaxy, and one can speculate that this may be due to the local self-modulation.
The self-modulation of a CR flux can be important for many other fundamental problems. In particular, this could cause the substantial reduction of CR ionization rates observed within dense molecular clouds [e.g., @caselli98], significantly lower than those measured toward diffuse clouds [@indri12]. We note that drops in the amount of CR flux, and the consequent drop in the CR ionization rate within (UV-)dark clouds, affect physical parameters crucial for the dynamical evolution of dense clouds: the ionization fraction, which controls the coupling between gas and magnetic fields, thus regulating star formation [e.g. @mckee89]; the gas temperature, which determines the thermal pressure, particularly important at the scales of dense cloud cores [e.g., @fuller92; @keto08] where stars form; internal MHD turbulence in molecular clouds, which could contribute to the observed magnetic and virial equilibrium and thus to the cloud dynamics and evolution [e.g., @myers88; @goodman98; @caselli02]. Last but not least, changes in the CR flux can significantly affect the chemistry, as gas-phase processes in dark clouds are dominated by ion-molecules reactions with rates depending on the ionization fraction [@herbst73], while surface chemistry can be modified by CRs directly (via the impulsive spot heating) or indirectly (via the UV-photons generated by the fluorescence of H$_2$ molecules).
Self-consistent numerical simulations of dynamically and chemically evolving magnetized interstellar clouds (with a proper treatment of CR propagation inclusive of their self-modulation and MHD turbulence generation) are needed to quantify our predictions for case-specific clouds within our Milky Way and external galaxies, as well as to test our theory against observations.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors are grateful to Andy Strong for reading the manuscript and giving useful comments. VAD and DOC are supported in parts by the grant RFBR 18-02-00075. DOC is supported in parts by foundation for the advancement of theoretical physics “BASIS”. PC acknowledges support from the European Research Council (ERC) Advanced Grant PALs 320620. CMK is supported in part by the ROC Ministry of Science and Technology grants MOST 104-2923-M-008-001-MY3 and MOST 105-2112-M-008-011-MY3. KSC is supported by the GRF Grant under HKU 17310916.
Appendix A\
Average pitch angle in the free-streaming regime {#<mu>}
================================================
Different transport zones are sketched in Figure \[figA1\]. For certainty, the zones are illustrated for the absorbing-wall setup (distance $=z$, see Figure \[fig1\]); the results are then readily applied to the setup with losses (distance $=|z|$, see Figure \[fig5\]). One can identify three free-streaming zones:
In zone I, corresponding to $E>E_{\rm ex}(\nu)$, CRs propagate across the envelope without experiencing scattering at any distance. The value of $\langle\mu\rangle$ in this case depends on mechanisms governing modification of isotropic IS spectrum $j_{\rm IS}(E)$ upon its entering into the envelope. (Since the strengths of the magnetic field inside and outside the envelopes are about the same, it is reasonable to assume that the magnetic field lines enter into the envelope without significant distortions.) Let us denote the spectrum formed upon entering as $j_{\rm IS}^*(E,\mu)$ with $\mu>0$. Then the average pitch angle, which determines free-streaming flux $S_{\rm free}(E)$ in Equation (\[sfree\]), is readily obtained: $$\label{<mu>eq}
\langle\mu\rangle=\int_{0}^{1}d\mu\:\mu \frac{j_{\rm IS}^*(E,\mu)}{{j_{\rm IS}(E)}}\,.$$ The exact form of $j_{\rm IS}^*(E,\mu)$ depends on unknown details of entering, but one can generally conclude that the resulting value of $\langle\mu\rangle$ is of the order of a few tenths. For instance, if $j_{\rm IS}^*(E,\mu)$ is simply a hemisphere $\mu>0$ of $j_{\rm IS}(E)$, then $\langle\mu\rangle=1/2$ (which corresponds to a well-known expression for a free-streaming flux through a flat surface). Generally, $\langle\mu\rangle$ may be a function of $E$.
Zone II is located “downstream” from the diffusion zone. The value of $\langle\mu\rangle$ is determined by modification of a local quasi-isotropic CR spectrum $j(E)$ leaving the diffusion zone. While details of this process may be different from those controlling $\langle\mu\rangle$ in zone I, one can still employ Equation (\[<mu>eq\]) with $j_{\rm IS}(E)$ replaced by $j(E)$. Using exactly the same line of arguments as before, we conclude that $\langle\mu\rangle$ in zone II should be about that in zone I.
Zone III “upstream” from the diffusion zone is unimportant for our analysis. For $E\ll E_{\rm ex}$, the flux propagating further toward the cloud the strongly modulated, i.e., the incident IS flux is almost entirely reflected back from the diffusion zone. Therefore, the value of $\langle\mu\rangle$ in zone III is very small, tending to $\sim v_{\rm A}/v$ when advection part of the (modulated) flux in Equation (\[smin\]) dominates over the diffusion part.
In the presence of losses, the “upstream” (“downstream”) zone corresponds to smaller (larger) distances (see Figure \[fig5\]). Essentially, in this case we only need to swap zones II and III in the shown sketch.
Appendix B\
Numerical solution of the governing equations {#numerical}
=============================================
Numerical results are deduced from the steady-state solution of time-dependent dimensionless Equations (\[Diff\_norm\]) and (\[Wave\_norm\]), obtained by adding terms $-\partial\tilde j/\partial\tilde t$ and $(2\tilde W)^{-1}\partial\tilde
W/\partial\tilde t$, respectively. Dimensionless time $\tilde t=t/t_*$ is determined by $t_*$ whose value is dictated by the used normalization. We employ an explicit finite difference method, which has straightforward implementation and reasonable convergence for our parameters.
To include the limitations on the CR flux velocity, we split this method into two steps: First, we evaluate the flux from $$\label{eq:SD_numerical}
\begin{array}{l}
{\displaystyle(\tilde{S}_{\rm diff})_{i,l}=\tilde{D}_{i,l}\frac{\tilde{j}_{i+1,l}
-\tilde{j}_{i,l}}{\tilde{z}_{i+1}-\tilde{z}_{i}}+\tilde{j}_{i+1,l}}\,,\vspace{.2cm}\\
\tilde{S}_{i,l} = \mbox{sign}\left\{(\tilde{S}_{\rm diff})_{i,l}\right\} \times \min \left\{|(\tilde{S}_{\rm diff})_{i,l}|,\:
(\tilde{S}_{\rm free})_{i+1,l}\right\},
\end{array}$$ and then calculate the evolution of the CR energy spectrum $\tilde j_{i,l}$. Here indices $i$ and $l$ represent discretization of the spatial coordinate and energy, respectively.
In fact, $(\tilde{S}_{\rm diff})_{i,l}$ in Equation (\[eq:SD\_numerical\]) is evaluated at an intermediate grid point, for which chose midpoint $\tilde z_{i+\frac12}=\frac12(\tilde z_i+\tilde z_{i+1})$. Therefore, also the diffusion coefficient $\tilde{D}_{i,l}$ and the density of MHD waves $\tilde{W}_{i,l}$ are calculated at $z_{i+\frac12}$. However, for brevity we omit $\frac12$ in the spatial index, keeping in mind that all these parameters actually correspond to the midpoint. Thus, a discrete equation for the energy spectrum is written as $$\frac{\tilde{j}_{i,l}(t+\Delta t)-\tilde{j}_{i,l}(t)}{\Delta t}
=2\frac{\tilde{S}_{i,l}-\tilde{S}_{i-1,l}}{\tilde{z}_{i+1}-\tilde{z}_{i-1}}
+\frac{(\tilde{L}_g)_{i,l+1}\tilde{j}_{i,l+1}-(\tilde{L}_g)_{i,l}\tilde{j}_{i,l}}{\tilde{p}_{l+1}-\tilde{p}_{l}}\,,$$ where relation $\tilde{z}_{i+\frac12} - \tilde{z}_{i-\frac12} = \frac12(\tilde z_{i+1} - \tilde z_{i-1})$ is taken into account. For small values of the diffusion coefficient, this becomes a standard explicit scheme for the heat transport equation with central difference, otherwise it transforms into an upwind scheme.
The evolution of density of the MHD waves is performed in a similar way. We have verified that results do not practically change when the advection wave transport, described by the first term on the lhs of Equation (\[Wave\_Eq\]), is taken into account. This allows us to omit this term and use the following upwind scheme: $$\frac{\tilde{W}_{i,l}(t+\Delta t)-\tilde{W}_{i,l}(t)}{\Delta t}
+\frac{\tilde{k}^{3}_{l+1}\tilde{W}^2_{i,l+1}-\tilde{k}^{3}_l \tilde{W}^2_{i,l}}{\tilde{k}_{l+1}-\tilde{k}_l}
=2(\tilde{\Gamma}_{i,l} - \tilde{\nu})\tilde{W}_{i,l}\,,$$ where $\tilde{\Gamma}_{i,l} = \tilde{S}_{i,l}/(2\tilde{k}_l)$; for $2(\tilde{\Gamma}_{i,l} - \tilde{\nu})\Delta t \ll 1$, the last term is replaced by $2(\tilde{\Gamma}_{i,l} - \tilde{\nu})\tilde{W}_{i,l}(t+\Delta t)$. To simplify the problem, we utilize the same grid for $\tilde j$ and $\tilde W$, and therefore $\tilde{p}_l$ and $\tilde{k}_l$ are related through the resonance condition.
Boundary conditions for the above equations are: $$\begin{aligned}
&&\tilde{j}_{1,l} = 0, \nonumber \\
&&\tilde{j}_{\mathcal I,l} = (\tilde{j}_{\rm IS})_l\,, \nonumber \\
&&\tilde{j}_{i,\mathcal L} = 0, \nonumber \\
&&\tilde{W}_{i,\mathcal L} = 0,\nonumber\end{aligned}$$ where $\mathcal I$ and $\mathcal L$ denote the number of points on $z$ and $E$ (or $k$) axes, respectively.
In order to accelerate the relaxation process, we assume that CRs are uniformly distributed at the initial moment, i.e., $\tilde{j}(t=0) = \tilde{j}_{\rm IS}$. As for the waves, we introduce a certain “zero-level” turbulence at the initial moment, and also ensure that $W$ never decreases below that level during its evolution. The choice of zero-level turbulence is dictated by two conditions: first, this should not affect CR propagation; second, this should be large enough for a fast convergence. The first condition is satisfied if the corresponding diffusion coefficient is $\sim\Theta vH$ with $\Theta \gg
1$, whereas the convergence time logarithmically depends on $\Theta$. Hence, a reasonably fast convergence can be archived for a wide range of $\Theta$, for our calculations $\Theta = 10^{10}$ was chosen.
The energy loss function $L_{\rm g}(E)$ is calculated as a sum of the ionization and pion production terms. Ionization losses, essential for non-relativistic protons, are taken from PSTAR NIST database [@star], while for losses due to the pion production we employ the expression proposed by @mann94.
Appendix C\
Expansion of the wave spectrum in series of $z$ {#linear_W}
===============================================
In Figure \[figA2\] we plot the wave spectrum $W(k,z)$ calculated numerically from Equation (\[W\_SC\]).
One can see that, when $\nu\gg1$ and $\eta_0\lesssim1$, Equation (\[field\]) reasonably approximates the numerical results except for a region near $z\simeq z_0$, where $W$ is relatively small. If needed, a quadratic term $\propto z^2$ can be included in Equation (\[field\]) to further improve the agreement with the numerical results. For relatively small values of $\nu$ and $\eta_0\gtrsim1$ the linear expansion fails to describe the results properly.
Abdalla, H., et al. \[H.E.S.S. Collaboration\] 2017, arXiv:1706.04535 Abramowski, A., et al. \[H.E.S.S. Collaboration\] 2016, Natur, 531, 476 Acero, F., Ackermann, M., Ajello, M., et al. 2016, ApJS, 224, 8 Aharonian, F. A. 2001, SSRv, 99, 187 Aharonian, F., et al. \[H.E.S.S. Collaboration\] 2006, Natur, 439, 695 Benson, P. J., & Myers, P. C. 1989, ApJS, 71, 89 Berger, M. J., Coursey, J. S., Zucker, M. A., & Chang, J. 2005, http://physics.nist.gov/Star Berezinskii, V. S., Bulanov, S. V., Dogiel, V. A., Ginzburg , V. L., & Ptuskin, V. S. 1990, in Astrophysics of Cosmic Rays, ed. V. L. Ginzburg (Norht-Holland: Amsterdam) Bloemen, J. B. G. M., Dogiel, V. A., Dorman, V. L., & Ptuskin, V. S. 1993, A&A, 267, 372 Breitschwerdt D., Dogiel, V. A., & V[ö]{}lk, H. J. 2002, A&A, 385, 216
Casanova, S., Aharonian, F. A., Fukui, Y., et al. 2010, PASJ, 62, 769
Caselli, P., Walmsley, C. M., Terzieva, R., & Herbst, E. 1998, ApJ, 499, 234
Caselli, P., Benson, P. J., Myers, P. C., & Tafalla, M. 2002, ApJ, 572, 238
Caselli, P., Keto, E., Bergin, E. A., et al. 2012, ApJL, 759, L37
Cecchi-Pestellini, C., & Aiello, S. 1992, MNRAS, 258, 125
Cesarsky, C. J., & Völk, H. J. 1978, A&A, 70, 367
Crutcher, R. M. 2012, ARA&A, 50, 29 Dalgarno, A. 2006, PNAS, 103, 12269 Digel, S. W., Grenier, I. A., Hunter, S. D., Dame, T. M., & Thaddeus, P. 2001, ApJ, 555, 12 Dogel \[Dogiel\], V. A., & Sharov, G. S. 1985, SvAL, 11, 346 Dogiel, V. A., Gurevich, A. V., & Zybin, K. P. 1994, A&A, 281, 937
Dogiel, V. A., Ichimura, A., Inoue, H., & Masai, K. 1998, PASJ, 50, 567 Dogiel, V., Chernyshov, D., Koyama, K. Nobukawa, M., & Cheng, K. S. 2011, PASJ, 63, 535
Fuller, G. A., & Myers, P. C. 1992, ApJ, 384, 523
Goldreich, P., & Sridhar, S. 1997, ApJ, 485, 680.
Goldsmith, P. F. 2001, ApJ, 557, 736
Goodman, A. A., Barranco, J. A., Wilner, D. J., & Heyer, M. H. 1998, ApJ, 504, 223
Ginzburg, V. L. 1979, Theoretical Physics and Astrophysics (Oxford: Pergamon Press) Hayakawa, S. 1969, Cosmic Ray Physics. Nuclear and Astrophysical Aspects (New York: Wiley-Interscience)
Herbst, E., & Klemperer, W. 1973, ApJ, 185, 505
Indriolo, N., & McCall, B. J. 2012, ApJ, 745, 91 Iroshnikov, P. 1964, SvA, 7, 566 Ivlev, A. V., Padovani, M., Galli, D., & Caselli, P. 2015a, ApJ, 812, 135 Ivlev, A. V., Röcker, T. B., Vasyunin, A., & Caselli, P. 2015b, ApJ, 805, 59 Ivlev, A. V., Akimkin, V. V., & Caselli, P. 2016, ApJ, 833, 92
Keto, E., & Caselli, P. 2008, ApJ, 683, 238
Ko, C.-M. 1992, A&A, 259, 377
Kraichnan, R. H. 1965, PhFl, 8, 1385 Krivonos, R. Clavel, M., Hong, J. S., et al. 2017, MNRAS, 468, 2822 Kulsrud, R., & Pearce, W. P. 1969, ApJ, 156, 445 Lagage, P. O., & Cesarsky, C. J. 1983, A&A, 118, 223 Léger, A., Jura, M., & Omont, A. 1985, A&A, 144, 147 Lerche, I. 1967, ApJ, 147, 689 Lis, D. C., & Goldsmith, P. F. 1990, ApJ, 356, 195
Mannheim, K., & Schlickeiser, R. 1994, A&A, 286, 983
McKee, C. F. 1989, ApJ, 345, 782
Morlino, G., & Gabici, S. 2015, MNRAS, 451, L100
Myers, P. C., & Goodman, A. A. 1988, ApJL, 326, L27
Nobukawa, K. K., Nobukawa, M., Uchiyama, H., et al. 2015, ApJ, 807, L10 Norman, C. A., & Ferrara, A. 1996, ApJ, 467, 280 Oka, T. 2006, PNAS, 103, 12235
Oka, T., Geballe, Th. R., Goto, M., et al. 2005, ApJ, 632, 882 Okuzumi, S. 2009, ApJ, 698, 1122 Padovani, M., Galli, D., & Glassgold, A. E. 2009, A&A, 501, 619 Padovani, M., Hennebelle, P., & Galli, D. 2013, A&A, 560A, 114 Protheroe, R. J., Ott, J., Ekers, R. D., Jones, D. I., & Crocker, R. M. 2008, MNRAS, 390, 683 Ptuskin, V. S., Moskalenko, I. V., Jones, F. C., Strong, A. W., & Zirakashvili, V. N. 2006, ApJ, 642, 902 Recchia, S., Blasi, P., & Morlino, G. 2016a, MNRAS, 462, L88 Recchia, S., Blasi, P., & Morlino, G. 2016b, MNRAS, 462, 4227 Schlickeiser, R. & Shalchi, A. 2008, ApJ, 686, 292 Schlickeiser, R., Caglar, M., & Lazarian, A. 2016, ApJ, 824, 89
Shen, C. J., Greenberg, J. M., Schutte, W. A., & van Dishoeck, E. F. 2004, A&A, 415, 203
Shingledecker, C. N., Le Gal, R., & Herbst, E. 2017, PCCP, 19, 11043
Shu, F. H., Adams, F. C., & Lizano, S. 1987, ARA&A, 25, 23
Skilling, J. 1975, MNRAS, 173, 255 Skilling, J., & Strong, A. W. 1976, A&A, 53, 253 Strong, A. W., Moskalenko, I. V., & Ptuskin, V. S. ARNPS, 57, 285, 2007
Tatischeff, V., Decourchelle, A., & Maurin, G. 2012, A&A, 546, 88
Testi, L., Birnstiel, T., Ricci, L., et al. 2014, in Protostars and Planets VI, ed. H. Beuther et al. (Tucson, AZ: Univ. Arizona Press ), 339
Tibaldo, L., Digel, S. W., Casandjian, J. M., et al. 2015, ApJ, 807, 161 Wentzel, D. G. 1974, ARAA, 12, 71
Yamamoto, S. 2017, Introduction to Astrochemistry: Chemical Evolution from Interstellar Clouds to Star and Planet Formation (Tokyo: Springer Japan)
Yang, R.-z., de Oña Wilhelmi, E., & Aharonian, F. 2014, A&A, 566, A142 Yang, R.-z., Jones, D. I., & Aharonian, F. 2015, A&A, 580, 90 Yang, R.-z., Aharonian, F., & Evoli, C. 2016, PhRvD, 93, 123007 Yusef-Zadeh, F., Hewitt, J. W., Wardle, M., et al. 2013, ApJ, 762, 33
Zhao, B., Caselli, P., Li, Z.-Y., Krasnopolksy, R., Shang, H., & Nakamura, F. 2016, MNRAS, 111, 22
[^1]:
[^2]: The CR flux and hence the excited MHD waves propagate from right to left, as sketched below in Figure \[fig1\]. Therefore, the minus sign is added in front of $S_{\rm free}$ and $v_{\rm A}N$ (note also that $\partial N/\partial z\geq0$ in this case).
[^3]: In the following we demonstrate that the modulated CR flux is insensitive to the particular model of cascade.
[^4]: For simplicity, the tilde sign over the dimensionless parameters $\nu$ and $\epsilon$ is omitted.
[^5]: In fact, the CR density remains finite in the diffusive regime: it is determined from the equality of the modulated and free-streaming fluxes in Equation (\[smin\]), i.e., from condition $S=-S_{\rm free}$.
[^6]: We note that the obtained maximum gas density is about the [*average*]{} density inside dense cores [e.g., @benson89].
[^7]: We remind that in the loss-free case, the excitation-damping balance always breaks down at the crossover point $\eta_0\sim 1$, see Sections \[approximate\] and \[CR\_flux\].
[^8]: In the presence of losses, the dependence of $|z_0|$ on the physical parameters is different in the non-relativistic and relativistic cases, as one can see by substituting Equations (\[solution1\]) and (\[solution2\]) in Equation (\[threshold2\]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cyber-Physical Systems (CPS) have the promise of presenting the next evolution in computing with potential applications that include aerospace, transportation, robotics, and various automation systems. These applications motivate advances in the different sub-fields of CPS (e.g. mobile computing and communication, control, and vision). However, deploying and testing complete CPSs is known to be a complex and expensive task. In this paper, we present the design, implementation, and evaluation of *Up and Away (UnA)*: a testbed for Cyber-Physical Systems that use UAVs as their physical component. *UnA* aims at abstracting the control of physical components of the system to reduce the complexity of UAV oriented Cyber-Physical Systems experiments. In addition, *UnA* provides an API to allow for converting CPS simulations into physical experiments using a few simple steps. We present a case study bringing a mobile-camera-based surveillance system simulation to life using *UnA*.'
author:
- |
\
\
\
title: |
Up and Away: A Cheap UAV Cyber-Physical Testbed\
(Work in Progress)
---
Introduction
============
Cyber-Physical Systems (CPS) have the promise of presenting the next evolution in computing by bridging the gap between the virtual world and the physical world. CPS potential applications include aerospace, transportation, and factory automation systems [@rajkumar2010cyber]. Advancements in cyber-physical systems research is supported by development in other research fields including mobile computing, embedded systems, computer vision, control, and communication. While this development ultimately contributes to advancing different components of cyber-physical systems, most of the work is evaluated either through simulations or through experiments that focus on the specific sub-problem. Furthermore, the complex and multidisciplinary nature of full scale experiments of cyber-physical systems makes the task even harder especially when it comes to testing individual “cyber” or “physical” components. From a different perspective, current generic CPS testbed [@fok2011pharos] are typically complex and incur large deployment cost.
In this paper, we present Up and Away (UnA), a testbed that allows for low-cost, rapid development of cyber-physical systems experiments. *UnA* abstracts the “physical” components of the system to allow for experiments that focus on “cyber” components with real physical interactions. We chose UAVs as the physical component of the system due to their deployment flexibility and maneuverability. Furthermore, UAVs are suitabile for many application (e.g. surveillance, transportation, and communications) allowing for cyber-physical experimentation with realistic objectives.
Several research groups have recently developed quadcoptor testbeds . The GRASP testbed [@michael2010grasp] uses off-the-shelf high-end Ascending Technologies Hummingbird quadcopters to demonstrate multirobot control algorithms. The ETH Flying Machine Arena [@lupashin2010simple] uses modified Hummingbird quadcopters to demonstrate several acrobatic and athletics control maneuvers. A testbed was developed by University of Colorado, Boulder for testing ad-hoc networking scenarios [@brown2004test]. The testbed uses UAVs which were deployed in a 7 $km^2$ area to test different operation conditions of the DSR protocol. The work in [@jung2005design] shows the development of a low-cost UAV testbed using Goldberg Decathlon ARF model airplane. The work aims at showing the design and development of hardware and software components to produce low-cost UAVs.
Unlike other UAV-based testbeds, *UnA*’s architecture is more concerned with abstracting the control of UAVs rather than improving it. Furthermore, *UnA* presents an architecture that provides default components and facilitates replacing any of them to test its impact on the overall performance of the system. To that end, AR Drones were selected as the UAVs to be used. We preferred using quadcopters to fixed-winged airplanes due to their maneuverability especially in highly constrained spaces (e.g. indoors) including their ability to hover. Moreover, AR Drones are extremely cheap[^1] UAVs that come with a large array of sensors and a widely supported open source API to control the drone and obtain sensory information for its sensors (the specifications of the AR Drone are better described in Section \[sec:arch\]).
UnA Architecture {#sec:arch}
================
In this section, we introduce the design and implementation of UnA. We start with our design goals followed by the architecture description. The, We describe the details of how *UnA* handles targets and UAVs localization, UAVs control, CPS processing and communication.
Design Goals
------------
The development of *UnA* is motivated by the need for a low-cost, easy to deploy experiments for CPS. Thus, *UnA*’s architecture is designed with the following goals in mind:
1\. *Multidisciplinary nature of CPS experiments:* The development of CPS experiments requires awareness of several fields including hardware design, control systems, software systems development [@rajkumar2010cyber]. This poses significant challenges for developing realistic experiments for such systems. *UnA* was designed with this unique feature of CPS in mind and therefore the *UnA* modules are extensible to support development of any of the CPS’s components.
2\. *Support of large scale experiments:* CPS experiments with a large number of physical nodes requires the testbed to support different communication models between the nodes. Furthermore, the processing power required to optimize the overall state of the system can grow exponentially with the number of nodes. The *UnA* architecture supports Ad-Hoc communication using AODV for routing to facilitate communication between the different UAVs. Furthermore, it supports off-loading processing tasks required by the UAVs to central or multiple computing nodes.
3\. *Seamless control of the UAVs:* *UnA* aims at reducing the complexities in CPS experiments design incurred by the automatically operated mechanical components of the system by abstracting the control algorithms and hardware design which allows for more focus on the “cyber” problems. Furthermore, autonomous control of off-the-shelf UAVs is a difficult task [@michael2010grasp; @lupashin2010simple]. Thus, *UnA* relies on the easy to use AR Drone API and extends it to use way-point navigation instead of controlling the rotation (i.e. roll, yaw and gaz) of the UAV.
Overview
--------
![UnA Architecture[]{data-label="fig:arch"}](figures/arch.pdf){width="54.00000%"}
Figure \[fig:arch\] shows a summary of the *UnA* architecture. *UnA* allows the UAVs that can communicate with each other, process sensory information locally and act accordingly (i.e. control itself). It also provides a base node that monitors the behavior of all UAVs. Moreover the central node can handle processing intensive tasks that UAVs off-load to it. The *UAV Node Modules* are responsible for obtaining and processing information about the drone and environment’s state, calculating the drone’s corresponding objectives accordingly, and controlling the drone’s motion to achieve these objectives. *UnA* is built on top of the Parrot AR Drone 2.0 [@krajnik2011ar]. The AR Drone is a quadrotor helicopter that are electerically powered with two camera: a front 720p camera with $93^{\circ}$ lens and a vertical QVGA camera with $64^{\circ}$ lens. These drones are controlled through an external computer through WiFi and are equipped with an on board ARM Cortex A8 processor with 1 Gbit RAM that runs a Linux 2.6.32 Busybox. The on-board computing machine is responsible for collecting and reporting the state of the drone to an external computer that controls the drone. The AR Drones are relatively cheap, easy to program and comes with an SDK that provides assisted maneuvers which satisfy most of our design goals.
The *Central Node Modules* are responsible for monitoring the state of each drone, manually controlling the drone, and providing part of the environment state information to the drones. Due to the limited processing capacity of current off-the-shelf UAVs, UnA’s architecture enables off-loading some of the processing tasks from the UAVs to the central node to enable real time responses to changes in the environment. The *Experiment Admin GUI* displays the state information of the different UAVs. It also supports the manual control of the UAVs for emergency cases. The main purpose of UnA’s architecture is to support the rapid development of CPS experiments that use UAVs as their Physical part. While all UnA can be extended and customized for different CPS applications, we assume that simulation code was already developed by the experiment administrator. Thus, the central node supports communication between its different modules with other applications through sockets. The *UnA Sim API* provides the appropriate wrappers to allow the simulation code to interface with UnA to bring the simulations to life by off-loading all the *UAV Node Modules* functionalities to the simulation engine through the central node.
For the rest of the section, we explain the different layers of the *UnA* architecture according to the color coding in Figure \[fig:arch\].
UAV and Target Localization
---------------------------
UnA provides two different approaches for obtaining location information of the UAVs: 1) Parrot Flight Recorder GPS module and 2) using a master camera mounted at a high point to track the location of the drone. While the first approach is straight forward, it doesn’t work within controlled indoor environments. Thus, computer vision is used for tracking the UAVs in *UnA*.
Multiple UAV tracking using computer vision with multiple cameras is a tricky task to implement [@oh2011indoor]. For that purpose, we propose a simpler approach that uses a single camera, mounted at a high point in the center of the area of interest, to track the drones by tagging each drone with a distinct color tag[^2]. Tracking objects of a certain color could be easily implemented using OpenCV [@bradski2008learning]. This approach is further extended to track other targets as well.
While earlier work [@lupashin2010simple; @michael2010grasp] relies on Motion Capturing Systems (MCS) for tracking UAVs. We chose the single camera approach as a much cheaper alternative that provides reasonable accuracy (Section \[sec:case\])[^3].
UAV Control {#sec:control}
-----------
*UAV Controller* and *Drones Central Controller* are responsible for moving the drone to a specific X, Y coordinates with a certain orientation within the area of interest. While quadcopters can move in any direction with any orientation, we use a simple controlling method: at the beginning of each command the orientation of the drone is set to $90^{\circ} w.r.t.$ the x-y plan of the area of interest. The change in orientation uses the compass for feedback. The UAV then moves the drone along the x-axis to its specific coordinate then along the y-axis and finally it’s rotated to the specified orientation. This control algorithm obtains feedback from the UAV localization module and the UAV internal compass in order to actuate any errors in placing the drone due to its inertia.
The AR Drone’s motion is controlled by the AT commands protocol. The AT commands are text based commands that are translated by the AR Drone’s firmware to control the drone’s roll, pitch, gaz, and yaw to specify the direction, speed, and duration of the drone’s motion. The controller converts the newly obtained objectives and the current location information to the corresponding set of AT commands. This conversion could be made locally on the drone (i.e. autopilot mode) or on the central node (i.e. remote control mode).
CPS Optimization
----------------
The *UAV CPS Optimizer* module is responsible for updating the objective (i.e. location and orientation) of the drone after obtaining updates on the state of the drone and the environment from the drone’s sensor, other drones and the central node. This module has three modes of operation: 1) distributed mode, 2) central mode and 3) emulation mode. In the distributed mode, a distributed objective function is used where each UAV calculates its own new objectives after each update which is suitable for lightweight distributed algorithms. In the central mode, the *UAV CPS Optimizer* reports the updates to the *Central CPS Optimizer* which calculates the new objectives based updates from all UAVs. Finally, in the emulation mode, the UAVs implement a distributed objective function but off-load its calculations to the central node to meet real-time responsiveness requirements.
Control and Coordination Networks
---------------------------------
One of our design choices is to separate the intercommunication between the UAVs and the communication between each UAV and the central node to two different networks. This design choice was motivated by two reasons: 1) While communication between UAVs can be established in ad-hoc, unreliable fashion, communication between the UAVs and the central nodes should be using an infrastructure that is connected to the UAVs at all times, 2) AR Drone 2.0 [@krajnik2011ar] provides an SDK for monitoring and controlling the UAVs that requires continuous communication between the UAVs and a central node. We rely on that SDK to facilitate the different tasks of the *Central Node Modules*.
For the implementation of the *Control Network*, we use the built in WiFi card but changed its configuration to work as a client instead of as an access point which is its default configuration. This way all nodes can connect to the central node through a WiFi infrastructure. As for the *Coordination Network*, we plug a WiFi dongle into the drone, compile the dongle’s driver to work on the ARM-based Busybox, and install the driver on the drone. The dongle is then configured to work in Ad-Hoc mode and join the drones SSID as soon as its visible.
Case Study {#sec:case}
==========
In this section, we demonstrate the steps of bringing a sample CPS simulation to life using *UnA*, namely a mobile-camera-based surveillance system. The system’s objective is to maximize the number of targets covered by a set of mobile cameras. The proposed algorithm is a central algorithm whose inputs are the location of all targets and the mobile cameras, and its output is a set of location and orientation directives for all mobile cameras. For the rest of the section, we will show the different steps of implementing the proposed CPS using *UnA*.
UAVs and Targets Localization
-----------------------------
To localize the different moving objects in the area of interest (i.e. targets and UAVs), each drone is tagged with a distinct color. On the other hand, because we do not care about the identity of the targets, all targets are tagged with the same color. An Axis 213 PTZ Network Camera is mounted at 5.2 meters to locate the different objects. The central node queries the camera for images at a rate of 20 Hz and uses OpenCV to search each image for all tags by iteratively performing the following steps for each color tag:
1. Filter the image based on the color range of the tag (`inRange()`). The tag’s color range is extracted by initial calibration.
2. Filter noisy parts of the image that fall within the same color range as the tag by eroding areas smaller than $3\times3$ pixels (`erode()`). By this step, the only remaining white area within the image is the tag we are trying to localize.
3. Detect contours using [@suzuki1985topological] to locate the tag within the image (`findContours()`). Pixel coordinates can then be scaled to actual x,y coordinates.
CPS Implementation
------------------
Our initial simulations results for the coverage algorithm were implemented on Matlab to benchmark the performance of the coverage algorithm using a different set of parameters (e.g. number of targets, number of mobile cameras and targets distribution). The Matlab code was used with a minor modification of obtainning its input from the localization module and sending its output to the UAV controller through sockets.
UAV Control {#uav-control}
-----------
To simplify the case study, all UAV controlling operations are off-loaded to the central node. The central node uses AR Drone API v2.1. to control the motion of the UAVs. The SDK provides the following command `ardrone_tool_set_ui` `_pad_start` for take off and landing and `ardrone_tool_set` `_progressive_cmd` for flying and hovering by controlling the UAVs yaw, gaz, roll and pitch. The controller is implemented following the motion methodology presented in Section \[sec:control\].
Experimental Scenario
---------------------
![The top left corner is the view from the UAV while the bottom left corner is the view from the master camera. Targets are circled in red when they appear in the view of either cameras.[]{data-label="fig:sample_run"}](figures/s0_1.png){width=".49\linewidth"}
![The top left corner is the view from the UAV while the bottom left corner is the view from the master camera. Targets are circled in red when they appear in the view of either cameras.[]{data-label="fig:sample_run"}](figures/s1_1.png){width=".49\linewidth"}
![The top left corner is the view from the UAV while the bottom left corner is the view from the master camera. Targets are circled in red when they appear in the view of either cameras.[]{data-label="fig:sample_run"}](figures/s2_1.png){width=".49\linewidth"}
![The top left corner is the view from the UAV while the bottom left corner is the view from the master camera. Targets are circled in red when they appear in the view of either cameras.[]{data-label="fig:sample_run"}](figures/s3_1.png){width=".495\linewidth"}
The area of deployment was 1.25 by 2.1 meters (4.1 by 6.8 feet). Figure \[fig:sample\_run\] show snapshots of an *UnA*. The scenario of this experiment is to first have the two targets (circled in red) positioned in a certain configuration and have the UAV cover them. Then, move the targets and make the UAV reposition itself to cover the targets.
The first two snapshots show the UAV moving itself to cover the targets in their initial configuration. The third snapshot shows the UAV covering both targets (i.e. both targets are within the view of the drone’s camera). The fourth snapshot shows the UAV after moving to cover the drones on their second configuration.
Limitations and Future Work {#sec:conc}
===========================
![CDF of distance error in drone placement.[]{data-label="fig:error"}](figures/error_cdf.pdf){width="30.00000%"}
*UnA* reduces the cost and overhead of CPS experimental deployments, however these advantages does not come without some limitations. One of the major limitations of *UnA* is the accuracy of the UAV placement. Errors in UAV control and maneuvering inherently appear due to inaccuracies in computer vision based localization system and assisted maneuvers provided by the AR Drone’s firmware. We benchmarked the *UnA*’s placement error by placing the drone arbitrarily and then moving it to a certain location over 14 times. Figure \[fig:error\] shows the CDF of the error in placing the drone (i.e. the distance between the drone’s actual location and its intended location). The effects of these placement inaccuracies are significantly increased due to the small size of the deployment area. Nonetheless, the current state of *UnA* could be useful for many applications.
Our future plans for *UnA* include larger deployments with two or three drones. We also plan to deploy Click Modular router on the UAVs to allow for a more extensive evaluation of communication protocols between the UAVs. Furthermore, we plan to extend the “physical” components abstraction to allow the seamless integration of several other mechnical systems (e.g. ground vehicles). We also plan to explore other approaches for UAV tracking (e.g. RFID tracking).
[^1]: An AR Drone 2.0 costs \$300 while an AscTec Hummingbird costs around \$4500 .
[^2]: While multiple color tags are used in [@oh2011indoor] to track the drone’s location and orientation, our proposed approach uses a single tag to track the location and the internal campus of the drone to detect the drone’s orientation.
[^3]: While PTZ cameras can cost several thousand dollars, MCS costs up to tens of thousands of dollars.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'With the increased complexity of modern computer attacks, there is a need for defenders not only to *detect* malicious activity as it happens, but also to *predict* the specific steps that will be taken by an adversary when performing an attack. However this is still an open research problem, and previous research in predicting malicious events only looked at binary outcomes ([e.g., ]{}whether an attack would happen or not), but not at the specific steps that an attacker would undertake. To fill this gap we present , a system that leverages Recurrent Neural Networks (RNNs) to predict future events on a machine, based on previous observations. We test on a dataset of 3.4 billion security events collected from a commercial intrusion prevention system, and show that our approach is effective in predicting the next event that will occur on a machine with a precision of up to 0.93. We also show that the models learned by are reasonably stable over time, and provide a mechanism that can identify sudden drops in precision and trigger a retraining of the system. Finally, we show that the long-term memory typical of RNNs is key in performing event prediction, rendering simpler methods not up to the task.'
author:
- 'Yun Shen$^\vardiamondsuit$, Enrico Mariconti$^\clubsuit$, Pierre-Antoine Vervier$^\vardiamondsuit$, and Gianluca Stringhini$^{\spadesuit\clubsuit}$'
bibliography:
- 'bibliography.bib'
title: ': Predicting Security Events Through Deep Learning'
---
Introduction {#sec:introduction}
============
The techniques used by adversaries to attack computer systems and networks have reached an unprecedented sophistication. Attackers use multiple steps to reach their targets [@chen2014study; @stringhini2015ain] and these steps are of heterogeneous nature, from sending spearphishing emails containing malicious attachments [@le2014look], to performing drive-by download attacks that exploit vulnerabilities in Web browsers [@cova2010detection; @provos2007ghost], to privilege escalation exploits [@provos2003preventing]. After the compromise, miscreants can monetize their malware infections in a number of ways, from remotely controlling the infected computers to stealing sensitive information [@farinholt2017catch; @stone2009your] to encrypting the victim’s data and holding it hostage [@kharraz2015cutting; @kolodenker2017paybreak].
Traditionally, the computer security community has focused on *detecting* attacks by using a number of statistical techniques [@cova2010detection; @ho2017detecting; @gu2007bothunter; @kruegel2003anomaly; @stringhini2017marmite; @warrender1999detecting]. While this is inherently an arms race, detection systems provide the foundation for network and system defense, and are therefore very important in the fight against network attacks. More recently, the attention of the community switched to *predicting* malicious events. Recent work focused on predicting whether a data breach would happen [@liu2015cloudy], whether hosts would get infected with malware [@leyla2017riskteller], whether a vulnerability would start being exploited in the wild [@sabottke2015vulnerability], and whether a website would be compromised in the future [@soska2014automatically]. These approaches learn the attack history from previous events ([e.g., ]{}historical compromise data) and use the acquired knowledge to predict future ones. Being able to predict whether an attack will happen or not can be useful in a number of ways. This can for example inform law enforcement on the next target that will be chosen by cybercriminals, enable cyber insurance underwriters to assess a company’s future security posture, or assist website administrators to prioritize patching tasks.
While useful to predictively assess risk in systems and organizations, existing prediction systems have two main limitations. First, they only focus on predicting events with a binary outcome ([e.g., ]{}whether an attack or a data breach will happen) but do not provide any insights on the techniques and the modus operandi that will be followed by attackers. Second, these systems need labeled data to build a model and be able to make a decision, and it is not always easy to obtain such labeled data at a scale that allows to train an accurate model. One additional issue is that attackers operate changes to their modus operandi over time, and therefore both feature engineering and binary detection systems themselves need to be updated and retrained to keep performing accurate detections [@mariconti2016mamadroid].
In this paper, we go beyond predicting a binary outcome on security events, but we rather want to predict the *exact actions* that will be taken by an attacker when performing a computer attack. To this end, we leverage recent advances in the area of deep learning to develop , a system that learns from past system events and can predict the specific event that will happen next. can provide much more precise predictive information compared to previous work, allowing companies to deploy truly proactive countermeasures based on the predicted information. For example, this system could predict which particular CVE will be used by an attacker when mounting a multi-step attack against a server, or assess the potential severity of an attack by only looking at its early steps.
{width="\linewidth"}
We test on a dataset of 3.4 billion security events collected by a commercial intrusion prevention system from a daily population of 740k machines over a period of 27 days. We show that is effective in predicting the specific malicious event that will occur with a precision of up to 0.93. We also show that the models trained by are fairly stable, with the detection performance changing slowly. We however identify sudden drops in prediction precision, due to unexpected changes in the observed attacks and systems (e.g., new vulnerabilities being discovered or old ones being patched). In these cases, we show that can automatically identify the issue and trigger a retraining of the system. We demonstrate that the long-term memory provided by Recurrent Neural Networks is key to performing an accurate prediction when dealing with complex multi-step attacks and the noise generated by legitimate or non-related malicious events in the wild; in fact, we show that clearly outperforms simpler systems that are only based on short-term memory ([e.g., ]{}Markov Chains).
Background and Motivation {#sec:motivation}
=========================
To illustrate the complexity of keeping track of events across different machines, consider the real-world example in Figure \[fig:endpoint\_sequences\]. We show three endpoints undergoing a coordinated attack to Apache Web servers (see Section \[sec:case\_studies\] for the detailed case study), where {$e_0, ..., e_{13}$} are events involved in this attack and are highlighted in bold. This coordinated attack consists of three parts: (i) run reconnaissance tasks if port `80/tcp` (HTTP) is open ([e.g., ]{}$e_4$ is default credential login, $e_5$ is Web server directory traversal), (ii) trigger a list of exploits against the Web application framework Struts ([e.g., ]{}$e_8$ is an exploit relating to Apache Struts CVE-2017-12611, $e_{11}$ is an attempt to use Apache Struts CVE-2017-5638, $e_{13}$ tries to exploit Apache Struts CVE-2017-9805, etc.) and (iii) execute a list of exploits against other possible applications running on the system ([e.g., ]{}$e_2$ exploits Wordpress arbitrary file download, $e_9$ targets Joomla local file inclusion vulnerability, etc).
The first challenge that we can immediately notice in Figure \[fig:endpoint\_sequences\] is that even though those three endpoints are going through the same type of attack, there is not an obvious pattern in which a certain event $e_i$ would follow or precede another event $e_j$ given $e_i, e_j \in \{e_0, ..., e_{13}\}$. For example, $e_{12}$ (Malicious OGNL Expression Upload) can be followed by $e_4$ (HTTP Apache Tomcat UTF-8 Dir Traversal) and $e_{13}$ (Apache Struts CVE-2017-9805) in $s_1$, yet, it is followed by $e_7$ (Wordpress Arbitrary File Download) and $e_{11}$ (Apache Struts CVE-2017-5638) in $s_2$.
The second challenge is that the endpoints may observe other security events not relating to the coordinated attack. For example, in $s_3$, we can observe a subsequence {$e_4, e_{41}, ..., e_{5}, e_{22}, e_{21}, e_7$} in which $e_5$ is followed by a number of unrelated events including $e_{41}$ (WifiCam Authentication Bypass) before reaching $e_5$. Note that the other noisy events are omitted for the sake of clarity. Between $e_5$ and $e_7$, there were two other noisy event $e_{22}$ (Novell ZENWorks Asset Management) and $e_{21}$ (ColdFusion Remote Code Exec).
More interestingly, some of these endpoints may potentially observe different attacks from various adversary groups happening at the same time. For example, we observe {$e_9, e_{19}, e_{24}, e_{25}, e_{26}, e_{12}$} in $s_1$, {$e_4, e_{19}, e_{30}, e_{25}, e_{24}, e_{31}, e_{12}$} in $s_2$, and {$e_6, e_{23}, e_{19}, e_{30}, e_{25}, e_{24}, e_{12}$} in $s_3$. It is possible that $e_{19}$ (SMB Validate Provider Callback CVE-2009-3103), $e_{25}$ (SMB Double Pulsar Ping), and $e_{24}$ (Microsoft SMB MS17-010 Disclosure Attempt) could be part of another coordinated attack. Facing these challenges, it is desirable to have a predictive model that is able to understand noisy events, recognize multiple attacks given different contexts in a given endpoint, and correctly forecast the upcoming security event. This is a more complex and difficult task than detecting each malicious event passively.
**Problem Formulation.** We formalize our security event prediction problem as follows. A security event $e_{j} \in E$ is a timestamped observation recorded at timestamp $j$, where $E$ denotes the set of all unique events and $|E|$ denotes the size of $E$. A security event sequence observed in an endpoint $s_i$ is a sequence of events ordered by their observation time, $s_i=\{e^{(i)}_{1}, e^{(i)}_{2}, ..., e^{(i)}_{n}\}$. We define the to-be-predicted event as *target event*, denoted as $e_{tgt}$. Each target event $e_{tgt}$ is associated with a number of already observed security events, denoted as $l$. The problem is to learn a sequence prediction function $f: \{e_{1}, e_{2}, ..., e_{l}\} \rightarrow e_{tgt}$ that accepts a *variable-length* input sequence $\{e_{1}, e_{2}, ..., e_{l}\}$ and predicts the target event $e_{tgt}$ for a given system. Note that our problem definition is a significant departure from previous approaches that accept only fixed-length input sequences. We believe that a predictive system should be capable of understanding and making predictions given *variable-length* event sequences as the contexts, hence our problem definition is a better formulation inline with real world scenarios.
Methodology {#sec:workflow}
===========
In this section we describe the system architecture and the technical details behind .
Architecture Overview {#sec:architecture}
---------------------
The architecture and workflow of is depicted in Figure \[fig:workflow\]. Its operation consists of four phases: data collection and preprocessing, model training & validation, security event prediction, and prediction performance monitoring.
{width="0.8\linewidth"}
**Data collection and preprocessing ().** takes as input a sequence of security events generated by endpoints ([e.g., ]{}computers that installed a security program). The goal of the data collection and preprocessing module is to prepare both the training and validation data to build ’ predictive model. then consumes that raw security event data generated from millions of machines that send back their activity reports. The collection and preprocessing module reconstructs the security events observed on a given machine $s_i$ as a sequence of events ordered by timestamps, in the format of $s_i=\{e^{(i)}_{1}, e^{(i)}_{2}, ..., e^{(i)}_{n}\}$. The output of the data collection and preprocessing module is $D = \{s_1, s_2, ..., s_m\}$ where $m$ denotes the number of machines. Finally, we build the training data $D_T$ and validation data $D_V$ from $D$ for the next stage, where $D_T \cap D_V = \emptyset$.
**Model training and validation.** The core of consists of the training of a recurrent neural network with recurrent memory cells (.a, see Section \[sec:lstm\_ma\_theory\] for technical details about recurrent memory cells). Essentially, specifies a probability distribution of $e_{w+1}$ possible events given historical observed events {$e_{1}, ..., e_{w}$}, where $w$ refers to the rollback window size, by applying an affine transformation to the hidden layer followed by a $softmax$,
$$Pr(e_{w+1} | e_{1:w}) = \frac{\exp(h^w \cdot p^j + q^j)}{\sum_{j' \in E} \exp(h^w \cdot p^{j'} + q^{j'}) }$$
where $p^j$ is the $j$-th column of output embedding $P \in R^{m \times |V|}$ and $q^j$ is a bias term. Given the training data $D_T$, ’ training objective is therefore to minimize the negative log-likelihood $\mathcal{L}$ of all the event sequences:
$$\mathcal{L} = - \sum_{t=1}^{|D_T|} Pr(e_{t} | e_{1:t-1} : \theta)$$
We use the validation data $D_V$ to verify if the parameters $\theta$ identified during the training phase can achieve reasonable prediction performance (.b). It is important to note that $D_T$ and $D_V$ come from different machines so as to verify the general prediction capability of on the endpoints that are not part of the training data. **Security event prediction ().** Once the model is trained, takes the historical events {$e_{0}$, ..., $e_{i}$} as the initial input ([i.e., ]{}a *variable-length* input sequence inline with the real-world scenario) and predicts the probability distribution of $e_{{i+1}}$ given $E$ as $Pr[e_{{i+1}} | e_{0:i}] = \{e_1 : p_1, ~ e_2 : p_2, ~ ..., ~ e_{|E|} : p_{|E|}\}$. Our strategy is to sort $Pr[e_{{i+1}} | e_{0:i}]$ and choose the event with the *maximum probabilistic score*. then verifies with the actual event sequence whether $e_{{i+1}}$ is the correct prediction. In case of a wrong prediction, updates its contextual information accordingly. Section \[sec:case\_studies\] provides a detailed case study of the security event prediction phase in a real-world scenario.
**Prediction performance monitoring ().** Finally, in an effort to maintain the prediction accuracy as high as possible, the prediction performance monitor tracks and reports the evolution of different metrics, such as the Precision, Recall, and F1 of the current model. It is possible to elaborate such metrics on ’ implementation in the wild as it is immediately possible to see whether predicted the right event. If the predictions precision is dropping below a certain threshold, the system would automatically understand that is necessary to retrain the model.
Recurrent Memory Array {#sec:lstm_ma_theory}
----------------------
Long short-term memory (LSTM) and variants such as gated recurrent units (GRU) are the most popular recurrent neural network models for sequential tasks, such as in character-level language modeling [@kim2016character]. One common approach to deal with complex sequential data is using a stacked RNNs architecture. Essentially, stacking RNNs creates a multi-layer feedforward network at each time-step, [i.e., ]{}the input to a layer being the output of the previous layer. In turn, stacking RNNs automatically creates different time scales at different levels, and therefore a temporal hierarchy [@hermans2013training]. This approach has been proven practical and achieving good accuracy in various cases, such as log prediction [@du2017deeplog], binary function recognition [@shin2015recognizing], and function type recovery [@chua2017neural]. Nevertheless, despite the proven success of stacked RNNs, one complication incurred by such strategy is the lack of generalization to new data, [e.g., ]{}stacking mechanisms chosen and tuned for current training data require vigorous evaluation and may not adapt well to the new data at run time [@zhang2016architectural]. Therefore, rather than stacking multiple layers of RNNs, it would be ideal to build more complex memory structures inside a RNN cell to retain temporal memories while keeping a single layer RNN network to maintain computational efficiency when training. To achieve both goals, we propose to leverage the recurrent memory array by Rocki [@rocki2016recurrent]; this is doable by modifying LSTM architectures, while it is not available on GRU architectures.
Following the notation in Rocki [@rocki2016recurrent], we can formally define the recurrent memory array as follows in Eq. \[lstm\_ma\_1\].
$$\begin{aligned}
\label{lstm_ma_1}
\nonumber
f_k^t & = & \sigma(W_{fk}x^t + U_{fk}h^{t-1} + b_{fk})\\ \nonumber
i_k^t & = & \sigma(W_{ik}x^t + U_{ik}h^{t-1} + b_{ik})\\ \nonumber
o_k^t & = & \sigma(W_{ok}x^t + U_{ok}h^{t-1} + b_{ok})\\ \nonumber
\tilde{c}_k^t & = & tanh(W_{ck}x^t + U_{ck}h^{t-1} + b_{ck}) \\ \nonumber
c_k^t & = & f_k^t \odot c_k^{t-1} + i_k^t \odot \tilde{c}_k^t \\
h^t & = & \sum_{k} o_k^t \odot tanh(c_k^t)\end{aligned}$$
where $f$ denotes forget gates, $i$ denotes inputs, $o$ denotes outputs, $c$ denotes cell states, and $h$ denotes the hidden states. Here, $\odot$ represents element-wise multiplication. It is straightforward to notice that parameter $k$ directly controls the number of cell memory vectors, which enables the recurrent memory array to build an array-like structure similar to the structure of the cerebellar cortex [@rocki2016recurrent].
To deal with noisy sequential input data (Section \[sec:dataset\]) as observed in the real world, we follow the **stochastic** design outlined in [@rocki2016recurrent] by treating initial output gate activations as inputs to a $softmax$ output distribution, sampling from this distribution, and selecting the most likely memory cell to activate (see Eq. \[lstm\_ma\_2\]).
$$\begin{aligned}
\label{lstm_ma_2}
\nonumber
p(i=k) & = & \frac{e^{o_k^t}}{\sum_k o_k^t} \\
h^t & = & o_i^t \odot tanh(c_i^t)\end{aligned}$$
Eq. \[lstm\_ma\_2\] identifies the probability of a memory cell $i$ to be activated and update $h^t$ accordingly using this cell while the rest of memory cells are deactivated. Hence, instead of summarizing all cell memory (see Eq. \[lstm\_ma\_1\]), only one output is used in this stochastic design that is resilient to noisy input. We refer interested readers to [@rocki2016recurrent] for theoretical proofs and empirical comparison studies with the other state-of-the-art RNN architectures.
Datasets {#sec:dataset}
========
![Summary of the security event datasets used in this paper. []{data-label="fig:ipsping_dataset"}](dataset_visualisation){width="0.8\linewidth"}
is a generic system that can be used to predict security events on different protection systems. To evaluate its performance in this paper, we focus specifically on security event data collected from Symantec’s intrusion prevention product. Symantec offers end users to explicitly opt in to its data sharing program to help improving its detection capabilities. To preserve the anonymity of users, client identifiers are anonymized and it is not possible to link the collected data back to the users that originated it. Meta-information associated with a security event is recorded when the product detects network-level or system-level activity that matches a pre-defined signature. From this data we extract the following information: anonymized machine ID, timestamp, security event ID, event description, system actions, and other information (e.g., file SHA2) if any. Note that we use the anonymized machine ID to reconstruct a series of security events detected in a given machine and discard it after the reconstruction process is done.
To thoroughly investigate the effectiveness, stability and reusability of , we collected 27 days of data, summarized in Figure \[fig:ipsping\_dataset\]. We compile two separate datasets. The first one, which we call $\mathbf{D1}$, spans a period of 17 days in November 2017 (1 November - 17 November), and is composed of over 2.2 billion security events. We use the first five days (1 November - 5 November) of $\mathbf{D1}$ to validate our approach and for a comparison study against three baseline methods (see Section \[sec:evaluation\]). We later use the first seven days (1 November - 7 November) of $\mathbf{D1}$ to build models with varied length of training period, and study the stability of our approach by evaluating the prediction accuracy for the rest of the 10 days of data (8 November - 17 November) from $\mathbf{D1}$. We also compile another dataset, which we call $\mathbf{D2}$. This dataset is composed of 1.2 billion security events collected on the 8th and 23rd day of each month between November 2017 and February 2018, and the first three days in January 2018. $\mathbf{D2}$ is used to understand whether the system retains its accuracy even in a longer term scale: training sets based on $\mathbf{D1}$ are months older than part of the data in $\mathbf{D2}$. We use the first three days in January 2018 to build new models and compare them to the models built with data from $\mathbf{D1}$ (1 November - 7 November) and study their prediction performance with a focus on ’ reusability (see Section \[sec:exp4\]). On average, we collect 131 million security events from 740k machines per day, roughly 176 security events per machine per day. In total, the monitored machines generated 4,495 unique security events over the 27 day observation period.
**Data Limitations.** It is important to note that the security event data is collected passively. That is, these security events are recorded only when corresponding signatures are triggered. Any events preemptively blocked by other security products cannot be observed. Additionally, any events that did not match the predefined signatures are also not observed. Hence the prediction model used in this paper can only predict the events observed by Symantec’s intrusion prevention product. We discuss more details on the limitations underlying the data used by in Section \[sec:limitations\].
Evaluation {#sec:evaluation}
==========
In this section we describe the experiments operated to evaluate . We designed a number of experiments that allow us to answer the following research questions:
- What is performance in identifying the upcoming security event (Section \[sec:overall\_prediction\]) and how does its performance compare to the baseline and state-of-the-art methods (Section \[sec:comparison\])?
- How do variations in the model’s training period affect the performance (Section \[sec:exp3\])?
- Can we reuse a trained model for a given period of time and when do we need to retrain the model (Section \[sec:exp4\])?
- What is the influence of the long-term memory of Recurrent Neural Network models to achieve security event prediction (Section \[sec:seqlen\])?
Experimental Setup {#sec:implementation}
------------------
**Implementation.** We implemented in Python 2.7 and TensorFlow 1.4.1. Experimentally, we set the number of unrolling $w$ to 20, the training batch size to 128, the number of memory array $k$ (see Section \[sec:lstm\_ma\_theory\]) to 4 and the number of hidden LSTM Memory Array units to 128. We find these parameters offering the best prediction performance given our dataset. All experiments were conducted on a server with 4 TITAN X (Pascal) 12GB 1.5G GPUs with the CUDA 8.0 toolkit installed. All baseline methods are implemented in Python 2.7 and experimented on a server with a 2.67GHz Xeon CPU X5650 and 128GB of memory.
**Evaluation setup.** To form a concrete evaluation setup, for both and other baseline methods experiments, we split the input data and use 80% for training, 10% for validation, and 10% for test. We strictly require that training, validation and test data to come from different machines so as to verify ’ general prediction capability in the endpoints that are not part of the training data. Specifically, we train for 100 epochs, validate model performance after *every* epoch and select the model that provides the best performance on validation data.
**Evaluation metrics.** We use the precision, recall, and F1 metrics to evaluate prediction results from the models. In our experimental setup, we calculate these metrics globally by counting the total true positives, false negatives and false positives. It is important to note that accepts *variable-length* security event sequences. We specially hold out the *last* event as the prediction target $e_{tgt}$ for evaluation purposes. Section \[sec:case\_studies\] showcases how can be leveraged to accomplish step-by-step prediction with a single event as the initial input.
Overall Prediction Results {#sec:overall_prediction}
--------------------------
![Experimental setup for ’ prediction evaluation (Section \[sec:overall\_prediction\]) and comparison study with baseline methods (Section \[sec:comparison\]).[]{data-label="fig:exp1_evaluation"}](exp1_visualisation){width="0.7\linewidth"}
In this section we evaluate the performance of our security event forecast model in predicting the exact upcoming event. This is a challenging task that a predictive system for security events aims at resolving due to the fact that there are 4,495 security events as possible candidates in our dataset (see Section \[sec:dataset\]) and an exact event should be correctly predicted.
[0.3]{}
[0.3]{}
[0.3]{}
**Experiment setup**. We use the experimental setup as illustrated in Figure \[fig:exp1\_evaluation\] for ’ performance evaluation. From $\mathbf{D1}$, we train our predictive model using **one day** of data and evaluate on both the same day and the following days until 5 November 2017. For example, we train using data from 2 November and evaluate its prediction performance from 2 November to 5 November.
**Experiment results.** Following our general experimental evaluation setup, we randomly select $14,396$ machines from the first days of November that are not part of the data used in the training set of the initial model. We focus on predicting the last event occurring on a machine given the sequence of previously-observed events. As shown in Figure \[fig:NextDays\], is able to achieve over 80% precision, recall, and F1-measure in predicting the exact upcoming security event when evaluating on the same day test data. Figure \[fig:NextDays\] shows that, when training on one day, and testing on the same day and the following ones, the values of the Precision, Recall and F1 do not decrease dramatically. When it does, it decreases, in the worse case, of less than $0.05$. The Figure also shows that Precision (Figure \[fig:NextDays\]a) and Recall (Figure \[fig:NextDays\]b) are well balanced and have very similar values and exactly the same scale (from $0.87$ to $0.795$). This result shows that can offer good prediction results. There is no security event dominating in our dataset, which may lead to biased but better prediction performance. The top 3 events in our training data are: (i) Microsoft SMB MS17-010 Disclosure Attempt (19.8%), (ii) SMB Double Pulsar Ping (16.4%), and (iii) Unimplemented Trans2 Subcommand (16.1%). The top 3 events in our test data are ranked as follows: (i) Microsoft SMB MS17-010 Disclosure Attempt (9.85%), (ii) HTTP PE Download (6.3%), and (iii) DNS Lookup Failures (3.5%). Interestingly, the dominant events in training and test are different, which makes ’ prediction results even stronger.
Over the days, we observe a trend that the prediction performance of drops slightly in terms of all three evaluation metrics. Take the model trained on 2 November for example, its prediction precision drops by 4% from 0.83 to 0.79. In Section \[sec:exp3\] we study if variations ([e.g., ]{}a longer training period) in the model’s training data would offer better performance and how stable the trained performs over consecutive days. Note that ‘micro’-averaging in a multi-class setting produces equal Precision, Recall and F1-Measure. For the rest of the evaluation process, we therefore use precision as the main evaluation metric.
Comparison Study {#sec:comparison}
----------------
[|c|c|c|c|c|c|]{} &\
**Method** & **01/Nov** & **02/Nov** & **03/Nov** & **04/Nov** & **05/Nov**\
Spectral & 0.05 & 0.031 & 0.023 & 0.013 & 0.02\
--------------
Markov Chain
--------------
: Prediction precision comparison study: vs. baseline approaches.[]{data-label="tab:sameday"}
& 0.62 & 0.56 & 0.56 & 0.53 & 0.52\
3-gram & 0.67 & 0.54 & 0.61 & 0.592 & 0.601\
& **0.83** & **0.82** & **0.83** & **0.82** & **0.81**\
In this section we aim at studying whether the higher complexity of Recurrent Neural Networks is required for the task of predicting security events, or whether simpler baseline methods would be enough for the task at hand. For comparison purposes, we implemented *first-order* Markov Chain [@norris1998markov] and 3-gram model [@brown1992class] (equivalent to the second order Markov Chain model) in Python 2.7.1. Note that it is natural to consider a higher order ([e.g., ]{}n-order where $n > 2$) Markov Chain model for security event prediction, however, due to the exponential states issue associated with high order Markov Chain models, it is computationally costly to build such a high order model for 4,495 events. Finally we use the sp2learning[^1] package to build a spectral learning model [@kamvar2003spectral] for sequence prediction as the third baseline prediction model. These three methods are often used to model sequences of elements in several fields and, being simpler than our RNN models and widely used in sequence prediction, they are relevantly good baselines to compare with.
**Experiment setup.** The comparison study uses daily data (1 November - 5 November) from $\mathbf{D1}$. To evaluate in this case, all training, validation, and test data come from the same day.
**Comparison study results.** Table \[tab:sameday\] shows the precision of compared to simpler systems. Table \[tab:sameday\] shows that outperforms the baseline methods but also that 3-grams perform better than Markov Chains, and Markov Chains perform better than the spectral learning method. This particular order shows the importance of sequence memory as the system that performs best among the baselines is the 3-grams. However, 3-grams are less effective than . This is due to two of the main characteristics of neural networks: the capacity of filtering noise and the longer term memory. As Table \[tab:sameday\] shows, has precision values higher than 0.8 in all the five days of tests showing a very good level of reliability. In Section \[sec:seqlen\] we show that the long-term memory that is an important feature of RNNs plays a key role in correctly predicting security events. Note that we didn’t report a comparison of the computation time among the methods due to the fact that leverages GPUs to train RNN models and the baseline methods rely on traditional CPUs, and therefore is in general much quicker to run. For example, our 3-gram implementation took over 10 days for training, yet requires only $\sim$10 minutes per epoch using GPUs.
Influence of Training Period Length {#sec:exp3}
-----------------------------------
![Experimental setup for multiple day evaluation of (Section \[sec:exp3\]).[]{data-label="fig:exp3_evaluation"}](exp3_visualisation){width="0.85\linewidth"}
In this section we look at whether training on longer periods of time achieves better prediction performance.
**Experiment setup.** We use the experimental setup as illustrated in Figure \[fig:exp3\_evaluation\] for ’ performance evaluation. From $\mathbf{D1}$, we train our predictive model using **one day** of data and evaluate on the test data from 8 November to 17 November. For example, we train using data from 2 November and evaluate its prediction performance on test data from 8 November to 17 November. To evaluate if a longer training period can offer better prediction performance, we also train our predictive model using **one week** of data (from 1 November to 7 November) and evaluate its performance in the aforementioned period.
**Experiment results.** In this experiment we evaluate the performance of our security event forecast model in predicting the exact upcoming event several days after the initial model was trained. The goal is to determine how well our predictive model ages in the short term and to make sure that it remains effective in predicting security events without the need to re-tune it after this period of time.
The question that this experiment is trying to answer is whether there is a difference in training the models over longer periods of time, such as one week, rather than one day. Table \[tab:NovTests\] provides some insights into this question. First, we used the first **five** days of November on their own to build five models. Second, we built one single model from the first **seven days** of the same month. We then tested the six different models (five based on one day of data and one based on one week of data) on **ten days** of data from 8 November to 17 November. *Overall, the training over one week of data produces similar results as those obtained using training over only one day of data.* On average, trained with one week data can achieve a precision score of 0.819, which is 0.3% higher than that of the models trained with one day data.
These results demonstrate that can offer good accuracy with stable performance over time since the standard deviation of precision scores over the measurement period of 10 days is small ($\sim$0.02). However, on 8 November and on 16 November the results are slightly different, exhibiting a higher accuracy for the week-long trained model. While in the first case (8 November) it is probably due to the proximity of the test day to the training week, the second case (16 November) appears to be an outlier. We further observe that the time proximity of the training and test data appears to have a positive impact on the prediction accuracy. Indeed, we can see that the model trained over one day of data is as efficient as the one trained over one week of data when tested on alerts generated only a few days later, probably due to the similarity among attack behaviors observed within a few days. The week-long trained model appears to be more efficient in the presence of deviating, or outlying attack behaviors in the test phase. This can easily be explained by the fact that the more data is used to build a model the more complete the model is. Hence it can better deal with rare events or deviating attack behaviors.
One of the reasons why ’ prediction precision might suddenly decrease is if the set of alerts significantly changes from a day to another, for example because a new vulnerability starts being exploited in the wild, a system patch fixes an existing one, or a major version of a popular software gets released. For this reason, in our architecture discussed in Section \[sec:workflow\] we included a component that monitors the performance of and can trigger a re-training of the system if it is deemed necessary. In the experiment discussed in Table \[tab:NovTests\], for example, the precision performance on 16 November drops by 6.9% on average from 8 November. This could indicate to the operator that something significant changed in the monitored systems and that needs to be retrained. As we will show in Section \[sec:perform\], this can be done in batch and it takes well less than a day to complete.
Stability Over Time {#sec:exp4}
-------------------
![Experimental setup for reliability evaluation (Section \[sec:exp4\]).[]{data-label="fig:exp4_evaluation"}](exp4_visualisation){width="0.9\linewidth"}
In this Section we evaluate ’ prediction accuracy when the training data is several months older than the test data. Our goal is to evaluate the reliability of the model in case there is no retraining for several months. As we discussed, is able to detect when it needs to be retrained, however this operation does not come for free and therefore it is desirable to minimize it as much as possible.
**Experiment setup.** The experimental setup is illustrated in Figure \[fig:exp4\_evaluation\]. We train our predictive model using both **one day** of data (from 1 November to 5 November respectively) and **one week** of data (from 1 November to 7 November) from $\mathbf{D1}$. Additionally for comparison purposes, we train three more predictive models using **one day** of data (from 1 January to 3 January respectively) from $\mathbf{D2}$. The test data consists of two days per month (on the 8th and the 23rd) so as to obtain a representative dataset from November 2017 until February 2018.
**Experiment results.** Table \[tab:MonthsTests\] shows the results obtained using the same training sets as in the previous Section augmented with three days in January, [i.e., ]{}one day-long model for each of the first five days of November 2017, one week-long model for the first seven days of the same month and one day-long model for each of the first three days of January 2018. The prediction precision results presented in Table \[tab:MonthsTests\] show consistency through the different training sets and a good level of stability, as the performance does not decrease dramatically over time. Moreover, the week-long training set does not show increased accuracy compared to the day-long ones. These new results thus confirm those from Section \[sec:overall\_prediction\] and show that *(i) the model quickly converges towards high accuracy with only one or a few days of training data, and (ii) the model ages very well even months after it was built*.
*December discontinuity.* Table \[tab:MonthsTests\] shows a particular behavior between 8 December and 23 December: ’ precision increases. We would normally expect the system’s precision to slightly decrease over time, possibly following a pattern, while in this case the precision increases. To investigate this phenomenon, we looked for potential differences in the raw data and noticed that the test data collected after 8 December exhibits a significant deviation with respect to one specific security event ID: the presence of one of the top three recorded alarms decreased by an order of magnitude, having a comparable number of occurrences to alerts occupying the 4th to 10th position. The alarm is related to DoublePulsar, a vulnerability disclosed in the first half of 2017. Such change may be due to different reasons. The most probable reason, however, could be the installation of patches: Microsoft releases monthly updates for Windows every 2nd Tuesday of the month ([e.g., ]{}12 December 2017) and many software- and hardware-related companies release patches immediately following Microsoft’s. Finally, a small change to the IPS signatures or to the attack modus operandi can heavily impact the hit rate of a given alarm.
*Comparison study.* To further investigate this December discontinuity phenomenon we decided to assess the impact of the training data on the model accuracy. To this end, we considered the training sets from data collected on the first three days of January and tested on the January and February dates (bottom part of Table \[tab:MonthsTests\]). We can see that trained in January performs slightly better than when trained in November. These results show that the results by remain reliable even months after the system was trained. Nevertheless, in the case of a sudden decrease in precision due to an adverse change in the data (e.g., the emergence of a new attack), would be able to detect this and prompt a retraining, as discussed in Section \[sec:workflow\].
Sequence Length Evaluation {#sec:seqlen}
--------------------------
In Section \[sec:comparison\] we showed that outperforms simpler systems that do not take advantage of long-term memory in the same way as the RNN model used by our approach. In general, understanding how Deep Learning models work is challenging, and they are often treated as black boxes. To make matters more complex, RNNs do not only rely on long-term memory, but also on short-term memory, in particular to filter out noise.
In this section we aim at understanding whether long-term memory is more influential in making decision than short-term memory or vice versa. With relying on short-term memory we mean a system that relies on a few elements of the sequence to make its decision, that is, the ones closest to the element that the system is trying to guess. With relying on long-term memory we mean when the system uses the whole sequence or a large part of it to take its decision on what the next security event could be. Intuitively, if short-term memory was predominant, we would not expect the performance of to increase with the number of observed events.
As looking into the Neural Network weights may be a complicated way to understand which type of memory is more important for the model, we decided to focus on the occurrences of successfully and unsuccessfully guessed events. Every event guessed by has a probability (confidence score) associated to it. First, we look at the distribution of the confidence scores among successfully guessed events (Figure \[fig:SuccOcc\]) and unsuccessfully guessed ones (Figure \[fig:FailOcc\]). As it can be seen, both types of events present a very skewed distribution in their confidence scores, with a negligible number of events being predicted with a probability of less than 0.5.
[0.45]{} {width="\textwidth"}
[0.45]{} {width="\textwidth"}
[0.45]{} ![The plots show the percentage of the sequences correctly guessed (a) or failed to guess (b) with respect to sequences that share all the events but the last. On the X axis, as for Figures \[fig:SuccOcc\] and \[fig:FailOcc\] there is the confidence level of the sequences used by the system. Figures show that sequences of at least 5 events (sl >=5) are quite unique, therefore long term memory is a crucial factor in the system accuracy.](SuccessGuessCut.pdf "fig:"){width="\textwidth"}
[0.45]{} ![The plots show the percentage of the sequences correctly guessed (a) or failed to guess (b) with respect to sequences that share all the events but the last. On the X axis, as for Figures \[fig:SuccOcc\] and \[fig:FailOcc\] there is the confidence level of the sequences used by the system. Figures show that sequences of at least 5 events (sl >=5) are quite unique, therefore long term memory is a crucial factor in the system accuracy.](FailureGuessCut.pdf "fig:"){width="\textwidth"}
To better understand if ’ results are mainly due to the use of long-term rather than short-term memory, we checked how unique the sequences on which makes its decisions are. These quantitative results can hint at which kind of approach is used by the algorithm. We try to evaluate the occurrences of the sequence in which the system tried to guess the last event compared to all those that differ from it for the last event (the one that tried to guess). This analysis is carried out for sequences of length $i+1, (i=2,...,9)$ where $i$ represents the number of events before the last we take into consideration. For all the successful/unsuccessful sequences we calculate the ratio between the times in which we had those $i$ events and predicted the last event correctly and the times in which we had the same $i$ events followed by any event (included the right one). According to the probability value of the guessed event, we calculate the average probability.
Figure \[fig:SuccGuess\] shows the data for all the sequences for which the last event has been correctly guessed by . Note that the X axis starts at 0.5 because, as Figures \[fig:SuccOcc\] and \[fig:FailOcc\] showed, the number of predictions with a lower confidence is very low. The values for $i$ (sequence length) less than 5 show that the system’s prediction is not very confident. Longer sequences ($i$ greater than 5), instead, are more unique and often correctly predict the last event. The opposite happens when we evaluate the sequences involving events not guessed correctly by our system (Figure \[fig:FailGuess\]). In fact, the left part of the graph presents sequences where the wrongly guessed event was rarely the one following the $i$ previous events. This may mean that in those cases there are sequences that differ only for the last event and a few events are quite frequent.
**Takeaways.** Long sequences including the guessed event are more frequent when we analyze the successful guesses. This situation is more common than the unsuccessful guesses as the system reaches high accuracy. Therefore, according to the graphs the system seem to rely more on long-term memory than short-term memory.
Runtime Performance {#sec:perform}
-------------------
We now discuss the specific characteristics of the system and its runtime performance.
The training phase is the longest one: building a model is a long process that can be performed offline. takes around 10 hours to retrain the model. Considered the stability of the model, which as shown in Section \[sec:exp4\] can be effective for long periods of time, rebuilding the model does not have to be done every day. We also showed that it is possible to identify when the system needs retraining because of a discontinuity in the distribution of events (see Section \[sec:exp3\]). Once trained, takes 25ms to 80ms to predict the upcoming event using the variable-length security events in a given system.
’ predictive model trained using one day of data is about 31MB. It can be easily pushed to the endpoints with limited network footprint. Note that with the advance of deep learning libraries, especially the recent development of TensorFlow, it is feasible for to be deployed not only in traditional endpoints ([e.g., ]{}PCs) but in mobile and embedded devices as well. This is another aspect that exemplifies the general applicability of .
Case Studies {#sec:case_studies}
============
In this section we describe a set of case studies showcasing the capabilities of in different real-world scenarios. We first show how can be used to detect a coordinated multi-step attack against a Web server (Section \[sec:case1\]). We then provide a number of real-world settings in which ’ prediction labels can be modified to achieve specific goals, for example predicting entire classes of attacks (Section \[sec:case2\]).
Predicting events in a multi-step attack {#sec:case1}
----------------------------------------
The first scenario where ’ prediction capability can be leveraged is when facing multi-step, coordinated attacks, [i.e., ]{}a single attack involving multiple steps performed sequentially or in parallel by an attacker and resulting in multiple alerts being raised by the IPS. The difficulty of identifying such attacks originates from the fact that some of the intermediate steps of a multi-step attack can be considered benign when seen individually by an IPS engine. Moreover, most attacks observed in the wild are the result of automated scripts, which are essentially programmed to check for some precondition on the victim systems and subsequently trigger the adequate exploit(s). For instance, an attack might consists of the following steps: (i) run reconnaissance tasks if port `80/tcp` (HTTP) is open, (ii) trigger a list of exploits against the Web application framework, e.g., Apache and (iii) execute a list of exploits against other possible applications running on top of it. Therefore, we may not observe all steps of an attack on every victim system, depending on which branches of the attack scripts were executed. This variability of observed events across systems hinders the identification of the global multi-step attack.
To identify candidate multi-step attacks in our dataset of IPS events we used the following approach. For each event $e_i$ observed on any of the monitored machines we compute its frequency of occurrence across all machines. We then consider a candidate multi-step attack any sequence of events $e_i$ occurring at the same frequency with an error margin of $10$% to capture the variability in attacks as explained above. We also set a *support* threshold on the number of machines exhibiting that sequence so as to avoid a biased frequency obtained from too few samples. To uncover the case study presented here we empirically set this threshold to $1000$ machines. For network-sourced attack steps, we also extract the source IP address to determine the likelihood of the global event sequence to be generated by a single attacker.
We present the case of a multi-step attack captured by the IPS and which was successfully learned by the prediction tool. The attack consists of multiple attempts to exploit a Web server and the Web application running on top of it. First, the attacker checks the Web server software for several vulnerabilities. In this case, it quickly identifies the server as running Apache and then attempts to exploit several vulnerabilities, such as Struts-related vulnerabilities. It then switches to checking the presence of a Web application running on top and then fingerprinting it. In this step, the attacker triggers different exploits against known vulnerabilities in various Content Management Systems, such as Wordpress, Drupal, Joomla. Eventually, the attack appears to fail as the various steps are individually blocked by the IPS.
[0.67]{}
[0.67]{}
To be able to visualize the decision process and explain how operates given the aforementioned multi-step attack, we feed a list of security events from a machine that was under the coordinated attack. By putting in this real-world environment, we are able to visualize how predicts the upcoming events as illustrated in Figure \[fig:model\_prediction\_visualisation\]. Note that events {$e_0$, ..., $e_{13}$} belong to the coordinated attack.
The process operates as follows. Take $machine_1$ in Figure \[fig:model\_prediction\_visualisation\]a for example, takes event $e_{10}$ as the initial feed and predicts the upcoming event $e_{12}$. It then verifies with the actual event to check if $e_{12}$ is the correct prediction. In our case, $e_{12}$ is the correct prediction with a confidence score higher than 0.5 (therefore $e_{12}$ in a green box in Figure \[fig:model\_prediction\_visualisation\]a). automatically leverages both $e_{10}$ and $e_{12}$ as “contextual information” to enable its internal memory array cells to better predict the next event. The same step is repeated: $e_{8}$ is correctly predicted, uses $e_{10}, e_{12}, ~\text{and}~ e_{8}$ to prepare its internal cells. In the case that makes a wrong prediction, [e.g., ]{}it predicts $e_{10}$ instead of the actual observed event $e_5$ ($e_{10}$ is enclosed in a red box in Figure \[fig:model\_prediction\_visualisation\]a), uses the actual event $e_5$ together with previous historical events. This enables to stay on track with the observed events and predict events that are closely relating to those of the coordinated attack. This may lead to incorrectly predict some random attacks the system experienced. For example, $e_{55}$ is a ‘PHP shell command execution’ attack which was observed in the security event sequences, but not part of this particular coordinated attack. It is important to notice that is able to correctly predict $e_{13}$ (an attack relating to CVE-2017-9805) that was not presented in the previous events, even thought its predecessors, such as $e_{12}$, $e_{3}$, $e_{10}$, appeared multiple times. We consider this a good example of using long-term memory to carry out the correct prediction as detailed in Section \[sec:seqlen\]. It is also interesting to see how adapts itself during its operation as shown in Figure \[fig:model\_prediction\_visualisation\]b (multiple $e_7$ and $e_6$ in dashed line rectangles). We can observe that did not correctly predict $e_6$ twice when given $e_7$ (see and ). Nevertheless, is capable of leveraging the contextual information ([i.e., ]{}the actual observed events) to rectify its behavior. As we can observe in , is able to make a borderline prediction and in makes a confident and correct prediction of $e_6$.
To further exemplify ’ sequential prediction capability in the above setup, we run it on 8 February (2018) test data using the model trained on 3 January (2018) and predict all upcoming events of 200 randomly selected machines (this effectively generates 32,391 sequences due to the step-by-step setup) with a precision of 94%, and against 8 December (2017) test data using a model trained on 4 November (2017) and obtain a precision of 80.89%. These experimental results provide additional evidence of ’ prediction capability in a real-world environment.
Adjusting the prediction granularity {#sec:case2}
------------------------------------
The goal of is primarily to accurately determine the next event that is going to occur on a given monitored system. In some cases multiple security events might share some common traits. For example, multiple IPS events can be used to describe different attacks against a particular software application, network protocol, etc. These shared traits can then be used to categorize such events. This categorization is specific to the security application that generated these events. Also, the categorization process undoubtedly results in a compressed and coarser-gained set of security events. In this section we describe several cases where such a categorization can be leveraged when the system fails to predict the exact security event but successfully predicts the exact traits, or categories of the attacks, such as the targeted network protocol and software application, or the attack type. To begin with, we extracted categories from the IPS signature labels and descriptions. These categories correspond to characteristics of attacks described by these signatures and are defined as follows. Whenever possible, we identify the *verdict* of the signature, the *severity* of the attack, the *type of attack*, [e.g., ]{}remote command execution (RCE), SQL injection, etc, the *targeted application*, if any, the *targeted network protocol*, if any, and whether the attack exploits a particular *CVE*. There is thus a one-to-many relationship between each signature and the categories it belongs to. We then uncover machines for which the categories of a mistakenly predicted event matches exactly the categories of the correct prediction. About $3.5$% of failed prediction results exhibit this pattern.
For our first example we consider a machine that was targeted by the Shellshock exploits targeting the Unix shell BASH. Several vulnerabilities were uncovered in the context of these infamous attacks, namely CVE-2014-6271, CVE-2014-6277, CVE-2014-6278, CVE-2014-7169, CVE-2014-7186 and CVE-2014-7187. These six vulnerabilities translate into six IPS signatures. These signatures all belong to the same categories, which include (i) `block`, (ii) `high`, (iii) `RCE`, (iv) `bash` and (v) `CVE`. These categories mean that the exploit attempt is meant to be (i) *blocked* because its potential security impact on the targeted machine is of (ii) *high* severity. This verdict is explained by the fact that, if successful, the exploit would enable the attacker to perform a *remote code execution (or RCE)* by exploiting a known *vulnerability* (with an assigned CVE identifier) against the Unix shell *BASH*. In this case study, a machine was targeted by several of the Shellshock exploits. After observing an attempt to exploit CVE-2014-6271, the system predicted another attempt to exploit the same vulnerability, instead of the correct prediction for CVE-2014-6278. While the event-level prediction result is wrong, the category-level prediction successfully identify an attempt to exploit a Shellshock vulnerability.
The second example is related to a machine that has apparently visited or have been triggered to visit a website distributing fake anti-virus software. Several IPS signatures have been defined to capture different aspects of these malicious websites, for instance, regular expressions matching specific HTML content, suspicious JavaScript code, etc. In this example, the system predicted that the machine would be redirected to a fake AV website containing a particular piece of malicious HTML code while in reality, the machine was redirected to a fake AV website containing a malicious piece of JavaScript code.
Additionally, we evaluate the performance of our security event forecast model in predicting if the upcoming event should be blocked or allowed, a relaxation as the aim is not to determine the exact event that will happen, but if it is a low-priority alarm (that is still allowed by the product we receive the data from) or if it is a high-priority one (that is blocked immediately). This is one of the essential tasks that a predictive intrusion prevention system needs to resolve. Our experiment shows that the proposed predictive model is able to achieve 88.9% precision in predicting if the upcoming event should be blocked or allowed. This represents a 8% precision increase comparing to the exact event prediction on the same day. Nevertheless, The added value of adjusting the prediction granularity obviously depends on the accuracy of the categorization and the expected level of granularity of the prediction.
Discussion {#sec:limitations}
==========
**Limitations of .** A recurrent neural network, broadly speaking, is a statistical model. The more the model “sees” (i.e., the more training data) the better the prediction performance is. For rare events, since the model does not have enough training samples, may not correctly predict these rare intrusion attempts. Existing statistical and machine learning methods are yet to offer a satisfactory solution to this problem [@weiss1998learning; @wu2004learning; @kwon2012unified]. It would also be interesting to understand whether the recent work by Kaiser [*et al*.]{} [@kaiser2017learning] that makes deep models learn to remember rare words can be applied to predict rare intrusion attempts. DeepLog, a previously proposed system [@du2017deeplog], focused on anomaly detection in regulated environments, such as Hadoop and OpenStack, with limited variety of events. In such a specific log environment, DeepLog is able to use a small fraction of normal log entries to train and achieves good detection results. Nevertheless, DeepLog still requires a small fraction of normal log entries would generate enough representative events and patterns. Another limitation following rare events prediction is model retraining when new security events ([e.g., ]{}new signatures) are created. This retraining is inevitable because machine learning models can only recognize events they have been trained upon. Our experimental results (Section \[sec:comparison\]) show that takes around 10 hours to retrain and can be redeployed in a timely fashion in a real-world scenario. As mentioned in Section \[sec:seqlen\] the nodes that are activated in an LSTM are not easy to examine. For this reason we cannot guarantee that the system does not take into consideration spurious correlations. At the same time we tried to limit this issue by extensively evaluate over a large amount of data and in different settings.
**Data limitations.** For its operation, relies on a dataset of pre-labeled security events. An inherent limitation of this type of data is that an event can be labeled only if it belongs to a known attack class. If, for example, a new zero-day vulnerability started being exploited in the wild, this would not be reflected in the data until a signature is created for it. To reduce the window between when an attack is being run and when it starts being detected by an intrusion prevention system security companies typically use threat intelligence systems and employ human specialists to analyze unlabeled data looking for new attack trends.
**performance.** Sections \[sec:evaluation\] and \[sec:case\_studies\] show the effectiveness of the system. The prediction of a security event in such a complicated environment is an important challenge. shows the ability of effectively tackling this challenge, showing stability, even when the training set is months older than the test set, and robustness to noise while detecting multistage attacks. We evaluated over different time periods to thoroughly prove its qualities; as we discussed, the system may need retraining only in case the data presents radical changes, while its precision does not decrease quickly if the training set is older than the test set. The system can support different dimensions of the training set as it has been tested using one day or one week of data. The differences are minimal: performance is extremely similar, but weekly training seems slightly more robust to anomalies on a specific day of data. However, weekly training sets require more time to build the model. Sections \[sec:seqlen\] and \[sec:case\_studies\] show how long-term memory and noise filtering are both important factors in the precision of the neural networks, explaining why the baseline methods used in Section \[sec:comparison\] are less precise than .
**Deployment.** The architecture of enables it to be reasonably flexible in terms of real-world deployment. can be deployed for each endpoint to proactively defend against attacks as we can see in Section \[sec:case\_studies\]. At the same time, can be tailored to protect an enterprise by training with the data coming from that enterprise only and thus better deal with the attacks targeting that enterprise. Note that ’ predictive model trained using one day of data is about 31MB. It can be easily pushed to the endpoints with limited network footprint. Together with the mobile TensorFlow library, it is practically feasible for to protect mobile/embedded devices by training with security event data coming from those devices only. For example, can be trained using the data collected by smart routers with an IPS installed and deployed in these routers to protect smart home environments.
**Evasion.** may be subject to evasion techniques from malicious agents. A vulnerability of deep learning systems is that while the system is classifying samples, it adapts its rules. Therefore, it may be subject to poisoning attacks from a criminal who influences the decision rules using fake actions before attacking the victim. However, to achieve such evasion, the attackers must apply such fake actions at a massive scale and target thousands of machines. A technique that could be used by adversaries is mimicry attacks, [i.e., ]{}evading security systems by injecting many irrelevant events to cover the alerts generated by a real attack. We argue that has the potential to be resistant to these attacks. Indeed, we have seen in the case studies that is able to filter out the noise from the sequences of events observed on the machines, and detect the important events correctly. An interesting future work would be to be able to quantify the amount of events necessary to evade systems like . Zero day attacks may be difficult to detect: a zero day attack may replicate known sequences of actions to exploit new vulnerabilities, but that would still be detected; when the zero day is applying a new kind of multi-step attack that has a different sequence of events, it may not be detected.
Related Work {#sec:relatedwork}
============
In this section, we broadly reviewed previous literature in both forecasting security events and deep learning, especially recurrent neural networks (RNNs), applications in security analytics.
Security Event Forecast
-----------------------
**System-level security event forecast.** Soska [*et al*.]{} [@soska2014automatically] proposed a general approach to predict with high certainty if a given website will become malicious in the future. The core idea of this study is building a list of features that best characterize a website, including traffic statistics, filesystem structure, webpage structure & contents and statistics heuristic of dynamic features ([e.g., ]{}contents). These features are later used to train an ensemble of C4.5 classifiers. This model is able to achieve operate with 66% true positives and 17% false positives given one-year data. Bilge [*et al*.]{} [@leyla2017riskteller] proposed a system that analyzes binary appearance logs of machines to predict which machines are at risk of infection. The study extracts 89 features from individual machine’s file appearance logs to produce machine profile, then leverages both supervised and semi-supervised methods to predict which machines are at risk. In terms of machine wise infection prediction, RiskTeller can predict 95% of the to-be-infected machines with only 5% false positives; regarding enterprise-wise infection prediction, Riskteller can, on average, achieve 61% TPR with 5% FPR.
**Organization-level security event forecast.** Liu [*et al*.]{} [@liu2015cloudy] explored the effectiveness of forecasting security incidents. This study collected 258 externally measurable features about an organization’s network covering two main categories: mismanagement symptoms (e.g., misconfigured DNS) and malicious activities (e.g., spam, scanning activities originated from this organization’s network). Based on the data, the study trained and tested a Random Forest classifier on these features, and are able to achieve with 90% True Positive (TP) rate, 10% False Positive (FP) rate and an overall accuracy of 90% in forecasting security incidents. Liu [*et al*.]{} [@liu2015predicting] carried out a follow-up study on externally observed malicious activities associated with network entities (e.g., spam, phishing, malicious attacks). It further proved that when viewed collectively, these malicious activities are stable indicators of the general cleanness of a network and can be leveraged to build predictive models (e.g., using SVM). The study extracts three features: intensity, duration, and frequency, from this activity data. It later trained a SVM model using these features and achieved reasonably good prediction performance over a forecasting window of a few months achieving 62% true positive rate with 20% false positive rate.
**Cyber-level security event forecast.** Sabottke [*et al*.]{} [@sabottke2015vulnerability] conducted a quantitative and qualitative exploration of the vulnerability-related information disseminated on Twitter. Based on the analytical results, the study designed a Twitter-based exploit detector, building on top 4 categories of features (Twitter Text, Twitter Statistics, CVSS Information and Database Information), for early detection of real-world exploits. This classifier achieves precision and recall higher than 80% for predicting the existence of private proof-of-concept exploits when only the vulnerabilities disclosed in Microsoft products and by using Microsoft’s Exploitability Index are considered.
Recurrent Neural Network Applications in Security Research
----------------------------------------------------------
**Binary Analysis.** Shin [*et al*.]{} [@shin2015recognizing] leveraged recurrent neural networks (RNN) to identify functions ([e.g., ]{}function boundaries, and general function identification) in binaries. For each training epoch, the RNN model is trained on $N$ examples (an example refers to a fixed-length sequence of bytes). The authors used one-hot encoding to convert each byte in a given example into a 256-vector, and associated a function start/end indicator with each byte (i.e., a 256-vector). Once the model is trained, it effectively serves as a binary classifier and outputs a decision for that byte as to whether it begins a function or not. The authors consequently combine the predictions from each model using simple heuristic rules to achieve aforementioned function identification tasks. It is claimed that this system is capable of halving the error rate on six out of eight benchmarks, and performs comparably on the remaining two. Chua [*et al*.]{} [@chua2017neural] presents [eklavya]{}, a RNN-based engine to recover function types ([e.g., ]{}identifying the number and primitive types of the arguments of a function) from x86/x64 machine code of a given function without prior knowledge of the compiler or the instruction set. On the condition that the boundaries of given functions are known, [eklavya]{} developed two primary modules - instruction embedding module and argument recovery module - to recover argument counts and types from binaries. The instruction embedding module takes a stream of functions as input and outputs a 256-vector representation of each instructions. After the instructions are represented as vectors, argument recovery module uses these sequences of vectors as training data and trains four RNNs for four tasks relating to function types recovery. The authors reported accuracy of around 84% and 81% for function argument count and type recovery tasks respectively.
**Anomaly Detection.** Du [*et al*.]{} [@du2017deeplog] proposed DeepLog, a deep neural network model utilizing Long Short-Term Memory (LSTM), to learn a system’s log patterns (e.g., log key patterns and corresponding parameter value patterns) from normal execution. At its detection stage, DeepLog uses both the log key and parameter value anomaly detection models to identify abnormal log entries. Its workflow model provides semantic information for users to diagnose a detect anomaly. The author reported that DeepLog outperformed other existing log-based anomaly detection methods achieving a $F$-measure of 96% in HDFS data and a $F$-measure of 98% in OpenStack data.
**Password Attack.** Melicher [*et al*.]{} [@melicher2016fast] used artificial neural networks to model text passwords’ resistance to guessing attacks and explore how different architectures and training methods impact neural networks’ guessing effectiveness. The authors demonstrated that neural networks guessed 70% of 4class8 (2,997 passwords that must contain all four character classes and be at least eight characters) passwords by $10^{15}$ guesses, while the next best performing guessing method (Markov models) guesses 57%.
**Malware Classification.** Pascanu [*et al*.]{} [@pascanu2015malware] model malware API calls as a sequence and use a recurrent model trained to predict next API call, and use the hidden state of the model (that encodes the history of past events) as the fixed-length feature vector that is given to a separate classifier (logistic regression or MLP) to classify malware.
The closest work to this paper is DeepLog [@du2017deeplog]. However, DeepLog focused on anomaly detection in regulated environment such as Hadoop and OpenStack with limited variety of events (e.g., 29 events in Hadoop environment and 40 events in OpenStack). In such a very specific log environment, DeepLog was able to use a small fraction of normal log entries to train and achieve good detection results. Our work aims at understanding multi-steps coordinated attacks in a noisy environment with a wide variety of events ([i.e., ]{}4,495 unique events in our dataset) and prediction in this setup is a far harder problem comparing to DeepLog. Additionally DeepLog considered an event *abnormal* if such event is *not* with top-$g$ probabilities to appear next. Our work does not employ this relaxed prediction criteria and focuses on the accurate prediction of the upcoming security event (out of 4,495 possible events) for a given system.
Conclusions {#sec:conclusion}
===========
This paper presented , a system for the prediction of security events. We evaluated it using an extensive dataset of alarms from an intrusion prevention system from a major security firm product. reaches a high precision for such a complex problem, showing stable results even when the model is trained months before the application to a test set.
We wish to thank the anonymous reviewers for their feedback and our shepherd Zhou Li for his help in improving our paper. Enrico Mariconti was supported by the EPSRC under grant 1490017.
[^1]: https://pypi.org/project/Sp2Learning/1.1.1/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Time Reversal Violation (TRV) interactions between quarks which appear in Standard Model (SM) and beyond-SM theories can induce TRV components in the nucleon-nucleon potential. The effects of these components can be studied by measuring the electric dipole moment (EDM) of light nuclei. In this work we present a complete derivation of the TRV nucleon-nucleon and three-nucleon potential up to next-to-next-to leading order (N2LO) in a chiral effective field theory ($\chi$EFT) framework. The TRV interaction is then used to evaluate the EDM of $\det$, $\tri$ and $\hel$ focusing in particular on the effects of the TRV three-body force and on the calculation of the theoretical errors. In case of a measurement of the EDM of these nuclei, the result of present work would permit to determine the values of the low energy constants and to identify the source of TRV.'
author:
- 'A. Gnech$^{\,{\rm a,b}}$, M. Viviani$^{\,{\rm b}}$'
title: Time Reversal Violation in Light Nuclei
---
Introduction
============
The violation of parity (P) with the conservation of charge conjugation (C) generates a CP violation that, using the CPT theorem, results in a time reversal violation (TRV). TRV is a key ingredient in the explanation of the observed matter-antimatter asymmetry in the Universe [@AS67]. The Standard Model (SM) has a natural source of CP-violation in the Cabibbo-Kobayashi-Maskawa (CKM) quark mixing matrix, however this mechanism is not sufficient for explaining the observed asymmetry [@AC93]. This discrepancy opens a window in possible TRV effect in the SM, such as the $\theta$-term in the Quantum Chromodynamics (QCD) sector [@GH76], or in other sources of beyond-SM (BSM) theories [@MP05].
The measurement of Electric Dipole Moments (EDMs) of particles is the most promising observable for studying TRV beyond CKM mixing matrix effects. Indeed, the EDM induced by the complex phase of the CKM matrix are suppressed since the EDM does not involve flavour changing [@MP05; @AC97; @MU12; @MU13]. Therefore, any non-vanishing EDM of a nuclear or an atomic system would highlight TRV effects beyond the CKM mixing matrix. The present experimental upper bounds on the EDMs of neutron and proton are $|d_n|<2.9\cdot 10^{-13}\ e$ fm [@CA06] and $|d_p|<7.9\cdot 10^{-12}\ e$ fm, where the proton EDM has been inferred from a measurement of the diamagnetic ${}^{199}{\rm Hg}$ atom [@WC09] using a calculation of the nuclear Schiff moment [@VF03]. For the electron, the most recent upperbound is $|d_e|<8.7\cdot10^{-16}\ e$ fm [@JB14], derived from the EDM of the ThO molecule.
In this context, there are proposals for the direct measurement of EDMs of electrons, single nucleons and light nuclei in dedicated storage rings [@YF06; @YK11; @AL13; @JP13; @FR13]. This new approach plans to reach an accuracy of $\sim 10^{-16}\ e$ fm, improving the sensitivity in particular in the hadronic sector. Any measurement of a non-vanishing EDM of this magnitude would be the evidence of TRV beyond CKM effects. However, a single measurement will not be sufficient to identify the source of TRV. For this reason, the measurement of EDM of various light nuclei such as $\det$, $\tri$ and $\hel$ can impose constrains on the TRV sources.
On the other hand, also spin rotation of polarized neutrons along the $y$-axis can be used as probe for TRV [@PK82; @LS82; @CP04; @CP06; @JDB14]. This explorative study is motivated by the new sensitivity reached by various cold neutron facilities such as the Los Alamos Neutron Science Center, the National Institute of Standards and Technology (NIST) Center for Neutron Research, and the Spallation Neutron Source (SNS) at Oak Ridge National Laboratory. Even if the sensitivity of the spin-rotation on TRV is far away from the present sensitivity of the EDM, this observable addresses the same physics of the EDM, posing, for the future, as a complementary and independent test for TRV in nuclei. Also other TRV observables in proton- and neutron-nucleus scattering have been proposed as probes of TRV effects [@YH11; @JDB14; @YNU16; @VG18; @PF19].
The use of light nuclei to study TRV results to be advantageous because the nuclear physics of the systems is theoretically under control and so the TRV effects can be easily highlighted. In particular, the chiral effective field theory ($\chi$EFT) has provided a practical and successful scheme to study two and many-nucleon interactions [@DR15; @EE15] and the interaction of electroweak probes with nuclei [@Park96; @Park03; @Koelling09; @Pastore09; @Koelling11; @Pastore11; @Piarulli13]. The $\chi$EFT approach is based on the observation that the chiral symmetry exhibited by QCD has a noticeable impact in the low-energy regime. Therefore, the form of the strong interactions of pions among themselves and other particles is severely constrained by the transformation properties of the fundamental Lagrangian under Lorentz, parity, time-reversal, and chiral symmetry [@W66; @CCWZ69]. The Lagrangian terms can be organized as an expansion in powers of $Q/\Lx$, where $\Lx\simeq1$ GeV specifies the chiral symmetry breaking scale and $Q$ is the exchanged pion momentum. Each term is associated to a low-energy constants (LECs) which are then determined by fits to experimental data.
The $\chi$EFT method permits to construct also an effective TRV Lagrangian treating all possible sources of TRV. The TRV Lagrangian induced by the $\theta$-term was derived in Refs. [@EM10; @JB150]. Also BSM terms such as supersymmetry, multi-Higgs scenarios, left-right symmetric model [*etc.*]{} induce TRV operators at the quark-gluon level which appear at level of dimension six (see for example Ref. [@BG10]). The $\chi$EFT Lagrangians for these sources were derived in Refs. [@JB150; @JV13; @WD13; @JV11]. This approach permits not only to determine the TRV interactions but also to estimate the chiral order of the LECs and their values as function of the fundamental parameters, providing a direct connection between the fundamental theories and the nuclear observables [@EM10; @JB150; @JV13]. To be noticed that the chiral order of the Lagrangian terms, which is determined by the products of the chiral order of the dynamical part and that of the LECs, really depends on the particular source of TRV. Therefore, when the LECs will be determined experimentally it will be possible to identify the dynamical properties of the TRV source [@JV11; @WD14].
Starting from the TRV Lagrangian, de Vries [*et al.*]{} [@CM11] and also Bsaisou [*et al.*]{} [@JB13] derived the chiral potential up to next-to-next leading order (N2LO) including only nucleon-pion interaction and contact interactions. In both works also the electromagnetic currents which play a role at N2LO for the EDM were derived but only in Ref. [@JB13] they were used to evaluate the EDM of the deuteron. In Ref. [@JV11b] the calculation of the EDM of $\det$, $\tri$ and $\hel$ was performed using only the one-pion-exchange part of the TRV potential coupled with phenomenological potential for the parity conserving (PC) part of the interaction.
Subsequent works showed the presence in the TRV Lagrangian of a three-pion term [@JB150], which was included in the calculation for the first time by Bsaisou [*et al.*]{} [@JB151]. This term generates at next-to-leading order (NLO) also a TRV three body force, which contribution to the $\tri$ and $\hel$ EDM was found to be smaller than expected by the chiral counting. The calculation reported in Ref. [@JB151] was also the first to use a complete chiral approach including the TRV potential up to NLO and the PC potential up to N2LO.
The aim of this work is twofold: First, the construction of a TRV potential up to N2LO considering all possible TRV interaction terms in the $\chi$EFT without making any assumption for the chiral order of the LECs. In this way all possible sources of TRV can be studied just setting the LECs to their estimated values and turning on and off the various terms in the Lagrangian. The second is the study of the EDM of $\det$, $\tri$ and $\hel$, providing a suitable framework for the future determination of the LECs. In our calculation the contribution of the TRV three-body force is found to be sizably larger than reported in Ref. [@JB151].
Finally, it is worthy to mention that exists a different approach to the derivation of the TRV nuclear forces based on meson-exchange model [@CP04]. This model includes pion and vector-meson exchange with 10 unknown meson constants. Such a theory, which has a wider energy range of validity but it is less systematic and with no direct connection with the fundamental Lagrangian, has been used to study the EDM of light nuclei [@CP04; @IS08; @NY15; @YH13] and the neutron spin rotation $\vec{n}-\vec{p}$ [@CP06] and $\vec{n}-\vec{d}$ [@CP04; @YH11] scattering.
The present paper is organized as follows. In Sec. \[sec:trvlag\] we will present the TRV chiral Lagrangian up to order $Q^2$ relevant for the calculation of the TRV potential, while in Sec. \[sec:trvpot\] we derive the TRV potential at N2LO. In Sec. \[sec:res\], we report the results obtained for the EDM of $\det$, $\tri$ and $\hel$ using the N2LO potential. Finally, in Sec. \[sec:conc\] we present our conclusions and perspectives. The technical details relative to the contributions of the various diagrams, and of the derivation of the potential in configuration space are given in Appendices \[app:vertex\], \[app:pot\] and \[app:rpot\]. Moreover, in Appendix \[app:NNNconv\] we will give some details of the calculation of the trinucleon wave function negative-parity component and about the convergence of the TRV three body force contribution.
The TRV Lagrangian {#sec:trvlag}
==================
The various possible sources of TRV in SM induce a TRV $NN$ and $NNN$ potential. This potential can be constructed starting from a pion-nucleon effective Lagrangian which includes, in principle, an infinite set of terms which violates the chiral symmetry as the fundamental (quark-level) Lagrangian. The effective Lagrangian can be ordered by a power counting scheme which permits to select the most important interactions. In literature the chiral order of the TRV Lagrangian is determined by considering the estimated order of the LECs [@EM10; @JV13; @JB150] which, however, is source dependent.
In this section we present only the TRV Lagrangian terms which can give some contribution to TRV $NN$ and $NNN$ potential up to N2LO in terms of pion field. In order to remain source independent we consider isoscalar, isovector and isotensor terms and we determine the chiral order of the TRV Lagrangian considering only the dynamical part. Namely, in the following we will consider all the TRV LECs to be equally important. To deal with a specific source of TRV, it will be sufficient to set some of the LECs to be zero, etc. At order $Q^0$ the TRV pion-nucleon Lagrangian includes three terms [@EM10; @JB150] $$\begin{aligned}
{\cal{L}}_{\rm TRV}^{\pi N\,(0)}=g_0\overline{\psi}\vec{\pi}\cdot\vec{\tau}
\psi+g_1\overline{\psi}\pi_3\psi+g_2\overline{\psi}\pi_3\tau_3\psi\,,\end{aligned}$$ where $\vec{\pi}$ is the pion field and $\psi$ is the nucleon field. To be noticed that the isotensor term is usually considered of higher order. At the same order a purely pionic interaction appears [@JB150], which reads $$\begin{aligned}
{\cal{L}}_{\rm TRV}^{3\pi\,(0)}=M\Delta_3 \pi_3\pi^2\,,\end{aligned}$$ where $M=938.88$ MeV is the average nucleon mass. At order $Q$ we have only one term that can give contribution to the potential up to N2LO and it reads, given explicitly as, $$\begin{aligned}
{\cal{L}}_{\rm TRV}^{\pi N\,(1)}=\frac{g_V^{(1)}}{2M\fp}\big[\overline{\psi}
\partial_\mu(\vec{\pi}\times\vec{\tau})_3\partial^\mu\psi+\ h.c.\ \big]\,,
\label{eq:ltrv1}\end{aligned}$$ while no Lagrangian terms for the pure pionic interaction are allowed. At order $Q^2$ we get six new contributions, =1.0pt $$\begin{aligned}
{\cal{L}}_{\rm TRV}^{\pi N\,(2)}&=&
{g_{S1}^{(2)}\over\fp^2}\overline{\psi}\partial_\mu\partial^\mu
(\pi\cdot\tau)\psi\nonumber\\
&+&{g_{S2}^{(2)}\over 2M^2\fp^2}\big[\overline{\psi}\partial_\mu\partial_\nu
(\pi\cdot\tau)\partial^\mu\partial^\nu\psi+{\rm h.c.} \big]\nonumber\\
&+&{g_{V1}^{(2)}\over\fp^2}\overline{\psi}\partial_\mu\partial^\mu
\pi_3\psi\nonumber\\
&+&{g_{V2}^{(2)}\over 2M^2\fp^2}\big[\overline{\psi}\partial_\mu\partial_\nu
\pi_3\partial^\mu\partial^\nu\psi+{\rm h.c.} \big]\nonumber\\
&+&{g_{T1}^{(2)}\over\fp^2}\overline{\psi}\partial_\mu\partial^\mu
\pi_3\tau_3\psi\nonumber\\
&+&{g_{T2}^{(2)}\over 2M^2\fp^2}\big[\overline{\psi}\partial_\mu\partial_\nu
\pi_3\tau_3\partial^\mu\partial^\nu\psi+{\rm h.c.} \big]\ ,
\label{eq:ltrv2}\end{aligned}$$ where $S$, $V$, $T$ stand for isoscalar, isovector and isotensor respectively and $\fp\simeq92$ MeV is the pion decay constant. The three-pion interaction that appears at this level gives contributions of order $Q^2$ to the TRV nuclear potential which is beyond our purpose, therefore we do not consider it. The Lagrangian contains also four-nucleon contact terms ${\cal L}^{\rm CT}_{\rm TRV}$ representing interactions originating from excitation of resonances and exchange of heavy mesons. At lowest order ${\cal L}^{\rm CT}_{\rm TRV}$ contains only five independent four-nucleon contact terms with a single gradient.
In the following we also need the PC Lagrangian up to N2LO: =1.0pt $$\begin{aligned}
{\cal L}^{PC}&=& {\cal L}_{\pi\pi}^{(2)}+\ldots\nonumber\\
&+& {\cal L}_{N\pi}^{(1)}+{\cal L}_{N\pi}^{(2)}+{\cal
L}_{N\pi}^{(3)}+\ldots+ {\cal L}^{PC}_{CT}\ ,\label{eq:Lpc} \\
{\cal L}_{\pi\pi}^{(2)}&=& {f_\pi^2\over 4} \langle\nabla_\mu U^\dag
\nabla^\mu U + \chi^\dag U+\chi U^\dag\rangle\ , \label{eq:Lpc_pipi2}\\
{\cal L}_{N\pi}^{(1)} &=& \overline{\psi}\Bigl(i\gamma^\mu D_\mu -M
+{g_A\over 2} \gamma^\mu\gamma^5 u_\mu \Bigr)\psi\
, \label{eq:Lpc_pin1}\\
{\cal L}_{N\pi}^{(2)} &=& c_1 \overline{\psi}\langle \chi_+\rangle
\psi - \frac{c_2}{8M^2}\big[\overline{\psi}\langle u_\mu u_\nu \rangle
D^\mu D^\nu \psi +\ {\rm h.c.}\big]\nonumber\\
&+&c_3\overline{\psi}\frac{1}{2}\langle u_\mu u^\mu \rangle \psi+
c_4 \overline{\psi}\frac{i}{4}[u_\mu,u_\nu]\sigma^{\mu\nu}\psi+
\ldots \ ,\label{eq:Lpc_pin2}\\
{\cal L}_{N\pi}^{(3)} &=& d_{16} \overline{\psi}{1\over 2}
\gamma^\mu\gamma^5 u_\mu \langle \chi_+\rangle \psi\nonumber \\
&+& d_{18} \overline{\psi}{i\over 2}
\gamma^\mu\gamma^5 [D_\mu, \chi_-] \psi+\ldots\ ,\label{eq:Lpc_pin3}\end{aligned}$$ where the building blocks $U$, $u_\mu$, $\chi_+$, $\chi_-$ and the covariant derivative $D_\mu$ are defined in Ref. [@viviani14]. We have omitted all the terms not relevant in the present work (for the complete expression for Lagrangian $ {\cal L}_{N\pi}^{(2)}$ and ${\cal L}_{N\pi}^{(3)}$ in Ref. [@NF00]). Four-nucleon contact terms (see, for example, Refs.[@EE09; @RM11]) are lumped into ${\cal L}^{PC}_{CT}$. The parameters $c_1$, $c_2$, $c_3$, $c_4$, $d_{16}$, and $d_{18}$ are LECs entering the PC Lagrangian. All the constants entering the terms discussed in this section have to be considered as “bare” parameters (i.e. unrenormalized).
The TRV potential up to order Q {#sec:trvpot}
===============================
In this section, we discuss the derivation of the TRV $NN$ and $NNN$ potential at N2LO. We provide, order by order in power counting the formal expressions for it in terms of the time-ordered perturbation theory (TOPT) amplitudes, following the scheme presented in Ref. [@viviani14]. Then, the various diagrams associated with these amplitudes are discussed (additional details are given in Appendix \[app:pot\]). We will give also some hints in the renormalization of the coupling constants.
From amplitudes to potentials
-----------------------------
We start considering the conventional $NN$ scattering amplitude ${\langle}N'N' |T|NN{\rangle}$, where $|NN{\rangle}$ and $|N'N'{\rangle}$ represent the initial and final two nucleon state and $T$ can be written as, $$\begin{aligned}
T = H_I \sum_{n=1}^\infty \left(
\frac{1}{E_i -H_0 +i\, \eta } H_I \right)^{n-1} \ ,
\label{eq:pt}\end{aligned}$$ where $E_i$ is the initial energy of the two nucleons, $H_0$ is the Hamiltonian describing free pions and nucleons, and $H_I$ is the Hamiltonian describing interactions among these particles. To be noticed that in Eq. (\[eq:pt\]) the interaction Hamiltonian $H_I$ is in the Schrödinger picture and that, at the order of interest here, it follows simply from $H_I=-\int {\rm d} \bmx\; {\cal L}_I (t=0,\bmx)$, where ${\cal L}_I$ is the interaction Lagrangian in interaction picture. Vertices from $H_I$ are listed in Appendix \[app:vertex\].
The $NN$ scattering amplitude can be organized as an expansion in powers of $Q/{\Lambda_\chi}\ll 1$, where ${\Lambda_\chi}\simeq1$ GeV is the typical hadronic mass scale, $${\langle}N'N' |T|NN{\rangle}=\sum_n T^{(n)}\,,
\label{eq:teft}$$ where $T^{(n)}\sim Q^n$.
The $T$ matrix in Eq. (\[eq:teft\]) is generated, order by order in the power counting, by the $NN$ potential $V$ from iterations of it in the Lippmann-Schwinger (LS) equation, $$V+V\, G_0\, V+V\, G_0 \, V\, G_0 \, V+\dots \ ,
\label{eq:lse}$$ where $G_0$ denotes the free two-nucleon propagator. If we assume that, $$\langle N^\prime N^\prime |V|NN\rangle=\sum_n V^{(n)}\ ,$$ with $V^{(n)}$ of order $Q^n$, it is possible to assign to any $T^{(n)}$ the terms in the LS equation that are of the same order. Generally in term like $[V^{(m)}G_0V^{(n)}]$ is of order $Q^{m+n+1}$ because $G_0$ is of order $Q^{-2}$ and the implicit loop integration brings a factor $Q^3$ (for a more detailed discussion, see Ref. [@Pastore11]).
In our case the two nucleons interact via a PC potential plus a very small TRV component. The $\chi$EFT Lagrangian implies the following expansion in powers of $Q$ for $T= T_{PC} + T_{TRV}$: $$\begin{aligned}
T_{PC}&=&T^{(0)}_{PC}+T^{(1)}_{PC}+T^{(2)}_{PC}+\ldots ,\label{tpc}\\
T_{TRV}&=&T^{(-1)}_{TRV}+T^{(0)}_{TRV}+T^{(1)}_{TRV}+\ldots .\label{ttrv}\end{aligned}$$ Assuming that the potential $V=V_{PC}+V_{TRV}$ have a similar expansion, $$\begin{aligned}
V_{PC} & = & V_{PC}^{(0)}+V_{PC}^{(1)}+V_{PC}^{(2)}+\dots \\
V_{TRV} & = & V_{TRV}^{(-1)}+V_{TRV}^{(0)}+V_{TRV}^{(1)}+\dots ,\end{aligned}$$ we can match order by order the $T$ and the terms in the LS equation obtaining the form definition of the TRV $NN$ potential from the scattering amplitude, $$\begin{aligned}
V_{TRV}^{(-1)} & = & T_{TRV}^{(-1)}\ ,\label{eq:vtrvml}\\
V_{TRV}^{(0)} & = & T_{TRV}^{(0)}-\left[V_{TRV}^{(-1)}G_0 V_{PC}^{(0)}\right]
-\left[V_{PC}^{(0)}G_0 V_{TRV}^{(-1)}\right]\ ,\label{eq:vtrv0}\\
V_{TRV}^{(1)} & = & T_{PV}^{(1)}-\left[V_{TRV}^{(0)}G_0 V_{PC}^{(0)}\right]
-\left[V_{PC}^{(0)}G_0 V_{TRV}^{(0)}\right]\nonumber\\
& - & \left[V_{TRV}^{(-1)}G_0 V_{PC}^{(1)}\right]-
\left[V_{PC}^{(1)}G_0 V_{TRV}^{(-1)}\right]\nonumber\\
& - & \left[V_{TRV}^{(-1)}G_0 V_{PC}^{(0)}G_0 V_{PC}^{(0)}\right]
- \left[V_{PC}^{(0)}G_0 V_{TRV}^{(-1)}G_0 V_{PC}^{(0)}\right]
\nonumber\\
& - & \left[V_{PC}^{(0)}G_0 V_{PC}^{(0)}G_0 V_{TRV}^{(-1)}\right]
\ .\label{eq:vtrv1}\end{aligned}$$ The generalization for the $NNN$ TRV potential is straightforward.
The $NN$ TRV potential
----------------------
We define the following momenta, $$\begin{aligned}
&&\bmK_j=(\bmp_j'+\bmp_j)/2\ , \quad
\bmk_j=\bmp_j'-\bmp_j\ ,\label{eq:notjb1}\end{aligned}$$ where $\bmp_j$ and $\bmp'_j$ are the initial and final momenta of nucleon $j$. From the overall momentum conservation $\bmp_1+\bmp_2=\bmp_1'+\bmp_2'$, we can define $\bmk=\bmk_1=-\bmk_2$. We also define $\bmK=(\bmK_1-\bmK_2)/2$, $\bmP=\bmp_1+\bmp_2=\bmK_1+\bmK_2$, in this way is possible to write the TRV $NN$ potential as, $$V_{TRV}(\bmk,\bmK_1,\bmK_2)= V_{TRV}^{(c.m.)}(\bmk,\bmK)+
V^{(\bmP)}_{TRV}(\bmk,\bmK)\ ,\label{eq:widetildev2}$$ where the term $ V^{(\bmP)}_{TRV}(\bmk,\bmK)$ represents boost corrections to $ V^{(c.m.)}_{TRV}(\bmk,\bmK)$ [@girlanda10], the potential in the center-of-mass (c.m.) frame. Below we ignore these boost corrections and provide expressions for $ V^{(c.m.)}_{TRV}(\bmk,\bmK)$ only.
![ \[fig:diagNN\] Time-ordered diagrams contributing to the TRV potential (only a single time ordering is shown). Nucleons and pions are denoted by solid and dashed lines, respectively. The open (solid) circle represents a PC (TRV) vertex.](diagNN.eps "fig:").
The diagrams that give contribution to the $NN$ TRV potential are shown in Fig. \[fig:diagNN\]. We do not consider diagrams which give contribution only to the renormalization of the LECs. In this section we write the final expression of the $NN$ TRV potential $ V^{(c.m.)}_{TRV}$ as, $$\begin{aligned}
V^{(c.m.)}_{TRV}&=&
V^{ ({\rm OPE})}_{TRV} +
V^{ ({\rm TPE})}_{TRV} +
V^{ ({\rm 3\pi,0})}_{TRV} +
V^{ ({\rm 3\pi,1})}_{TRV} \nonumber\\
&&+ V^{ ({\rm RC})}_{TRV} +
V^{ ({\rm 3\pi,RC})}_{TRV} +
V^{ ({\rm CT})}_{TRV} \ ,\label{eq:dec}\end{aligned}$$ namely as a sum of terms due to one-pion exchange (OPE), two-pion exchange (TPE), three-pion exchange at NLO (3$\pi$,0) and at N2LO (3$\pi$,1), relativistic corrections derived from the OPE (RC) and from the $3\pi$-exchange ($3\pi$,RC), and contact contributions (CT). From now on we define also $g_0^*=g_0+g_2/3$ (see Appendix \[app:pot\] for details). The OPE term is the contribution of diagram (a) of order $Q^{-1}$ (LO) and it reads, =0.7pt $$\begin{aligned}
{V_{TRV}}^{(\rm OPE)}
&=&\frac{\overline{g}_A\overline{g}_0^*}{2\overline{f}_\pi}({\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2})\frac{i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})}{\omk^2}\nonumber\\
&+&
\frac{\overline{g}_A\overline{g}_1}{4\overline{f}_\pi}\Big[(\tau_{1z}+\tau_{2z})
\frac{i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})}{\omega_k^2}\nonumber\\
&&\qquad
+(\tau_{1z}-\tau_{2z})\frac{i \bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})}{\omega_k^2}\Big]
\nonumber \\
&+&\frac{\overline{g}_A\overline{g}_2}{6\overline{f}_{\pi}}
(3\tau_{1z}\tau_{2z}-{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2})\frac{i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})}{\omk^2}\ ,\nonumber\\
\label{eq:isotope}\end{aligned}$$ where there are an isoscalar, an isovector and an isotensor components and $\overline{g}_0^*=\overline{g}_0+\overline{g}_2/3$. The coupling constants $\overline{g}_A/\overline{f}_{\pi}$, $\overline{g}_0^*$, $\overline{g}_1$, $\overline{g}_2$ are renormalized coupling constants, having reabsorbed the various infinites generated by loops and given as combinations of the “bare” LECs entering the Lagrangian. The expression for $\overline{g}_A/\overline{f}_{\pi}$ is the same as reported in Ref. [@viviani14]. The LECs $g_{S1}^{(2)}$, $g_{S2}^{(2)}$, etc. enter in Eq. (\[eq:isotope\]) through the renormalized constants $\overline{g}_0$, etc. (see Appendix \[app:pot\] for more details). Note that, as mentioned before, the correct expressions of the renormalized constants in term of the bare ones should contain also the contributions of additional diagrams not considered here (see Ref. [@viviani14] for the procedure to be followed for the PC and PV potentials, respectively).
The TPE term comes from the not singular contribution of panels (e)-(h) of Fig. \[fig:diagNN\], taking also into account the subtracting terms given in Eq. (\[eq:vtrv1\]). As discussed in Appendix \[app:pot\] this term has no isovector component, in agreement with the result reported in [@CM11]. Therefore, as reported in Eqs. (\[eq:pote\]) and (\[eq:box\]), we have, =1.0pt $$\begin{aligned}
{V_{TRV}}^{(\rm TPE)}&=\frac{\ga{g}_0^*}{\fp\Lx^2}
{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\ i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\ L(k)\nonumber\\
&+\frac{\ga^3 {g}_0^*}{\fp\Lx^2}\ {\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\ i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})
\ (H(k)-3L(k))\nonumber\\
&-\frac{\ga {g}_2}{3\fp\Lx^2}
(3\tau_{1z}\tau_{2z}-{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2})
\ i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\ L(k)\nonumber\\
&-\frac{\ga^3 {g}_2}{3\fp\Lx^2}(3\tau_{1z}\tau_{2z}-{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2})\nonumber\\
&\qquad\qquad\times i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\
(H(k)-3L(k))\ ,\end{aligned}$$ where the loop functions $L(k)$ and $H(k)$ are defined in Eqs. (\[eq:sL\]) and (\[eq:H\]). In this and the following NLO and N2LO terms the bare coupling constants $\ga$, $\fp$, $g_0^*$, $\ldots$ can be safely replaced by the corresponding physical (renormalized) values.
The $3\pi$-exchange term gives a NLO contribution through the diagram (i) of Fig. \[fig:diagNN\], =1.0pt $$\begin{aligned}
{V_{TRV}}^{(3\pi,0)}=&-\frac{5\ga^3\deltatre M}{4\fp\Lx^2}\pi
\Big[(\tau_{1z}+\tau_{2z})
\frac{i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})}{\omega_k^2}\nonumber\\
&+
(\tau_{1z}-\tau_{2z})\frac{i \bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})}{\omega_k^2}\Big]
\nonumber\\
&\qquad\times\Big(\big(1-\frac{2\mp^2}{s^2}\big)s^2A(k)+\mp\Big)\label{eq:3pnlo}\ ,\end{aligned}$$ where $A(k)$ is defined in Eq. (\[eq:A\]). The expression obtained in Eq. (\[eq:3pnlo\]) is in agreement with the expression derived in Refs. [@JV13; @JB150].
The diagram in panel (l) of Fig. \[fig:diagNN\] contributes to ${V_{TRV}}^{(3\pi)}$ at N2LO. The expression for this diagram is derived in Appendix \[app:pot\], here we report the final result only, =1.0pt $$\begin{aligned}
&{V_{TRV}}^{(3\pi,1)}=\frac{5\ga\deltatre Mc_1}{2\fp\Lx^2}
\Big[(\tau_{1z}+\tau_{2z})i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&\qquad+(\tau_{1z}-\tau_{2z})i \bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\Big]
\ 4\frac{\mp^2}{\omega_k^2}L(k)\nonumber\\
&\qquad-\frac{5\ga\deltatre Mc_2}{6\fp\Lx^2}
\Big[(\tau_{1z}+\tau_{2z})i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&\qquad+(\tau_{1z}-\tau_{2z}) i \bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\Big]
\Big(2L(k)+6\frac{\mp^2}{\omega_k^2}L(k)\Big)\nonumber\\
&\qquad-\frac{5\ga\deltatre Mc_3}{4\fp\Lx^2}
\Big[(\tau_{1z}+\tau_{2z})i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&\qquad+(\tau_{1z}-\tau_{2z})i \bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\Big]
\Big(3L(k)+5\frac{\mp^2}{\omega_k^2}L(k)\Big)\label{eq:3pc3}\ .\end{aligned}$$ Note in Eq. (\[eq:3pc3\]) the presence of the $c_1$, $c_2$ and $c_3$ LECs, which belong to the PC sector. In Eqs. (\[eq:3pnlo\]) and (\[eq:3pc3\]) the $\deltatre$ is the renormalized LEC to the relative order of the expressions.
The ${V_{TRV}}^{ ({\rm RC})}$ term of the potential takes into account contributions from the RC of the vertices in the OPE of panel (a), $$\begin{aligned}
{V_{TRV}}^{(\rm RC)}&=&\frac{\ga {g}_0^*}{8\fp M^2}
{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\ \frac{1}{\omk^2}\nonumber\\
&&\times\Big[-\frac{i}{2} \left(8K^2+k^2\right)\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&&+\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\nonumber\\
&+&\frac{\ga {g}_1}{16\fp M^2}\ \frac{1}{\omk^2}\nonumber\\
&&\times\Big\{(\tau_{1z}-\tau_{2z})
\Big[-\frac{i}{2} \left(8K^2+k^2\right)\bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\nonumber\\
&&+\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}+\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\nonumber\\
&&+(\tau_{1z}+\tau_{2z})
\Big[-\frac{i}{2} \left(8K^2+k^2\right)\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&&+\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\Big\}\nonumber\\
&+&\frac{\ga {g}_2}{24\fp M^2}(3\tau_{1z}\tau_{2z}-{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2})
\ \frac{1}{\omk^2}\nonumber\\
&&\times\Big[-\frac{i}{2} \left(8K^2+k^2\right)\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&&+\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\label{eq:rco2}\ ,\nonumber\\\end{aligned}$$ where, as for the OPE, we have an isoscalar, an isovector and an isotensor term. Also the diagram of Fig. \[fig:diagNN\] (i) gives a contribution to the at N2LO, both from the RC of the vertices, and from NLO in the pion propagators (see Appendix \[app:pot\]). The final expression we obtain is $$\begin{aligned}
{V_{TRV}}&^{(\rm 3\pi,RC)}=-\frac{5\ga^3\deltatre}{16\fp\Lx^2}
\Big[(\tau_{1z}+\tau_{2z})i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+(\tau_{1z}-\tau_{2z})i \bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\Big]\nonumber\\
&\times\Big(\frac{25}{6}L(k)-\frac{7}{2}\frac{\mp^2}{\omega_k^2}L(k)
+2\frac{\mp^2}{\omega_k^2}H(k)\Big)\nonumber\\
&-\frac{25\ga^3\deltatre}{12\fp\Lx^2}\frac{1}
{\omega_k^2}\ \Big\{(\tau_{1z}+\tau_{2z})
\Big[\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}\nonumber\\
&+\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]
+(\tau_{1z}+\tau_{2z})\nonumber\\
&\times\Big[\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\Big\}\label{eq:rc3p}\ ,
\nonumber\\\end{aligned}$$ where in Eqs. (\[eq:rco2\]) and (\[eq:rc3p\]) the TRV coupling constants are renormalized to order $Q^2$.
Last, the potential $V^{ ({\rm CT})}_{TRV}$, derived from the $NN$ contact diagrams (b) of Fig. \[fig:diagNN\], reads $$\begin{aligned}
V^{\text{(CT)}}_{TRV}&=&\frac{1}{\Lambda_\chi^2f_\pi}\big\{\overline{C}_1\ i\bmk\cdot
\left({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}\right)\nonumber\\
&+&\overline{C}_2\ i\bmk\cdot
\left({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}\right){\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\nonumber\\
&+&{\overline{C}_3\over 2}\
\big[i\bmk\cdot\left({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}\right)\left(\tau_{z1}+\tau_{z2}\right)
\nonumber\\
&&-i\bmk\cdot\left({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2}\right)
\left(\tau_{z1}-\tau_{z2}\right)\big]\nonumber\\
&+&{\overline{C}_4\over 2}\
\big[i\bmk\cdot\left({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}\right)\left(\tau_{z1}+\tau_{z2}\right)
\nonumber\\
&&+i\bmk\cdot\left({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2}\right)
\left(\tau_{z1}-\tau_{z2}\right)\big]\nonumber\\
&+&\overline{C}_5\ i\bmk\cdot
\left({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}\right)\left(3\tau_{z1}\tau_{z2}-{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\right)
\big\}\label{eq:ct}\ .\end{aligned}$$ Above $\overline{C}_1$, $\overline{C}_2$, $\overline{C}_3$, $\overline{C}_4$ and $\overline{C}_5$ are renormalized LECs since they have reabsorbed various singular terms coming from the TPE diagrams, 3$\pi$ diagrams and the relativistic corrections. Note that it is possible to write ten operators which can enter $V_{TRV}^{(\text{CT})}$ at order $Q$ but only five of them are independent. In this work we have chosen to write the operators in term of $\bmk$, so that the $r$-space version of $V_{TRV}^{(\text{CT})}$ will assume a simple local form with no gradients. We want also to remark that the $C_2$, $C_4$ and $C_5$ LECs are needed in order to reabsorb the divergences coming from the TPE and $3\pi$ exchange diagrams.
In the calculation of the EDM in Sec. \[sec:res\], the configuration space version of the potential is needed. This formally follows from =1.0pt $$\begin{aligned}
\!\!\!\!\!\!\!\!\!\!\!\! \langle \bmr_1'\bmr_2'|V|\bmr_1\bmr_2\rangle&=&
\delta (\bmR-\bmR') \int {d^3k\over (2\pi)^3} {d^3K\over (2\pi)^3}
\nonumber\\
&\times& e^{i(\bmK+\bmk/2)\cdot\bmr'} V(\bmk,\bmK)
e^{-i(\bmK-\bmk/2)\cdot\bmr}\ ,
\label{eq:vrsp}\end{aligned}$$ where $\bmr=\bmr_1-\bmr_2$ and $\bmR=(\bmr_1+\bmr_2)/2$, and similarly for the primed variables. In order to carry out the Fourier transforms above, the integrand is regularized by including a cutoff of the form $$C_\Lambda(k)={\rm e}^{-(k/\Lambda)^4}\ ,\label{eq:cutoff}$$ where the cutoff parameter $\Lambda$ is taken in the range 450–550 MeV. With such a choice the $V^{ ({\rm OPE})}_{TRV}$, $V^{ ({\rm TPE})}_{TRV}$, $V^{ ({\rm 3\pi,0})}_{TRV}$, $V^{ ({\rm 3\pi,1})}_{TRV}$, and $V^{ ({\rm CT})}_{TRV}$ components of the resulting potential are local, i.e., $\langle \bmr_1'\bmr_2'|V|\bmr_1\bmr_2\rangle=
\delta (\bmR-\bmR')\, \delta (\bmr-\bmr') V(\bmr)$, while the RC component contains mild non-localities associated with linear and quadratic terms in the relative momentum operator $-i {{\bm \nabla}}$. Explicit expressions for all these components are listed in Appendix \[app:rpotNN\].
The $NNN$ TRV potential {#sec:NNNtrv}
-----------------------
The $3\pi$ TRV vertex gives rise to a three body contribution through the diagram (m) in Fig. \[fig:diagNN\]. The lowest contribution appears at NLO while at N2LO the various time ordering cancel out (see Appendix \[app:pot\]). The final expression for the NLO of the $NNN$ TRV potential is, $$\begin{aligned}
{V_{TRV}}^{\rm{NNN}}&=&\frac{\deltatre g_A^3 M}{4\fp^3}({\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\ \tau_{3z}+
{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_3}\ \tau_{2z}+{\boldsymbol{\tau}_2}\cdot{\boldsymbol{\tau}_3}\ \tau_{1z})\nonumber\\
&&\times\frac{(i\bmk_1\cdot{\boldsymbol{\sigma}_1})\ (i\bmk_2\cdot{\boldsymbol{\sigma}_2})\ (i\bmk_3\cdot{\boldsymbol{\sigma}_3})}
{\omega^2_{k_1}\omega^2_{k_2}\omega^2_{k_3}}\ ,
\label{eq:NNNpot}\end{aligned}$$ which is in agreement with the expression of Ref. [@JV13; @JB150].
Also in this case we need the Fourier transform of the potential. Using the overall momentum conservation $\bmp_1+\bmp_2+\bmp_3=\bmp_1'+\bmp_2'+\bmp_3'$, which give us that $\bmk_3=-\bmk_1-\bmk_2$, and defining $\bmQ=\bmk_1+\bmk_2$ and $\bmq=\bmk_1-\bmk_2$ the Fourier transform becomes, $$\begin{aligned}
\langle \bmr_1'\bmr_2'\bmr_3'|& V|\bmr_1\bmr_2\bmr_3\rangle=
\delta (\bmr_1-\bmr_1')\delta (\bmr_2-\bmr_2')\delta (\bmr_3-\bmr_3')
\nonumber\\
&\times\int {d^3q\over (2\pi)^3} {d^3Q\over (2\pi)^3}
V(\bmq,\bmQ)
e^{-i(\bmq/2)\cdot\bmx_2}\,e^{-i\bmQ\cdot\bmx_1}\ ,
\label{eq:vrnnn}\nonumber\\\end{aligned}$$ where $\bmx_1=\bmr_2-\bmr_1$ and $\bmx_2=\bmr_3-(\bmr_1+\bmr_2)/2$. Note that with this choice of $\bmQ$ and $\bmq$, the form of the Jacobi vectors appear automatically. We use the regularization function reported in Eq. (\[eq:cutoff\]), where we replace $k$ with $Q$, namely, $$C_\Lambda(Q)={\rm e}^{-(Q/\Lambda)^4}\ ,\label{eq:cutoffnnn}$$ which gives a local form of the $NNN$ TRV potential. A complete description of how to carry out the integration in the Fourier transform is given in Appendix \[sec:NNNr\].
Results {#sec:res}
=======
In this section, we report results for the EDM of $\det$, $\tri$ and $\hel$. Hereafter, we do not use the barred notation for the renormalized LECs but all the coupling constant must be considered as renormalized. The calculations are based on the TRV $NN$ potential derived in the previous section (and summarized in Appendix \[app:pot\]) and on the (strong interaction) PC $NN$ potential obtained by Entem and Machleidt at next-to-next-to-next-to-next-to-leading order (N4LO) [@DR15]. These potentials are regularized with a cutoff function depending on a parameter $\Lambda$; its functional form, however, is different from the adopted here for $V_{TRV}$. Below we consider the versions with $\Lambda=450$ MeV, $500$ MeV, and $550$ MeV. The calculations of $\tri$ and $\hel$ EDMs also include the $NNN$ TRV nuclear potential derived in Sec. \[sec:NNNtrv\] and the PC $NNN$ potential derived in $\chi$EFT at N2LO. As for the $NN$ PC potential, it depends on a cutoff parameter $\Lambda$ which is chosen to be consistent with those in the PC and TRV $NN$ potentials. The three-nucleon PC potential depends, in addition, on two unknown LECs, denoted as $c_D$ and $c_E$ and also on the LECs $c_1$, $c_3$ and $c_4$. In this work, we use the values reported in Table III of Ref. [@LE18]. The ${V_{TRV}}^{(3\pi,1)}$ term of the TRV potential depends on the LECs $c_1$, $c_2$ and $c_3$ which are taken from Table II of Ref. [@DR15] and summarized here in Table \[tab:cecd\].
PC interactions $c_1$ $c_2$ $c_3$
----------------- ------- ------- -------
N2LO -0.74 - -3.61
N3LO -1.07 3.20 -5.32
N4LO -1.10 3.57 -5.54
: \[tab:cecd\] Values of the coefficients and $c_1$, $c_2$, $c_3$ in unit of GeV$^{-1}$ taken from Ref. [@DR15].
In the following the values $g_A=1.267$ and $\fp=92.4$ MeV are adopted.
This section is organized as follows. In Sec. \[sec:dedm\], we present the general expression for the EDM operator and the results for the deuteron EDM, while in Sec. \[sec:3Hedm\] we present the calculation of the $\tri$ and $\hel$ EDMs.
Deuteron EDM {#sec:dedm}
------------
The EDM operator ${\hat{\boldsymbol{D}}}$ is composed by two parts, $${\hat{\boldsymbol{D}}}={\hat{\boldsymbol{D}}}_{\rm PC}+{\hat{\boldsymbol{D}}}_{\rm TRV}.$$ The ${\hat{\boldsymbol{D}}}_{\rm PC}$ receives contribution at LO from the nuclear EDM polarization operator $${\hat{\boldsymbol{D}}}_{\rm PC}=e\sum_i\frac{1+\tau_z(i)}{2}\boldsymbol{r}_i,
\label{eq:dpc}$$ where $e>0$ is the electric unit charge, $\tau_z(i)$ and $\bmr_i$ are the $z$ component of the isospin and the position of the i-th particle. The ${\hat{\boldsymbol{D}}}_{\rm TRV}$ LO contribution comes from the intrinsic nucleon EDM, $${\hat{\boldsymbol{D}}}_{\rm TRV}=\frac{1}{2}\sum_i\left[(d_p+d_n)+(d_p-d_n)\tau_z(i)
\right]\boldsymbol{\sigma}(i)\ ,
\label{eq:dtrv}$$ where $d_p$ and $d_n$ are the EDM of proton and neutron respectively and $\boldsymbol{\sigma}(i)$ is the spin operator which act on the i-th particle. As discussed in Refs. [@JV11b; @JB13] both the ${\hat{\boldsymbol{D}}}_{\rm PC}$ and ${\hat{\boldsymbol{D}}}_{\rm TRV}$ receive contributions from transition currents at N2LO. Of course, a complete treatment of the EDM up to N2LO needs to take care of them but hereafter we neglect their contribution, showing only the effects of N2LO TRV potential. In future work we plan to include the N2LO current contributions in the calculations.
The nucleon EDM of an $A$ nucleus can be expressed as $$\begin{aligned}
d^A & = & {\langle}\psi^A_{+} | \hat{D}_{\rm TRV} | \psi^A_{+} {\rangle}+2\, {\langle}\psi^A_{+} | \hat{D}_{\rm PC} | \psi^A_{-} {\rangle}\nonumber\\
& \equiv & d_{\rm TRV}^A+d_{\rm PC}^A\ ,\end{aligned}$$ where $|\psi^A_{+}{\rangle}$ $(|\psi^A_{-}{\rangle})$ is defined to be the even-parity (odd-parity) component of the wave function. In general, due to the smallness of the LECs, the EDM can be expressed as linear on the TRV LECs, $$\begin{aligned}
d^A_{\rm TRV}&=&d_pa_p+d_na_n\\
d^A_{\rm PC}&=&g_0a_0+g_1a_1+g_2a_2+\Delta_3 a_\Delta\nonumber\\
&+&C_1A_1+C_2A_2+C_3A_3+C_4A_4+C_5A_5\ ,
\label{eq:da}\end{aligned}$$ where the $a_i$ for $i=0,1,2$, $a_\Delta$, $A_i$ for $i=1,\dots,5$ and $a_p$, $a_n$ are numerical coefficients independent on the LECs values (however, they do depend on the cutoff $\Lambda$ in the PC and TRV chiral potentials).
We evaluate also the theoretical errors associated with the chiral expansion of the nuclear potential. We express the error on the numerical coefficients for the deuteron as, $$\left(\delta a_i\right)^2=\left(\delta a_i^{\text{PC}}\right)^2
+\left(\delta a_i^{\text{TRV}}\right)^2\ ,
\label{eq:errd}$$ where $\delta a_i^{\text{PC}}$ is the error associated to the chiral expansion of the PC potential and $\delta a_i^{\text{TRV}}$ the error associated with the chiral expansion of the TRV potential. Both the contributions were evaluated following the prescriptions of Epelbaum [*et al.*]{} [@EE15] where as reference momentum in the calculation of the errors we used the mass of the pion. It is straightforward to understand that the errors are dominated by the TRV part because we are using the N2LO potential for it and the N4LO potential for the PC part.
The coefficients for the deuteron, evaluated with the N4LO PC potential and the N2LO TRV potential and the associated errors for the three different choices of the cutoff parameters, are given in Table \[tab:2hdpc\]. The coefficients $a_p$ and $a_n$ multiplying the intrinsic neutron and proton EDM, as already pointed out first in Ref. [@NY15] and then in Ref. [@JB151], are given by, $$a_n=a_p=\left(1-\frac{3}{2}P_D\right)\, ,$$ where $P_D$ is the percentage of D-wave present in the deuteron wave function. The values of $a_n$ and $a_p$ obtained using the Entem and Machleidt at N4LO for three different choices of the cut off are reported in Table \[tab:2hdpc\]. The operator ${\hat{\boldsymbol{D}}}_{\rm PC}$ for the deuteron reduces to ${\hat{\boldsymbol{D}}}_{\rm PC}=(\tau_z(1)-\tau_z(2))\bmr$ therefore $d^2_{\rm PC}$ in Eq. (\[eq:da\]) receives contribution only from the component $^3P_1$ which has $T=1$. This component is generated by the TRV potential components proportional to $({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})$, namely those proportional to $g_1$, $\Delta_3$, $C_3$ and $C_4$.
$\Lambda$(MeV) $450$ $500$ $550$
---------------- ------------------- ------------------- -------------------
$a_n(a_p)$ $\m0.934\pm0.001$ $\m0.939\pm0.001$ $\m0.938\pm0.001$
$a_1$ $\m0.192\pm0.006$ $\m0.197\pm0.004$ $\m0.194\pm0.003$
$a_\Delta$ $ -0.306\pm0.174$ $ -0.341\pm0.153$ $ -0.349\pm0.137$
$A_3$ $\m0.013\pm0.004$ $\m0.013\pm0.004$ $\m0.013\pm0.004$
$A_4$ $ -0.013\pm0.004$ $ -0.013\pm0.004$ $ -0.013\pm0.004$
: \[tab:2hdpc\] Values of the deuteron coefficients $a_n$ and $a_p$ in units of $d_p$ and $d_n$ and of $a_1$, $a_\Delta$, $A_3$, $A_4$ in units of $e$ fm for the three different choices of cutoff parameters $\Lambda$.
The coefficient $a_\Delta$ can be written as, $$a_\Delta=a_\Delta(0)+\big(c_1a_\Delta(1)+c_2a_\Delta(2)+c_3a_\Delta(3)\big)+
a_\Delta(RC)\ ,\label{eq:adelta}$$ where the first term comes from $V_{TRV}^{(3\pi,0)}$, the term in parenthesis from $V_{TRV}^{(3\pi,1)}$, where the LECs $c_1$, $c_2$ and $c_3$ appear, and the last term from $V_{TRV}^{(3\pi,{\rm RC})}$. In Table \[tab:2had\] we report the values of the coefficients $a_\Delta(i)$ evaluated using the N4LO PC potential for the various cut-off.
$\Lambda$(MeV) $450$ $500$ $550$
---------------- ----------- ----------- -----------
$a_\Delta(0) $ $ -0.872$ $ -0.894$ $ -0.894$
$a_\Delta(1) $ $\m0.117$ $\m0.120$ $\m0.120$
$a_\Delta(2) $ $ -0.119$ $ -0.119$ $ -0.117$
$a_\Delta(3) $ $ -0.209$ $ -0.207$ $ -0.203$
$a_\Delta(RC)$ $ -0.042$ $ -0.037$ $ -0.032$
: \[tab:2had\] Values of the various components of the deuteron coefficients $a_\Delta$ as given in Eq. (\[eq:adelta\]) in units of $e$ fm evaluated using the N4LO PC potential for the three different choices of cutoff parameters $\Lambda$.
We observe that, taking individually the coefficients which multiplies the LECs $c_i$, their correction to the NLO is between $13-23\%$ which is in line with what we expect from the chiral expansion. However, using the values of $c_i$ given in Table \[tab:cecd\] the total correction of $V^{(3\pi,1)}_{TRV}$ is about $66\% $, which is very large compared to what we expect adding a new order in the potential. Therefore, the value of $a_\Delta$ is very sensitive to the $V_{TRV}^{(3\pi,1)}$ component of the potential and on the choice of the values of the constant $c_i$ and in particular of $c_3$. This is reflected also in the large error associated with this coefficient reported in Table \[tab:2hdpc\]. On the other hand, the $V_{TRV}^{(3\pi,{\rm RC})}$ contribution is about $\sim4\%$ as expected for a relativistic corrections. We notice also that the values of the coefficients for the three different cut-off are compatible within the error bars.
We now compare our results with the values reported in Ref. [@JB151] where the Authors used the same TRV potential at NLO with a N2LO PC potential with three-body forces [@EE09; @EE05]. In Table \[tab:2hdcomp\] we compare our results obtained with our TRV potential up to NLO and N2LO with a cutoff $\Lambda=500$ MeV with the ones reported in Ref. [@JB151].
TRV pot. $a_n(=a_p)$ $a_1$ $a_\Delta$ $A_3$ $A_4$
-------------------- ------------- --------- ------------ ----------- ----------
Ref. [@JB151](NLO) $0.939$ $0.183$ $-0.748$ $\m0.006$ $-0.006$
This work (NLO) $0.939$ $0.200$ $-0.893$ $-$ $-$
This work (N2LO) $0.939$ $0.197$ $-0.341$ $0.013$ $-0.013$
: \[tab:2hdcomp\] Comparison of the coefficients $a_1$, $a_\Delta$, $A_3$ and $A_4$ for the deuteron with the result of Refs. [@JB151]. To be notice that in this work we use $e>0$. For this work we report the calculation up to NLO and N2LO.
In order to compare the values of the $A_i$ coefficient we divide the reported values for $\frac{2(\hbar c)^3}{\Lx^2\fp}$ which permits to connect the two potentials. As can be seen from Table \[tab:2hdcomp\], our values at NLO seem to be systematically larger compared to the values reported in Ref. [@JB151]. Even if we use a different PC potential, the reason should be found in the different regularization function. However, from a qualitative point of view there is a substantial agreement with Ref. [@JB151]. Similar agreement has been founded for $a_1$ with the result reported in Refs. [@CP04; @NY15; @YH13] while are smaller compare the results reported in Refs. [@JV11; @IS08].
$\tri$ and $\hel$ EDMs {#sec:3Hedm}
----------------------
In this section we report the results obtained for the EDM of the $\tri$ and $\hel$. The wave functions of $\tri$ and $\hel$ have been obtained with the hyperspherical harmonics (HH) [@AK08; @LE09] from the Hamiltonians N4LO/N2LO-500, N4LO/N2LO-450 and N4LO/N2LO-550 discussed in Sec. \[sec:res\]. Moreover, we evaluated also the errors on the numerical coefficients due to the chiral expansion as, $$\left(\delta a_i\right)^2=\left(\delta a_i^{\text{PC}}\right)^2
+\left(\delta a_i^{\text{TRV}}\right)^2+\left(\delta a_i^{\psi}\right)^2\ ,
\label{eq:errh}$$ where $\delta a_i^{\text{PC}}$ and $\delta a_i^{\text{TRV}}$ are the same as in Eq. (\[eq:errd\]), while $\delta a_i^{\psi}$ is the error associated to the numerical accuracy of the 3-body wave function which we estimated to be of the order of $\sim 1\%$. The calculated values of the numerical coefficients for the three choices of the cutoff with their associated errors are reported in Table \[tab:3htot\].
The $a_\Delta$ can be written as, $$\begin{aligned}
a_\Delta&=&a_\Delta(0)+\big(c_1a_\Delta(1)+c_2a_\Delta(2)+c_3a_\Delta(3)\big)+
\nonumber\\
&&a_\Delta(RC)+a_\Delta(3N) ,\label{eq:adelta3}\end{aligned}$$ where $a_\Delta(0)$, $a_\Delta(1)$, $a_\Delta(2)$, $a_\Delta(3)$ and $a_\Delta(RC)$ are defined as in Eq. (\[eq:adelta\]) while $a_\Delta(3N)$ represents the TRV 3-body potential contribution. In Table \[tab:3had\] we report the values of the coefficients $a_\Delta(i)$ evaluated using the N4LO/N2LO-$\Lambda$ PC potential for the various cut-off.
---------------- ----------- ----------- ----------- ----------- ----------- -----------
$\Lambda$(MeV) $450$ $500$ $550$ $450$ $500$ $550$
$a_\Delta(0) $ $ -0.716$ $ -0.751$ $ -0.758$ $ -0.716$ $ -0.749$ $ -0.755$
$a_\Delta(1) $ $\m0.093$ $\m0.098$ $\m0.099$ $\m0.093$ $\m0.098$ $\m0.099$
$a_\Delta(2) $ $ -0.107$ $ -0.110$ $ -0.110$ $ -0.106$ $ -0.109$ $ -0.109$
$a_\Delta(3) $ $ -0.194$ $ -0.198$ $ -0.198$ $ -0.192$ $ -0.196$ $ -0.196$
$a_\Delta(RC)$ $ -0.048$ $ -0.046$ $ -0.042$ $ -0.048$ $ -0.044$ $ -0.041$
$a_\Delta(3N)$ $ -0.202$ $ -0.190$ $ -0.205$ $ -0.193$ $ -0.180$ $ -0.196$
---------------- ----------- ----------- ----------- ----------- ----------- -----------
: \[tab:3had\] Values of the various components of $a_\Delta$ as given in Eq. (\[eq:adelta3\]) for $\tri$ and $\hel$ in units of $e$ fm evaluated using the N4LO/N2LO PC potential for the three different choices of cutoff parameters $\Lambda$.
The $a_\Delta(3N)$ give a correction to $a_\Delta(0)$ of the order of the $\sim25\%$, which is in line with the chiral perturbation theory prediction because both the contributions appear at the same order. For completeness we report in Table \[tab:3Nconv\] of Appendix \[app:NNNconv\] the convergence pattern of this contribution as function of the HH basis expansion. For both $\tri$ and $\hel$, using the values of the $c_i$ reported in Table \[tab:cecd\], we have found a correction to $a_\Delta(0)$ due to ${V_{TRV}}^{(3\pi,1)}$ of $\sim79\%$, and of order $\sim6\%$ due to ${V_{TRV}}^{(3\pi,{\rm RC})}$. While the RC are in line with what we expect, the ${V_{TRV}}^{(3\pi,1)}$ corrections have much more impact on the values of $a_\Delta$ than predicted by the chiral perturbation theory. The large correction due to ${V_{TRV}}^{(3\pi,1)}$ that appears at N2LO is also reflected in the large uncertainties associated with this coefficient. Again, the estimated uncertainties depends critically on the adopted values of $c_1$, $c_2$, and $c_3$ (see Table \[tab:cecd\]).
The contribution of the TPE diagrams to $a_0$ and $a_2$ are of the order of $\sim45\%$ and $\sim40\%$ respectively which is larger than expected from the chiral convergence and they are due mainly to the box diagram in Fig. \[fig:diagNN\]. On the other hand, the RC to $a_0$, $a_1$ and $a_2$ are of the order of $\sim1-3\%$ perfectly consistent with the prediction of the chiral perturbation theory. The effects of the PC $NNN$ potential on the values of all the coefficients is around $\sim2\%$. From Table \[tab:3htot\] it is possible also to observe that the values of the numerical coefficients are mostly equal in modulus between $\tri$ and $\hel$ except $a_p$ and $a_n$. Moreover, it is possible to observe that the isovector terms have the same sign for $\tri$ and $\hel$. The isovector part of the TRV potential depends on the third component of the isospin, therefore the $|\psi^A_{-}{\rangle}$ wave function component generated has a sign $-$ for $\tri$ and a $+$ for $\hel$. However, in the ${\hat{\boldsymbol{D}}}_{\rm PC}$ there is the $\tau_z(i)$ operator which bring again a sign $-$ for $\tri$ and $+$ for $\hel$ giving in total a $+$.
In Table \[tab:3hcomp\] we compare our calculations at NLO for a value of the cutoff $\Lambda=500$ MeV with the values reported in Ref. [@JB151]. As can be seen inspecting the values of the coefficients evaluated at NLO, there is a nice agreement with the results of Ref. [@JB151]. The numerical differences (which however are within the error bars reported in Ref. [@JB151]) must be searched in the different PC potential and in the different regularization function in the TRV potential employed here. On the other hand for $a_\Delta^{(3)}$, the pure three body part of $a_\Delta$, the difference is of one order of magnitude which can not be explained by the different regularization function used. We found a good agreement for the values of $a_0$ and $a_1$ with the results obtained using a phenomenological potential and the meson exchange TRV potential in Refs. [@NY15; @YH13] while we obtain smaller values compare to Refs. [@JV11; @IS08].
In the case of $\tri$ and $\hel$ we studied also the dependence of the values of the coefficients on the chiral order of the PC potential. As example, in Fig. \[fig:a0coeff\] and \[fig:Dcoeff\] we show the values of the coefficient $a_0$ and $a_\Delta^{(3)}$ for $\tri$ evaluated using the N2LO TRV potential with different chiral order of the PC potential and for the three different choices of the cutoff.
![\[fig:a0coeff\] Values of the coefficient $a_0$ for the $\tri$ nucleus and for the three choice of the cutoff when varying the chiral order of the PC potential.](figure2.eps){width="8cm"}
![\[fig:Dcoeff\] The same as Fig. \[fig:a0coeff\] but for the coefficient $a_\Delta^{(3)}$.](figure3.eps){width="8cm"}
In all the case we studied, the values evaluated with the N2LO, N3LO and N4LO PC potential differ by less than $5\%$ which confirm the robustness of the calculation.
----------------- ------------------- ------------------- ------------------- ------------------- ------------------- -------------------
$\Lambda$ (MeV) $450$ $500$ $550$ $450$ $500$ $550$
$a_n$ $ -0.032\pm0.001$ $ -0.033\pm0.001$ $ -0.035\pm0.001$ $\m0.900\pm0.009$ $\m0.908\pm0.009$ $\m0.906\pm0.009$
$a_p$ $\m0.901\pm0.009$ $\m0.909\pm0.009$ $\m0.907\pm0.009$ $ -0.032\pm0.001$ $ -0.033\pm0.001$ $ -0.034\pm0.001$
$a_0$ $ -0.052\pm0.012$ $ -0.055\pm0.013$ $ -0.052\pm0.012$ $\m0.053\pm0.012$ $\m0.056\pm0.013$ $\m0.052\pm0.012$
$a_1$ $\m0.147\pm0.005$ $\m0.154\pm0.004$ $\m0.155\pm0.003$ $\m0.148\pm0.005$ $\m0.155\pm0.004$ $\m0.155\pm0.003$
$a_2$ $ -0.114\pm0.010$ $ -0.120\pm0.009$ $ -0.121\pm0.008$ $\m0.112\pm0.009$ $\m0.118\pm0.009$ $\m0.119\pm0.008$
$a_\Delta$ $ -0.378\pm0.105$ $ -0.388\pm0.101$ $ -0.407\pm0.088$ $ -0.373\pm0.106$ $ -0.383\pm0.102$ $ -0.402\pm0.089$
$A_1$ $\m0.005\pm0.001$ $\m0.006\pm0.001$ $\m0.006\pm0.001$ $ -0.005\pm0.002$ $ -0.006\pm0.002$ $ -0.006\pm0.001$
$A_2$ $ -0.009\pm0.003$ $ -0.010\pm0.003$ $ -0.010\pm0.003$ $\m0.009\pm0.003$ $\m0.010\pm0.003$ $\m0.010\pm0.002$
$A_3$ $ -0.008\pm0.002$ $ -0.008\pm0.002$ $ -0.008\pm0.002$ $ -0.008\pm0.003$ $ -0.008\pm0.002$ $ -0.008\pm0.002$
$A_4$ $\m0.012\pm0.004$ $\m0.013\pm0.004$ $\m0.013\pm0.004$ $\m0.012\pm0.004$ $\m0.013\pm0.004$ $\m0.013\pm0.003$
$A_5$ $ -0.021\pm0.006$ $ -0.022\pm0.006$ $ -0.022\pm0.006$ $\m0.020\pm0.006$ $\m0.022\pm0.006$ $\m0.022\pm0.005$
----------------- ------------------- ------------------- ------------------- ------------------- ------------------- -------------------
: \[tab:3htot\] Values of the numerical coefficients for $\tri$ and $\hel$ in units of $e$ fm ($a_n$ $(a_p)$ in units of $d_n$ ($d_p$)) for the three different choices of the cutoff.
------------------ ----------- ----------- --------------- ----------- ----------- ---------------
This work This work Ref. [@JB151] This work This work Ref. [@JB151]
(NLO) (N2LO) (NLO) (NLO) (N2LO) (NLO)
$a_n$ $ -0.033$ $ -0.033$ $ -0.030$ $\m0.908$ $\m0.908$ $\m0.904$
$a_p$ $\m0.909$ $\m0.909$ $\m0.918$ $ -0.033$ $ -0.033$ $ -0.029$
$a_0$ $ -0.101$ $ -0.055$ $ -0.108$ $\m0.101$ $\m0.056$ $\m0.111$
$a_1$ $\m0.158$ $\m0.154$ $\m0.139$ $\m0.158$ $\m0.155$ $\m0.139$
$a_2$ $\m0.087$ $\m0.120$ n.a. $\m0.086$ $\m0.118$ n.a.
$a_\Delta^{(2)}$ $ -0.751$ $ -0.198$ $-0.598$ $ -0.749$ $ -0.202$ $-0.608$
$a_\Delta^{(3)}$ $ -0.190$ $ - $ $-0.017$ $ -0.180$ $ - $ $-0.017$
$A_1$ $-$ $\m0.006$ $\m0.005$ $-$ $ -0.006$ $ -0.005$
$A_2$ $-$ $ -0.010$ $ -0.011$ $-$ $\m0.010$ $\m0.011$
$A_3$ $-$ $ -0.008$ $ -0.005$ $-$ $ -0.008$ $ -0.005$
$A_4$ $-$ $\m0.013$ $\m0.009$ $-$ $\m0.013$ $\m0.009$
$A_5$ $-$ $ -0.022$ n.a. $-$ $\m0.022$ n.a.
------------------ ----------- ----------- --------------- ----------- ----------- ---------------
: \[tab:3hcomp\] Comparison of the values of the coefficients for $\tri$ and $\hel$ obtained for $\Lambda=500$ MeV with the results of Ref. [@JB151]. $a_\Delta^{(2)}$ and $a_\Delta^{(3)}$ correspond respectively to the 2-body and 3-body contribution to $a_\Delta$. To be noticed that in this work we use $e>0$.
Conclusions {#sec:conc}
===========
In this work we derived the TRV $NN$ and $NNN$ potential at N2LO. In order to derive the potential we have considered the most generic Lagrangian without any specific hypothesis for the TRV source. With the derived potential, we have calculated the EDM of $\det$, $\tri$ and $\hel$ investigating the effect of the N2LO components. We have found that the sensitivity of the light nuclei EDM to the LEC $\Delta_3$ found at NLO, is well reduced by the N2LO contribution which is a quite unexpected behavior inside the chiral perturbation framework. We also checked the robustness of our calculation, evaluating the EDM of the nuclei using different chiral orders in the PC potential. The discrepancy between the use of the N2LO and the N4LO PC potential is approximately $5\%$.
We have compared our results with the existing other values reported in literature and in particular with the calculation of Ref. [@JB151]. We have found a substantial agreement with the results reported in Ref. [@JB151] where the small numerical differences can be originated by the different function used to regularize the potential. We have found a qualitative agreement with those of Refs. [@CP04; @NY15; @YH13] while we have obtained smaller values compared to Refs. [@JV11; @IS08].
Our results depend on eleven coupling constants that should be determined by comparing with experimental data. As many authors already pointed out [@EM10; @JB150; @JV13], the size of the coupling constant depends on the CP violating model. Using our study it will be possible, in case of more than one measurements, to determine the LECs and then individuate the TRV source which generates the EDM by comparing the values of the calculated LECs and their predicted sizes.
In future, we plan to use $\chi$EFT to derive the TRV currents which give contribution at N2LO [@JV11b; @JB13]. This would allow to have a fully consistent calculation of the EDM of light nuclei up to N2LO. We also plan to study the $\vec{n}-\vec{p}$ and $\vec{n}-\vec{d}$ spin rotation for an independent and complementary study of TRV effects respect to EDM.
acknowledgments {#acknowledgments .unnumbered}
===============
Computational resources provided by the INFN-Pisa Computer Center are gratefully acknowledged.
Interaction vertices {#app:vertex}
====================
It is convenient to decompose the interaction Hamiltonian $H_I$ as follows =0.5pt $$\begin{aligned}
H_I\!=\!H^{00}\!+\!H^{01}\!+\!H^{10}\!+\!H^{02}\!+\!H^{11}\!+\!H^{20}\!+\!\cdots\ ,\end{aligned}$$ where $H^{nm}$ has $n$ creation and $m$ annihilation operators for the pion. Explicitly, $$\begin{aligned}
H^{00} & = & {1\over \Omega}\sum_{\a_1' \a_1\a_2' \a_2}
b^{\dag}_{\a_1'}b_{\a_1} b^{\dag}_{\a_2'} b_{\a_2}
M^{00}_{\a_1'\a_1\a_2'\a_2}
\delta_{\bmp_1'+\bmp_2' ,\bmp_1+\bmp_2}\ , \label{eq:m00}\nonumber\\
\\
H^{01}& = & \frac{1}{\sqrt{\Omega}} \sum_{\a' \a}\sum_{\bmq \, a}
b^{\dag}_{\a'}b_{\a}a_{\bmq\, a} M^{01}_{\a'\a,\bmq\, a}
\delta_{\bmq+\bmp,\bmp'}\ ,\label{eq:m01}\\
H^{10}& = & \frac{1}{\sqrt{\Omega}} \sum_{\a' \a}\sum_{\bmq\, a}
b^{\dag}_{\a'}b_{\a}a^\dag_{\bmq\, a} M^{10}_{\a'\a,\bmq\, a}
\delta_{\bmq+\bmp',\bmp}\ ,\label{eq:m10}\\
H^{02} &= & \frac{1}{\Omega}\sum_{\a' \a}\sum_{\bmq' a'\,\bmq\, a}
b_{\a'}^{\dag}b_{\a}a_{\bmq' a'}a_{\bmq\, a}
M^{02}_{\a'\a,\bmq' a'\, \bmq \, a}
\delta_{\bmq+\bmq'+\bmp,\bmp'}\ ,\label{eq:m02}\nonumber\\
\\
H^{11} &= & \frac{1}{\Omega}\sum_{\a' \a}\sum_{\bmq' a'\, \bmq\, a}
b_{\a'}^{\dag}b_{\a}a^\dag_{\bmq' a'}a_{\bmq\,a }
M^{11}_{\a'\a,\bmq'a'\bmq \, a}
\delta_{\bmq+\bmp,\bmq'+\bmp'}\ ,\label{eq:m11}\nonumber\\
\\
H^{20} &= & \frac{1}{\Omega}\sum_{\a' \a}\sum_{\bmq'a' \,\bmq\,a}
b_{\a'}^{\dag}b_{\a}a^\dag_{\bmq' a'}a^\dag_{\bmq \, a}
M^{20}_{\a'\a,\bmq' a'\, \bmq \,a}
\delta_{\bmp,\bmq+\bmq'+\bmp'}\ ,\label{eq:m20}\nonumber\\\end{aligned}$$ etc. Here $\alpha_j\equiv \bmp_j,s_j,t_j$ denotes the momentum, spin projection, isospin projection of nucleon $j$ with energy $E_j=\sqrt{p_j^2+M^2}$, $\bmq$ and $a$ denote the momentum and isospin projection of a pion with energy $\omega_q=\sqrt{\bmq^2+m_\pi^2}$, and $M^{nm}$ are the vertex functions listed below. The various momenta are discretized by assuming periodic boundary conditions in a box of volume $\Omega$. We note that in the expansion of the nucleon field $\psi$ we have only retained the nucleon degrees of freedom, since anti-nucleon contributions do not enter the TRV $NN$ and $NNN$ potential at the order of interest here. We note also that in general the creation and annihilation operators are not normal-ordered. After normal-ordering them, tadpole-type contributions result, which are relevant only for renormalization, therefore we discard them hereafter.
The vertex functions $M^{nm}$ involve products of Dirac 4-spinors, which are expanded non-relativistically in powers of momenta, and terms up to order $Q^3$ are retained. Useful formulas are reported in Appendix F of Ref. [@viviani14].
![ \[fig:vertex\] Vertices entering the TRV potential at N2LO. The solid (dash) lines represent nucleons (pions). The open (solid) symbols denote PC (TRV) vertices.](vertex.eps)
The interaction vertices needed for the construction of the TRV potential, without considering renormalization contributions, are summarized in Fig. \[fig:vertex\]. Note that in the power counting of these vertices below, we do not include the $1/\sqrt{\omega_k}$ normalization factors in the pion fields. We obtain:
1. $\pi NN$ vertices. The PC interaction term is derived in Appendix F of Ref. [@viviani14]. For completeness, here we report the final formula for the PC vertex up to order $Q^3$ that reads, $$\begin{aligned}
{}^{PC}M^{\pi NN,01}_{\alpha' \alpha,
\bmq\,a} &=& {g_A \over 2 f_\pi}
{\tau_a\over \sqrt{2\omega_q}}\Bigl[ i\, \bmq\cdot{{\bm \sigma}}-{i\over
M}\omega_q\;\bmK\cdot{{\bm \sigma}}\nonumber \\
&+& {i\over 4M^2}\Bigl(2\bmK\cdot\bmq\;\bmK\cdot{{\bm \sigma}}-
2K^2\;\bmq\cdot{{\bm \sigma}}\nonumber\\
&& \qquad -{1\over 2}\bmk\cdot{{\bm \sigma}}\; \bmq\cdot\bmk\Bigr)\Bigr]
\nonumber\\
&+& {m_\pi^2 \over f_\pi}(2d_{16}-d_{18})
{\tau_a\over \sqrt{2\omega_q}}\; i\bmq\cdot{{\bm \sigma}}\ ,\label{eq:MpiNN01b}\\
{}^{PC}M^{\pi NN, 10}_{\alpha'\alpha,
\bmq\,a} &= & -{}^{PC}M^{\pi NN,01}_{\alpha' \alpha,\bmq\,a} .
\label{eq:MpiNN10b}
\end{aligned}$$ In diagrams, these PC vertex functions are represented as open circles. The TRV $\pi NN$ vertices are due to interaction terms proportional to LECs $g_0$, $g_1$ and $g_2$ which corresponds to isoscalar, isovector and isotensor interaction, plus terms which derive from ${\cal L}_{\text{TRV}}^{\pi N\, (1)}$ and ${\cal L}_{\text{TRV}}^{\pi N\, (2)}$ given in Eqs. (\[eq:ltrv1\]) and (\[eq:ltrv2\]). They read (up to order $Q^2$), $$\begin{aligned}
{}^{TRV}&M^{\pi NN,01}_{\alpha' \alpha,
\bmq\,a}=-\frac{(g_0 \tau_a+g_1\delta_{a,3}+g_2\tau_3\delta_{a,3})}
{\sqrt{2\omega_q}}\nonumber\\
&\qquad\times\Bigl[ 1-{1\over 4M^2}\Bigl(
2K^2+i\ \bmk\times\bmK\Bigr)\Bigr]\nonumber\\
&\qquad-ig^{(1)}_V\frac{\epsilon_{ab3}\tau_b}{\fp\sqrt{2\omega_q}}\big[\omega_q
-{1\over 2M}\bmK\cdot\bmq\big]\nonumber\\
&\qquad+\frac{\mp^2}{\fp^2\sqrt{2\omega_q}}\big(g^{(2)}_{S1}\tau_a
+g^{(2)}_{V1}\delta_{a,3}+g^{(2)}_{T1}\tau_3\delta_{a,3}\big)\nonumber\\
&\qquad-\frac{\omega_q^2}{\fp^2\sqrt{2\omega_q}}\big(g^{(2)}_{S2}\tau_a
+g^{(2)}_{V2}\delta_{a,3}+g^{(2)}_{T2}\tau_3\delta_{a,3}\big)
\ ,\nonumber\\\label{eq:MpiNN01a}
\end{aligned}$$ $$\begin{aligned}
{}^{TRV}&M^{\pi NN, 10}_{\alpha'\alpha,
\bmq\,a}=-\frac{(g_0 \tau_a+g_1\delta_{a,3}+g_2\tau_3\delta_{a,3})}
{\sqrt{2\omega_q}}\nonumber\\
&\qquad\times\Bigl[ 1-{1\over 4M^2}\Bigl(
2K^2+i\ \bmk\times\bmK\Bigr)\Bigr]\nonumber\\
&\qquad+ig^{(1)}_V\frac{\epsilon_{ab3}\tau_b}{\fp\sqrt{2\omega_q}}\big[\omega_q
-{1\over 2M}\bmK\cdot\bmq\big]\nonumber\\
&\qquad+\frac{\mp^2}{\fp^2\sqrt{2\omega_q}}\big(g^{(2)}_{S1}\tau_a
+g^{(2)}_{V1}\delta_{a,3}+g^{(2)}_{T1}\tau_3\delta_{a,3}\big)\nonumber\\
&\qquad-\frac{\omega_q^2}{\fp^2\sqrt{2\omega_q}}\big(g^{(2)}_{S2}\tau_a
+g^{(2)}_{V2}\delta_{a,3}+g^{(2)}_{T2}\tau_3\delta_{a,3}\big)\ ,\nonumber\\
\label{eq:MpiNN10a}
\end{aligned}$$
2. $\pi\pi NN$ vertices. The PC interaction is needed up to NLO in the NR expansion at the order we are interested. The corresponding vertex functions will receive contribution from the Weinberg-Tomozawa term and from ${\cal L}^{(2)}_{N\pi}$. The vertex functions read, $$\begin{aligned}
{}^{PC}M&^{\pi\pi NN,02}_{\alpha' \alpha,\bmq' a'\,\bmq\, a}=
{i\over 8f_\pi^2} {\epsilon_{aa'b}\tau_b
\over\sqrt{2\omega_q}\sqrt{2\omega_{q'}}}\Big[(\omega_q-\omega_{q'})
\nonumber\\
&-\frac{2\bmK\cdot(\bmq-\bmq')-i(\bmk\times{{\bm \sigma}})\cdot(\bmq-\bmq')}
{2M}\Big]\nonumber\\
&+\frac{\delta_{ij}}{\fp^2}\frac{c_1(2m_\pi^2)+c_2\omega_q\omega_{q'}+c_3
(\omega_q\omega_{q'}-\bmq\cdot\bmq')}{\sqrt{2\omega_q}\sqrt{2\omega_{q'}}}
\nonumber\\
&-\frac{c_4}{2\fp^2}{\epsilon_{aa'b}\tau_b
\over\sqrt{2\omega_q}\sqrt{2\omega_{q'}}}(\bmq\times\bmq')\cdot{{\bm \sigma}}\ , \\
{}^{PC}M&^{\pi\pi NN, 20}_{\alpha' \alpha,\bmq' a'\,\bmq\, a}=
{i\over 8f_\pi^2} {\epsilon_{aa'b}\tau_b
\over\sqrt{2\omega_q}\sqrt{2\omega_{q'}}}\Big[(\omega_{q'}-\omega_{q})
\nonumber\\
&-\frac{2\bmK\cdot(\bmq'-\bmq)-i(\bmk\times{{\bm \sigma}})\cdot(\bmq'-\bmq)}
{2M}\Big]\nonumber\\
&+\frac{\delta_{ij}}{\fp^2}\frac{c_1(2m_\pi^2)+c_2\omega_q\omega_{q'}+c_3
(\omega_q\omega_{q'}-\bmq\cdot\bmq')}{\sqrt{2\omega_q}\sqrt{2\omega_{q'}}}
\nonumber\\
&-\frac{c_4}{2\fp^2}{\epsilon_{aa'b}\tau_b
\over\sqrt{2\omega_q}\sqrt{2\omega_{q'}}}(\bmq\times\bmq')\cdot{{\bm \sigma}}\ .
\end{aligned}$$ The ${}^{PC}M^{\pi\pi NN,11}_{\alpha' \alpha,\bmq' a'\,\bmq\, a}$ vertex is not needed for the evaluation of the time-ordered diagrams.
3. $3\pi$ vertices. The TRV Lagrangian has a three-pion interaction term shown in diagram (4) of Fig. \[fig:vertex\]. The vertices, neglecting the tadpole terms, are given by $H_I^{3\pi}=H_I^{3\pi,03}+H_I^{3\pi,12}
+H_I^{3\pi,21}+H_I^{3\pi,30}$, where $$\begin{aligned}
H_I^{3\pi,03} &={1\over\Omega^{3/2}}
\sum_{\substack{\bmq a\,\bmq'a'\\ \bmp b}} a_{\bmq\,a}a_{\bmq'\,a'}a_{\bmp\,b}\;
M^{3\pi,03}_{\bmq a\,\bmq'a'\,\bmp b}\,\delta_{\bmq+\bmq'+\bmp,0}\ ,\nonumber\\
\\
H_I^{3\pi,12} &= {1\over\Omega^{3/2}}
\sum_{\substack{\bmq a\,\bmq'a'\\ \bmp b}} a_{\bmq\,a}a_{\bmq'\,a'}a_{\bmp\,b}^\dagger\;
M^{3\pi,12}_{\bmq a\,\bmq'a'\, \bmp b}\,\delta_{\bmq+\bmq',\bmp}\ ,\nonumber\\
\\
H_I^{3\pi,21} &= {1\over\Omega^{3/2}}
\sum_{\substack{\bmq a\,\bmq'a'\\\bmp b}} a_{\bmq\,a}a_{\bmq'\,a'}^\dagger a_{\bmp\,b}^\dagger\;
M^{3\pi,21}_{\bmq a\,\bmq'a'\,\bmp b}\,\delta_{\bmq,\bmq'+\bmp}\ ,\nonumber\\
\\
H_I^{3\pi,30} & ={1\over\Omega^{3/2}}
\sum_{\substack{\bmq a\,\bmq'a'\\ \bmp b}} a_{\bmq\,a}^\dagger a_{\bmq'\,a'}^\dagger a_{\bmp\,b}^\dagger\;
M^{3\pi,30}_{\bmq a\,\bmq'a'\,\bmp b}\,\delta_{0,\bmq+\bmq'+\bmp}\ ,\nonumber\\
\end{aligned}$$ with, $$\begin{aligned}
{}^{TRV}M^{3\pi,03}_{\bmq a\,\bmq'a'\,\bmp b}&=-\frac{\Delta M}
{3\sqrt{2\omega_q\,2\omega_{q'}\,2\omega_p}}\nonumber\\
&\times\bigl(\delta_{a,a'}\delta_{b,3}+
\delta_{a,b}\delta_{a',3}+\delta_{a',b}\delta_{a,3}\bigr)\ ,\nonumber\\
\\
{}^{TRV}M^{3\pi,12}_{\bmq a\,\bmq'a'\,\bmp b}&=
3\,{}^{TRV}M^{3\pi,03}_{\bmq a\,\bmq'a'\,\bmp b}\ ,\\
{}^{TRV}M^{3\pi,21}_{\bmq a\,\bmq'a'\,\bmp b}&=
3\,{}^{TRV}M^{3\pi,03}_{\bmq a\,\bmq'a'\,\bmp b}\ ,\\
{}^{TRV}M^{3\pi,30}_{\bmq a\,\bmq'a'\,\bmp b}&=
{}^{TRV}M^{3\pi,03}_{\bmq a\,\bmq'a'\,\bmp b}\ .
\end{aligned}$$
4. $4N$ contact interaction. The EFT Hamiltonian includes also the term given in Eq. (\[eq:m00\]) derived from a contact Lagrangian. We only need its TRV part of order $Q$, which includes five independent interaction terms. With a suitable choice of the LECs, the vertex function ${}^{TRV}M^{00}$ can be written as $$\begin{aligned}
{}^{TRV}&M^{00}_{\a_1'\a_1\a_2'\a_2}=
{1 \over 2\Lambda_\chi^2 f_\pi} \Bigl[
C_1 i \bmk_1\cdot ({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}) \nonumber\\
&\qquad+ C_2\,{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\, i\bmk_1\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&\qquad+ {C_3\over 2}\, \bigl(\left(\tau_{1z}+\tau_{2z}\right)
\, i \bmk_1\cdot\left({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}\right)\nonumber\\
&\qquad\qquad-\left(\tau_{1z}-\tau_{2z}\right)\,
i\bmk_1\cdot\left({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2}\right)\bigr)\nonumber\\
&\qquad+ {C_4\over 2}\, \bigl(\left(\tau_{1z}+\tau_{2z}\right)
\, i \bmk_1\cdot\left({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}\right)\nonumber\\
&\qquad\qquad+\left(\tau_{1z}-\tau_{2z}\right)\,
i\bmk_1\cdot\left({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2}\right)\bigr)\nonumber\\
&\qquad+C_5\, \left(3\tau_{1z}\tau_{2z}-{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\right) i\bmk_1\cdot
\left({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}\right)
\Bigr]\ , \label{eq:m00trv}\nonumber\\
\end{aligned}$$ where $\bmk_1=\bmp_1'-\bmp_1=-\bmp_2'+\bmp_2$.
The TRV $NN$ potential {#app:pot}
======================
In this section we discuss the derivation of the TRV $NN$ and $NNN$ potential, providing explicit expressions of the diagrams given in Fig. \[fig:diagNN\]. The power counting is as follows: (i) a PC (TRV) $\pi NN$ vertex is of order $Q$ ($Q^{0}$); (ii) a PC $\pi\pi NN$ vertex is of order $Q^1$; (iii) a TRV $3\pi$ vertex is of order $Q^{-3}$; (iv) a PC (TRV) $NN$ contact vertex is of order $Q^0$ ($Q$); (v) an energy denominator without (with one or more) pions is of order $Q^{-2}$ ($Q^{-1}$); (vi) factors $Q^{-1}$ and $Q^{3}$ are associated with, respectively, each pion line and each loop integration. The momenta are defined as given in Eq. (\[eq:notjb1\]), and in what follows use is made of the fact that $\bmk\cdot\bmK$ vanishes in the c.m. frame. It is useful to define the isospin operator as, $$\begin{aligned}
\TO &=&{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\ ,\\
\ttp&=&(\tau_{1z}+\tau_{2z})\ ,\\
\ttm&=&(\tau_{1z}-\tau_{2z})\ ,\\
\tten&=&(3\tau_{1z}\tau_{2z}-{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2})\ .\end{aligned}$$
The TRV $NN$ potential is derived from the amplitudes in Fig. \[fig:diagNN\] via Eqs. (\[eq:vtrvml\])–(\[eq:vtrv1\]). Up to order $Q$ included, we obtain for the OPE component in panel (a) of Fig. \[fig:diagNN\] : $$V({\rm a}) =
V^{(-1)}({\rm NR})+ V^{(1)}({\rm RC}) + V^{(1)}({\rm LEC})\ , \label{eq:ope2}$$ where, $$\begin{aligned}
V^{(-1)}&({\rm NR})=\frac{\ga g_0^*}{2\fp}\TO\frac{i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})}{\omk^2} \nonumber\\
&+\frac{g_Ag_1}{4f_\pi}\Big[\ttp
\frac{i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})}{\omega_k^2}+
\ttm\frac{i \bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})}{\omega_k^2}\Big]\nonumber\\
&+\frac{\ga g_2}{6\fp}\tten\frac{i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})}{\omk^2}
\!\!\ ,\label{eq:opeloapp}
\end{aligned}$$ $$\begin{aligned}
{V_{TRV}}^{(1)}&{(\rm RC)}=\frac{\ga g_0^*}{8\fp M^2}
\TO\ \frac{1}{\omk^2}\nonumber\\
&\times\Big[-\frac{i}{2} \left(8K^2+k^2\right)\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\nonumber\\
&+\frac{\ga g_1}{16\fp M^2}\ \frac{1}{\omk^2}\nonumber\\
&\times\Big\{\ttm
\Big[-\frac{i}{2} \left(8K^2+k^2\right)\bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\nonumber\\
&+\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}+\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\nonumber\\
&+\ttp
\Big[-\frac{i}{2} \left(8K^2+k^2\right)\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\Big\}\nonumber\\
&+\frac{\ga g_2}{24\fp M^2}\tten
\ \frac{1}{\omk^2}
\Big[-\frac{i}{2} \left(8K^2+k^2\right)\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\Big]\label{eq:rcopeapp}\ ,\\
{V_{TRV}}^{(1)}&({\rm LEC})={V_{TRV}}^{(-1)}({\rm NR})
\frac{2m^2_\pi}{\ga}(2d_{16}-d_{18})\nonumber\\
&+\frac{\ga}{2\fp^3}\TO i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})
\Big[g^{(2)}_{S2}-g^{(2)}_{S1}\frac{\mp^2}{\omk^2}\Big]\nonumber\\
&+\frac{g_A}{4f_\pi^3}\Big[\ttp
i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})+
\ttm i \bmk\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\Big]\nonumber\\
&\qquad \times\Big[g^{(2)}_{V2}-g^{(2)}_{V1}\frac{\mp^2}{\omk^2}\Big]\nonumber\\
&+\frac{\ga}{6\fp^3}\tten i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})
\Big[g^{(2)}_{T2}-g^{(2)}_{T1}\frac{\mp^2}{\omk^2}\Big]\nonumber\\
&+\frac{g^{(1)}_Vg_A}{2\fp^2M}({\boldsymbol{\tau}_1}\times{\boldsymbol{\tau}_2})_z\,\bmK\cdot({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\,
\label{eq:lec},\end{aligned}$$ where $g_0^*=g_0+g_2/3$. The contribution given in the first line of Eq. (\[eq:lec\]) renormalizes the coupling constants $g_0^*$, $g_1$ and $g_2$ in the OPE term. The terms of Eq. (\[eq:lec\]) which are multiplied by the factor $\mp^2/\omk^2$ are also reabsorbed in the constant $g_0^*$, $g_1$ and $g_2$ while all the other terms are reabsorbed in the LECs $C_2$, $C_4$ and $C_5$ in Eq. (\[eq:ct\]). Regarding the last term in Eq. (\[eq:lec\]), it is possible to use a Fierz transformation obtaining a combination of the operators which multiplies the LECs $C_3$ and $C_4$. Therefore all ${V_{TRV}}^{(1)}(\text{LEC})$ can be reabsorbed in the OPE and contact potentials. The factor $k^2=\omega_k^2-m_\pi^2$ in the isoscalar, isovector and isotensor component of $V^{(1)}({\rm RC})$ leads to a piece that can be reabsorbed in the contact term proportional to $C_2$, $C_4$ and $C_5$ in Eq. (\[eq:ct\]) and a piece proportional to $m_\pi^2$ that simply renormalizes the LECs $g_0^*$, $g_1$ and $g_2$.
The component of the TRV potential coming from the contact terms in panel (b) of Fig. \[fig:diagNN\] derives directly from the vertex function ${}^{TRV}M^{00}$ given in Eq. (\[eq:m00trv\]). The final expression has already been given in Eq. (\[eq:ct\]). The diagrams reported in panels (c) and (d) contain a combination of a contact interaction with the exchange of a pion. However, it can be shown that their contribution is at least of order $Q^3$ .
Next we consider the TPE components in panels (e)-(h). The contribution from diagrams (e) reads, $$\begin{aligned}
V^{(1)}(\rm e)&=&-\frac{\ga g_0^*}{4\fp^3}\TO \, i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\int_\bmq\frac{1}{\omp\omm(\omp+\omm)} \nonumber\\
&&+\frac{\ga g_2}{12\fp^3}\tten \, i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\int_\bmq\frac{1}{\omp\omm(\omp+\omm)}\ ,
\nonumber\\
\end{aligned}$$ where $\omega_\pm=\sqrt{(\bmk\pm \bmq)^2 + 4 \, m_\pi^2}$ and $\int_\bmq=\int \frac{d\bmq}{(2\pi)^3}$. The isovector component of the OPE vertex vanishes since the integrand is proportional to $\bmq$, therefore there is no isovector component from panel (e). Dimensional regularization allows one to obtain the finite part as, $$\begin{aligned}
\overline{V}^{(1)}(\rm e)&=&\frac{\ga g_0^*}{\fp\Lx^2}\TO i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}) L(k)\nonumber\\
&&-\frac{\ga g_2}{3\fp\Lx^2}\tten i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}) L(k)\ ,\label{eq:pote}
\end{aligned}$$ where $\Lambda_\chi=4\pi f_\pi$ and the loop function $L(k)$ is defined as $$L(k)= {1\over 2} {s\over k} \ln\left({s+k\over s-k}\right)\ ,\quad
s=\sqrt{k^2+4\, m^2_\pi}\ .\label{eq:sL}$$ The singular part is given by, $$\begin{aligned}
V^{(1)}_\infty({\rm e}) &=&\frac{\ga g_0^*}{2\fp\Lx^2}\TO i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})(\deps-2)\nonumber\\
&&-\frac{\ga g_2}{6\fp\Lx^2}\tten i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})(\deps-2)\ ,
\end{aligned}$$ where $$\deps = -{2\over \epsilon}+\gamma-\ln
\pi +\ln\left({m^2_\pi\over\mu^2}\right)
\ ,$$ $\epsilon=3-d$, $d$ being the number of dimensions ($d\rightarrow
3$), and $\mu$ is a renormalization scale. This singular contribution is absorbed in the ${V_{TRV}}^{({\rm CT})}$ term proportional to $C_2$ for the isoscalar part and to $C_5$ for the isotensor part.
The contributions from panels (f)-(h) in Fig. \[fig:diagNN\] are collectively denoted as “box” below, and the non-iterative pieces in reducible diagrams of type (h) are identified via Eq. (\[eq:vtrv1\]). From the panel (f) we obtain, $$\begin{aligned}
V^{(1)}&({\rm f})=\Big\{-\frac{\ga^3 g_0^*}{16\fp^3}(3+2\ \TO)i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\frac{\ga^3 g_2}{48\fp^3}\tten \ i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&-\frac{\ga^3 g_1}{32\fp^3}
\Big[\ttp \ i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2)+
\ttm \ i\bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\Big]\Big\}\nonumber\\
&\times\int_\bmq\frac{\omp^2+\omp\omm+\omm^2}{\omp^3\omm^3(\omp+\omm)}
(k^2-q^2)\ ,\label{eq:panelf}\end{aligned}$$ while the contribution of panel (g) results, $$\begin{aligned}
V^{(1)}&({\rm g})=\Big\{\frac{\ga^3 g_0^*}{16\fp^3}(3-2\ \TO)i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\frac{\ga^3 g_2}{48\fp^3}\tten \ i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\frac{\ga^3 g_1}{32\fp^3}
\Big[\ttp \ i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2)+
\ttm \ i\bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\Big]\Big\}\nonumber\\
&\times\int_\bmq\frac{\omp^2+\omp\omm+\omm^2}{\omp^3\omm^3(\omp+\omm)}
(k^2-q^2)\ .\label{eq:panelg}\end{aligned}$$ The complete “box” contribution is given by the sum of $V({\rm f})$ and $V({\rm g})$ and it reads, $$\begin{aligned}
V^{(1)}({\rm box})&=\Big\{-\frac{\ga^3 g_0^*}{4\fp^3}\ \TO \ i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\frac{\ga^3 g_2}{12\fp^3}\ \tten \ i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\Big\}\nonumber\\
&\times\int_\bmq\frac{\omp^2+\omp\omm+\omm^2}{\omp^3\omm^3(\omp+\omm)}
(k^2-q^2)\ ,\label{eq:panelbox}\end{aligned}$$ where all the isovector terms cancel out. After dimensional regularization, the finite part reads, $$\begin{aligned}
\overline{V}^{(1)}({\rm box})&=\frac{\ga^3 g_0^*}{\fp\Lx^2}\ \TO \ i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}) [H(k)-3\, L(k)]\nonumber\\
&-\frac{\ga^3 g_2}{3\fp\Lx^2}\ \tten \ i\bmk\cdot
({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2}) [H(k)-3\, L(k)]\ ,\label{eq:box}\nonumber\\\end{aligned}$$ where $$H(k)= {4\, m^2_\pi\over s^2} L(k)\ ,\label{eq:H}$$ while the singular part is given by, $$\begin{aligned}
V^{(1)}_{\infty}&({\rm box})=-\frac{\ga^3 g_0^*}{\fp\Lx^2}\TO\
i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\Big(\frac{3}{2}\deps-1\Big)\nonumber\\
&+\frac{\ga^3 g_2}{3\fp\Lx^2}\tten\
i\bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\Big(\frac{3}{2}\deps-1\Big)\ .\end{aligned}$$ The latter is absorbed in the ${V_{TRV}}^{({\rm CT})}$ term proportional to $C_2$ for the isoscalar part and to $C_5$ for the isotensor part.
Now we consider the contributions that come from the panels (i) and (l) of Fig. \[fig:diagNN\]. At NLO the contributions of the panel (l) cancel out due to the isospin structure of the vertices. The contribution of diagrams (i) result, $$\begin{aligned}
V^{(0)}({\rm i})&=-\frac{5\ga^3\Delta_3M}{32\fp^3}\frac{1}{\omk^2}
\Big[\ttp\ i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2)\nonumber\\
&+\ttm \ i \bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\Big]
\int_\bmq \frac{k^2-q^2}{\omp^2\omm^2}\ .\label{eq:3piapp}\end{aligned}$$ Using dimensional regularization we obtain, $$\begin{aligned}
\overline{V}^{(0)}&({\rm i})=-\frac{5\ga^3\Delta_3 M}{4\fp\Lx^2}
\frac{\pi}{\omega_k^2}
\big(\ttm\ i \bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\ttp\ i \bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\big)
\Big[\Big(1-\frac{2\mp^2}{s^2}\Big)s^2A(k)+\mp\Big]\ ,\nonumber\\\end{aligned}$$ where, $$\begin{aligned}
A(k)=\frac{1}{2k}\arctan\Big({\frac{k}{2\mp}}\Big)\ .\label{eq:A}\end{aligned}$$ To be noticed that the use of dimensional regularization does not give the divergent part of the integral in Eq. (\[eq:3piapp\]). This is due to the fact that the dimensional regularization cannot deal with linear divergences. To explicit the linear divergence we use a simple regularization of Eq. (\[eq:3piapp\]), namely we integrate over $q$ up to a (large) value $\Lambda_R$. The result for the non divergent part is equal to the one reported in Eq. (\[eq:3piapp\]) while the divergent part reads, $$\begin{aligned}
V^{(0)}_{\infty}&({\rm i})=-\frac{5\ga^3\Delta_3 M}{4\fp\Lx^2}
\frac{1}{\omega_k^2}
\big(\ttm\ i \bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\nonumber\\
&+\ttp\ i \bmk\cdot({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\big)
\Big[\Lambda_R+{\cal O}\big(\frac{k^2}{\Lambda_R}\big)\Big]\ ,\end{aligned}$$ where spurious contributions of order $Q^2/\Lambda_R$ or more appear but they can be neglected for $\Lambda_R\rightarrow\infty$. The divergent part can be reabsorbed in the ${V_{TRV}}^{({\rm CT})}$ term proportional to $C_4$.
At N2LO the contribution of panel (i) comes both from the second order in the pion propagator (PP) and in the pion-nucleon vertex (PNV). For the former we obtain, $$\begin{aligned}
&V^{(1)}({\rm i-PP})=-\frac{5\ga^3\Delta_3}{128\fp^3}\frac{1}{\omk^2}
\Big[\ttp\
i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2)\nonumber\\
&\quad+\ttm \ i \bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\Big]
\int_\bmq \frac{\omp^2+\omp\omm+\omm^2}{\omp^3\omm^3(\omp+\omm)}
\big(k^2-q^2\big)^2\nonumber\\
&\quad+\frac{5\ga^3\Delta_3}{8\fp^3}\frac{1}{\omk^2}
\int_\bmq \frac{\omp^2+\omp\omm+\omm^2}{\omp^3\omm^3(\omp+\omm)}\nonumber\\
&\qquad\qquad\times
\big[\bmk\cdot{\boldsymbol{\sigma}_1}(\bmq\times\bmk)\cdot{\boldsymbol{\sigma}_2}(\bmq\cdot\bmK)\tau_{1z}
\nonumber\\
&\qquad\qquad
-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmq\times\bmk)\cdot{\boldsymbol{\sigma}_1}(\bmq\cdot\bmK)\tau_{2z}\big]\end{aligned}$$ while for the latter, $$\begin{aligned}
&V^{(1)}({\rm i-PNV})=\frac{5\ga^3\Delta_3}{64\fp^3}\frac{1}{\omk^2}
\Big[\ttp\
i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2)\nonumber \\
&\qquad+\ttm \ i \bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\Big]
\int_\bmq \frac{k^2+q^2}{\omp\omm(\omp+\omm)}\nonumber\\
&\qquad+\frac{5\ga^3\Delta_3}{16\fp^3}\frac{1}{\omk^2}\nonumber\\
&\qquad\times\Big[\big(\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}+\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\big)\ttm\nonumber\\
&\qquad+\big(\bmk\cdot{\boldsymbol{\sigma}_1}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}-\bmk\cdot{\boldsymbol{\sigma}_2}(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\big)\ttp\Big]\nonumber\\
&\qquad\times\int_\bmq \frac{1}{\omp\omm(\omp+\omm)}\ .\end{aligned}$$ Also in this case we use the dimensional regularization that permits us to write the finite contribution as, $$\begin{aligned}
\overline{V}&^{(1)}({\rm i})=-\frac{5\ga^3\Delta_3}{16\fp\Lx^2}
(\ttm\ i\bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2) \nonumber\\
&+\ttp\ i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2))\nonumber\\
&\times\Big(\frac{25}{6}L(k)-\frac{7}{2}\frac{\mp^2L(k)}{\omega_k^2}
+2\frac{\mp^2H(k)}{\omega_k^2}\Big)
-\frac{25\ga^3\Delta_3}{12\fp\Lx^2}\frac{1}
{\omega_k^2}\nonumber\\
&\times\Big[\ttm \ \big((\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}\bmk\cdot{\boldsymbol{\sigma}_1}+(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\bmk\cdot{\boldsymbol{\sigma}_2}\big)\nonumber\\
&+\ttp \ \big((\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_2}\bmk\cdot{\boldsymbol{\sigma}_1}-(\bmk\times\bmK)\cdot{\boldsymbol{\sigma}_1}\bmk\cdot{\boldsymbol{\sigma}_2}\big)\Big]\ ,\end{aligned}$$ while the divergent part reads, $$\begin{aligned}
V^{(1)}_\infty&({\rm i})=
\frac{5\ga^3\Delta_3}{16\fp\Lx^2}
(\ttm\ i\bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\nonumber\\
&+\ttp\ i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2))\nonumber\\
&\times\Big[\frac{\mp^2}
{\omega_k^2}\Big(\frac{151}{12}\deps-\frac{305}{18}\Big)
-\frac{25}{12}\deps-\frac{2}{9}\Big]\ .
\label{eq:divpanli}\end{aligned}$$ The divergences present in Eq. (\[eq:divpanli\]) are reabsorbed in the ${V_{TRV}}^{({\rm OPE})}$ term proportional to $g_1$ for the part which multiply $\mp^2/\omega_k^2$ and in the ${V_{TRV}}^{({\rm CT})}$ term proportional to $C_4$ for the rest. All the divergences related to the term where $\bmK$ is present cancel out.
The N2LO contribution of panel (l) in Fig. \[fig:diagNN\] is proportional to the LECs $c_1$, $c_2$ and $c_3$ of the PC sector and it reads, $$\begin{aligned}
V^{(1)}&({\rm l})=-\frac{5\ga\Delta_3M}{2\fp^3}\frac{1}{\omk^2}
\Big[\ttp \ i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2)\nonumber\\
&+\ttm \ i \bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\Big]
\Big(c_1 m_\pi^2\int_\bmq\frac{1}{\omp\omm(\omp+\omm)}\nonumber\\
&+\frac{c_2+c_3}{8}\int_\bmq\frac{1}{\omp+\omm}
-\frac{c_3}{8}\int_\bmq\frac{k^2-q^2}{\omp\omm(\omp+\omm)}\Big)\ .\end{aligned}$$ Using dimensional regularization we obtain for the finite part, $$\begin{aligned}
\overline{V}^{(1)}&({\rm l})=\frac{5\ga\Delta_3M}{2\fp\Lx^2}
(\ttm\ i\bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\nonumber\\
&+\ttp\ i\bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2))\nonumber\\
&\times\Big[4c_1\ \frac{\mp^2L(k)}{\omega_k^2}
-\frac{c_2}{3}\
\Big(2L(k)+6\frac{\mp^2}{\omega_k^2}L(k)\Big)\nonumber\\
&-\frac{c_3}{2}\
\Big(3L(k)+5\frac{\mp^2}{\omega_k^2}L(k)\Big)\Big]\ ,\end{aligned}$$ while the divergent part is given by, $$\begin{aligned}
V^{(1)}_\infty&(l)=\frac{5\ga\Delta_3M}{2\fp\Lx^2}
(\ttm\ i\bmk\cdot({{\bm \sigma}}_1+{{\bm \sigma}}_2)\nonumber\\
&+\ttp\ i\bmk\cdot({{\bm \sigma}}_1-{{\bm \sigma}}_2))
\Big[c_1\frac{1}{\omega_k^2}(-2\deps+4)\nonumber\\
&-\frac{c_2}{3}\Big(\frac{\mp^2}{\omega_k^2}
\big(5\deps-\frac{19}{3}\big)
+\big(\deps-\frac{5}{3}\big)\Big)\nonumber\\
&-\frac{c_3}{2}\Big(\frac{\mp^2}
{\omega_k^2}\big(\frac{11}{2}\deps-\frac{25}{3}\big)
+\big(\frac{3}{2}\deps-\frac{2}{3}\big)\Big)\Big]\ .\end{aligned}$$ As for the panel (i) the divergences which multiply $\mp^2/\omega_k^2$ are reabsorbed in the $V^{({\rm OPE})}$ term proportional to $g_1$ while all the others in the $V^{({\rm CT})}$ term proportional to $C_4$. At N2LO panel (l) can receive contribution also from the second order in the $\pi NN$ and $\pi\pi NN$ vertices but due to the isospin structure these contributions vanish.
As regarding the $NNN$ TRV potential all the contributions come from diagrams (m) of Fig. \[fig:diagNN\]. The expression we obtain at NLO is given in Eq. (\[eq:NNNpot\]). The N2LO component would come from NLO PC $\pi NN$ vertex or in the pion propagators. In both cases the different time-order diagrams cancel out each-other completely.
The potential in configuration space {#app:rpot}
====================================
The $NN$ potential {#app:rpotNN}
------------------
In this subsection we present the $NN$ potential part in the configuration space which follows directly from Eq. (\[eq:vrsp\]) and it reads, $$\begin{aligned}
V_{TRV}(\bmr,\bmp)&=&
V^{({\rm OPE})}(\bmr)+V^{({\rm TPE})}(\bmr)+V^{(3\pi,0)}(\bmr)
\nonumber\\
&+&V^{(3\pi,1)}(\bmr)+
V^{({\rm CT})}(\bmr)+V^{({\rm RC})}(\bmr,\bmp)
\nonumber\\
&+&V^{(3\pi,{\rm RC})}(\bmr,\bmp)\ ,
\label{eq:trvnnr}\end{aligned}$$ where $\bmp=-i{{\bm \nabla}}$ is the relative momentum operator. It is convenient to define the operators, $$\begin{aligned}
\ssp&=&({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\cdot\bmvr\ ,\\
\ssm&=&({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\cdot\bmvr\ ,\\
\slp&=&-i({\boldsymbol{\sigma}_1}+{\boldsymbol{\sigma}_2})\cdot(\bmvr\times\hat{\bm L})\ ,\\
\slm&=&-i({\boldsymbol{\sigma}_1}-{\boldsymbol{\sigma}_2})\cdot(\bmvr\times\hat{\bm L})\ ,\\
\srx&=&({\boldsymbol{\sigma}_1}\times{\boldsymbol{\sigma}_2})\cdot\bmvr\ ,\\
\srlp&=&{\boldsymbol{\sigma}_1}\cdot\bmvr\, {\boldsymbol{\sigma}_2}\cdot\hat{\bm L}+
{\boldsymbol{\sigma}_2}\cdot\bmvr\, {\boldsymbol{\sigma}_1}\cdot\vec{\bm L}\ ,\\
\srlm&=&{\boldsymbol{\sigma}_1}\cdot\bmvr\, {\boldsymbol{\sigma}_2}\cdot\hat{\bm L}-
{\boldsymbol{\sigma}_2}\cdot\bmvr\, {\boldsymbol{\sigma}_1}\cdot\vec{\bm L}\ ,\end{aligned}$$ where $\hat\bmL=\hat\bmr\times\bmp$ is the “reduced” orbital angular momentum operator. In terms of these, $V_{TRV}(\bmr,\bmp)$ can be written as =1.0pt $$\begin{aligned}
V^{({\rm OPE})}(\bmr) &=\frac{\ga g_0^*\mp}{2\fp}\TO\ \ssm \ g'(r)\nonumber\\
&+
\frac{g_Ag_1\mp}{4f_\pi}\left(\ttp\ \ssm+\ttm\ \ssp\right)\ g'(r)\nonumber\\
&+\frac{\ga g_2\mp}{6\fp}\tten\ \ssm \ g'(r)\ ,\\
V^{({\rm TPE})}(\bmr) &=
\frac{\ga g_0^*\mp^3}{\fp\Lx^2}\TO\ \ssm\ L'(r)\nonumber\\
&+\frac{\ga^3 g_0^*\mp^3}{\fp\Lx^2}\TO\ \ssm\ (H'(r)-3L'(r))\nonumber\\
&-\frac{\ga g_2\mp^3}{3\fp\Lx^2}\tten\ \ssm\ L'(r)\nonumber\\
&-\frac{\ga^3 g_2\mp^3}{3\fp\Lx^2}\tten\ \ssm\ (H'(r)-3L'(r))\ ,\\
V^{(3\pi,0)}(\bmr)&=
-\frac{5\ga^3\Delta_3 M \mp^2}{4\fp\Lx^2}\pi\nonumber\\
&\times(\ttp\ \ssm+\ttm\ \ssp)(A'(r)+g'(r))\\
V^{(3\pi,1)}(\bmr)&=\frac{5\ga\Delta_3M\mp^3}{2\fp\Lx^2}(\ttp\ \ssm+\ttm\ \ssp)
\nonumber\\
&\Big[4c_1\ L'_\omega(r)-\frac{c_2}{3}(2L'(r)+6L'_\omega(r))\nonumber\\
&\qquad-\frac{c_3}{2}(3L'(r)+5L'_\omega(r))\Big]\ ,\\
V^{({\rm CT})}(\bmr) &=
{m_\pi^2 \over \Lambda_\chi^2 f_\pi} \Bigl[
C_1\ \ssm\ Z'(r)+C_2\ \TO\ \ssm\ Z'(r)\nonumber\\
&+{C_3 \over 2}\ (\ttp\ \ssm-\ttm\ \ssp)\ Z'(r)\nonumber\\
&+{C_4 \over 2}\ (\ttp\ \ssm+\ttm\ \ssp)\ Z'(r)\nonumber\\
&+C_5\ \tten\ \ssm\ Z'(r) \Bigr]\ ,\end{aligned}$$ $$\begin{aligned}
V^{({\rm RC})}&(\bmr,\bmp)=\nonumber\\
&+\frac{\ga g_0^*\mp^3}{2\fp M^2 }\ \TO\ \Big[\frac{\ssm}{4}\Big(g'''(r)
+2\frac{g''(r)}{r}-2\frac{g'(r)}{r^2}\Big)\nonumber\\
&+\slm\frac{g'(r)}{r^2}-\frac{g'(r)}{r^2}\ssm L^2
-\frac{1}{4}\Big(\frac{g''(r)}{r}
+\frac{g'(r)}{r^2}\Big)\srlm\nonumber\\
&+\Big(\ssm\Big(g''(r)+2\frac{g'(r)}{r}\Big)
-i\frac{\srx}{2}\frac{g'(r)}{r}\Big)
\frac{d}{dr}\nonumber\\
&+g'(r)\ssm\frac{d^2}{dr^2}\Big]\frac{1}{\mp^2}\nonumber\\
&+\frac{\ga g_1\mp^3}{4\fp M^2}\
\ttp\ \Big[\frac{\ssm}{4}\Big(g'''(r)
+2\frac{g''(r)}{r}-2\frac{g'(r)}{r^2}\Big)\nonumber\\
&+\slm\frac{g'(r)}{r^2}-\frac{g'(r)}{r^2}\ssm L^2
-\frac{1}{4}\Big(\frac{g''(r)}{r}
+\frac{g'(r)}{r^2}\Big)\srlm\nonumber\\
&+\Big(\ssm\Big(g''(r)+2\frac{g'(r)}{r}\Big)
-i\frac{\srx}{2}\frac{g'(r)}{r}\Big)
\frac{d}{dr}\nonumber\\
&+g'(r)\ssm\frac{d^2}{dr^2}\Big]\frac{1}{\mp^2}\nonumber\\
&+\frac{\ga g_1\mp^3}{4\fp M^2}\
\ttm\ \Big[\frac{\ssp}{4}\Big(g'''(r)
+2\frac{g''(r)}{r}-2\frac{g'(r)}{r^2}\Big)+\nonumber\\
&+\slp\frac{g'(r)}{r^2}-\frac{g'(r)}{r^2}\ssp L^2
-\frac{1}{4}\Big(\frac{g''(r)}{r}
-\frac{g'(r)}{r^2}\Big)\srlp\nonumber\\
&+\ssp\Big(g''(r)+2\frac{g'(r)}{r}\Big)
\frac{d}{dr}+g'(r)\ssp\frac{d^2}{dr^2}\Big]\frac{1}{\mp^2}\nonumber\\
&+\frac{\ga g_2\mp^3}{6\fp M^2}\ \tten\ \Big[\frac{\ssm}{4}\Big(g'''(r)
+2\frac{g''(r)}{r}-2\frac{g'(r)}{r^2}\Big)\nonumber\\
&+\slm\frac{g'(r)}{r^2}-\frac{g'(r)}{r^2}\ssm L^2
-\frac{1}{4}\Big(\frac{g''(r)}{r}
+\frac{g'(r)}{r^2}\Big)\srlm\nonumber\\
&+\Big(\ssm\Big(g''(r)+2\frac{g'(r)}{r}\Big)
-i\frac{\srx}{2}\frac{g'(r)}{r}\Big)
\frac{d}{dr}\nonumber\\
&+g'(r)\ssm\frac{d^2}{dr^2}\Big]\frac{1}{\mp^2}\ ,\label{eq:rcoper}\end{aligned}$$ $$\begin{aligned}
&V^{(3\pi-{\rm RC})}(\bmr,\bmp)=
-\frac{5\ga^3\Delta_3\mp^3}{16\fp\Lx^2}(\ttp\ \ssm+\ttm\ \ssp)\nonumber\\
&\qquad \Big(\frac{25}{6}L'(r)-\frac{7}{2}L'_\omega(r)+2H'_\omega(r)\Big)
\nonumber\\
&\qquad+\frac{25\ga^3\Delta_3\mp^3}{12\fp\Lx^2}\Big[\ttp\ \Big(\Big(\frac{g''(r)}{r}
+\frac{g'(r)}{r^2}\Big)\srlm\nonumber\\
&\qquad+2i\srx\frac{g'(r)}{r}\frac{d}{dr}\Big)
+\ttm\ \Big(\frac{g''(r)}{r}
-\frac{g'(r)}{r^2}\Big)\srlp\Big]\frac{1}{\mp^2}\ ,\label{eq:rc3pr}\end{aligned}$$ with $$\begin{aligned}
g(r) &=& \int {{\rm d}^3k\over (2\pi)^3}\; {C_\Lambda(k)\over \mp}
\frac{1}{\omega_k^2}\,
e^{i\bmk\cdot\bmr} \ ,\label{eq:g}\\
L(r) &=& \int {{\rm d}^3k\over (2\pi)^3}\; {C_\Lambda(k)\over m_\pi^3} L(k)\,
e^{i\bmk\cdot\bmr} \ ,\label{eq:lr}\\
H(r) &=& \int {{\rm d}^3k\over (2\pi)^3}\; {C_\Lambda(k)\over m_\pi^3} H(k)\,
e^{i\bmk\cdot\bmr} \ ,\label{eq:hr}\\
A(r) &=& \int {{\rm d}^3k\over (2\pi)^3}\; {C_\Lambda(k)\over m_\pi^2}
{s^2\ A(k)\over \omega_k^2}\,
e^{i\bmk\cdot\bmr} \ ,\label{eq:ar}\\
L_\omega(r) &=&
\int {{\rm d}^3k\over (2\pi)^3}\; {C_\Lambda(k)\over m_\pi}
{L(k) \over \omega_k^2}\,
e^{i\bmk\cdot\bmr} \ ,\label{eq:lrw}\\
H_\omega(r) &=&
\int {{\rm d}^3k\over (2\pi)^3}\; {C_\Lambda(k)\over m_\pi}
{H(k) \over \omega_k^2}\,
e^{i\bmk\cdot\bmr} \ ,\label{eq:hrw}\\
Z(r) &=& \int {{\rm d}^3k\over (2\pi)^3}\; {C_\Lambda(k)\over m_\pi^2}\,
e^{i\bmk\cdot\bmr} \ .\label{eq:z}\end{aligned}$$ The functions $g(r)$, $L(r)$, $H(r)$, $A(r)$, $L_\omega(r)$, $H_\omega(r)$ and $Z(r)$ are calculated numerically by standard quadrature techniques.
The $NNN$ potential {#sec:NNNr}
-------------------
In this section we present the explicit derivation of the $NNN$ potential in configuration space. Writing explicitly the integral in Eq. (\[eq:vrnnn\]), neglecting the deltas for simplicity, we get,
$$\begin{aligned}
V&(\bmx_1,\bmx_2)=\nonumber\\
&\quad-\frac{\Delta_3 g_A^3 M}{4\fp^3}\ T_3\
\frac{1}{2}\int {d^3q\over (2\pi)^3} {d^3Q\over (2\pi)^3}
e^{-i(\bmq/2)\cdot\bmx_2}\,e^{-i\bmQ\cdot\bmx_1}\,\nonumber\\
&\quad\times\frac{i(\bmQ+\bmq)\cdot{\boldsymbol{\sigma}_1}\, i(\bmQ-\bmq)\cdot{\boldsymbol{\sigma}_2}\, i\bmQ\cdot{\boldsymbol{\sigma}_3}}
{\omega_+^2\omega_-^2\omega_Q^2}\ ,\label{eq:qQxx}\end{aligned}$$
where $T_3={\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_2}\ \tau_{3z}+
{\boldsymbol{\tau}_1}\cdot{\boldsymbol{\tau}_3}\ \tau_{2z}+{\boldsymbol{\tau}_2}\cdot{\boldsymbol{\tau}_3}\ \tau_{1z}$. The momenta $\bmq$ and $\bmQ$ in Eq. (\[eq:qQxx\]) can be rewritten applying the gradient $i{{\bm \nabla}}$ to the exponential functions and it reads, $$\begin{aligned}
V&(\bmx_1,\bmx_2)=\nonumber\\
&\frac{\Delta_3 g_A^3 M}{4\fp^3}\ T_3\
\frac{1}{2}\big[ ({{\bm \nabla}}_{\bmx_1}+2{{\bm \nabla}}_{\bmx_2})\cdot{\boldsymbol{\sigma}_1}\
({{\bm \nabla}}_{\bmx_1}-2{{\bm \nabla}}_{\bmx_2})\cdot{\boldsymbol{\sigma}_2}\nonumber\\
&\times{{\bm \nabla}}_{\bmx_1}\cdot{\boldsymbol{\sigma}_3}\big]
\int{d^3Q\over (2\pi)^3}\ \frac{e^{-i\bmQ\cdot\bmx_1}}{\omega_Q^2}
\ \int {d^3q\over (2\pi)^3}\ \frac{e^{-i(\bmq/2)\cdot\bmx_2}}
{\omega_+^2\omega_-^2}\ .\label{eq:qQxxd}
\end{aligned}$$ The integral in $\bmq$ in Eq. (\[eq:qQxxd\]) can be solved using the Feynman tricks. The final result is $$\begin{aligned}
V(\bmx_1,&\bmx_2)=
\frac{\Delta_3 g_A^3 M}{4\fp^3}\ T_3\
\frac{1}{32\pi}\big[ ({{\bm \nabla}}_{\bmx_1}+2{{\bm \nabla}}_{\bmx_2})\cdot{\boldsymbol{\sigma}_1}\nonumber\\
&\times({{\bm \nabla}}_{\bmx_1}-2{{\bm \nabla}}_{\bmx_2})\cdot{\boldsymbol{\sigma}_2}\
{{\bm \nabla}}_{\bmx_1}\cdot{\boldsymbol{\sigma}_3}\big]
\int{d^3Q\over (2\pi)^3}\ \frac{e^{-i\bmQ\cdot\bmx_1}}{\omega_Q^2}\nonumber\\
&\times \int_{-1}^1 dx\ e^{i(x/2)\bmQ\bmx_2} \frac{e^{-(L/2)x_2}}
{L}\ ,\label{eq:qxxxd}
\end{aligned}$$ where $L=\sqrt{Q^2(1-x^2)+4\mp^2}$. The derivative are then evaluated obtaining five integral operators $I_i$, $$\begin{aligned}
V(\bmx_1,\bmx_2)&=&\frac{\Delta_3 g_A^3 M}{4\fp^3}\ T_3\ \sum_{i=1,5} I_i
\end{aligned}$$ where, $$\begin{aligned}
I_1&=-\frac{i}{16\pi}\int{d^3Q\over (2\pi)^3}\
\frac{e^{-i\bmQ\cdot\bmx_1}}{\omega_Q^2}
\int_{-1}^1 dx\, \nonumber\\
&\quad\times e^{i(x/2)\bmQ\cdot\bmx_2}\frac{e^{-(L/2)x_2}}{x_2}
\,({\boldsymbol{\sigma}_1}\cdot{\boldsymbol{\sigma}_2})\bmQ\cdot{\boldsymbol{\sigma}_3}\ ,\\
I_2&=+\frac{i}{32\pi}\int{d^3Q\over (2\pi)^3}\
\frac{e^{-i\bmQ\cdot\bmx_1}}{\omega_Q^2}
\int_{-1}^1 dx\, (1-x^2) \nonumber\\
&\quad\times e^{i(x/2)\bmQ\cdot\bmx_2}\frac{e^{-(L/2)x_2}}{L}
\,\bmQ\cdot{\boldsymbol{\sigma}_1}\,\bmQ\cdot{\boldsymbol{\sigma}_2}\,\bmQ\cdot{\boldsymbol{\sigma}_3}\ ,\\
I_3&=-\frac{1}{32\pi}\int{d^3Q\over (2\pi)^3}\
\frac{e^{-i\bmQ\cdot\bmx_1}}{\omega_Q^2}
\int_{-1}^1 dx\, (1-x)\nonumber\\
&\quad\times e^{i(x/2)\bmQ\cdot\bmx_2}e^{-(L/2)x_2}
\bmQ\cdot{\boldsymbol{\sigma}_1}\,\bmx_2\cdot{\boldsymbol{\sigma}_2}\,\bmQ\cdot{\boldsymbol{\sigma}_3}\ ,\\
I_4&=+\frac{1}{32\pi}\int{d^3Q\over (2\pi)^3}\
\frac{e^{-i\bmQ\cdot\bmx_1}}{\omega_Q^2}
\int_{-1}^1 dx\, (1+x)\nonumber\\
&\quad\times e^{i(x/2)\bmQ\cdot\bmx_2}e^{-(L/2)x_2}
\,\bmx_2\cdot{\boldsymbol{\sigma}_1}\,\bmQ\cdot{\boldsymbol{\sigma}_2}\,\bmQ\cdot{\boldsymbol{\sigma}_3}\ ,\\
I_5&=+\frac{i}{32\pi}\int{d^3Q\over (2\pi)^3}\
\frac{e^{-i\bmQ\cdot\bmx_1}}{\omega_Q^2}
\int_{-1}^1 dx\, \Big(L+\frac{2}{x_2}\Big)\nonumber\\
&\quad\times e^{i(x/2)\bmQ\cdot\bmx_2} \frac{e^{-(L/2)x_2}}{L}
\,\bmx_2\cdot{\boldsymbol{\sigma}_1}\,\bmx_2\cdot{\boldsymbol{\sigma}_2}\,\bmQ\cdot{\boldsymbol{\sigma}_3}\ ,
\end{aligned}$$ To be noticed that integral $I_4=I_3$ exchanging particle 1 and 2.
In order to divide the angular integration in $\hat{\bmQ}$ from the radial integration in $Q$, the exponential functions are expanded in plane waves. Therefore the integrals $I_i$ result, $$I_i=\sum_{l,l'}O_i^{ll'}J_i^{ll'}\ ,$$ where $J_i$ is the integral over $Q$ while in $O_i$ the integral is over $\hat{Q}$. The expressions of $J_i$ are explicitly given below: $$\begin{aligned}
J_1^{ll'}&=\frac{1}{8\pi}\int_{-1}^1 dx\ \int_0^\infty dQ \frac{Q^2}{\omega_Q^2}
\ C_\Lambda(Q)e^{-(L/2)x_2}\nonumber\\
&\qquad\qquad\times j_l\big({xQx_2\over 2}\big)j_{l'}(Qx_1)\frac{Q}{x_2}
\ ,\\
J_2^{ll'}&=\frac{1}{16\pi}\int_{-1}^1 dx\ \int_0^\infty dQ \frac{Q^2}{\omega_Q^2}
\ C_\Lambda(Q)e^{-(L/2)x_2}\nonumber\\
&\qquad\qquad\times j_l\big({xQx_2\over 2}\big)j_{l'}(Qx_1)
\frac{Q^3}{L}(1-x^2)\ ,\\
J_3^{ll'}&=\frac{1}{16\pi}\int_{-1}^1 dx\ \int_0^\infty dQ \frac{Q^2}{\omega_Q^2}
\ C_\Lambda(Q)e^{-(L/2)x_2}\nonumber\\
&\qquad\qquad\times j_l\big({xQx_2\over 2}\big)j_{l'}(Qx_1)
Q^2(1-x)\ ,\\
J_4^{ll'}&=\frac{1}{16\pi}\int_{-1}^1 dx\ \int_0^\infty dQ \frac{Q^2}{\omega_Q^2}
\ C_\Lambda(Q)e^{-(L/2)x_2}\nonumber\\
&\qquad\qquad\times j_l\big({xQx_2\over 2}\big)j_{l'}(Qx_1)
Q^2(1+x)\ ,\\
J_5^{ll'}&=\frac{1}{16\pi}\int_{-1}^1 dx\ \int_0^\infty dQ \frac{Q^2}{\omega_Q^2}
\ C_\Lambda(Q)e^{-(L/2)x_2}\nonumber\\
&\qquad\qquad\times j_l\big({xQx_2\over 2}\big)j_{l'}(Qx_1)
Q\Big(L-\frac{2}{x_2}\Big)\ ,
\end{aligned}$$ where $j_l(x)$ are spherical Bessel functions and we include a regularization function $C_\Lambda(Q)$ which is defined in Eq. (\[eq:cutoffnnn\]). The functions $J_i$ depend only on the modules of $\bmx_1$ and $\bmx_2$ and they are evaluated numerically by standard quadrature techniques.
For the $O_i$ integrals we get, $$\begin{aligned}
O_1^{ll'}&=(-i)i^{l'}(-i)^{l}\hat{l'}\hat{l}\int d\hat{Q}
\Big[Y_{l'}(\hat{Q})Y_{l'}(\hat{x}_1)\Big]_0\nonumber\\
&\times
\Big[Y_{l }(\hat{Q})Y_{l}(\hat{x}_2)\Big]_0
{\boldsymbol{\sigma}_1}\cdot{\boldsymbol{\sigma}_2}\, \hat{Q}\cdot{\boldsymbol{\sigma}_3}\ ,\\
O_2^{ll'}&=(+i)i^{l'}(-i)^{l}\hat{l'}\hat{l}\int d\hat{Q}
\Big[Y_{l'}(\hat{Q})Y_{l'}(\hat{x}_1)\Big]_0\nonumber\\
&\times
\Big[Y_{l }(\hat{Q})Y_{l}(\hat{x}_2)\Big]_0
\hat{Q}\cdot{\boldsymbol{\sigma}_1}\,\hat{Q}\cdot{\boldsymbol{\sigma}_2}\,\hat{Q}\cdot{\boldsymbol{\sigma}_3}\ ,\\
O_3^{ll'}&=-i^{l'}(-i)^{l}\hat{l'}\hat{l}\int d\hat{Q}
\Big[Y_{l'}(\hat{Q})Y_{l'}(\hat{x}_1)\Big]_0\nonumber\\
&\times
\Big[Y_{l }(\hat{Q})Y_{l}(\hat{x}_2)\Big]_0
\hat{Q}\cdot{\boldsymbol{\sigma}_1}\,\hat{x}_2\cdot{\boldsymbol{\sigma}_2}\,\hat{Q}\cdot{\boldsymbol{\sigma}_3}\ ,\\
O_4^{ll'}&=+i^{l'}(-i)^{l}\hat{l'}\hat{l}\int d\hat{Q}
\Big[Y_{l'}(\hat{Q})Y_{l'}(\hat{x}_1)\Big]_0\nonumber\\
&\times
\Big[Y_{l }(\hat{Q})Y_{l}(\hat{x}_2)\Big]_0
\hat{x}_2\cdot{\boldsymbol{\sigma}_1}\,\hat{Q}\cdot{\boldsymbol{\sigma}_2}\,\hat{Q}\cdot{\boldsymbol{\sigma}_3}\ ,\\
O_5^{ll'}&=(+i)i^{l'}(-i)^{l}\hat{l'}\hat{l}\int d\hat{Q}
\Big[Y_{l'}(\hat{Q})Y_{l'}(\hat{x}_1)\Big]_0\nonumber\\
&\times
\Big[Y_{l }(\hat{Q})Y_{l}(\hat{x}_2)\Big]_0
\hat{x}_2\cdot{\boldsymbol{\sigma}_1}\,\hat{x}_2\cdot{\boldsymbol{\sigma}_2}\,\bmQ\cdot{\boldsymbol{\sigma}_3}\ ,
\end{aligned}$$ which depend only to the angular part of the spatial coordinates $\bmx_1$ and $\bmx_2$. The matrix element of the $O_i$ operators between HH functions can easily expressed in terms of products of 9-j, 6-j and 3-j Wigner symbols.
Details of the calculation and convergence of the $a_\Delta(3N)$ coefficient {#app:NNNconv}
============================================================================
In order to compute the even-parity $|\psi^A_{+}{\rangle}$ and odd-parity $|\psi^A_{-}{\rangle}$ component of the wave function we need to solve the following eigenvalue problem, $$\begin{aligned}
\label{eq:egp}
\begin{pmatrix}
V_{\rm PC}^{++} & V_{\rm TRV}^{+-}\\
V_{\rm TRV}^{-+} & V_{\rm PC}^{--}
\end{pmatrix}
\begin{pmatrix}
|\psi^A_{+}{\rangle}\\
|\psi^A_{-}{\rangle}\end{pmatrix}
=E \begin{pmatrix}
|\psi^A_{+}{\rangle}\\
|\psi^A_{-}{\rangle}\end{pmatrix}\,.
\end{aligned}$$ Using the fact that $||V_{\rm TRV}||<<||V_{\rm PC}||$ we can rewrite the eigenvalue problem of Eq. (\[eq:egp\]) as, $$\begin{aligned}
\label{eq:egp2}
\begin{cases}
V_{\rm PC}^{++}|\psi^A_{+}{\rangle}=E|\psi^A_{+}{\rangle}\\
|\psi^A_{-}{\rangle}=-(V_{\rm PC}^{--}-E)^{-1} V_{\rm TRV}^{-+}
|\psi^A_{+}{\rangle}\end{cases}\,,
\end{aligned}$$ where the first equation is the standard eigenvalue problem, while with the second we can compute the odd component of the wave function. This approach from the numerical point of view results more stable than solving directly Eq. (\[eq:egp\]) and permits to study the odd component of the wave function without solving every time the eigenvalue problem.
Let us study now the convergence pattern of the $a_\Delta(3N)$ coefficient, defined in Eq. (\[eq:adelta3\]), in term of the grandangular momentum $K$ of the HH basis (for more details, see Ref. [@AK08; @LE09]). Increasing $K$ is equivalent of enlarging the expansion basis. To be definite, in this appendix, we have considered the $\tri$ case and used the N4LO/N2LO-500 PC interaction.
Let us denote with $K^+$ ($K^-$) the maximum value of the grandangular momentum of the HH functions used to describe the even (odd) part of the wave function $|\psi^A_{+}{\rangle}$ ($|\psi^A_{-}{\rangle}$). We have computed the even part of the wave function up to complete convergence ($K^+=50$) obtaining a binding energy $B(\tri)=8.476$ MeV using the first formula in Eq. (\[eq:egp2\]). For solving the second equation of Eq. (\[eq:egp2\]) we have performed different calculations varying $K^-$ and we have reported the corresponding results for $a_\Delta(3N)$ in Table \[tab:3Nconv\]. As can be seen by inspecting the table, the pattern of convergence is very smooth. A safe convergence at the third digit is reached for $K^-\sim21$, which was the value also selected for performing the final calculations. On the other hand, the value of $a_\Delta(3N)$ is not much sensitive to $K^+$, in particular when $K^+>20$. Therefore, the value of the coefficient $a_\Delta(3N)$ appears to be well under control in our calculation.
$K^-$ $a_\Delta(3N)$
------- ----------------
5 -0.1728
9 -0.1842
13 -0.1879
17 -0.1892
21 -0.1897
: \[tab:3Nconv\] Convergence pattern of the $\tri$ $a_\Delta(3N)$ coefficient as function of $K^-$, the maximum grandangular momentum of the HH functions used for constructing the odd-parity component of the $\tri$ wave function. For the even-parity we have used $K^+=50$, a value sufficient to reach full convergence of the even-parity component of the wave function. The reported calculations are performed using N4LO/N2LO-500 PC interaction.
[100]{}
A.D. Sakarhov, Pisma Zh. Eksp. Teor. Fiz. [**5**]{}, 32 (1967).
A.G. Cohen, D.B. Kaplan, and A.E. Nelson, Ann. Rev. Nucl. Part. Sci. [**43**]{}, 27 (1993).
G. ’t Hooft, Phys. Rev. Lett. [**37**]{}, 8 (1976).
M. Pospelov, A. Ritz, Annals Phys. [**318**]{}, 119 (2005).
A. Czarnecki and B. Krause, Phys. Rev. Lett. [**78**]{}, 4339 (1997).
T. Mannel and N. Uraltsev, Phys. Rev. D [**85**]{}, 096002 (2012).
T. Mannel and N. Uraltsev, JHEP [**03**]{}, 064 (2013).
C.A. Baker [*et al.*]{}, Phys. Rev. Lett. [**97**]{} 131801 (2006).
W.C. Griffith, M.D. Swallows, T.H. Loftus, M.V. Romalis, B.R. Heckel, and E.N. Fortson, Phys. Rev. Lett. [**102**]{}, 101601 (2009).
V.F. Dmitriev and R.A. Sen’kov Phys. Rev. Lett. [**91**]{}, 212303 (2003).
J. Baron [*et al.*]{} Science [**343**]{} (6168), 269 (2014).
Y.F. Orlov, W.M. Morse and Y.K. Semertzidis, Phys. Rev. Lett. [**96**]{}, 214802 (2006).
Y.K. Semertzidis, Proceedings of the DPF 2011 Conference (2011).
A. Lehrach, B. Lorentz, W. Morse, N. Nikolaev and F. Rathmann, arXiv:1201.5773 (2013)
J. Pretz, Hyperfine Interact. [**214**]{} (1-3), 111 (2013).
F. Rathmann, A. Saleev and N.N. Nikolaev, J. Phys. Conf. Ser. [**447**]{}, 012011 (2013).
P.K. Kabir, Phys. Rev. D [**25**]{}, 2013 (1982).
L. Stodolsky, Nucl. Phys. B [**197**]{}, 213 (1982).
C.-P. Liu and R.G.E. Timmermans, Phys. Rev. C [**70**]{}, 055501 (2004).
C.-P. Liu, R.G.E. Timmermans, Phys. Lett. B [**634**]{}, 488 (2006).
J. D. Bowman and V. Gudkov, Phys. Rev. C [**90**]{}, 065503 (2014).
Y.-H. Song, R. Lazauskas, and V. Gudkov, Phys. Rev. C [**83**]{}, 065503 (2011).
Yu. N. Uzikov and A. A. Temerbayev, Int. Journal of Modern Physics: Conf. Series [**40**]{}, 1660080 (2016); Yu. N. Uzikov and J. Haidenbauer, Phys. Rev. C [**94**]{}, 035501 (2016).
V. Gudkov and H. M. Shimizu, Phys. Rev. C [**97**]{}, 065502 (2018).
P. Fadeev and V. V. Flambaum, [arXiv:1903.08937]{}.
D.R. Entem, N. Kaiser, R. Machleidt, Y. Nosyk, Phys. Rev. C [**[91]{}**]{}, 014002 (2015).
E. Epelbaum, H. Krebs and U.-G. Meissner, Phys. Rev. Lett. [**115**]{}, 122301 (2015).
T.-S. Park, D.-P. Min, and M. Rho, Nucl. Phys. A [**596**]{}, 515 (1996); Y.-H. Song, R. Lazauskas, and T.-S. Park, Phys. Rev. C [**79**]{}, 064002 (2009). T.-S. Park [*et al.*]{}, Phys. Rev. C [**67**]{}, 055206 (2003). S. Kölling, E. Epelbaum, H. Krebs, and U.-G. Meissner, Phys. Rev. C [**80**]{}, 045502 (2009). S. Pastore, L. Girlanda, R. Schiavilla, M. Viviani, and R.B. Wiringa Phys. Rev. C [**80**]{}, 034004 (2009). S. Kölling, E. Epelbaum, H. Krebs, and U.-G. Meissner, Phys. Rev. C [**84**]{}, 054008 (2011). S. Pastore, L. Girlanda, R. Schiavilla, and M. Viviani, Phys. Rev. C [**84**]{}, 024001 (2011). M. Piarulli [*et al.*]{}, Phys. Rev. C [**87**]{}, 014006 (2013). S. Weinberg, Phys. Rev. Lett. [**17**]{}, 616 (1966); Phys. Rev. [**166**]{}, 1568 (1968); Physica A [**96**]{}, 327 (1979). C. Callan, S. Coleman, J. Wess, and B. Zumino, Phys. Rev. [**177**]{}, 2247 (1969).
E. Mereghetti, W.H. Hockings and U. van Kolck, Annals Phys. [**[325]{}**]{}, 2363 (2010).
J. Bsaisou, Ulf-G. Meissner, A. Nogga and A. Wirzba, Annals Phys. [**[359]{}**]{}, 317 (2015).
B. Grzadkowski, M. Iskrzynski, M. Misiak, and J. Rosiek, JHEP [**1010**]{}, 085 (2010).
J. de Vries, E. Mereghetti, R.G.E. Timmermans and U. van Kolck, Annals Phys. [**338**]{}, 50 (2013).
W. Dekens and J. de Vries, JHEP [**1305**]{}, 149 (2013).
J. de Vries, E. Mereghetti, R.G.E. Timmermans, and U. van Kolck, Phys. Rev. Lett. [**107**]{}, 091804 (2011).
W. Dekens [*et al.*]{}, JHEP [**07**]{}, 069 (2014).
C. M. Maekawa, E. Mereghetti, J. de Vries, and U. van Kolck, Nucl. Phys. A [**872**]{}, 117 (2011).
J. Bsaisou, C. Hanhart, S. Liebig, U.-G. Meissner, A. Nogga, and A. Wirzba, Eur. Phys. J. [**A 49**]{}, 31 (2013).
J. de Vries, [*et al.*]{}, Phys. Rev. C [**84**]{}, 065501 (2011)
J. Bsaisou, J. de Vries, C. Hanhart, S. Liebig, Ulf-G. Meissner, D. Minossi, A. Nogga, and A. Wirzba, JHEP [**03**]{}, 104 (2015), erratum JHEP [**05**]{}, 083 (2015).
I. Stetcu, C.-P. Liu, J.L. Friar, A.C. Hayes, and P. Navratil, Phys. Lett. B [**665**]{}, 168 (2008).
N. Yamanaka and E. Hiyama, Phys. Rev. C [**91**]{}, 054005 (2015)
Y.-H. Song, R. Lazauskas, and V. Gudkov Phys. Rev. C [**87**]{}, 015501 (2013)
M. Viviani, [*et al.*]{}, Phys. Rev. C [**89**]{}, 064004 (2014).
N. Fettes, [*et al.*]{}, Ann. Phys. [**283**]{}, 273 (2000).
E. Epelbaum, H. W. Hammer, and U.-G. Meissner, Rev. Mod. Phys. [**81**]{}, 1773 (2009).
R. Machleidt and D. R. Entem, Phys. Rep. [**503**]{}, 1 (2011).
L. Girlanda, S. Pastore, R. Schiavilla, and M. Viviani, Phys. Rev. C [**81**]{}, 034005 (2010).
L.E. Marcucci, F. Sammarruca, M. Viviani, R. Machleidt, arXiv:1809.01849.
E. Epelbaum, W. Glöckle, and U.-G. Meissner, Nucl. Phys. A [**747**]{}, 362 (2005)
A. Kievsky, S. Rosati, M. Viviani, L.E. Marcucci, and L. Girlanda, J. Phys. G: Nucl. Part. Phys. [**35**]{}, 063101 (2008).
L.E. Marcucci, A. Kievsky, L. Girlanda, S. Rosati, and M. Viviani, Phys. Rev. C [**80**]{}, 034003 (2009).
| {
"pile_set_name": "ArXiv"
} |
=-0cm=23.cm=16.cm =0.25cm=0.25cm
@twoside 0.5 cm 0.5cm 0.5cm
0truecm
16.5cm
=cmr17 scaled4 [U0.8mm N0.8mm I0.8mm V0.8mm E0.8mm R0.8mm S0.8mm I0.8mm T0.8mm 0=1= 0=0 0 by -1 A0.8mm T B0.8mm O0.8mm N0.8mm N0.8mm ]{}\
[ P0.8mm h0.8mm y0.8mm s0.8mm i0.8mm k0.8mm a0.8mm l0.8mm i0.8mm s0.8mm c0.8mm h0.8mm e0.8mm s I0.8mm n0.8mm s0.8mm t0.8mm i0.8mm t0.8mm u0.8mm t0.8mm ]{}\
[**Chiral fermion action\
with (8,0) worldsheet supersymmetry**]{}\
[**E. Ivanov${}^{(a)}$ and E. Sokatchev${}^{(b)}$**]{}\
[ ${}^{(a)}$*Laboratory of Theoretical Physics, Joint Institute for Nuclear Research, 141980 Dubna near Moscow, Russia*]{}\
[${}^{(b)}$*Physikalisches Institut, Universität Bonn, Nussallee 12, D-53115 Bonn, Germany*]{}\
[**Abstract**]{}\
[We propose an action describing chiral fermions with an arbitrary gauge group and with manifest $(8,0)$ worldsheet supersymmetry. The form of the action is inspired by and adapted for completing the twistor-like formulation of the $D=10$ heterotic superstring.]{}
---------------
Post address:
Nussallee 12
D-53115 Bonn
Germany
---------------
-----------------
BONN-TH-94-10
Bonn University
June 1994
-----------------
[**1. Introduction.**]{} Chiral fermions are a necessary ingredient of the self-consistent quantum heterotic superstring theory. They serve to compensate the conformal anomaly and ensure that the critical dimension of space-time is 10 [@kniga].
Recently various twistor-like formulations of the $D=10$ heterotic string at the classical level have been proposed [@Ton; @DGHS; @gr]. The basic attractive feature of twistor-like formulations of the heterotic superstring [^1] is the trading of the local $\kappa$-symmetry of the $D=10$ superstring action for $N=(8,0)$ local supersymmetry of the worldsheet. The twistor-like mechanism was first discovered by Sorokin et al [@volk] in the context of a $D=3,4$ superparticle with local $N=1,2$ worldline supersymmetry and then generalized to the $D=6$ [@DS] and $D=10$ [@GS] superparticle with $N=4$ and $N=8$ worldline supersymmetry and to heterotic superstrings in $D=4$ [@IK] and $D=6$ [@DIS] with $N=2,4$ worldsheet supersymmetry (see also [@berko]). It unveiled for the first time the transparent geometric meaning of $\kappa$-symmetry. The fact that all symmetries of the superstring action become manifest and geometrically interpretable within a twistor-like formulation gives us hope that the latter is most appropriate for covariant quantization of the heterotic superstring.
Keeping in mind this ambitious though still remote prospect, one may wonder how to consistently incorporate, to begin with at the classical level, the chiral fermions into the twistor-like formulation of the heterotic superstring while preserving all symmetries and geometric features of the latter. An early attempt in this direction was made by Tonin [@Ton]. There the chiral fermion superfield equation of motion was introduced into the action with a Lagrange multiplier. Unfortunately, the latter turned out to propagate unwanted degrees of freedom. Another approach to this problem was proposed by Sorokin and Tonin [@ST]. This time the field equations of the chiral fermions are derived from a quadratic $(8,0)$ superfield constraint, once again added to the superstring action with a Lagrange multiplier. To make the derivation unambiguous and to avoid the propagation of extra degrees of freedom, they have to restrict themselves only to supersymmetric solutions of their constraint and to resort to a rather subtle positiveness argument.
A different $(8,0)$ action for chiral fermions has recently been proposed by Howe [@Howe]. It is a standard bilinear superfield action with a specially chosen operator made out of up to 7 spinor derivatives. The action has a large abelian gauge invariance which helps to restrict the on-shell content of the theory to just chiral fermions. The superfields incorporating these fermions must satisfy a certain constraint which effectively ties up the internal symmetry index on the fermion fields to the (local) $SO(8)$ symmetry group of the worldsheet. As remarked in [@Howe], this results in a severe restriction on the internal symmetry group (it should not be bigger than $SO(4)$).
In the present letter we propose another chiral fermion action with local $(8,0)$ worldsheet supersymmetry and with an arbitrary gauge group. It is a modification of the one of Sorokin and Tonin [@ST]. Like in [@ST], an appropriate superfield terms is included into the Wess-Zumino term of the twistor superstring action sharing the same Lagrange multiplier superfield. As shown in [@DGHS], after fixing a certain gauge this Lagrange multiplier is reduced just to the string coupling constant. Then the new term produces the standard kinetic term for chiral fermions as well as an auxiliary field term. We stress that the chiral fermion modification of the superstring Wess-Zumino term we propose is essentially different from that in [@ST]. It is chosen so that requiring compatibility with the standard constraints on the background supergravity three-form [*does not*]{} imply equations of motion for the chiral fermions (as it does in [@ST]). Instead, the integrability condition results in a quadratic constraint on the chiral fermions superfields which simply reduces the off-shell component content of the superfields (quite similarly to standard linear irreducibility constraints). This constraint turns out to be also necessary for the off-shell local worldsheet supersymmetry of the chiral fermion action. The realization of off-shell supersymmetry in this action has some unusual features which we discuss in detail.
[**2. Preliminaries.**]{} We start by briefly recalling the structure of the heterotic superstring action in the twistor-like formulation as it has been given in [@DGHS]. The action consists of three parts, each of them given by an integral over the $(8,0)$ worldsheet superspace $$\label{twistact}
S = S_1 + S_2 + S_{wz} \;.
{\end{equation}}For our purposes it will be important to know
the explicit form of $S_{wz}$
\begin{equation} \label{wz}
S_{wz} = \int d^2 x d^8 \theta P^{MN} [\; B_{MN} +
E_{[M}^{\;\;+} E_{N \}}^{\;\;-} {\mbox
e}^\Phi E^{\underline{a}}_- E_{+\;\underline{a}}
- \partial_{[M}
Q_{N\}}\;]\;.$$ Here and in what follows non-underlined indices refer to the worldsheet superspace $\{ z^M \} = \{ x^-, x^+, \theta^\mu \}$, $\mu =1,...,8 $, i.e. $M = -,+,\mu$, while underlined ones refer to the target $D=10$ superspace $z^{\underline{M}}$. For instance, $\underline{a} =0,1,...,9$ is the vector tangent space index. The symbol $[MN\}$ means graded antisymmetrization. The quantities $B_{MN}$ and $E_{\pm}^{\underline{a}}$ are, respectively, the worldsheet superspace pull-backs of the background $D=10$ supergravity two-form $B_{\underline{MN}}$ and vielbeins $E_{\underline{M}}^{\;\;
\underline{A}}$ $$B_{MN} = (-)^{M (N + \underline{N})} \partial_N z^{\underline{N}}
\partial_M z^{\underline{M}} B_{\underline{MN}} \;, \;\;\;
E_{\pm}^{\underline{A}} = E_\pm^M\partial_M z^{\underline{N}}
E_{\underline{N}}^{\underline{A}}\;,$$ $E_{M}^{\;\; \pm}$ are elements of the vielbein and $E_\pm^M$ of the inverse vielbein matrix on the worldsheet, $\Phi$ is the dilaton field of background supergravity, $Q_N$ and $P^{MN}$ are worldsheet superfield Lagrange multipliers, whose rôle will be explained later on.
The term $S_1$ in (\[twistact\]) produces the so called geometro-dynamical constraint $$\label{gd}
E_\alpha^{\underline{a}} \equiv E_\alpha^M \partial_M z^{\underline{N}}
E_{\underline{N}}^{\underline{a}} = 0$$ on the spinor pull-back of the vector target superspace vielbeins. It has as a corollary a twistor representation for the left-handed Virasoro vector $$\label{twlk}
E_-^{\underline{a}} = {1\over 8} E_\alpha^{\underline \alpha}
\gamma^{\underline{a}}_{\underline{\alpha\beta}} E_\alpha^{\underline\beta}$$ which provides a [*twistor-like solution*]{} to the left-handed Virasoro constraint $E_-^{\underline{a}} E_{-\underline{a}} = 0$ of the superstring. The term $S_2$ in (\[twistact\]) enforces an irreducibility condition on the spinor component $e_{\alpha}^+$ of the worldsheet inverse vielbein (see eq. (\[cnstr\]) below); the latter covariantizes the worldsheet superspace derivatives with respect to arbitrary shifts of the right-handed coordinate $x^+$ and thus provides the Lagrange multiplier for the right-handed Virasoro constraint. More details on these two terms can be found in [@DGHS].
The term (\[wz\]) is the leading one in the twistor superstring action, since it produces the entire superstring component action. It is instructive to discuss in some detail how this works, because we will use the same mechanism for generating the chiral fermion component action.
The structure of $S_{wz}$ is entirely specified by the requirement that the result of varying with respect to the Lagrange multiplier $P^{MN}$, $$\label{bequat}
B_{MN} = \partial_{[M} Q_{N \}} -
E_{[M}^{\;\;+} E_{N \}}^{\;\;-} {\mbox
e}^\Phi E^{\underline{a}}_- E_{+\;\underline{a}}\;.$$ is consistent with the $D=10$ supergravity constraints on the three-form field strength $H_{\underline{KMN}}$ of the two-form $B_{\underline{MN}}$ [@10sg]. This means that eq. (\[bequat\]) implies the integrability condition $$\label{hequat}
H_{KMN} \equiv \partial_{[K} B_{MN \}} =
- \partial_{[K} (E_{M}^{\;\;+} E_{N \}}^{\;\;-} {\mbox
e}^\Phi E^{\underline{a}}_- E_{+\;\underline{a}})\;.
{\end{equation}}Using the geometro-dynamical constraint (\ref{gd}) and its corollary
(\ref{twlk}), one
can show that equation (\ref{hequat}) is consistent with the
$D=10$ background supergravity constraints and is actually equivalent to them.
When (\ref{hequat}) holds the action term $S_{wz}$ is invariant
under the following gauge transformation
\begin{equation} \label{gauge1}
\delta P^{MN} = \partial_K \Sigma^{MNK}\;,
{\end{equation}}where $\Sigma^{KMN}(z)$ is a totally (graded) antisymmetric
superfield.
To see how (\ref{wz}) produces the standard component heterotic
superstring action, one varies with respect to the Lagrange multiplier
$Q_{M}$ and obtains
\begin{equation}
\partial_N P^{MN} = 0\;.$$ The general solution of this equation is given by $$\label{cohom1}
P^{MN} = \partial_K \Lambda^{MNK} + \theta^8 \delta_+^{\;\;[M}
\delta_-^{\;\;N \}} T \;.$$ The second (cohomological) term contains the constant $T$, $$\partial_+ T = \partial_- T = 0\;.$$ Then one substitutes this solution back into $S_{wz}$ and observes that the first term in (\[cohom1\]) can be completely gauged away using the gauge freedom (\[gauge1\]). As a result in this gauge in $P^{MN}$ there survives only one component $$\label{pgauge}
P^{+-} = \theta^8 T$$ and $S_{wz}$ (\[wz\]) is reduced to the ordinary $x$-space integral of the first component of the superfield expression within the square brackets. It contains, in particular, the superstring WZ term $$S_{wz} = T \int d^2 x \;\partial_- z^{\underline{N}}
\partial_+ z^{\underline{M}} B_{\underline{MN}} + ... \;,$$ as well as the superstring kinetic term. The constant $T$ is naturally interpreted as the superstring tension.
To close this introductory section, we present the worldsheet gauge group (including local $(8,0)$ supersymmetry) in the twistor formulation. The whole action [(\[twistact\])]{} and in particular $S_{wz}$ are invariant under the following restricted class of diffeomorphisms of the $(8,0)$ superspace $$\begin{aligned}
\label{susy}
\delta \theta_\alpha &=& -{i\over 2} D_{\alpha} \Lambda^-\;, {\nonumber}\\
\delta x^- &=& \Lambda^- - {1\over 2} \theta_\alpha D_{\alpha} \Lambda^-\;,
{\nonumber}\\
\delta x^+ &=& \Lambda^+ \;,\end{aligned}$$ with $\Lambda^{\pm} (z)$ being arbitrary unconstrained superfield parameters. This group is chosen so that it leaves the “almost flat" derivatives $$\begin{aligned}
\label{covder}
D_{\alpha} &=& \partial_{\alpha} + i\theta_{\alpha} \partial_- +
e_\alpha^{\;\;+}\partial_+ {\nonumber}\\
D_- &=& \partial_- - {i\over 8} D_\alpha e_\alpha^{\;\;+} \partial_+ {\nonumber}\\
D_+ &=& \partial_+ \;\end{aligned}$$ covariant. In (\[covder\]) the single non-trivial component of the worldsheet superspace inverse vielbeins (defined by $D_A = E_A^M\partial_M$) is $e^{\;\;+}_\alpha$. It is needed to ensure covariance under the shifts of $x^+$ and obeys the constraint ($\{ \}$ denotes the symmetric traceless part) $$\label{cnstr}
D_{\{ \alpha} e_{\beta \}}^{\;\;+} = 0\;,$$ so that the covariant derivatives $D_\alpha, D_-$ still form the flat algebra $$\label{susyalg}
\{ D_\alpha, D_\beta \} = 2i \delta_{\alpha\beta}D_-\;, \;\;\;
[D_\alpha, D_-] =
0\;.$$ The constraint (\[cnstr\]) is produced by the term $S_2$ in the superstring action (\[twistact\]).
From the explicit form of the covariant derivatives $D_\alpha, D_-,
D_+$ above one can read off the form of the worldsheet zweibein matrix. In what follows we shall make use of the vielbeins $E^{\pm}_M$ entering the WZ term (\[wz\]): $$\begin{aligned}
\label{zweibein}
E^+_+ &=& 1\;,\;\; E^+_- = {i\over 8}\; D_\alpha e_\alpha^+\;,\;\;E^+_\mu =
- e^+_\mu + {1\over 8} \theta_\mu\; D_\alpha e^+_\alpha \;, {\nonumber}\\
E^-_+ &=& 0\;, \;\; E^-_- = 1\;, \;\; E^-_\mu = -i \theta_\mu \;.\end{aligned}$$
Finally, we give the transformation laws of the covariant derivatives and $e^+_\alpha$ under the restricted diffeomorphism group (\[susy\]) $$\begin{aligned}
\label{transder}
\delta D_\alpha &=& {i\over 2} (D_\alpha D_\beta)\Lambda^- D_\beta {\nonumber}\\
\delta D_- &=& -(D_-\Lambda^-)D_- + {i\over 2}(D_- D_\alpha \Lambda^-)
D_\alpha {\nonumber}\\
\delta D_+ &=& -\left( D_+\Lambda^+ + {i\over 2}(D_+D_\alpha \Lambda^-)
e_\alpha^{\;+} + {i\over 8} (D_+\Lambda^-) D_\alpha e_\alpha^{\;+} \right)
D_+ {\nonumber}\\
&& - (D_+\Lambda^-) D_- + {i\over 2}(D_+D_\alpha \Lambda^-)D_\alpha {\nonumber}\\
\delta e_\alpha^{\;+} &=& D_\alpha \Lambda^+ + {i\over 2} D_\alpha D_\beta
\Lambda^- e_\beta^{\;+}\;.\end{aligned}$$
[**3. Chiral fermion action.**]{} Let us now turn to our basic task. We shall identify the chiral fermions with the first components of some anticommuting $(8,0)$ superfields $\Psi^i (z)$. The index $i$ is in general arbitrary, but for definiteness we shall regard it as a vector $SO(n)$ index, for instance as the vector index of $SO(32)$. The action which we propose for this superfield will be the following addition to the superstring WZ term [(\[wz\])]{} $$\label{fermact}
S_f = \int d^2x d^8 \theta
\left\{ P^{MN}
E^{\;+}_{[M}\; (i\; \partial_{N\}}
\Psi^i \Psi^i + {1\over 8} E^-_{N\}}
\nabla_\alpha \Psi^i \nabla_\alpha \Psi^i)
+
P^{\{\alpha \beta\}}
\nabla_\alpha \Psi^i \nabla_\beta \Psi^i \right\}\;.$$ Here $P^{MN}$ is the same Lagrange multiplier as in the superstring action (\[wz\]) and $P^{\{\alpha \beta \}}$ is a new symmetric traceless Lagrange multiplier, whose meaning will be clarified later on. The covariant derivative $\nabla_\alpha$ is defined as follows: $$\label{nabla}
\nabla_\alpha = D_\alpha - (\partial_+ e^+_\alpha) \; w \;.$$ The operator $w$ measures the weight with respect to local dilatations with the superfield parameter $D_+\Lambda^+ + \ldots$ (see the first term in the transformation law $\delta D_+$ in (\[transder\])). As follows from [(\[transder\])]{}, the derivative $D_+$ has weight $-1$, so the vielbein $E^{\;+}_M$ present in [(\[fermact\])]{} has weight $+1$. Thus, to achieve invariance of [(\[fermact\])]{} we should ascribe to the superfield $\Psi^i$ the weight $-{1/2}$. The derivative [(\[nabla\])]{} is covariant when acting on a superfield with an arbitrary weight $w$. It still satisfies the flat algebra [(\[susyalg\])]{} together with the modified covariant $x^-$ derivative (its own weight is zero) $$\nabla_- = D_- + {i\over 8}\; \partial_+ (D_\alpha e^+_\alpha) \; w \; .$$ Note that the term with $\partial_N\Psi$ in [(\[fermact\])]{} is covariant as it stands, since the superfields $\Psi^i$ anticommute.
In the superstring action the internal symmetry group of the chiral fermions, e.g., $SO(32)$ is gauged. Therefore one needs to add the pull-back of the $D=10$ super-Yang-Mills connection $E_\alpha^{\underline M} A_{\underline M}(\underline z)$ to the covariant derivative (\[nabla\]). However, this does not affect the algebra of the derivatives, provided the connection $A_{\underline M}(\underline z)$ satisfies the constraints of $D=10$ super-Yang-Mills theory [@SYM]. Therefore, for simplicity in what follows we shall omit the gauge field term in $\nabla_\alpha$.
Now we are ready to explain why we propose the chiral fermion superfield action just in the form [(\[fermact\])]{}. First of all, we wish to produce the standard component kinetic term of the chiral fermions $$\sim \psi^i \partial_- \psi^i\;,\;\;\; \psi^i = \Psi^i|_{\theta =0}\;,$$ using the same mechanism as in the case of the superstring WZ term, i.e. by passing to the gauge [(\[pgauge\])]{}. This accounts for the term $\partial_N\Psi^i\Psi^i$ in [(\[fermact\])]{}. It is of the same form as the one proposed in [@ST]. However, the second $\Psi$ term in [(\[fermact\])]{} was not included in the action in [@ST]. The fact that it should be there follows from two requirements:
\(i) the integrability condition [(\[hequat\])]{}, which is equivalent to the $D=10$ supergravity constraints, should not be affected by adding the action [(\[fermact\])]{} to the original superspace WZ term [(\[wz\])]{};
\(ii) the constraint on the superfields $\Psi^i$ following from requirement (i) [*should not put the theory on shell*]{}. Precisely this point accounts for the main difference between our action and that of [@ST].
Let us first explain how requirement (i) is satisfied. In the presence of the new action term (\[fermact\]) varying with respect to $P^{MN}$ leads to modifications of the two- and three-forms in eq. [(\[bequat\])]{} and its consequence [(\[hequat\])]{}. We denote them by $\hat B_{MN}$ and $\hat H_{KMN}$, respectively: $$B_{MN} \Rightarrow B_{MN} + \hat B_{MN}\;,\;
H_{KMN} \Rightarrow H_{KMN } + \hat H_{KMN}\;,\;\;
\hat H_{KMN}= \partial_{[K } \hat B_{MN \}}\;.$$ So, we require $$\label{hvanish}
\hat H_{[KMN \}}= 0 \;.$$ To find the constraint on $\Psi$ following from (\[hvanish\]), we replace for the moment the expression $\nabla_\alpha \Psi^i \nabla_\alpha \Psi^i$ in [(\[fermact\])]{} by an unknown $X$ and then write down $$\hat B_{MN} =
E^{\;+}_M \;(i\; \partial_N
\Psi^i \Psi^i + {1\over 8}\; E^-_N\; X) - (-)^{MN}\; (M \leftrightarrow N)\;.$$ Using the explicit form of the zweibein matrix [(\[zweibein\])]{} it is straightforward though a bit tedious to compute all the components of $\hat H_{KMN}$ and to see that the necessary and sufficient condition for them to vanish is $$\label{constr}
\nabla_\alpha \Psi^i \nabla_\beta \Psi^i - {1\over 8}\;
\delta_{\alpha \beta}\; X =
0 \;\Rightarrow X = \nabla_\alpha \Psi^i \nabla_\alpha \Psi^i \;.$$ Let us give explicitly several components of $\hat H_{KMN}$ $$\begin{aligned}
\hat H_{\alpha \beta \gamma} &=& -2i \;E^+_{\alpha}\; [\;(
\nabla_\beta \Psi^i \nabla_\gamma \Psi^i - {1\over 8}\;
\delta_{\beta \gamma}\; X)
{\nonumber}\\
&& - {1\over 16}\; \theta_\beta\; (
16i\; \nabla_- \Psi^i \nabla_\gamma \Psi^i - \nabla_\gamma X )\;] +
symmetrization \label{1}\\
\hat H_{+\alpha \beta} &=& -2i \; (
\nabla_\alpha \Psi^i \nabla_\beta \Psi^i - {1\over 8}\;
\delta_{\alpha \beta}\; X) {\nonumber}\\
&&+ {i\over 8}\; \theta_\alpha\; (
16i\; \nabla_- \Psi^i \nabla_\beta \Psi^i - \nabla_\beta X) +
(\alpha \leftrightarrow \beta) \label{2} \\
\hat H_{\alpha + -} &=& {1\over 8}\; (\nabla_\alpha X - 16i\; \nabla_-\Psi^i
\nabla_\alpha \Psi^i)\;. \label{3}\end{aligned}$$ From the vanishing of the last component it follows that $$\label{conseq}
16i\; \nabla_- \Psi^i \nabla_\beta \Psi^i - \nabla_\beta X = 0\;,$$ then from the vanishing of the second one follows the constraint [(\[constr\])]{}. Eq. [(\[conseq\])]{} is in fact a corollary of [(\[constr\])]{}.
The most essential difference from the approach of Sorokin and Tonin becomes clear at this point. They do not include the $X$ term in the superstring WZ action and so from the same consistency condition [(\[hvanish\])]{} they derive the constraint $$\label{ston}
\nabla_\alpha \Psi^i \nabla_\beta \Psi^i = 0.$$ It is much stronger than [(\[constr\])]{} in that it implies the equations of motion for $\Psi^i$. Below we will show that our constraint [(\[constr\])]{} is purely kinematical and serves to reduce the off-shell field content of $\Psi^i$.
We would also like to comment on the approach of Howe [@Howe]. His action is based on a [*linear constraint*]{} on the chiral fermion superfield $\Psi^{i' q}$: $$\label{howe}
D_\alpha \Psi^{i'q} = (\gamma^{i'})_{\alpha\dot\alpha} P^{\dot\alpha\; q}.$$ Here the $SO(32)$ index is split into an $SO(8)$ index $i'$ and an $SO(4)$ index $q$. The indices $\alpha$ and $\dot\alpha$ are $s$ and $c$ spinor indices of $SO(8)$ and $\gamma^{i'}$ is the $SO(8)$ gamma matrix. It should be pointed out that eq. (\[howe\]) provides a solution (but not the general one) to our constraint (\[constr\]) for this particular arrangement of the indices. The advantage of the constraint (\[howe\]) is its linearity, but its serious drawback (noted in [@Howe] as well) is the strong restriction on the chiral fermion gauge group, which is at most $SO(4)$ (or a tensor product of several $SO(4)$’s).
Coming back to our action [(\[fermact\])]{}, it remains to explain the rôle of the second Lagrange multiplier term in it. On the one hand, it serves to impose the constraint (\[constr\]) on shell. On the other hand, its presence helps to maintain the gauge freedom [(\[gauge1\])]{} which allows one to bring $P^{MN}$ to the form [(\[pgauge\])]{}. The gauge variation of the $P^{MN}$ term in [(\[fermact\])]{} is proportional to the constraint [(\[constr\])]{}, so it can be compensated by an appropriate variation of the new Lagrange multiplier $P^{\{\alpha \beta\}}$. The explicit form of this variation is not too enlightening, so we do not give it here. Note that $P^{\{\alpha \beta\}}$ has its own gauge freedom $$\label{gauge2}
\delta P^{\{\alpha \beta \}} = \nabla_\gamma \; \Lambda^{\{\gamma \alpha
\beta \}}\;,$$ which allows one to gauge away some components of this superfield.
The meaning of the constraint [(\[constr\])]{} and how it helps to make the action [(\[fermact\])]{} supersymmetric in the gauge [(\[pgauge\])]{} will be discussed in the next Section. In this gauge [(\[fermact\])]{} takes the form $$\label{fermact2}
S_f =T \int d^2 x d^8\theta
\left\{\theta^8 \; (i\; \nabla_-
\Psi^i \Psi^i + {1\over 8}
\nabla_\alpha \Psi^i \nabla_\alpha \Psi^i) +
P^{\{\alpha \beta\}}
\nabla_\alpha \Psi^i \nabla_\beta \Psi^i \right\}\;.$$
[**4. Peculiarities of the chiral fermion action.**]{} Here we discuss some unusual features of the action [(\[fermact2\])]{} as it stands, leaving aside its superstring-inspired appearance. It is worth mentioning that though we have introduced it in the framework of the $D=10$ heterotic superstring and of an $(8,0)$ worldsheet superspace, its basic features actually do not depend on the Grassmann dimension of the worldsheet superspace. So we may consider a general $(N,0)$ superspace, which corresponds to the change $8\rightarrow N$ in [(\[fermact2\])]{}. We assume that the $\theta$’s are always real and transform, in general, according to the vector representation of the automorphism group $SO(N)$ [^2]. We also assume that the derivatives $\nabla_\alpha,
\nabla_-$ in [(\[fermact2\])]{} are flat, i.e. we put $e^+_\alpha = 0$. Actually, one can always choose a gauge in which all covariant derivatives are almost flat [@DIS], [@DGHS], $$e^+_\alpha = i\theta_\alpha \;g_{--}\;, \;\;\; D_- = \partial_- +
g_{--}\;\partial_+\;,$$ where $g_{--}$ is the component of the worldsheet metric responsible for the second Virasoro constraint. For simplicity we put $g_{--}$ equal to zero (which is only possible locally), although this is not essential for what follows.
First we demonstrate that the action $$\label{Nferm}
S = T\int d^2 x d^N\theta
\left\{ \theta^N \; (i\;
\partial_-
\Psi^i \Psi^i + {1\over N} \;
D_\alpha \Psi^i D_\alpha \Psi^i) +
P^{\{\alpha \beta\}}
D_\alpha \Psi^i D_\beta \Psi^i \right\}$$ is supersymmetric despite the presence of explicit $\theta$’s. The easiest way to see this is to observe that $$i\;\partial_-
\Psi^i \Psi^i + {1\over N}
D_\alpha \Psi^i D_\alpha \Psi^i = {1\over N}\;D_\alpha (\Psi^i D_\alpha
\Psi^i)\;,$$ then to integrate in both terms of the action by parts and, finally, to rewrite it in a very simple form resembling Chern-Simons type actions [@HT] $$\label{Nferm2}
S = T \int d^2x d^N \theta P^{\alpha}\;\Psi^i D_\alpha \Psi^i$$ with $$\label{p}
P^{\alpha} = \theta^{N-1}_\alpha + D_\beta P^{\{\beta \alpha\}}\;,\;\;
\theta^{N-1}_\alpha \equiv D_\alpha \theta^N\;$$ (actually, when passing from [(\[Nferm\])]{} to [(\[Nferm2\])]{}, there appears a minus sign in the case of an odd $N$, but it can be absorbed into the constant $T$).
At this point we can forget the particular structure of $P^\alpha$ [(\[p\])]{} coming from the original definition of the chiral fermion action. Instead, we can define the action by eq. [(\[Nferm2\])]{} with $P^\alpha$ satisfying the constraint $$\label{constrP}
D_\alpha P^\alpha = 0\;.$$ It is easy to check that [(\[p\])]{} provides the general solution to this constraint up to an arbitrary right-moving function $a(x^+)$ in front of the first (cohomological) term in (\[p\]). Note, however, that the action (\[Nferm2\]) and the constraint [(\[constrP\])]{} are invariant under $x^+$ dependent scale transformations $$\Psi^i \rightarrow \lambda^{1/2} (x^+) \Psi^i\;,\; P^{\alpha}
\rightarrow \lambda^{-1} (x^+) P^\alpha$$ which allow us to gauge $a(x^+)$ into any non-zero constant. So, the action [(\[Nferm2\])]{} with the additional constraint [(\[constrP\])]{} describe the general situation.
In this Chern-Simons-like representation the chiral fermion action does not include explicit $\theta$’s (they are hidden in the solution of the constraint [(\[constrP\])]{}) and is invariant (together with the constraint) with respect to the local transformations [(\[susy\])]{} with $\Lambda^+ = \Lambda^+(x^+)$, provided $P^\alpha$ transforms according to the law $$\label{Psusy}
\delta P^\alpha = -{i\over 2}\;(D_\beta D_\alpha \Lambda^-) P^\beta
+ {1\over 2} (N-2) (\partial_- \Lambda^-) P^\alpha \;.$$ The second term in [(\[Psusy\])]{} cancels the variation of the superspace integration measure in [(\[Nferm2\])]{}.
Now let us discuss some peculiar features of the realization of rigid supersymmetry in the above action. To this end we will go to components by varying in (\[Nferm\]) with respect to the Lagrange multiplier $P^{\{\alpha \beta \}}$, substituting the constraint back into the action (this will be justified below) and finally integrating over the $\theta$’s. The $\theta$ integration is now trivial because of the presence of the factor $\theta^N$. In this way we obtain that for an arbitrary $N$ the chiral fermions are described off shell by the action $$\label{fermcomp}
S_f = T\int d^2 x (i\;\partial_- \psi^i \psi^i + {1\over N}\;
b^i_\alpha b^i_\alpha )\;;\;\;\;
b^i_\alpha \equiv D_\alpha \Psi^i |_{\theta = 0}\;,$$ supplemented by the nonlinear superfield constraint $$\label{constr2}
D_{\alpha} \Psi^i D_\beta \Psi^i - {1\over N}\;\delta_{\alpha \beta}\;
D_{\gamma} \Psi^i D_\gamma \Psi^i = 0\;.$$
The first unusual feature of the action [(\[fermcomp\])]{} is that, irrespective of the value of $N$, it involves only the two first fields from the $\theta$ expansion of $\Psi^i$. Nevertheless, it is off-shell supersymmetric! To see this, we first write down explicitly the first two component constraints following from the superfield one [(\[constr2\])]{} $$\begin{aligned}
&& b^i_{\{\alpha} b^{i}_{\beta \}} = 0\;, \label{compconstr1} \\
&& \phi^i_{[\gamma \{\alpha]} b^i_{\beta\}} +
i \delta_{\gamma \{\alpha} \partial_- \psi^i b^i_{\beta\}} = 0
\label{compconstr2}\end{aligned}$$ with $$\phi^i_{[\alpha \beta]} \equiv {1\over 2}\;D_{[\alpha} D_{\beta]} \Psi^i|_
{\theta = 0}\;.$$ The $\gamma\beta$ trace of [(\[compconstr2\])]{} yields the important relation $$\label{constr3}
\phi^i_{[\alpha \beta]} b^i_\beta - i(N-1)\;\partial_-\psi^i b^i_\alpha
= 0 \;,$$ while the part with mixed symmetry results in some further restriction on $ \phi^i_{[\alpha \beta]} $ which does not involve the physical field $\psi^i$ (its explicit form is not needed for our purposes).
Now, let us make a rigid supersymmetry transformation in [(\[fermcomp\])]{} (the transformation laws for the component fields follow from the expansion of the superfield $\Psi^i$), $$\label{supsi}
\delta \psi^i = \epsilon_\alpha\;b^i_\alpha\;,\;\;\;\;
\delta b^i_\alpha = \epsilon_\gamma\;(\phi^i_{[\gamma \alpha]} + i\delta
_{\gamma \alpha}\;\partial_-\psi^i)\;.$$ The variation of the action is $$\delta S_f = T\int d^2x {2\over N} \epsilon_\gamma
[\phi^i_{[\gamma \alpha]}b^i_\alpha - i(N-1)\;\partial_-\psi^i b^i_\gamma
]$$ and it vanishes in virtue of [(\[constr3\])]{}. Thus, the action [(\[fermcomp\])]{} is supersymmetric due to the superfield constraint [(\[constr2\])]{}. As we have just seen, in fact only the component [(\[constr3\])]{} of this constraint is involved in achieving the supersymmetry of the action; the other component constraints are needed to ensure that the whole set of them are supersymmetric in their own right [^3].
Now we discuss the meaning of the component constraints. Splitting $b^i_\alpha$ into a “radial” and an “angular” parts, $$\label{decompb}
b^i_\alpha = m \hat b^i_\alpha\;,$$ we can rewrite [(\[compconstr1\])]{} as $$\label{constrb}
m\;(\hat b^i_\alpha \hat b^i_\beta - {1\over N}\;\delta_{\alpha \beta}) =
0\;.$$ If one assumes $m$ to be non-singular off-shell (this is not so on-shell, see below), then [(\[constrb\])]{} implies that $b^i_\alpha$ is an orthogonal $SO(N)$ matrix if $n = N$ or it represents the coset $SO(n)/SO(n-N)$ if $ n > N$. In order for this to make sense we have to take $n \geq N$. With the help of $\hat b^i_\alpha$ one can covariantly split any $SO(n)$ vector into an $SO(n-N)$ projection and an orthogonal $N$-dimensional one which is inert under $SO(n)$, but is transformed in a proper way by the automorphism $SO(N)$. So $\hat b^i_\alpha$ is a sort of a “bridge” relating these two groups. The meaning of the constraint [(\[constr3\])]{} becomes clear now: it states that such an $SO(N)$ projection of the component $\phi^i_{[\gamma \alpha]}$ for some irreducible combination of its $SO(N)$ indices is not independent, but is expressed in terms of $\partial_-\psi^i$. As was already mentioned, [(\[compconstr2\])]{} implies more constraints on this projection which amount to the vanishing of some other of its $SO(N)$ irreducible components. The rest of $\phi^i_{[\gamma \alpha]}$ remains arbitrary. For instance, in the case $N=2$ we have only the constraint [(\[constr3\])]{} on the $n$ fields $\phi^i = \epsilon_{\alpha \beta}\phi^i_{\alpha \beta}$, so in the latter there remain $n-2$ independent components; in the case $N=3$ we have five more constraints of the second type alongside [(\[constr3\])]{}, which together leave $3n - 8$ independent components in the $3n$ fields $\phi_{[\alpha \beta]}$, etc. It can be shown that all of the subsequent component constraints in [(\[constr2\])]{} have the typical structure of [(\[compconstr2\])]{}: they mean that some $b$ projections of the corresponding higher components of $\Psi^i$ either vanish or are expressed in terms of the lower ones.
It should be pointed out that a priori it is not so evident that the constraint [(\[constr2\])]{} is purely kinematical and does not produce differential conditions of the kind of equations of motion for some higher components of $\Psi^i$ and for certain values of $N$. If it were the case, we would not be allowed to substitute the constraint back into the action. Fortunately, simple arguments show that for any $N$ this constraint remains purely kinematical off shell.
To convince ourselves, let us pull out a constant part from $\Psi_i$, $$\Psi_i = \theta_\alpha \delta_{\alpha i} + \tilde \Psi_i$$ and split the index $i$ into $i = ({\alpha}, {i'})$, where ${\alpha}$ and ${i'}$ run, respectively, from 1 to $N$ and from $N+1$ to $n$. Then $\tilde \Psi_i$ splits into a pair of superfields, $$\tilde \Psi_i = (\tilde \Psi_{{\alpha}}, \tilde \Psi_{{i'}}).$$ In terms of these superfields the original constraint (\[constr2\]) can be rewritten as a linear one, $$\label{linconstr2}
D_{\alpha} \Phi_{\beta} + D_{\beta} \Phi_{\alpha} -
\delta_{\alpha \beta} {2\over N} D_{\gamma} \Phi_{\gamma} = 0,$$ where $$\label{rel}
\Phi_{\alpha} \equiv \tilde \Psi_{\alpha} +
{1\over 2} \tilde \Psi_{{\beta}} D_{\alpha} \tilde \Psi_{{\beta}}
+ {1\over 2} \tilde \Psi_{{i'}} D_{\alpha} \tilde \Psi_{{i'}}\; .$$ Thus, we have a purely kinematical constraint for the superfield $\Phi_{\beta}$ which has a simple general solution: $$\Phi_\alpha = D_\alpha G (z) + \theta_\alpha F(x^+, x^-)\;.$$ On the other hand, the relation (\[rel\]) is just a canonical redefinition of $\tilde
\Psi_{\alpha}$: one can reexpress it in terms of $\Phi_{\alpha}$ and $\tilde \Psi_{{i'}}$ from (\[rel\]) by means of iterations. The superfield $\tilde \Psi_{{i'}}$ remains entirely unconstrained. So, it becomes clear that for any $N$ the constraint (\[constr2\]) does not contain any dynamics off shell and merely expresses some components of $\Psi_i$ in terms of others (or puts them equal to zero). The corresponding Lagrange multiplier $P^{\{\alpha \beta\}}$ does not contain propagating degrees of freedom. We note that just because the constraint [(\[constr2\])]{} is kinematical, there is no local symmetry of Siegel’s type [@Siegel] associated with $P^{\{\alpha \beta\}}$ in the actions [(\[fermact\])]{}, [(\[fermact2\])]{}, [(\[Nferm\])]{}, [(\[Nferm2\])]{}. Once again, this is in contrast to ref. [@ST], where the constraint (\[ston\]) is dynamical and its Lagrange multiplier needs some sort of Siegel’s invariance in order not to propagate.
Thus we have seen that there are no problems with the off-shell supersymmetry of the chiral fermion action. However, going on shell in this action is rather subtle. As follows from [(\[fermcomp\])]{}, on shell $$\label{eqmob}
b^i_\alpha = 0$$ and we are no longer allowed to divide by $m$ in the constraints [(\[compconstr1\])]{}, [(\[compconstr2\])]{} and the subsequent ones. At first sight this does not lead to difficulties as [(\[compconstr1\])]{}, [(\[compconstr2\])]{} are satisfied identically if $b^i_\alpha$ satisfies [(\[eqmob\])]{}. However, beginning with $N=3$, in the higher-order constraints there appear terms which are not multiplied by $b^i_\alpha$ and so do not vanish on shell. For instance, in the $N=3$ case on shell there remains the following quadratic constraint on the field $\phi^i_{[\gamma \alpha]}$ $$\label{badconstr}
\phi^i_{[\gamma \alpha]} \phi^i_{[\gamma \beta]}
\epsilon_{\alpha\beta\sigma} = 0\;.$$ For higher $N$ there appear quadratic constraints of this type including derivatives of fields. This would seem to create difficulties if these fields were somehow present in the action. However, the only auxiliary field entering the component action (\[fermcomp\]) is in fact the “radial part" $m$ of $b^i_\alpha$. The others are needed only to supersymmetrize the constraint [(\[constr3\])]{} (itself required for the supersymmetry of the action). This indicates that the appearance of such strange constraints on shell is harmless. When we vary with respect to the fields $\psi^i$ and $m$ in the action to obtain the equations of motion, we should vary the constraints as well. The latter simply serve to partially express the variations of the higher-order components of $\Psi^i$ through $\delta \psi^i$ and $\delta m$. The presence of the terms of the type [(\[badconstr\])]{} in the component constraints results in factors of $m^{-1}$ in the expressions for the variations of the higher-order fields. The more we approach the minimum, the bigger the variations of these fields become, tending to infinity at the minimum itself. But we should not care about these variations as they do not contribute to the action functional, i.e. do not generate any equations of motion. Another argument is that on shell $(N,0)$ supersymmetry is trivially realized. Indeed, there the supersymmetry variation (\[supsi\]) of $\psi^i$ vanishes, but this does not contradict the algebra of supersymmetry as on shell $$\partial_-\psi^i = 0.$$
In conclusion we can say that the form of the chiral fermion action with manifest (local) worldsheet supersymmetry presented here completes the twistor-like formulation of the heterotic string in $D=10$ and can serve as a starting point for a new attempt to covariantly quantize the theory.
[**Acknowledgements**]{} The authors have profited from discussions with F. Delduc, A. Galperin and V. Fateev. E.I. would like to thank Prof. V. Rittenberg for hospitality at the Physics Institute of the University of Bonn, where this work has been done. E.I. is also grateful to the Russian Foundation of Fundamental Research, grant 93-02-3821, and to the International Science Foundation, grant M9T000, for financial support.
[22]{}
M. B. Green, J. H. Schwarz and E. Witten, [*Superstring Theory*]{}, CUP, 1987. M. Tonin, [*Phys. Lett.*]{} [**283B**]{} (1992) 213; [*Int. J. Mod. Phys.*]{} [**A7**]{} (1992) 6013. F. Delduc, A. Galperin, P. Howe and E. Sokatchev, [*Phys. Rev.*]{} [**D47**]{} (1992) 578. S. Aoyama, P. Pasti and M. Tonin, Phys. Lett. [**283B**]{} (1992) 213;\
I. Bandos, D. Sorokin, M. Tonin and D. Volkov, [*Phys. Lett.*]{} [**319B**]{} (1993) 445. A. Galperin and E. Sokatchev, [*Phys. Rev.*]{} [**D48**]{} (1993) 4810;\
V. Chikalov and A. Pashnev, [*Mod. Phys. Lett.*]{} [**A8**]{} (1993) 285;\
P. Pasti and M. Tonin, preprint DFPD/94/TH/05, Padova (1994);\
I. Bandos, M. Cederwall, D. Sorokin and D. Volkov, preprint ITP-94-10, Göteborg (1994). P. Pasti and M. Tonin, [*Nucl. Phys.*]{} [**B418**]{} (1994) 337;\
E. Bergshoeff and E. Sezgin, preprint CTP TAMO-67/93. D. P. Sorokin, V. I. Tkach and D. V. Volkov, [*Mod. Phys. Lett.*]{} [**A4**]{} (1989) 901;\
D. P. Sorokin, V. I. Tkach, D. V. Volkov and A. A. Zheltukhin, [*Phys. Lett.*]{} [**216B**]{} (1989) 302. F. Delduc and E. Sokatchev, [*Class. Quantum Grav.*]{} [**9**]{} (1991) 361. A. Galperin and E. Sokatchev, [*Phys. Rev.*]{} [**D46**]{} (1992) 714. E. A. Ivanov and A. A. Kapustnikov, [*Phys. Lett.*]{} [**267B**]{} (1991) 175. N. Berkovits, [*Phys. Lett.*]{} [**232B**]{} (1989)184; [**241B**]{} (1990) 497; [*Nucl. Phys.*]{} [**B350**]{} (1991) 193; [**B358**]{} (1991) 169. F. Delduc, E. Ivanov and E. Sokatchev, [*Nucl. Phys.*]{} [**B384**]{} (1992) 334. D. Sorokin and M. Tonin, [*Phys. Lett.*]{} [**326B**]{} (1994) 84. P. Howe, [A note on chiral fermions and heterotic strings]{}, King’s College preprint, 1994. B. E. W. Nilsson, [*Nucl. Phys.*]{} [**B188**]{} (1981) 176. E. Witten, [*Nucl. Phys.*]{} [**B266**]{} (1986) 245. P. Howe and P. Townsend, Phys. Lett. [**259B**]{} (1991) 285. W. Siegel, [*Nucl. Phys.*]{} [**B238**]{} (1984) 307.
[^1]: Recently several attempts have been made to adapt the twistor-like approach to non-heterotic superstrings [@nonhet] or $p$-branes [@p-b], but in our opinion it is still not entirely clear to what extent these formulations are self-consistent.
[^2]: In the case $N=8$ (as well as $N=4,2,1$), due to the triality property of $SO(8)$, the $\theta$’s can equally well be placed in either the $s$ or $c$ spinor representations of $SO(8)$. This is in fact necessary if one wants to fix a light-cone gauge in which the worldsheet $\theta$’s are identified with one half of the target superspace ones (see [@DGHS] for details).
[^3]: Note that for the supersymmetry of [(\[fermcomp\])]{} it is enough to impose the weaker constraint $D_\alpha D_{\{\alpha}\Psi^i D_{\beta\}}\Psi^i = 0$, the first component of which is just eq. [(\[constr3\])]{}. The reason for choosing [(\[constr2\])]{} is the requirement of consistency with the twistor-like formulation of the superstring, as explained earlier.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'To measure the similarity of two documents in the bag-of-words (BoW) vector representation, different term weighting schemes are used to improve the performance of cosine similarity—the most widely used inter-document similarity measure in text mining. In this paper, we identify the shortcomings of the underlying assumptions of term weighting in the inter-document similarity measurement task; and provide a more fit-to-the-purpose alternative. Based on this new assumption, we introduce a new simple but effective similarity measure which does not require explicit term weighting. The proposed measure employs a more nuanced probabilistic approach than those used in term weighting to measure the similarity of two documents w.r.t each term occurring in the two documents. Our empirical comparison with the existing similarity measures using different term weighting schemes shows that the new measure produces (i) better results in the binary BoW representation; and (ii) competitive and more consistent results in the term-frequency-based BoW representation.'
author:
- |
Sunil Aryal$^{1}$ [email protected]\
Kai Ming Ting$^2$ [email protected]\
Takashi Washio$^3$ [email protected]\
Gholamreza Haffari$^4$ [email protected]\
$^1$Deakin University, Australia\
$^2$Federation University, Australia\
$^3$Osaka University, Japan\
$^4$Monash University, Australia
bibliography:
- 'SP.bib'
title: |
A new simple and effective measure for bag-of-word\
inter-document similarity measurement
---
**Keywords:** Inter-document similarity, tf-idf term weighting, cosine similarity, BM25, weighted Jaccard
Introduction {#sec_intro}
============
Pairwise similarity measurements of documents is a fundamental task in many text mining problems such as query-by-example, document classification and clustering.
In the bag-of-words (BoW) [@IntroIR_Salton1986; @IntroIR_Manning2008] vector space model, a document ${\bf x}$ is represented by an $M$-dimensional vector where $M$ is the number of terms in a given dictionary, i.e., ${\bf x}=\langle x_1, x_2, \cdots, x_M\rangle$; and it has the following two representations:
1. Term-frequency-based representation: each $x_i\in \mathbb Z_+$ ($\mathbb Z_+$ is a set of non-negative integers) is the occurrence frequency of term $t_i$ in document ${\bf x}$.
2. Binary representation: each $x_i\in \{0,1\}$ where 0 represents the absence of term $t_i$ in document ${\bf x}$ and 1 represents the presence of $t_i$ in ${\bf x}$.
Because the number of terms in a document is significantly less than that in the dictionary, every document is represented as a sparse BoW vector, where many entries are zero. Because of sparsity, Euclidean distance is not a good similarity measure and the angular distance, aka cosine distance, is a preferred choice of inter-document similarity measure [@IntroIR_Salton1986; @TFIDF_Salton1988].
Because all terms in a document are not equally important to represent its subject, different ‘term weighting’ schemes [@IntroIR_Manning2008; @TFIDF_Salton1988] are used to adjust vector components based on the importance of their terms.
The idea of term weighting was first introduced in the field of Information Retrieval (IR) where the task is to measure the relevance of documents in a given collection $D$ for a given query phrase consisting of a few terms. It is based on the following two assumptions [@IntroIR_Manning2008; @TFIDF_Salton1988; @SimilaritySpace_Zobel1998]:
1. A term is important in a document if it occurs multiple times in the document.
2. A rare term (that occurs in a few documents in the collection) is more important than frequent terms (that occur in many documents in the collection).
The importance of terms in a document are estimated independent of the query. Because a query in the IR task is short and each term generally occurs only once, it is not an issue that the weights are determined independent of the query.
However, it can be counter-productive in the query-by-example task where the query itself is a document, and terms often occur more than once in the query document. For example, to a query document ${\bf q}$, a document ${\bf x}$ having more occurrences of the terms in ${\bf q}$ may not be more similar than ${\bf y}$ which has exactly the same occurrences of terms in ${\bf q}$. Prior research in the BoW inter-document similarity measurement task were focused on developing effective term weighting schemes to improve the task specific performances of existing measures such as cosine and Best Match 25 (BM25) [@TFIDF_Salton1988; @Okapi_Robertson1994; @ProbTFIDF_Joachims1997; @PhDThesis_Singhal1997; @ProbRelFramework_Robertson2009; @IRWeightSenAna_Paltoglou2010; @kNN_LSHTC_Han2012; @ICF_Wang2013]. In contrast, we investigate an alternative similarity measure where an adjustment of vector components using term weighting is not required.
This paper makes the following contributions:
1. Identify the shortcomings of the underlying assumptions of term weighting schemes employed in existing measures; and provide an alternative which is more congruous with the requirements of inter-document similarity measurements.
2. Introduce a new simple but effective inter-document similarity measure which is based on the new assumption and does not require explicit term weighting. It uses a more nuanced probabilistic approach than those used in term weighting to measure the similarity of two documents w.r.t each term occurring in the two documents under measurement.
3. Compare the performance of the new measure with existing measures (which use different term weighting schemes) in the query-by-example task. Our result reveals that the new measure produces (i) better results than existing measures in the binary BoW representation; and (ii) competitive and more consistent results to existing measures in the term-frequency-based BoW representation.
The rest of the paper is organized as follows. Related work in the areas of term weighting and inter-document similarity measures are discussed in Section \[sec\_relatedWork\]. Issues of term weighting in inter-document similarity measurements are discussed Section \[sec\_Issues\]. The proposed new inter-document similarity measure is presented in Section \[sec\_sidf\], followed by empirical results in Section \[sec\_exp\], related discussion in Section \[sec\_discussion\], and the last section presents the conclusions.
The key notations used in this paper are defined in Table \[tbl\_notations\].
--------------------------------------------------------------------- ---------------------------------------------------------------------------------
$D$ A collection of $N$ documents (i.e., $|D|=N$)
${\bf x}$ BoW vector of a document $\langle x_1,x_2,\cdots,x_M\rangle$
$t_i$ The $i^{th}$ term in the dictionary
$n_i$ The number of documents in $D$ having $t_i$
$T_{\bf x}$ The set of terms in ${\bf x}$
$w_i({\bf x})$ The importance or weight of $t_i$ in ${\bf x}$
$tf_i({\bf x})$ Term frequency factor of $t_i$ in ${\bf x}$
$idf(t_i)$ Inverse document frequency factor of $t_i$
$s({\bf x}, {\bf y})$ The similarity of two documents ${\bf x}$ and ${\bf y}$
$dl({\bf x})$ The length of document ${\bf x}$ (i.e., $\sum_{i=1}^M x_i$)
$avgdl$ The average length of documents in $D$
${\bf x} { \;{\mathop{\succ}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@ ${\bf x}$ is more similar to ${\bf q}$ than ${\bf y}$ w.r.t $t_i\in T_{\bf q}$
\hbox{\scriptsize {\bf q}\{i\}}\vss}}}\;} {\bf y}$
${\bf x} { \;{\mathop{=}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@ ${\bf x}$ is equally similar to ${\bf q}$ as ${\bf y}$ w.r.t $t_i\in T_{\bf q}$
\hbox{\scriptsize {\bf q}\{i\}}\vss}}}\;} {\bf y}$
--------------------------------------------------------------------- ---------------------------------------------------------------------------------
: Key notations
\[tbl\_notations\]
Related work {#sec_relatedWork}
============
In this section, we present the pertinent details of term weighting and some widely used existing BoW inter-document similarity measures.
[**Term weighting**]{} {#subsec_tf-idf}
----------------------
In the field of IR, there has been considerable research investigating the effective term weighting scheme. The importance of a term $t_i$ in document ${\bf x}$, $w_i(\bf x)$, is estimated using different variants and combinations of two factors [@IntroIR_Manning2008; @TFIDF_Salton1988; @ProbTFIDF_Joachims1997; @Okapi_Robertson1994; @PhDThesis_Singhal1997; @ProbRelFramework_Robertson2009; @IRWeightSenAna_Paltoglou2010; @kNN_LSHTC_Han2012; @ICF_Wang2013]: (i) document-based factor based on the frequency of $t_i$ in ${\bf x}$, $x_i$; and (ii) collection-based factor based on the number of documents where $t_i$ occurs, $n_i$.
The most widely used term weighting scheme is tf-idf (term frequency - inverse document frequency) where $w_i({\bf x})=tf_i({\bf x}) \times idf(t_i)$ [@IntroIR_Manning2008; @TFIDF_Salton1988]; and it consists of:
1. Document-based factor: $tf_i({\bf x})=1+\log(x_i)$ if $x_i>0$, and 0 otherwise;
2. Collection-based factor: $idf(t_i) = \log\left(\frac{N}{n_i}\right)$.
In the IR task, the idea of tf-idf term weighting is based on the following assumptions [@SimilaritySpace_Zobel1998]:
1. Documents with multiple occurrences of query terms are more relevant than documents with single occurrence of query terms \[the tf assumption\].
2. Documents having rare query terms (occurring in a few documents in the collection) are more relevant to the query than documents having frequent query terms (occurring in many documents in the collection) \[the idf assumption\].
The tf factor considers the importance of $t_i$ in a document. Even though a document with multiple occurrences of a query term is more likely to be relevant to the given query, a document having higher occurrences of one query term is not necessarily more relevant than a document having lower occurrences of two query terms. Therefore, the logarithmic scaling of raw term frequencies is used to reduce the over influence of high frequencies of query terms [@IntroIR_Manning2008; @TFIDF_Salton1988].
The idf factor considers the importance of $t_i$ in the given collection. Basically, it ranks the importance of terms in the given dictionary based on the number of documents where they occur. Terms occurring only in a few documents (i.e., rare terms) are considered to be more important in documents; and they are given more weights than the terms occurring in many documents (i.e., frequent terms) [@IntroIR_Manning2008; @TFIDF_Salton1988].
[**Inter-document similarity measures**]{} {#subsec_exisitingMeasures}
------------------------------------------
Here, we discuss three commonly used measures to estimate the similarity of two document vectors ${\bf x}$ and ${\bf y}$, $s({\bf x}, {\bf y})\to \mathbb R$ where $\mathbb R$ is a real domain.
### Cosine similarity {#subsec_cosine}
The cosine similarity measure with the tf-idf term weighting is the most commonly used inter-document similarity measure. Using term weighted vectors, the cosine similarity of two documents ${\bf x}$ and ${\bf y}$ is estimated as:
$$\label{eqn_cos_sim}
s_{cos}({\bf x}, {\bf y}) = \frac{\sum_{i=1}^M w_i({\bf x})\times w_i({\bf y})}{\sqrt{\sum_{i=1}^M {w_i({\bf x})}^2}\times \sqrt{\sum_{i=1}^M {w_i({\bf y})}^2}}$$
Note that the two terms in the denominator of Eqn \[eqn\_cos\_sim\] are the Euclidean lengths ($\ell_2$-norms) of the term weighted vectors.
It is important to normalize the similarity of documents by their lengths, otherwise it favors longer documents which have higher probability of having more terms in common with the query document over shorter documents [@IntroIR_Salton1986; @IntroIR_Manning2008; @TFIDF_Salton1988; @PivotedLenghtNormalisation_Singhal1996].
### Best Match 25 (BM25) {#subsec_bm25}
BM25 [@ProbRelFramework_Robertson2009; @ProbIR_Jones2000] is the state-of-the-art document ranking measure in IR. It is based on the probabilistic framework of term weighting by [@Okapi_Robertson1994]. [@kNN_LSHTC_Han2012] used BM25 to measure the similarity of two documents ${\bf x}$ and ${\bf y}$ as follows:
$$\begin{multlined}
s_{bm25}({\bf x},{\bf y}) = \displaystyle \sum_{i=1}^M idf_{bm25}(t_i) \times \frac{x_i\cdot (a + 1)}{x_i + a\cdot\left(1-b+b\cdot\frac{dl({\bf x})}{avgdl}\right)} \times \\ \frac{y_i\cdot (a + 1)}{y_i + a\cdot\left(1-b+b\cdot\frac{dl({\bf y})}{avgdl}\right)}
\end{multlined}
\label{eqn_bm25}$$
where $dl({\bf x})=\sum_{i=1}^M x_i$ is the normal length of document ${\bf x}$ (i.e., $\ell_1$-norm of the unweighted vector); $avgdl=\frac{1}{N}\sum_{{\bf x}\in D}dl({\bf x})$ is the average normal document length; $a$ and $b$ are free parameters that control the influence of the term frequencies and document lengths; and $idf_{bm25}(t_i)$ is the idf factor of term $t_i$ defined as follows: $$idf_{bm25}(t_i) = \log \left(\frac{N-n_i+0.5}{n_i+0.5}\right)
\label{eqn_idf_bm25}$$
It uses different variants of tf and idf factors in the similarity measure. The pivoted normal document length [@PivotedLenghtNormalisation_Singhal1996] is used in the tf factor so that longer documents which have higher probability of having more terms in common with the query document are not favored over shorter documents.
### Jaccard similarity {#subsec_jaccard}
The Jaccard similarity [@Jaccard1901] of two documents ${\bf x}$ and ${\bf y}$ is estimated as follows:
$$s_{jac}({\bf x}, {\bf y}) = \frac{|T_{\bf x} \cap T_{\bf y}|}{|T_{\bf x} \cup T_{\bf y}|}
\label{eqn_jaccardSim}$$
where $T_{\bf x} = \{t_i: x_i > 0\}$ is the set of terms in document ${\bf x}$ and $|\cdot|$ is the cardinality of a set.
It only considers the number of terms occurring in both ${\bf x}$ and ${\bf y}$ and does not take into account the importance of terms in documents. The similarity is normalized by the number of distinct terms occurring in either ${\bf x}$ or ${\bf y}$ to take into account that ${\bf x}$ and ${\bf y}$ have higher chance of having terms in common if they have more terms.
The weighted or generalized version of Jaccard similarity [@WeightedJacard_Chierichetti2010] of two documents using term weighted vectors is defined as follows:
$$s_{wjac}({\bf x}, {\bf y}) = \frac{\sum_{i=1}^M\min\{w_i({\bf x}), w_i({\bf y})\}}{\sum_{i=1}^M\max\{w_i({\bf x}), w_i({\bf y})\}}
\label{eqn_wtJaccardSim}$$
The similarity of ${\bf x}$ and ${\bf y}$ w.r.t $t_i\in T_{\bf x} \cap T_{\bf y}$ depends on the importance of $t_i$ in the two documents. The similarity is normalized by the sum of maximum weights of all $t_i\in T_{\bf x} \cup T_{\bf y}$.
Note that the weighted Jaccard similarity of ${\bf x}$ and ${\bf y}$ (Eqn \[eqn\_wtJaccardSim\]) in the binary BoW vector representation without any term weighting is equivalent to the traditional Jaccard similarity (Eqn \[eqn\_jaccardSim\]).
Issues of the tf-idf assumptions in inter-document similarity measurements {#sec_Issues}
==========================================================================
Even though the tf and idf assumptions discussed in Section \[subsec\_tf-idf\] are intuitive in the IR task to rank documents for a given query phrase of a few terms, they can be counter-intuitive in the query-by-example task which requires inter-document similarity measurements to rank documents in $D$ w.r.t a given query document.
In the literature, the query-by-example task is treated as the IR task where query is a document; and the same idea of the tf-idf term weighting is used. However, there is a fundamental difference between the two tasks — unlike in the typical IR task where the query comprises of a few distinct terms (i.e., each term generally occurs only once in the query phrase), the query in the query-by-example task is a long document which often has multiple occurrences of terms.
[**Issue of the tf assumption**]{} {#subsec_tf-issue}
----------------------------------
For a query document ${\bf q}$ with terms $T_{\bf q}$, a document ${\bf x}$ having more occurrences of terms in $T_{\bf q}$ than in ${\bf q}$, may not be more similar to ${\bf q}$ than another document ${\bf y}$, which has similar occurrences of terms in $T_{\bf q}$ as in ${\bf q}$. For example, let’s assume ${\bf x}$ and ${\bf y}$ have frequencies of $t_r\in T_{\bf q}$ as $x_r=10$ and $y_r=1$, respectively. If ${\bf q}$ has $q_r=1$, it is difficult to say that ${\bf x}$ is more similar to ${\bf q}$ than ${\bf y}$ w.r.t $t_r$ $\left(i.e., {\bf x} { \;{\mathop{\succ}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@
\hbox{\scriptsize {\bf q}\{r\}}\vss}}}\;} {\bf y}\right)$, simply because of $x_r>q_r$ (and $q_r=y_r$). It might be the case that ${\bf y}$ is exactly the same document as ${\bf q}$.
Because of the tf-based term weighting factor, ${\bf x}\neq {\bf q}$ can be more similar to ${\bf q}$ than ${\bf q}$ itself using some existing measure such as BM25[^1]. Thus, the tf assumption can be counter-intuitive in inter-document similarity measurements.
[**Issue of the idf assumption**]{} {#subsec_idf-issue}
-----------------------------------
Similarly, ${\bf x}$ having rare terms of $T_{\bf q}$ may not be more similar to ${\bf q}$ than ${\bf y}$ having frequent terms of $T_{\bf q}$. For example, let’s assume the scenario presented in Table \[tbl\_scenario\]:
$n$ $idf(t)$ $\bf x$ $\bf y$ $\bf q$
----------- --------------- ----------- --------- --------- ---------
[$t_g$]{} $\frac{N}{2}$ $\log(2)$ 1 1 1
[$t_h$]{} $N$ 0 1 10 10
: A scenario to demonstrate the issue of the idf assumption. Note that all $\frac{N}{2}$ documents having $t_g$ have its frequency of 1; and all $N$ documents having $t_h$ have its frequency of 1 except ${\bf y}$ where $y_h=10$
\[tbl\_scenario\]
[@ r @ l @ l l @ l @]{} &\
\
tf: & ${\bf y} { \;{\mathop{\succ}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@
\hbox{\scriptsize {\bf {q}}\{i\}}\vss}}}\;} {\bf x}$ & if $y_i > x_i$ & ${\bf y} { \;{\mathop{\succ}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@
\hbox{\scriptsize {\bf q}\{i\}}\vss}}}\;} {\bf x}$ & if $|\{{\bf z} \in D:\alpha(y_i,q_i) \le z_i \le \beta(y_i,q_i)\}|<$\
idf: & ${\bf y} { \;{\mathop{\succ}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@
\hbox{\scriptsize {\bf {q}}\{i,j\}}\vss}}}\;} {\bf x}$ & if $n_i < n_j$ (for $q_i>0, q_j>0$;& & $|\{{\bf w} \in D:\alpha(x_i,q_i) \le w_i \le \beta(x_i,q_i)\}|$\
& & where $\alpha(\cdot,\cdot)=\min(\cdot,\cdot); \beta(\cdot,\cdot)=\max(\cdot,\cdot)$\
\
& ${\bf y} { \;{\mathop{\succ}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@
\hbox{\scriptsize {\bf q}\{r\}}\vss}}}\;} {\bf x}$ & because $|\{{\bf z} \in D:q_r = z_r = y_r\}|<$\
& & $y_r=q_r$) & & $|\{{\bf w} \in D:q_r \le w_r \le x_r\}|$\
\
& ${\bf y} { \;{\mathop{\succ}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@
\hbox{\scriptsize {\bf q}\{g,h\}}\vss}}}\;} {\bf x}$ & because (i) $|\{{\bf z} \in D:q_g = z_g = y_g\}|=$\
& & and & & $|\{{\bf w} \in D:q_g = w_g = x_g\}|$\
& & (ii) $idf(t_h)=0$ (even though & & (ii) $|\{{\bf z} \in D:q_h = z_h = y_h\}|<$\
& & $q_h=y_h=10$ and $x_h=1$) & & $|\{{\bf w} \in D:x_h \le w_h \le q_h\}|$\
\[tbl\_compare\]
Because of $idf(t_h)=0$, the term $t_h$ will be completely ignored. However, $q_h=y_h=10$ is more useful than $q_g=x_g=y_g=1$ because ${\bf y}$ is the only document in $D$ which has as many occurrences of $t_h$ as ${\bf q}$. Even though there is no discrimination between documents w.r.t $t_g$ (all $\frac{N}{2}$ documents with $t_i$ have its frequency of 1), $t_g$ is assigned more weight with $idf(t_g)=\log 2$ than $t_h$ with $idf(t_h)=0$. As a result, ${\bf x}$ and ${\bf y}$ become equally similar to ${\bf q}$ w.r.t $t_g$ and $t_h$ $\left(i.e., {\bf x} { \;{\mathop{=}\limits_{\vbox to 5\ex@{\kern-\tw@\ex@
\hbox{\scriptsize {\bf q}\{g,h\}}\vss}}}\;} {\bf y}\right)$ even though ${\bf y}$ has exactly the same occurrences of $t_g$ and $t_h$ as ${\bf q}$. This example shows that the idf assumption can be counter-intuitive in document similarity measurements.
Our proposal to overcome the issues of tf-idf based term weighting in document similarity measurements {#sec_sidf}
======================================================================================================
The main problem of the tf-idf term weighting in inter-document similarity measurements is that the importance of $t_i$ in ${\bf x}$, $w_i({\bf x})$, is estimated without considering the frequency of $t_i$ in ${\bf q}$, $q_i$. It is not an issue in the IR task because $q_i$ is almost always 1 if $t_i$ occurs in the given query phrase ${\bf q}$. In a query document, $q_i$ can be larger than 1. Therefore, judging the importance of $t_i$ in ${\bf x}$, without considering $q_i$, can be counter-productive in inter-document similarity measurements.
A more fit-to-the-purpose approach would be to evaluate the importance of $t_i$ in ${\bf x}$ by examining the similarity of $x_i$ w.r.t. $q_i$. However, as discussed in Section \[subsec\_idf-issue\], simply having similar occurrences of $t_i$ (i.e., $x_i=q_i$) is not sufficient to consider them to be similar. The similarity shall also consider how rare the frequency of $t_i$ is in the collection.
Putting the above requirements together, the similarity of ${\bf x}$ and ${\bf q}$ w.r.t $t_i$ shall be based on the number of documents in $D$ which have similar occurrence frequencies of $t_i$ as in both ${\bf x}$ and ${\bf q}$. More formally, ${\bf x}$ and ${\bf q}$ are more likely to be similar w.r.t $t_i$ if $|\{{\bf z} \in D: \min(x_i, q_i) \leq z_i \leq \max(x_i, q_i) \}|$ is small. The first part in Table \[tbl\_compare\] compares the underlying assumptions of the tf-idf term weighting (used in existing measures) and the proposed approach called Sp, to be introduced in the next subsection.
This approach addresses the limitations of both the tf and idf assumptions discussed in Section \[sec\_Issues\]. The results of the new approach using the same examples discussed in Sections \[subsec\_tf-issue\] and \[subsec\_idf-issue\] are provided in the second part of Table \[tbl\_compare\]. The comparisons demonstrate that the new approach provides more intuitive outcomes than the tf-idf term weighting.
[**Sp: A new document similarity measure**]{}
---------------------------------------------
Recently, [@mp_Aryal2014; @MpKAIS_Aryal2017] introduced a data dependent measure where the similarity of two data objects $u$ and $v$ depends on the distribution of data between $u$ and $v$ [@mp_Aryal2014; @MpKAIS_Aryal2017]. The intuition is that $u$ and $v$ are more likely to be similar if there are less data between them. For example, two individuals earning 800k and 900k are judged to be more similar by human than two individuals earning 50k and 150k because a lot more people earn in \[50k, 150k\] than \[800k, 900k\].
Using the similar idea, the similarity of two documents ${\bf x}$ and ${\bf y}$ can be estimated as: $$ s_{sp}({\bf x}, {\bf y}) = \frac{1}{|T_{\bf x} \cup T_{\bf y}|} \sum_{t_i \in T_{\bf x} \cap T_{\bf y}} \log\frac{N}{|\{{\bf z} \in D:\min(x_i, y_i)\leq z_i\leq \max(x_i, y_i) \}|}
\label{eqn_sidf}$$
where $\frac{1}{|T_{\bf x} \cup T_{\bf y}|}$ is a normalization term to account for the probability of a term occurring in both ${\bf x}$ and ${\bf y}$. It reduces the bias towards documents having more terms because they have a higher probability of having terms, which also exist in a query document, than documents having fewer terms.
The number of distinct terms is used as a normalization factor (as in the traditional Jaccard similarity) because it is not sensitive to multiple occurrences of the terms in a document which do not occur in the query document. In the IR task, [@PivotedLenghtNormalisation_Singhal1996] have shown that it is more effective than the cosine or normal length normalizations which penalize more to documents having multiple occurrences of the terms which are not in the query phrase.
Sp can be interpreted as a simple probabilistic measure where the similarity of two documents w.r.t $t_i \in T_{\bf x} \cap T_{\bf y}$ is assigned based on the probability of the frequency of $t_i$ to be in \[$\min(x_i, y_i)$, $\max(x_i, y_i)$\], $P(\min(x_i, y_i) \leq \chi_i \leq \max(x_i, y_i))$ (where $\chi_i$ be a random variable representing the occurrence frequency of term $t_i$ in a document). In practice, $P(\min(x_i, y_i) \leq \chi_i \leq \max(x_i, y_i)) = \frac{|\{{\bf z} \in D\mbox{ }:\mbox{ }\min(x_i, y_i)\leq z_i\leq \max(x_i, y_i) \}|}{N}$, which is the inverse of the term used in Eqn \[eqn\_sidf\].
[**Characteristics of Sp**]{}
-----------------------------
The proposed measure has the following characteristics:
1. [*Term weighting is not required*]{}:
Unlike in existing measures such as cosine and BM25, $x_i$ and $y_i$ are not used directly in the similarity measure. They are used just to define $\min(x_i, y_i)$ and $\max(x_i, y_i)$; and the similarity is based on $|\{{\bf z} \in D:\min(x_i, y_i)\leq z_i\leq \max(x_i, y_i)\}|$ which is invariant to the monotonic scaling of frequency values. Hence, Sp does not require additional term weighting to adjust frequency values.\
2. [*Self-similarity is data dependent and the upper bound of similarity*]{}:
Unlike cosine and both variants of Jaccard similarity where the self-similarity of documents is fixed with the maximum of 1, Sp has data dependent self similarity because $s_{sp}({\bf x}, {\bf x})$ depends on the $P(x_i)$ for all $t_i \in T_{\bf x}$. Thus, $s_{sp}({\bf x}, {\bf x})$ and $s_{sp}({\bf y}, {\bf y})$ can be different.
The similarity in Sp is bounded by its self-similarity i.e., $\forall_{{\bf y} \neq {\bf x}} \mbox{ } s_{sp}({\bf x}, {\bf x}) > s_{sp}({\bf x}, {\bf y})$. Although BM25 also has data dependent self-similarity, it is possible to have similarity of different documents to be larger than the self-similarity, i.e., there may be ${\bf y}\neq {\bf x}$ with $s_{bm25}({\bf x}, {\bf y}) > s_{bm25}({\bf x}, {\bf x})$[^2].\
3. [*Relationship with the traditional Jaccard similarity and idf term weighting*]{}:
The formulation of Sp (Eqn \[eqn\_sidf\]) looks similar to the formulation of the traditional Jaccard similarity (Eqn \[eqn\_jaccardSim\]) except that the similarity of ${\bf x}$ and ${\bf y}$ w.r.t $t_i \in T_{\bf x} \cap T_{\bf y}$ is based on $|\{{\bf z} \in D:\min(x_i,y_i)\leq z_i\leq \max(x_i,y_i)\}|$ in Sp, whereas it is the fixed constant of 1 in the traditional Jaccard similarity.
In the binary BoW vector representation, when $t_i \in T_{\bf x} \cap T_{\bf y}$ and $|\{{\bf z} \in D: z_i = 1\}|=n_i$, Sp assigns the similarity of ${\bf x}$ and ${\bf y}$ w.r.t $t_i$ based on $idf(t_i)$, whereas in the traditional Jaccard similarity, it is 1, irrespective of whether $t_i$ is rare or frequent in $D$.
In the term-frequency-based BoW representation, Sp is different from the idf weighting because $idf(t_i)$ is based on $|\{{\bf z} \in D:z_i > 0\}|$, whereas Sp is based on $|\{{\bf z} \in D:\min(x_i,y_i) \leq z_i\leq \max(x_i, y_i)\}|$, where $x_i>0$ and $y_i > 0$ when $t_i \in T_{\bf x} \cap T_{\bf y}$.
[**Computational complexity**]{} {#subsec_sidf_complexity}
--------------------------------
In the term-frequency-based BoW vector representation, it appears that computing $|\{{\bf z} \in D:\min(x_i,y_i) \leq z_i\leq \max(x_i, y_i)\}|$ naively can be expensive as it requires a range search to find the number of documents having the frequencies of $t_i$ between $x_i$ and $y_i$. Since, all $x_i$ are integers (term occurrence frequency counts), it can be computed in constant time by the following simple preprocessing.
Let $m_i$ be the maximum occurrence frequency of term $t_i$ in any document in the given collection $D$. We can maintain a cumulative frequency count array $F_i$ of size $m_i+1$ where $F_i[j]$ contains the number of documents having occurrences of $t_i$ fewer than or equal to $j$.
Using $F_i$, it can be estimated in constant time as $|\{{\bf z} \in D:\min(x_i,y_i) \leq z_i\leq \max(x_i, y_i)\}|= F_i[\max(x_i, y_i)]-F_i[\min(x_i, y_i)-1]$. Note that $\min(x_i, y_i)$ can not be 0 because $|\{{\bf z} \in D:\min(x_i,y_i) \leq z_i\leq \max(x_i, y_i)\}|$ is computed only if $t_i \in T_{\bf x} \cap T_{\bf y}$ (i.e., $x_i>0$ and $y_i>0$) and thus $\min(x_i, y_i) > 0$.
The above preprocessing requires $O(MN)$ time complexity and $O(Mm)$ space complexity, where $m$ is the average maximum frequency of terms.
With the above preprocessing, the runtime of computing the similarity of ${\bf x}$ and ${\bf y}$ using Sp is the same as that of the existing similarity measures which is $O(M)$.
Empirical evaluation {#sec_exp}
====================
In this section, we present the results of experiments conducted to evaluate the task specific performances of Sp, BM25, weighted Jaccard and cosine similarity in the query-by-example task to retrieve documents similar to a given query document. We did experiments with both the term-frequency-based and binary BoW vector representations. We used different combinations of tf and idf based term weighting factors with the weighted Jaccard and cosine similarity measures.
[**Datasets and experimental setup**]{}
---------------------------------------
We used 10 datasets from 6 benchmark document collections. The characteristics of data sets are provided in Table \[tbl\_data\]. NG20, R8, R52 and Webkb are from [@PhDThesis_Ana2007][^3]; and others are from [@Centroid_Han2000][^4].
Name $N$ $M$ $C$ Collection
-------- -------- -------- ----- --------------------------
Fbis 2,463 2,000 17 TREC collection
La1s 3,204 13,195 6 TREC collection
La2s 3,075 12,432 6 TREC collection
New3s 9,558 26,832 44 TREC collection
Ng20 18,821 5,489 20 20 Newsgroup collection
Ohscal 11,162 11,465 10 Ohsumed patients records
R8 7,674 3,497 8 Reuters collection
R52 9,100 7,379 52 Reuters collection
Wap 1,560 8,460 20 Yahoo web pages
Webkb 4,199 1,817 4 University web pages
: Characteristics of datasets ($N$: Number of documents, $M$: Number of terms, $C$: Number of classes).
\[tbl\_data\]
Each dataset was divided into two subsets ${\mathcal D}$ and ${\mathcal Q}$ using a 10-fold cross validation which have 90% and 10% of the documents, respectively. ${\mathcal D}$ was used as a given collection from which similar documents were extracted for each query document in ${\mathcal Q}$. For each ${\bf q}\in {\mathcal Q}$, documents in ${\mathcal D}$ were ranked in descending order of their similarities to ${\bf q}$ using different contending similarity measures. The top $k$ documents were presented as the similar documents to ${\bf q}$.
For performance evaluation, a document was considered to be similar to ${\bf q}$ if they have the same class label. In order to demonstrate the consistency of a measure at different top $k$ retrieved results, we evaluated the precision at the top $k$ retrieved results ($P@k$ in terms of percentage) with $k=1,2,\cdots,25$ and used the mean average precision up to $k=25$. The performance evaluation criterion is: $MAP@25=\frac{\sum_{k=1}^{25} P@k}{25}$.
We repeated the experiment 10 times using each of the 10 folds as ${\mathcal Q}$ and the remaining 9 folds as ${\mathcal D}$. The average $MAP@25$ and standard error over the 10 runs were reported. The average $MAP@25$ of two measures were considered to be significantly different if their confidence intervals based on two standard errors were not overlapping.
The free parameters $a$ and $b$ in BM25 were set to 1.2 and 0.95, respectively, as recommended by [@IRWeightSenAna_Paltoglou2010] and [@ProbIR_Jones2000].
All the experimental setups and similarity measures were implemented in Java using the WEKA platform [@Weka_2009]. All the experiments were conducted in a Linux machine with 2.27 GHz processor and 16 GB memory. We discuss the experimental results with the term-frequency-based and binary BoW vector representations separately in the following two subsections.
[**Results in the term-frequency-based BoW vector representation**]{}
---------------------------------------------------------------------
Here we used two term weighting schemes - tf factor only and tf-idf factors, with weighted Jaccard and cosine. The six contending measures are: [*Sp*]{}, [*BM25*]{}, [*Cos.tf-idf*]{} (cosine with tf-idf), [*Cos.tf*]{} (cosine with tf only), [*WJac.tf-idf*]{} (weighted Jaccard with tf-idf) and [*WJac.tf*]{} (weighted Jaccard with tf only).
The average $MAP@25$ and standard error over 10 runs of six contending measures are provided in Table \[tbl\_tf\_mapl\] and the summarized results in terms of pairwise win-loss-draw counts of contending measures based on the two standard errors significance test over the 10 datasets used in the experiment are provided in Table \[tbl\_tf\_wld\].
[*BM25*]{} [*Cos.tf-idf*]{} [*Cos.tf*]{} [*WJac.tf-idf*]{} [*WJac.tf*]{} [*Sp*]{}
-------- ---------------- ------------------------ ------------------------ ------------------------ ---------------- ------------------------
Fbis 65.12$\pm$0.62 [**68.42$\pm$0.61**]{} [**68.28$\pm$0.58**]{} 66.75$\pm$0.54 [**67.77$\pm$0.51**]{}
La1s 74.41$\pm$0.32 75.97$\pm$0.42 73.08$\pm$0.49 [**79.18$\pm$0.33**]{} 77.54$\pm$0.47
La2s 76.42$\pm$0.49 78.11$\pm$0.42 75.24$\pm$0.44 79.45$\pm$0.37 [**80.89$\pm$0.40**]{}
New3s 67.01$\pm$0.18 68.31$\pm$0.19 69.36$\pm$0.16 68.45$\pm$0.15 68.98$\pm$0.16
Ng20 74.81$\pm$0.24 67.80$\pm$0.28 73.67$\pm$0.23 64.28$\pm$0.24 72.30$\pm$0.20
Ohscal 59.72$\pm$0.22 53.59$\pm$0.21 59.68$\pm$0.21 60.81$\pm$0.20 60.14$\pm$0.19
R52 85.50$\pm$0.20 80.80$\pm$0.27 84.55$\pm$0.21 84.72$\pm$0.19 84.39$\pm$0.22
R8 91.05$\pm$0.14 86.14$\pm$0.22 91.03$\pm$0.18 91.94$\pm$0.21 91.40$\pm$0.17
Wap 19.67$\pm$0.42 65.33$\pm$0.34 61.97$\pm$0.41 [**70.54$\pm$0.46**]{} 65.10$\pm$0.48
Webkb 70.28$\pm$0.23 68.55$\pm$0.24 73.04$\pm$0.27 73.90$\pm$0.31 [**74.91$\pm$0.33**]{}
: Term-frequency-based BoW representation: Average $MAP@25$ and standard error over 10 runs. The best result is underlined and the results equivalent (insignificant difference based on two standard errors) to the best result are bold faced.
\[tbl\_tf\_mapl\]
[*Sp*]{} [*WJac.tf*]{} [*WJac.tf-idf*]{} [*Cos.idf*]{} [*Cos.tf-idf*]{}
------------------- ---------- --------------- ------------------- --------------- ------------------
[*BM25*]{} 8-2-0 8-2-0 6-2-2 7-0-3 5-5-0
[*Cos.tf-idf*]{} 8-1-1 6-2-2 8-1-1 5-4-1
[*Cos.tf*]{} 5-4-1 4-5-1 5-4-1
[*WJac.tf-idf*]{} 3-2-5 3-6-1
[*WJac.tf*]{} 5-2-3
: Term-frequency-based BoW representation: Win-loss-draw counts of measures in columns against those in rows based on the two standard error significance test over 10 runs.
\[tbl\_tf\_wld\]
Table \[tbl\_tf\_mapl\] shows that [*Sp*]{} and [*Cos.tf*]{} produced the best or competitive to the best result in five datasets each; followed by [*WJac.tf-idf*]{} in four; whereas [*Cos.tf-idf*]{}, [*BM25*]{} and [*WJac.tf*]{} were best or competitive to the best measure in only one dataset each.
The first column in Table \[tbl\_tf\_wld\] shows that [*Sp*]{} had more wins than losses over all contending measures. It had one more wins than losses against the closest contenders [*Cos.tf*]{} and [*WJac.tf-idf*]{}.
Of the two cosine measures, [*Cos.tf*]{} had more wins than losses to [*Cos.tf-idf*]{}. This shows that the idf term weighting can be counter-productive with cosine in inter-document similarity measurements. It is mainly due to the cosine normalization which penalizes more to documents having rare terms (with high idf weights) which are not in ${\bf q}$. In comparison to [*BM25*]{}, [*Cos.tf*]{} produced better results with seven wins and no loss; and [*Cos.tf-idf*]{} was competitive with five wins versus five losses.
It is interesting to note that, in the Wap dataset, [*BM25*]{} produced significantly worse result than other contenders. It is due to the idf factor used in [*BM25*]{}. If a term $t_i$ occurs in more than half of the documents in ${\mathcal D}$ (i.e., $n_i>\frac{N}{2}$), $idf_{bm25}(t_i)$ is negative and $t_i$ has negative contribution in the similarity of two documents. When $idf_{bm25}(t_i)$ was replaced by the traditional $idf(t_i)$ in the formulation of [*BM25*]{} (Eqn \[eqn\_bm25\]), it produced $MAP@25$ = 67.04% which was still worse than those of [*Sp*]{} and [*WJac.tf-idf*]{}.
In weighted Jaccard similarity, [*WJac.tf-idf*]{} produced better retrieval results than [*WJac.tf*]{}. It is interesting to note that [*WJac.tf-idf*]{} produced better retrieval results than [*Cos.tf-idf*]{}, [*Cos.tf*]{} and [*BM25*]{}. This could be mainly due to the vector length normalization used in [*BM25*]{} and [*cosine*]{} that penalizes more to documents having higher frequencies of terms which are not in ${\bf q}$.
It is interesting to note that [*Sp*]{} and [*WJac.tf-idf*]{} produced more consistent results than the other contending measures. They did not produce the worst result in any dataset whereas [*WJac.tf*]{} produced the worst result in one dataset (NG20) followed by [*Cos.tf*]{} in two datasets (La1s and La2s); [*BM25*]{} in three datasets (Fbis, New3s and Wap); and [*Cos.tf-idf*]{} in four datasets (Ohscal, R8, R52 and Webkb).
In terms of runtime, all measures had runtime in the same order of magnitude. For example, in the NG20 dataset, the average total runtime of one run (including preprocessing) using [*Sp*]{} took 15935 seconds; whereas [*BM25*]{}, [*Cos.tf-idf*]{} and [*WJac.tf-idf*]{} took 27432, 16089 and 14875 seconds, respectively.
[**Results in the binary BoW vector representation**]{}
-------------------------------------------------------
Here, six contending measures are: [*Sp*]{}, [*BM25*]{}, [*Cos.idf*]{} (cosine with idf), [*Cos*]{} (cosine without idf), [*WJac.idf*]{} (weighted Jaccard with idf) and [*WJac*]{} (weighted Jaccard without idf). Note that [*WJac*]{} which is not using any term weighting is equivalent to the traditional Jaccard similarity defined in Eqn \[eqn\_jaccardSim\].
The average $MAP@25$ and standard error over 10 runs of the six contending measures are provided in Table \[tbl\_bin\_mapl\]; and the summarized results in terms of pairwise win-loss-draw counts of contending measures based on the two standard errors significance test over the 10 datasets used in the experiment are provided in Table \[tbl\_bin\_wld\].
[*BM25*]{} [*Cos.idf*]{} [*Cos*]{} [*WJac.idf*]{} [*WJac*]{} [*Sp*]{}
-------- ------------------------ ---------------- ---------------- ------------------------ ------------------------ ------------------------
Fbis 66.46$\pm$0.50 63.24$\pm$0.56 [**67.17$\pm$0.46**]{} 64.58$\pm$0.52 [**66.94$\pm$0.47**]{}
La1s 74.78$\pm$0.25 76.78$\pm$0.34 75.96$\pm$0.38 [**78.54$\pm$0.34**]{} 77.55$\pm$0.39
La2s 76.71$\pm$0.48 78.48$\pm$0.42 77.55$\pm$0.38 [**80.02$\pm$0.39**]{} 79.12$\pm$0.35
New3s 66.73$\pm$0.16 64.88$\pm$0.15 67.76$\pm$0.15 65.66$\pm$0.16 68.13$\pm$0.16
Ng20 73.80$\pm$0.17 64.12$\pm$0.20 72.26$\pm$0.19 63.07$\pm$0.21 72.61$\pm$0.20
Ohscal [**58.95$\pm$0.19**]{} 55.06$\pm$0.17 58.56$\pm$0.18 58.66$\pm$0.21 58.45$\pm$0.17
R52 [**83.87$\pm$0.24**]{} 79.01$\pm$0.28 83.23$\pm$0.22 83.36$\pm$0.21 [**83.80$\pm$0.22**]{}
R8 90.54$\pm$0.16 86.03$\pm$0.19 90.24$\pm$0.17 91.10$\pm$0.20 90.92$\pm$0.18
Wap 16.47$\pm$0.34 66.97$\pm$0.47 59.16$\pm$0.44 65.09$\pm$0.48 [**70.02$\pm$0.53**]{}
Webkb 73.29$\pm$0.39 70.86$\pm$0.23 74.19$\pm$0.37 [**75.59$\pm$0.29**]{} 74.97$\pm$0.35
\[tbl\_bin\_mapl\]
[*Sp*]{} [*WJac*]{} [*WJac.idf*]{} [*Cos*]{} [*Cos.idf*]{}
---------------- ---------- ------------ ---------------- ----------- ---------------
[*BM25*]{} 5-2-3 5-5-0 4-3-2 4-4-2 3-7-0
[*Cos.idf*]{} 8-1-1 5-4-1 8-1-1 4-6-0
[*Cos*]{} 7-2-1 5-3-2 6-3-1
[*WJac.idf*]{} 5-0-5 2-6-2
[*WJac*]{} 8-0-2
: Binary BoW representation: Win-loss-draw counts of measures in columns against those in rows based on the two standard error significance test over 10 runs.
\[tbl\_bin\_wld\]
Table \[tbl\_bin\_mapl\] shows that [*Sp*]{} produced the best or competitive to the best result in six datasets; followed by [*BM25*]{} in five; [*WJac.idf*]{} in four; [*Cos*]{} in two; and [*WJac*]{} in one dataset only. [*Cos.idf*]{} did not produce competitive result to the best performing measure in any dataset.
In terms of pairwise win-loss-draw counts as shown in the first column in Table \[tbl\_bin\_wld\], [*Sp*]{} had many more wins than losses against all other contending measures.
It is interesting to note that [*BM25*]{}, [*Cos.idf*]{} and [*Cos*]{} using the binary BoW representation produced better retrieval results than their respective counterparts using the term-frequency-based BoW representation in some datasets. For example: (i) [*BM25*]{} in Fbis, New3s and Webkb; (ii) [*Cos.idf*]{} in La1s, Ohscal, Wap and Webkb; and (iii) [*Cos*]{} in La1s, La2s and Webkb. In contrast, [*WJac.idf*]{}, [*WJac*]{} and [*Sp*]{} using binary BoW vectors did not produce better retrieval results than their respective counterparts using term-frequency-based BoW vectors.
Like in the term-frequency-based BoW representation, all measures had runtimes in the same order of magnitude.
Discussion {#sec_discussion}
==========
Even though some studies have used different variants of tf and idf term weighting factors with the most widely used cosine similarity, the tf and idf factors discussed in Section \[subsec\_tf-idf\] have been shown to be the most consistent in the IR task [@PhDThesis_Singhal1997].
For the tf factor, instead of using the logarithmic scaling of $x_i$, some researchers have used other scaling approaches such as augmented $\left(0.5+0.5\times\frac{x_i}{\max(x_1,x_2,\cdots,x_M)}\right)$ [@TFIDF_Salton1988] and Okapi $\left(\frac{x_i}{2+x_i}\right)$ [@Okapi_Robertson1994]. Similarly, for the idf factor, instead of using $\frac{N}{n_i}$, some researchers have used the probabilistic idf factor based on $\frac{N-n_i}{n_i}$ [@Okapi_Robertson1994; @PhDThesis_Singhal1997]. Note that BM25 (Eqn \[eqn\_bm25\]) uses tf factor similar to Okapi and idf factor similar to the probabilistic idf factor [@ProbRelFramework_Robertson2009].
In the supervised text mining task of document classification, different approaches utilizing class information are proposed to estimate the collection-based term weighting factors [@ICF_Wang2013; @SupervisedTW_Debole2003; @TermWeigthingTC_Lan2009]. Inverse category frequency (icf) [@ICF_Wang2013] has been shown to produce better classification result than the traditional idf factor with the cosine similarity measure. It considers the distribution of a term among classes rather than among documents in the given collection. The intuition behind icf is that the fewer classes a term $t_i$ occurs in, the more discriminating power the term $t_i$ contributes to classification [@ICF_Wang2013]. If $C$ and $c_i$ are the total number of classes and the number of classes in which $t_i$ occurs at least once in at least one document, then the icf factor is estimated as: $icf(t_i)=\log \left(1+\frac{C}{c_i}\right)$.
We have evaluated the performance of Sp in the kNN document classification task with existing measures using the supervised term weighting scheme of icf [@ICF_Wang2013]. Sp produced either better or competitive classification results with existing measures using supervised or unsupervised term weighting in the 5NN classification task. The classification results are provided in the Appendix.
Even though the weighted Jaccard similarity has been used in other application domains [@WeightedJacard_Chierichetti2010], it is not widely used in the literature to measure similarities of BoW documents. Our experimental results in Section \[sec\_exp\] show that the weighted Jaccard similarity with tf-idf term weighting scheme can be an effective alternative of cosine and BM25 in inter-document similarity measurements.
Sp has superior performance over all contenders in the binary BoW vector representation. It can be very useful in application domains such as legal and medical where the exact term frequency information may not available due to privacy issue because it is possible to infer information in a document from its term frequencies [@bigram_zhu2008].
Concluding remarks
==================
For the purpose of inter-document similarity measurements task, we identify the limitations of the underlying assumptions of the most widely used tf-idf term weighting scheme employed in existing measures such as cosine and BM25; and provide an alternative which is more intuitive in this task.
Based on the new assumption, we introduce a new simple but effective inter-document similarity measure called Sp.
Our empirical evaluation in the query-by-example task shows that:
1. Sp produces better or at least competitive results to the existing similarity measures with the state-of-the-art term weighting schemes in the term-frequency-based BoW representations. Sp produces more consistent results than the existing measures across different datasets.
2. Sp produces better results than the existing similarity measures with or without idf term weighting in the the case of binary BoW representation.
When cosine and BM25 are employed, our result shows that it is important to use an appropriate BoW vector representation (binary or term-frequency-based) and also an appropriate term weighting scheme. Using inappropriate representation and term weighting scheme can result in poor performance.
In contrast, using Sp, users do not have to worry about applying any additional term weighting to measure the similarity of two documents and still get better or competitive results in comparison to the best results obtained by cosine or BM25.
Acknowledgement {#acknowledgement .unnumbered}
===============
The preliminary version of this paper was published in the Proceedings of the 11th Asia Information Retrieval Societies Conference 2015 [@tfidf_Aryal2015].
Appendix A: kNN classification results {#appendix-a-knn-classification-results .unnumbered}
======================================
In order to predict a class label for a test document ${\bf q}$, its $k$ nearest neighbour (or most similar) documents were searched in the given labelled training set of documents using a contending similarity measure and the majority class among the $k$NNs was predicted as the class label for ${\bf q}$.
All classification experiments were conducted using a 10-fold cross validation (10 runs with each one out of the 10 folds as the test set and the remaining 9 folds as the training set). The average classification accuracy and standard error over a 10-fold cross validation were reported. All collection-based term weighting factors (idf and icf) were computed from the training set and used in both the training and test documents. The parameter $k$ was set to a commonly used value of 5 (i.e., 5NN classification was used).
We discuss the 5NN classification results with the term-frequency-based and binary BoW vector representations separately in the following two subsections.
[**Term-frequency-based BoW vector representation**]{} {#term-frequency-based-bow-vector-representation .unnumbered}
------------------------------------------------------
We used term weighting based on tf only, tf-idf and tf-icf with weighted Jaccard and cosine resulting in eight contending measures: [*Sp*]{}, [*BM25*]{}, [*Cos.tf-icf*]{}, [*Cos.tf-idf*]{}, [*Cos.tf*]{}, [*WJac.tf-icf*]{}, [*WJac.tf-idf*]{} and [*WJac.tf*]{}.
The average classification accuracies and standard errors over a 10-fold cross validation of the eight contending measures are provided in Table \[tbl\_tf\_acc\] and the summarized results in terms of pairwise win-loss-draw counts of contending measures based on the two standard errors significance test in the 10 datasets used in the experiment are provided in Table \[tbl\_tf\_acc\_wld\].
[*BM25.tf*]{} [*Cos.tf.icf*]{} [*Cos.tf.idf*]{} [*Cos.tf*]{} [*WJac.tf.icf*]{} [*WJac.tf.idf*]{} [*WJac.tf*]{} [*Sp*]{}
-------- ---------------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------
Fbis 76.98$\pm$1.04 [**79.33$\pm$1.05**]{} [**80.23$\pm$0.78**]{} [**79.29$\pm$0.95**]{} [**79.54$\pm$0.91**]{} [**79.21$\pm$0.84**]{} [**79.21$\pm$0.80**]{}
La1s 83.68$\pm$0.58 83.43$\pm$0.84 86.70$\pm$0.66 82.30$\pm$0.80 87.30$\pm$0.80 87.05$\pm$0.70 [**88.48$\pm$0.46**]{}
La2s 86.28$\pm$0.50 84.52$\pm$0.47 87.93$\pm$0.77 84.23$\pm$0.61 88.46$\pm$0.63 88.03$\pm$0.56 [**89.59$\pm$0.48**]{}
New3s 79.31$\pm$0.30 79.54$\pm$0.34 78.99$\pm$0.34 [**80.90$\pm$0.30**]{} [**80.86$\pm$0.33**]{} 80.60$\pm$0.36 80.55$\pm$0.36
Ng20 87.57$\pm$0.22 86.92$\pm$0.22 84.74$\pm$0.34 86.28$\pm$0.27 87.41$\pm$0.25 83.05$\pm$0.28 86.62$\pm$0.17
Ohscal 72.63$\pm$0.29 72.04$\pm$0.43 66.95$\pm$0.45 [**74.25$\pm$0.46**]{} 72.36$\pm$0.21 [**74.22$\pm$0.33**]{} 73.19$\pm$0.34
R52 91.20$\pm$0.28 87.72$\pm$0.54 [**92.18$\pm$0.18**]{} 91.69$\pm$0.22 91.17$\pm$0.22 90.63$\pm$0.21 90.94$\pm$0.25
R8 95.19$\pm$0.21 [**95.34$\pm$0.17**]{} 90.80$\pm$0.25 [**95.39$\pm$0.20**]{} 94.98$\pm$0.23 [**95.27$\pm$0.31**]{} 95.28$\pm$0.27
Wap 17.76$\pm$0.79 75.90$\pm$0.46 76.92$\pm$0.76 72.44$\pm$0.58 80.70$\pm$0.80 [**82.31$\pm$0.92**]{} 76.22$\pm$0.58
Webkb 81.16$\pm$0.38 81.86$\pm$0.48 77.92$\pm$0.43 81.58$\pm$0.57 [**84.14$\pm$0.42**]{} 83.33$\pm$0.61 [**84.33$\pm$0.53**]{}
\[tbl\_tf\_acc\]
[*BM25*]{} [*Cos.icf*]{} [*Cos.idf*]{} [*Cos*]{} [*WJac.icf*]{} [*WJac.idf*]{} [*WJac*]{} [*Sp*]{}
-------- ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------ ------------------------
Fbis [**79.29$\pm$0.65**]{} [**79.62$\pm$0.99**]{} 77.75$\pm$0.95 [**78.20$\pm$0.88**]{} [**79.82$\pm$1.08**]{} [**78.28$\pm$0.81**]{} [**79.05$\pm$0.82**]{}
La1s 84.27$\pm$0.63 [**87.45$\pm$0.79**]{} [**87.70$\pm$0.63**]{} 85.89$\pm$0.72 [**88.05$\pm$0.73**]{} 87.55$\pm$0.62 [**88.67$\pm$0.49**]{}
La2s 86.08$\pm$0.58 87.48$\pm$0.49 [**89.43$\pm$0.42**]{} 86.67$\pm$0.50 88.52$\pm$0.68 [**89.99$\pm$0.43**]{} 88.00$\pm$0.60
New3s 79.10$\pm$0.36 78.31$\pm$0.41 78.19$\pm$0.35 79.79$\pm$0.36 [**80.03$\pm$0.41**]{} 78.74$\pm$0.29 [**80.15$\pm$0.40**]{}
Ng20 87.19$\pm$0.20 87.25$\pm$0.19 82.80$\pm$0.20 85.64$\pm$0.12 86.61$\pm$0.18 82.16$\pm$0.20 86.84$\pm$0.24
Ohscal 72.02$\pm$0.32 72.36$\pm$0.31 68.54$\pm$0.32 72.89$\pm$0.24 72.36$\pm$0.32 72.89$\pm$0.29 72.79$\pm$0.33
R52 89.74$\pm$0.38 86.14$\pm$0.43 90.21$\pm$0.27 90.51$\pm$0.25 89.81$\pm$0.24 89.75$\pm$0.17 [**90.80$\pm$0.19**]{}
R8 [**94.80$\pm$0.21**]{} [**95.05$\pm$0.13**]{} 90.98$\pm$0.44 [**94.99$\pm$0.23**]{} 94.54$\pm$0.30 [**94.86$\pm$0.17**]{} [**95.05$\pm$0.31**]{}
Wap 15.51$\pm$0.69 76.86$\pm$0.65 78.27$\pm$1.01 69.68$\pm$0.82 80.00$\pm$0.56 76.28$\pm$0.62 [**81.60$\pm$0.81**]{}
Webkb [**83.97$\pm$0.49**]{} [**84.68$\pm$0.54**]{} 81.45$\pm$0.60 [**84.26$\pm$0.49**]{} [**84.16$\pm$0.47**]{} [**84.73$\pm$0.42**]{} [**84.71$\pm$0.53**]{}
\[tbl\_bin\_acc\]
[*Sp*]{} [*WJac.tf*]{} [*WJac.tf-idf*]{} [*WJac.tf-icf*]{} [*Cos .tf*]{} [*Cos.tf-idf*]{} [*Cos.tf-icf*]{}
------------------- ---------- --------------- ------------------- ------------------- --------------- ------------------ ------------------
[*BM25*]{} 6-2-2 7-2-1 6-2-2 7-2-1 5-2-3 4-5-1 2-3-5
[*Cos.tf-icf*]{} 6-1-3 5-2-3 5-0-5 6-1-2 4-2-4 2-5-3
[*Cos.tf-idf*]{} 8-0-2 5-1-4 9-0-1 6-1-3 5-4-1
[*Cos.tf*]{} 5-4-1 4-3-3 5-3-2 5-1-4
[*WJac.tf-icf*]{} 2-2-6 0-3-7 3-2-5
[*WJac.tf-idf*]{} 1-1-8 2-5-3
[*WJac.tf*]{} 4-1-5
: Term-frequency-based BoW representation: Win-loss-draw counts of measures in columns against those in rows based on the two standard errors significance test over a 10-fold cross validation of 5NN classification.
\[tbl\_tf\_acc\_wld\]
[*Sp*]{} [*WJac*]{} [*WJac.idf*]{} [*WJac.icf*]{} [*Cos*]{} [*Cos.idf*]{} [*Cos.icf*]{}
---------------- ---------- ------------ ---------------- ---------------- ----------- --------------- ---------------
[*BM25*]{} 4-1-5 4-3-3 3-2-5 4-3-3 3-3-4 3-6-1 3-3-4
[*Cos.icf*]{} 4-0-6 0-1-8 3-2-5 3-1-6 0-4-6 1-5-4
[*Cos.idf*]{} 6-0-4 4-3-3 7-1-2 7-1-2 4-4-2
[*Cos*]{} 6-0-4 3-2-5 5-0-5 6-0-4
[*WJac.icf*]{} 3-1-6 0-5-5 3-3-4
[*WJac.idf*]{} 1-0-9 0-4-6
[*WJac*]{} 6-0-4
: Binary BoW representation: Win-loss-draw counts of measures in columns against those in rows based on the two standard errors significance test over a 10-fold cross validation of 5NN classification.
\[tbl\_bin\_acc\_wld\]
The 5NN classification accuracies in Table \[tbl\_tf\_mapl\] show that [*Sp*]{}, [*WJac.tf-idf*]{}, [*WJac.tf-icf*]{} and [*Cos.tf*]{} produced the best or competitive to the best result in five datasets each followed by [*WJac.tf*]{} in four; [*Cos.tf-icf*]{} and [*BM25*]{} in two datasets each; and [*Cos.tf-idf*]{} in one dataset only.
The pairwise win-loss-draw counts of [*Sp*]{} in the first column of Table \[tbl\_tf\_acc\_wld\] shows that it had more wins than losses over all contending measures except [*Wjac.tf-idf*]{} and [*Wjac.tf-icf*]{} where it had competitive results with the same number of wins and losses.
It is interesting to note that [*Sp*]{} and all three variants of weighted Jaccard similarity produced better classification results than all three variants of cosine and BM25. Like in the similar document retrieval task discussed in Section \[sec\_exp\], BM25 produced the worst classification accuracy in the Wap dataset because of $idf_{bm25}(t_i)$. The classification accuracy was increased to 79.42% when $idf_{bm25}(t_i)$ was replaced by the traditional idf $idf(t_i)$.
The supervised term weighting using icf (tf-icf) did not always produce better classification results than the traditional tf-idf based term weighting with both cosine and weighted Jaccard. It had five wins and two losses with cosine whereas it had two wins and three losses with weighted Jaccard.
[**Binary BoW vector representation**]{} {#binary-bow-vector-representation .unnumbered}
----------------------------------------
We used weighted Jaccard and cosine similarities with and without idf and icf weighting resulting in eight contending measures: [*Sp*]{}, [*BM25*]{}, [*Cos.idf*]{}, [*Cos.icf*]{}, [*Cos*]{}, [*WJac.idf*]{}, [*WJac.icf*]{} and [*WJac*]{}.
The average classification accuracies and standard errors over a 10-fold cross validation of the eight contending measures are provided in Table \[tbl\_bin\_acc\] and the summarized results in terms of pairwise win-loss-draw counts of contending measures based on the two standard errors significance test in the 10 datasets used in the experiment are provided in Table \[tbl\_bin\_acc\_wld\].
The 5NN classification accuracies in Table \[tbl\_bin\_acc\] show that [*Sp*]{} produced the best or competitive to the best result in eight datasets. The closest contenders [*BM25*]{} and [*WJac.idf*]{} produced the best or competitive to the best result in six datasets each followed by [*WJac.icf*]{} in five; [*Cos.icf*]{} in four; [*WJac*]{} and [*Cos*]{} in three datasets each; and [*Cos.idf*]{} in two datasets only.
In terms of pairwise win:loss:draw counts as shown in the first column in Table \[tbl\_bin\_acc\_wld\], [*Sp*]{} had more wins than losses against all other contending measures. It had one win and no loss against [*WJac.idf*]{} and three wins and one loss against [*WJac.icf*]{}.
Like in the term-frequency-based BoW representation, the supervised weighting scheme based on icf did not always produce better classification results than the traditional idf based term weighting scheme with both cosine and weighted Jaccard in the binary BoW vector presentation as well. It had five wins and one loss with cosine whereas it had three wins and three losses with the weighted Jaccard.
It is interesting to note that [*BM25*]{}, [*Cos.icf*]{}, [*Cos.idf*]{} and [*Cos*]{} which are using the binary BoW vector representation produced better classification accuracies than their respective counterparts using the term-frequency-based BoW representation in some datasets; e.g., [*BM25*]{} was better in three datasets (Fbis, New3s, WebKb); [*Cos.icf*]{} and [*Cos*]{} in three datasets (La1s, La2s, Webkb); and [*Cos.idf*]{} in two datasets (La2s, Webkb). However, all three variants of weighted Jaccard and [*Sp*]{} with the term-frequency-based BoW representation produced either better or competitive results with the binary BoW representation.
\[lastpage\]
[^1]: It depends on the lengths of documents and parameters $a$ and $b$.
[^2]: It depends on the lengths of ${\bf x}$ and ${\bf y}$ and parameters $a$ and $b$.
[^3]: BoW vectors available at: http://web.ist.utl.pt/acardoso/datasets/
[^4]: BoW vectors available at: http://www.cs.waikato.ac.nz/ml/weka/datasets.html
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A completely reducible subcomplex of a spherical building is a spherical building.'
author:
- Linus Kramer
date: April 15 2008
title: '[**Completely reducible subcomplexes of spherical buildings**]{}'
---
By a *sphere* we mean metric space isometric to the unit sphere ${\mathbb{S}}^m\subseteq{\mathbb{R}}^{m+1}$, endowed with the spherical metric $d$. The distance of $u,v\in{\mathbb{S}}^m\subseteq{\mathbb{R}}^{m+1}$ is given by $\cos(d(u,v))=u\cdot v$ (standard inner product). Recall that a geodesic metric space $Z$ is CAT$(1)$ if no geodesic triangle of perimeter $<2\pi$ in $Z$ is thicker than its comparison triangle in the $2$-sphere ${\mathbb{S}}^2$. It follows that any two points at distance $<\pi$ can be joined by a unique geodesic segment (an isometric copy of a closed interval). A subset $Y$ of a CAT(1)-space $Z$ is *convex* if the following holds: for all $x,y\in Y$ with $d(x,y)<\pi$, the geodesic segment $[x,y]$ is contained in $Y$. It is clear from the definition that arbitrary intersections of convex subsets are convex (and CAT$(1)$).
Convex sets in Coxeter complexes
================================
Let $\Sigma$ be an $n$-dimensional spherical Coxeter complex, let $\bar\Sigma$ be a simplicial complex which refines the triangulation of $\Sigma$ and which is invariant under the Coxeter group $W$ and $\pm id$. Examples of such triangulations are $\Sigma$ itself and its barycentric subdivisions. In the geometric realization, the simplices are assumed to be spherical. The *span* of a subset of a sphere is the smallest subsphere containing the set.
We assume now that $A\subseteq\bar\Sigma$ is an $m$-dimensional subcomplex whose geometric realization $|A|$ is convex.
Let $a\in A$ be an $m$-simplex. Then $|A|\subseteq \mathrm{span}|a|$.
*Proof.*
Assume this is false. Let $u\in|A|\setminus\mathrm{span}|a|$. Then $-u\not\in|a|$ and $Y=\bigcup\{[u,v]\mid v\in|a|\}$ is contained in $|A|$. But $Y$ is a cone over $|a|$ and in particular $m+1$-dimensional, a contradiction. [ plus 1filll $\Box$]{}
We choose an $m$-simplex $a\in A$ and put $$S=\mathrm{span}|a|\cap|\bar\Sigma|;$$ this is an $m$-sphere containing $|A|$. Recall that an $m$-dimensional simplicial complex is called *pure* if every simplex is contained in some $m$-simplex.
$A$ is pure.
*Proof.*
It suffices to consider the case $m\geq 1$. Let $a\in A$ be an $m$-simplex, and assume that $b\in A$ is a lonely simplex of maximal dimension $\ell<m$. Then $int(-b)$ is disjoint from $int(a)$. Let $v$ be an interior point of $a$ and $u$ an interior point of $b$ and consider the geodesic segment $[u,v]\subseteq|A|$. If $b$ is a point, the existence of the geodesic shows that $b$ is contained in some higher dimensional simplex, a contradiction. If $\ell\geq 1$, then $[u,v]$ intersects $int(b)$ in more than two points (because $b$ is lonely), so $v$ is in the span of $b$. This contradicts $dim(a)>dim(b)$. [ plus 1filll $\Box$]{}
\[OppLem\] If there exists an $m$-simplex $a\in A$ with $-a\in A$, then $|A|=S$.
*Proof.*
Then any point in $S$ lies on some geodesic of length $<\pi$ joining a point in $|a|$ with a point in $|-a|$. [ plus 1filll $\Box$]{}
Topologically, the convex set $|A|$ is either an $m$-sphere or homeomorphic to a closed $m$-ball. For $m\geq 2$, these spaces are strongly connected (i.e. they cannot be separated by $m-2$-dimensional subcomplexes [@Alex]). It follows that $A$ is a chamber complex, i.e. the chamber graph $C(A)$ (whose vertices are the $m$-simplices and whose edges are the $m-1$ simplices) is connected [@Alex]. If $m=1$, then $|A|$ is a connected graph and hence strongly connected.
If $m\geq 1$, then $A$ is a chamber complex. [ plus 1filll $\Box$]{}
Results by Balser-Lytchak and Serre
===================================
We now assume that $X$ is a simplicial spherical building modeled on the Coxeter complex $\Sigma$. By means of the coordinate charts for the apartments we obtain a metric simplicial complex $\bar X$ refining $X$, which is modeled locally on $\bar\Sigma$. In this refined complex $\bar X$, we call two simplices $a,b$ *opposite* if $a=-b$ in some (whence any) apartment containing both. We let $opp(a)$ denote the collection of all simplices in $\bar X$ opposite $a$. The geometric realization $|\bar X|$ is CAT$(1)$. Furthermore, any geodesic arc is contained in some apartment.
We assume that $A\subseteq\bar X$ is an $m$-dimensional subcomplex and that $|A|$ is convex. For any two simplices $a,b\in A$, we can find an apartment $\bar\Sigma$ containing $a$ and $b$. The intersection $|A|\cap|\bar\Sigma|$ is then convex, so we may apply the results of the previous section to it. We note also that $|A|$ is CAT$(1)$.
$A$ is a pure chamber complex.
*Proof.*
Let $a\in A$ be an $m$-simplex and let $b\in A$ be any simplex. Let $\bar\Sigma$ be an apartment containing $a$ and $b$. Since $|\bar\Sigma|\cap|A|$ is $m$-dimensional and convex, we find an $m$-simplex $c\in A\cap\bar\Sigma$ containing $b$. Similarly we see that $A$ is a chamber complex.
The next results are due to Serre [@Serre] and Balser-Lytchak [@BL1; @BL2].
If there is a simplex $a\in A$ with $opp(a)\cap A={\varnothing}$, then $|A|$ is contractible.
*Proof.*
We choose $u$ in the interior of $a$. Then $d(u,v)<\pi$ for all $v\in|A|$, so $|A|$ can be contracted to $u$ along these unique geodesics. [ plus 1filll $\Box$]{}
\[SerreProp\] If there is an $m$-simplex $a$ in $A$ with $opp(a)\cap A\neq{\varnothing}$, then every simplex $a\in A$ has an opposite in $A$.
*Proof.*
Let $a,b\in A$ be opposite $m$-simplices, let $\bar\Sigma$ be an apartment containing both and let $S\subseteq|\bar\Sigma|$ denote the sphere spanned by $a,b$. Then $S\subseteq|A|$. Let $c$ be any $m$-simplex in $A$. If $c$ is not opposite $a$, we find interior points $u,v$ of $c,a$ with $d(u,v)<\pi$. The geodesic arc $[u,v]$ has a unique extension in $S$. Along this extension, let $w$ be the point with $d(u,w)=\pi$ and let $c'$ be the smallest simplex containing $w$. Then $c'$ is opposite $c$.
Thus every $m$-simplex in $A$ has an opposite, and therefore every simplex in $A$ has an opposite. [ plus 1filll $\Box$]{}
In this situation where every simplex has an opposite, $A$ is called $A$ *completely reducible*. If every simplex of a fixed dimension $k\leq m$ has an opposite in $A$, then clearly every vertex in $A$ has an opposite. Serre [@Serre] observed that the latter already characterizes complete reducibility.
\[SerreProp2\] If every vertex in $A$ has an opposite, then $A$ is completely reducible.
*Proof.*
We show inductively that $A$ contains a pair of opposite $k$-simplices, for $0\leq k\leq m$. This holds for $k=0$ by assumption, and we are done if $k=m$ by \[SerreProp\]. So we assume that $0\leq k<m$.
Let $a,a'$ be opposite $k$-simplices in $A$ and let $b\in A$ be a vertex which generates together with $a$ a $k+1$-simplex (recall that $A$ is pure, so such a vertex exists). We fix an apartment $\bar\Sigma$ containing $a$, $b$ and $a'$. The geodesic convex closure $Y$ of $b$ and $|a|\cup|a'|$ in the sphere $|\bar\Sigma|$ is a $k+1$-dimensional hemisphere (and is contained in $|A|$). Let $b'\in A$ be a vertex opposite $b$. A small ${\varepsilon}$-ball in $Y$ about $b$ generates together with $b'$ a $k+1$-sphere $S\subseteq|A|$. Because $\dim S=k+1$, there exists a point $u\in S$ such that the minimal simplex $c$ containing $u$ has dimension at least $k+1$. Let $u'$ be the opposite of $u$ in $S$, and $c'$ the minimal simplex containing $u'$. Then $c,c'$ is a pair of opposite simplices in $A$ of dimensions at least $k+1$. [ plus 1filll $\Box$]{}
Completely reducible subcomplexes are buildings
===============================================
We assume that $A$ is $m$-dimensional, convex and completely reducible. If $m=0$, then $A$ consists of a set of vertices which have pairwise distance $\pi$. This set is, trivially, a $0$-dimensional spherical building. So we assume now that $1\leq m\leq n$. Two opposite $m$-simplices $a,b\in A$ determine an $m$-sphere $S(a,b)$ which we call a *Levi sphere*.
If $a,b\in A$ are $m$-simplices, then there is a Levi sphere containing $a$ and $b$.
*Proof.*
This is true if $b$ is opposite $a$. If $b$ is not opposite $a$, we choose interior points $u\in int(a)$ and $v\in int(b)$, and a simplex $c\in A$ opposite $b$. The geodesic $[u,v]$ has a unique continuation $[v,w]$ in the Levi sphere $S(b,c)$, such that $d(u,w)=\pi$. Let $\bar\Sigma$ be an apartment containing the geodesic arc $[u,v]\cup[v,w]$ and let $d$ be the smallest simplex in $\bar\Sigma$ containing $w$. Then $d$ is in $A$ and opposite $a$, so the there is a Levi sphere $S(a,d)$ containing $[u,v]\cup[v,w]$. Since $b$ is the smallest simplex containing $v$, it follows that $b\in S(a,d)$. [ plus 1filll $\Box$]{}
Since $A$ is pure, we have the following consequence.
Any two simplices $a,b\in A$ are in some Levi sphere. [ plus 1filll $\Box$]{}
We call an $m-1$-simplex $b\in A$ *singular* if it is contained in three different $m$-simplices. The following idea is taken from Caprace [@Cap]. Two $m$-simplices are *t-equivalent* if there is a path between them in the dual graph which never crosses a singular $m-1$-simplex. The t-class of $a$ is contained in all Levi spheres containing $a$.
Let $b$ be a singular $m-1$-simplex. Let $S$ be a Levi sphere containing $b$ and let $H\subseteq S$ denote the great $m-1$-sphere spanned by $|b|$. Then $H$ is the union of singular $m-1$-simplices.
*Proof.*
Let $a$ be an $m$-simplex containing $b$ which is not in $S$ and let $-b$ denote the opposite of $b$ in $S$. Let $S'$ be a Levi sphere containing $a$ and $-b$ and consider the convex hull $Y$ of $|a|\cup|-b|$ in $S'$. Then $Y$ is an $m$-hemisphere. The intersection $Y\cap S$ is convex, contains the great sphere $H$, and is different from $Y$, so $Y\cap S=H$. [ plus 1filll $\Box$]{}
We call $H$ a *singular great sphere*. Along singular great spheres, we can do ’surgery’:
Let $S,H,Y$ be as in the previous lemma. Let $Z\subseteq S$ be a hemisphere with boundary $H$. Then $Z\cup Y$ is a Levi sphere.
*Proof.*
We use the same notation as in the previous lemma. Let $c\subseteq Z$ be an $m$-simplex containing $-b$, then $|c|\cup H$ generates $Z$. Let $S'$ be a Levi sphere containing $c$ and $a$. Then $Z\cup Y\subseteq S'$ and $Z\cup Y=H$, whence $S'=Z\cup Y$. [ plus 1filll $\Box$]{}
Let $S$ be a Levi sphere and let $H,H'\subseteq S$ be singular great spheres. Let $s$ denote the metric reflection of $S$ along $H$. Then $s(H')$ is again a singular great sphere.
*Proof.*
We use the notation of the previous lemma. Let $b'$ be a singular $m-1$-simplex in $H'\cap Z$. Let $-b'$ denote its opposite in the Levi sphere $S'=Z\cup Y$. We note that the interior of $b$ is disjoint from $S$. Let $b''$ be the opposite of $-b'$ in the Levi sphere $S''=(S\setminus Z)\cup Y$. Then $b''$ is a singular $m-1$-simplex in $S$, and $b''$ is precisely the reflection $s(b')$ of $b'$ along $H$. [ plus 1filll $\Box$]{}
For every Levi sphere $S$ we obtain in this way a finite reflection group $W_S$ which permutes the singular great spheres in $S$. As a reprentation sphere, $S$ may split off a trivial factor $S_0$, the intersection of all singular great spheres in $S$. We let $S_+$ denote its orthogonal complement, $S=S_0*S_+$. The intersections of the singular great spheres with $S_+$ turn $S_+$ into a spherical Coxeter complex, with Coxeter group $W_S$. Let $F\subseteq S$ be a fundamental domain for $W_S$, i.e. $F=C*S_0$, where $C\subseteq S_+$ is a Weyl chamber. The geometric realization of the t-class of any $m$-simplex in $F$ is precisely $F$.
If two Levi spheres $S,S'$ have an $m$-simplex $a$ in common, then there is a unique isometry ${\varphi}:S\rTo S'$ fixing $S\cap S'$ pointwise. The isometry fixes $S_0$ and maps $W_S$ isomorphically onto $W_{S'}$.
*Proof.*
The intersection $Y=S\cap S'$ contains the fundamental domain $F$. Since $F$ is relatively open in $S$, there is a unique isometry ${\varphi}:S\rTo S'$ fixing $Y$. The Coxeter group $W_s$ is generated by the reflections along the singular $m-1$-simplices in $Y$. Therefore ${\varphi}$ conjugates $W_S$ onto $W_{S'}$. Finally, $Y$ contains $S_0$. [ plus 1filll $\Box$]{}
If two Levi spheres $S,S'$ have a point $u$ in common, then $S_0=S_0'$. Furtheremore, there exists an isometry ${\varphi}:S\rTo S'$ which fixes $S\cap S'$ and which conjugates $W_S$ to $W_{S'}$.
*Proof.*
Let $a,a'$ be $m$-simplices in $S$ and $S'$ containing $u$, and let $S''$ be a Levi sphere containing $a$ and $a'$. We compose $S\rTo S''\rTo S'$.
Let $A$ be completely reducible. Then there is a thick spherical building $Z$ such that $|A|$ is the metric realization of $Z*{\mathbb{S}}^0*\cdots*{\mathbb{S}}^0$.
*Proof.*
Let $S$ be a Levi sphere and let $k=\dim S_0+1$. We make $S_0$ into a Coxeter complex with Coxeter group $W_0={\mathbb{Z}}/2^k$ (we fix an action, this is not canonical). By the previous Corollary, we can transport the simplicial structure on $S$ unambiguously to any Levi sphere in $A$. [ plus 1filll $\Box$]{}
For $A=X$, this is Scharlau’s reduction theorem for weak spherical buildings [@Scha] [@Cap].
[www]{}
P. S. Alexandrov, [*Combinatorial topology. Vol. 1, 2 and 3*]{}, Translated from the Russian, Reprint of the 1956, 1957 and 1960 translations, Dover, Mineola, NY, 1998. MR1643155 (99g:55001)
A. Balser and A. Lytchak, Building-like spaces, J. Math. Kyoto Univ. [**46**]{} (2006), no. 4, 789–804. MR2320351 (2008b:53052)
A. Balser and A. Lytchak, Centers of convex subsets of buildings, Ann. Global Anal. Geom. [**28**]{} (2005), no. 2, 201–209. MR2180749 (2006g:53049)
M. R. Bridson and A. Haefliger, [*Metric spaces of non-positive curvature*]{}, Springer, Berlin, 1999. MR1744486 (2000k:53038)
P.-E. Caprace, The thick frame of a weak twin building, Adv. Geom. [**5**]{} (2005), no. 1, 119–136. MR2110465 (2006a:51010)
R. Charney and A. Lytchak, Metric characterizations of spherical and Euclidean buildings, Geom. Topol. [**5**]{} (2001), 521–550 (electronic). MR1833752 (2002h:51008)
B. Kleiner and B. Leeb, Rigidity of quasi-isometries for symmetric spaces and Euclidean buildings, Inst. Hautes Études Sci. Publ. Math. No. 86 (1997), 115–197 (1998). MR1608566 (98m:53068)
R. Scharlau, A structure theorem for weak buildings of spherical type, Geom. Dedicata [**24**]{} (1987), no. 1, 77–84. MR0904550 (89b:51012)
J.-P. Serre, Complète réductibilité, Astérisque No. 299 (2005), Exp. No. 932, viii, 195–217. MR2167207 (2006d:20084)
J. Tits, [*Buildings of spherical type and finite BN-pairs*]{}, Lecture Notes in Math., 386, Springer, Berlin, 1974. MR0470099 (57 \#9866)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the dispersive properties for the wave equation in the Kerr space-time with small angular momentum. The main result of this paper is to establish Strichartz estimates for solutions of the aforementioned equation. This follows a local energy decay result for the Kerr space-time obtained in the earlier work [@TT] and uses the techniques and results from [@MMTT] by the author and collaborators. As an application, we then prove global well-posedness and uniqueness for the energy critical semilinear wave equation.'
address: 'Department of Mathematics, Purdue University, West Lafayette, IN 47907-2067'
author:
- Mihai Tohaneanu
title: Strichartz estimates on Kerr black hole backgrounds
---
Introduction
============
Understanding the decay properties of solutions to the linear wave equation on Kerr backgrounds is considered a crucial first step in proving the stability of the Kerr solution to the Einstein equations. Until recently even the problem of obtaining uniform bounds for such solutions was completely open, and only some partial results (pointwise decay and energy bounds for azimuthal solutions away from the event horizon) were obtained in [@FKSY], [@FS]. Recently in [@TT] and independently in [@DR2], [@DR3], [@AB] uniform pointwise bounds as well as local energy decay were established for small angular momentum. The aim of this paper is to prove Strichartz estimates under the same assumption of small angular momentum.
The Kerr geometry in Boyer-Lindquist coordinates is given by $$ds^2 = g_{tt}dt^2 + g_{t\phi}dtd\phi + g_{rr}dr^2 + g_{\phi\phi}d\phi^2
+ g_{\theta\theta}d\theta^2$$ where $t \in {{\mathbb R}}$, $r > 0$, $(\phi,\theta)$ are the spherical coordinates on ${{\mathbb S}}^2$ and $$g_{tt}=-\frac{\Delta-a^2\sin^2\theta}{\rho^2}, \qquad
g_{t\phi}=-2a\frac{2Mr\sin^2\theta}{\rho^2}, \qquad
g_{rr}=\frac{\rho^2}{\Delta}$$ $$g_{\phi\phi}=\frac{(r^2+a^2)^2-a^2\Delta
\sin^2\theta}{\rho^2}\sin^2\theta, \qquad g_{\theta\theta}={\rho^2}$$ with $$\Delta=r^2-2Mr+a^2, \qquad \rho^2=r^2+a^2\cos^2\theta.$$
Here $M$ represents the mass of the black hole, and $aM$ its angular momentum.
A straightforward computation gives us the inverse of the metric: $$g^{tt}=-\frac{(r^2+a^2)^2-a^2\Delta\sin^2\theta}{\rho^2\Delta},
\qquad g^{t\phi}=-a\frac{2Mr}{\rho^2\Delta}, \qquad
g^{rr}=\frac{\Delta}{\rho^2},$$ $$g^{\phi\phi}=\frac{\Delta-a^2\sin^2\theta}{\rho^2\Delta\sin^2\theta}
, \qquad g^{\theta\theta}=\frac{1}{\rho^2}.$$
The case $a = 0$ corresponds to the Schwarzschild space-time. We shall subsequently assume that $a$ is small $a \ll M$, so that the Kerr metric is a small perturbation of the Schwarzschild metric. We let $\Box_{\mathbf K} $ denote the d’Alembertian associated to the Kerr metric.
In the above coordinates the Kerr metric has singularities at $r = 0$ on the equator $\theta = \pi/2$ and at the roots of $\Delta$, namely $r_{\pm}=M\pm\sqrt{M^2-a^2}$. The singularity at $r=r_{+}$ is just a coordinate singularity, and corresponds to the event horizon. The singularity at $r = r_-$ is also a coordinate singularity; for a further discussion of its nature, which is not relevant for our results, we refer the reader to [@Ch; @HE]. To remove the singularities at $r = r_{\pm}$ we introduce functions $r^*$, $v_{+}$ and $\phi_{+}$ so that (see [@HE]) $$dr^*=(r^2+a^2)\Delta^{-1}dr,
\qquad
dv_{+}=dt+dr^*,
\qquad
d\phi_{+}=d\phi+a\Delta^{-1}dr.$$
The metric then becomes $$\begin{split}
ds^2= &\
-(1-\frac{2Mr}{\rho^2})dv_{+}^2+2drdv_{+}-4a\rho^{-2}Mr\sin^2\theta
dv_{+}d\phi_{+} -2a\sin^2\theta dr d\phi_{+} +\rho^2 d\theta^2 \\
& \ +\rho^{-2}[(r^2+a^2)^2-\Delta a^2\sin^2\theta]\sin^2\theta
d\phi_{+}^2
\end{split}$$ which is smooth and nondegenerate across the event horizon up to but not including $r = 0$. Just like in [@MMTT] and [@TT], we introduce the function $${{\tilde{v}}}= v_{+} - \mu(r)$$ where $\mu$ is a smooth function of $r$. In the $({{\tilde{v}}},r,\phi_{+},
\theta)$ coordinates the metric has the form $$\begin{split}
ds^2= &\ (1-\frac{2Mr}{\rho^2}) d{{\tilde{v}}}^2
+2\left(1-(1-\frac{2Mr}{\rho^2})\mu'(r)\right) d{{\tilde{v}}}dr \\
&\ -4a\rho^{-2}Mr\sin^2\theta d{{\tilde{v}}}d\phi_{+} + \Bigl(2 \mu'(r) -
(1-\frac{2Mr}{\rho^2}) (\mu'(r))^2\Bigr) dr^2 \\
&\ -2a\theta (1+2\rho^{-2}Mr\mu' (r))\sin^2dr d\phi_{+} +\rho^2
d\theta^2 \\
&\ +\rho^{-2}[(r^2+a^2)^2-\Delta a^2\sin^2\theta]\sin^2\theta
d\phi_{+}^2.
\end{split}$$
On the function $\mu$ we impose the following two conditions:
\(i) $\mu (r) \geq {{r^*}}$ for $r > 2M$, with equality for $r >
{5M}/2$.
\(ii) The surfaces ${{\tilde{v}}}= const$ are space-like, i.e. $$\mu'(r) > 0, \qquad 2 - (1-\frac{2Mr}{\rho^2}) \mu'(r) > 0.$$ As long as $a$ is small, we can work with the same function $\mu$ as in the case of the Schwarzschild space-time.
For convenience we also introduce $${{\tilde{\phi}}}= \zeta(r)\phi_{+}+(1-\zeta(r))\phi$$ where $\zeta$ is a cutoff function supported near the event horizon and work in the $({{\tilde{v}}},r,{{\tilde{\phi}}}, \theta)$ coordinates which are identical to $(t,r,\phi,\theta)$ outside of a small neighborhood of the event horizon.
Given $r_{-} < r_e <2M$ we consider the wave equation $$\Box_{\mathbf K} u = f
\label{boxsinhom}$$ in the cylindrical region $${{\mathcal}M}_{R} = \{ {{\tilde{v}}}\geq 0, \ r \geq r_e \}
\label{mr}$$ with initial data on the space-like surface $$\Sigma_R^- = {{\mathcal}M}_{R} \cap \{ {{\tilde{v}}}= 0 \}
\label{mr-}$$ The lateral boundary of ${{\mathcal}M}_R$, $$\Sigma_R^+ = {{\mathcal}M}_R \cap \{ r = r_e\}
\label{mr+}$$ is also space-like, and can be thought of as the exit surface for all waves which cross the event horizon.
We define the initial (incoming) energy on $\Sigma_R^-$ as $$\label{energy1}
E[u](\Sigma_R^-) = \int_{\Sigma_R^-} \left( |\partial_r u|^2 +
|\partial_{{\tilde{v}}}u|^2 + |{{\not\negmedspace\nabla}}u|^2 \right) r^2 dr d\omega$$ the outgoing energy on $\Sigma_R^+$ as $$\label{energy2}
E[u](\Sigma_R^+) = \int_{\Sigma_R^+}
\left( |\partial_r u|^2 + |\partial_{{\tilde{v}}}u|^2 +
|{{\not\negmedspace\nabla}}u|^2 \right) r_e^2 d{{\tilde{v}}}d\omega$$ and the energy on an arbitrary ${{\tilde{v}}}$ slice as $$\label{energy3}
E[u]({{\tilde{v}}}_0) = \int_{ {{\mathcal}M}_R \cap \{{{\tilde{v}}}= {{\tilde{v}}}_0\}}
\left(
|\partial_r u|^2 + |\partial_{{\tilde{v}}}u|^2 + |{{\not\negmedspace\nabla}}u|^2
\right) r^2 dr d\omega$$
For solutions to the constant coefficient wave equation on ${{\mathbb R}}\times {{\mathbb R}}^3$, the well-known Strichartz estimates state that $$\label{strich}
\||D_x|^{-\rho_1} \nabla u\|_{L^{p_1}_tL^{q_1}_x} \lesssim
\|\nabla u(0)\|_{L^2} + \||D_x|^{\rho_2} f\|_{L^{p_2'}_tL^{q_2'}_x}.$$ Here the exponents $(\rho_i,p_i,q_i)$ are subject to the scaling relation $$\frac{1}p+\frac{3}q = \frac{3}2 -\rho
\label{scalingpq}$$ and the dispersion relation $$\frac{1}p + \frac{1}q \leq \frac12, \qquad 2 < p \leq \infty.
\label{dispersionpq}$$ All pairs $(\rho,p,q)$ satisfying and are called Strichartz pairs. Those for which the equality holds in are called sharp Strichartz pairs. Such estimates first appeared in the seminal works [@Brenner], [@Strichartz1; @Strichartz2] and as stated include contributions from, e.g., [@GV], [@P], [@K], [@LiSo], and [@KT].
If one allows variable coefficients, such estimates are well-understood locally-in-time. For smooth coefficients, this was first shown in [@MSS] and later for $C^2$ coefficients in [@Smith] and [@T1; @T2; @T3].
Globally-in-time, the problem is more delicate. Even a small, smooth, compactly supported perturbation of the flat metric may refocus a group of rays and produce caustics. Thus, constructing a parametrix for incoming rays proves to be quite difficult. At the same time, one needs to contend with the possibility of trapped rays at high frequencies and with eigenfunctions/resonances at low frequencies.
Our main theorem is the following global in time estimate:
\[Strichartz.theorem\] If $u$ solves $\Box_{\mathbf K} u = f$ in ${{\mathcal}M}_R$ then for all nonsharp Strichartz pairs $(\rho_1,p_1,q_1)$ and $(\rho_2,p_2,q_2)$ we have $$E[u](\Sigma_R^+) + \sup_{{{\tilde{v}}}} E[u]({{\tilde{v}}}) + \left\| \nabla u\right\|^2_{L^{p_1}_{{\tilde{v}}}\dot H^{-\rho_1,q_1}_x}
\lesssim E[u](\Sigma_R^-) +
\left\| f \right\|^2_{L^{p_2'}_{{\tilde{v}}}\dot H^{\rho_2, q_2'}_x }.$$
Here the Sobolev-type spaces $\dot H^{s,p}$ coincide with the usual $\dot H^{s,p}$ homogeneous spaces in ${{\mathbb R}}^3$ expressed in polar coordinates $(r,\omega)$.
As a corollary of this result one can consider the global solvability question for the energy critical semilinear wave equation in the Kerr space. Let $$\Sigma_0 = \{ t = 0 \}$$ Note that $\Sigma_0$ is a smooth, spacelike surface.
We now consider the Cauchy problem with initial data on $\Sigma_0$: $$\left\{
\begin{array}{lc}
\Box_{\mathbf K} u = \pm u^5 & \text{in } {{\mathcal}M}\cr \cr
u = u_0, \ \tilde T u = u_1 & \text{in } \Sigma_0.
\end{array}
\right.
\label{nonlin}$$ Here $\tilde T$ is a vector field that is smooth, everywhere timelike and equals $\partial_{{\tilde{v}}}$ on $\Sigma_0$ outside ${{\mathcal}M}_C$. Observe that we cannot use $\partial_{{\tilde{v}}}$ on all of $\Sigma_0$ since it becomes spacelike inside the ergosphere (i.e. when $g_{tt}>0$).
\[tnlw\] Let $r_e > r_{-}$. Then there exists $\epsilon > 0$ so that for each initial data $(u_0,u_1)$ which satisfies $$E[u](\Sigma_0) \leq \epsilon$$ the equation admits an unique solution $u$ in the region $\{ r > r_e\}$ which satisfies the bound $$E[u](\Sigma_{r_e}) + \| u \|_{\dot H^{s,p}(\{r > r_e\})} \lesssim
E[u](\Sigma_0)$$ for all indices $s,p$ satisfying $$\frac{4}p = s+\frac12, \qquad 0 \leq s < \frac12.$$ Furthermore, the solution has a Lipschitz dependence on the initial data in the above topology.
Theorem \[tnlw\] follows from Theorem \[Strichartz.theorem\] exactly like in Section 5 of [@MMTT], hence we will now focus on proving the latter.
[**Acknowledgements:**]{} The author would like to thank Daniel Tataru for many useful conversations and suggestions regarding the paper.
The Strichartz estimates
========================
In this section we will prove Theorem \[Strichartz.theorem\]. The key to the proof is the approach developed in [@MT] for small perturbations of the Minkowski metric and [@MMTT] for the Schwarzschild metric which allows one to establish global-in-time Strichartz estimates provided that a strong form of the local energy estimates holds.
A weaker local energy result has been proved in [@TT]. We first improve this to a stronger result that only requires logarithmic losses in the energy norm. We then apply the techniques from [@MT] and [@MMTT] to obtain Strichartz estimates for all nonsharp exponents.
Let us first recall the setup and results from [@TT]. Let $\tau, \xi, \Phi$ and $\Theta$ be the Fourier variables corresponding to $t, r, \phi$ and $\theta$, and $$p( r, \phi,\tau, \xi, \Phi,\Theta)
=g^{tt}\tau^2+2g^{t\phi}\tau\Phi+g^{\phi\phi}\Phi^2
+g^{rr}\xi^2 +g^{\theta\theta}\Theta^2$$ be the principal symbol of $\Box_{\mathbf K} $. On any null geodesic one has $$\label{Ham}
p(t, r, \phi,\theta,\tau, \xi, \Phi,\Theta)=0.$$
Moreover, all trapped null geodesics in the exterior $r>r_+$ must also satisfy (see [@TT] for more details): $$\label{xi}
\xi= 0$$ $$\label{rct}
r= \text{constant} \;
\text{ so that } \; (2r\Delta - (r-M)\rho^2)^2\leq 4a^2 r^2\Delta\sin^2 \theta$$ $$\label{tauphi}
R_a(r,\tau,\Phi)= 0$$ where $$\begin{aligned}
R_a(r,\tau,\Phi)=&(r^2+a^2)(r^3-3Mr^2+a^2r+a^2M)\tau^2 \\
&{}- 2aM(r^2-a^2)\tau\Phi- a^2(r-M)\Phi^2
\end{aligned}$$
By we can bound $\Phi$ in terms of $\tau$, $$|\Phi| \leq 4 M |\tau|$$ For $\Phi$ in this range and small $a$ the polynomial $\tau^{-2} R_a(r,\tau,\Phi)$ can be viewed as a small perturbation of $$\tau^{-2} R_0(r,\tau,\Phi) = r^4(r-3M)$$ which has a simple root at $r = 3M$. Hence for small $a$ the polynomial $R_a$ has a simple root close to $3M$, which we denote by $r_a(\tau,\Phi)$.
In [@TT] the following result is proved:
\[KerrLE\] Let $u$ solve $\Box_{\mathbf K} u = f$ in ${{\mathcal}M}_R$. Then $$\|u \|_{LEW_{\mathbf K}}^2 + \sup_{\tilde v} E[u](\tilde v) + E[u](\Sigma_R^+)
\lesssim E[u](\Sigma_R^-)+ \|f \|_{LEW_{\mathbf K}^{*}}^2.$$
The exact definition of the norms $LEW_{\mathbf K}$ is not important; what matters is that it is equivalent to $H^1_{t,x}$ on compact sets outside an $O(1)$ neighborhood of $r=3M$ and is degenerate on the trapped set (described above). Near infinity (say for $r>4M$) the norm $LEW_{\mathbf K}$ is defined on dyadic regions as follows: $$\label{LEinfty}
\|u\|_{LEW_{\mathbf K}} = \sup_{j \in {{\mathbb Z}}}
2^{-\frac{j}2}\|\nabla_{t, x} u\|_{L^2 ({{\mathbb R}}\times \{|x|\in [2^{j-1}, 2^j]\})} + 2^{-\frac{3j}2}\|u\|_{L^2 ({{\mathbb R}}\times \{|x|\in [2^{j-1}, 2^j]\})}$$
The norm $LEW_{\mathbf K}^{*}$ is the dual of $LEW_{\mathbf K}$; in particular, near infinity it will be defined as $$\|f\|_{LEW_{\mathbf K}^*}= \sum_{j \in {{\mathbb Z}}}
2^{\frac{j}2}\|f\|_{L^2 ({{\mathbb R}}\times \{|x|\in [2^{j-1}, 2^j]\})}
\label{le*m}$$
We first improve Theorem \[KerrLE\] around $r=3M$. We will work with the usual Boyer-Lindquist coordinates $(t, r, \phi, \theta)$ since we are only interested in improving the estimate around $r=3M$. In these coordinates we can write $$\rho^2 \sqrt{\Delta} \Box_{\mathbf K} \frac{1}{\sqrt{\Delta}} = L_{K} =
\Delta \partial_r^2 + (-\frac{(r^2+a^2)^2}{\Delta}\partial_t^2 -a\frac{2Mr}{\Delta}\partial_t
\partial_{\phi} -\frac{a^2}{\Delta}\partial_{\phi}^2 + L_a) + V(r),
\label{KerrRW}$$ where $$L_a = a^2 \sin^2 \theta \partial_t^2 + \frac{1}{\sin^2 \theta}\partial_{\phi}^2 + \frac{1}{\sin \theta}\partial_{\theta} (\sin \theta \partial_{\theta})$$ and $$V(r) = \sqrt{\Delta}\partial_{r}(\Delta\partial_{r}\frac{1}{\sqrt{\Delta}})$$
We can now take the Fourier transform in $t$ (this is allowed since by Theorem \[KerrLE\] since $u$ is a tempered distribution in $t$) and $\phi$, and expand $\mathcal{F}_t L_a$ (which is an elliptic operator on ${{\mathbb S}}^2$) as the countable sum of its eigenvectors and corresponding eigenvalues $\lambda_a^2$ (note that $\lambda_0$ correspond to the usual spherical harmonics). This is possible since the operators $\partial_\phi$ and $L_a$ commute. We are left with the ordinary differential equation $$\label{psKeq}
(\Delta \partial_r^2 + V_{\lambda_a, \tau, \Phi}(r))w(r) = g$$ where $$\label{Vdiff}
V_{\lambda_a,\tau, \Phi}(r) = \tau^2 W(r - r_a (\tau, \Phi)) - \lambda_a^2 + V.$$ and $W(0)=W'(0)=0$, $W''(0)>0$.
Let $\gamma_0: {{\mathbb R}}\to {{\mathbb R}}^+$ be a smooth increasing function so that $$\gamma_0(y) = \left\{ \begin{array}{cc} 1& y < 1 \cr y & y \geq 2.
\end{array}
\right.$$
Let $\gamma_1: {{\mathbb R}}^+ \to {{\mathbb R}}^+$ be a smooth increasing function so that $$\gamma_1(y) = \left\{ \begin{array}{cc} y^\frac12 & y < 1/2 \cr 1 & y \geq 1.
\end{array}
\right.$$
Let $\gamma: {{\mathbb R}}^2 \to {{\mathbb R}}^+$ be a smooth function with the following properties: $$\label{fnpdo}
\gamma(y,z) = \left\{ \begin{array}{cc} 1 & z < C \cr
\gamma_0(y) & y< \sqrt{z/2},\ z \geq C \cr
z^\frac12 \gamma_1(y^2/z) & y \geq \sqrt{z/2}, \ z \geq C
\end{array}
\right.$$ where $C$ is a large constant. In the sequel $z$ is either a discrete parameter or very large (see Remark \[roleoflambda\]), so the lack of smoothness at $z=C$ is of no consequence.
We define the symbol $$b_{ps}(r, \tau, \xi, \Phi, \lambda_a) = \gamma(-\psi(\frac{\lambda_a}{\tau})\ln((r-r_a(\tau, \Phi))^2 +\lambda_a^{-2} \xi^2), \ln \lambda_a),$$ and its inverse $$b_{ps}^{-1}(r, \tau, \xi, \Phi, \lambda_a) =
\frac{1}{\gamma(-\psi(\frac{\lambda_a}{\tau})\ln((r-r_a(\tau, \Phi))^2 +\lambda_a^{-2} \xi^2), \ln \lambda_a)}.$$ Here $\psi:{{\mathbb R}}\to{{\mathbb R}}$ is a smooth cutoff such that $$\psi(y) = \left\{ \begin{array}{cc} 1 & y \in [-4, -2]\cup [2, 4] \cr 0 & y\in (-\infty, 8]\cup [-1, 1] \cup [8, \infty)
\end{array}
\right.$$ The role of $\psi$ is to make sure that $b=1$ when $\tau \ll \lambda_a$ and $\tau \gg \lambda_a$.
Observe that, as opposed to their Schwarzschild counterparts $a_{ps}$ and $a_{ps}^{-1}$ defined in [@MMTT], the symbols $b_{ps}$ and $b_{ps}^{-1}$ will depend on $\tau$ and $\Phi$.
We note that if $\lambda_a$ is small then $b_{ps}$ and $b_{ps}^{-1}$ both equal $1$, while if $\lambda_a$ is large then they satisfy the bounds $$\begin{split}
1 \leq b_{ps}(r, \tau, \xi, \Phi, \lambda_a) \leq b_{ps}(r, \tau, 0, \Phi, \lambda_a)\leq (\ln
\lambda_a)^\frac12,
\\
(\ln
\lambda_a)^{-\frac12}\leq b_{ps}^{-1}(r, \tau, 0, \Phi, \lambda_a) \leq b_{ps}^{-1}(r, \tau, \xi, \Phi, \lambda_a)
\leq 1.
\end{split}
\label{bpsbds}$$
We also observe that the region where $y^2 > z$ corresponds to $(r-r_a(\tau, \Phi))^2
+\lambda_a^{-2} \xi^2 < e^{-\sqrt{ \ln \lambda_a}}$. Thus differentiating the two symbols we obtain the following bounds $$\begin{aligned}
|\partial_{r}^\alpha \partial_\xi^\beta \partial_{\lambda_a}^\nu \partial_\tau^\eta
b_{ps}(r, \tau, \xi, \Phi, \lambda_a)| \leq c_{\alpha,\beta,\nu} (1+|\ln((r-r_a(\tau, \Phi))^2 +\lambda_a^{-2} \xi^2)|)\lambda_a^{-\beta-\nu-\eta} \\
((r-r_a(\tau, \Phi))^2 +\lambda_a^{-2} \xi^2 + e^{-\sqrt{ \ln \lambda_a}} )^{-\frac{\alpha+\beta+\eta}2},
\label{bpsbd}\end{aligned}$$ respectively $$\begin{aligned}
|\partial_{r}^\alpha \partial_\xi^\beta \partial_{\lambda_a}^\nu \partial_\tau^\eta
b_{ps}^{-1}(r, \tau, \xi, \Phi, \lambda_a)| \leq c_{\alpha,\beta,\nu}
b_{ps}^{-2}(r, \tau, \xi, \Phi, \lambda_a) (1+|\ln((r-r_a(\tau, \Phi))^2 +\lambda_a^{-2} \xi^2)|) \\ \lambda_a^{-\beta-\nu-\eta}((r-r_a(\tau, \Phi))^2 +\lambda_a^{-2}
\xi^2 + e^{-\sqrt{ \ln \lambda_a}} )^{-\frac{\alpha+\beta+\eta}2},
\label{bps1bd}\end{aligned}$$ when $\alpha + \beta + \nu + \tau> 0$ and $\ln \lambda \geq C$. These show that we have a good operator calculus for the corresponding pseudodifferential operators. In particular in terms of the classical symbol classes we have $$\label{apssymb}
b_{ps}, b_{ps}^{-1} \in S^{\delta}_{1- \delta, 0}, \qquad \delta > 0.$$
By and one easily sees that these operators are approximate inverses. More precisely for small $\lambda_a$, $\ln \lambda_a < C$, they are both the identity, while for large $\lambda_a$ $$\label{almostid}
\| b_{ps}^w(\lambda_a) (b_{ps}^{-1})^w(\lambda_a)-I\|_{L^2 \to L^2} \lesssim
\lambda_a^{-1} e^{\sqrt{\ln \lambda_a}}, \qquad \ln \lambda_a \geq C.$$ Choosing $C$ large enough we insure that the bound above is always much smaller than $1$.
\[roleoflambda\] The role played by $\lambda_a$ changes from the proof of Theorems \[theorem.3\] to the proof of the Strichartz estimates. Since all of our $L^2$ estimates admit orthogonal decompositions with respect to the eigenfunctions associated to $\lambda_a$, it suffices for Theorems \[theorem.3\] to fix $\lambda_a$ and work with the operators $b_{ps}^w(\lambda_a)$. However, in the proof of the Strichartz estimates we need kernel bounds for operators of the form $B_{ps}$, which is why we think of $\lambda_a$ as a fourth Fourier variable (besides $\tau$, $\Phi$ and $\xi$) and track the symbol regularity with respect to $\lambda_a$ as well. Of course, this is meaningless for $\lambda_a$ in a compact set; only the asymptotic behavior as $\lambda_a \to \infty$ is relevant.
We can now introduce the Weyl operators $$B_{ps} = \sum_{\lambda_a} b_{ps}^w(\lambda_a) \Pi_{\lambda_a},$$ respectively $$B_{ps}^{-1} = \sum_{\lambda_a} (b_{ps}^{-1})^w(\lambda_a) \Pi_{\lambda_a}.$$ where $\Pi_{\lambda_a}$ are the spectral projectors on the eigenspaces of $L_a$ determined by $\lambda_a$.
We use these two operators in order to define the improved local energy norms around $r=3M$. Let $I $ be a small neighborhood of $3M$ (which does not depend on $a$). We say that $\tilde u \in LEK_{ps}$ if $\tilde u : {{\mathbb R}}\times I \times {{\mathbb S}}^2 \to {{\mathbb R}}$ and $$\|\tilde u\|_{LEK_{ps}} = \|B^{-1}_{ps}\tilde u\|_{H^1} \approx
\| A_{ps}^{-1} \nabla\tilde u\|_{L^2} < \infty,$$
Similarly $\tilde f \in LEK_{ps}^*$ if $\tilde f : {{\mathbb R}}\times I \times {{\mathbb S}}^2 \to {{\mathbb R}}$ and $$\|\tilde f \|_{LEK_{ps}^*} = \| B_{ps}\tilde f\|_{L^2} < \infty.$$
Let $\chi(r)$ be a cutoff function supported on $I$ which equals $1$ on a smaller neighborhood of size $O(1)$ near $3M$, and $u: {{\mathcal}M}_R \to {{\mathbb R}}$. We say that $u \in LE_{\mathbf K}$ if $\chi u$ is the restriction of some $\tilde u \in LEK_{ps}$ on ${{\mathcal}M}_R$, and $$\| u\|_{LE_{\mathbf K}} = \inf_{\tilde u \mid_{{{\mathcal}M}_R} = \chi u}\|\tilde u\|_{LEK_{ps}} + \|(1-\chi) u\|_{LEW_{\mathbf K}}$$
Similarly if $f: {{\mathcal}M}_R \to {{\mathbb R}}$, then $f \in LE_{\mathbf K}^*$ if $\chi f$ is the restriction of some $\tilde f \in LEK_{ps}^*$ on ${{\mathcal}M}_R$, and $$\| f\|_{LE_{\mathbf K}^*} = \inf_{\tilde f \mid_{{{\mathcal}M}_R} = \chi u} \|\chi \tilde f\|_{LEK_{ps}^*} + \|(1-\chi) f\|_{LEW_{\mathbf K}^{*}}$$
Our improved local energy estimate reads:
\[theorem.3\] For all functions $u$ which solve $\Box_{\mathbf K} u = f$ in ${{\mathcal}M}_R$ we have $$\sup_{{{\tilde{v}}}> 0} E[u]({{\tilde{v}}}) + E[u](\Sigma_R^+) +
\| u \|_{LE_{\mathbf K}}^2 \lesssim E[u](\Sigma_R^-) + \| f \|_{LE_{\mathbf K}^*}^2.$$
We start with the following estimate near $r=3M$, which is the equivalent of Proposition 3.3 from [@MMTT]:
a\) Let $u$ be a function supported in $\{ 5M/2 < r < 5M \}$ which solves $\Box_{\mathbf K} u = f$. Then $$\label{Kerrpsb}
\| u\|_{LEK_{ps}}^2\lesssim \| f\|^2_{LEK_{ps}^*}.$$
b\) Let $f \in {LEK_{ps}}^*$ be supported in $\{ 11M/4 < r < 4M \}$. Then there is a function $u$ supported in $\{ 5M/2 < r < 5M \}$ so that $$\sup_t E[u] + \| u\|_{LEK_{ps}}^2
+ \| \Box_{\mathbf K} u -f\|_{LE_{\mathbf K}^{1*}}^2
\lesssim \| f\|^2_{ LEK_{ps}^*}.
\label{Kerrppsb}$$ \[Kerrppsbprop\]
We start with part (a). By Plancherel’s formula and the fact that $$\label{Phi<lambda_a}
\Phi < \lambda_a$$ will follow if we can prove that $$\| \partial_r w(r) \|_{L^2} + (|\tau|+|\lambda_a|)\|(b_{ps}^{-1})^w(\tau, \Phi, \lambda_a) w\|_{L^2} \lesssim \| b_{ps}^w(\tau, \Phi, \lambda_a) g\|_{L^2}.
\label{Kerrpsbhard}$$ for any fixed $\tau$, $\Phi$ and $\lambda_a$ (here $b_{ps}^w(\tau, \Phi, \lambda_a)$ is the one dimensional pseudodifferential operator obtained from $b_{ps}$ by fixing $\tau$, $\Phi$ and $\lambda_a$.)
Depending on the relative sizes of $\lambda_a$, $\tau$ and $\Phi$ and taking into account we consider several cases. In the easier cases it suffices to replace the bound with a simpler bound $$\| \partial_{r} w \|_{L^2} + (|\tau|+|\lambda_a|)\|w\|_{L^2} \lesssim \|g\|_{L^2}.
\label{Kerrpsbeasy}$$
[**Case 1:**]{} $\lambda_a,\tau, \Phi \lesssim 1$. Then $V_{\lambda_a,\tau, \Phi}(r) \approx 1$. We solve as a Cauchy problem with data on one side; namely, if $$E[w](r) = \Delta(\partial_ r w)^2 + w^2$$ then an easy computation shows that
$$\frac{d}{dr}E[w](r) \lesssim E[w](r) + g^2$$ By Gronwall’s inequality and the fact that $E[w](r) = 0$ away from the photonsphere we obtain the pointwise bound $$|w|+|\partial_{r} w| \lesssim \|g\|_{L^2}$$ which easily implies .
[**Case 2:**]{} $\Phi, \lambda_a \ll \tau$. Then $V_{\lambda_a,\tau, \Phi}(r) \approx
\tau^2$, $\partial_r V_{\lambda_a,\tau, \Phi}(r) \approx \tau^2$ for $r$ in a compact set; therefore is hyperbolic in nature. Hence we can solve as a Cauchy problem with data on one side; namely, if $$E[w](r) = \Delta (\partial_r w)^2 + V_{\lambda_a,\tau,\Phi}(r)w^2$$ then an easy computation shows that
$$\frac{d}{dr}E[w](r) \lesssim E[w](r) + g^2$$ By Gronwall’s inequality and the fact that $E[w](r) = 0$ away from the photonsphere we obtain the pointwise bound
$$\tau |w|+|\partial_r w| \lesssim \|g\|_{L^2}$$ which implies .
[**Case 3:**]{} $\lambda_a \gg \tau$. Then $V_{\lambda_a,\tau, \Phi}(r) \approx
-\lambda_a^2$ for $r$ in a compact set; therefore is elliptic. Then we solve as an elliptic problem with Dirichlet boundary conditions on a compact interval and obtain $$\lambda_a^{\frac32}\|w\|_{L^2}+\lambda_a^{\frac12}\|\partial_r w\|_{L^2} \lesssim \|g\|_{L^2}$$ which again gives .
[**Case 4:**]{} $\lambda_a \approx \tau \gg 1$. In this case no longer holds. However, we can now use Proposition 3.4 from [@MMTT] which states
\[ODE\] Let $W$ be a smooth function satisfying $W(0)=W'(0)=0$, $W''(0)>0$, and $|\epsilon| \lesssim 1$. Let $w$ be a solution of the ordinary differential equation $$(\partial_{{r^*}}^2 + \lambda^2(W({{r^*}}) +\epsilon)) w({{r^*}}) = g$$ supported near ${{r^*}}= 0$. Then we have $$\label{psahard}
\| \partial_{{{r^*}}} u \|_{L^2} + \lambda \| (a_{ps}^{-1})^w(\lambda) u\|_{L^2}
\lesssim \| a_{ps}^w(\lambda) g\|_{L^2} ,$$ Here the symbols $a_{ps}(\lambda)$ and $(a_{ps}^{-1})(\lambda)$ are defined as follows $$a_{ps}(\lambda)({{r^*}},\xi) = \gamma(-\ln({{r^*}}^2 +\lambda^{-2} \xi^2), \ln \lambda),$$ $$a_{ps}^{-1}(\lambda)({{r^*}},\xi) =
\frac{1}{\gamma(-\ln({{r^*}}^2 +\lambda^{-2} \xi^2), \ln \lambda)}.$$
We can now apply Proposition \[ODE\] with ${{r^*}}= r-r_a(\tau,\Phi)$, $W$ as in and with $b_{ps}^w(\tau, \Phi, \lambda_a)$ and $(b_{ps}^{-1})^w(\tau, \Phi, \lambda_a)$ replacing $a_{ps}^w(\lambda)$ and $(a_{ps}^{-1})^w(\lambda)$ respectively. Note that there the extra $\Delta$ coefficient in front of $\partial_r^2$ in plays no role, since $\Delta \approx \partial_r \Delta \approx 1$ near $r=3M$.
Part (b) follows now from part (a) exactly like in Proposition 3.3 from [@MMTT] with $\lambda_a$ replacing the spherical harmonics $\lambda$ and ${{r^*}}= r-r_a(\tau,\Phi)$.
We now repeat the arguments in Section 4 of [@MMTT] to turn the improved local energy estimate, Proposition \[Kerrppsbprop\] into Strichartz estimates, Theorem \[Strichartz.theorem\]. The only (minor) differences appear in proving Case II of Proposition 4.10, which we will settle below. We thus need to prove
\[Kerrps\] For $u$ supported in $\{ 5M/2 < r < 5M\}$ we have $$\|\nabla u\|_{L^p_{{\tilde{v}}}H^{-\rho,q}}^2 \lesssim
E[u](0)+\|u\|_{LEK_{ps}}^2 + \|\Box_K u\|_{LEK^*_{ps}}^2.$$
Clearly the operator $\Box_K$ can be replaced by $L_{K}$. The potential $V$ can be neglected due to the straightforward bound $$\|V u\|_{LE^*_{ps}} \lesssim \|u\|_{LE_{ps}}.$$ Indeed, for $u_{\lambda_a}$ the eigenfunction corresponding to the eigenvalue $\lambda_a^2$ we have $$\|V u_{\lambda_a}\|_{LEK^*_{ps}} \lesssim |\ln (2+\lambda_a)|^\frac12 \|u_{\lambda_a}\|_{L^2}
\lesssim \lambda_a |\ln (2+\lambda_a)|^{-\frac12} \|u_{\lambda_a}\|_{L^2} \lesssim
\|u_{\lambda_a}\|_{LE_{ps}}$$
We introduce the auxiliary function $$\psi = B_{ps}^{-1} u.$$ By the definition of the $LEK_{ps}$ norm we have $$\| \psi \|_{H^1} \lesssim \|u\|_{LEK_{ps}}.
\label{h1Kpsi}$$ We also claim that $$\|L_K \psi\|_{L^2} \lesssim \| u\|_{LEK_{ps}} + \| L_K u\|_{L^2}.
\label{Kppsi}$$ Since $B_{ps}^{-1}$ is $L^2$ bounded, this is a consequence of the commutator bound $$[ B_{ps}^{-1}, L_K ] : LEK_{ps} \to L^2,$$ or equivalently $$[ B_{ps}^{-1}, L_K] B_{ps}: H^1 \to L^2.
\label{Kcmt}$$ It suffices to consider the first term in the symbol calculus, as the remainder belongs to $OPS^\delta_{1-\delta,\delta}$, mapping $H^\delta$ to $L^2$ for all $\delta > 0$. The symbol of the first term is $$q(\tau, \Phi, \xi,r,\lambda_a) = \{ b_{ps}^{-1}(\lambda_a), \Delta \xi^2 -\frac{(r^2+a^2)^2}{\Delta}\tau^2 -a\frac{2Mr}{\Delta}\tau\Phi -\frac{a^2}{\Delta}\Phi^2 + \lambda_a^2 \}
b_{ps}(\lambda_a)$$ and a-priori we have $q \in S^{1+\delta}_{1-\delta,\delta}$. For a better estimate we compute the Poisson bracket $$\begin{split}
q(\tau, \Phi, \xi,r,\lambda_a) = &\ b_{ps}^{-1}(\lambda_a) \gamma_y(y,\ln \lambda_a)
\psi(\frac{\Phi}{\tau}) \\ &\ \frac{ 4 \xi (r-r_a(\tau, \Phi)) - 2 \xi \lambda_a^{-2}(\tau^2 W(r-r_a(\tau, \Phi)) + 2(r-M)\xi^2}{(r-r_a(\tau, \Phi))^2 + \lambda_a^{-2}\xi^2}
\end{split}$$ where $y = (r-r_a(\tau, \Phi))^2 + \lambda_a^{-2}\xi^2$ and $W\approx r^2$. The first three factors on the right are bounded. The fourth is bounded by $\max \{\lambda_a, \tau \}$ by Cauchy Schwarz and the fact that $q$ is supported in $|\xi| \lesssim \lambda_a$. Hence we obtain $q \in
(\tau + \lambda_a) S^{0}_{1-\delta,\delta}$, and the commutator bound follows.
Given and , we localize as before $\psi$ to time intervals of unit length and then apply the local Strichartz estimates. By summing over these strips we obtain $$\| \nabla \psi \|_{L^p H^{-\rho,q}} \lesssim \| u\|_{LEK_{ps}} + \|
L_K u\|_{L^2}$$ for all sharp Strichartz pairs $(\rho,p,q)$.
To return to $u$ we invert $B_{ps}^{-1}$, $$u = B_{ps} \psi + (I - B_{ps} B_{ps}^{-1}) u.$$ The second term is much more regular, since by we have $$I - B_{ps} B_{ps}^{-1} \in S^{-1 + \delta}_{1- \delta, 0}, \qquad \delta > 0.$$ This leads to $$\| \nabla (I - B_{ps} B_{ps}^{-1}) u\|_{L^2 H^{1-\delta}} \lesssim
\| u \|_{LEK_{ps}}, \qquad \delta > 0;$$ therefore all the Strichartz estimates are satisfied simply by Sobolev embeddings.
For the main term $B_{ps}\psi$ we take advantage of the fact that we only seek to prove the nonsharp Strichartz estimates for $u$. The nonsharp Strichartz estimates for $\psi$ are obtained from the sharp ones via Sobolev embeddings, $$\|\nabla \psi\|_{H^{-\rho_2,q_2}} \lesssim \|\nabla
\psi\|_{H^{-\rho_1,q_1}}, \qquad \frac{3}q_2 + \rho_2 = \frac{3}q_1 +
\rho_1, \quad \rho_1 < \rho_2.$$ To obtain the nonsharp estimates for $u$ instead, we need a slightly stronger form of the above bound, namely
Assume that $(\rho_1, p_1, q_1)$, $(\rho_2, p_2, q_2)$ are Strichartz pairs with $p_1 < p_2$, $q_1 < q_2 < \infty$. Then $$\label{Kps.Sobolev}
\|B_{ps} w \|_{L^{p_2}H^{-\rho_2,q_2}} \lesssim \| w\|_{L^{p_1}H^{-\rho_1,q_1}}, \qquad
\frac{3}q_2 + \rho_2 = \frac{3}q_1 + \rho_1.$$
We need to prove that the operator $$\tilde B = Op^w(\xi^2+\lambda_0^2+1)^{-\frac{\rho_2}2} B_{ps}
Op^w(\xi^2+\lambda_0^2+1)^{\frac{\rho_1}2}$$ maps $L^{q_1}$ into $L^{q_2}$. The principal symbol of $\tilde B$ is $$\tilde b(r,\tau, \xi,\Phi, \lambda_a) = (\xi^2+\lambda_0^2+1)^{\frac{\rho_1-\rho_2}2}
b_{ps}({{r^*}},\xi,\lambda_a),$$ and by the pdo calculus the remainder is easy to estimate, $$\tilde B - \tilde b^w \in OPS^{ \rho_1-\rho_2 -1 +\delta}_{1-\delta, 0}, \qquad \delta > 0.$$ The conclusion of the lemma will follow from the Hardy-Littlewood-Sobolev inequality (applied separately for $t$, $r$ and $\omega$) if we prove a suitable pointwise bound on the kernel $K$ of $b_0^w$, namely $$|K(t_1, r_1,\omega_1, t_2, r_2,\omega_2)| \lesssim |t_1-t_2|^{-1 + \frac{1}{p_1} -\frac{1}{p_2}}(|r_1-r_2|
|\omega_1-\omega_2|^2)^{-1 + \frac{1}{q_1} -\frac{1}{q_2}}.
\label{KerrKbd}$$ For fixed $r$ we consider a smooth dyadic partition of unity in frequency as follows: $$1 = (1-\chi_{\{\tau \approx \lambda_a \}})\chi_{\{|\xi| > \lambda_a\}}\chi_{\{\mu < 1\}}
+ \chi_{\{\tau \approx \lambda_a \}}\sum_{\mu\geq 1 \, \text{dyadic}} \chi_{\{\lambda_a \approx \mu\}}
\sum_{\nu = 0}^\mu \chi_{\{|\xi| \approx \nu\}}$$
This leads to a similar decomposition for $\tilde b$, namely $$\tilde b = {\tilde b}_{00} + \sum_{\mu\geq 1} \sum_{\nu = 0}^\mu
\tilde b_{\mu \nu}.$$ In the region $\{|\xi| \gtrsim \lambda\}\cup \{\mu < 1\}\cup \{\tau \gg \lambda \}\cup \{\tau \ll \lambda \}$ the symbol $\tilde b$ is of class $S^{\rho_1-\rho_2}$, which yields a kernel bound for $\tilde b_{00}$ of the form $$|K_{00}(t_1, r_1,\omega_1,t_2, r_2,\omega_2)| \lesssim (|t_1-t_2| + |r_1-r_2| +
|\omega_1-\omega_2|)^{-3 + \rho_2 - \rho_1}.$$ The symbols of $\tilde b_{\mu \nu}$ are supported in $\{ |\xi| \approx \nu, \
\lambda \approx \tau \approx \mu \}$, are smooth on the same scale and have size at most $\ln (\nu^{-1} \mu) \mu^{\rho_1-\rho_2}$. Clearly on their support we also have $\lambda_0 \approx \lambda_a \approx \mu$. Hence after integration by parts we get that their kernels satisfy bounds of the form $$\begin{aligned}
& |K_{\mu,\nu} (t_1, r_1,\omega_1, t_2, r_2,\omega_2)|\\
\lesssim &\ln (\nu^{-1}\mu) \mu^{\rho_1-\rho_2} \nu (|r_1-r_2|\nu + 1)^{-N} \mu^2 (|\omega_1-\omega_2|\mu + 1)^{-N}(|t_1-t_2|\mu + 1)^{-N} \\
\lesssim & |t_1-t_2|^{-1 + \frac{1}{p_1} -\frac{1}{p_2}} \ln (\nu^{-1}
\mu) \mu^{\frac{3}{q_1}-\frac{3}{q_2}} \nu (|r_1-r_2|\nu + 1)^{-N} \mu^2
(|\omega_1-\omega_2|\mu + 1)^{-N}\end{aligned}$$ for all nonnegative integers $N$. Then follows after summation.
[MA]{} L. Andersson and P. Blue: [*Hidden symmetries and decay for the wave equation on the Kerr spacetime*]{}, arXiv:0908.2265v1 P. Brenner: [*On $L_p-L_{p'}$ estimates for the wave-equation*]{}. Math. Z. [**145**]{} (1975), 251–254. S. Chandrasekhar: [*The Mathematical Theory of Black Holes*]{} Oxford Classic Texts in the Physical Sciences. Oxford University Press, 1992 M. Dafermos and I. Rodnianski: [*A proof of the uniform boundedness of solutions to the wave equation on slowly rotating Kerr backgrounds*]{}, arXiv:0805.4309v1 M. Dafermos and I. Rodnianski: [*Lectures on black holes and linear waves*]{}, arXiv:0811.0354v1 F. Finster, N. Kamran, J. Smoller, S. T. Yau: [*Decay of solutions of the wave equation in Kerr geometry*]{}. Comm. Math. Phys. [**264**]{} (2006), 465-503. F. Finster, J. Smoller: [*A Time Independent Energy Estimate for Outgoing Scalar Waves in the Kerr Geometry*]{}, J. Hyperbolic Differ. Equ. [**5**]{} (2008), no. 1, 221-255 J. Ginibre and G. Velo: [*Generalized Strichartz inequalities for the wave equation*]{}. J. Funct. Anal. [**133**]{} (1995), 50–68. S. W. Hawking and G. F. R. Ellis: [*The large scale structure of space-time*]{}. Cambridge Monographs on Mathematical Physics, No. 1. London, New York: Cambridge University Press 1973. L. Kapitanski: [*Some generalizations of the Strichartz-Brenner inequality*]{}. Leningrad Math. J. [**1**]{} (1990), 693–726. M. Keel and T. Tao: [*Endpoint Strichartz estimates*]{}. Amer. J. Math. [**120**]{} (1998), 955–980. H. Lindblad and C. D. Sogge: [*On existence and scattering with minimal regularity for semilinear wave equations*]{}. J. Funct. Anal. [**130**]{} (1995), 357–426. J. Marzuola, J. Metcalfe, D. Tataru, and M. Tohaneanu: [*Strichartz estimates on Schwarzschild black hole backgrounds*]{}, arXiv:0802.3942. Accepted to Communications in Mathematical Physics, 2009 J. Metcalfe and D. Tataru: [*Global parametrices and dispersive estimates for variable coefficient wave equations*]{}, arXiv:0707.1191. G. Mockenhaupt, A. Seeger, and C. D. Sogge: [*Local smoothing of Fourier integral operators and Carleson-Sjölin estimates*]{}. J. Amer. Math. Soc. [**6**]{} (1993), 65–130. H. Pecher: [*Nonlinear small data scattering for the wave and Klein-Gordon equations*]{}. Math. Z. [**185**]{} (1984), 261–270. H. F. Smith: [*A parametrix construction for wave equations with $C\sp{1,1}$ coefficients*]{}. Ann. Inst. Fourier (Grenoble) [**48**]{} (1998), 797–835. R. S. Strichartz: [*A priori estimates for the wave equation and some applications*]{}. J. Funct. Anal. [**5**]{} (1970), 218–235. R. S. Strichartz: [*Restrictions of Fourier transforms to quadratic surfaces and decay of solutions of wave equations*]{}. Duke Math. J. [**44**]{} (1977), 705–714. D. Tataru: [*Strichartz estimates for operators with nonsmooth coefficients and the nonlinear wave equation*]{}. Amer. J. Math. [**122**]{} (2000), 349–376. D. Tataru: [*Strichartz estimates for second order hyperbolic operators with nonsmooth coefficients II*]{}. Amer. J. Math. [**123**]{} (2001), 385–423. D. Tataru: [*Strichartz estimates for second order hyperbolic operators with nonsmooth coefficients III*]{}. J. Amer. Math. Soc. [**15**]{} (2002), 419–442. D. Tataru and M. Tohaneanu: [*Local energy estimate on Kerr black hole backgrounds*]{}, arXiv:0810.5766v2
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We show that there is no algorithm deciding whether the maximal residually free quotient of a given finitely presented group is finitely presentable or not.
Given a finitely generated subgroup $G$ of a finite product of limit groups, we discuss the possibility of finding an explicit set of defining equations (i.e. of expressing $G$ as the maximal residually free quotient of an explicit finitely presented group).
author:
- 'Vincent Guirardel, Gilbert Levitt'
bibliography:
- 'published.bib'
- 'unpublished.bib'
date:
title: Computing equations for residually free groups
---
Introduction
============
Any countable group $G$ has a largest residually free quotient $RF(G)$, equal to $G/\bigcap_{f\in {{\mathcal {H}}}}\ker f$ where ${{\mathcal {H}}}$ is the set of all homomorphisms from $G$ to a non-abelian free group ${{\mathbb {F}}}$.
In the language of [@BMR_algebraicI], if $R$ is a finite set of group equations on a finite set of variables $S$, then $G=RF({\langle S\mid R \rangle})$ is the *coordinate group* of the variety defined by the system of equations $R$. We say that $R$ is a *set of defining equations* of $G$ over $S$. Equational noetherianness of free groups implies that any finitely generated residually free group $G$ has a (finite) set of defining equations [@BMR_algebraicI].
On the other hand, any finitely generated residually free group embeds into a finite product of limit groups (also known as finitely generated fully residually free groups), which correspond to the *irreducible components* of the variety defined by $R$ [@BMR_algebraicI; @KhMy_irreducible2; @Sela_diophantine1]. Conversely, any subgroup of a finite product of limit groups is residually free.
This gives three possibilites to define a finitely generated residually free group $G$ in an explicit way:
1. give a finite presentation of $G$ (if $G$ is finitely presented);
2. give a set of defining equations of $G$: write $G=RF({\langle S\mid R \rangle})$, with $S$ and $R$ finite;
3. write $G$ as the subgroup of $L_1\times\dots\times L_n$ generated by a finite subset $S$, where $L_1,\dots,L_n$ are limit groups given by some finite presentations.
We investigate the algorithmic possibility to go back and forth between these ways of defining $G$.
One can go from 2 to 3: given a set of defining equations of $G$, one can find an explicit embedding into some product of limit groups [@KhMy_irreducible2; @KhMy_effective; @BHMS_fprf; @GrWi_enumerating].
Conversely, if $G$ is given as a subgroup of a product of limit groups, *and if one knows that $G$ is finitely presented,* one can compute a presentation of $G$ [@BHMS_fprf]. Obviously, a finite presentation is a set of defining equations.
Since residually free groups are not always finitely presented, we investigate the following question:
Let $L=L_1\times\dots\times L_n$ be a product of limit groups. Let $G$ be the subgroup generated by a finite subset $S{\subset}L $. Can one algorithmically find a finite set of defining equations for $G$, [i.e. ]{}find a finite presentation $\langle S \mid R\rangle$ such that $G=RF(\langle S \mid R\rangle)$?
We will prove that this question has a negative answer. On the other hand, we introduce a closely related notion which has better algorithmic properties.
Let $RF_{na}(G)$ be the quotient $G/\bigcap_{f\in {{\mathcal {H}}}_{na}}\ker f$ where ${{\mathcal {H}}}_{na}$ is the set of all homomorphisms from $G$ to ${{\mathbb {F}}}$ with *non-abelian image*. Of course, $RF_{na}(G)$ is a quotient of $RF(G)$, which forgets the information about morphisms to ${{\mathbb {Z}}}$. In fact (Lemma \[lem\_centre\]), it is the quotient of $ RF(G)$ by its center.
We say that $G$ is a *residually non-abelian free* group if $G=RF_{na}(G)$, [i.e. ]{}if every non-trivial element of $G$ survives in a non-abelian free quotient of $G$; equivalently, $G$ is residually non-abelian free if and only if $G$ is residually free and has trivial center. Given a residually non-abelian free group $G$, we say that $R$ is a *set of $na$-equations* of $G$ over $S$ if $G=RF_{na}({\langle S\mid R \rangle})$.
We write $Z(G)$ for the center of $G$, and $b_1(G)$ for the torsion-free rank of $H_1(G,{{\mathbb {Z}}})$.
\[thm1\]
- There is an algorithm which takes as input presentations of limit groups $L_1,\dots,L_n$, and a finite subset $S
\subset L_1\times\dots\times L_n$, and which computes a finite set of $na$-equations for $G/Z(G)=RF_{na}(G)$, where $G=\langle S\rangle$.
- One can compute a finite set of defining equations for $G=\langle S\rangle$ if and only if one can compute $b_1(G)$.
Since there is no algorithm computing $b_1(\langle S\rangle)$ from $S{\subset}{{\mathbb {F}}}_2\times {{\mathbb {F}}}_2$ [@BrMi_structure], we deduce:
\[cor2\] There is no algorithm which takes as an input a finite subset $S{\subset}{{\mathbb {F}}}_2\times {{\mathbb {F}}}_2$ and computes a finite set of equations for $\langle S\rangle$.
We also investigate the possiblity to decide whether a residually free quotient is finitely presented. Using Theorem \[thm1\] and [@Grunewald_fp], we prove:
\[thm2\] There is no algorithm with takes as an input a finite group presentation ${\langle S\mid R \rangle}$, and which decides whether $RF({\langle S\mid R \rangle})$ is finitely presented.
The residually non-abelian free quotient $RF_{na}$
==================================================
We always denote by $G$ a finitely generated group, and by ${{\mathbb {F}}}$ a non-abelian free group.
$RF(G)$ is the quotient of $G$ by the intersection of the kernels of all morphisms $G\to {{\mathbb {F}}}$.
$RF_{na}(G)$ is the quotient of $G$ by the intersection of the kernels of all morphisms $G\to {{\mathbb {F}}}$ with non-abelian image.
One may view $RF(G)$ as the image of $G$ in ${{\mathbb {F}}}^{{\mathcal {H}}}$, where ${{\mathcal {H}}}$ is the set of all morphisms $G\to {{\mathbb {F}}}$, and $RF_{na}(G)$ as the image in ${{\mathbb {F}}}^{{{\mathcal {H}}}_{na}}$, where ${{\mathcal {H}}}_{na}$ is the set of all morphisms with non-abelian image.
Every homomorphism $G\to {{\mathbb {F}}}$ factors through $RF(G)$ (through $RF_{na}(G)$ if its image is not abelian). By definition, $G$ is residually free if and only if $G=RF(G)$, residually non-abelian free if and only if $G=RF_{na}(G)$.
\[lem\_centre\] There is an exact sequence $$1{\rightarrow}Z(RF(G)){\rightarrow}RF(G){\rightarrow}RF_{na}(G){\rightarrow}1 .$$
In particular, $G$ is residually non-abelian free if and only if $G$ is residually free and $Z(G)=1$. If $G$ is a non-abelian limit group, it has trivial center and $RF_{na}(G)=RF(G)=G$.
Let $H=RF(G)$. Consider $a\in Z(H)$ and $f:H{\rightarrow}{{\mathbb {F}}}$ with $f(a)\neq 1$. The image of $f$ centralizes $f(a)$, so is abelian by commutative transitivity of ${{\mathbb {F}}}$. Thus $a$ has trivial image in $RF_{na}(H)=RF_{na}(G)$.
Conversely, consider $a\in H\setminus Z(H)$, and $b\in H$ with $[a,b]\neq 1$. There exists $f:H{\rightarrow}{{\mathbb {F}}}$ such that $f([a,b])\neq 1$. Then $f(H)$ is non-abelian, and $f(a)\neq 1$. This means that the image of $a$ in $RF_{na}(G)$ is non-trivial.
Any epimorphism $f:G{\rightarrow}H$ induces epimorphisms $f_{RF}:RF(G){\rightarrow}RF(H)$ and $f_{na}:RF_{na}(G){\rightarrow}RF_{na}(H)$.
\[lem\_b1\] Let $f:G{\rightarrow}H$ be an epimorphism. Then $f_{RF}:RF(G){\rightarrow}RF(H)$ is an isomorphism if and only if $f_{na}:RF_{na}(G){\rightarrow}RF_{na}(H)$ is an isomorphism and $b_1(G)=b_1(H)$.
Note that $f_{RF}$ (resp. $f_{na}$) is an isomorphism if and only if any morphism $G{\rightarrow}{{\mathbb {F}}}$ (resp. any such morphism with non-abelian image) factors through $f$. The lemma then follows from the fact that the embedding ${{\mathrm{Hom}}}(H, {{\mathbb {Z}}}){\hookrightarrow}{{\mathrm{Hom}}}(G, {{\mathbb {Z}}})$ induced by $f$ is onto if and only if $b_1(G)=b_1(H)$.
Given a product $L_1\times\dots\times L_n$, we denote by $p_i$ the projection onto $L_i$.
\[lem\_nonab\] Let $G\subset L=L_1\times\dots\times L_n$ with $L_i$ a limit group. Let $I\subset\{1,\dots,n\}$ be the set of indices such that $p_i(G)$ is abelian. Then $RF_{na}(G)$ is the image of $G$ in $L'=\prod_{i\notin I} L_i$ (viewed as a quotient of $L_1\times \dots\times L_n$).
Note that $G=RF(G)$. An element $(x_1,\dots,x_n)\in G$ is in $Z(G)$ if and only if $x_i$ is central in $p_i(G)$ for every $i$. Since $p_i(G)$ is abelian or has trivial center, $Z(G)$ is the kernel of the natural projection $L\to L'$. The result follows from Lemma \[lem\_centre\].
\[lem\_fp\] $RF(G)$ is finitely presented if and only if $RF_{na}(G)$ is.
If $H$ is any residually free group, the abelianization map $H\to H_{ab}$ is injective on $Z(H)$ since any element of $Z(H)$ survives in some free quotient of $H$, which has to be cyclic (see [@BHMS_fprf Lemma 6.2]). In particular, $Z(H)$ is finitely generated if $H$ is. Applying this to $H=R(G)$, the exact sequence of Lemma \[lem\_centre\] gives the required result.
Proof of the theorems
=====================
Let $S$ be a finite set of elements in a group. We define $S_0=S\cup\{1\}$. If $R,R'$ are sets of words on $S\cup S{^{-1}}$, then $R^{S_0}$ is the set of all words obtained by conjugating elements of $R$ by elements of $S_0$, and $[R^{S_0},R']$ is the set of all words obtained as commutators of words in $R^{S_0}$ and words in $R'$.
\[prop\_eqna\] Let $A_1,\dots,A_n$ be arbitrary groups, with $n\ge2$. Let $G\subset A_1\times\dots \times A_n$ be generated by $S=\{s_1,\dots,s_k\}$. Let $p_i:G{\rightarrow}A_i$ be the projection. Assume that $p_i(G)=RF_{na}({\langle S\mid R_i \rangle})$ for some finite set of relators $R_i$.
Then the set $$\Tilde R= [R_n^{S_0},[R_{n-1}^{S_0},\dots [R_3^{S_0}, [R_2^{S_0},R_1]]\dots]]$$ is a finite set of $na$-equations of $RF_{na}(G)$ over $S$, [i.e. ]{}$RF_{na}(G)=RF_{na}({\langle S\mid \Tilde R \rangle})$.
An equality such as $p_i(G)=RF_{na}({\langle S\mid R_i \rangle})$ means that there is an isomorphism commuting with the natural projections $F(S)\to p_i(G)$ and $F(S)\to RF_{na}({\langle S\mid R_i \rangle})$, where $F(S)$ denotes the free group on $S$.
Recall that a free group ${{\mathbb {F}}}$ is CSA: commutation is transitive on ${{\mathbb {F}}}\setminus\{1\}$, and maximal abelian subgroups are malnormal. In particular, if two non-trivial subgroups commute, then both are abelian. If $A, B$ are non-trivial subgroups of ${{\mathbb {F}}}$, and if $A$ commutes with $B,B^{x_1},\dots,B^{x_p}$ for elements $x_1,\dots ,x_p\in {{\mathbb {F}}}$, then ${\langle A,B,x_1,\dots ,x_p \rangle}$ is abelian.
We write $$\Tilde G={\langle S\mid \Tilde R \rangle}={\langle S\mid [ R_n^{S_0},[R_{n-1}^{S_0},\dots[R_2^{S_0},R_1]\dots]] \rangle}.$$ We always denote by $\varphi:F(S)\to{{\mathbb {F}}}$ a morphism with non-abelian image. We shall show that *such a $\varphi$ factors through $G$ if and only if it factors through $\tilde G$.* This implies the desired result $RF_{na}(G)=RF_{na}( \Tilde G )$: both groups are equal to the image of $F(S)$ in ${{\mathbb {F}}}^{{{\mathcal {H}}}_{na}}$, where ${{\mathcal {H}}}_{na}$ is the set of all $\varphi$’s which factor through $G$ and $\tilde G$.
We proceed by induction on $n$. We first claim that $\varphi$ is trivial on $\Tilde R$ if and only if it is trivial on some $R_i$. The if direction is clear. For the only if direction, observe that the image of $[R_{n-1}^{S_0},\dots [R_2^{S_0},R_1]\dots]$ commutes with all conjugates of $\varphi(R_n)$ by elements of $\varphi(F(S))$, so $R_n$ or $[R_{n-1}^{S_0},\dots[R_2^{S_0}, R_1]\dots]$ has trivial image. The claim follows by induction.
Now suppose that $\varphi$ factors through $\Tilde G$. Then $\varphi$ kills $\Tilde R$, hence some $R_i$. It follows that $\varphi$ factors through $p_i(G)$, hence through $G$.
Conversely, suppose that $\varphi$ factors through $f: G\to {{\mathbb {F}}}$. Consider the intersection of $G$ with the kernel of $p_n:G\to A_n$ and the kernel of $p_{1,\dots,n-1}:G\to A_1\times \dots\times A_{n-1}$. These are commuting normal subgroups of $G$. If both have non-trivial image in ${{\mathbb {F}}}$, the CSA property implies that the image of $f$ is abelian, a contradiction. We deduce that $f$ factors through $p_n $ or through $p_{1,\dots,n-1}$, and by induction that it factors through some $p_i$. Thus $\varphi$ kills $R_i$, hence $\Tilde R$ as required.
\[Proof of Theorem \[thm1\]\]
Let $L_1, \dots, L_n$ and $G=\langle S\rangle$ be as in Theorem \[thm1\]. Using a solution of the word problem in a limit group, one can find the indices $i$ for which $p_i(G){\subset}L_i$ is abelian (this amounts to checking whether the elements of $p_i(S)$ commute).
First assume that no $p_i(G)$ is abelian. As pointed out in [@BHMS_fprf Lemma 7.5], one deduces from [@Wilton_hall] an algorithm yielding a finite presentation $ \langle S \mid R_i\rangle$ of $p_i(G)$. Since $p_i(G)$ is not abelian, one has $p_i(G)=RF_{na}(\langle S \mid R_i\rangle)$, and Proposition \[prop\_eqna\] yields a finite set of na-equations for $RF_{na}(G)$ over $S$ (if $n=1$, then $RF_{na}(G)=p_1(G)$). If some $p_i(G)$’s are abelian, we simply replace $G$ by its image in $L'$ as in Lemma \[lem\_nonab\]. This proves the first assertion of the theorem.
Now suppose that $b_1(G)$ is known. We want a finite set $R$ such that $RF(G)=RF({\langle S\mid R \rangle})$. If $n=1$, then $G$ is a subgroup of the limit group $L_1$, and one can find a finite presentation of $G$ as explained above. So assume $n\geq 2$. Consider the finite presentation $\Tilde G={\langle S\mid \Tilde R \rangle}$ given by Proposition \[prop\_eqna\], so that $RF_{na}(\Tilde G)=RF_{na}(G)$.
We claim that $G$ is a quotient of $\Tilde G$. To see this, we consider an $x\in F(S)$ which is trivial in $\Tilde G$ and we prove that it is trivial in $G$. If not, residual freeness of $G$ implies that $x$ survives under a morphism $\varphi:F(S)\to {{\mathbb {F}}}$ which factors through $G$. If $\varphi$ has non-abelian image, it factors through $RF_{na}(G)=RF_{na}(\Tilde G)$, hence through $\Tilde G$, contradicting the triviality of $x$ in $\Tilde G$. If the image is abelian, $\varphi$ also factors through $\Tilde G$ because all relators in $\Tilde R$ are commutators.
Since $\Tilde R$ is finite, we can compute $b_1(\Tilde G)$. If $b_1(\Tilde G)=b_1(G)$, we are done by Lemma \[lem\_b1\] since $G$ is a quotient of $\Tilde G$. If $b_1(\Tilde G)>b_1(G)$, we enumerate all trivial words of $G$ (using an enumeration of trivial words in each $p_i(G)$), and we add them to the presentation of $\Tilde G$ one by one. We compute $b_1$ after each addition, and we stop when we reach the known value $b_1(G)$.
Conversely, if we have a finite set of defining equations for $G$, so that $G=RF(\langle S \mid R \rangle)$, we can compute $b_1(\langle S \mid R \rangle)$, which equals $b_1(G)$ by Lemma \[lem\_b1\].
There is no algorithm with takes as input a finite group presentation ${\langle S\mid \Tilde R \rangle}$, and which decides whether $ RF({\langle S\mid \Tilde R \rangle})$ is finitely presented.
Given a finite set $S\subset {{\mathbb {F}}}_2\times {{\mathbb {F}}}_2$, Theorem \[thm1\] provides a finite set $\Tilde R$ such that $RF_{na}(\langle S\rangle)=RF_{na}({\langle S\mid \Tilde R \rangle})$. Using Lemma \[lem\_fp\], we see that finite presentability of $RF ({\langle S\mid \Tilde R \rangle})$ is equivalent to that of $RF_{na}({\langle S\mid \Tilde R \rangle})$, hence to that of $RF (\langle S\rangle)=\langle S\rangle$. But it follows from [@Grunewald_fp] that there is no algorithm which decides, given a finite set $S\subset {{\mathbb {F}}}_2\times {{\mathbb {F}}}_2$, whether $\langle S\rangle$ is finitely presented.
Vincent Guirardel\
Institut de Mathématiques de Toulouse\
Université de Toulouse et CNRS (UMR 5219)\
118 route de Narbonne\
F-31062 Toulouse cedex 9\
France.\
*e-mail:*`[email protected]`\
Gilbert Levitt\
Laboratoire de Mathématiques Nicolas Oresme\
Université de Caen et CNRS (UMR 6139)\
BP 5186\
F-14032 Caen Cedex\
France\
*e-mail:*`[email protected]`\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We address the following decision problem. Given a numeration system $U$ and a $U$-recognizable set $X\subseteq\mathbb{N}$, i.e. the set of its greedy $U$-representations is recognized by a finite automaton, decide whether or not $X$ is ultimately periodic. We prove that this problem is decidable for a large class of numeration systems built on linearly recurrent sequences. Based on arithmetical considerations about the recurrence equation and on $p$-adic methods, the DFA given as input provides a bound on the admissible periods to test.'
address:
- '$^1$ University of Liège, Department of Mathematics, Allée de la découverte 12 (B37), B-4000 Liège, Belgium'
- '$^2$ Department of Mathematics, Hofstra University, Hempstead, NY, USA'
author:
- 'É. Charlier$^1$, A. Massuir$^1$, M. Rigo$^1$'
- 'E. Rowland$^2$'
title: Ultimate periodicity problem for linear numeration systems
---
Introduction
============
Let us first recall the general setting of linear numeration systems that are used to represent, in a greedy way, non-negative integers by words over a finite alphabet of digits. See, for instance, [@Fraenkel]. Let ${\mathbb{N}}= \{0, 1, 2, \dots\}$.
\[def:numsys\] A [*numeration system*]{} is given by an increasing sequence $U=(U_i)_{i\ge 0}$ of integers such that $U_0=1$ and $C_U:=\sup_{i\ge 0} \lceil \frac{U_{i+1}}{U_i}\rceil$ is finite. Let $A_U=\{0,\ldots,C_U-1\}$ be the canonical alphabet of digits. The [*greedy $U$-representation*]{} of a positive integer $n$ is the unique finite word $\operatorname{rep}_U(n)=w_\ell\cdots w_0$ over $A_U$ satisfying $$n=\sum_{i=0}^\ell w_i\, U_i,\ w_\ell\neq 0 \text{ and } \sum_{i=0}^t w_i\,
U_i<U_{t+1},\ t=0,\ldots,\ell.$$ We set $\operatorname{rep}_U(0)$ to be the empty word $\varepsilon$. A set $X\subseteq\mathbb{N}$ of integers is [*$U$-recognizable*]{} if the language $\operatorname{rep}_U(X)$ over $A_U$ is regular (i.e. accepted by a finite automaton).
Recognizable sets of integers are considered as particularly simple because membership can be decided by a deterministic finite automaton in linear time with respect to the length of the representation. It is well-known that such a property for a subset of $\mathbb{N}$ depends on the choice of the numeration system. For a survey on integer base systems, see [@BHMV]. For generalized numeration systems, see [@FLANS].
If $x=x_\ell\cdots
x_0$ is a word over an alphabet of integers, then the [*$U$-numerical value*]{} of $x$ is $$\operatorname{val}_U(x)=\sum_{i=0}^\ell
x_i\, U_i.$$
From the point of view of formal languages, it is quite desirable that $\operatorname{rep}_U(\mathbb{N})$ is regular ; we want to be able to check whether or not a word is a valid greedy $U$-representation. This implies that $U$ must satisfies a linear recurrence relation. See, for instance, [@Sha] or [@cant Prop. 3.1.5].
A numeration system $U$ is said to be [*linear*]{} if it ultimately satisfies a homogeneous linear recurrence relation with integer coefficients. There exist $k\ge 1$, $a_{k-1},\ldots,a_0\in\mathbb{Z}$ such that $a_0\neq 0$ and $N\ge 0$ such that for all $i\ge N$, $$\label{eq:linrec}
U_{i+k}=a_{k-1} U_{i+k-1}+\cdots +a_0 U_i.$$ The polynomial $X^N(X^k-a_{k-1}X^{k-1}-\cdots -a_0)$ is called the [*characteristic polynomial*]{} of the system.
The regularity of $\operatorname{rep}_U(\mathbb{N})$ is also important for another reason. The language $\operatorname{rep}_U(\mathbb{N})$ is regular if and only if every ultimately periodic set of integers is $U$-recognizable [@LR Thm. 4]. In particular, as recalled in Proposition \[pro:up\], if an ultimately periodic set $X$ is given, then a DFA accepting $\operatorname{rep}_U(X)$ can effectively be obtained.
In this paper, we address the following decidability question. Our aim is to prove that this problem is decidable for a large class of numeration systems.
\[pb\] Given a linear numeration system $U$ and a (deterministic) finite automaton $\mathcal{A}$ whose accepted language is contained in the numeration language $\operatorname{rep}_U(\mathbb{N})$, decide whether the subset $X$ of $\mathbb{N}$ that is recognized by $\mathcal{A}$ is ultimately periodic, i.e. whether or not $X$ is a finite union of arithmetic progressions (along a finite set).
This question about ultimately periodic sets is motivated by the celebrated theorem of Cobham. Let $p,q\ge 2$ be integers. If $p$ and $q$ are multiplicatively independent, i.e. $\frac{\log(p)}{\log(q)}$ is irrational, then the ultimately periodic sets are the only sets that are both $p$-recognizable and $q$-recognizable [@cobham]. These are exactly the sets definable by a first-order formula in the Presburger arithmetic $\langle\mathbb{N},+\rangle$. Cobham’s result has been extended to various settings ; see [@Durand; @Mitrofanov] for an application to morphic words. See [@DurRig] for a survey.
In this paper, we write greedy $U$-representations with most significant digit first (MSDF convention): the leftmost digit is associated with the largest $U_\ell$ occurring in the decomposition. Considering least significant digit first would not affect decidability (a language is regular if and only if its reversal is) but this could have some importance in terms of complexity issues not discussed here.
**What is known.** Let us quickly review cases where the decision problem is known to be decidable. Relying on number theoretic results, the problem was first solved by Honkala for integer base systems [@Honkala]. An alternative approach bounding the syntactic complexity of ultimately periodic sets of integers written in base $b$ was studied in [@LRRV]. Recently a deep analysis of the structure of the automata accepting ultimately periodic sets has lead to an efficient decision procedure for integer base systems [@MSaka; @BMMR; @Marsault2019]. An integer base system is a particular case of a Pisot system, i.e. a linear numeration system whose characteristic polynomial is the minimal polynomial of a Pisot number (an algebraic integer larger than $1$ whose conjugates all have modulus less than one). For these systems, one can make use of first-order logic and the decidable extension $\langle \mathbb{N},+,V_U\rangle$ of Presburger arithmetic [@BH]. For an integer base $p$, $V_p(n)$ is the largest power of $p$ dividing $n$. A typical example of Pisot system is given by the Zeckendorf system based on the Fibonacci sequence $1,2,3,5,8,\ldots$. Given a $U$-recognizable set $X$, there exists a first-order formula $\varphi(n)$ in $\langle \mathbb{N},+,V_U\rangle$ describing $X$. The formula $$(\exists N)(\exists p)(\forall n\ge N)(\varphi(n)\Leftrightarrow \varphi(n+p))$$ thus expresses when $X$ is ultimately periodic, $N$ being a preperiod and $p$ a period of $X$. The logic formalism can be applied to systems such that the addition is $U$-recognizable by an automaton, i.e. the set $\{(x,y,z)\in\mathbb{N}^3:x+y=z\}$ is $U$-recognizable. This is the case for Pisot systems [@Frou].
When addition is not known to be $U$-recognizable, other techniques must be sought. Hence the problem was also shown to be decidable for some non-Pisot linear numeration systems satisfying a gap condition $\lim_{i\to+\infty} U_{i+1}-U_i=+\infty$ and a more technical condition $\lim_{m\to+\infty} N_U(m)=+\infty$ where $N_U(m)$ is the number of residue classes that appear infinitely often in the sequence $(U_i\bmod{m})_{i\ge 0}$ ; see [@BCFR]. An example of such a system is given by the relation $U_i=3\, U_{i-1}+2\, U_{i-2}+3\, U_{i-3}$ [@Frougny97]. For extra pointers to the literature (such as an extension to a multidimensional setting), the reader can follow the introduction in [@BCFR].
**Our contribution.** In view of the above summary, we are looking for a decision procedure that may be applied to non-Pisot linear numeration systems such that $N_U(m)\not\to \infty$ when $m$ tends to infinity. Hence we want to take into account systems where we are not able to apply a decision procedure based on first-order logic nor on the technique from [@BCFR]. We follow Honkala’s original scheme: if a DFA $\mathcal{A}$ is given as input (the question being whether the corresponding recognized subset of $\mathbb{N}$ is ultimately periodic), the number of states of $\mathcal{A}$ should provide an upper bound on the admissible preperiods and periods. If there is a finite number of such pairs to test, then we build a DFA $\mathcal{A}_{N,p}$ for each pair $(N,p)$ and one can test whether or not two automata $\mathcal{A}$ and $\mathcal{A}_{N,p}$ accept the same language. This provides us with a decision procedure. Roughly speaking, if the given DFA is “small”, then it cannot accept an ultimately periodic set with a minimal period being “overly complicated”, i.e. “quite large”.
\[exa:ppp\] Here is an example of a numeration system based on a Parry (the $\beta$-expansion of $1$ is finite or ultimately periodic, see [@cant Chap. 2]) non-Pisot number $\beta$: $$U_{i+4}=2\, U_{i+3}+2\, U_{i+2}+2\, U_i.$$ Indeed, the largest root $\beta$ of the characteristic polynomial is roughly $2.804$, and $-1.134$ is another root of modulus larger than one. With the initial conditions $1,3,9,23$, $\operatorname{rep}_U(\mathbb{N})$ is the regular language over $\{0,1,2\}$ of words avoiding factors $2202$, $221$ and $222$. For details, see [@cant Ex. 2.3.37]. When $m$ is a power of $2$, there is a unique congruence class visited infinitely often by the sequence $(U_i\bmod{m})_{i\ge 0}$ because $U_i\equiv 0\pmod{2^r}$ for large enough $i$. For such an example, $N_U(m)$ does not tend to infinity and thus the previously known decision procedures may not be applied. This is a perfect candidate for which no decision procedures are known.
This paper is organized as follows. In Section \[sec2\], we make clear our assumptions on the numeration system. In Section \[sec3\], we collect several known results on periodic sets and $U$-representations. In particular, we relate the length of the $U$-representation an integer to its value. The core of the paper is made of Section \[sec4\] where we discuss cases to bound the admissible periods. In particular, we consider two kinds of prime factors of the admissible periods: those that divide all the coefficients of the recurrence and those that don’t, see . In Section \[sec5\], we apply the discussion of the previous section. First, we obtain a decision procedure when the gcd of the coefficients of the recurrence relation is $1$, see Theorem \[thm:main1\]. This extends the scope of results from [@BCFR]. On the other hand, if there exist primes dividing all the coefficients, our approach heavily relies on quite general arithmetic properties of linear recurrence relations. It has therefore inherent limitations because of notoriously difficult results in $p$-adic analysis such as finding bounds on the growth rate of blocks of zeroes in $p$-adic numbers of a special logarithmic form. We discuss the question and give illustrations of these $p$-adic techniques in Section \[sec:ap\]. The paper ends with some concluding remarks.
Our setting {#sec2}
===========
We have minimal assumptions on the considered linear numeration system $U$.
- $\mathbb{N}$ is $U$-recognizable ;
- There are arbitrarily large gaps between consecutive terms: $$\limsup_{i\to+\infty}(U_{i+1}-U_i)=+\infty.$$
- The gap sequence $(U_{i+1}-U_i)_{i\ge 0}$ is ultimately non-decreasing: there exists $N\ge 0$ such that for all $i\ge N$, $$U_{i+1}-U_i \le U_{i+2}-U_{i+1}.$$
Let us make a few comments. (H1) gives sense and meaning to our decision problem ; under that assumption, ultimately periodic sets are $U$-recognizable. As recalled in the introduction, it is a well-known result that (H1) implies that the numeration system $(U_i)_{i\ge 0}$ must satisfy a linear recurrence relation with integer coefficients. The assumptions (H2) and (H3) imply that $\lim_{i\to+\infty}(U_{i+1}-U_i)=+\infty$. However, in many cases, even if $\lim_{i\to+\infty}(U_{i+1}-U_i)=+\infty$, the gap sequence may decrease from time to time.
The main reason why we introduce (H3) is the following one. Let $1 0^j w$ be a greedy $U$-representation for some $j\ge 0$. Assume (H3) and $i=|w|+\ell\ge N$. Then for all $\ell'\ge \ell$, $10^{\ell'} w$ is a greedy $U$-representation as well. Indeed, if $n$ is a non-negative integer such that $U_i+n<U_{i+1}$, then $U_{i+1}+n=U_{i+1}-U_i+U_i+n\le U_{i+2}-U_{i+1}+U_i+n<U_{i+2}$. Hence $U_{i'}+n<U_{i'+1}$ for all $i'\ge i$, meaning that as soon as the greediness property is fulfilled, one can shift the leading $1$ at every larger index. This is not always the case, as seen in Example \[exa:noth2\].
This property will be used in Corollary \[cor:h1h2\], which in turn will be crucial in the proofs of Propositions \[pro:cas1\] and \[pro:cas2a\] as well as Theorem \[the:main\], where we construct $U$-representations with leading $1$’s in convenient positions.
Note that Example \[exa:ppp\] satisfies the above assumptions.
\[exa:toy\] Our toy example that will be treated all along the paper is given by the recurrence $U_{i+3}=12\, U_{i+2}+6\, U_{i+1}+12\, U_i$. Even though the system is associated with a Pisot number, it is still interesting because $N_U(m)$ does not tend to infinity (so we cannot follow the decision procedure from [@BCFR]) and the gcd of the coefficients of the recurrence is larger than $1$. Let $r\ge 1$. If $m$ is a power of $2$ or $3$, then $U_i\equiv 0\pmod{2^r}$ (resp. $U_i\equiv 0\pmod{3^r}$) for large enough $i$. By taking the initial conditions $1,13,163$, the language of greedy $U$-representations is regular. For the reader aware of $\beta$-numeration systems, let us mention that this choice of initial conditions corresponds to the Bertrand initial conditions, in which case the language $\operatorname{rep}_U(\mathbb{N})$ is equal to the set of factors (with no leading zeroes) occurring in the $\beta$-expansions of real numbers where $\beta$ is the dominant root of the characteristic polynomial $X^3-12X^2-6X-12$ of the recurrence relation of the system $U$ [@Bertrand-Mathis:1989].
Some classical lemmas {#sec3}
=====================
A set $X\subseteq\mathbb{N}$ is [*ultimately periodic*]{} if its characteristic sequence $\mathbf{1}_X\in\{0,1\}^\mathbb{N}$ is of the form $uv^\omega$ where $u,v$ are two finite words over $\{0,1\}$. It is assumed that $u,v$ are chosen of minimal length. Hence the [*period*]{} of $X$ denoted by $\pi_X$ is the length $|v|$ and its preperiod is the length $|u|$. We say that $X$ is [*(purely) periodic*]{} whenever the preperiod is zero. The following lemma is a simple consequence of the minimality of the period chosen to represent an ultimately periodic set.
\[lem:per\] Let $X\subseteq\mathbb{N}$ be an ultimately periodic set of period $\pi_X$ and let $i,j$ be integers greater than or equal to the preperiod of $X$. If $i\not\equiv j\pmod{
\pi_X}$ then there exists $r<\pi_X$ such that either $i+r\in X$ and $j+r\not\in X$ or, $i+r\not\in X$ and $j+r\in X$.
Our assumption (H2) permits us to extend greedy $U$-representations with some extra leading digits. See [@BCFR Lemma 7] for a proof.
\[lem:condlim\] Let $U$ be a numeration system satisfying (H2). Then for all $i\ge 0$ and all $L\ge i$, there exists $\ell\ge L$ such that $$1 0^{\ell-{\lvert{\operatorname{rep}_U(t)}\rvert}}\operatorname{rep}_U(t),\ t=0,\ldots,U_i-1$$ are greedy $U$-representations. Otherwise stated, if $w$ is a greedy $U$-representation, then there exist arbitrarily large $r$ such that the word $10^{r}w$ is also a greedy $U$-representation.
When $\mathbb{N}$ is $U$-recognizable, using a pumping-like argument, we can give an upper bound on the number of zeroes to be inserted.
\[cor:h1h2\] Let $U$ be a numeration system satisfying (H1) and (H2). Then there exists an integer constant $C> 0$ such that if $w$ is a greedy $U$-representation then, for some $\ell<C$, $10^\ell w$ is also a greedy $U$-representation. If furthermore $U$ satisfies (H3) then $10^{\ell'}w$ is greedy for all $\ell'\ge \ell$.
By assumption (H1), there exists a DFA, say with $C$ states, accepting the language $\operatorname{rep}_U(\mathbb{N})$. Let $w$ be a greedy $U$-representation. Then from Lemma \[lem:condlim\], there exists $r\ge C$ such that $10^rw\in \operatorname{rep}_U(\mathbb{N})$. The path of label $10^rw$ starting from the initial state is accepting. Since $r\ge C$, a state is visited at least twice when reading the block $0^r$. Thus there exists an accepting path of label $10^\ell w$ with $\ell<C$.
We now turn to the special case. We proceed by induction. If $10^\ell w$ is a greedy $U$-representation, then $$\operatorname{val}_U(10^\ell w)=U_{\ell+|w|}+\operatorname{val}_U(w)<U_{\ell+|w|+1}.$$ Under (H3), $$U_{\ell+|w|+1}-U_{\ell+|w|}\le U_{\ell+|w|+2}-U_{\ell+|w|+1},$$ adding together both sides of the two inequalities leads to $U_{\ell+|w|+1}+\operatorname{val}_U(w)<U_{\ell+|w|+2}$ meaning that $10^{\ell+1}w$ is a greedy $U$-representation.
\[exa:noth2\] The sequence $1,2,4,5,16,17,64,65,\ldots$ is a solution of the linear recurrence $U_{i+4}=5U_{i+2}-4U_i$ but it does not satisfy (H3). The property stated in the last part of Corollary \[cor:h1h2\] does not hold: only some shifts to the left of the leading coefficient $1$ lead to valid greedy expansions. The word $1001$ is the greedy representation of $6$ but for all $t\ge 1$, $1(00)^t1001$ is not a greedy representation.
\[exa:noth2-newh2ok\] The sequence $1,2,3,4,8,12,16,32,48,64,128,\ldots$ is a solution of the linear recurrence $U_{i+3}=4U_i$. The numeration language $0^*\operatorname{rep}_U(\mathbb{N})$ is the set of suffixes of $\{000,001,010,100\}^*$, hence (H1) holds. For all $i\ge 0$, $U_{i+1}-U_i=4^{\lfloor i/3\rfloor}$. Therefore, (H2) and (H3) are also verified.
We will also make use of the following folklore result. See, for instance, [@cant Prop. 3.1.9]. It relies on the fact that a linearly recurrent sequence is ultimately periodic modulo $Q$.
\[pro:up\] Let $Q,r\ge 0$. Let $A\subseteq\mathbb{N}$ be a finite alphabet. If $U$ is a linear numeration system, then $$\left\{w\in A^*\mid
\operatorname{val}_U(w)\in Q\, \mathbb{N}+r\right\}$$ is accepted by a DFA that can be effectively constructed. In particular, whenever $\mathbb{N}$ is $U$-recognizable, i.e. under (H1), then any ultimately periodic set is $U$-recognizable.
Under assumption (H1) the formal series $\sum_{i\ge 0}U_i\, X^i$ is $\mathbb{N}$-rational because $U_i$ is the number of words of length less than or equal to $i$ in the regular language $\operatorname{rep}_U(\mathbb{N})$. One can therefore make use of Soittola’s theorem [@SS Thm. 10.2]: The series is the merge of rational series with dominating eigenvalues and polynomials. We thus define the following quantities.
\[def:uT\] We introduce an integer $u$ and a real number ${\beta}$ depending only on the numeration system. From Soittola’s theorem, there exist an integer $u\ge 1$, real numbers $\beta_0,\ldots,\beta_{u-1}\ge 1$ and non-zero polynomials $P_0,\ldots,P_{u-1}$ such that for $r\in\{0,\ldots,u-1\}$ and large enough $i$, $$U_{ui+r}= P_r(i)\, \beta_r^i +Q_r(i)$$ where $\frac{Q_r(i)}{\beta_r^i}\to 0$ when $i\to\infty$. Since $(U_i)_{i\ge 0}$ is increasing, for $r<s<u$, for all $i$, we have $$U_{ui+r}<U_{ui+s}<U_{u(i+1)+r}.$$ By letting $i$ tend to infinity, this shows that we must have $\beta_0=\cdots=\beta_{u-1}$ that we denote by $\beta$ and $\deg(P_0)=\cdots=\deg(P_{u-1})$ that we denote by $d$. Otherwise stated, $U_{ui+r}\sim c_r i^{d} \beta^i$ for some constant $c_r$. Finally, let $T$ be such that $c_T=\max_{0\le r<u} c_r$. Otherwise stated, we highlight with $T$ a subsequence $(U_{ui+T})_{i\ge 0}$ with the maximal dominant coefficient.
Note that if a numeration system has a dominant root, i.e. the minimal recurrence relation satisfied by $(U_i)_{i\ge 0}$ has a unique root $\beta>1$, possibly with multiplicity greater than 1, of maximum modulus, then $u=1$.
\[lem:length\] With the notation of Definition \[def:uT\], if $\beta>1$ then there exists nonnegative constants $K$ and $L$ such that for all $n$, $${\lvert{\operatorname{rep}_U(n)}\rvert} < u \log_{\beta}(n) + K$$ and $${\lvert{\operatorname{rep}_U(n)}\rvert} > u \log_\beta(n)- u \log_\beta(P_T(\log_\beta(n)+K/u))-L.$$
This lemma shows that the length of the greedy $U$-representation of $n$ grows at most like $\log_{\beta^{1/u}}(n)$. If $P_T$ is a constant polynomial, the lower bound is of the form $u \log_\beta(n) +L'$ for some constant $L'$. From this result, we may express the weaker information (on ratios instead of differences) that ${\lvert{\operatorname{rep}_U(n)}\rvert}\sim u \log_\beta(n)$. The intricate form of the lower bound can be seen on an example such as $(U_i)_{i\ge 0}=(i^d\, 2^i)_{i\ge 0}$. In such a case, we get $\log_2(n)<{\lvert{\operatorname{rep}_U(n)}\rvert}+d\, \log_2({\lvert{\operatorname{rep}_U(n)}\rvert})$. Hence a lower bound for ${\lvert{\operatorname{rep}_U(n)}\rvert}$ is less than $\log_2(n)$.
We have ${\lvert{\operatorname{rep}_U(n)}\rvert}=\ell$ if and only if $U_{\ell-1}\le n <U_\ell$. We make use of Definition \[def:uT\] for $u$, $\beta$ and $T$. Let $j=\lfloor\frac{\ell-1-T}{u}\rfloor$. Since $U$ is increasing, $$U_{\ell-1}\ge U_{ju+T} =P_T(j) \beta^j + Q_T(j).$$ We get $$\log_{\beta}(n)
\ge \log_{\beta}(U_{\ell-1})
\ge j+\log_\beta(P_T(j))+\log_{\beta}\left(1 + \frac{Q_T(j)}{P_T(j)\beta^j}\right).$$ We also have $j>\frac{\ell-1-T}{u}-1\ge \frac{\ell-u}{u}-1=\frac{\ell}{u}-2$. From Definition \[def:uT\], we know that $P_T(i)>0$ for all $i$ and that $Q_T(i)$ is in $o(\beta^i)$. So there exists a constant $K\ge 0$ such that $$\begin{aligned}
\ell &< u(j+2)\\
&\le u\log_{\beta}(n) +2u - u \log_\beta(P_T(j)) - u\log_{\beta}\left(1 + \frac{Q_T(j)}{P_T(j)\beta^j}\right)\\
&\le u\log_{\beta}(n) +K.\end{aligned}$$
We proceed similarly to get a lower bound for $\ell$. Let $k=\lfloor\frac{\ell-T}{u}\rfloor$. Since $U$ is increasing, $$U_\ell<U_{u(k+1)+T}=P_T(k+1)\beta^{k+1} + Q_T(k+1).$$ We get $$\log_\beta(n)
<\log_\beta(U_\ell)
<k+1+\log_\beta(P_T(k+1))
+\log_{\beta}\left(1+\frac{Q_T(k+1)}{P_T(k+1)\beta^{k+1}}\right).$$ Observe that $k\le j+1$. Hence, from the first part, we get $$k+1\le j+2 \le \log_\beta(n)+\frac{K}{u}.$$ We also have $k\le \frac{\ell-T}{u}\le \frac{\ell}{u}$. Similarly as in the first part of the proof and since $P_T$ is a non-decreasing polynomial, there exists a constant $L\ge 0$ such that $$\ell \ge uk
> u\log_\beta(n)- u \log_\beta\left(P_T\left(\log_\beta(n)+\frac{K}{u}\right)\right)-L.$$
\[exa:merge\] Consider the sequence $1,2,6,12,36,72,\ldots$ defined by $U_0=1$, $U_{2i+1}=2 U_{2i}$ and $U_{2i+2}=3U_{2i+1}$. Then for all $i\ge 0$, $U_{i+2}=6U_{i}$. It is easily seen that $U_{2i}=6^i$ and $U_{2i+1}=2\cdot 6^i$. With the notation of Definition \[def:uT\], $u=2$, $\beta=6$, $d=0$ and $P_T=c_T=2$. The language $0^*\operatorname{rep}_U(\mathbb{N})$ is made of words where in even (resp. odd) positions when reading from right to left (i.e. least significant digits first), we can write $0,1$ (resp. $0,1,2$). If ${\lvert{\operatorname{rep}_U(n)}\rvert}=2 \ell+1$ then $U_{2\ell}=6^{\ell}\le n<U_{2\ell+1}=2\cdot 6^{\ell}$, so ${\lvert{\operatorname{rep}_U(n)}\rvert}\le 2\log_6(n)+1$ and ${\lvert{\operatorname{rep}_U(n)}\rvert}>2\log_6(\frac{n}{2})+1=2\log_6(n)-2\log_6(2)+1$. If ${\lvert{\operatorname{rep}_U(n)}\rvert}=2 \ell$ then $U_{2\ell-1}=2\cdot 6^{\ell-1}\le n<U_{2\ell}= 6^{\ell}$, so ${\lvert{\operatorname{rep}_U(n)}\rvert}\le 2\log_6(3n)= 2\log_6(n)+2\log_6(3)$ and ${\lvert{\operatorname{rep}_U(n)}\rvert}>2\log_6(n)$.
Consider the sequence $1,3,8,20,48,112,\ldots$ defined by $U_0=1$, $U_1=3$ and $U_{i+2}=4U_{i+1}-4U_i$. Then $U_i=(\frac{i}{2}+1)2^i$. With the notation of Definition \[def:uT\], $u=1$, $\beta=2$, $d=1$ and $P_T=\frac{X}{2}+1$. If ${\lvert{\operatorname{rep}_U(n)}\rvert}=\ell$ then $U_{\ell-1}=(\frac{\ell-1}{2}+1)2^{\ell-1}\le n<U_{\ell}=(\frac{\ell}{2}+1)2^{\ell}$, so ${\lvert{\operatorname{rep}_U(n)}\rvert}< \log_2(n)+1$ and ${\lvert{\operatorname{rep}_U(n)}\rvert}> \log_2(n)-\log_2(\frac{\ell}{2}+1)>\log_2(n)-\log_2(\frac{1}{2}\log_2(n)+\frac{3}{2})$. With the notation of Lemma \[lem:length\], $K=1$ and $P_T(\log_2(n)+K+2)=\frac{1}{2}\log_2(n)+\frac{5}{2}$.
As shown by the next result. It is enough to obtain a bound on the possible period of $X$. In [@BCFR Prop. 44], the result is given in a more general setting (i.e. for abstract numeration systems) and we restate it in our context.
\[bound:prep\] Let $U$ be a numeration system satisfying (H1), let $X\subseteq\mathbb{N}$ be an ultimately periodic set and let $\mathcal{A}_X$ be a DFA accepting $\operatorname{rep}_U(X)$. Then the preperiod of $X$ is bounded by a computable constant depending only on the size of $\mathcal{A}_X$ and the period $\pi_X$ of $X$.
Number of states {#sec4}
================
We follow Honkala’s strategy introduced in [@Honkala]. A DFA $\mathcal{A}_X$ accepting $\operatorname{rep}_U(X)$ is given as input. Assuming that $X$ is ultimately periodic, the number of states of $\mathcal{A}_X$ should provide an upper bound on the possible period and preperiod of $X$. Roughly speaking, the minimal preperiod/period should not be too large compared with the size of $\mathcal{A}_X$. This should leave us with a finite number of candidates to test. Thanks to Proposition \[pro:up\], one therefore builds a DFA for each pair of admissible preperiod/period. Equality of regular languages being decidable, we compare the language accepted by this DFA and the one accepted by $\mathcal{A}_X$. If an agreement is found, then $X$ is ultimately periodic, otherwise it is not. As a consequence of Proposition \[bound:prep\], we only focus on the admissible periods.
Assume that the minimal automaton $\mathcal{A}_X$ of $\operatorname{rep}_U(X)$ is given. Let $\pi_X$ be a potential period for $X$. We consider the prime decomposition of $\pi_X$. There are three types of prime factors.
1. Those that do not divide $a_0$.
2. Those that divide $a_0$ but that do not simultaneously divide all the coefficients of the recurrence relation.
3. The remaining ones are the primes dividing all the coefficients of the recurrence relation.
Our strategy is to bound those three types of factors separately.
Factors of the period that are coprime with $a_0$
-------------------------------------------------
The next result shows that given $\mathcal{A}_X$, the possible period cannot have a large factor coprime with $a_0$. So it provides a bound on this kind of factor that may occur in a candidate period.
\[pro:cas1\] Assume (H1), (H2) and (H3). Let $X\subseteq\mathbb{N}$ be an ultimately periodic $U$-recognizable set and let $q$ be a divisor of the period $\pi_X$ such that $(q,a_0)=1$. Then the minimal automaton of $\operatorname{rep}_U(X)$ has at least $q$ states.
Since $(q,a_0)=1$, the sequence $(U_i\bmod q)_{i\ge 0}$ is purely periodic. In particular, $1$ occurs infinitely often in this sequence.
We will make use of Corollary \[cor:h1h2\]. Let us define $q$ integers $k_1,\ldots,k_q\ge 0$ and thus $q$ words $w_1,\ldots,w_q\in\{0,1\}^*$ of the following form $$w_j:=10^{k_j}10^{k_{j-1}} \cdots 10^{k_1}0^{{\lvert{\operatorname{rep}_U(\pi_X)}\rvert}}.$$ Thanks to Corollary \[cor:h1h2\], we may impose the following conditions.
- First, $k_1$ is taken large enough to ensure that $\operatorname{val}_U(w_1)$ is larger than the preperiod of $X$ and $10^{k_1}\operatorname{rep}_U(\pi_X)$ is a valid greedy $U$-representation.
- Second, $k_2,\ldots,k_q$ are taken large enough to ensure that $w_j\in\operatorname{rep}_U(\mathbb{N})$ for all $j$.
- Third, we can choose $k_1,\ldots,k_q$ so that the $1$’s occur at indices $m$ such that $U_m\equiv 1 \pmod q$.
Observe that $\operatorname{val}_U(w_j)\equiv j\pmod q$. Since $q$ divides $\pi_X$, the words $w_1,\ldots,w_q$ have pairwise distinct values modulo $\pi_X$.
Let $i,j\in\{1,\ldots,q\}$ such that $i\neq j$. By Lemma \[lem:per\], we can assume that there exists $r_{i,j}<\pi_X$ such that $\operatorname{val}_U(w_i)+r_{i,j}\in X$ and $\operatorname{val}_U(w_j)+r_{i,j}\not\in X$ (the symmetric situation is handled similarly). In particular, ${\lvert{\operatorname{rep}_U(r_{i,j})}\rvert}\le {\lvert{\operatorname{rep}_U(\pi_X)}\rvert}$. Consider the two words $$w_i 0^{-{\lvert{\operatorname{rep}_U(r_{i,j})}\rvert}} \, \operatorname{rep}_U(r_{i,j})
\quad\text{ and }\quad
w_j 0^{-{\lvert{rep_U(r_{i,j})}\rvert}} \, \operatorname{rep}_U(r_{i,j})$$ where, in the above notation, it should be understood that we replace the rightmost zeroes in $w_i$ and $w_j$ by $\operatorname{rep}_U(r_{i,j})$. The first word belongs to $\operatorname{rep}_U(X)$ and the second does not. Consequently, the number of states of the minimal automaton of $\operatorname{rep}_U(X)$ is at least $q$: $w_1,\ldots,w_q$ belong to pairwise distinct Nerode equivalence classes.
Prime factors of the period that divide $a_0$ but do not divide all the coefficients of the recurrence relation {#sec:fpc}
---------------------------------------------------------------------------------------------------------------
We depart from the strategy developed in [@BCFR] and now turn to a particular situation where a prime factor $p$ of the candidate period for $X$ is such that, for some integer $\mu\ge1$, the sequence $(U_i\bmod p^\mu)_{i\ge 0}$ has a period containing a non-zero element. Again, this will provide us with an upper bound on $p$ and its exponent in the prime decomposition of the period.
We say that an ultimately periodic sequence has a *zero period* if it has period $1$ and the repeated element is $0$. Otherwise stated, the sequence has a tail of zeroes.
\[rem:zp\] Let $\mu\ge 1$. Observe that if the periodic part of $(U_i\bmod p^\mu)_{i\ge 0}$ contains a non-zero element, then the same property holds for the sequence $(U_i\bmod p^{\mu'})_{i\ge 0}$ with $\mu'\ge \mu$.
Furthermore, assume that for infinitely many $\mu$, $(U_i\bmod p^\mu)_{i\ge 0}$ has a zero period. Then from the previous paragraph, we conclude that $(U_i\bmod p^\mu)_{i\ge 0}$ has a zero period for all $\mu\ge 1$.
We give a sequence where only finitely many sequences modulo $p^\mu$ have a zero period. Take the sequence $U_0=1$, $U_1=4$, $U_2=8$ and $U_{i+2}=U_{i+1}+U_i$ for $i\ge 1$, then the sequence $(U_i\bmod 2^\mu)_{i\ge 0}$ has a zero period for $\mu=1,2$ because of the particular initial conditions. But it is easily checked that it has a non-zero period for all $\mu\ge 3$.
The next result is a special instance of [@BCFR Thm. 32] and its proof turns out to be much simpler. It precisely describes the case where a zero period occurs infinitely often.
\[the:jason\] Let $p$ be a prime. The sequence $(U_i\bmod p^\mu)_{i\ge 0}$ has a zero period for all $\mu\ge 1$ if and only if all the coefficients $a_0,\ldots,a_{k-1}$ of the linear relation are divisible by $p$.
It is clear that if $a_0,\ldots,a_{k-1}$ are divisible by $p$, then for any choice of initial conditions, $U_k,\ldots,U_{2k-1}$ are divisible by $p$, hence $U_{2k},\ldots,U_{3k-1}$ are divisible by $p^2$, and so on and so forth. Otherwise stated, for all $\mu\ge 1$ and all $i\ge \mu k$, $U_i$ is divisible by $p^{\mu}$.
We turn to the converse. Since the sequence $(U_i)_{i\ge 0}$ is linearly recurrent, the power series $$\mathsf{U}(x):=\sum_{i\ge 0}U_i\, x^i$$ is rational. By assumption, $(U_i\bmod p^\mu)_{i\ge 0}$ has a zero period for all $\mu\ge 1$. Otherwise stated, with the $p$-adic absolute value notation, $|U_i|_p\le p^{-\mu}$ for large enough $i$, i.e. $|U_i|_p\to 0$ as $i\to+\infty$. Recall that a series $\sum_{i\ge 0}\gamma_i$ converges in $\mathbb{Q}_p$ if and only if $\lim_{i\to+\infty}|\gamma_i|_p=0$. Hence the series $\mathsf{U}(x)$ converges in $\mathbb{Q}_p$ in the closed unit disc. Therefore, the poles $\rho_1,\ldots,\rho_r\in\mathbb{C}_p$ of $\mathsf{U}(x)$ must satisfy $|\rho_j|_p>1$ for $1\le j\le r$.
Let $P(x)=1-a_{k-1}x-\ldots -a_0x^k$ be the reciprocal polynomial of the linear recurrence relation . By minimality of the order $k$ of the recurrence, the roots of $P$ are precisely the poles of $\mathsf{U}(x)$ with the same multiplicities. If we factor $$P(x)=(1-\delta_1x)\cdots (1-\delta_kx)$$ each of the $\delta_j$ is one of the $\frac{1}{\rho_1},\ldots,\frac{1}{\rho_r}$. For $n>0$, the coefficient of $x^n$ is an integer equal to a sum of product of elements of $p$-adic absolute value less than $1$. Since $|a+b|_p\le \max\{|a|_p,|b|_p\}$, this coefficient is an integer with a $p$-adic absolute value less than $1$, i.e. a multiple of $p$.
Thanks to Remark \[rem:zp\] and Theorem \[the:jason\], if $p$ is a prime not dividing all the coefficients of the recurrence relation then there exists a least integer $\lambda$ (depending only on $p$) such that $(U_i\bmod{p^\lambda})_{i\ge 0}$ has a period containing a non-zero element.
\[pro:cas2a\] Assume (H1), (H2) and (H3). Let $p$ be a prime not dividing all the coefficients of the recurrence relation and let $\lambda\ge 1$ be the least integer such that $(U_i\bmod{p^\lambda})_{i\ge 0}$ has a period containing a non-zero element. If $X\subseteq\mathbb{N}$ is an ultimately periodic $U$-recognizable set with period $\pi_X=p^\mu\cdot r$ where $\mu\ge\lambda$ and $r$ is not divisible by $p$, then the minimal automaton of $\operatorname{rep}_U(X)$ has at least $p^{\mu-\lambda+1}$ states.
We will make use of the following observation. Let $n\ge 1$. In the additive group $\left(\mathbb{Z}/p^n\mathbb{Z},+\right)$, an integer $a$ has order $p^s$ with $0\le s\le n$ if and only if $a=p^{n-s}\cdot m$ where $m$ is not divisible by $p$.
By assumption $(U_i\bmod p^\lambda)_{i\ge 0}$ has a period containing a non-zero element $R$ of order $\operatorname{ord}_{p^\lambda}(R)=p^{\theta}$ for some $\theta$ such that $0<\theta\le\lambda$. Consider a large enough index $K$ such that it is in the periodic part of $(U_i\bmod p^\mu)_{i\ge 0}$ and $U_K\equiv R\pmod{p^\lambda}$. Using the above observation twice, $U_K=m\cdot p^{\lambda-\theta}$ for some $m$ coprime with $p$ and therefore, $U_K$ has order $\operatorname{ord}_{p^\mu}(U_K)=p^{\mu-\lambda+\theta}$ modulo $p^\mu$.
We can again apply the same construction as in the proof of Proposition \[pro:cas1\]. We define words of the form $$w_j:=10^{k_j}10^{k_{j-1}} \cdots 10^{k_1}0^{{\lvert{\operatorname{rep}_U(\pi_X)}\rvert}}$$ with the same properties, except for the second one: the ones occur at indices $t$ such that $U_t\equiv U_K\pmod{p^\mu}$. Note that $$\operatorname{val}_U(w_j)\equiv j\cdot U_K \pmod{p^\mu}.$$ Hence the number of distinct numerical values modulo $p^\mu$ that are taken by those words is given by the order of $U_K$ in $\mathbb{Z}/p^\mu\mathbb{Z}$, i.e. $p^{\mu-\lambda+\theta}$. The conclusion follows in a similar way as in the proof of Proposition \[pro:cas1\].
Prime factors of the period that divide all the coefficients of the recurrence relation
---------------------------------------------------------------------------------------
We can factor the period $\pi_X$ as $$\label{eq:fQ}
\pi_X=Q_X\cdot p_1^{\mu_1}\cdots p_t^{\mu_t}$$ where every $p_j$ divides all the coefficients of the recurrence relation and, for every prime factor $q$ of $Q_X$, at least one of the coefficients of the recurrence relation is not divisible by $q$. Otherwise stated, the factor $Q_X$ collects the prime factor of types $(1)$ and $(2)$.
There is a finite number of primes dividing all the coefficients of the recurrence relation. Thus, we only have to obtain an upper bound on the corresponding exponents $\mu_1,\ldots,\mu_t$ that may appear in .
Let $j\in\{1,\ldots,t\}$ and $\mu\ge 1$. From Theorem \[the:jason\], the sequence $(U_i\bmod p_j^{\mu})_{i\ge 0}$ has a zero period. We denote by $\mathsf{f}_{p_j}(\mu)$ the length of the preperiod, i.e. $U_{\mathsf{f}_{p_j}(\mu)-1}\not\equiv 0\pmod{ p_j^{\mu}}$ and $U_{i}\equiv 0\pmod{p_j^{\mu}}$ for all $i\ge \mathsf{f}_{p_j}(\mu)$.
Let us consider the numeration system from Example \[exa:ppp\]. The sequence $(U_i\bmod 2)_{i\ge 0}$ is $1,1,1,1,0^\omega$. Hence $\mathsf{f}_2(1)=4$. The sequence $(U_i\bmod 4)_{i\ge 0}$ is $1, 3, 1, 3, 2, 0, 2, 2, 0^\omega$. Hence $\mathsf{f}_2(2)=8$. Continuing this way, we have $\mathsf{f}_2(3)=12$ and $\mathsf{f}_2(4)=16$.
Note that $\mathsf{f}_{p_j}$ is non-decreasing: $\mathsf{f}_{p_j}(\mu+1)\ge \mathsf{f}_{p_j}(\mu)$.
We denote by ${M_X}$ the maximum of the values $\mathsf{f}_{p_j}(\mu_j)$ for $j\in\{1,\ldots,t\}$: $${M_X}=\max_{1\le j\le t}\mathsf{f}_{p_j}(\mu_j).$$ Thus, ${M_X}$ is the least index such that for all $i\ge M_X$ and all $j\in\{1,\ldots,t\}$, $U_i\equiv 0 \pmod{p_j^{\mu_j}}$. By the Chinese remainder theorem, ${M_X}$ is also the least index such that for all $i\ge M_X$, $U_i\equiv 0 \pmod{\frac{\pi_X}{Q_X}}$. In particular, $U_{M_X}\ge \frac{\pi_X}{Q_X}$ and thus, ${\lvert{\operatorname{rep}_U(\frac{\pi_X}{Q_X}-1)}\rvert}\le M_X$.
Let us consider the numeration system from Example \[exa:toy\]. Here we have two prime factors $2$ and $3$ to take into account. Computations show that $\mathsf{f}_2(1)=3$, $\mathsf{f}_2(2)=5$, $\mathsf{f}_2(3)=7$ and $\mathsf{f}_3(1)=3$, $\mathsf{f}_3(2)=6$, $\mathsf{f}_3(3)=9$. Assume that we are interested in a period $\pi_X=72=2^3.3^2$. With the above definition, $M_X=\max(\mathsf{f}_2(3),\mathsf{f}_3(2))=7$. One can check that $(U_i\bmod 72)_{i\ge 0}$ is $1, 13, 19, 30, 54, 48, 36, 0^\omega$.
We introduce a quantity $\gamma_{Q_X}$ which only depends on the numeration system $U$ and the number $Q_X$ defined in . Since we are only interested in decidable issues, there is no need to find a sharp estimate on this quantity.
\[def:gamma\] Under (H1), for each $r\in\{0,\ldots,Q-1\}$, a DFA accepting the language $\operatorname{rep}_U(Q\, \mathbb{N}+r)$ can be effectively built (see Proposition \[pro:up\] or the construction in [@cant Prop. 3.1.9]). We let $\gamma_Q$ denote the maximum of the numbers of states of these DFAs for $r\in\{0,\ldots,Q-1\}$.
The crucial point in the next statement is that the most significant digit $1$ occurs for $U_{M_X-1}$ in a specific word. The proof makes use of the same kind of arguments built for definite languages as in [@LRRV Lemma 2.1].
\[the:main\] Assume (H1), (H2) and (H3). Let $X\subseteq\mathbb{N}$ be an ultimately periodic $U$-recognizable set with period $\pi_X$ factored as in . Assume that ${M_X}-{\lvert{\operatorname{rep}_U(\frac{\pi_X}{Q_X}-1)}\rvert} \ge C$ where $C$ is the constant given in Corollary \[cor:h1h2\]. Then the minimal automaton of $\operatorname{rep}_U(X)$ has at least $\frac{1}{\gamma_{Q_X}}({\lvert{\operatorname{rep}_U(\frac{\pi_X}{Q_X}-1)}\rvert}+1)$ states.
This result will provide us with an upper bound on $\mu_1,\ldots,\mu_t$ (details are given in Section \[ss52\]). Since $Q_X$ has been bounded in the first part of this paper, if $\max(\mu_1,\ldots,\mu_t)\to\infty$, then $\frac{\pi_X}{Q_X}\to\infty$ but therefore the number of states of the minimal automaton of $\operatorname{rep}_U(X)$ should increase.
We may apply Corollary \[cor:h1h2\] and consider the given positive constant $C$: we will assume that if $w$ is a greedy $U$-representation, then, for all $n\ge C-1$, $10^nw$ also belongs to $\operatorname{rep}_U(\mathbb{N})$.
For every $r\in\{0,\ldots,Q_X-1\}$, the set $X\cap (Q_X\mathbb{N}+r)$ has a period dividing $\frac{\pi_X}{Q_X}$ and at least one of these subsets has period exactly $\frac{\pi_X}{Q_X}$. So we can choose an $r\in\{0,\ldots,Q_X-1\}$ such that the set $X\cap (Q_X\mathbb{N}+r)$ has period $\frac{\pi_X}{Q_X}$.
Let $\mathcal{B}$ be the minimal automaton of $\operatorname{rep}_U(X\cap (Q_X\mathbb{N}+r))$. We will provide a lower bound on the number of states of this automaton. By definition of ${M_X}$, we have $U_{{M_X}-1}\not\equiv 0\pmod{\frac{\pi_X}{Q_X}}$. Let $g\ge C-1$ be large enough so that $U_{M_X+g}$ is larger than the preperiod of $X\cap (Q_X\mathbb{N}+r)$. By Lemma \[lem:per\] applied to the set $X\cap (Q_X\mathbb{N}+r)$, since $U_{{M_X}+g}+U_{{M_X}-1}\not\equiv U_{{M_X}+g}\pmod{\frac{\pi_X}{Q_X}}$, we may suppose that there exists $s<\frac{\pi_X}{Q_X}$ such that $$U_{{M_X}+g}+U_{{M_X}-1}+s\in X\cap (Q_X\mathbb{N}+r) \quad \text{and}\quad U_{{M_X}+g}+s\not\in X\cap (Q_X\mathbb{N}+r)$$ (the symmetrical situation is treated in the same way). Let $\ell_X:={\lvert{\operatorname{rep}_U(\frac{\pi_X}{Q_X}-1)}\rvert}$. Note that ${\lvert{\operatorname{rep}_U(s)}\rvert}\le \ell_X$ and then by assumption, $M_X-1-{\lvert{\operatorname{rep}_U(s)}\rvert}\ge M_x-1-\ell_X\ge C-1$. Thanks to Corollary \[cor:h1h2\], both words $$u:=10^g 1 0^{{M_X}-1-{\lvert{\operatorname{rep}_U(s)}\rvert}}\operatorname{rep}_U(s)$$ and $$v:=10^{g} 0 0^{{M_X}-1-{\lvert{\operatorname{rep}_U(s)}\rvert}}\operatorname{rep}_U(s)$$ are greedy $U$-representations. For all $\ell\ge 0$, define an equivalence relation $E_\ell$ on the set of states of $\mathcal{B}$: $$E_\ell(q,q')\Leftrightarrow (\forall x\in A_U^*)\bigl[|x|\ge \ell\Rightarrow (\delta(q,x)\in \mathcal{F}\Leftrightarrow \delta(q',x)\in \mathcal{F})\bigr]$$ where $\delta$ (resp. $\mathcal{F}$) is the transition function (resp. the set of final states) of $\mathcal{B}$. Let us denote the number of equivalence classes of $E_\ell$ by $P_\ell$. Clearly, $E_\ell(q,q')$ implies $E_{\ell+1}(q,q')$, and thus $P_\ell\ge P_{\ell+1}$.
Let $i\in\{0,\ldots,\ell_X\}$. By assumption, $\ell_X<M_X$. Since $u$ and $v$ have the same suffix of length $M_X-1$, we can factorize these words as $$u=u_i w_i
\quad\text{and}\quad
v=v_i w_i$$ where $|w_i|=i$. Let $q_0$ be the initial state of $\mathcal{B}$. By construction, $\delta(q_0,u_i w_i)\in\mathcal{F}$ whereas $\delta(q_0,v_i w_i)\notin\mathcal{F}$, hence the states $\delta(q_0,u_i)$ and $\delta(q_0,v_i)$ are not in relation with respect to $E_i$. But for all $j> i$, they satisfy $E_{j}$. It is enough to show that $$\label{eq:Ei+1}
E_{i+1}(\delta(q_0,u_i),\delta(q_0,v_i)).$$
Figures \[fig:ti\] and \[fig:ti2\] can help the reader. Let $x$ be such that $|x|=i+t$, with $t\ge 1$. Let $p$ be the prefix of $\operatorname{rep}_U(s)$ of length ${\lvert{\operatorname{rep}_U(s)}\rvert}-i$, this prefix $p$ being empty whenever this difference is negative. If we replace $w_i$ by $x$ in $u$ and $v$, we get $$u_ix=10^g10^{M_X-1-|px|+t}px
\quad\text{and}\quad
v_ix=10^g00^{M_X-1-|px|+t}px.$$ Then $$\operatorname{val}_U(u_i x)-\operatorname{val}_U(v_i x)=U_{M_X+t-1}$$ and by definition of ${M_X}$, $U_{M_X+t-1}\equiv 0\pmod{\frac{\pi_X}{Q_X}}$. Hence, $\operatorname{val}_U(u_i x)$ and $\operatorname{val}_U(v_i x)$ belong to the periodic part of $X\cap (Q_X\mathbb{N}+r)$ and they differ by a multiple of the period. Therefore, $\operatorname{val}_U(u_i x)$ belongs to $X\cap (Q_X\mathbb{N}+r)$ if and only if so does $\operatorname{val}_U(v_i x)$.
(3.05,.6) – (3.05,-2.4); (.7,2.2) – (.7,1.2); (.7,.5) – (.7,-0.5); (-1,0) rectangle (3,.6); at (-1,0) [$u$:]{}; (3.1,.8) – (5.1,.8); at (4.1,.8) [$i$]{}; (2,1.5) – (5.1,1.5); at (3.55,1.5) [$\le\ell_X$]{}; (1,2.2) – (5.1,2.2); at (3.55,2.2) [$M_X-1$]{}; at (.95,.9) [$1$]{}; at (.95,-.7) [$0$]{}; (3.1,0) rectangle (5.1,.6); at (1,0) [$u_i$]{}; at (4.1,0) [$w_i$]{}; (2,-.8) rectangle (5.1,-.2); at (-1,-.8) [$\operatorname{rep}_U(s)$:]{}; at (2.5,-.8) [$p$]{}; (-1,-1.6) rectangle (3,-1); (3.1,-1.6) rectangle (5.1,-1); at (1,-1.6) [$v_i$]{}; at (-1,-1.6) [$v$:]{}; at (4.1,-1.6) [$w_i$]{}; (3.1,-2.4) rectangle (6.5,-1.8); (5.2,-1.6) – (6.5,-1.6); at (5.85,-1.6) [$t$]{}; at (4.8,-2.4) [$x$]{};
(3.05,.6) – (3.05,-2.4); (.7,2.2) – (.7,1.2); (.7,.5) – (.7,-0.5); (-1,0) rectangle (3,.6); at (-1,0) [$u$:]{}; (3.1,.8) – (5.1,.8); at (4.1,.8) [$i$]{}; (2,1.5) – (5.1,1.5); at (3.55,1.5) [$\le\ell_X$]{}; (1,2.2) – (5.1,2.2); at (3.55,2.2) [$M_X-1$]{}; at (.95,.9) [$1$]{}; at (.95,-.7) [$0$]{}; (3.1,0) rectangle (5.1,.6); at (1,0) [$u_i$]{}; at (4.1,0) [$w_i$]{}; (3.5,-.8) rectangle (5.1,-.2); at (-1,-.8) [$\operatorname{rep}_U(s)$:]{}; (-1,-1.6) rectangle (3,-1); (3.1,-1.6) rectangle (5.1,-1); at (1,-1.6) [$v_i$]{}; at (-1,-1.6) [$v$:]{}; at (4.1,-1.6) [$w_i$]{}; (3.1,-2.4) rectangle (6.5,-1.8); (5.2,-1.6) – (6.5,-1.6); at (5.85,-1.6) [$t$]{}; at (4.8,-2.4) [$x$]{};
In order to obtain , it remains to show that either both $u_ix$ and $v_ix$ are valid greedy $U$-representations or both are not. If the word $px$ is not a greedy $U$-representation then neither $u_ix$ nor $v_ix$ can be valid. Assume now that $px$ is a greedy $U$-representation. Note that in both situations described in Figures \[fig:ti\] and \[fig:ti2\], $|px|\le \ell_X+t$. Thanks to the assumption, $M_X-1-|px|+t\ge M_X-1-\ell_X \ge C-1$. The greediness of $px$ and Corollary \[cor:h1h2\] imply that $10^{M_X-1-|px|+t}px$ is a greedy $U$-representation. Since $g\ge C-1$, $u_ix$ is also a greedy $U$-representation and the same observation trivially holds for $v_ix$.
We conclude that $$P_0 > P_1 > \cdots > P_{\ell_X}\ge 1.$$ Since $P_0$ is the number of states of $\mathcal{B}$, the automaton $\mathcal{B}$ has at least $\ell_X+1$ states.
Let $\mathcal{A}_X$ and $\mathcal{A}_r$ be the minimal automata of $\operatorname{rep}_U(X)$ and $\operatorname{rep}_U(Q_X\mathbb{N}+r)$ respectively. The number of states of $\mathcal{A}_r$ is bounded by $\gamma_{Q_X}$. The DFA $\mathcal{B}$ is a quotient of the product automaton $\mathcal{A}_X\times \mathcal{A}_r$, hence the number of states of $\mathcal{B}$ is at most the number of states of $\mathcal{A}_X$ times $\gamma_{Q_X}$. We thus obtain that the number of states of $\mathcal{A}_X$ is at least $\frac{\ell_X+1}{\gamma_{Q_X}}$.
Cases we can deal with {#sec5}
======================
The gcd of the coefficients of the recurrence relation is $1$.
--------------------------------------------------------------
In this case, for any ultimately periodic set $X$, the factorization of the period $\pi_X$ given in has the special form $\pi_X=Q_X$ and the addressed decision problem turns out to be decidable.
\[thm:main1\] Let $U$ be a linear numeration system satisfying (H1), (H2) and (H3), and such that the gcd of the coefficients of the recurrence relation is $1$. Given a DFA $\mathcal{A}$ accepting a language contained in the numeration language $\operatorname{rep}_U(\mathbb{N})$, it is decidable whether this DFA recognizes an ultimately periodic set.
Assume that $X$ is an ultimately periodic set with period $\pi_X$. Let $p$ be a prime that divides $\pi_X$. Either $p$ divides the last coefficient of the recurrence relation $a_0$, or it does not.
In the latter case, thanks to Proposition \[pro:cas1\], for any $n\ge 1$, if $p^n$ divides $\pi_X$ then $p^n$ is bounded by the number of states of $\mathcal{A}$.
In the former case, $p$ divides $a_0$. Note that there is only a finite number of such primes. By assumption, $p$ does not divide all the coefficients of the recurrence relation. Then thanks to Theorem \[the:jason\], there exists $\mu\ge 1$ such that the periodic part of the sequence $(U_i\bmod{p^\mu})_{i\ge 0}$ contains a non-zero element. Let $\lambda$ be the least such $\mu$. By an exhaustive search, one can determine the value of $\lambda$: one finds the period of a sequence $(U_i\bmod{p^\mu})_{i\ge 0}$ as soon as two $k$-tuples $(U_i \bmod{p^\mu},\ldots,U_{i+k-1} \bmod{p^\mu})$ are identical (where $k$ is the order of the recurrence). We then apply Proposition \[pro:cas2a\]. For any $n\ge 1$, if $p^n$ divides $\pi_X$ then either $n<\lambda$ or $p^{n-\lambda+1}$ is bounded by the number of states of $\mathcal{A}$.
The previous discussion provides us with an upper bound on $\pi_X$, i.e. on the admissible periods for $X$. Then from Proposition \[bound:prep\], associated with each admissible period, there is a computable bound for the corresponding admissible preperiods for $X$. We conclude that there is a finite number of pairs of candidates for the preperiod and period of $X$. Similar to Honkala’s scheme, we therefore have a decision procedure by enumerating a finite number of candidates. For each pair $(a,b)$ of possible preperiods and periods, there are $2^a2^b$ corresponding ultimately periodic sets $X$. For each such candidate $X$, we build a DFA accepting $\operatorname{rep}_U(X)$ and compare it with $\mathcal{A}$. We can conclude since equality of regular languages is decidable.
There exist recurrence relations with that property but that were not handled in [@BCFR]. Take [@BCFR Example 35] $$U_{i+5} = 6 U_{i +4} + 3U_{i+3} - U_{i+2} + 6U_{i+1} + 3U_{i},\ \forall i\ge 0.$$ For this recurrence relation, $N_U(3^i)\not\to\infty$. The characteristic polynomial has the dominant root $3+2\sqrt{3}$ and it also has three roots of modulus $1$. Therefore, no decision procedure was known. But thanks to Theorem \[thm:main1\], we can handle such new cases under our mild assumptions (H1), (H2) and (H3). Indeed, by applying Bertrand’s theorem with the initial conditions $1,7,45,291,1881$, the numeration language $0^*\operatorname{rep}_U(\mathbb{N})$ is the set of words over $\{0,1,\ldots,6\}$ avoiding the factors $63,64,65,66$, hence (H1) holds. Moreover, it is easily checked that for all $i\ge 0$, $U_{i+1}-U_i\ge 5 U_i$. Therefore, the system $U$ also satisfies (H2) and (H3).
The gcd of the coefficients of the recurrence relation is larger than 1. {#ss52}
------------------------------------------------------------------------
If $X$ is an ultimately periodic set with period $\pi_X=Q_X\cdot p_1^{\mu_1}\cdots p_t^{\mu_t}$ with $t\ge 1$ as in , then the quantity $M_X$ is well-defined. Theorem \[the:main\] has a major assumption. The quantity $$n_X:=M_X-{\lvert{\operatorname{rep}_U\Big(\frac{\pi_X}{Q_X}-1\Big)}\rvert}$$ should be larger than some positive constant $C$, which only depends on the numeration system $U$.
\[thm:main2\] Let $U$ be a linear numeration system satisfying (H1), (H2) and (H3), and such that the gcd of the coefficients of the recurrence relation is larger than 1. Let $C$ be the constant given in Corollary \[cor:h1h2\]. Assume there exists a computable positive integer $D$ such that for all ultimately periodic sets $X$ of period $\pi_X=Q_X\cdot p_1^{\mu_1}\cdots p_t^{\mu_t}$ as in with $t\ge 1$, if $\max(\mu_1,\ldots,\mu_t)\ge D$ then $n_X\ge C$. Then, given a DFA $\mathcal{A}$ accepting a language contained in the numeration language $\operatorname{rep}_U(\mathbb{N})$, it is decidable whether this DFA recognizes an ultimately periodic set.
Assume that $X$ is an ultimately periodic set with period $\pi_X=Q_X\cdot p_1^{\mu_1}\cdots p_t^{\mu_t}$ as in . Note that there are only finitely many primes dividing all the coefficients of the recurrence relation , hence the possible $p_1,\ldots,p_t$ belongs to a finite set depending only on the numeration system $U$.
Applying the same reasoning as in the proof of Theorem \[thm:main1\], $Q_X$ is bounded by a constant deduced from $\mathcal{A}$. So the quantity $\gamma_{Q_X}$ introduced in Definition \[def:gamma\] is also bounded.
By hypothesis, there exists a computable positive integer constant $D$ such that if $\max(\mu_1,\ldots,\mu_t)\ge D$ then $n_X\ge C$. The number of $t$-uples $(\mu_1,\ldots,\mu_t)$ in $ \{0,\ldots,D-1\}^t$ is finite. So there is a finite number of periods $\pi_X$ of the form $Q_X\cdot p_1^{\mu_1}\cdots p_t^{\mu_t}$ with $Q_X$ bounded and $(\mu_1,\ldots,\mu_t)$ in this set. We can enumerate them and proceed as in the last paragraph of the proof of Theorem \[thm:main1\].
We may now assume that $\max(\mu_1,\ldots,\mu_t)\ge D$. Thanks to the assumption, $n_X\ge C$ and we are able to apply Theorem \[the:main\]: it provides a bound on $\frac{\pi_X}{Q_X}$ and thus on the possible exponents $\mu_1,\ldots,\mu_t$ depending only on $\mathcal{A}$. We conclude in the same way as in the proof of Theorem \[thm:main1\].
In the last part of this section, we present a possible way to tackle new examples of numeration systems by applying Theorem \[thm:main2\]. We stress the fact that when $\pi_X$ is increasing then both terms $M_X$ and ${\lvert{\operatorname{rep}_U(\frac{\pi_X}{Q_X}-1)}\rvert}$ are increasing. If $\beta>1$ (see Definition \[def:uT\]), then the growth of the second one has a logarithmic bound thanks to Lemma \[lem:length\], so we need insight on $\mathsf{f}_{p_j}(\mu)$ to be able to guarantee $n_X\ge C$.
The *$p$-adic valuation* of an integer $n$, denoted $\nu_p(n)$, is the exponent of the highest power of $p$ dividing $n$. There is a clear link between $\nu_{p_j}$ and $\mathsf{f}_{p_j}$: for all non-negative integers $\mu$ and $N$, $$\mathsf{f}_{p_j}(\mu)=N \iff
(\nu_{p_j}(U_{N-1})<\mu\ \wedge\ \forall i\ge N,\, \nu_{p_j}(U_i)\ge \mu).$$
With our Example \[exa:toy\] and initial conditions $1,2,3$, computing the first few values of $\nu_2(U_i)$ might suggest that it is bounded by a function of the form $\frac{i}{2}+c$, for some constant $c$. Nevertheless, computing more terms we get the following pairs $(i,\nu_2(U_i))$: $(67, 44)$, $(2115, 1070)$, $(10307, 5172)$, $(534595, 267318)$, $(2631747, 1315896)$. The constant $c$ suggested by each of these points is respectively $\frac{21}{2}$, $\frac{25}{2}$, $\frac{37}{2}$, $\frac{41}{2}$, $\frac{45}{2}$, which is increasing. This example explains the second term $g(i)$ in the function bounding $\nu_{p_j}(U_i)$ in the next statement.
In the next statement, the reader can think about logarithm function instead of a general function $g$. Indeed, for any $\epsilon>0$, for large enough $i$, $\log(i)< \epsilon\, i$. We also keep context and notation from .
\[lem:valp\] Let $j\in\{1,\ldots,t\}$ and let $\beta$ as in Definition \[def:uT\]. Assume that $\beta>1$ and that there exist $\alpha,\epsilon\in\mathbb{R}_{>0}$ and a non-decreasing function $g$ such that $$\nu_{p_j}(U_i)<\lfloor \alpha i\rfloor +g(i)$$ and there exists $N$ such that $g(i)<\epsilon\, i$ for all $i>N$. Then, for large enough $\mu$, $$\mathsf{f}_{p_j}(\mu)> \frac{\mu}{\alpha+\epsilon}.$$
By definition of the $p$-adic valuation, $p_j^{\nu_{p_j}(U_i)} \mid U_i$ and $p_j^{\nu_{p_j}(U_i)+1}\nmid U_i$. Thus, By definition of $\mathsf{f}_{p_j}$, for all $i$, $$\mathsf{f}_{p_j}(\nu_{p_j}(U_i)+1)\ge i+1.$$ For all $\mu$, there exists $i$ such that $$\lfloor \alpha i\rfloor+g(i) \le \mu < \lfloor \alpha (i+1)\rfloor+g(i+1).$$ Take $\mu$ large enough so that $i\ge N$. Using the right-hand side inequality, $\mu< \alpha(i+1)+\epsilon(i+1)$ and we get $$i> \frac{\mu}{\alpha+\epsilon}-1.$$ Using the left-hand side inequality, $\mu\ge \lfloor \alpha i\rfloor+g(i) > \nu_{p_j}(U_i)$. Since we have integers on both sides, $\mu\ge \nu_{p_j}(U_i)+1$. Since $\mathsf{f}_{p_j}$ is non-decreasing, for all large enough $\mu$, $$\mathsf{f}_{p_j}(\mu) \ge \mathsf{f}_{p_j}(\nu_{p_j}(U_i)+1)\ge i+1> \frac{\mu}{\alpha+\epsilon}. \qedhere$$
We look for a lower bound for $n_X$. Suppose that for each $j\in\{1,\ldots,t\}$, there exists $\alpha_j,\varepsilon_j,g_j$ and $N_j$ as in the above lemma. Then $${M_X}=\max_j \mathsf{f}_{p_j}(\mu_j)
>\max_j \left(\frac{\mu_j}{\alpha_j+\epsilon_j}\right)
\ge \frac{\max_j \mu_j}{\max_j (\alpha_j+\epsilon_j)}.$$ Second, let $u
$ and $\beta$ as in Definition \[def:uT\]. By hypothesis, $\beta>1$. Applying Lemma \[lem:length\], there exists a constant $K$ such that $${\lvert{\operatorname{rep}_U(\frac{\pi_X}{Q_X}-1)}\rvert}\le u \log_{\beta} \left(\prod_j p_j^{\mu_j}\right)+K.$$ The right hand side is $$u \sum_j \mu_j \log_{\beta} (p_j) +K \le u (\max_j \mu_j) \sum_j \log_{\beta} p_j +K.$$ Consequently, $$n_X
\ge \max_j \mu_j \left( \frac{1}{\max_j (\alpha_j+\epsilon_j)} - u \sum_j \log_{\beta} p_j\right) -K.$$
If $\pi_X$ tends to infinity (and assuming that the corresponding factor $Q_X$ remains bounded as explained in the proof of Theorem \[thm:main2\]), then $\max_j \mu_j$ must also tend to infinity. So we are able to conclude, i.e. $n_X$ tends to infinity and in particular, $n_X$ will become larger than $C$ (the constant from Corollary \[cor:h1h2\]) whenever $$\label{eq:test}
\frac{1}{\max_j (\alpha_j+\epsilon_j)} > u \sum_j \log_{\beta} p_j.$$
Actually, we don’t need $n_X$ tending to infinity, we have the weaker requirement $n_X\ge C$. The constant $D$ from Theorem \[thm:main2\] can be obtained as follows. To ensure that $n_X\ge C$, it is enough to have $$\label{eq:test2}
\max_j \mu_j \ge \frac{C+K}{\frac{1}{\max_j (\alpha_j+\epsilon_j)} - u \sum_j \log_{\beta} p_j}$$ and the right hand side only depends on the numeration system $U$.
As a conclusion, we simply define the constant $D$ as the right hand side in and, under the assumption of Lemma \[lem:valp\] about the behavior of the $p_j$-adic valuations of $(U_i)_{i\ge 0}$, the decision procedure of Theorem \[thm:main2\] may thus be applied. From a practical point of view, even though $n_X$ tending to infinity is not required, testing is relatively easy to estimate as seen in the following remark. This is not a formal proof, simply rough computations suggesting what could be the value of $\alpha$ in Lemma \[lem:valp\].
One can first make some computational experiments. Take the numeration system of Example \[exa:ppp\]. If we compute $\nu_2(U_i)$, the values for $41\le i\le 60 $ are given by $$10, 10, 10, 11, 12, 11, 11, 12, 12, 12, 12, 13, 16, 13, 13, 14, 14, 14, 14, 15.$$ Hence, one can conjecture that $\alpha_1=\frac{1}{4}$ and the above condition becomes ($u=1$), assuming $\epsilon_1$ to be negligible, $$4>\log_{2.804}(2)\simeq 0.672.$$
Take the numeration system of Example \[exa:toy\]. If we compute $\nu_2(U_i)$, the values for $41\le i\le 60 $ are given by $$24, 20, 21, 21, 24, 22, 23, 23, 27, 24, 25, 25, 28, 26, 27, 27, 33,
28, 29, 29$$ and, similarly, for $\nu_3(U_i)$ $$13, 14, 14, 14, 15, 15, 15, 16, 17, 16, 17, 17, 17, 18, 18, 18, 19,
20, 19, 20.$$ Hence, one can conjecture that $\alpha_1=\frac{1}{2}$ and $\alpha_2=\frac{1}{3}$. The recurrence has a real dominant root $\beta\simeq 12.554$. Assuming $\epsilon_1$ and $\epsilon_2$ to be negligible, the condition is therefore $$2>\log_{12.554}(2)+\log_{12.554}(3)\simeq 0.708.$$
An incursion into $p$-adic analysis {#sec:ap}
===================================
In this section, we discuss the requirement on the $p$-adic valuation given in Lemma \[lem:valp\]. To that end, we reconsider our toy example.
A second-order sequence
-----------------------
Throughout this section, let $U_{i + 3} = 12 U_{i + 2} + 6 U_{i + 1} + 12 U_i$ with initial conditions $U_0 = 1, U_1 = 13, U_2 = 163$ be the sequence of Example \[exa:toy\]. The $3$-adic valuation of $U_i$ has a simple structure.
\[p = 3 valuation\] For all $i \geq 0$, $$\nu_3(U_i)
= \left\lfloor\frac{i}{3}\right\rfloor +
\begin{cases}
0 & \text{if }i {\mathrel{\not\equiv}}4 \pmod 9 \\
1 & \text{if }i \equiv 4 \pmod 9.
\end{cases}$$
Let $T_i = U_i / 3^{\frac{i - 2}{3}}$. Since $U_{i + 3} = 12 U_{i + 2} + 6 U_{i + 1} + 12 U_i$, the sequence $(T_i)_{i \geq 0}$ satisfies the recurrence $T_{i + 3} = 4 \cdot 3^{2/3} T_{i + 2} + 2 \cdot 3^{1/3} T_{i + 1} + 4 T_i$. The initial terms are $T_0 = 3^{2/3}, T_1 = 13 \cdot 3^{1/3}, T_2 = 163$, so it follows that $T_i \in {\mathbb{Z}}[3^{1/3}]$ for all $i \geq 0$. Modulo $9 {\mathbb{Z}}[3^{1/3}]$, one computes that the sequence $(T_i)_{i \geq 0}$ is periodic with period length $27$ and period $$\begin{array}{ccccccccc}
3^{2/3}, & 4 \cdot 3^{1/3}, & 1, & 7 \cdot 3^{2/3}, & 3 \cdot 3^{1/3}, & 1, & 2 \cdot 3^{2/3}, & 2 \cdot 3^{1/3}, & 4, \\
3^{2/3}, & \phantom{1 \cdot {}} 3^{1/3}, & 7, & 7 \cdot 3^{2/3}, & 3 \cdot 3^{1/3}, & 7, & 8 \cdot 3^{2/3}, & 5 \cdot 3^{1/3}, & 1, \\
3^{2/3}, & 7 \cdot 3^{1/3}, & 4, & 7 \cdot 3^{2/3}, & 3 \cdot 3^{1/3}, & 4, & 5 \cdot 3^{2/3}, & 8 \cdot 3^{1/3}, & 7.
\end{array}$$ Therefore the sequence $(\nu_3(T_i))_{i \geq 0}$ of $3$-adic valuations is $$\frac{2}{3},\ \frac{1}{3},\ 0,\ \frac{2}{3},\ \frac{4}{3},\ 0,\ \frac{2}{3},\ \frac{1}{3},\ 0,\ \dots$$ with period length $9$. (Here we use the natural extension of $\nu_3$ to a function $\nu_3 \colon {\mathbb{Z}}[3^{1/3}] \to \frac{1}{3} {\mathbb{Z}}$.) Equivalently, $$\nu_3(T_i)
= \left\lfloor\frac{i}{3}\right\rfloor - \frac{i - 2}{3} +
\begin{cases}
0 & \text{if }i {\mathrel{\not\equiv}}4 \pmod 9 \\
1 & \text{if }i \equiv 4 \pmod 9.
\end{cases}$$ It follows that $$\nu_3(U_i)
= \frac{i - 2}{3} + \nu_3(T_i)
= \left\lfloor\frac{i}{3}\right\rfloor +
\begin{cases}
0 & \text{if }i {\mathrel{\not\equiv}}4 \pmod 9 \\
1 & \text{if }i \equiv 4 \pmod 9
\end{cases}$$ for all $i \geq 0$.
Theorem \[p = 3 valuation\] implies $\frac{i - 2}{3} \leq \nu_3(U_i) \leq \frac{i + 2}{3}$ for all $i \geq 0$. In particular, $\nu_3(U_i) < \lfloor\frac{i}{3}\rfloor + 2$, so the condition of Lemma \[lem:valp\] is satisfied, and therefore for every $\epsilon > 0$ we have $$\mathsf{f}_3(\mu) > \frac{\mu}{\frac{1}{3} + \epsilon}$$ for large enough $\mu$. This takes care of one of the two primes dividing the coefficients of the recurrence relation. We still have to discuss $\nu_2(U_i)$.
However, Theorem \[p = 3 valuation\] is not representative of the behavior of $\nu_p(s_i)$ for a general sequence $(s_i)_{i \geq 0}$ satisfying a linear recurrence with constant coefficients. For instance, the $2$-adic valuation of the sequence $(U_i)_{i \geq 0}$ is (much) more complicated. To study the more general setting, we will make use of the field ${\mathbb{Q}}_p$ of $p$-adic numbers and its ring of integers ${\mathbb{Z}}_p$. The $p$-adic valuation $\nu_p(x)$ of an element $x \in {\mathbb{Q}}_p$ is related to its $p$-adic absolute value ${\lvert{x}\rvert}_p$ by ${\lvert{x}\rvert}_p = p^{-\nu_p(x)}$.
Let ${\lvert{\operatorname{rep}_p(n)}\rvert}$ be the number of digits in the standard base-$p$ representation of $n$. For all $n \geq 1$, we can bound $\nu_p(n)$ as $$\nu_p(n) \leq {\lvert{\operatorname{rep}_p(n)}\rvert} - 1 = \left\lfloor \tfrac{1}{\log(p)} \log(n) \right\rfloor \leq \tfrac{1}{\log(p)} \log(n).$$ (We avoid writing “$\log_p(n)$” here to reserve $\log_p$ for the $p$-adic logarithm, which will come into play shortly.) Proposition \[log upper bound\] below gives the analogous upper bound on $\nu_p(n - \zeta)$ when $\zeta$ is a $p$-adic integer whose sequence of base-$p$ digits does not have blocks of consecutive $0$s that grow too quickly.
Let $p$ be a prime, and let $\zeta \in {\mathbb{Z}}_p \setminus {\mathbb{N}}$. Write $\zeta = \sum_{i \geq 0} d_i p^i$, where each $d_i \in \{0, 1, \dots, p - 1\}$. For each $a \geq 0$, let $\ell_\zeta(a) \geq 0$ be maximal such that $0 = d_a = d_{a + 1} = \dots = d_{a + \ell_\zeta(a) - 1}$.
\[log upper bound\] Let $p$ be a prime, and let $\zeta \in {\mathbb{Z}}_p \setminus {\mathbb{N}}$. If there exist real numbers $C, D$ such that $C > 0$, $D \geq -(C + 1)$, and $\ell_\zeta(a) \leq C a + D$ for all $a \geq 2$, then $\nu_p(n - \zeta) \leq \frac{2 C + D + 2}{\log(p)} \log(n)$ for all $n \geq p$.
Write $\zeta = \sum_{i \geq 0} d_i p^i$, where each $d_i \in \{0, 1, \dots, p - 1\}$. For each $a \geq 0$, define the integer $\zeta_a {\mathrel{\mathop:}=}\left(\zeta \bmod p^a\right) = \sum_{i = 0}^{a - 1} d_i p^i$. Then $\nu_p(\zeta_a - \zeta) = a + \ell_\zeta(a)$.
Let $n \geq p$, and let $a {\mathrel{\mathop:}=}{\lvert{\operatorname{rep}_p(n)}\rvert} \geq 2$. Since $\zeta \notin {\mathbb{N}}$, the $p$-adic valuation $b {\mathrel{\mathop:}=}\nu_p(n - \zeta)$ is an integer. There are two cases.
If $n \leq \zeta_b$, then in fact $n = \zeta_b$; this is because $n \leq \zeta_b < p^b$, so $n \neq \zeta_b$ implies $n - \zeta_b {\mathrel{\not\equiv}}0 \pmod{p^b}$, which contradicts $b = \nu_p(n - \zeta)$. Since ${\lvert{\operatorname{rep}_p(n)}\rvert} = a$ and $n = \zeta_b$, we have $0 = d_a = \dots = d_{b - 1}$. Therefore $\zeta_a = \zeta_b = n \geq p^{a - 1}$, and $$\frac{\nu_p(n - \zeta)}{\log(n)}
= \frac{\nu_p(\zeta_a - \zeta)}{\log(\zeta_a)}
\leq \frac{a + \ell_\zeta(a)}{\log(p^{a - 1})}
\leq \frac{a + C a + D}{(a - 1) \log(p)}
\leq \frac{2 + 2 C + D}{\log(p)},$$ where the final inequality follows from $1 + C + D \geq 0$.
If $n > \zeta_b$, then $n = \zeta_b + p^b m$ for some positive integer $m$. Therefore $n \geq p^b$, so $$\frac{\nu_p(n - \zeta)}{\log(n)}
\leq \frac{b}{\log(p^b)}
= \frac{1}{\log(p)}
< \frac{1 + C}{\log(p)}
\leq \frac{2 + 2 C + D}{\log(p)}$$ if $b \geq 1$ and $\frac{\nu_p(n - \zeta)}{\log(n)} = 0 < \frac{2 + 2 C + D}{\log(p)}$ if $b = 0$.
We now turn our attention to the sequence of $2$-adic valuations $\nu_2(U_i)$.
\[p = 2 zero\] There exists a unique $2$-adic integer $\zeta$ with the property that if $(i_n)_{n \geq 0}$ is a sequence of non-negative integers such that $\nu_2(U_{i_n}) \to \infty$ then $i_n \to \zeta$ in ${\mathbb{Z}}_2$.
A formula for $\zeta$ is given by Equation in the proof. In particular, $\zeta$ is a computable number, and one computes $\zeta \equiv 660098850944665 \pmod{2^{50}}$.
Let $p = 2$. To analyze the $2$-adic behavior of $(U_i)_{i \geq 0}$, we construct a piecewise interpolation of $U_i$ to ${\mathbb{Z}}_2$ using the method described by Rowland and Yassawi [@Rowland--Yassawi]. Let $P(x) = x^3 - 12 x^2 - 6 x - 12$ be the characteristic polynomial of $(U_i)_{i \geq 0}$. The polynomial $P(x)$ has a unique root $\beta_1 \in {\mathbb{Z}}_2$ satisfying $\beta_1 \equiv 2 \pmod 4$; this can be shown by an application of Hensel’s lemma (checking ${\lvert{P(2)}\rvert}_2 < {\lvert{P'(2)}\rvert}_2^2$). Polynomial division shows that $P(x)$ factors in ${\mathbb{Z}}_2[x]$ as $$P(x) = (x - \beta_1) \left(x^2 + (\beta_1 - 12) x + (\beta_1^2 - 12 \beta_1 - 6)\right).$$ One checks that $P(x)$ has no roots in ${\mathbb{Z}}_2$ congruent to $0$, $1$, $3$, $4$, $5$, or $7$ modulo $8$. Since $\beta_1$ has multiplicity $1$, this implies that the splitting field $K$ of $P(x)$ is a quadratic extension of ${\mathbb{Q}}_2$. Let $\beta_2$ and $\beta_3$ be the other two roots of $P(x)$ in $K = {\mathbb{Q}}_2(\beta_2)$. Since $\beta_1 \equiv 2 \pmod 4$, the $2$-adic absolute value of $\beta_1$ is ${\lvert{\beta_1}\rvert}_2 = \frac{1}{2}$. Using the quadratic factor of $P(x)$ and an approximation to $\beta_1$, one computes ${\lvert{\beta_2}\rvert}_2 = {\lvert{\beta_3}\rvert}_2 = \frac{1}{\sqrt{2}}$.
Let $c_1, c_2, c_3 \in K$ be such that $$U_i = c_1 \beta_1^i + c_2 \beta_2^i + c_3 \beta_3^i$$ for all $i \geq 0$. Using the initial conditions, we solve for $c_1, c_2, c_3$ to find $$\begin{aligned}
c_1 &= \frac{-U_0 \beta_2 \beta_3 + U_1 (\beta_2 + \beta_3) - U_2}{(\beta_2 - \beta_1) (\beta_1 - \beta_3)} \\
c_2 &= \frac{-U_0 \beta_3 \beta_1 + U_1 (\beta_3 + \beta_1) - U_2}{(\beta_3 - \beta_2) (\beta_2 - \beta_1)} \\
c_3 &= \frac{-U_0 \beta_1 \beta_2 + U_1 (\beta_1 + \beta_2) - U_2}{(\beta_1 - \beta_3) (\beta_3 - \beta_2)},\end{aligned}$$ where $U_0 = 1, U_1 = 13, U_2 = 163$. One computes ${\lvert{c_1}\rvert}_2 = 2$ and ${\lvert{c_2}\rvert}_2 = 2 \sqrt{2} = {\lvert{c_3}\rvert}_2$. Factoring out $\beta_2^i$ gives $$\label{2-adic valuation}
U_i
= \beta_2^i \left(c_1 \, (\tfrac{\beta_1}{\beta_2})^i + c_2 + c_3 \, (\tfrac{\beta_3}{\beta_2})^i\right).$$ Since ${\lvert{\frac{\beta_1}{\beta_2}}\rvert}_2 = \frac{1}{\sqrt{2}}$ and ${\lvert{\frac{\beta_3}{\beta_2}}\rvert}_2 = 1$, the power $(\tfrac{\beta_1}{\beta_2})^i$ approaches $0$ as $i \to \infty$, while $(\tfrac{\beta_3}{\beta_2})^i$ does not. Therefore the size of $\nu_2(U_i / \beta_2^i)$ is limited by the proximity of $c_2 + c_3 \, (\tfrac{\beta_3}{\beta_2})^i$ to $0$.
To analyze the size of $c_2 + c_3 \, (\tfrac{\beta_3}{\beta_2})^i$, we interpret $(\frac{\beta_3}{\beta_2})^i$ as a function of a $p$-adic variable. For this we need the $p$-adic exponential and logarithm, which are defined on extensions of ${\mathbb{Q}}_p$ by their usual power series; $\log_p(1 + x)$ converges if $\lvert x \rvert_p < 1$, and $\exp_p x$ converges if $\lvert x \rvert_p < p^{-1/(p - 1)}$. Moreover, $\log_p$ is an isomorphism from the multiplicative group $\{x : \lvert x - 1 \rvert_p < p^{-1/(p - 1)}\}$ to the additive group $\{x : \lvert x \rvert_p < p^{-1/(p - 1)}\}$, and its inverse map is $\exp_p$ [@Gouvea Proposition 4.5.9 and Section 6.1]. Direct computation shows ${\lvert{(\frac{\beta_3}{\beta_2})^4 - 1}\rvert}_2 = \frac{1}{8} < \frac{1}{2} = p^{-1/(p - 1)}$. Therefore, for all $m \geq 0$ and $r \in \{0, 1, 2, 3\}$, $$\begin{aligned}
(\tfrac{\beta_3}{\beta_2})^{r + 4 m}
&= (\tfrac{\beta_3}{\beta_2})^r (\tfrac{\beta_3}{\beta_2})^{4 m}\\
&= (\tfrac{\beta_3}{\beta_2})^r \exp_2 \log_2((\tfrac{\beta_3}{\beta_2})^{4 m}) \\
&= (\tfrac{\beta_3}{\beta_2})^r \exp_2\!\left(m \log_2((\tfrac{\beta_3}{\beta_2})^4)\right).\end{aligned}$$ Denote $L {\mathrel{\mathop:}=}\log_2((\frac{\beta_3}{\beta_2})^4)$. Using the power series for $\log_2$, one computes ${\lvert{L}\rvert}_2 = \frac{1}{8}$. For each $x \in {\mathbb{Z}}_2[\beta_2]$ and $r \in \{0, 1, 2, 3\}$, define $$f_r(r + 4 x) {\mathrel{\mathop:}=}c_2 + c_3 \, (\tfrac{\beta_3}{\beta_2})^r \exp_2(L x).$$ For all $x \in {\mathbb{Z}}_2$, we have ${\lvert{L x}\rvert}_2 = \frac{1}{8} {\lvert{x}\rvert}_2 \leq \frac{1}{8} < \frac{1}{2} = p^{-1/(p - 1)}$, so $f_r$ is well defined on $r + 4 {\mathbb{Z}}_2$. The four functions $f_0, f_1, f_2, f_3$ comprise a piecewise interpolation of $c_2 + c_3 \, (\frac{\beta_3}{\beta_2})^i$. Namely, $c_2 + c_3 \, (\frac{\beta_3}{\beta_2})^i = f_{i \bmod 4}(i)$ for all $i \geq 0$.
Since each $f_r$ is a continuous function, from Equation we see that $\nu_2(U_i / \beta_2^i)$ is large when $i$ is close to a zero of $f_{i \bmod 4}$. The equation $f_r(r + 4 x) = 0$ is equivalent to $$\exp_2(L x)
= -\tfrac{c_2}{c_3} (\tfrac{\beta_2}{\beta_3})^r.$$ For $r \in \{0, 2, 3\}$, one computes ${\left\lvert{-\frac{c_2}{c_3} (\frac{\beta_2}{\beta_3})^r - 1}\right\rvert}_2 \geq \frac{1}{2}$, so there is no solution $x$ for these values of $r$. For $r = 1$, ${\left\lvert{-\frac{c_2}{c_3} (\frac{\beta_2}{\beta_3})^r - 1}\right\rvert}_2 = \frac{1}{16} < \frac{1}{2}$, so there is a unique solution, namely $x = \frac{1}{L} \log_2\!\left(-\frac{c_2 \beta_2}{c_3 \beta_3}\right)$, which has size ${\lvert{x}\rvert}_2 = \frac{1}{2}$. Let $$\label{zeta definition}
\zeta {\mathrel{\mathop:}=}1 + 4 \tfrac{1}{L} \log_2\!\left(-\tfrac{c_2 \beta_2}{c_3 \beta_3}\right),$$ so that $f_1(\zeta) = 0$ and ${\lvert{\zeta}\rvert}_2 = 1$.
It remains to show that $\zeta \in {\mathbb{Z}}_2$. Let $\sigma : K \to K$ be the Galois automorphism that non-trivially permutes $\beta_2$ and $\beta_3$. The formulas for $c_2$ and $c_3$ imply $\frac{c_2}{c_3} \cdot \frac{\sigma(c_2)}{\sigma(c_3)} = 1$; this implies $$\begin{aligned}
\log_2\!\left(-\tfrac{c_2 \beta_2}{c_3 \beta_3}\right)
+ \sigma\!\left(\log_2\!\left(-\tfrac{c_2 \beta_2}{c_3 \beta_3}\right)\right)
&= \log_2\!\left(\tfrac{c_2 \beta_2}{c_3 \beta_3} \cdot \tfrac{\sigma(c_2) \beta_3}{\sigma(c_3) \beta_2}\right) \\
&= \log_2(1)
= 0.\end{aligned}$$ Similarly, $$\log_2((\tfrac{\beta_3}{\beta_2})^4) + \sigma\!\left(\log_2((\tfrac{\beta_3}{\beta_2})^4)\right)
= \log_2(1)
= 0.$$ Therefore $$\frac{
\log_2\!\left(-\tfrac{c_2 \beta_2}{c_3 \beta_3}\right)
}{
\log_2((\frac{\beta_3}{\beta_2})^4)
}
= \frac{
-\sigma\!\left(\log_2\!\left(-\tfrac{c_2 \beta_2}{c_3 \beta_3}\right)\right)
}{
-\sigma\!\left(\log_2((\frac{\beta_3}{\beta_2})^4)\right)
}
= \sigma\!\left(\frac{
\log_2\!\left(-\tfrac{c_2 \beta_2}{c_3 \beta_3}\right)
}{
\log_2((\frac{\beta_3}{\beta_2})^4)
}\right)$$ is invariant under $\sigma$ and thus is an element of ${\mathbb{Q}}_2$. It follows from ${\lvert{\zeta}\rvert}_2 = 1$ that $\zeta \in {\mathbb{Z}}_2$.
The interpolation in the previous proof depends on appropriate powers of $\frac{\beta_3}{\beta_2}$ satisfying $x = \exp_2(\log_2(x))$. We verified this by directly checking ${\lvert{(\frac{\beta_3}{\beta_2})^4 - 1}\rvert}_2 < \frac{1}{2}$. In general, an appropriate exponent is given by [@Rowland--Yassawi Lemma 6], namely $$\begin{cases}
1 & \text{if $e < p - 1$} \\
p^{\lceil \log(e+1) / \log p \rceil} & \text{if $e \geq p - 1$,}
\end{cases}$$ where $e$ is the ramification index of the field extension. The ramification index of the extension $K$ in the proof of Theorem \[p = 2 zero\] is $e = 2$; this follows from the fact that $e$ is a divisor of the degree of the extension and that $e \neq 1$ since we identified an element $\beta_2 \in K$ with $2$-adic valuation $\nu_2(\beta_2) = \frac{1}{2}$. Therefore the exponent $2^{\lceil \log(3) / \log(2) \rceil} = 4$ suffices. Since ${\lvert{\frac{\beta_3}{\beta_2}}\rvert}_2 = 1$, [@Rowland--Yassawi Lemma 6] implies ${\lvert{(\frac{\beta_3}{\beta_2})^4 - 1}\rvert}_2 < \frac{1}{2}$. (In general, one must divide by a root of unity before raising to the appropriate exponent, but this root of unity is $1$ for $\frac{\beta_3}{\beta_2}$ since the ramification index of $K$ is equal to its degree.)
By Proposition \[log upper bound\], the growth rate of $\nu_2(U_i)$ is determined by the approximability of $$\zeta = \cdots 10010110000101101100111101100001101111011010011001_2$$ by non-negative integers.
\[p = 2 block lengths\] Let $\zeta \in {\mathbb{Z}}_2$ be defined as in Equation . The lengths of the $0$ blocks of the $2$-adic digits of $\zeta$ satisfy $\ell_\zeta(a) \leq \frac{2}{95} a + \frac{18}{5}$ for all $a \geq 0$.
Conjecture \[p = 2 block lengths\] is weak in the sense that it is almost certainly far from sharp. One expects the digits of $\zeta$ to be randomly distributed, in which case $\ell_\zeta(a) = \frac{1}{\log(2)} \log(a) + O(1)$. Indeed, among the first $1000$ base-$2$ digits of $\zeta$, the longest block of $0$s has length $10$. However, results concerning digits of irrational numbers are notoriously difficult to prove. Bugeaud and Kekeç [@Bugeaud--Kekec Theorem 1.6] give a lower bound on the number of non-zero digits among the first $a$ digits of an irrational algebraic number in ${\mathbb{Q}}_p$; see also Theorem 2.1 in the same paper. However, there are no known results of this form for transcendental numbers.
The conjectural bound was obtained by computing the line through $\ell_\zeta(19) = 4$ and $\ell_\zeta(304) = 10$. If Conjecture \[p = 2 block lengths\] is true, then an explicit formula for $\nu_2(U_i)$ is given by the following theorem. In particular, the approximation $\zeta \equiv 660098850944665 \pmod {2^{50}}$ is sufficient to compute $\nu_2(U_i)$ for all $i \leq 2^{49}$.
\[p = 2 valuation\] Let $\zeta \in {\mathbb{Z}}_2$ be defined as in Equation . Conjecture \[p = 2 block lengths\] implies that, for all $i \geq 10$, $$\nu_2(U_i)
= \left\lfloor\frac{i - 1}{2}\right\rfloor +
\begin{cases}
\nu_2(i - \zeta) & \text{if }i {\mathrel{\not\equiv}}1 \pmod 4 \\
0 & \text{if }i \equiv 1 \pmod 4.
\end{cases}$$
We start as in the proof of Theorem \[p = 3 valuation\]. Let $T_i = U_i / 2^{\frac{i}{2} - 1}$. Since $U_{i + 3} = 12 U_{i + 2} + 6 U_{i + 1} + 12 U_i$, the sequence $(T_i)_{i \geq 0}$ satisfies the recurrence $T_{i + 3} = 6 \sqrt{2} T_{i + 2} + 3 T_{i + 1} + 3 \sqrt{2} T_i$. The initial terms are $T_0 = 2, T_1 = 13 \sqrt{2}, T_2 = 163$, so it follows that $T_i \in {\mathbb{Z}}[\sqrt{2}]$ for all $i \geq 0$. Modulo $2 {\mathbb{Z}}[\sqrt{2}]$, the sequence $(T_i)_{i \geq 2}$ is periodic with period length $4$: $1, \sqrt{2}, 1, 0, 1, \sqrt{2}, 1, 0, \dots$. It follows that if $i \geq 2$ and $i {\mathrel{\not\equiv}}1 \pmod 4$ then $$\begin{aligned}
\nu_2(U_i)
= \frac{i}{2} - 1 + \nu_2(T_i)
&= \frac{i}{2} - 1 +
\begin{cases}
0 & \text{if }i \equiv 0 \pmod 4 \\
0 & \text{if }i \equiv 2 \pmod 4 \\
\frac{1}{2} & \text{if }i \equiv 3 \pmod 4
\end{cases} \\
&= \left\lfloor\frac{i - 1}{2}\right\rfloor.\end{aligned}$$
It remains to determine $\nu_2(U_i)$ when $i \equiv 1 \pmod 4$. We continue to use the $2$-adic numbers $\beta_1, \beta_2, \beta_3, c_1, c_2, c_3$ and the function $f_1$ defined in the proof of Theorem \[p = 2 zero\]. When $i \equiv 1 \pmod 4$, Equation gives $${\left\lvert{U_i}\right\rvert}_2
= 2^{-\frac{i}{2}} {\left\lvert{c_1 \, (\tfrac{\beta_1}{\beta_2})^i + f_1(i)}\right\rvert}_2.$$ To obtain a simpler formula for ${\left\lvert{U_i}\right\rvert}_2$, we compare the sizes of the two terms being added and use the fact that ${\lvert{x + y}\rvert}_p = \max\{{\lvert{x}\rvert}_p, {\lvert{y}\rvert}_p\}$ if ${\lvert{x}\rvert}_p \neq {\lvert{y}\rvert}_p$. For the first, we have ${\left\lvert{c_1 \, (\tfrac{\beta_1}{\beta_2})^i}\right\rvert}_2 = 2^{1 - \frac{i}{2}}$. For the second, $${\left\lvert{f_1(i)}\right\rvert}_2
= {\left\lvert{c_2 + \tfrac{c_3 \beta_3}{\beta_2} \exp_2\!\left(L \cdot \tfrac{i - 1}{4}\right)}\right\rvert}_2.$$ Since the function $f_1(1 + 4 x) = c_2 + \tfrac{c_3 \beta_3}{\beta_2} \exp_2(L x)$ has a unique zero $\frac{\zeta - 1}{4}$, the $p$-adic Weierstrass preparation theorem [@Gouvea Theorem 6.2.6] implies the existence of a power series $h(x) \in K\llbracket x \rrbracket$ such that $h(0) = 1$, ${\lvert{h(x)}\rvert}_2 = 1$ for all $x \in {\mathbb{Z}}_2[\beta_2]$, and $$f_1(1 + 4 x)
= \frac{c_2 + \tfrac{c_3 \beta_3}{\beta_2}}{-\tfrac{\zeta - 1}{4}} \left(x - \tfrac{\zeta - 1}{4}\right) h(x).$$ Therefore $$\begin{aligned}
{\left\lvert{f_1(i)}\right\rvert}_2
&= {\left\lvert{\frac{c_2 + \tfrac{c_3 \beta_3}{\beta_2}}{-\tfrac{\zeta - 1}{4}}}\right\rvert}_2
{\left\lvert{\tfrac{i - 1}{4} - \tfrac{\zeta - 1}{4}}\right\rvert}_2 \\
&= \sqrt{2} {\left\lvert{i - \zeta}\right\rvert}_2.\end{aligned}$$ Conjecture \[p = 2 block lengths\] and Proposition \[log upper bound\] imply ${\lvert{i - \zeta}\rvert}_2 \geq i^{-536/95}$ for all $i \geq 2$. The functions $2^{1 - \frac{i}{2}}$ and $\sqrt{2} i^{-536/95}$ intersect at $i \approx 70.21$. For all $i \geq 73$ such that $i \equiv 1 \pmod 4$, $${\left\lvert{c_1 \, (\tfrac{\beta_1}{\beta_2})^i}\right\rvert}_2
= 2^{1 - \frac{i}{2}}
< \sqrt{2} i^{-536/95}
\leq {\left\lvert{f_1(i)}\right\rvert}_2$$ and therefore $${\left\lvert{U_i}\right\rvert}_2
= 2^{-\frac{i}{2}} {\left\lvert{c_1 \, (\tfrac{\beta_1}{\beta_2})^i + f_1(i)}\right\rvert}_2
= 2^{-\frac{i}{2}} {\left\lvert{f_1(i)}\right\rvert}_2
= 2^{\frac{1 - i}{2}} {\left\lvert{i - \zeta}\right\rvert}_2.$$ Moreover, explicit computation shows that $2^{1 - \frac{i}{2}} < \sqrt{2} {\lvert{i - \zeta}\rvert}_2$ for all $i \equiv 1 \pmod 4$ satisfying $13 \leq i \leq 69$, so ${\left\lvert{U_i}\right\rvert}_2 = 2^{\frac{1 - i}{2}} {\left\lvert{i - \zeta}\right\rvert}_2$ for these values as well. Therefore $\nu_2(U_i) = \frac{i - 1}{2} + \nu_2(i - \zeta)$ for all $i \geq 13$ such that $i \equiv 1 \pmod 4$.
Conjecture \[p = 2 block lengths\] implies that $\nu_2(U_i) \leq \frac{i}{2} + \frac{536}{95 \log(2)} \log(i)$ for all $i \geq 10$.
Since $U_i \neq 0$ for all $i \geq 0$, we have ${\left\lvert{U_i}\right\rvert}_2 \neq 0$ for all $i \geq 0$. Since ${\left\lvert{f_1(\zeta)}\right\rvert}_2 = 0$, this implies $\zeta \notin {\mathbb{N}}$. Conjecture \[p = 2 block lengths\] and Proposition \[log upper bound\] imply $\nu_2(i - \zeta) \leq \frac{536}{95 \log(2)} \log(i)$ for all $i \geq 2$. By Theorem \[p = 2 valuation\], $\nu_2(U_i) \leq \frac{i}{2} + \frac{536}{95 \log(2)} \log(i)$ for all $i \geq 10$.
This is sufficient to apply Lemma \[lem:valp\]. Under Conjecture \[p = 2 block lengths\], we have the right behavior for both $\nu_2(U_i)$ and $\nu_3(U_i)$.
A fourth-order sequence
-----------------------
The general case is even more complicated than Theorem \[p = 2 zero\]. For example, let $p = 2$ and consider the sequence $(U_i)_{i \geq 0}$ satisfying the recurrence $U_{i + 4} = 2 U_{i + 3} + 2 U_{i + 2} + 2 U_n$ with initial conditions $U_0 = 1, U_1 = 3, U_2 = 9, U_2 = 23$ from Example \[exa:ppp\]. By the Eisenstein criterion, the characteristic polynomial $P(x) = x^4 - 2 x^3 - 2 x^2 - 2$ is irreducible over ${\mathbb{Q}}_2$. Let $K$ be the splitting field of $P(x)$ over ${\mathbb{Q}}_2$. Let $\beta_1, \beta_2, \beta_3, \beta_4$ be the four roots of $P(x)$ in $K$, and let $c_1, c_2, c_3, c_4$ be the elements of $K$ such that $U_i = \sum_{j = 1}^4 c_j \beta_j^i$ for all $i \geq 0$.
To compute with the roots $\beta_i$, we would want to write $K$ as a simple extension ${\mathbb{Q}}_2(\alpha)$. For this, we need to determine the degree $d$ of the extension and a polynomial $Q(x) \in {\mathbb{Q}}_2[x]$ of degree $d$ such that $Q(x)$ is irreducible over ${\mathbb{Q}}_2$ and $Q(\alpha) = 0$. Then we could compare the sizes ${\lvert{\beta_j}\rvert}_2$ of the roots to each other. Experiments suggest that ${\lvert{\beta_1}\rvert}_2 = {\lvert{\beta_2}\rvert}_2 = {\lvert{\beta_3}\rvert}_2 = {\lvert{\beta_4}\rvert}_2 = 2^{-1/4}$ and ${\lvert{(\frac{\beta_j}{\beta_1})^8 - 1}\rvert}_2 = \frac{1}{4} < \frac{1}{2} = p^{-1/(p - 1)}$ for each $j \in \{2, 3, 4\}$. Assuming this is the case, $U_i / \beta_1^i = \sum_{j = 1}^4 c_j (\frac{\beta_j}{\beta_1})^i$ can be interpolated piecewise to ${\mathbb{Z}}_2$ using $8$ analytic functions. However, we cannot solve $c_1 + b_2 \exp_2(L_2 x) + b_3 \exp_2(L_3 x) + b_4 \exp_2(L_4 x) = 0$ explicitly, as we solved $c_2 + c_3 \, (\tfrac{\beta_3}{\beta_2})^r \exp_2(L x) = 0$ in the proof of Theorem \[p = 2 zero\]. Instead, we could use the $p$-adic Weierstrass preparation theorem [@Gouvea Theorem 6.2.6] to determine the number of solutions and compute approximations to them. However, we would also need to determine which of these solutions belong to ${\mathbb{Z}}_2$. We do not carry out this step here, but this would give an analogue of Theorem \[p = 2 zero\], with some finite set $Z$ of $2$-adic integers such that every sequence $(i_n)_{n \geq 0}$ of non-negative integers with $\nu_2(U_{i_n}) \to \infty$ satisfies $i_n \to \zeta$ for some $\zeta \in Z$. If the blocks of zeroes in the digit sequences of each $\zeta \in Z$ satisfy $\ell_\zeta(a) \leq C a + D$ for some $C, D$ as in Conjecture \[p = 2 block lengths\], then Proposition \[log upper bound\] gives an upper bound on $\nu_2(U_i)$. This same approach applies to a general constant-recursive sequence and a general prime $p$.
Concluding remarks
==================
The case of integer base $b$ numeration systems is not treated in this paper. Let $b\ge 2$. Assume first for the sake of simplicity that $b$ is a prime. Consider the sequence $U=(b^i)_{i\ge 0}$. If $X$ is an ultimately periodic set with period $\pi_X=b^\lambda$ for some $\lambda$, then with our notation $Q_X=1$ and ${\lvert{\operatorname{rep}_U(\pi_X-1)}\rvert}=\lambda$. The sequence $(b^i \bmod{b^\lambda})_{i\ge 0}$ has a zero period and $\mathsf{f}_b(\lambda)=\lambda$. Hence we don’t have the required assumption to apply Theorem \[the:main\]: for every such set $X$, $n_X=0$. Let us also point out that the technique of Propositions \[pro:cas1\] or \[pro:cas2a\] cannot be applied: adding $1$ as a most significant digit will not change the value of a representation modulo $\pi_X$ when words are too long, $U_i\equiv 0\pmod{b^\lambda}$ for large enough $i$. Of course, integer base systems can be handled with other decision procedures [@BMMR; @BH; @Honkala; @LRRV; @Marsault2019; @MSaka]. If the base $b$ is now a composite number of the form $p_1^{s_1}\cdots p_t^{s_t}$, the same observation holds. The length of the non-zero preperiod of $(b^i\bmod p_j^\mu)_{i\ge 0}$ is $\lfloor \frac{\mu}{s_j}\rfloor$. Taking again an ultimately periodic set with period $\pi_X=b^\lambda$, we get $Q_X=1$ and $\mathsf{f}_{p_j}(\lambda s_j)=\lambda$, hence $M_X=\lambda$ and we still have ${\lvert{\operatorname{rep}_U(\pi_X-1)}\rvert}=\lambda$, so $n_X=0$.
A similar situation occurs in a slightly more general setting: the merge of $r$ sequences that ultimately behave like $b^i$. Let $b\ge 2$, $u\ge 1$, $N\ge 0$. If the recurrence relation is of the form $U_{i+u}=b U_i$ for $i\ge N$ (as for instance in Example \[exa:noth2-newh2ok\]), then again $n_X\not\to \infty$ as $\pi_X\to\infty$. Indeed, if $X$ is an ultimately periodic set with period $\pi_X=b^\lambda$, then $Q_X=1$ and applying Lemma \[lem:length\] (here the polynomial $P_T$ with the notation of Definition \[def:uT\] is just a constant), ${\lvert{\operatorname{rep}_U(\pi_X-1)}\rvert}\ge u\lambda-L$, for some constant $L$, and with the same reasoning as for a composite integer base, $M_X\le N+u\lambda$. Thus $n_X$ remains bounded for all $\lambda$. So there is no way to ensure that $n_X$ can be larger than $C$.
Trying to figure out the limitations of our decision procedure and assuming that we are under the assumption of Lemma \[lem:valp\], this type of linear numeration systems is the only one that we were able to find where our procedure cannot be applied. Moreover, as shown by the following proposition, these systems are sufficiently close to the classical base-$b$ system so usual decision procedures can still be applied. It is an open problem to determine if there exist linear numeration systems satisfying (H1), (H2) and (H3) where the decision procedure may not be applied and not of the above type.
Take $b=4$, $u=2$ and $N=0$. Start with the first two values $1$ and $3$. We get the sequence $1,3,4,12,16,48,64,\ldots$. We have $\mathsf{f}_2(\mu)=\mu$ if $\mu$ is even and $\mathsf{f}_2(\mu)=\mu+1$ if $\mu$ is odd. Hence, for a set of period $\pi_X=4^\lambda$, $M_X=\mathsf{f}_2(2\lambda)=2\lambda$. Moreover, ${\lvert{\operatorname{rep}_U(4^\lambda-1)}\rvert}=2\lambda$. So, $n_X=0$ for all $\lambda$.
Let $b\ge 2$, $u\ge 1$, $N\ge 0$. Let $U$ be a linear numeration system $U=(U_i)_{i\ge 0}$ such that $U_{i+u}=b U_i$ for all $i\ge N$. If a set is $U$-recognizable then it is $b$-recognizable. Moreover, given a DFA accepting $\operatorname{rep}_U(X)$ for some set $X$, we can compute a DFA accepting $\operatorname{rep}_b(X)$.
We build in two steps a sequence of transducers reading least significant digit first that maps any $U$-representation $c_{\ell-1}\cdots c_1 c_0\in A_U^*$ (here written with the usual convention that the most significant digit is on the left) to the corresponding $b$-ary representation. Adding leading zeroes, we may assume that the length $\ell$ of the $U$-representation is of the form $N+mu$. The idea is to read the first $N+u$ (least significant) digits and to output a single digit (over a finite alphabet in $\mathbb{N}$) equal to $$d_0=\operatorname{val}_U(c_{N+u-1}\cdots c_0).$$ Then we process blocks of size $u$, each such block of the form $$c_{N+(j+1)u-1}\cdots c_{N+ju}$$ gives as output a single digit equal to $$d_j=c_{N+(j+1)u-1}U_{N+u-1}+\cdots +c_{N+ju} U_N.$$ So the digits $d_0,d_1,\ldots,d_{m-1}$ all belong to the finite set $$\{\operatorname{val}_U(w)\colon w\in A_U^* \text{ and } |w|\le N+u\}.$$ From the form of the recurrence, we have $$\operatorname{val}_U(c_{N+mu-1}\cdots c_0)=\sum_{j=0}^{m-1} d_j b^j
=\operatorname{val}_b(d_{m-1}\cdots d_0).$$ So this transducer $\mathcal{T}$ maps any $U$-representation to a non-classical $b$-ary representation of the same integer. Precisely, when a DFA accepting $\operatorname{rep}_U(X)$ is given, we build a DFA accepting the language $$L=0^*\operatorname{rep}_U(X)\cap\{w\in A_U^*\colon |w|\equiv N\pmod u,\ |w|\ge N\}.$$ Recall that if $L$ is a regular language then its image $\mathcal{T}(L)$ by a transducer is again regular. Moreover, $\operatorname{val}_b(\mathcal{T}(L))=X$.
Then, it is a classical result that normalization in base $b$, i.e. mapping a representation over a non-canonical finite set of digits to the canonical expansion over $\{0,\ldots,b-1\}$ can be achieved by a transducer $\mathcal{N}$ [@Frou] (or [@FLANS p. 104]). To conclude with the proof, we compose these two transducers and consider the image $\mathcal{N}(0^*\mathcal{T}(L))=0^*\operatorname{rep}_b(X)$.
With the above proposition, the decision problem for the merge of sequences ultimately behaving like $b^i$ (such as the numeration systems of Examples \[exa:noth2-newh2ok\] and \[exa:merge\]) can be reduced to the usual decision problem for integer bases.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Yann Bugeaud for pointing out relevant theorems in [@Bugeaud--Kekec].
[99]{}
J. Bell, É. Charlier, A. Fraenkel, M. Rigo, A decision problem for ultimately periodic sets in nonstandard numeration system, [*Internat. J. Algebra Comput.*]{} [**19**]{} (2009), no. 6, 809–839. V. Berthé, M. Rigo (Eds), [*Combinatorics, Automata, and Number Theory*]{}, Encycl. Math. and its Appl. [**135**]{}, Cambridge Univ. Press (2010). A. Bertand-Mathis, Comment écrire les nombres entiers dans une base qui n’est pas entière, [*Acta Math. Hungar.*]{} [**54**]{} (1989), 237–241. B. Boigelot, I. Mainz, V. Marsault, M. Rigo, An efficient algorithm to decide periodicity of $b$-recognisable sets using MSDF convention, 44th International Colloquium on Automata, Languages, and Programming, Art. No. [**118**]{}, 14 pp., LIPIcs. Leibniz Int. Proc. Inform., 80, Schloss Dagstuhl. Leibniz-Zent. Inform., Wadern, 2017. V. Bruyère, G. Hansel, Bertrand numeration systems and recognizability, [*Theoret. Comput. Sci.*]{} [**181**]{} (1997), 17–43. V. Bruyère, G. Hansel, Ch. Michaux, R. Villemaire, Logic and $p$-recognizable sets of integers, [*Bull. Belg. Math. Soc. Simon Stevin*]{} [**1**]{} (1994), 191–238. Y. Bugeaud, G. Kekeç, On Mahler’s classification of $p$-adic numbers, *Bulletin of the Australian Mathematical Society* **98** (2018) 203–211. A. Cobham, On the base-dependence of sets of numbers recognizable by finite automata, [*Math. Systems Theory*]{} [**3**]{} (1969), 186–192. F. Durand, Decidability of the HD0L ultimate periodicity problem, [*RAIRO Theor. Inform. Appl.*]{} [**47**]{} (2013), 201–214. F. Durand, M. Rigo, On Cobham’s theorem, in Handbook of Automata, to appear. A. S. Fraenkel, Systems of numeration, [*Amer. Math. Monthly*]{} [**92**]{} (1985), 105–114. Ch. Frougny, Representations of numbers and finite automata, Math. Systems Theory [**25**]{} (1992), 37–60. Ch. Frougny, On the sequentiality of the successor function, [*Inform. Comput.*]{} [**139**]{} (1997), 17–38. F. Q. Gouvêa, *$p$-adic Numbers: An Introduction* second edition, Universitext, Springer–Verlag, Berlin, 1997. J. Honkala, A decision method for the recognizability of sets defined by number systems, [*Theoret. Inform. Appl.*]{} [**20**]{} (1986), 395–403. A. Lacroix, N. Rampersad, M. Rigo, É. Vandomme, Syntactic complexity of ultimately periodic sets of integers and application to a decision procedure, [*Fund. Infor.*]{} [**116**]{} (2012), 175–187. P. B. A. Lecomte, M. Rigo, Numeration Systems on a Regular Language, [*Theory Comput. Systems*]{} [**34**]{} (2001), 27–44. V. Marsault, An efficient algorithm to decide periodicity of [$b$]{}-recognisable sets using [LSDF]{} convention, [*Log. Methods Comput. Sci.*]{} [**15**]{} (2019), Paper No. 8, 30. V. Marsault, J. Sakarovitch, Ultimate periodicity of $b$-recognisable sets: a quasilinear procedure, Developments in language theory, 362–373, [*Lect. Notes in Comput. Sci.*]{} [**7907**]{}, Springer, Heidelberg, (2013). I. V. Mitrofanov, Almost periodicity of morphic words, [*Dokl. Math.*]{} [**93**]{} (2016), 207–210. M. Rigo, [*Formal Languages, Automata and Numeration Systems: Applications to Recognizability and Decidability*]{}, vol. 2, Networks and Telecommunications series, ISTE-Wiley, 2014. E. Rowland, R. Yassawi, $p$-adic asymptotic properties of constant-recursive sequences, [*Indag. Math.*]{} [**28**]{} (2017), 205–220. A. Salomaa, M. Soittola, [*Automata-theoretic aspects of formal power series*]{}, Texts and Monographs in Computer Science, Springer-Verlag, New York, Heidelberg, Berlin, 1978. J. Shallit, Numeration systems, linear recurrences and regular sets, [*Inform. Comput.*]{} [**113**]{} (1994), 331–347.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'João C. Pinto Barros, Michele Burrello and Andrea Trombettoni'
bibliography:
- 'biblio\_PB\_B\_T.bib'
title: Gauge Theories with Ultracold Atoms
---
Introduction {#sec:1}
============
During the School on “Strongly Coupled Field Theories for Condensed Matter and Quantum Information Theory”, held in Natal (Brazil) in the days 2 - 14 August 2015, one of the authors gave a course on “Quantum Simulations of Gauge Fields with Ultracold Atoms”. The course was meant to be informal, at the blackboard, with time for discussions and to interact with the younger part of the audience. Subsequently, in 2017 J. C. Pinto Barros obtained the PhD in SISSA (Trieste) defending a Thesis on “Field and Gauge Theories with Ultracold Atoms”, under the supervision of Andrea Trombettoni and Marcello Dalmonte and with Michele Burrello and Enrique Rico Ortega acting as external referees. The present chapter is based on the Natal’s course and on the above mentioned PhD Thesis of J. C. Pinto Barros [@jpbthesis]. The latter Thesis is available at `http://www.statphys.sissa.it/wordpress/?page_id=1095`
The course gave an introductory discussion on lattice gauge theories, and then moved to explain how to simulate gauge potentials and gauge fields. A prior knowledge of ultracold atomic systems was assumed, even though during the lectures the corresponding concepts and notions were briefly reminded.
Gauge theories\[sec:Gauge-Theories\]
====================================
A gauge theory is a model that has a gauge symmetry. Such symmetry can be seen as a redundancy in the description of the degrees of freedom. In other words, this means that one can have two mathematically distinct solutions of the equations describing the system and nonetheless they describe the same physical situation. The paradigmatic example is classical electrodynamics. It describes the behavior of the electric field $\vec{E}\left(t,\vec{x}\right)$ and the magnetic field $\vec{B}\left(t,\vec{x}\right)$ in the presence of an electric charge density $\rho\left(t,\vec{x}\right)$ and the current density $\vec{j}\left(t,\vec{x}\right)$. The system is governed by the Maxwell equations: $$\begin{array}{cc}
\nabla\cdot\vec{E}\left(t,\vec{x}\right)=\rho\left(t,\vec{x}\right)\,;\qquad\hfill & \nabla\times\vec{B}\left(t,\vec{x}\right)-\partial_{t}\vec{E}\left(t,\vec{x}\right)=\vec{j}\left(t,\vec{x}\right)\,;\\
\nabla\cdot\vec{B}\left(t,\vec{x}\right)=0\,;\hfill & \nabla\times\vec{E}\left(t,\vec{x}\right)+\partial_{t}\vec{B}\left(t,\vec{x}\right)=0\,.\hfill
\end{array} \label{maxwell}$$ In the above equations and in the rest of this Chapter, natural units shall be adopted. The homogeneous equations, which are independent of charges and currents, can be straightforwardly solved by introducing a scalar potential $\phi\left(t,\vec{x}\right)$ and a vector potential $\vec{A}\left(t,\vec{x}\right)$: $$\vec{E}\left(t,\vec{x}\right)=-\nabla\phi\left(t,\vec{x}\right)-\partial_{t}\vec{A}\left(t,\vec{x}\right)\,, \qquad \vec{B}\left(t,\vec{x}\right)=\nabla\times\vec{A}\left(t,\vec{x}\right)\,. \label{eq:EB_phiA}$$ Using these two relations the last two equations in are fulfilled and the ones from the first row can be written in terms of $\phi\left(t,\vec{x}\right)$ and $\vec{A}\left(t,\vec{x}\right)$. After a solution is found, it can be plugged in Equation \[eq:EB\_phiA\] in order to obtain the electric and magnetic fields. However, not all different $\phi\left(t,\vec{x}\right)$ and $\vec{A}\left(t,\vec{x}\right)$ will give different electric and magnetic fields. In fact if two fields $\phi\left(t,\vec{x}\right)^{\prime}$ and $\vec{A}\left(t,\vec{x}\right)^{\prime}$ are related to another solution $\phi\left(t,\vec{x}\right)$ and $\vec{A}\left(t,\vec{x}\right)$ by: $$\phi\left(t,\vec{x}\right)^{\prime}=\phi\left(t,\vec{x}\right)+\partial_{t}\alpha\left(t,\vec{x}\right)\,, \qquad \vec{A}\left(t,\vec{x}\right)^{\prime}=\vec{A}\left(t,\vec{x}\right)-\nabla\alpha\left(t,\vec{x}\right)\,,\label{eq:EMG_gauge}$$ for some regular function $\alpha\left(t,\vec{x}\right)$, then the electric and magnetic fields, given by Equation \[eq:EB\_phiA\], remain unchanged. This means that the solutions $\phi,\vec{A}$ and $\phi^{\prime},\vec{A}^{\prime}$ correspond to the same physical situation and therefore they are just redundant descriptions of the same physics. The transformations of Equation \[eq:EMG\_gauge\] are called gauge transformations.
The existence of a gauge symmetry does not require the field to be dynamical. Consider a charged quantum particle in a background of a classical electromagnetic field. The Schrödinger equation for this system can be written as the equation in the absence of any field and “correcting” the canonical momentum $\vec{p}\rightarrow\vec{p}-e\vec{A}$. In the presence of an electromagnetic field the mechanical momentum, associated with the kinetic energy of the particle and denoted here by $\vec{\pi}$, is no longer the canonical momentum given by $\vec{p}$. The relation between them is $\vec{\pi}=\vec{p}-e\vec{A}$ which is at the core of this substitution. The same happens for the time derivative with the scalar potential $i\partial_{t}\rightarrow i\partial_{t}-e\phi$. Therefore, the Schrödinger equation reads, in the absence of any other interactions: $$\left(i\partial_{t}-e\phi\right)\psi\left(t,\vec{x}\right)=\left(-i\nabla-e\vec{A}\right)^{2}\psi\left(t,\vec{x}\right)\,.$$ Also this equation is invariant under the transformation \[eq:EMG\_gauge\] provided that the wave function is transformed by a position-dependent phase: $$\psi\left(t,\vec{x}\right)=e^{ie\alpha\left(t,\vec{x}\right)}\psi\left(t,\vec{x}\right)\,.$$ Given the space and time dependence of this transformation, it is denoted as a local gauge symmetry.
In quantum field theory an illustrative example is provided by QED. The Lagrangian is given by: $${\cal L}=\bar{\psi}\left(\gamma^{\mu}\left(i\partial_{\mu}-eA_{\mu}\right)-m\right)\psi-\frac{1}{4}F_{\mu\nu}F^{\mu\nu}\,.\label{eq:WED_Lagrangian}$$ Implicit sum over repeated indices is assumed. $\gamma^{\mu}$ are the gamma matrices satisfying the Clifford algebra $\left\{ \gamma^{\mu},\gamma^{\nu}\right\} =2\eta^{\mu\nu}$, $\eta^{\mu\nu}$ is the Minkowski metric $\eta=\mathrm{Diag}\left(1,-1,-1,-1\right)$, $\psi$ the Dirac spinor and $\bar{\psi}=\psi^{\dagger}\gamma^{0}$. The indices $\mu$ run from $0$ to $3$ where $0$ corresponds to the time index. $A_{\mu}$ is called gauge field and the last term of the Lagrangian corresponds to its kinetic term where $F_{\mu\nu}=\partial_{\mu}A_{\nu}-\partial_{\nu}A_{\mu}$. Also in this case there is a local set of transformations that leave this Lagrangian invariant. Explicitly: $$A_{\mu}\left(x\right)\rightarrow A_{\mu}\left(x\right)-\frac{1}{e}\partial_{\mu}\alpha\left(x\right)\,, \qquad \psi\left(x\right)\rightarrow e^{i\alpha\left(x\right)}\psi\left(x\right)\,,\label{eq:U1QED}$$
One can define the covariant derivative $D_{\mu}=\partial_{\mu}+ieA_{\mu}^{a}$ such that, under a gauge transformation, $D_{\mu}\psi \rightarrow e^{i\alpha\left(x \right)} D_{\mu}\psi$. In this way, the local gauge symmetry becomes apparent.
This is an example of a $U\left(1\right)$ gauge theory: a gauge transformation is defined, at each point, by phases $\alpha\in\left[0,2\pi\right[$ which combine according to the group $U\left(1\right)$. This construction can be generalized to other gauge groups, like $\mathbb{Z}_{N}$, or even non-Abelian, like $SU\left(N\right)$, for $N$ an integer number. For example, the Kitaev toric code is a $\mathbb{Z}_{2}$ (Abelian) gauge theory [@kitaev2003] whereas Quantum Chromodynamics (QCD), the theory that describes strong interactions in particle physics, is a $SU\left(3\right)$ (non-Abelian) gauge theory [@gross1973; @fritzsch1973; @politzer1973]. In the following a brief description of non-Abelian $SU\left(N\right)$ gauge invariance in quantum field theory is provided. For more details see, for example, [@peskin1995].
In order to explore these other symmetries, extra indices must be inserted (in the paradigmatic example of QCD these are the color indices). To simplify the notation, whenever $\psi$ it is used it is meant: $$\psi\equiv\left(\begin{array}{c}
\psi_{1}\\
\psi_{2}\\
\vdots\\
\psi_{n}
\end{array}\right)\label{eq:psi_color}$$ where each one of the $\psi_{i}$ corresponds to a (four-component in $3+1$ dimensions) Dirac spinor. Consider then a general symmetry group and a respective set of generators represented by Hermitian $n\times n$ matrices $t^{a}$. The goal is to build a Lagrangian which is invariant under the set of local transformations $$\psi\left(x\right)\rightarrow e^{i\alpha^{a}\left(x\right)t^{a}}\psi\left(x\right)\,.$$ This is a unitary transformation that mixes the $n$ components of the vector \[eq:psi\_color\] following a $n-$dimensional representation of the gauge group element $e^{i\alpha^{a}\left(x\right)t^{a}}$. The gauge field becomes in turn a matrix which can be parametrized as $A_\mu^a t^a$. Under a gauge transformation the field transforms as $$A_\mu^a t^a\rightarrow e^{i\alpha^{a}\left(x\right)t^{a}}\left(A_\mu^a t^a+\frac{i}{g}\partial_\mu\right)e^{-i\alpha^{a}\left(x\right)t^{a}}.$$ Writing the covariant derivative as $D_{\mu}=\partial_{\mu}-igA_{\mu}^{a}t^{a}$ one finds $D_{\mu}\psi\left(x\right)\rightarrow e^{i\alpha^{a}\left(x\right)t^{a}}D_{\mu}\psi\left(x\right)$. In this way $\bar{\psi}\left(\gamma^{\mu}D_{\mu}-m\right)\psi$ is a gauge invariant operator which includes the fermionic kinetic term and the matter-gauge coupling. Note that $\bar{\psi}$ is to be interpreted as line vector with components $\bar{\psi}_{i}$ and $\gamma^{\mu}$ are diagonal on the color indices, i.e. act the same for every color by standard matrix multiplication $\gamma^{\mu}\psi_{i}$. In order to define the gauge field dynamics, its gauge invariant kinetic term must be inserted. A possible way to derive its form is by considering the commutator $\left[D_\mu,D_\nu\right]=it^a F_{\mu\nu}^a$. Putting it differently, a general form for $F_{\mu\nu}^a$ can be obtained from this formula. Explicit computation yields $F_{\mu\nu}^{a}=\partial_{\mu}A_{\nu}^{a}-\partial_{\nu}A_{\mu}^{a}+gf^{abc}A_{\mu}^{b}A_{\nu}^{c}$ where the structure constants $f^{abc}$ are given by $\left[t^{a},t^{b}\right]=it^{c}f^{abc}$ and depend only on the symmetry group. From the transformation law for the covariant derivatives, one can see that $F_{\mu\nu}^{a}F^{a\mu\nu}$ is gauge invariant. The full Lagrangian can then be written as $${\cal L}=\bar{\psi}\left(\gamma^{\mu}D_{\mu}-m\right)\psi-\frac{1}{4}F_{\mu\nu}^{a}F^{a\mu\nu}\label{L_matter-gauge}$$
The perspective of implementing these kind of models in table top experiments is very appealing. First of all, it could give answers to very fundamental questions in physics like, for example, the exploration of the phase diagram of QCD. That is certainly a long term challenge and the path envisioned towards it involves the implementation of simpler intermediate steps. While QCD has a $SU\left(3\right)$ gauge symmetry and involves $3+1$ dimensions, this does not need to be the main target. A much simpler case of a $U\left(1\right)$ gauge symmetry in $1+1$ dimensions is already of great interest. In fact, this was the target of the first experimental implementation of a lattice gauge theory [@martinez2016](to be discussed in Section \[subsec:Encoding-in-\]). Step by step one may think to be able to realize more and more complex models. It is clear that if these models are realized they become interesting on their own both theoretically and experimentally. In particular, for example, it may also be advantageous to have situations where only certain degrees of freedom live in higher dimensions keeping others in lower dimensionality [@amaral1992; @PintoBarros2017LR; @PintoBarros1017Robustness] which could be used to simulate systems with long-range interactions, which have been the subject of an intense investigation in the last years [@PhysRevE.89.062120; @Brezin2014; @PhysRevE.92.052113; @PhysRevB.94.224411; @PAULOS2016246; @PhysRevE.95.012143; @PhysRevE.96.012108; @PhysRevLett.118.241601; @Behan_2017; @PhysRevB.96.104432].
Gauge symmetry on the lattice
-----------------------------
### Static fields \[sec:static\]
Following the discussion in the previous Section, a many-body Hamiltonian in the presence of a magnetic field can be obtained by replacing the momentum components for each particle by $p_{i}\rightarrow p_{i}-eA_{i}$. On the lattice, instead, this can be approximated by the Peierls substitution where the hopping parameters become complex. This is valid in a tight-binding regime and for a slow varying magnetic field. Explicitly the kinetic term is modified according to $$K=\underset{\vec{r},j}{\sum}t_{j}\hat{a}_{\vec{r}+\hat{j}}^{\dagger}\hat{a}_{\vec{r}}+\mathrm{h.c.}\rightarrow\underset{\vec{r},j}{\sum}t_{j}\hat{a}_{\vec{r}+\hat{j}}^{\dagger}e^{i\theta_{j}\left(\vec{r}\right)}\hat{a}_{\vec{r}}+\mathrm{h.c.}$$ In the previous equation the sum of $\vec{r}$ is taken over the lattice sites and the sum of $j$ is taken over all $d$ directions corresponding to the dimensionality of the system. The angle $\theta_{j}\left(\vec{r}\right)$ is just a phase that can depend, on general grounds, on both the direction of the hopping and the position. The key difference is that this phase here is non dynamical, so there is no kinetic term for it. This simply corresponds to allow for the hopping parameter of the particles on the lattice to be complex. Similarly to the models in continuum space, not all complex hoppings represent different physical scenarios as there is gauge invariance. In Section \[sim\_gauge\_pot\], several examples of techniques to engineer complex phases on the hopping parameters are discussed. Reviews can be found in [@dalibard2011; @Goldman2014b; @dalibard2015; @burrello2017].
### Dynamical fields \[sec:dynamical\]
In order to study a dynamical quantum (lattice) gauge theory, the lattice system under analysis must include also the degrees of freedom for the gauge fields and the complex hopping parameters are therefore promoted to operators acting on these degrees of freedom. Such degrees of freedom are usually associated to the lattice edges and their kinetic term must be supplied.
A constructive way to define such a system consists on taking the Lagrangian \[L\_matter-gauge\], write an Hamiltonian and perform a naive discretization. This offers in turn a recipe to engineer possible quantum simulations of these systems: a straightforward way of proceeding is indeed to create a system implementing the specific Hamiltonian of the lattice gauge theory, therefore it is useful to consider such theories in their Hamiltonian formulation [@kogut1975hamiltonian]. To describe the $U\left(1\right)$ case, it is useful to introduce the following link operators, acting on the gauge degrees of freedom: $$U_{\vec{r},j}=\exp\left(ie\overset{a\left(\vec{r}+\hat{j}\right)}{\underset{a\vec{r}}{\int}}dxA_{j}\left(x\right)\right)\,,\qquad L_{\vec{r},j}=\frac{E_{\vec{r},j}}{e}\,.$$ $U$ and $L$ are operators corresponding respectively to the connection and electric field of the theory (see, for example, [@ZoharBurrello2015] for more details). Based on these operators, we can define the Hamiltonian $$H=-\frac{i}{2a}\underset{\vec{r},j}{\sum}\left(\psi_{\vec{r}}^{\dagger}U_{\vec{r},j}\psi_{\vec{r}+\hat{j}}-\mathrm{h.c.}\right)+m\underset{n}{\sum}\psi_{\vec{r}}^{\dagger}\psi_{\vec{r}}+\frac{ae^{2}}{2}\underset{n}{\sum}L_{\vec{r},j}^{2}\,,$$ which reproduces the correct continuum theory when the *naive* continuum limit is taken. In the expressions above $a$ is the lattice spacing, $\vec{r}$ are the lattice points and $j$ labels the links connected to it. $E_{\vec{r}j}$ is the discretized version of the electric field which is the conjugate momentum of $A_j$ in the continuum version. The commutation relations between the link operators are $$\left[L_{\vec{r},i},U_{\vec{r}^{\prime},j}\right]=\delta_{\vec{r}\vec{r}^{\prime}}\delta_{ij}U_{\vec{r},i},\quad \left[L_{\vec{r},i},U_{\vec{r}^{\prime},j}^{\dagger}\right]=-\delta_{\vec{r}\vec{r}^{\prime}}\delta_{ij}U_{\vec{r},i}^\dagger,\quad
\left[U_{\vec{r},i},U_{\vec{r}^{\prime},j}\right]=\left[U^\dag_{\vec{r},i},U_{\vec{r}^{\prime},j}\right]=0.
\label{eq:com-rel}$$ We pause here to point out a couple of subtleties. The first concerns the so-called *naive* continuum limit, obtained by simply sending $a\rightarrow0$. While this works well for bosons, fermions suffer from the so-called “fermion doubling problem”. When this limit is taken with more care each fermion flavor on the lattice gives rise to $2^{d}$ fermion flavors on the continuum, being $d$ the number of discretized dimensions. The Nielsen-Ninomiya Theorem [@Nielsen1981; @Nielsen1981I; @Nielsen1981II] states that this is always the case when the fermion action is real, local and invariant under lattice translations and chiral transformations. There are alternative approaches to evade the Nielsen-Ninomiya Theorem which have their own advantages and disadvantages. A possible choice, popular among the quantum simulation community, is provided by staggered fermions [@Susskind1977] (also known as Kogut-Susskind fermions). The idea consists on distributing the spinor components among different lattice sites. In this way, instead of a spinor per site, one has only one fermion. Only for the Hamiltonian formulation of the $1+1D$ theory the fermion doubling problem can be completely solved in this way. In this case the Hamiltonian becomes $$H=-\frac{i}{2a}\underset{n}{\sum}\left(c_{n}^{\dagger}U_{n}c_{n+1}-\mathrm{h.c.}\right)+m\underset{n}{\sum}\left(-1\right)^{n}c_{n}^{\dagger}c_{n}+\frac{ae^{2}}{2}\underset{n}{\sum}L_{n}^{2}.$$ Spinors can be reconstructed from $\psi_{n}=\left(c_{2n}, c_{2n+1}\right)^T{/\sqrt{a_{st}}}$.
In higher dimensions the most non-trivial step consists on the existence of plaquette terms or, in other words, an energy cost for magnetic fields. These are gauge-invariant terms which must be present in order to fully represent the gauge theory. The absence of these terms is related to the strong coupling limit of the theory. On a 2D square lattice, the plaquette term originating at the point $\vec{r}$ is $U_{\square}=U_{\vec{r},x}U_{\vec{r}+\hat{x},y}U_{\vec{r}+\hat{y},x}^{\dagger}U_{\vec{r},y}^{\dagger}$ consisting on the smallest loops possible to draw on the lattice. The Hamiltonian for $d$ spatial dimensions takes the form: $$\begin{split}
H=&-\frac{i}{2a}\underset{\vec{r},i}{\sum}\left(-1\right)^{r_{1}+\ldots+r_{i-1}}\left(c_{\vec{r}}^{\dagger}U_{\vec{r},i}c_{\vec{r}+\hat{i}}-\mathrm{h.c.}\right)+m\underset{\vec{r}}{\sum}\left(-1\right)^{r_{1}+\ldots+r_{d}}c_{\vec{r}}^{\dagger}c_{\vec{r}}\\
&+\frac{a^{2-d}e^{2}}{2}\underset{\vec{r},i}{\sum}L_{\vec{r},i}^{2}-\frac{a^{d-4}}{4e^{2}}\underset{\square}{\sum}\left(U_{\square}+U_{\square}^{\dagger}\right)\label{eq:H_KG_general_d}
\end{split}$$ The extra alternating signs on the first term are required to obtain the correct Dirac Hamiltonian in the continuum limit with staggered fermions [@Susskind1977].
Another fundamental point, associated with the Hamiltonian formulation, consists on the restriction of the Hilbert space to physical states. This can be derived from the Lagrangian formulation by noting that the component $A_{0}$ is non-dynamical (there is no term $\partial_{0}A_{0}$). As a consequence it acts as a Lagrange multiplier enforcing the Gauss’ law as a constraint. Therefore, on the one spatial dimensional lattice, the physical states are defined by the relation: $$\label{gausslaw}
G_{\vec{r}}\left|\Psi\right\rangle =0$$ for each lattice site $\vec{r}$, where $$G_{\vec{r}}=\underset{i}{\sum}\left(L_{\vec{r},j}-L_{\vec{r}-\hat{j},j}\right)-Q_{\vec{r}}\,, \label{eq:u1-generators}$$ where $Q_{\vec{r}}$ is the dynamical matter charge. For the $1+1D$ case, for example, $Q_{n}=c_{n}^{\dagger}c_{n}+\frac{1-\left(-1\right)^{n}}{2}$. The alternating tem, which may look odd, is related to the staggered formulation. Considering a state with no electric field. The Gauss law demands that fermions populate the odd sites while leaving the even empty. This is because the spinor degrees of freedom are distributed along the lattice. Occupied odd sites have the interpretation of a filled Dirac sea. When a fermion hops from an odd to an even site it creates a hole in the Dirac sea while creating a particle above the Dirac sea. This is interpreted as the creation of particle/anti-particle pair where the hole plays the role of an anti-particle. In the presence of gauge fields, the hopping described above must be accompanied by a change on the electric field preserving Gauss’ law, as described by the connection operator in the first term of the Hamiltonian .
The $G_{\vec{r}}$ are also generators of the gauge transformation and can be extended for the $U\left(N\right)$ and $SU\left(N\right)$ gauge theories. To this purpose, one can consider matter fields $\psi_{\vec{r}}$ that transform under the gauge symmetries under a suitable representation of dimension $n$ of the gauge group. The generators of the gauge symmetries must therefore satisfy the relation $\left[G_{\vec{r}}^{a},\psi_{\vec{r}}\right]=t^{a}\psi_{\vec{r}}$ where $t^{a}$ are $n-$dimensional representations of the (left) group generators.
In order to preserve the gauge-invariance of the Hamiltonian, the connection operators must transform like tensors under the gauge transformations and they must follow the same representation of the matter fields: $U_{\vec{r},{j}} \to e^{i\alpha_{\vec{r}}^a t^a} U_{\vec{r},{j}} e^{-i\alpha_{\vec{r}+\hat{j}}^at^a}$. In particular the connection is multiplied on the left side by the transformation inherited from the lattice site on its left and on the right side by the inverse of the transformation inherited from the lattice site on its right. When we deal with a non-Abelian group, it is thus useful to distinguish left and right generators for the group transformations [@kogut1975hamiltonian], labelled by $L_{\vec{r},i}$ and $R_{\vec{r},i}$ respectively (see, for example, [@ZoharBurrello2015]). The local generators of the gauge transformation can therefore be defined as: $$G_{\vec{r}}^{a}=\sum_{i}\left(L_{\vec{r},i}^{a}+R_{\vec{r}-\hat{i},i}^{a}\right)+\psi_{\vec{r}}^{\dagger}t^{a}\psi_{\vec{r}}.$$ Finally, the lattice Hamiltonian for the non-Abelian theory will be: $$\begin{gathered}
H=-\frac{i}{2a}\underset{\vec{r},i}{\sum}\left(-1\right)^{r_{1}+\ldots+r_{i-1}}\left(\psi_{\vec{r}}^{\dagger}U_{\vec{r}i}\psi_{\vec{r}+\hat{i}}-\mathrm{h.c.}\right)+m\underset{\vec{r}}{\sum}\left(-1\right)^{r_{1}+\ldots+r_{d}}\psi_{\vec{r}}^{\dagger}\psi_{\vec{r}} \\
+\frac{a^{2-d}g^{2}}{2}\underset{\vec{r},i,a}{\sum}\left(\left(L_{\vec{r}i}^{a}\right)^{2}+\left(R_{\vec{r}i}^{a}\right)^{2}\right)-\frac{a^{d-4}}{4g^{2}}\underset{\square}{\sum}\mathrm{Tr}\left(U_{\square}+U_{\square}^{\dagger}\right)
.\label{eq:General_Abelian_QL}\end{gathered}$$ Again, the Gauss law should be imposed on physical states $G_{\vec{r}}^a\left|\Psi\right\rangle =0$.
Often times, in the proceeding Sections, the matter-gauge correlated hopping will be written as $\psi_{\vec{r}}^{\dagger}U_{\vec{r}i}\psi_{\vec{r}+\hat{i}}+\mathrm{h.c.}$ rather than $i\left(\psi_{\vec{r}}^{\dagger}U_{\vec{r}i}\psi_{\vec{r}+\hat{i}}-\mathrm{h.c.}\right)$ as above. While the latter reproduces the familiar continuum Hamiltonian in the naive continuum limit, both are related by a gauge transformation.
### Challenges, limitations and quantum link models\[sec:QL\]
Cold atom systems offer the possibility to construct Hubbard-like Hamiltonians with tunable hopping parameters and on-site interactions. However, gauge potentials and gauge fields demand more than that.
When the field is static, as described in Section \[sec:static\], the hopping parameters become complex. This is not readily available in simple optical lattices, but, thanks to recent experimental developments, it is nowadays possible to engineer static gauge fields, as we will discuss in Section \[sim\_gauge\_pot\].
For dynamical gauge fields, as discussed in Section \[sec:dynamical\], the matter hopping and link operators must be correlated in such a way to guarantee the existence of the local gauge symmetry (at each lattice site). Such kind of hopping is not natural in a cold atomic system and a discussion on how to implement is presented in Section \[sec:Simulating-gauge-fields\].
There is, yet, a further difficulty for dynamical gauge fields. Take, for example, the commutation relations in \[eq:com-rel\] pertaining a certain link. The operator $U_{\vec{r}}$ acts as a raising operator of the electric field (or equivalently of $L_{\vec{r}}$). But, for a $U(1)$ theory, this corresponds to an infinite Hilbert space per link. Constructing such links is certainly a challenge for its implementation, even for small lattice sizes. A solution of this problem is provded by quantum link models which are characterized by a finite Hilbert space per link, without violating the required gauge symmetry. These models were introduced by Horn in 1981 [@horn1981] and were further studied in [@orland1990; @orland1992; @chandrasekharan1997; @brower1999; @brower2004]. Proposed as an alternative formulation to Wilson gauge field theories on the lattice, they became an attractive realization of gauge symmetries for quantum simulation purposes.
In quantum link models the link degrees of freedom are replaced by quantum spins, such that the algebra in \[eq:com-rel\] is replaced by the algebra of angular momentum. In particular this correspond to considering alterative link operators: $$L_{+\vec{r},i}=S_{\vec{r},i}^{x}+iS_{\vec{r},i}^{y}\,,\quad L_{\vec{r},i}=S_{\vec{r},i}^{z}$$ where the raising and lowering operators $L_{\pm \vec{r},i}$ replace $U_{\vec{r}}$ and its conjugate. With this construction, the first two relations of \[eq:com-rel\] are still satisfied. However, the last no longer holds because $U$ and $U^\dag$ do not commute any longer. In particular, $L_{\pm \vec{r}i}$ are not unitary whereas $U_{\vec{r}}$ was.
Even though the algebra itself is different, the angular momentum operators can be equally used to construct a gauge theory without compromising the gauge symmetry. In particular, we can choose the dimension of the Hilbert space in each of the links to be $2S+1$ with $S$ a positive half integer (corresponding, in the spin language, to the total spin). It is expected that in the limit of large $S$ the Wilson formulation should be recovered. Explicitly one can use the following link variables $U_{\vec{r}i}\rightarrow L_{+\vec{r}i}/\sqrt{S\left(S+1\right)}$. The new non-zero commutation relation is $\left[U_{\vec{r}i},U_{\vec{r}i}^{\dagger}\right]=2L_{\vec{r}i}/S\left(S+1\right)$. In the limit of $S\rightarrow+\infty$ the right hand side goes to zero and the initial algebra is recovered.
There is an analogous construction for $U\left(N\right)$ non-Abelian symmetries. One can see that the symmetry can be realized using an $SU\left(2N\right)$ algebra (note that for $N=1$ this gives, correctly, $SU\left(2\right)$). It is possible to construct the new algebra using the so-called “rishon fermions” [@brower2004]. They are written in terms of pairs of fermionic operators $l_{\vec{r},j}^{m}$ and $r_{\vec{r}+\hat{j},j}^{m}$ for each link between the sites $\vec{r}$ and $\vec{r} + \hat{j}$. These operators define additional left and right gauge modes laying on the lattice sites $\vec{r}$ and $\vec{r} + \hat{j}$, with the aim of describing the link degrees of freedom. $m$ labels their color index. We can write: $$\begin{aligned}
&L_{\vec{r},j}^{a}=\frac{1}{2}l_{\vec{r},j}^{m\dagger}t_{mn}^{a}l_{\vec{r},j}^{n}\,,\qquad R_{\vec{r},j}^{a}=\frac{1}{2}r_{\vec{r}+\hat{j},j}^{m\dagger}t_{mn}^{a}r_{\vec{r}+\hat{j},j}^{n}\,,\\&
E_{\vec{r},j}=\frac{1}{2}\left(r_{\vec{r}+\hat{j},j}^{m\dagger}r_{\vec{r}+\hat{j},j}^{m}-l_{\vec{r},j}^{m\dagger}l_{\vec{r},j}^{m}\right)\,,\\
&U_{\vec{r},j}^{mn}=l_{\vec{r},j}^{m}r_{\vec{r}+\hat{j},j}^{n\dagger}\,.\label{eq:rishon_representation}\end{aligned}$$ The finiteness of the Hilbert space is a feature desirable for future quantum simulation schemes. Even though not a primary concern at this stage, it is reassuring that the effective continuum limit can be achieved even if one uses quantum link models [@chandrasekharan1997].
Simulation of gauge potentials \[sim\_gauge\_pot\]
==================================================
In accordance with the previous discussion, the goal of this Section is to show specific examples on how a complex hopping parameter can be engineered. The two main strategies described will be two contrasting situations. In one external parameters are varied adibatically (Section \[sec:adiabatic\]), while in the other fast modes are integrated out (Section \[sec:driven\]).
Adiabatic change of external parameters \[sec:adiabatic\]
---------------------------------------------------------
The idea of this approach has, in its core, the tight relation between the Aharonov-Bohm phase [@aharonov1959] and the Berry phase which was a concept introduced by Berry in [@berry1984]. The first is the phase acquired by a particle traveling around a closed contour. At the end of the path, when it is back to the initial position, the wave function acquires a new phase which is independent of the details of how the path was done and only depending on the total magnetic flux through the contour. On the other side, the Berry phase corresponds to the phase acquired when some external parameters of the system are varied in time, “slowly”, coming back again to their initial value for a non-degenerate state. In a more precise way, the starting point is an Hamiltonian $H\left(q^{a},\lambda_{i}\right)$ where $q^{a}$ are degrees of freedom and $\lambda_{i}$ are a set of external parameters. If these parameters are varied sufficiently slowly returning, in the end, to their initial value, and if the initial state is an eigenstate non degenerate in energy, then the system will be back to its initial state. During the process, however, it will acquire a phase: $$\left|\psi\right\rangle \underset{\mathrm{adiabatic\ change}}{\longrightarrow}e^{i\gamma}\left|\psi\right\rangle \,.$$ The phase $\gamma$ can be derived by computing the time evolution operator and subtract the “trivial” dynamical phase acquired simply due to the time evolution. Let us consider the adiabatic evolution of a system such that each energy eigenstates remain non-degenerate during the whole process. In this case, starting from one of the eigenstates of the initial Hamiltonian, the system will continuosly evolve remaining in the corresponding eigenstate, with energy $E(t)$, at each time. Therefore the dynamical phase results $e^{-i\int E\left(t\right)dt}$. The additional Berry phase reads instead: $$\gamma=\ointctrclockwise_{{\cal C}}\tilde{A}_{i}\left(\lambda\right)d\lambda_{i}\,,$$ where ${\cal C}$ is the closed path in the space of the parameters $\lambda_{i}$ and $\tilde{A}_{i}$ are given by: $$\tilde{A}_{i}\left(\lambda\right)=i\left\langle \phi\left(\lambda\right)\right|\frac{\partial}{\partial\lambda_{i}}\left|\phi\left(\lambda\right)\right\rangle\,, \label{eq:artificial_A}$$ where $\left|\phi\left(\lambda\right)\right\rangle $ are reference eigenstates taken with an arbitrary choice of their overall phases. $\tilde{A}\left(\lambda\right)$ is the Berry connection. Different choices of the reference eigenstates with different phases, for example ${|{\phi'(\lambda)}\rangle} = e^{i\alpha\left(\lambda\right)}\left|\phi\left(\lambda\right)\right\rangle $, would just reproduce a gauge transformation on $\tilde{A}$: $$\tilde{A}_i\rightarrow\tilde{A}_i' = \tilde{A}_i-\frac{\partial\alpha}{\partial\lambda_{i}}\,.$$ This principle can be applied in multi-level atomic systems in order to reproduce artificial gauge fields in an ultracold atomic setting. As an example, the computation can be done for a two level atom, where it is shown how this vector potential appears explicitly at the Hamiltonian level. These two levels correspond to two internal states of the atom, the ground state $\left|g\right\rangle $ and an excited state $\left|e\right\rangle $. The center of mass Hamiltonian, assumed diagonal on the internal states, is taken to be simply the free particle Hamiltonian. The total Hamiltonian is $H=H_{0}+U$. By a suitable shift of the energy spectrum we can assume that the ground and excited state energies are related by $E_{g}=-E_{e}$. Then $U$ can be written as $$U=\frac{\Omega}{2}\left(\begin{array}{cc}
\cos\theta & e^{i\phi}\sin\theta\\
e^{i\phi}\sin\theta & -\cos\theta
\end{array}\right)$$ where $\theta$ and $\phi$ may depend on the position. The frequency $\Omega$ characterizes the strength of the coupling between the two states and it is assumed to be position independent. The eigenstates of this operator, denoted to as “dressed states”, are given by: $$\left|\chi_{1}\right\rangle =\left(\begin{array}{c}
\cos\frac{\theta}{2}\\
e^{i\phi}\sin\frac{\theta}{2}
\end{array}\right)\,,\qquad \left|\chi_{2}\right\rangle =\left(\begin{array}{c}
-e^{-i\phi}\sin\frac{\theta}{2}\\
\cos\frac{\theta}{2}
\end{array}\right)$$ with eigenvalues $\pm\hbar\Omega/2$ respectively. We assume that the initial internal state is $\left|\chi_{1}\right\rangle $ and that the evolution is adiabatic, such that the system remains in the state ${|{\chi_1}\rangle}$ at all times. Hence the state of the system can be described by a wave function $\left|\psi\left(t,\vec{r}\right)\right\rangle =\varphi\left(t,\vec{r}\right)\left|\chi_{1}\left(\vec{r}\right)\right\rangle $ where $\varphi\left(t,\vec{r}\right)$ will obey a modified Schrödinger equation due to the dependence of $\left|\chi_{1}\left(\vec{r}\right)\right\rangle $ on the position. Plugging this into the Schrödinger equation and projecting on $\left|\chi_{1}\left(\vec{r}\right)\right\rangle $, we find the effective Hamiltonian governing $\varphi$: $$H_{\mathrm{eff}}=\frac{\left(p_{i}-i\left\langle \chi_{1}\left(\vec{r}\right)\right|\frac{\partial}{\partial x_{i}}\left|\chi_{1}\left(\vec{r}\right)\right\rangle \right)^{2}}{2m}+\frac{\left|\left\langle \chi_{2}\left(\vec{r}\right)\right|\frac{\partial}{\partial x_{i}}\left|\chi_{1}\left(\vec{r}\right)\right\rangle\right| ^{2}}{2m}+\frac{\Omega}{2}$$ As expected, a vector potential $\tilde{A}_{i}\left(\vec{r}\right)$ corresponding to the Berry connection is found. Additionally a potential $\tilde{V}\left(\vec{r}\right)$ is also created and related to virtual transitions to the other state $\left|\chi_{2}\left(\vec{r}\right)\right\rangle $. In this two level approximation, these two quantities are given by $\tilde{A}_{i}\left(\vec{r}\right)=\frac{\cos\theta-1}{2}\frac{\partial\phi}{\partial x_{i}}$ and $\tilde{V}\left(\vec{r}\right)=\frac{\left(\nabla\theta\right)^{2}+\sin^{2}\theta\left(\nabla\phi\right)^{2}}{8m}$. Discussions about the practical implementation on optical lattices can be found in [@dalibard2011; @dum1996; @visser1998; @Goldman2014b]. First experimental evidence of scalar potentials in quantum optics was found in [@Dutta1999] and the first observation of geometric magnetic fields in cold atomic physics was done in [@Lin2009]. By considering a set of degenerate or quasi-degenerate dressed states it is possible to achieve non-Abelian gauge potentials as well [@dalibard2011; @Goldman2014b].
Effective Hamiltonian in periodic driven system \[sec:driven\]
--------------------------------------------------------------
In contrast to the approach of the previous Subsection, where the creation of the magnetic field relied on a slow change in time (i.e. the particle moves slowly enough such that the position dependent internal state is followed adiabatically), the following technique relies on fast oscillations. The basic principle consists on having two very distinct timescales. A fast oscillating time dependent potential will give rise to an effective time independent Hamiltonian which will present the desired complex hopping term. A general technique was proposed in [@goldman2014] and it is based on a generic time-dependent periodic Hamiltonian: $$H=H_{0}+V\left(t\right)$$ where all the the time dependence is relegated to $V\left(t\right)=V(t+\tau)$ where $\tau$ is the time period. $V(t)$ can be decomposed as: $$V\left(t\right)=\underset{n}{\sum}\left(V_{n+}e^{in\omega t}+V_{n-}e^{in\omega t}\right)$$ where $V_{n\pm}$ are operators and $\omega= 2\pi/\tau$. The condition $V_{n+}=V_{n-}^{\dagger}$ guarantees the Hermiticity of the Hamiltonian.
A unitary transformation $e^{iK\left(t\right)}$ generates an effective Hamiltonian given by: $$H_{\mathrm{eff}}=e^{iK\left(t\right)}He^{-iK\left(t\right)}+i\left(\frac{\partial}{\partial t}e^{iK\left(t\right)}\right)e^{-iK\left(t\right)}$$ We choose a periodic operator $K\left(t\right)$ such that the effective Hamiltonian is time independent. Under this requirement, the time evolution operator can be represented as: $$U\left(t_{i}\rightarrow t_{f}\right)=e^{iK\left(t_{f}\right)}e^{-iH_{\mathrm{eff}}\left(t_{f}-t_{i}\right)}e^{-iK\left(t_{i}\right)}\,,$$ and it can be shown that, at lowest order, the effective Hamiltonian can be written as [@goldman2014]: $$H_{\mathrm{eff}}=H_{0}+\tau\underset{n}{\sum}\frac{1}{n}\left[V_{n+},V_{n-}\right]+{\cal O}\left(\tau^{2}\right)\label{eq:Heff_oscilator}.$$ This expansion relies on the small parameter $\tilde{V}\tau$ where $\tilde{V}$ is the typical energy scale of $V(t)$. This expansion turns out to be very useful in the effective description of ultracold atomic systems though care should be taken, in a case by case scenario, in order to be sure about the convergence of the series.
### Lattice Shaking
The lattice shaking approach consists on having an external time dependent optical potential that is changing in time in accordance to the previous description. Then a change of basis is performed for a co-moving frame that, along with a time average, will create an effective Hamiltonian with the desired complex hopping. As an example, a brief prescription is presented along the lines of the realization in a $\mathrm{Rb}$ Bose-Einstein condensate [@Struck2012]. The Hamiltonian considered is the usual tight-biding Hamiltonian in 2D with the usual hopping and on-site part $H_{os}$ (by on-site it is intended one body potential and scattering terms that act in single sites). In addition, there is an extra time dependent potential: $$H=-\underset{\vec{r},j}{\sum}t_{\vec{r}j}\hat{a}_{\vec{r}+\hat{j}}^{\dagger}\hat{a}_{\vec{r}}+H_{\mathrm{os}}+\underset{\vec{r}}{\sum}v_{\vec{r}}\left(t\right)\hat{a}_{\vec{r}}^{\dagger}\hat{a}_{\vec{r}}.$$ The function $v_{\vec{r}}\left(t\right)$ is periodic on time with period $\tau$: $v_{\vec{r}}\left(t\right)=v_{\vec{r}}\left(t+\tau\right)$. A unitary transformation on the states is performed and plugged in on the Schrödinger equation, thus defining new states $\left|\psi^{\prime}\right\rangle $ such that $\left|\psi\left(t\right)\right\rangle =U\left(t\right)\left|\psi^{\prime}\left(t\right)\right\rangle $. The Hamiltonian becomes $H^{\prime}\left(t\right)=U\left(t\right)^{\dagger}HU\left(t\right)-iU\left(t\right)^{\dagger}\dot{U}\left(t\right)$ (where the dot stands for time derivative). The transformation is given by $$U\left(t\right)=e^{-i\int_0^t dt^{\prime} \sum_{\vec{r}} v_{\vec{r}}\left(t^{\prime}\right)\hat{a}_{\vec{r}}^{\dagger}\hat{a}_{\vec{r}}}\,.$$ It is straightforward to see that this transformation cancels the part of $H$ (which will be present also on $U^{\dagger}HU$) corresponding to $v_{i}\left(t\right)\hat{a}_{\vec{r}}^{\dagger}\hat{a}_{\vec{r}}$. On the other side, since this does not commute with the kinetic term, a time dependence will be inherited by the hopping terms. For a set of rapidly oscillating function $v_{i}\left(t\right)$ the Hamiltonian can be replaced by an effective one, resulting from time averaging over a period. The new hopping parameters will read: $$t_{\vec{r}j}\rightarrow t_{\vec{r}j}\left\langle e^{i\Delta v_{\vec{r}j}}\right\rangle _\tau$$ where $\left\langle \right\rangle _{\tau}$ stands for the average over a period: $\tau^{-1}\int_{0}^{\tau}dt$ and $\Delta v_{\vec{r}j}=v_{\vec{r}}\left(t\right)-v_{\vec{r+}\hat{j}}\left(t\right)-\left\langle v_{\vec{r}}\left(t\right)-v_{\vec{r+}\hat{j}}\left(t\right)\right\rangle_\tau $.
### Laser-assisted hopping
In this case the effective dynamics is induced by the coupling of the atoms on the optical lattice with a pair of Raman lasers. A fundamental ingredient consists on introducing an energy offset $\Delta$ on neighboring sites. It is enough to consider such scenario along a single direction. Considering a $2D$ lattice: $$H=-t\underset{\vec{r},j}{\sum}\left(\hat{c}_{\vec{r}+\hat{j}}^{\dagger}\hat{c}_{\vec{r}}+\mathrm{h.c.}\right)+\frac{\Delta}{2}\underset{\vec{r}}{\sum}\left(-1\right)^{x}c_{\vec{r}}^{\dagger}\hat{c}_{\vec{r}}+V\left(t\right)$$ where $\vec{r}=\left(x,y\right)$ runs through the lattice sites. The offset term characterized by $\Delta$ can be obtained by tilting the lattice, introducing magnetic gradients or through superlattices. The potential $V\left(t\right)$ is the result of the two external lasers that induce an electric field $E_{1}\cos\left(\vec{k}_{1}\cdot\vec{r}_{1}-\omega_{1}t\right)+E_{2}\cos\left(\vec{k}_{2}\cdot\vec{r}_{2}-\omega_{2}t\right)$. It is assumed that the frequencies are fine-tuned such that they match the offset $\omega_{1}-\omega_{2}=\Delta$. Neglecting fast oscillating terms the potential is written as: $$V\left(t\right)=2E_{1}E_{2}\underset{\vec{r}}{\sum}e^{i\left(\vec{k}_{R}\cdot\vec{r}-\Delta t\right)}c_{\vec{r}}^{\dagger}c_{\vec{r}}+\mathrm{h.c.}$$ with $\vec{k}_{R}=\vec{k}_{1}-\vec{k}_{2}$. Then one can get the effective Hamiltonian in two steps. First performing an unitary transformation $\exp[{-it\frac{\Delta}{2}\sum_{\vec{r}}\left(-1\right)^{x}c_{\vec{r}}^{\dagger}\hat{c}_{\vec{r}}}]$ will create oscillatory hopping terms (with $\exp\left({\pm i\Delta t}\right)$ in front). Then one may apply the previous formalism building an effective Hamiltonian using Equation \[eq:Heff\_oscilator\]: $$\begin{split}
H=&-t\underset{x,y}{\sum}\left(\hat{c}_{x,y}^{\dagger}\hat{c}_{x,y+1}+\mathrm{h.c.}\right)\\
&-\frac{2tE_{1}E_{2}}{\Delta}\underset{x\ \mathrm{even},y}{\sum}\left[\left(e^{i\vec{k}_{R}\cdot\vec{r}}-1\right)\left(e^{i\vec{k}_{R}\cdot\vec{r}}\hat{c}_{x,y}^{\dagger}\hat{c}_{x+1,y}+e^{-i\vec{k}_{R}\cdot\vec{r}}\hat{c}_{x-1,y}^{\dagger}\hat{c}_{x,y}\right)+\mathrm{h.c.}\right]+{\cal O}\left(\Delta^{-2}\right)
\end{split}$$ It is clear that this generates complex hopping and looking more carefully one finds that the lattice has a staggered flux. With a choice $\vec{k}_{R}=\left(\Phi,\Phi\right)$ (as also made in the experiment [@Aidelsburger2011]) one can write upon a gauge transformation: $$\begin{split}
H=&-t\underset{x,y}{\sum}\left(\hat{c}_{x,y}^{\dagger}\hat{c}_{x,y+1}+\mathrm{h.c.}\right)\\
&-\frac{2tE_{1}E_{2}\sin\Phi/2}{\Delta}\underset{x\ \mathrm{even},y}{\sum}\left[\left(e^{i\Phi y}\hat{c}_{x,y}^{\dagger}\hat{c}_{x+1,y}+e^{-i\Phi y}\hat{c}_{x-1,y}^{\dagger}\hat{c}_{x,y}\right)+\mathrm{h.c.}\right]+{\cal O}\left(\Delta^{-2}\right)
\end{split}$$ where it is clear that a sequence of fluxes $\pm\Phi$ alternates in the plaquettes along the $x$ direction. More refined techniques allow for the realization of systems with uniform fluxes [@Goldman2015]. In such systems the Chern number of the Hofstadter bands was measured in [@Aidelsburger2014]. It is worth noting that other kind of one-body terms, beyond the staggered term, can be used as it was done in the first quantum simulations of this model with ultracold atoms [@Aidelsburger2013; @Miyake2013]. In that case a linear potential is used. These kind of approaches can be adapted to more general scenarios including different geometries and multi-component species. The latter, for example, can be achieved by introducing spin dependent potentials as done in [@Aidelsburger2013].
Simulation of gauge fields\[sec:Simulating-gauge-fields\]
=========================================================
In the context of Abelian gauge theories, the goal of simulating gauge fields consists in attributing dynamics to the complex phases on the hopping parameters that were identified in the previous Section. In order to construct such dynamics one should identify degrees of freedom that will play the role of the gauge field. Several proposals have been put forward which map the gauge degrees of freedom into some other controllable variables. The platforms used include ultracold atoms, trapped ions and superconducting qubits. They may be analogue or digital quantum simulators and include Abelian or non-Abelian symmetries [@Weimer2010; @glaetzle2014; @tagliacozzo2013Abelian; @Banerjee2012; @notarnicola2015; @kasper2016; @kapit2011; @zohar2013_2d+1; @bazavov2015; @Zohar2012; @tewari2006; @tagliacozzo2013; @banerjee2013; @stannigel2014; @Zohar2013SU(2); @Zohar2013; @hauke2013; @marcos2013; @marcos2014; @mezzacapo2015; @doucot2003; @laflamme2015; @laflamme2016; @Zohar2017; @Zohar2017digitalgauge; @Brennen2016; @dehkharghani2017]. A more detailed description of two particular approaches in analogue cold atomic simulators will follow: the gauge invariance will be obtained by either penalizing with a large energy cost the non-physical states or by exploiting microscopic symmetries. The symmetries addressed will be $U\left(N\right)$ and $SU(N)$.
There are other symmetries which have been explored, namely $\mathbb{Z}_{n}$ [@Zohar2013; @notarnicola2015] which, in particular, can provide an alternative route towards $U\left(1\right)$ symmetry in the large $n$ limit [@notarnicola2015] and can be addressed with similar approaches. Proposal for the realization of $\mathbb{C}P\left(N-1\right)$ [@adda19781; @eichenherr1978] models have been put forward in [@laflamme2015; @laflamme2016]. These models can serve as toy models for QCD and are also relevant in studying the approach to the continuum limit, in the context of D-theories, where the continuum limit is taken via dimensional reduction [@brower1999; @brower2004]. Furthermore other formulations are possible for specific groups [@Tagliacozzo2014; @mathur2005; @anishetty2009; @ZoharBurrello2015; @ZoharBurrello2016]. Gauge theories with Higgs fields have also been the target of quantum simulation proposals [@Kuno2016; @Kasamatsu2013; @Kuno2017; @GonzlezCuadra2017].
Another relevant approach is the so-called quantum Zeno dynamics which takes inspiration on the quantum Zeno effect, stating that a system being continuously observed does not evolve on time. Furthermore, if the measurement commutes with a certain part of the Hamiltonian, then it can freeze a certain part of the Hilbert space but still enables the dynamics in another subspace [@facchi2002]. This feature can be used in order to freeze gauge dependent quantities and let the system evolve in the gauge invariant subspace. The Hamiltonian to be implemented has the form $H_{\mathrm{noise}}=H_{0}+H_{1}+\sqrt{2\kappa}\sum_{x,a}G_{x}^{a}$ where $H_{0}$ and $H_{1}$ are time independent and are, respectively, gauge invariant and gauge variant parts of the Hamiltonian. The operators $G_{x}^{a}$ are associated to the constraint one wishes to impose $G_{x}^{a}\left|\psi\right\rangle =0$. In the case of gauge theories $G_{x}^{a}$ are the generators of gauge transformations. An advantage of this approach, with respect to the energy punishment approach of the next Section, is that only linear terms on the generators must be imposed on the Hamiltonian (energy punishment requires quadratic terms). By other side leakage from the gauge invariant subspace of the Hilbert space happens as a function of time, which does not happen in the energy penalty approach. This approach was developed in [@stannigel2014].
Another approach, that was successfully implemented in the first quantum simulator of a gauge theory using trapped ions [@martinez2016], is the digital quantum simulator [@lloyd1996]. The key idea consists on in dividing the full time evolution operator $e^{-iHt}$ into smaller pieces of sizes $\tau=t/N$ and apply time evolution of smaller parts of the Hamiltonian at a time. Consider for example an Hamiltonian which is a sum of $M$ contributions : $H=\sum_{\alpha}^{M}H_{\alpha}$. Each part $H_{\alpha}$ can represent, for example, a nearest neighbor spin interaction in which case only two spins are coupled on each $H_{\alpha}.$ For large enough $N$ one can write: $$e^{-iHt}=\left(e^{-iH\tau}\right)^{N}\simeq\left(\underset{\alpha=1}{\overset{M}{\prod}}e^{-iH_{\alpha}\tau}\right)^{N}$$ Each time step can now be interpreted as an individual gate. While in the analogue simulation the great difficult lies on building the appropriate gauge invariant Hamiltonian, in digital quantum simulations that is not a problem. The difficulty lies, however, in building an efficient sequence of gates. Other then the scheme used in the first experimental realization [@Muschik2017], other proposals towards digital quantum simulations of lattice gauge theories have been put forward [@tagliacozzo2013Abelian; @Weimer2010; @tagliacozzo2013; @mezzacapo2015; @Zohar2017; @Zohar2017digitalgauge]
Gauge invariance from energy punishment\[subsec:EnergyPunishement\]
-------------------------------------------------------------------
The energy punishment approach is a quite general approach which allows for the theoretical construction of models that will exhibit a given symmetry in its low energy sector. It consists on building a Hamiltonian which does not prohibit the symmetry violation to occur but instead punishes it with a large energy. In a more concrete way, let suppose one wants to implement a set of symmetries corresponding to a set of generators $\left\{ G_{x}\right\} $ commuting with each other $\left[G_{x},G_{y}\right]=0$. Furthermore consider a typical Hamiltonian $H_{0}$ which does not respect these symmetries. Then one constructs the following Hamiltonian: $$H=H_{0}+\Gamma\underset{x}{\sum}G_{x}^{2}\label{eq:Hpenalty}$$ where $\Gamma$ is a large energy scale, meaning much larger than the energy scales involved in $H_{0}$. Since $G_{x}$ are Hermitian $G_{x}^{2}$ have non-negative eigenvalues. One can choose the lowest eigenvalue to be zero by an appropriate definition of $G_{x}$. Then, at low energy $\left(\ll\Gamma\right)$, the states will respect approximately the condition $G_{x}\left|\psi\right\rangle \simeq0$. If not, this would give a state automatically in an energy scale $\sim\Gamma$. It is then possible to construct an effective Hamiltonian, valid in low energy, which will respect the symmetries generated by $\left\{ G_{x}\right\} $. Let $G$ be the projector operator on the subspace of the total Hilbert space obeying $G_{x}\left|\psi\right\rangle =0$ and let $P=1-G$. Then the low energy Hamiltonian can be written as: $$H_{\mathrm{eff}}=GH_{0}G-\frac{1}{\Gamma}GH_{0}P\frac{1}{\underset{x}{\sum}G_{x}^{2}}PH_{0}G+{\cal O}\left(\Gamma^{-2}\right)\label{eq:Heff_pert_theory}$$ which fulfills the symmetries. Within this framework an effective Abelian gauge theory can be constructed. In non-Abelian theories the generators of the gauge transformation do not commute and this construction fails. There are, of course, several possible drawbacks even on the theoretical level. For example the Hamiltonian \[eq:Heff\_pert\_theory\], even though gauge invariant, may contain unwanted interactions or miss some particular terms which are present on the target system.
In order to construct a quantum simulator the first task is naturally to map the degrees of freedom of the target theory into the laboratory controlled ones, in this case the atomic variables. The matter fields, which are fermionic, will naturally be described by fermionic atomic species. Regarding gauge fields, the target will be the quantum links formulation discussed in Section \[sec:QL\]. Therefore the goal consists on building the quantum links satisfying the algebra $\left[L_{\vec{r},i},U_{\vec{r}',j}\right]=\delta_{ij}\delta_{\vec{r}\vec{r}'}U_{\vec{r}',j}$ and $\left[U_{\vec{r},i},U_{\vec{r}',j}^{\dagger}\right]=\delta_{ij}\delta_{\vec{r}\vec{r}'}2L_{\vec{r},i}/S\left(S+1\right)$.
This can be achieved using the Schwinger representation. Given two bosonic species $b^{\left(\sigma\right)}$ with $\sigma=1,2$ which are associated to each link, one can write $$U_{\vec{r}i}=\frac{1}{\sqrt{S\left(S+1\right)}}b_{\vec{r}i}^{\left(2\right)}{}^{\dagger}b_{\vec{r}i}^{\left(1\right)},\ L_{\vec{r}i}=\frac{1}{2}\left(b_{\vec{r}i}^{\left(2\right)}{}^{\dagger}b_{\vec{r}i}^{\left(2\right)}-b_{\vec{r}i}^{\left(1\right)}{}^{\dagger}b_{\vec{r}i}^{\left(1\right)}\right)\label{eq:Schwinger_bosons}$$ Each link is loaded with a total of $2S$ bosons where $S$ is an half integer. Then one has the desired representation for the quantum links in terms of atomic variables. Now the variables are identified. One then can then build a $d$ dimensional optical lattice where fermions are allowed to hop among lattice points and in each links there are a total of $2S$ bosons. For $1D$, the target Hamiltonian is of the form: $$\label{hamstag}
H=-t\underset{n}{\sum}\left(c_{n}^{\dagger}U_{n}c_{n+1}+\mathrm{h.c.}\right)+m\underset{n}{\sum}\left(-1\right)^{n}c_{n}^{\dagger}c_{n}+\frac{g^{2}}{2}\underset{n}{\sum}L_{n}^{2}$$ When comparing to the general structure of \[eq:General\_Abelian\_QL\] there are two differences: the plaquette term and the position-dependent coefficient of the kinetic term. The plaquettes are naturally absent in $1D$, whereas the tunneling amplitude can be fixed by a gauge transformation $c_{n}\rightarrow\left(-i\right)^{n}c_{n}$. The Hamilton has therefore the required structure and can be targeted with the Schwinger boson approach and it assumes the form: $$H=-\frac{t}{\sqrt{S\left(S+1\right)}}\underset{n}{\sum}\left(c_{n}^{\dagger}b_{n}^{\left(\bar{\sigma}\right)}{}^{\dagger}b_{n}^{\left(\sigma\right)}c_{n}+\mathrm{h.c.}\right)+m\underset{n}{\sum}\left(-1\right)^{n}c_{n}^{\dagger}c_{n}+\frac{g^{2}}{8}\underset{n}{\sum}\left(b_{n}^{\left(2\right)}{}^{\dagger}b_{n}^{\left(2\right)}-b_{n}^{\left(1\right)}{}^{\dagger}b_{n}^{\left(1\right)}\right)^{2}\label{eq:H_U(1)_SchwingerBosons}$$ The two last terms can be, in principle, implemented directly using a proper tuning of the interactions between the bosons and the potential for the fermions. The first term, instead, is a correlated hopping between bosons and fermions which is obtained less easily. Furthermore the additional terms that are not gauge invariant, like $b_{n}^{\left(\bar{\sigma}\right)}{}^{\dagger}b_{n}^{\left(\sigma\right)}$ and $c_{n}^{\dagger}c_{n+\hat{i}}$, must be suppressed. This is solved by the energy punishment approach. In general the non-gauge invariant Hamiltonian with the ingredients described has the form: $$\begin{split}
H_{0}=&-\underset{n,i}{\sum}\left[t_{F}\left(c_{n}^{\dagger}c_{n+1}+\mathrm{h.c.}\right)-t_{B}\left(b_{n}^{\left(2\right)}{}^{\dagger}b_{n}^{\left(1\right)}+\mathrm{h.c.}\right)\right]\\
&+\underset{n}{\sum}\left(v_{n}^{F}c_{n}^{\dagger}c_{n}+\underset{\sigma}{\sum}v_{n}^{B\sigma}b_{ni}^{\left(\sigma\right)}{}^{\dagger}b_{ni}^{\left(\sigma\right)}\right)+U\underset{n}{\sum}\left(b_{n}^{\left(2\right)}{}^{\dagger}b_{n}^{\left(2\right)}-b_{n}^{\left(1\right)}{}^{\dagger}b_{n}^{\left(1\right)}\right)^{2}\label{eq:H0_energypunishment}
\end{split}$$ Using the generators for the $U\left(1\right)$ gauge symmetry in Equation \[eq:u1-generators\] one considers the full Hamiltonian: $$H=H_{0}+\Gamma\underset{n}{\sum}\left(L_{n}-L_{n-1}-c_{n}^{\dagger}c_{n}+\frac{1-\left(-1\right)^{n}}{2}\right)^{2}$$ It is crucial that one has access to the interactions that are introduced on the last term corresponding to the energy punishment. To see that this is the case it useful to be more specific about the labels $\sigma$. One can take, as in [@Banerjee2012], the labels $\sigma=1,2$ meaning respectively left and right part of the link, which can be thought to coincide with the lattice site. In this way $b_{n}^{\left(2\right)}{}^{\dagger}b_{n}^{\left(1\right)}$ are just regular hopping terms. Furthermore it is recalled that the total number of bosons associated to each link is conserved. Therefore one can write: $L_{n}=-S+b_{n}^{\left(2\right)}{}^{\dagger}b_{n}^{\left(2\right)}=S-b_{n}^{\left(1\right)}{}^{\dagger}b_{n}^{\left(1\right)}$. This means that terms like $L_{n}^{2}$ and and $L_{n}L_{n-1}$ can be written as a density-density interaction. Regarding the last case, recall that $b_{n}^{\left(1\right)}$ and $b_{n-1}^{\left(2\right)}$ are effectively in the same site, see Figure \[fig:Superlattice-punishement\]. Now Equation \[eq:Heff\_pert\_theory\] can be applied. The number of particles in each site is a good quantum number to describe the eigenstates of $G_{x}$. The number of particles in the site $j$ are denoted by $n_{j}^{F}=c_{j}^{\dagger}c_{j}$, $n_{j}^{1}=b_{j}^{\left(1\right)}{}^{\dagger}b_{j}^{\left(1\right)}$ and $n_{j}^{2}=b_{j-1}^{\left(2\right)}{}^{\dagger}b_{j-1}^{\left(2\right)}$. The subspace of gauge invariant states is then characterized by: $$n_{j}^{F}+n_{j}^{1}+n_{j}^{2}=2S+\frac{1-\left(-1\right)^{j}}{2}$$
![Superlattice configurations for the two boson species and the fermionic one. Bosons of the species $1$ at an even site $2j$ can only hop to $2j-1$ while a boson of species $2$ has only access to the site $2j+1$. The Figure presents a an example of a gauge invariant state configuration (on these three sites) where $G_{x}\left|\psi\right\rangle =0$. \[fig:Superlattice-punishement\]](figure1)
In the lowest order only the two last terms of \[eq:H0\_energypunishment\] survive as any single hopping destroys the above relation. At the next order there are three possible virtual processes that preserve this condition. Up to some linear terms on the particle density operator, they are:
1. \[cond1\] Boson-boson hopping: a boson hops to the neighboring site on the same link and another boson hops back. Gives rise to a boson density-density interaction.
2. \[cond2\] Fermion-Fermion hopping: a fermion hops to a neighboring site and then hops back. Only possible if neighboring site is unoccupied and gives rise to a nearest neighbor fermion density-density interaction.
3. \[cond3\] Boson-Fermion hopping: a fermion hops to a neighboring site and a boson belonging to the link that connects the two sites does the opposite path. Gives rise to a correlated hopping.
The terms coming from \[cond1\] should be joined with the last term of \[eq:H0\_energypunishment\] in order to form the correct kinetic term for the gauge fields. The terms in \[cond2\] are somehow unwanted and correspond to a repulsion between neighbor fermions $n_{j}^{F}n_{j+1}^{F}$. Naturally, they do not spoil gauge invariance and their inclusion should not be a problem [@Banerjee2012]. Finally the terms originating from \[cond3\] give rise to the correlated hopping responsible for the matter-gauge coupling as written on the first term of \[eq:Schwinger\_bosons\]. There is another issue which should be addressed. From the beginning it was assumed that the the number of bosons in each link is conserved. In particular this means that bosons are not allowed to pass to a neighboring link. In order to guarantee this condition in an experiment one should introduce an extra bosonic species and this is the reason that bosons in neighboring links were represented with different colors on Figure \[fig:Superlattice-punishement\]. Then one bosonic species is trapped on the even links and the other in the odd links. This will prevent bosonic hopping between links. A numerical study of real dynamics of the the model as well as accuracy of the effective gauge invariance obtained was also done in [@Banerjee2012].
Finally, in a possible experimental realization, the first fundamental step is to guarantee that the system is initialized on a gauge invariant state. This can be done by loading the atoms in a deep lattice such that they are in Mott phase. Afterwards the system should evolve according to the fine tuned Hamiltonian described above (after lowering the lattice barriers). Finally measures of relevant quantities can be performed.
This principle is valid in higher dimensionality where one has to face the difficulty of generating plaquette terms. This was done for the pure gauge in [@Zohar2011] and [@Zohar2012] by suitably allowing hopping between links. In the first case each link has an infinite dimensional Hilbert space that is represented by a Bose-Einstein condensate. In the second the proposal is simplified by considering a quantum link model.
Gauge invariance from many body interaction symmetries
------------------------------------------------------
This approach consists on building a lattice which will have the necessary local gauge invariance arising from microscopic symmetries. Specific proposals may vary significantly even though the same principle is used. For example in [@dehkharghani2017] the simulation is built upon the global symmetry conserving the total number of excitations and is achieved via a state-dependent hopping. In turn, see for example [@Zohar2013; @Kasper2017], are built upon conservation of angular momentum. For concreteness the later approach will be described in more detail below. In the case of [@banerjee2013] $SU\left(N\right)$ symmetries of the ground state manifold of alkaline-earth-like atoms could be exploited in order to built non-Abelian gauge theories. Symmetries only allow for certain type of processes to occur and, by exploiting these constraints, one can build a gauge symmetry. This can be done, as said before, considering angular momentum conservation. The Schwinger model is taken as an illustrative example. Bosons, that will make up the gauge fields, are placed at the two boundaries of the links. Because the goal consists, partially, in forbidding gauge dependent terms like simple boson or fermion hopping, the lattice should be spin dependent. In this way a single hopping is forbidden as it does not conserve angular momentum. By other side one should guarantee that correlated spin between bosons and fermions is allowed. This can be achieved by a judicious choice of respective hyperfine angular momentum in each lattice site. For concreteness, consider a single link connecting two sites and a total of two bosonic ($b^{\left(1\right)}$,$b^{\left(2\right)}$) and two fermionic species ($c$,$d$). The site at the left of the link can only be populated by $c$ while the right side by $d$. Analogously the left end of the link can only be populated by $b^{\left(1\right)}$ while the right end can only be populated by $b^{\left(2\right)}$. Then the conditions described above for allowed/forbidden hopping are automatically satisfied if one chooses the hyperfine angular momentum of each atomic species to satisfy: $$m_{F}\left(d\right)-m_{F}\left(c\right)=m_{F}\left(b^{\left(1\right)}\right)-m_{F}\left(b^{\left(2\right)}\right)$$ It is intended that the lattice is, indeed, spin dependent so that $m_{F}\left(d\right)\neq m_{F}\left(c\right)$ and $m_{F}\left(b^{\left(1\right)}\right)\neq m_{F}\left(b^{\left(2\right)}\right)$. In other words, what this means is that the difference of angular momentum caused by a fermion hop can be exactly compensated by a bosonic hop in the opposite direction. This leads directly to the correlated hopping desired which, in fact, comes from the scattering terms between bosons and fermions. The only other allowed scattering term between fermions and bosons correspond to density-density interactions like $c^{\dagger}c\left(b^{\left(2\right)}{}^{\dagger}b^{\left(2\right)}+b^{\left(1\right)}{}^{\dagger}b^{\left(1\right)}\right)$. These are just linear terms on the fermionic number operator due to the conservation of the total number of bosons per link. Summing over all lattice sites will give just a constant shift of the energy. The scattering terms between bosons give rise to the gauge kinetic term as before (in $1+1$ dimensions).
Again, for higher dimensionality, there is a non-trivial extra step consisting on building plaquette interactions. If plaquettes are ignored and the model described above is loaded on an higher dimensional lattice the result corresponds to the strong coupling limit of the gauge theory.
The plaquette terms can be achieved by the so-called loop method. It uses perturbation theory in a similar way that was used in the energy penalty approach. In order to discuss the essence of the construction of the plaquette terms, one can consider just the pure gauge theory. The target Hamiltonian is $$H_\mathrm{target}=\frac{g^{2}}{2}\underset{\vec{r},i}{\sum}L_{\vec{r}i}^{2}-\frac{1}{4g^{2}}\underset{\square}{\sum}\left(U_{\square}+U_{\square}^{\dagger}\right).$$ The description will be specialized for $2+1$ dimensions but the theoretical construction for higher dimensions is analogous. The construction of the plaquette term relies on a perturbative expansion similar to \[eq:Hpenalty\] but, in this case, $H_{0}$ is already a gauge invariant Hamiltonian. For reasons that will be explained below one should have two fermionic species, say $\chi$ and $\psi$, and build the trivial part of the generalization of the $1+1$ process: $$H_{0}=-t\underset{\vec{r},i}{\sum}\left(\psi_{\vec{r}}^{\dagger}U_{\vec{r}i}\psi_{\vec{r}+\hat{i}}+\chi_{\vec{r}}^{\dagger}U_{\vec{r}i}\chi_{\vec{r}+\hat{i}}+\mathrm{h.c.}\right)+\frac{g^{2}}{2}\underset{\vec{r},i}{\sum}L_{\vec{r}i}^{2}\,.\label{eq:H0_LoopMethod}$$ The fermionic species are auxiliary and in the effective model they will be integrated out. There should be no interacting term between them. Here the energy penalty must enforce the following conditions at each site $\vec{r}=\left(r_{1},r_{2}\right)$:
- there is a fermion $\psi$ if both $r_{1}$ and $r_{2}$ are even
- there is a fermion $\chi$ if both $r_{1}$ and $r_{2}$ are odd
- no fermion otherwise
The positions of these fermions is represented on Figure \[fig:loop\_method\] a). This kind of constraint can be obtained, for large $\Gamma$, with a Hamiltonian of the form: $$H_\mathrm{penalty}=-\Gamma\underset{\vec{r}}{\sum}\left[\frac{\left(1+\left(-1\right)^{r_{1}}\right)\left(1+\left(-1\right)^{r_{2}}\right)}{4}\psi_{\vec{r}}^{\dagger}\psi_{\vec{r}}+\frac{\left(1-\left(-1\right)^{r_{1}}\right)\left(1-\left(-1\right)^{r_{2}}\right)}{4}\chi_{\vec{r}}^{\dagger}\chi_{\vec{r}}\right]\label{eq:H_loop}$$ Through perturbation theory, according to \[eq:Heff\_pert\_theory\], one gets the plaquette terms at fourth order. This process is “cleaner” if the $U_{\vec{r}}$ in \[eq:H0\_LoopMethod\] are considered unitary. IN particular we may consider a unitary limit,in which the total spin of the quantum link goes to infinity: $S\rightarrow =+\infty$. Order by order:
1. Only the pure gauge part of \[eq:H0\_LoopMethod\] contributes, no fermionic term occurs.
2. Trivial constant contribution assuming that $U_{n}$ are unitary. The virtual process giving rise to this contribution is a single link interaction where a fermionic-bosonic correlated hopping occurs back and forth restoring the initial state. There are never fermions on the neighbor lattice site. In turn in the unitary limit there is an infinite number of bosons such that $\left[U,U^{\dagger}\right] \to 0$. In the case of finite bosonic number, extra contribution corresponding to a renormalization of the pure gauge term of \[eq:H0\_LoopMethod\] will appear, together with another term which can be discarded by application of the Gauss law.
3. Trivial constant contribution assuming that $U_{n}$ are unitary. Virtual contributions evolving links constitute again back and forth hopping plus a pure gauge term at any stage of the process. The extra contributions coming from considering a finite number of boson per link cannot be disregarded trivially as second order for this case.
4. Gives the desired plaquette term plus renormalization of the pure gauge term of \[eq:H0\_LoopMethod\] assuming that $U_{n}$ are unitary. The last case corresponds to the virtual process where a fermion goes around a plaquette and returns to the initial place. This virtual process is represented on Figure \[fig:loop\_method\] b). Naturally, in the non-unitary case, more terms appear.
![Loop method for obtaining the plaquette terms. In the panel a) it is depicted the positions of the auxiliary fermions that are used to construct the plaquette term using gauge invariant building blocks. One of the species, say $\psi$, is represented in red and placed on sites with both coordinates even. In turn $\chi$, in pink, is placed on sites with both coordinates odd. This correspond to the ground-state of \[eq:H\_loop\]. In the panel b) it is represented a virtual process that gives rise to a plaquette term\[fig:loop\_method\].](figure2)
Plaquette terms only appear at fourth order. However, in the unitary limit, most contributions are trivial. One can then see that it is effectively a second order contribution [@zohar2015].
When one considers a finite number of bosons in the links there are extra contributions appearing which cannot be disregarded. As in the case of the energy penalty, these contributions, even though unwanted, can be tolerated as they are naturally gauge invariant. However one should guarantee that these extra contributions are not more important than the plaquette term which is the target term. That can be achieved if the coupling term is parameterized is $g^{2}$ is taken to be small in units of $t$. By taking $g^{2}\sim t^{2}/\Gamma$ one makes the unwanted terms at third order effectively of the same order as the plaquettes and unwanted terms of the fourth order effectively of higher order than the plaquettes.
On top of these, an extra species of fermions can be introduced to play the role of matter fields. They will consist, in the initial Hamiltonian, to the usual correlated hopping with the bosons. Furthermore the staggered mass term (of Equation \[eq:H\_KG\_general\_d\]) should also be introduced. In the unitary case this extra piece commutes with the interacting part of \[eq:H0\_LoopMethod\] and no further contribution is obtained in perturbation theory. In the truncated case there is an extra (gauge invariant) correlated hopping coming at third order. Another different aspect of the introduction of dynamical fermions is that the Gauss law ($\sum_{i}L_{\vec{r}i}-L_{\vec{r}-\hat{i},i}=\mathrm{const}$) can no longer be used to trivialize terms. The divergence of the electric gets a contribution from the charge density of the dynamical fermions. Nonetheless it can still be employed and the extra charge density terms can be compensated on the initial Hamiltonian if proper fine tuning is available experimentally.
In [@Kasper2017] it was proposed a realization of the Schwinger ($1+1$) model using a mixture of $^{23}\mathrm{Na}$ for the bosons and $^{6}\mathrm{Li}$ for the fermions as well as an extensive study on the influence of the finiteness of the number of bosons per link in that case.
Encoding in $1+1$ fermions \[subsec:Encoding-in-\]
--------------------------------------------------
The case of the Schwinger model, $1+1$ Dirac fermions coupled to a gauge field, is an interesting experimental and theoretical playground. It shares some non-trivial features with QCD like confinement, chiral symmetry breaking and a topological theta vacuum [@wiese2013]. However, due to its simplicity, it allows for analytical and numerical studies which may become significantly harder in more complicated theories. Furthermore it was the target of the first experimental implementation of a lattice gauge theory [@martinez2016]. In the context of quantum simulations it may not only provide the entrance door towards more complicated experimental realizations but also a way of benchmarking experimental techniques.
One of the reasons why this model bares an intrinsic simplicity, as mentioned previously, is the fact that the gauge fields are non-dynamical. This is reflected on the absence of plaquette terms in the Hamiltonian formulation. Furthermore the Gauss law fixes the gauge field and can be used to integrate out its degrees of freedom. This results in a long-range interacting model which will be addressed next. In the following the lattice Hamiltonian formulation is considered for $N$ lattice sites: $$H=-it\overset{N-1}{\underset{n=1}{\sum}}\left[c_{n}^{\dagger}U_{n}c_{n+1}-\mathrm{h.c}\right]+m\overset{N}{\underset{n=1}{\sum}}\left(-1\right)^{n}c_{n}^{\dagger}c_{n}+\frac{g^{2}}{2}\overset{N-1}{\underset{n=1}{\sum}}L_{n}^{2}\label{eq:PBC_KG_Schwinger}$$ Here an infinite dimensional Hilbert space per link is considered, therefore the operators $U_{n}$ are unitary and the non-trivial commutation relations on the links are given by $\left[L_{m},U_{n}\right]=U_{n}\delta_{mn}$. Equivalently the link can be written as $U_{n}=e^{i\theta_{n}}$. The Gauss law is imposed in accordance with the relations and . This model can be formulated in terms of Pauli spin operators [@Banks1976] through the Jordan-Wigner transformation: $$\left\{ \begin{array}{c}
c_{n}=\underset{l<n}{\prod}\left(i\sigma_{z}\left(l\right)\right)\sigma^{-}\left(n\right)\\
c_{n}^{\dagger}=\underset{l<n}{\prod}\left(-i\sigma_{z}\left(l\right)\right)\sigma^{+}\left(n\right)
\end{array}\right.\label{eq:Jordan_Wigner}$$ where $\sigma_{i}\left(l\right)$ represent the Pauli matrices in the site $l$ and $\sigma^{\pm}\left(n\right)=\sigma_{x}\left(n\right)\pm i\sigma_{y}\left(n\right)$. In terms of the spins the Gauss law is determined by: $$G_{n}=L_{n}-L_{n-1}-\frac{1}{2}\left(\sigma_{z}\left(n\right)+\left(-1\right)^{n}\right)\,.$$ By restricting ourselves to the physical space, through the Gauss law $G_{n}\left|\psi\right\rangle =0$, the link variables can be almost completely eliminated. Using periodic boundary conditions ($L_{0}=L_{N}$) one finds: $$L_{n}=L_{0}+\frac{1}{2}\underset{l=1}{\overset{n}{\sum}}\left(\sigma_{z}\left(l\right)+\left(-1\right)^{n}\right)\,.$$ The value of $L_{0}$ is a parameter of the theory and corresponds to a background field. For simplicity it will be taken to zero at the present discussion. By using the above relations the Hamiltonian \[eq:PBC\_KG\_Schwinger\] can be rewritten as: $$H=t\overset{N}{\underset{n=1}{\sum}}\left[\sigma^{+}\left(n\right)e^{i\theta_{n}}\sigma^{-}\left(n+1\right)+\mathrm{h.c}\right]+\frac{m}{2}\overset{N}{\underset{n=1}{\sum}}\left(-1\right)^{n}\sigma_{z}\left(n\right)+\frac{g^{2}}{8}\overset{N}{\underset{n=1}{\sum}}\left[\underset{l=1}{\overset{n}{\sum}}\left(\sigma_{z}\left(l\right)+\left(-1\right)^{n}\right)\right]^{2}\label{eq:encoded_H}$$ where a trivial constant term was dropped. The remaining gauge field variable $\theta_{n}$ can be eliminated by a residual gauge transformation [@Hamer1997]: $$\sigma^{\pm}\left(n\right)\rightarrow\sigma^{\pm}\left(n\right)\underset{j<n}{\prod}e^{\pm i\theta_{j}}\,.$$ This is a non-trivial transformation as $\theta$’s are operators. More precisely, the above relation should be seen as defining a new set of operators $\bar{\sigma}^{\pm}\left(n\right) =\sigma^{\pm}\left(n\right)\prod_{j<n}e^{\pm i\theta_{j}}$. The $\bar{\sigma}$ still respect the angular momentum algebra between each other. Therefore they are still spin operators on the sites of the lattice, despite acting non-trivially on the links. Since the links degrees of freedom are being traced out using the Gauss law, one can arrive at an effective spin model for the sites. Plugging this transformation and expanding the interaction term, the resulting model is a long-range interacting spin model: $$H=t\overset{N}{\underset{n=1}{\sum}}\left[\sigma^{+}\left(n\right)\sigma^{-}\left(n+1\right)+\mathrm{h.c}\right]+\overset{N}{\underset{n=1}{\sum}}\left(\frac{m}{2}\left(-1\right)^{n}-\frac{g^{2}}{8}\left(1-\left(-1\right)^{n}\right)\right)\sigma_{z}\left(n\right)+\frac{g^{2}}{4}\overset{N-2}{\underset{n=1}{\sum}}\underset{l=1}{\overset{N-1}{\sum}}\left(N-l\right)\sigma_{z}\left(n\right)\sigma_{z}\left(l\right)$$
This is a useful formulation for quantum simulations since the total of $N$ particles and $N-1$ gauge fields are simulated by just $N$ spins (with exotic long-range interactions), thanks to the gauge invariance. The difficulty was moved towards an efficient way of implementing the long-range asymmetric interaction between spins. This Hamiltonian was implemented as a digital quantum simulator in [@martinez2016] using trapped ions ($^{40}\mathrm{Ca}^{+}$). The system was composed of four qubits. The Schwinger mechanism of pair creation of particle-antiparticle was explored, as well as real time evolution of entanglement in the system. Based on the staggering prescription in Section \[sec:dynamical\], a particle on an odd site corresponds to the vacuum and a hole as an antiparticle (the contrary holds for particles in the even sites). Following this picture the number of particles at the site $n$ is given by $\nu_{n}=\left(1-\left(-1\right)^{n}\right)/2+\left(-1\right)^{n}c_{n}^{\dagger}c_{n}$ and therefore a relevant observable is the particle density $\nu\left(t\right)=\left(2N\right)^{-1}\sum_{n}\left\langle 1+\left(-1\right)^{n}\sigma_{z}\left(n\right)\right\rangle $. Starting from a bare vacuum ($\nu\left(0\right)=0$) it is observed a rapid increase of the particle density followed by a decrease which is due to particle/anti-particle recombination. Also the vacuum persistence $G\left(t\right)=\left\langle 0\right|e^{-iHt}\left|0\right\rangle $ and entanglement were evaluated. The latter is done by reconstructing the density matrix and evaluating the entanglement in one half of the system with the other half through logarithmic negativity. Entanglement is produced through particle creation that get distributed across the two halves. More detail on the simulation and experimental results can be found in [@martinez2016; @Muschik2017]. Future challenges include the simulation of larger systems as well higher dimensionality and non-Abelian symmetries.
The authors want to thank the organizers of the Natal’s school: Pasquale Sodano, Alvaro Ferraz, Kumar S. Gupta and Gordon Semenoff. They are also pleased to thank the participants to the course, in particular T. J. G. Apollaro, V. E. Korepin, T. Macrì, G. Mussardo, E. Tonni and J. Viti, for useful discussions and to have contributed to set a stimulating and pleasant atmosphere. Finally, special thanks go to Marcello Dalmonte for discussions and common work on the topics discussed in this chapter.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'combined with computer vision systems, such as convolutional neural networks, offer a flexible and affordable solution for terrain monitoring, mapping, and detection tasks. However, a key challenge remains the collection and annotation of training data for the given sensors, application, and mission. We introduce an system that incorporates novelty estimation into its objective function, based on research for uncertainty estimation in deep learning. The system is designed for data collection to reduce both the number of flights and of annotated images. We evaluate the approach on real world terrain mapping data and show significantly smaller collected training dataset compared to standard lawnmower data collection techniques.'
author:
- |
Hermann Blum$^{1}$, Silvan Rohrbach$^{1}$, Marija Popović$^{1}$, Luca Bartolomei$^{2}$, Roland Siegwart$^{1}$\
\
$^{1}$Autonomous Systems Lab, ETH Zürich\
$^{2}$Vision for Robotics Lab, ETH Zürich [^1]
bibliography:
- 'PaperpileMay15.bib'
title: '**Active Learning for UAV-based Semantic Mapping**'
---
Introduction
============
offer cost-efficient, flexible and automated delivery of high-quality sensing data in various applications, search-and-rescue, inspection and agricultural monitoring. Image sensors in particular are well suited for based sensing because of their low cost, size, and power demand. Recent advances in computer vision and deep learning enable automated analysis of the gathered data and make the system applicable to a range of monitoring and mapping tasks in large or hard-to-access environments.
In this work, we look into the problem of data collection for environmental monitoring. Especially in mentioned applications like long-term agricultural monitoring, imagery can change drastically over time and differ from the training data. Cases of novelty require tedious data collection, annotation and retraining, with two time-costly challenges: On the one hand the flight time is limited due to the energy consumption of the , on the other hand the time investment for manually annotating the large pool of acquired images is huge. By flying over an area in a conventional predetermined lawnmower-fashion, energy is wasted on gathering repetitive and similar images that will not significantly improve the segmentation and classification quality. We propose an system that actively searches for and gathers data different to the training distribution of the available semantic segmenter.
Based on results from the recent ‘Fishyscapes’ benchmark for novelty detection in semantic segmentation [@Blum2019-eh], we propose an system that maximises the novelty of the gathered images in a single flight mission. Given the available resources, we reduce the number of flights and annotated data, while achieving faster improvements of the semantic segmentation.
An illustration of our approach is given in figure \[fig:teaser\]. For every new image captured by the , we estimate the novelty of the image and follow this information to high-interest regions.
The contributions of this work are the following:
- We propose an algorithm that uses novelty estimation from deep learning as primary source of information.
- We evaluate the proposed solution and the novelty estimation towards the problem of active learning.
{width="\linewidth"}
Related Work {#sec:related_work}
============
[ [**Novelty detection **]{}]{} or uncertainty estimation for deep learning models is a very active area of research. Uncertainty and mistakes in prediction algorithms can come from noise in the input, but also from differences between the training data and the input. In our case, we are exclusively interested in the second part, as we want to decrease the distance between input and training distributions by gathering a broader range of training data with our setup.\
Hendrycks [@Hendrycks2017-ua] give a comprehensive overview of the problem and evaluate the baseline of the softmax output. Bayesian Deep Learning [@Gal2016-mx; @Kendall2017-jy] estimates uncertainty from deep models whose outputs and weights are probability distributions. Different works compare the flow of data through the network to the training data and estimate uncertainty as deviation from the training distribution [@Mandelbaum2017-ti; @Papernot2018-xz]. [@Blum2019-eh] adapted and evaluated these approaches towards novelty detection in semantic segmentation. As a third direction of research, learned representations of the input are reconstructed and compared in input space [@Pidhorskyi2018-qf; @Lis2019-vd].
[ [**Active Sampling **]{}]{} is known in machine learning as a technique for data reduction to speed up training times. Wang [@Wang2017-tz] show a system that samples from training images based on the softmax confidence. Gal [@Gal2017-hn] develop a similar system based on Bayesian Deep Learning, which is more suited for novelty detection. Both systems focus on reducing the expensive labelling and do not take into account the problem of data acquisition.
[ [**Informative Path Planning **]{}]{} has recently experienced increasing interest for a variety of applications, such as environmental monitoring [@Hitz2017-kh; @Hollinger2014-md; @Ghaffari2019], surveillance [@vivaldini2016] and inspection [@Bircher2018]. The aim is efficient continuous or discrete data acquisition in complex environments using a mobile robot, whose motions are constrained by its sensing and mobility capabilities. Popular techniques include Partially Observable Markov Decision Processes (POMDPs) [@Kaelbling1998] or Gaussian Processes [@Rasmussen2005]. However, the sensing scenario represents an extreme case where only information about past places in the trajectory is available, which makes path planning for more than a few steps unfeasible and reduces the applicability of the mentioned methods.\
While they do not use novelty as input to a path planner, Richter [@Richter2017-wg] proposed one of the first path planning systems with novelty detection based on deep learning. They use reconstruction based novelty to safely switch between a neural network-based obstacle avoidance controller and a slower conservative planner. In our method, we directly exploit the novelty information to steer the robot towards more informative regions.
Method {#sec:method}
======
Novelty Detection
-----------------
Our novelty estimation approach follows results from [@Mandelbaum2017-ti; @Papernot2018-xz; @Blum2019-eh] and uses density estimation in the feature space. In particular, we use kernel density estimation to produce patchwise uncertainty estimates. Lower resolution feature vectors from a convolutional neural network are compared to their nearest neighbors from the training distribution. Novelty is here defined as the average cosine distance to neighbors from the training distribution, a metric that was shown to work well in different scenarios in the aforementioned works. However, we note that the proposed system setup is independent of the underlying technique for novelty detection as long as it produces pixel- or patchwise values.
Given a training set of images $\bm{A}$, we extract embeddings $\bm{Z}_l = f_l(\bm{A})$ at layer $l$ from the segmentation network and store them in a database. For a given input image $\bm{a}'$ and embeddings $\bm{z}_l' = f_l(\bm{a}')$, we then approximate the probability density of the input image with respect to the distribution of training data using a kernel density estimation of $\bm{z}_l'$ with respect to the $k$ nearest neighbors in $\bm{Z}_l$. This can be found through $$\begin{aligned}
D(\bm{z}') = \sum \limits_{i=0}^{k-1} \frac{\bm{z}'\bm{z}^{(i)}}{|\bm{z}'|\,|\bm{z}^{(i)}|}.\end{aligned}$$
$D(\bm{z}')$ is a patchwise uncertainty estimation of the current input image. The size of the patches is dependent on the layer $l$ where the embeddings are extracted, usually the resolution is 8 or 16 times smaller than the input image. An example of the uncertainty estimation is shown in figure \[fig:teaser\].
For our input image size, the above approach requires 64 nearest neighbor searches per input image and is therefore not feasible for real-time. However, it can be directly switched out with flow-based density estimation as was very recently shown in [@Blum2019-eh], which only requires a single pass through a network.
Path Planning
-------------
The objective in our problem is to find new, informative images for training. The difficulty of approaching this problem is twofold. First, the information computed by the novelty detection is available only in locations that have already been visited by the ; second, the distribution of novelty over the explorable space is unknown. In this work, we adopt the assumption of spatial correlation, while we do not pose any hypothesis on where to find novel inputs, we assume that there are regions of connected novelty cells scattered over the map, rather than an i.i.d. uniform distribution.\
Given a grid discretization of the world, the path-planning problem is stated as active sampling from the adjacent cells of the current position in the grid map. An illustration of the problem is given in figure \[fig:DiscGrid\].
At each re-planning step, the following information is available to the path planner:
- number of additional explorable cells ( battery life);
- explored cells in the map;
- distances to the borders of the explorable world;
- average novelty score of the current cell image;
- gradient direction of the novelty score from the current cell image.
We implement the path planning on basis of potential fields, where each grid cell $\bm{p}$ on the map has an assigned potential $\varphi(\bm{p})$: $$\begin{aligned}
\varphi(\bm{p}) &= D(\bm{p}) + \textrm{penalty}_\textrm{border}(\bm{p}) + \textrm{penalty}_\textrm{visited}(\bm{p})\end{aligned}$$
The novelty of a grid cell image $D(\bm{p})$ is initialized uniformly to a constant value for all unknown patches and updated once a patch has been observed. $\textrm{penalty}_\textrm{border}(\bm{p})$ increases the potential towards the border of the observable world. This term is required only in simulated environments, in order to avoid situations where the gets stuck in corners of the simulation environment. $\textrm{penalty}_\textrm{visited}(\bm{p})$ is a constant penalty applied to cells that have already been visited and therefore should be avoided.
$t \gets 0$ follow shortest path to non-visited patch update $\varphi(\bm{x}_t)$ with observed novelty $\hat{\varphi}(\bm{p}) \gets \varphi(\bm{p})$ $\hat{\varphi}(\bm{p}) \gets \hat{\varphi}(\bm{p}) - f(\bm{p}, \nabla D(\bm{x}_t))$ $\hat{\varphi}(\bm{p}) \gets \hat{\varphi}(\bm{p}) \circledast {\tiny\frac{1}{9}\begin{bmatrix}1\hspace{-7pt} & 1\hspace{-7pt}& 1 \\ 1\hspace{-7pt} & 1\hspace{-7pt}& 1 \\ 1\hspace{-7pt} & 1\hspace{-7pt} & 1\end{bmatrix}}$ $\hat{\varphi}(\bm{x}_\textrm{straight}) \gets \hat{\varphi}(\bm{x}_\textrm{straight}) + \textrm{penalty}_\textrm{forward}$ choose direction with lowest $\hat{\varphi}(\bm{p})$ $t \gets t + 1$ return home
Based on the defined potential field, the drone selects one of the 4 adjacent grid cells to its current location $\bm{x}_t$ according to algorithm \[alg:IPP\]. The algorithm follows a scheme of fast traversion of low-novelty areas and exhaustive exploration of high-novelty areas. We distinguish 3 different cases, depending on the novelty of position $\bm{x}_t$:
low novelty
: is defined as $D(\bm{x}_t) \leq \alpha$. These are images uninteresting for training. The path planner tries to avoid these regions and follows the gradient of the potential field towards more informative areas.
medium novelty
: is defined by $\alpha < D(\bm{x}_t) \leq \beta$. It can depict borders between different classes, as well as borders to more informative areas. The path planner takes the gradient of the current novelty heatmap as additional information into account, as it might be directed towards high novelty regions.
high novelty
: is defined by $D(\bm{x}_t) > \beta$. It identifies regions that contain crucial data for training. Instead of following the gradient, we found that it is more helpful to explore larger areas of high-average novelty. In order to encourage exploration, we smooth the potential field to not disturb the path planner with local fluctuation of the novelty estimation function. Moreover, we add $\textrm{penalty}_\textrm{forward}$ for moving out of high-novelty regions.
We set $\alpha$ and $\beta$ based on the lower and upper quartile thresholds on a validation set. The overall goal of the algorithm is to catch as many high-novelty cells in a mission as possible.
Evaluation {#sec:evaluation}
==========
We evaluate our approach on the remote sensing task of the RIT18 [@Kemker2018-kc] dataset. The dataset contains high-resolution hyperspectral images of the same location for two points in time, suitable as a training and validation dataset. For both images, ground truth annotations are given. We use the classes *grass*, *tree*, *beach*, and *other*. We simulate the flight by laying a grid over the image with cell size $128 \times 128 \textrm{ px}$.
The images gathered from each grid cell are segmented using a fully convolutional network [@Long2015-fm] with a VGG-16 encoder [@Simonyan2015-zf], in particular the implementation from [@Blum2018-kp]. We use the embeddings from the *conv5-1* layer and sample the density over 20 nearest neighbors.
The experiment is set up as follows. The is sent on multiple missions, each time with the objective to gather new training data. After each mission, we add the new images together with annotations to the pool of training data, retrain the semantic segmenter, and build a new kNN database. The new semantic segmenter is then used for the novelty estimation in the next mission.\
At every iteration we evaluate the accuracy of the segmenter on the whole map measured in .
To evaluate our approach, we compare against two different lawnmower baselines:
a - big lawnmower
: For the conventional approach flights with the UAV are simulated in lawnmower fashion across the whole site. A starting point near the edge of the site is chosen and from there line after line, back and forth across the map images are acquired until there is no energy left and the (re-)training of the network is performed.
b - small lawnmower
: For more diverse image acquisition with less flights, the small-scale lawnmower approach works by manually choosing different starting positions spread out across the whole site. From each of the starting positions a flight with a small-scale lawnmower approach is executed which results in the collection of imagery in a rectangle with a size depending on the energy capacity of the UAV. After each flight, the gathered image patches are then used for (re-)training.
c -
: To test the developed IPP system the same starting positions as the one used for the small lawnmower approach are used. Instead of predetermined paths, from the second flight on the IPP system is used to guide the UAV autonomously until the energy is depleted, then the vehicle heads back to the starting position and the network is (re-)trained with the acquired images.
![Comparison of the mean IoU on the full map after every retraining of the networks following the three different approaches.[]{data-label="fig:iou_plot"}](plots/segmentation_miou.pdf){width="\linewidth"}
The experimental results are shown in figure \[fig:iou\_plot\]. The experiment validates that our system collects useful data faster than the lawnmower based approaches. To reach a good segmentation performance of $\textrm{mIoU}\geq 90\,\%$, only 3 missions were necessary. As a comparison, the prediction maps and generated paths after 3 missions of all methods are shown in figure \[fig:planned\_paths\]. In particular, we note that the novelty plotted over the whole map after 3 missions in \[fig:planned\_paths:c\] highlights the lake on the upper left as a region of high novelty, which is also the only region that is wrongly classified. The corresponding path shows that the was aborted before collecting more lake data due to energy constraints and went back to the starting position.
Discussion
==========
In our experiments with the RIT18 dataset, we found that one of the main differences among the evaluated path planners was the data balance of the different classes captured in each mission, because the classes are geographically very separated on the map. This makes the experiment in general sensitive to the choice of starting positions for each mission. Choosing positions in all 4 quadrants of the map came to the advantage of the smaller lawnmower approach for the particular dataset. It remains to test how our framework performs with a starting position fixed over several missions. To disentangle the evaluation of our system and the novelty detection, we plan experiments on datasets with a different class distribution, as well as experiments where we exchange the novelty estimation to a randomly generated heatmap.
On top of the points above, we restricted the path planning problem by the available information and the possible actions. For other scenarios, the information available can vary, it is also possible to have novelty maps similar to the one on figure \[fig:planned\_paths:c\] from previous missions available. In our experiments the height above ground was kept constant. While informative path planning with variable height was explored in different works before, it remains subject of further research how the height affects segmentation performance and novelty estimation.
Conclusion
==========
In this work we present an system to collect valuable training data with a . We show how to incorporate novelty estimation from deep learning into a path planning objective and evaluate our system on a real world terrain monitoring map. The results indicate a significantly faster useful data acquisition with improved performance compared to traditional lawnmower approaches.
Acknowledgment {#acknowledgment .unnumbered}
==============
We thank Cesar Cadena and Juan Nieto for their valuable inputs.
[^1]: This work was supported through funding by Hilti Group.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present a determination of the CKM parameter $|V_{us}|$ based on new measurements of the six largest $K_L$ branching fractions and semileptonic form factors by the KTeV (E832) experiment at Fermilab. We find $|{\mbox{$V_{us}$}}| = \vusave \pm \vusavein_{\rm KTeV} \pm \vusaveex_{\rm ext}$, where the errors are from KTeV measurements and from external sources. We also use the measured branching fractions to determine the CP violation parameter $|{\mbox{$\eta_{+-}$}}| = (\etapmv \pm \etapmint_{\rm KTeV}
\pm \etapmext_{\rm ext})\times 10^{-3} $.
author:
- 'T. Alexopoulos'
- 'M. Arenton'
- 'R.F. Barbosa'
- 'A.R. Barker'
- 'L. Bellantoni'
- 'A. Bellavance'
- 'E. Blucher'
- 'G.J. Bock'
- 'E. Cheu'
- 'S. Childress'
- 'R. Coleman'
- 'M.D. Corcoran'
- 'B. Cox'
- 'A.R. Erwin'
- 'R. Ford'
- 'A. Glazov'
- 'A. Golossanov'
- 'J. Graham'
- 'J. Hamm'
- 'K. Hanagaki'
- 'Y.B. Hsiung'
- 'H. Huang'
- 'V. Jejer'
- 'D.A. Jensen'
- 'R. Kessler'
- 'H.G.E. Kobrak'
- 'K. Kotera'
- 'J. LaDue'
- 'A. Ledovskoy'
- 'P.L. McBride'
- 'E. Monnier'
- 'H. Nguyen'
- 'R. Niclasen'
- 'V. Prasad'
- 'X.R. Qi'
- 'E.J. Ramberg'
- 'R.E. Ray'
- 'M. Ronquest'
- 'E. Santos'
- 'P. Shanahan'
- 'J. Shields'
- 'W. Slater'
- 'D. Smith'
- 'N. Solomey'
- 'E.C. Swallow'
- 'P.A. Toale'
- 'R. Tschirhart'
- 'Y.W. Wah'
- 'J. Wang'
- 'H.B. White'
- 'J. Whitmore'
- 'M. Wilking'
- 'B. Winstein'
- 'R. Winston'
- 'E.T. Worcester'
- 'T. Yamanaka'
- 'E. D. Zimmerman'
title: 'A Determination of the CKM Parameter $|{\mbox{$V_{us}$}}|$ '
---
The Cabibbo-Kobayashi-Maskawa (CKM) matrix [@cabibbo; @km] describes the charged current couplings of the $u$, $c$, and $t$ quarks to the $d$, $s$, and $b$ quarks. The first row of this matrix provides the most stringent test of the unitarity of the matrix. Current measurements [@pdg02] deviate from unitarity at the 2.2 sigma level: $1-(|V_{ud}|^2+|V_{us}|^2+|V_{ub}|^2)= 0.0043 \pm 0.0019$. $|V_{us}|$, which contributes an uncertainty of 0.0010 to this unitarity test, has been determined from charged and neutral kaon semileptonic decay rates. This determination is based on the partial width for semileptonic $K$ decay, $\Gamma_{K\ell 3}$: $$\Gamma_{K\ell 3}=\frac{\textstyle G_F^2
M_K^5}{\textstyle 192\pi^3} S_{EW} (1+\delta^\ell_K)C^2
\left|{\mbox{$V_{us}$}}\right|^2f^2_+(0) I^\ell_K,
\label{eq:vus}$$ where $\ell$ refers to either $e$ or $\mu$, $G_F $ is the Fermi constant, $M_K$ is the kaon mass, $S_{EW}$ is the short-distance radiative correction, $\delta^\ell_K$ is the mode-dependent long-distance radiative correction, $f_+(0)$ is the calculated form factor at zero momentum transfer for the $\ell \nu$ system, and $I^\ell_K$ is the phase-space integral, which depends on measured semileptonic form factors. $C^2$ is 1 (1/2) for neutral (charged) kaon decays. The current PDG determination of $|{\mbox{$V_{us}$}}|$ is based only on $K \to \pi e \nu$ decays; $K \to \pi \mu \nu$ decays have not been used because of large uncertainties in $I^{\mu}_K$.
In this Letter, we present a determination of $|V_{us}|$ by the KTeV (E832) experiment at Fermilab based on measurements of the ${\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}}$ and ${\mbox{$K_{L}\to\pi^{\pm}{\mu}^{\mp}\nu$}}$ partial widths and form factors. These measurements are described in detail elsewhere [@ktev_kbr; @ktev_kl3ff]; a brief summary is given here. Our $|V_{us}|$ determination also makes use of a new treatment of radiative corrections [@Troy].
To determine the ${\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}}$ and partial widths, we measure the following five ratios: $$\begin{aligned}
{\mbox{$\Gamma_{K\mu3}$}}/{\mbox{$\Gamma_{Ke3}$}}& \equiv & {\mbox{$\Gamma({\mbox{$K_{L}\to\pi^{\pm}{\mu}^{\mp}\nu$}})$}}/ {\mbox{$\Gamma({\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}})$}}\ \\
{\mbox{$\Gamma_{+-0}$}}/{\mbox{$\Gamma_{Ke3}$}}& \equiv & {\mbox{$\Gamma(K_{L}\rightarrow \pi^+\pi^-\pi^0)$}}/ {\mbox{$\Gamma({\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}})$}}\ \ \ \\
{\mbox{$\Gamma_{000}$}}/{\mbox{$\Gamma_{Ke3}$}}& \equiv & {\mbox{$\Gamma(K_{L}\rightarrow \pi^0\pi^0\pi^0)$}}/ {\mbox{$\Gamma({\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}})$}}\ \\
{\mbox{$\Gamma_{+-}$}}/{\mbox{$\Gamma_{Ke3}$}}& \equiv & {\mbox{$\Gamma(K_{L}\rightarrow \pi^+\pi^-)$}}/ {\mbox{$\Gamma({\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}})$}}\ \\
{\mbox{$\Gamma_{00}$}}/{\mbox{$\Gamma_{000}$}}& \equiv & {\mbox{$\Gamma(K_{L}\rightarrow \pi^0\pi^0)$}}/ {\mbox{$\Gamma(K_{L}\rightarrow \pi^0\pi^0\pi^0)$}}, \end{aligned}$$ where internal bremsstrahlung contributions are included for all decay modes with charged particles. Since the six decay modes listed above account for more than 99.9% of the total decay rate, the five partial width ratios may be converted into measurements of the branching fractions for the six decay modes. The $K_L$ lifetime is then used to convert these branching fractions into partial widths. The branching fraction measurements also can be used to determine the CP violation parameter $|{\mbox{$\eta_{+-}$}}|^2 \equiv \Gamma({\mbox{$K_{L}$}}\to \pi^+\pi^-)/
\Gamma({\mbox{$K_{S}$}}\to \pi^+\pi^-)$.
The semileptonic form factors describe the distribution of $t$, the momentum transfer to the $\ell\nu$ system. This $t$ dependence increases the decay phase space integrals, $I^{e}_K$ and $I^{\mu}_K$, by about 10%. We use the following parametrization for the two independent semileptonic form factors: $$\begin{array}{l}
{\mbox{${f}_+(t)$}}= f_+(0)\left(1 + {\mbox{$\lambda'_+$}}\frac{\textstyle t}{\textstyle M^2_\pi} +
\frac{\textstyle 1}{\textstyle 2} {\mbox{$\lambda''_+$}}\frac{\textstyle t^2}{\textstyle M^4_\pi}\right) \\
{\mbox{${f}_0(t)$}}= f_+(0)\left(1 + {\mbox{$\lambda_0$}}\frac{\textstyle t}{\textstyle
M^2_\pi}\right), \\
\end{array}$$ where $f_+(0)$ is obtained from theory, and we measure , , and .
In principle, the form factors can be measured directly from the $t$ distribution. The undetected neutrino and unknown kaon momentum, however, result in a twofold ambiguity in the reconstructed value of $t$. To avoid systematic uncertainties associated with this ambiguity, we use a technique based only on components of particle momenta measured transverse to the kaon momentum [@ktev_kl3ff].
The KTeV experiment (Fig. \[fig:detector\]) and associated event reconstruction techniques have been described in detail elsewhere [@Alavi-Harati:2002ye]. An 800 GeV/c proton beam striking a BeO target is used to produce two almost parallel neutral beams. The regenerator beam, which includes $K_S$, is not used in this analysis; the vacuum beam provides $K_L$ decays used for these measurements. A large vacuum decay region surrounded by photon veto detectors extends to 159 m from the primary target. Following the vacuum region is a drift chamber spectrometer, trigger hodoscope, pure CsI electromagnetic calorimeter, and a muon system consisting of scintillator hodoscopes behind 4 m and 5 m of steel. The analyses presented in this Letter make use of the detector calibration and Monte Carlo simulation from the KTeV $\epe$ analysis [@Alavi-Harati:2002ye].
Simple event reconstruction and selection may be used to distinguish different kaon decay modes from each other, and to reduce background to a negligible level for all decay modes. The reconstruction of charged decay modes (, , , and ) begins with the identification of two oppositely charged tracks coming from a single vertex. Note that for the ${\mbox{$K_{L}\rightarrow \pi^{+}\pi^{-}\pi^{0}$}}$ decay, we choose not to reconstruct the $\pi^0 \to \gamma\gamma$ decay to reduce the acceptance uncertainty in the ${\mbox{$\Gamma_{+-0}$}}/ {\mbox{$\Gamma_{Ke3}$}}$ ratio.
The charged decay modes are separated from each other on the basis of particle identification and kinematic requirements. To select ${\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}}$ decays, the electron is identified using the calorimeter energy measurement ($E$), combined with the spectrometer momentum ($p$). is separated from other two-track decays based on the two-track invariant mass, $m_{\pi\pi}$, and the square of the two-track momentum transverse to the $K_L$ direction, $p^2_t$. To isolate we use an additional variable, , described in [@ktev_kbr]. Note that the background to each of these decay modes is suppressed by rejecting events with hits in the muon system. Two-track events that are not identified as , , or , are selected as candidates. To reduce the acceptance uncertainty in the ${\mbox{$\Gamma_{K\mu3}$}}/{\mbox{$\Gamma_{Ke3}$}}$ ratio, we do not require a signal in the muon hodoscope to identify decays.
The reconstruction of the ${\mbox{$K_{L}\rightarrow\pi^{0}\pi^{0}$}}$ and ${\mbox{$K_{L}\rightarrow \pi^{0}\pi^{0}\pi^{0}$}}$ decay modes, where $\pi^0 \to \gamma\gamma$, is based on energies and positions of photons measured in the CsI electromagnetic calorimeter as described in [@Alavi-Harati:2002ye]. Exactly four (six) clusters, each with a transverse profile consistent with a photon, are required for ${\mbox{$K_{L}\rightarrow\pi^{0}\pi^{0}$}}$ (${\mbox{$K_{L}\rightarrow \pi^{0}\pi^{0}\pi^{0}$}}$). Photons are paired to reconstruct two or three neutral pions consistent with a single decay vertex.
All reconstructed decay modes are required to have kaon energy, $E_K$, between 40 and 120 GeV, and decay position, $z$, between 123 and 158 m from the target. For the reconstruction of semileptonic and decays, there is a missing particle ($\nu$ or $\pi^0$), which results in multiple kaon energy solutions. Each of these solutions is required to be in the accepted range.
After all event selection requirements and background subtraction, we have between $10^5$ and $10^6$ events per decay mode. The background is 0.7% for and much smaller for the other decay modes. After correcting each of the ratios for acceptance differences between numerator and denominator, we find the partial width ratios given in Table \[tab:ratios\]. The precision is 1.2% for ${\mbox{$\Gamma_{000}$}}/ {\mbox{$\Gamma_{Ke3}$}}$ and about 0.5% for other ratios. A comparison of data and MC $z$-vertex distributions for the semileptonic decay modes (Fig. \[fig:zvtx\_slopes\]) demonstrates the quality of the MC simulation used for the acceptance correction.
Decay Modes Partial Width Ratio
------------------------------------------------------ ---------------------------------------------------------------------
${\mbox{$\Gamma_{K\mu3}$}}/ {\mbox{$\Gamma_{Ke3}$}}$ $\RPMNvalue \pm \RPMNerrstat \pm \RPMNerrsyst $
${\mbox{$\Gamma_{000}$}}/ {\mbox{$\Gamma_{Ke3}$}}$ $\RZZZvalue \pm \RZZZerrstat \pm \RZZZerrsyst $
${\mbox{$\Gamma_{+-0}$}}/ {\mbox{$\Gamma_{Ke3}$}}$ $\RPPPvalue \pm \RPPPerrstat \pm \RPPPerrsyst $
${\mbox{$\Gamma_{+-}$}}/ {\mbox{$\Gamma_{Ke3}$}}$ $(\RPPvalue \pm \RPPerrstat \pm \RPPerrsyst) \times 10^{-3} $
${\mbox{$\Gamma_{00}$}}/ {\mbox{$\Gamma_{000}$}}$ $(\RNEUTvalue \pm \RNEUTerrstat \pm \RNEUTerrsyst ) \times 10^{-3}$
: \[tab:ratios\]Measured partial width ratios. The first error is statistical and the second systematic. The five statistical errors are independent; correlations among the systematic errors are treated in [@ktev_kbr].
The five partial width ratios may be combined to determine the branching fractions shown in Table \[tab:br\] [@ktev_kbr]. Using the PDG average for the neutral kaon lifetime [@pdg02], $\tau_L = (5.15 \pm 0.04) \times 10^{-8}$ s, our branching fraction measurements correspond to the partial decay widths shown in Table \[tab:br\].
Figure \[fi:pdg2\] shows a comparison of the KTeV and PDG values for the six branching fractions. The new KTeV measurements are on average a factor of two more precise than the current world average values, but are not in good agreement with these averages. Compared to the PDG fit [@pdg02], the KTeV measurement of $B({\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}})$ is higher by 5%, $B({\mbox{$K_{L}\rightarrow \pi^{0}\pi^{0}\pi^{0}$}})$ is lower by 8%, $B({\mbox{$K_{L}\rightarrow\pi^{+}\pi^{-}$}})$ is lower by 5%, and $B({\mbox{$K_{L}\rightarrow\pi^{0}\pi^{0}$}})$ is lower by 8%. Our measurements of $B({\mbox{$K_{L}\to\pi^{\pm}{\mu}^{\mp}\nu$}})$ and $B({\mbox{$K_{L}\rightarrow \pi^{+}\pi^{-}\pi^{0}$}})$ are consistent with the PDG fit. A detailed comparison between the KTeV measurements and previous results is given in [@ktev_kbr].
Using the measured branching fractions for ${\mbox{$K_{L}\rightarrow\pi^{+}\pi^{-}$}}$ and ${\mbox{$K_{L}\rightarrow\pi^{0}\pi^{0}$}}$ together with $\tau_L$, $\tau_S$, $Re(\epe)$, and $B(K_S\to\pi\ell\nu)$, we determine the CP violation parameter $|\eta_{+-}| = (\etapmv \pm
\etapmint_{\rm KTeV} \pm \etapmext_{\rm ext})\times 10^{-3}$; most of the external error results from the uncertainty in $\tau_L$. Our result is 2.6% lower than the PDG average. A comparison of $|{\mbox{$\eta_{+-}$}}|$ determinations is given in [@ktev_kbr].
Decay Mode Branching fraction $\Gamma_i$ ($10^7 s^{-1}$)
------------ ---------------------------------------------- ----------------------------------------------
$\BKEvalue \pm \BKEerrtot $ $\WKEvalue \pm \WKEerrtot $
$\BKMvalue \pm \BKMerrtot $ $\WKMvalue \pm \WKMerrtot $
$\BPMZvalue \pm \BPMZerrtot $ $\WPMZvalue \pm \WPMZerrtot $
$\BZZZvalue \pm \BZZZerrtot $ $\WZZZvalue \pm \WZZZerrtot $
$(\BPMvalue \pm \BPMerrtot)\times 10^{-3} $ $(\WPMvalue \pm \WPMerrtot)\times 10^{-3} $
$(\BZZvalue \pm \BZZerrtot)\times 10^{-3} $ $(\WZZvalue \pm \WZZerrtot)\times 10^{-3} $
: \[tab:br\] $K_L$ branching fractions and partial widths ($\Gamma_i$). Correlations among uncertainties in these measurements are given in [@ktev_kbr].
The form factor is measured in both semileptonic decay modes; the effect of is proportional to the lepton mass, so it is only measured in ${\mbox{$K_{L}\to\pi^{\pm}{\mu}^{\mp}\nu$}}$ decays. The measured parameters for the semileptonic form factors are ${\mbox{$\lambda'_+$}}= (\lplabv \pm \lplabe)\times 10^{-3}$, ${\mbox{$\lambda''_+$}}= (\lplbbv \pm \lplbbe)\times 10^{-3}$, and ${\mbox{$\lambda_0$}}= (\lzbv \pm \lzbe )\times 10^{-3}$. The corresponding phase space integrals are $\textstyle I_K^{e} = \ikebv \pm \ikebetot $ and $\textstyle I_K^{\mu} = \ikmbv \pm \ikmbetot$, where the quoted errors include an additional uncertainty related to the form factor parameterization [@ktev_kl3ff]. Compared to phase space integrals based on PDG form factors, KTeV’s $I_K^{e}$ and $I_K^{\mu}$ integrals are 1.7% and 4.2% lower, respectively. If we fit our data without the ${\mbox{$\lambda''_+$}}$ term, our $I_K^{e}$ and $I_K^{\mu}$ integrals are increased by 1%, and are consistent with PDG averages that use only linear terms.
To check the consistency of our branching fraction and form factor measurements with lepton universality, we compare $G_F$ for the two decay modes by taking the ratio of Eq. \[eq:vus\] for ${\mbox{$K_{L}\to\pi^{\pm}{\mu}^{\mp}\nu$}}$ and ${\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}}$: $$\Big(\frac{G^{\mu}_F}{G^e_F}\Big)^2 =
\Big[\frac{\textstyle {\mbox{$\Gamma({\mbox{$K_{L}\to\pi^{\pm}{\mu}^{\mp}\nu$}})$}}}{\textstyle {\mbox{$\Gamma({\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}})$}}} \Big]
\Big/
\Big(
\frac{\textstyle 1+ {\mbox{$\delta_K^\mu$}}}{\textstyle 1 +{\mbox{$\delta_K^e$}}}
\cdot
\frac{\textstyle I_K^\mu}{\textstyle I_K^e}
\Big)
\label{eq:universal}.$$ Many common uncertainties cancel in this ratio. The ratio of radiative corrections is calculated to be $(1+\delta_K^{\mu})/(1+\delta_K^e) =
\DeltaRat \pm \DeltaRatE$ [@Troy], the ratio of the phase space integrals is $I_K^{\mu}/I_K^e =\ikrbv \pm \ikrbetot$, and ${\mbox{$\Gamma_{K\mu3}$}}/{\mbox{$\Gamma_{Ke3}$}}$ is from Table \[tab:br\]. The resulting ratio of couplings squared is $(G^{\mu}_F / G^e_F)^2= \LeptUni \pm \LeptUniE$, consistent with lepton universality. The same ratio calculated from PDG widths and form factors is $(G^{\mu}_F / G^e_F)^2= \PDGLeptUni \pm \PDGLeptUniE$. Note that the 0.5% uncertainty in our universality test is much smaller than the 5% difference between the KTeV and PDG values of ${\mbox{$\Gamma_{K\mu3}$}}/ {\mbox{$\Gamma_{Ke3}$}}$.
The measured partial widths and phase space integrals for semileptonic decays can be combined with theoretical corrections to calculate $|V_{us}|$ using Eq. \[eq:vus\]. The short-distance radiative correction, $S_{EW} = 1.022$ [@sew], is evaluated with a cutoff at the proton mass. The long-distance radiative corrections are taken from [@Troy]: $\delta^e_K = 0.013\pm 0.003$ and $\delta^\mu_K = 0.019\pm 0.003$. For $f_+(0)$, we use the same value used in the PDG evaluation of $|V_{us}|$: $f_+(0) = 0.961 \pm 0.008$ [@leut-roos].
The resulting values of $|{\mbox{$V_{us}$}}|$ are $\vuske \pm \vuskee$ for $K_{e3}$ and $\vuskm \pm \vuskme$ for $K_{\mu 3}$, where the errors include an external uncertainty of from $f_+(0)$, the $K_L$ lifetime, and radiative corrections. Assuming lepton universality, we average the ${\mbox{$K_{L}\to\pi^{\pm}e^{\mp}\nu$}}$ and ${\mbox{$K_{L}\to\pi^{\pm}{\mu}^{\mp}\nu$}}$ results (accounting for correlations): $$|{\mbox{$V_{us}$}}| = \vusave \pm \vusavein_{\rm KTeV} \pm \vusaveex_{\rm ext}.
\label{eq:vusresult}$$ The KTeV error comes from uncertainties in the KTeV branching fraction and form factor measurements.
To compare our result with previous charged and neutral kaon measurements, we use the product of $|{\mbox{$V_{us}$}}|$ and $f_+(0)$ rather than $|{\mbox{$V_{us}$}}|$ to avoid significant common uncertainties from $f_+(0)$. Figure \[fig:vus\] shows a comparison of our measurement of $$|{\mbox{$V_{us}$}}|f_+(0) = \vusfave \pm \vusfavee$$ with values from the PDG and Brookhaven E865 [@bnl865]. Our value of $|V_{us}|f_+(0)$ is inconsistent with previous $K_L$ determinations, but is consistent with $K^+$ results (both E865 and earlier measurements). The figure also shows $f_+(0)(1-|V_{ud}|^2-|V_{ub}|^2)^{1/2}$, the expectation for $f_+(0)|V_{us}|$ assuming unitarity, based on $|V_{ud}| = 0.9734 \pm 0.0008$, $|V_{ub}| = (3.6 \pm 0.7)\times 10^{-3}$, and several recent calculations of $f_+(0)$. Our value of $|V_{us}|$ (Eq. \[eq:vusresult\]), based the Leutwyler and Roos calculation of $f_+(0)$, is consistent with unitarity: $1-(|V_{ud}|^2+|V_{us}|^2+|V_{ub}|^2)= \deltauni \pm \unite $. For other calculations of $f_+(0)$, the consistency with unitarity ranges from 1 to 1.7 sigma, as shown in Fig. \[fig:vus\]. Our improved form factor measurements may help to reduce theoretical uncertainties in $f_+(0)$ [@bij; @becirevic; @jamin].
In summary, KTeV has made improved measurements of the six largest $K_L$ branching fractions and the semileptonic form factors. We use these results to determine $|{\mbox{$\eta_{+-}$}}| = (\etapmv \pm \etapme) \times 10^{-3}$ and $|{\mbox{$V_{us}$}}|f_+(0) = \vusfave \pm \vusfavee$. Using $f_+(0) = 0.961 \pm 0.008$ [@leut-roos], we find $|{\mbox{$V_{us}$}}| = \vusave \pm \vusavee$, consistent with unitarity of the CKM matrix.
We gratefully acknowledge the support and effort of the Fermilab staff and the technical staffs of the participating institutions for their vital contributions. We also thank T. Andre and J. Rosner for useful discussions of radiative corrections. This work was supported in part by the U.S. Department of Energy, The National Science Foundation, and The Ministry of Education and Science of Japan.
[14]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, ****, ().
(), (), , .
() (), , .
(), , .
(), ****, (), .
, ****, ().
, ****, ().
(), ****, ().
, ****, ().
, **** ().
, , , ****, (), .
, **** ().
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Department of Physics, Sun Yat-Sen University, Guangzhou 510275,\
People’s Republic of China\
and\
China Center of Advanced Science and Technology (World Laboratory),\
P.O.Box 8730, Beijing 100080, People’s Republic of China [^1]
author:
- 'Qiong-Gui Lin[^2]'
title: 'Anisotropic harmonic oscillator in a static electromagnetic field[^3]'
---
0 cm -0.5 cm
0.4cm
PACS numbers: 03.65.Ge, 42.50.Dv\
Keywords: anisotropic harmonic oscillator; electromagnetic field; exact solutions; coherent states; squeezed states
Introduction
============
The anisotropic harmonic oscillator potential is of physical interest in quantum mechanics, since it may describe the motion of, say, an electron in an anisotropic metal lattice. With external electromagnetic fields imposed on, such models may also be useful in semiconductor physics [@01; @02]. For an isotropic harmonic oscillator potential and a homogeneous static magnetic field the Schrödinger equation can be easily solved in the cylindrical coordinates. If the harmonic oscillator potential is anisotropic, the problem is not so easy. Nevertheless, it can be solved analytically because the Hamiltonian is quadratic in the canonical coordinates and momenta. There exists quite general formalisms for solving systems with such quadratic Hamiltonians [@1], and the above problem represents a typical example [@2]. In Ref. [@2] only the isotropic case is considered as an example for illustrating the general method. The anisotropic case with a homogeneous static magnetic field was studied in Refs. [@3; @4]. In these works the magnetic field was arranged to point in one of the rectangular axis directions. In this paper we will solve the problem in a somewhat easier formalism. A homogeneous static electric field is also included, though its effect is trivial. We will also deal with a case where the magnetic field has a more general direction. This seems not having been considered before. Coherent and squeezed states of these systems will be discussed in some detail. In terms of these concepts the time evolution of some wave packets can be discussed very conveniently.
In the next section we consider a charged harmonic oscillator with arbitrary frequencies $\omega_x$, $\omega_y$ and $\omega_z$ in three rectangular axis directions. A homogeneous static magnetic field in the $z$ direction and a homogeneous static electric field in an arbitrary direction are imposed to the system. The problem can be easily reduced to one equivalent to the case without an electric field by making some coordinate transformation. In Sec. III we diagonalize the reduced Hamiltonian and give the energy eigenvalues. Although this has been previously done, the method used here is different and seems more convenient, especially for the subsequent discussions of coherent and squeezed states. The wave functions are explicitly worked out in Sec. IV. In Sec. V we discuss some special cases, some of which may need different handling, or even cannot be solved. In Sec. VI we consider a somewhat different case where $\omega_z=0$ and $E_z=0$ ($E_z$ is the $z$ component of the electric field), but the magnetic field has a $x$ or $y$ component, in addition to the $z$ component. To our knowledge, this was not considered before. It is solved analytically by the formalism of Sec. III. In Sec. VII we discuss the coherent and squeezed states and their time evolution in detail. A brief summary is given in Sec. VIII.
Reduction of the Hamiltonian
============================
Consider a charged particle with charge $q$ and mass $M$, moving in an anisotropic harmonic oscillator potential, a homogeneous static magnetic field ${\mathbf B}=B{\mathbf e}_z$, and a homogeneous static electric field ${\mathbf E}=E_x {\mathbf e}_x+E_y {\mathbf e}_y +E_z
{\mathbf e}_z$, where ${\mathbf e}_x$, ${\mathbf e}_y$ and ${\mathbf
e}_z$ are unit vectors of the rectangular coordinate system, and $B$, $E_x$, $E_y$ and $E_z$ are all constants. The stationary Schrödinger equation is $$\label{1}
H^{\rm T}\Psi({\mathbf x})={\cal E}^{\rm T}\Psi({\mathbf x}),$$ where the Hamiltonian is $$\label{2}
H^{\rm T}={1 \over 2M}\left({\mathbf p}-{q\over c}{\mathbf
A}\right)^2 +{1\over 2}M(\omega_x^2 x^2 +\omega_y^2 y^2 +\omega_z^2
z^2) -q{\mathbf E}\cdot{\mathbf x},$$ where ${\mathbf E}\cdot{\mathbf x}=E_x x+E_y y +E_z z$. Since the magnetic field points in the $z$ direction, one may take $A_x$ and $A_y$ independent of $z$, and $A_z=0$. Then the Hamiltonian can be decomposed as $H^{\rm T}=H^{xy}+H^z$, where $$\label{3}
H^{xy}={1 \over 2M}\left[\left(p_x-{q\over c}A_x\right)^2
+\left(p_y-{q\over c}A_y\right)^2\right] +{1\over 2}M(\omega_x^2 x^2
+\omega_y^2 y^2)-q(E_x x+E_y y),$$ and $$\label{4}
H^z={1 \over 2M}p_z^2+{1\over 2}M\omega_z^2 z^2-qE_z z.$$ Therefore the total wave function can be factorized as $\Psi({\mathbf
x}) =\psi(x,y) Z(z)$, where $\psi(x,y)$ and $Z(z)$ satisfy the following equations: $$\label{5}
H^{xy}\psi(x,y)={\cal E}^{xy}\psi(x,y),$$ $$\label{6}
H^z Z(z)={\cal E}^z Z(z),$$ with ${\cal E}^{xy}+{\cal E}^z={\cal E}^{\rm T}$. Obviously, $H_z$ can be recast in the form $$\label{7}
H^z={1 \over 2M}p_z^2+{1\over 2}M\omega_z^2 (z-z_0)^2-{q^2 E_z^2
\over 2M\omega_z^2}$$ where $z_0=qE_z/M\omega_z^2$, so that Eq. (\[6\]) can be easily solved with the following results. $$\label{8}
{\cal E}_{n_3}^z=\left(n_3+\frac 12\right)\hbar\omega_z-{q^2 E_z^2
\over 2M\omega_z^2}, \quad n_3=0,1,2,\ldots,$$ $$\label{9}
Z_{n_3}(z)=\psi_{n_3}^{\omega_z}(z-z_0),\quad n_3=0,1,2,\ldots,$$ where we have quoted the standard wave functions for harmonic oscillators: $$\label{10}
\psi_n^\omega(x)=\left({\sqrt{M\omega}\over 2^n n!\sqrt{\pi\hbar}}
\right)^{1/2}\exp\left(-\frac {M\omega}{2\hbar} x^2\right)
H_n\left(\sqrt{M\omega\over\hbar} x\right),\quad n=0,1,2,\ldots.$$ Therefore our main task is to solve Eq. (\[5\]). We take the gauge $$\label{11}
A_x=-{\textstyle\frac 12}B\tilde y,\quad A_y={\textstyle\frac
12}B\tilde x,$$ where $\tilde x=x-x_0$, $\tilde y=y-y_0$, and $x_0=qE_x/M\omega_x^2$, $y_0=qE_y/M\omega_y^2$, then $H^{xy}$ can be written as $$\label{12}
H^{xy}={1 \over 2M}(p_x^2 +p_y^2) +{1\over 2}M\omega_1^2 \tilde x^2
+{1\over 2}M\omega_2^2 \tilde y^2-\omega_B\tilde L_z-{q^2 E_x^2 \over
2M\omega_x^2} -{q^2 E_y^2 \over 2M\omega_y^2},$$ where $\tilde L_z=\tilde xp_y-\tilde yp_x$, $\omega_B=qB/2Mc$, and $$\label{13}
\omega_1=\sqrt{\omega_x^2+\omega_B^2}, \quad
\omega_2=\sqrt{\omega_y^2+\omega_B^2}.$$ Note that $\omega_B$ may be either positive or negative, depending on the signs of $q$ and $B$. Now Eq. (\[5\]) can be written as
\[14\] $$\label{14a}
H\psi(x,y)={\cal E}\psi(x,y),$$ where $$\label{14b}
H={1 \over 2M}(p_x^2 +p_y^2) +{1\over 2}M\omega_1^2 \tilde x^2
+{1\over 2}M\omega_2^2 \tilde y^2-\omega_B\tilde L_z,$$
$$\label{15}
{\cal E}={\cal E}^{xy}+{q^2 E_x^2 \over 2M\omega_x^2}+{q^2 E_y^2
\over 2M\omega_y^2}.$$
Eq. (\[14\]) has the same form as that for an anisotropic harmonic oscillator moving on the $xy$ plane under the influence of a homogeneous static magnetic field perpendicular to the plane, except that $x$, $y$ are replaced by $\tilde x$, $\tilde y$. The main feature is that the reduced Hamiltonian is quadratic in the canonical variables. Though this equation has been studied by some authors, we will solve it in a somewhat different and simpler way in the next two sections. Here we point out that the calculations below will be simpler if one take the gauge $A_x=-B\tilde y$, $A_y=0$ or $A_x=0$, $A_y=B\tilde x$. However, we prefer the above gauge (\[11\]), since in this gauge it would be convenient to compare the results with known ones when $\omega_x=\omega_y$, or $\omega_1=\omega_2$.
Diagonalization of the reduced Hamiltonian
==========================================
If $\omega_x=\omega_y$, Eq. (\[14\]) can be easily solved in the cylindrical coordinates. In the general case this does not work, so other methods are necessary. We define a column vector $$\label{16}
X=(\tilde x, p_x, \tilde y, p_y)^\tau,$$ where $\tau$ denotes matrix transposition, then the reduced Hamiltonian (\[14b\]) can be written as $$\label{17}
H={\textstyle\frac 12}X^\tau{\cal H}X ={\textstyle\frac 12}
X^\dagger{\cal H}X,$$ where ${\cal H}$ is a c-number matrix which we take to be symmetric. We do not write down ${\cal H}$ here since this is easy. If one can diagonalize ${\cal H}$ then the Hamiltonian would become a sum of two one-dimensional Hamiltonians for harmonic oscillators or something similar (repulsive harmonic oscillators or free particles). However, a crucial point here is to preserve the commutation relation $$\label{18}
[X_\alpha, X_\beta]=-\hbar(\Sigma_y)_{\alpha\beta}, \quad \alpha,
\beta=1,2,3,4$$ after the linear transformation that diagonalizes ${\cal H}$, where $\Sigma_y={\rm diag}(\sigma_y,\sigma_y)$, and $\sigma_y$ is the second Pauli matrix. Therefore the needed transformations are the so called symplectic ones [@1; @2]. Here we will transform the Hamiltonian into a form expressed in terms of raising and lowering operators, so the formalism is somewhat different. Let us proceed as follows.
Because $H$ is quadratic in the canonical variables $X$, it is easy to show that $$\label{19}
[{\mathrm i}H,X]=\hbar\Omega X,$$ where $$\label{20}
\Omega={\mathrm i}\Sigma_y{\cal H}$$ is another c-number matrix. As both ${\cal H}$ and $i\Sigma_y$ are real, so is $\Omega$. Our first step is to diagonalize $\Omega$, so we write down it here. $$\label{21}
\Omega=\left(
\begin{array}{cccc}
0 & 1/M & \omega_B & 0 \\
-M\omega_1^2 & 0 & 0 & \omega_B \\
-\omega_B & 0 & 0 & 1/M \\
0 & -\omega_B & -M\omega_2^2 & 0
\end{array}
\right).$$ The characteristic polynomial for this matrix is
\[22\] $$\label{22a}
\det(\lambda I-\Omega)=\lambda^4+b\lambda^2+c,$$ where $$\label{22b}
b=\omega_x^2+\omega_y^2+4\omega_B^2,\quad c=\omega_x^2\omega_y^2.$$
Since $$\label{23}
\Delta\equiv
b^2-4c=(\omega_x^2-\omega_y^2)^2+8\omega_B^2(\omega_x^2+\omega_y^2
+2\omega_B^2)\ge 0,$$ we have two real roots for $\lambda^2$. If $\omega_x=0$ or $\omega_y=0$, one of the two roots is zero, and the following discussions are not valid. We will return to this case latter. Currently we assume that both $\omega_x$ and $\omega_y$ are nonzero, then $c>0$ and $\sqrt\Delta<b$, and both roots for $\lambda^2$ are negative. Therefore the above characteristic polynomial has four pure imaginary roots: $$\label{24}
\{\lambda_1,\lambda_2,\lambda_3,\lambda_4\}=\{-{\mathrm i}\sigma_1,
{\mathrm i}\sigma_1, -{\mathrm i}\sigma_2, {\mathrm i}\sigma_2\},$$ where $$\label{25}
\sigma_1=\left(b+\sqrt\Delta\over 2\right)^{1/2},\quad
\sigma_2=\left(b-\sqrt\Delta\over 2\right)^{1/2},$$ and $\sigma_1\ge\sigma_2>0$. The equal sign appears when $\omega_B=0$ and $\omega_x=\omega_y$.
Because $\Omega$ is not symmetric, right eigenvectors and left ones are different. We define two left eigenvectors (row vectors) $u_i$ corresponding to the eigenvalues $-{\mathrm i}\sigma_i$ ($i=1,2$) by $$\label{26}
u_i\Omega=-{\mathrm i}\sigma_i u_i,\quad i=1,2,$$ then the other two are $u_i^*$, corresponding to the eigenvalues ${\mathrm i}\sigma_i$ ($i=1,2$). Similarly, the right eigenvectors (column vectors) are $v_i$ and $v_i^*$ ($i=1,2$), satisfying the equations $$\label{27}
\Omega v_i=-{\mathrm i}\sigma_i v_i,\quad i=1,2$$ and the complex conjugate ones. From these eigenvalue equations it can be shown that
\[28\] $$\label{28a}
u_i^* v_j=u_i v_j^*=0, \quad \forall\; i,j=1,2,$$ and $u_i v_j=u_i^* v_j^*=0$ for $\sigma_i\ne\sigma_j$ (this should be checked individually when $\sigma_1=\sigma_2$, see below). By appropriately choosing the normalization constants we have $$\label{28b}
u_i v_j=u_i^* v_j^*=\delta_{ij}, \quad \forall\; i,j=1,2.$$
The specific expression for the above eigenvectors are not necessary in this section. However, a relation between the left and right eigenvectors is very useful in the following. Using Eqs. (\[20\]), (\[26\]) and (\[27\]), it is easy to show that $v_i\propto
\Sigma_y u_i^\dagger$, so we choose $$\label{29}
v_i=-\Sigma_y u_i^\dagger, \quad i=1,2.$$ Note that the condition (\[28b\]) only determine the product of the normalization constants in $u_i$ and $v_i$. With this relation both constants are fixed (up to a phase factor). Also note that Eqs. (\[28b\]) and (\[29\]) lead to $-u_i\Sigma_y u_i^\dagger=1$ (no summation over $i$ on the left-hand side). The left-hand side of this equation is real (which can be easily shown), but not necessarily positive. Actually it can be shown by using the eigenvalue equation that $-u_i\Sigma_y u_i^\dagger=(u_i\Sigma_y){\cal H}
(u_i\Sigma_y)^\dagger/\sigma_i$. Thus the sign of this quantity depends on the matrix ${\cal H}$, and in general Eq. (\[29\]) should be replaced by $v_i=\epsilon_i\Sigma_y u_i^\dagger$ where $\epsilon_i=\pm 1$. For the present case, however, Eq. (\[29\]) is sound.
Now we define a $4\times 4$ matrix $Q$ by arranging the column vectors in the following order: $$\label{30}
Q=(v_1, v_1^*, v_2, v_2^*).$$ Using Eq. (\[28\]) it is easy to show that $$\label{31}
Q^{-1}=(u_1^\tau, u_1^{*\tau}, u_2^\tau, u_2^{*\tau})^\tau.$$ With the help of the eigenvalue equations (\[26\]) and (\[27\]), we have $$\label{32}
Q^{-1}\Omega Q={\rm diag}(-{\mathrm i}\sigma_1, {\mathrm i}\sigma_1,
-{\mathrm i}\sigma_2, {\mathrm i}\sigma_2).$$ Therefore the matrix $\Omega$ is diagonalized. Using the relation (\[29\]), it is not difficult to show that $$\label{33}
Q^\dagger=-\Sigma_z Q^{-1} \Sigma_y,$$ where $\Sigma_z={\rm diag}(\sigma_z,\sigma_z)$, and $\sigma_z$ is the third Pauli matrix. This relation is very useful in the following.
Next we will diagonalize the Hamiltonian. With the above results this is easy. We define two operators
\[34\] $$\label{34a}
a_i=u_i X/\sqrt\hbar,\quad i=1,2.$$ Their hermitian conjugates are $$\label{34b}
a_i^\dagger=u_i^* X/\sqrt\hbar,\quad i=1,2.$$
Using Eqs. (\[18\]), (\[28\]) and (\[29\]) it is easy to show that the nonvanishing commutators among these operators are $$\label{35}
[a_i, a_j^\dagger]=\delta_{ij}, \quad i,j=1,2.$$ Therefore they are similar to the raising and lowering operators for harmonic oscillators. We define a column vector $$\label{36}
A=(a_1, a_1^\dagger, a_2, a_2^\dagger)^\tau,$$ then Eq. (\[34\]) can be written in the matrix form $$\label{34'}\eqnum{$34'$}
A=Q^{-1}X/\sqrt\hbar.$$ The inverse is $$\label{37}
X=\sqrt\hbar QA,$$ and thus $$\label{38}
X^\tau=X^\dagger=\sqrt\hbar A^\dagger Q^\dagger.$$ Substituting these into Eq. (\[17\]), and using Eqs. (\[20\]), (\[32\]) and (\[33\]), it is rather easy to show that $$\label{39}
H=\textstyle\frac 12 \hbar A^\dagger\Sigma A,$$ where $\Sigma={\rm diag}(\sigma_1, \sigma_1, \sigma_2, \sigma_2)$. Using the commutation relations (\[35\]) this becomes $$\label{40}
H=\textstyle\hbar\sigma_1(a_1^\dagger a_1 +\frac 12)+
\hbar\sigma_2(a_2^\dagger a_2 +\frac 12).$$ Therefore the Hamiltonian is diagonalized. More specifically, it becomes the sum of two one-dimensional harmonic oscillator Hamiltonians. The energy levels are readily available: $$\label{41}
{\cal E}_{n_1 n_2}=\textstyle\hbar\sigma_1(n_1 +\frac 12)+
\hbar\sigma_2(n_2 +\frac 12),\quad n_1, n_2=0,1,2,\ldots.$$ When $\omega_B=0$, we have $\sigma_1=\max(\omega_x, \omega_y)$ and $\sigma_2=\min(\omega_x, \omega_y)$, and these energy levels reduce to known results, as expected. Substituting these into Eq. (\[15\]) we obtain the energy levels for the motion on the $xy$ plane: $$\label{42}
{\cal E}^{xy}_{n_1 n_2}=\hbar\sigma_1\left(n_1 +\frac 12\right)+
\hbar\sigma_2\left(n_2 +\frac 12\right)-{q^2 E_x^2 \over
2M\omega_x^2} -{q^2 E_y^2 \over 2M\omega_y^2},\quad n_1,
n_2=0,1,2,\ldots.$$ The total energy is the sum of ${\cal E}^{xy}_{n_1 n_2}$ and ${\cal
E}_{n_3}^z$ given by Eq. (\[8\]). The wave functions on the $xy$ plane will be worked out in the next section.
The method employed in this section is similar to that used in Ref. [@5] which dealt with a charged particle moving in a rotating magnetic field (also studied in Refs. [@6] and [@7]). However, we present here a relation between the right and left eigenvectors, which leads to the very useful relation (\[33\]), so that the calculations here seems more straightforward and easier. The method can be easily extended to systems with more degrees of freedom.
The wave functions in the coordinate representation
===================================================
The eigenstates of the Hamiltonian (\[40\]) corresponding to the eigenvalues (\[41\]) are $$\label{43}
|n_1n_2\rangle={1\over \sqrt{n_1! n_2!}}
(a_1^\dagger)^{n_1}(a_2^\dagger)^{n_2}|00\rangle, \quad n_1,
n_2=0,1,2,\ldots,$$ where $|00\rangle$ is the ground state satisfying $$\label{44}
a_1|00\rangle=a_2|00\rangle=0.$$ The task of this section is to work out these states in the coordinate representation, that is, the wave functions $$\label{45}
\psi_{n_1n_2}(x,y)=\langle xy|n_1n_2\rangle.$$ For this purpose the left eigenvectors are necessary. These are given by
\[46\] $$\label{46a}
u_i=K_i{\bbox(} -{\mathrm i}
M\sigma_i(\sigma_i^2-\omega_y^2-2\omega_B^2),\;
\sigma_i^2-\omega_y^2,\; M\omega_B(\sigma_i^2+\omega_y^2),\; {\mathrm
i}2\omega_B\sigma_i {\bbox)},\quad i=1,2$$ and their complex conjugates, where the normalization constants are $$\label{46b}
K_i=\{{2M\sigma_i}[(\sigma_i^2-\omega_y^2)^2+4\omega_B^2\omega_y^2]\}
^{-1/2}.$$
When $\omega_B=0$, $\{\sigma_i|i=1,2\}=\{\omega_x, \omega_y\}$. For $\omega_x$, the last two components of $u_i$ vanish. For $\omega_y$, the first two components of $u_i$ vanish. Thus the orthogonality is ensured. If the further condition $\omega_x=\omega_y$ holds (which yields $\sigma_1=\sigma_2$), one may take the limit $\omega_y\to
\omega_x$ after everything is worked out.
First we should work out the ground state. In the coordinate representation, Eq. (\[44\]) takes the form $$\label{47}
(\xi_{ij}\tilde x_j-{\mathrm
i}\hbar\eta_{ij}\partial_j)\psi_{00}(x,y)=0,\quad i=1,2,$$ where $x_1=x$, $x_2=y$ and similarly for $\tilde x_1$ and $\tilde
x_2$, $\partial_j=\partial/\partial x_j=\partial/\partial\tilde x_j$, and we have defined two $2\times 2$ matrices $$\label{48}
\xi=(\xi_{ij})=\left(
\begin{array}{cc}
u_{11} & u_{13} \\
u_{21} & u_{23}
\end{array}
\right),\quad \eta=(\eta_{ij})=\left(
\begin{array}{cc}
u_{12} & u_{14} \\
u_{22} & u_{24}
\end{array}
\right),$$ where $u_{i\beta}$ ($i=1,2$; $\beta=1,2,3,4$) is the $\beta$th component of $u_i$. Suppose that $$\label{49}
\psi_{00}(x,y)=N_0\exp[-s(x, y)],\quad s(x, y)={\tilde x_i
\Lambda_{ij} \tilde x_j/2\hbar},$$ where $S$ is a $2\times 2$ symmetric matrix whose elements are complex numbers, and $N_0$ is a normalization constant. Substituting this into Eq. (\[47\]) we obtain $$\label{50}
\Lambda={\mathrm i}\eta^{-1}\xi.$$ This can be easily worked out. The elements are
\[51\] $$\label{51a}
\Lambda_{11}={M\omega_x(\sigma_1+\sigma_2) \over \omega_x+\omega_y}
\equiv \hbar\lambda_x^2,$$ $$\label{51b}
\Lambda_{22}={M\omega_y(\sigma_1+\sigma_2) \over \omega_x +\omega_y}
\equiv \hbar\lambda_y^2,$$ $$\label{51c}
\Lambda_{12}=\Lambda_{21}={{\mathrm i}M\omega_B(\omega_x-\omega_y)
\over \omega_x +\omega_y} \equiv {\mathrm i}\hbar\lambda_{xy},$$
where the relation $\sigma_1\sigma_2=\omega_x\omega_y$ has been used. Therefore the ground-state wave function is
\[52\] $$\label{52a}
\psi_{00}(x,y)=N_0\exp(\textstyle -\frac 12 \lambda_x^2\tilde x^2
-\frac 12 \lambda_y^2\tilde y^2-{\mathrm i}\lambda_{xy}\tilde x\tilde
y),$$ where the normalization constant is $$\label{52b}
N_0=\sqrt{\lambda_x\lambda_y\over\pi} =\left[{M(\sigma_1+\sigma_2)
\sqrt{\omega_x\omega_y}\over \pi\hbar(\omega_x+\omega_y)}
\right]^{1/2}.$$
When $\omega_B=0$, we have $\lambda_x=\sqrt{M\omega_x/\hbar}$, $\lambda_y=\sqrt{M\omega_y/\hbar}$, and $\lambda_{xy}=0$. On the other hand, if $\omega_x=\omega_y$, we have $\omega_1=\omega_2$, $\sigma_1=\omega_1+|\omega_B|$, $\sigma_2=\omega_1-|\omega_B|$, and $\lambda_x=\lambda_y=\sqrt{M\omega_1/\hbar}$, $\lambda_{xy}=0$. These are all expected results.
It is remarkable that the uncertainty relation for the above ground state is $$\label{53}
\Delta x\Delta p_x=\Delta y\Delta p_y={\hbar\over 2}\left(1+
{\lambda_{xy}^2\over\lambda_x^2\lambda_y^2}\right)^{1/2}\ge
{\hbar\over 2}.$$ The equal sign holds only when $B=0$ or $\omega_x=\omega_y$. Therefore the ground state is in general not a minimum wave packet, though it is of the Gaussian type. It seems that this is not noticed in the previous literature.
For any function $F(x,y)$, it is easy to show that $$\label{54}
(\xi^*_{ij}\tilde x_j-{\mathrm i}\hbar\eta^*_{ij}\partial_j)F(x,y)
=-{\mathrm i}\hbar\exp[s^*(x,y)]\eta^*_{ij}\partial_j
\{\exp[-s^*(x,y)]F(x,y)\},\quad i=1,2.$$ According to Eq. (\[43\]), and using the above relation we obtain $$\begin{aligned}
\label{55}
\psi_{n_1n_2}(x,y)=&&{(-{\mathrm i}\sqrt\hbar)^{n_1+n_2}
\over\sqrt{n_1!n_2!}} N_0K_1^{n_1}K_2^{n_2} \exp(\textstyle \frac 12
\lambda_x^2\tilde x^2+ \frac 12 \lambda_y^2\tilde y^2-{\mathrm i}
\lambda_{xy}\tilde x\tilde y)
\nonumber\\
&&[(\sigma_1^2-\omega_y^2)\partial_x -{\mathrm i}
2\omega_B\sigma_1\partial_y] ^{n_1}
[(\sigma_2^2-\omega_y^2)\partial_x -{\mathrm i}
2\omega_B\sigma_2\partial_y] ^{n_2} \exp( -\lambda_x^2\tilde
x^2-\lambda_y^2\tilde y^2).\end{aligned}$$
Some special cases
==================
In this section we discuss some special cases where some of the parameters are zero.
1\. If $\omega_z=0$ and $E_z=0$, the motion in the $z$ direction is free, and ${\cal E}_z\ge 0$ is continuous.
2\. If $\omega_z=0$ but $E_z\ne 0$, the motion in the $z$ direction is that in a homogeneous electric field, which is also well known, and ${\cal E}_z$ is also continuous, and can take on any real value [@8].
3\. In the preceding sections we have assumed that both $\omega_x$ and $\omega_y$ are nonzero. In this case, if any one of $E_x$ or $E_y$ is zero, it can be substituted into the above results directly and no modification to the formalism is needed. However, if one of $\omega_x$ and $\omega_y$ is zero, the situation is different. We would deal with $\omega_y=0$ in the following. The other case could be discussed in a similar way.
4\. If $\omega_y=0$ but $E_y\ne 0$, the Hamiltonian could not be reduced to a quadratic form, and to our knowledge the problem could not be solved analytically.
5\. If $\omega_y=0$ and $E_y=0$, we choose the gauge $A_x=0$, $A_y=Bx$, and let $$\label{56}
\psi(x,y)={1\over\sqrt{2\pi}}\exp({\mathrm i}ky)\phi(x),$$ then $\phi(x)$ satisfies the equation $$\label{57}
\left[{1 \over 2M}p_x^2 +{1\over 2}M\tilde\omega_1^2 (x-x_k)^2\right]
\phi(x) ={\cal E}^x\phi(x),$$ where $$\label{58}
\tilde\omega_1=\sqrt{\omega_x^2+4\omega_B^2},\quad
x_k={qE_x+2k\hbar\omega_B \over M\tilde\omega_1^2},$$ and $$\label{59}
{\cal E}^x={\cal E}^{xy}+{(qE_x+2k\hbar\omega_B)^2 \over
2M\tilde\omega_1^2}-{\hbar^2 k^2 \over 2M}.$$ The energy levels and wave functions for Eq. (\[57\]) are readily available, so the final results are $$\label{60}
{\cal E}^{xy}_{n_1 k}=\hbar\tilde\omega_1\left(n_1 +\frac 12\right)-
{(qE_x+2k\hbar\omega_B)^2 \over 2M\tilde\omega_1^2}+{\hbar^2 k^2
\over 2M},\quad n_1=0,1,2,\ldots,\; k\in(-\infty,+\infty),$$ and the corresponding wave functions are $$\label{61}
\psi_{n_1 k}(x,y)={1\over\sqrt{2\pi}}\exp({\mathrm i}ky)
\psi_{n_1}^{\tilde\omega_1}(x-x_k).$$ If further $\omega_x$ or $E_x$ or both are zero, the results can be obtained from the above ones directly.
6\. Finally, if both $\omega_x$ and $\omega_y$ are zero, one can appropriately choose the coordinate axes of the $xy$ plane such that $E_y=0$. Thus it is a special case of the above case 5.
A more general magnetic field
=============================
In this section we set $\omega_z=0$ and $E_z=0$, and consider a magnetic field with an $x$ or $y$ component (but not both) in addition to the $z$ component. To our knowledge this was not considered previously. Without loss of generality we take $$\label{62}
{\mathbf B}=B'{\mathbf e}_x+B{\mathbf e}_z.$$ The gauge convenient for the present case might be $$\label{63}
A_x=0,\quad A_y=B\tilde x=B(x-x_0),\quad A_z=B'y,$$ where $x_0$ is the same as defined before. In this gauge the total Hamiltonian takes the form $$\label{64}
H^{\rm T}={1 \over 2M}(p_x^2 +p_y^2) +{1 \over
2M}(p_z-2M\omega_B'y)^2 +{1\over 2}M\tilde\omega_1^2\tilde x^2
+{1\over 2}M\omega_y^2 y^2 -qE_y y- 2\omega_B\tilde xp_y-{q^2 E_x^2
\over 2M\omega_x^2},$$ where $\omega_B$ and $\tilde\omega_1$ are the same as defined before, and $\omega_B'=qB'/2Mc$. $p_z$ is obviously a conserved quantity, so we make the factorization $$\label{65}
\Psi({\mathbf x}) ={1\over\sqrt{2\pi}}\exp({\mathrm i}kz)\psi(x,y).$$ The Schrödinger equation (\[1\]) is then reduced to the form (\[14a\]) where now $$\label{66}
H={1 \over 2M}(p_x^2 +p_y^2) +{1\over 2}M\tilde\omega_1^2 \tilde x^2
+{1\over 2}M\tilde\omega_2^2 \tilde y^2-2\omega_B\tilde xp_y,$$ and $$\label{67}
{\cal E}={\cal E}^{\rm T}+{q^2E_x^2 \over
2M\omega_x^2}+{(qE_y+2k\hbar\omega_B')^2 \over
2M\tilde\omega_2^2}-{\hbar^2 k^2 \over 2M}.$$ In these equations $\tilde y=y-y_k$ which is different from the previous one, and $$\label{68}
\tilde\omega_2=\sqrt{\omega_y^2+4\omega_B'^2},\quad
y_k={qE_y+2k\hbar\omega_B' \over M\tilde\omega_2^2}.$$ Now the reduced Hamiltonian $H$ is quadratic, and different from Eq. (\[14b\]) only in the last term. Thus the problem can be solved in much the same way as before. We only give the results here. The energy levels are $$\begin{aligned}
\label{69}
{\cal E}^{\rm T}_{n_1 n_2 k}=&&\hbar\sigma_1\left(n_1 +\frac
12\right)+ \hbar\sigma_2\left(n_2 +\frac 12\right)-{q^2 E_x^2 \over
2M\omega_x^2} -{(qE_y+2k\hbar\omega_B')^2 \over
2M\tilde\omega_2^2}+{\hbar^2 k^2 \over 2M},\nonumber\\ && n_1,
n_2=0,1,2,\ldots, \; k\in(-\infty,+\infty),\end{aligned}$$ where $\sigma_1$ and $\sigma_2$ are given by Eqs. (\[25\]), (\[22b\]) and (\[23\]), but with $\omega_y$ replaced by $\tilde\omega_2$. The wave functions are given by $$\label{70}
\Psi_{n_1n_2k}({\mathbf x}) ={1\over\sqrt{2\pi}}\exp({\mathrm i}kz)
\psi_{n_1n_2k}(x,y),$$ where $\psi_{n_1n_2k}(x,y)$ has the same form as Eq. (\[55\]), with $\omega_y$ replaced by $\tilde\omega_2$ everywhere, including in the expressions for $\lambda_x$, $\lambda_y$, $K_1$ and $K_2$, but here $$\label{71}
\lambda_{xy}=-{2M\omega_B\tilde\omega_2 \over \hbar(\omega_x
+\tilde\omega_2)}.$$ Also note that $\tilde y$ depends on $k$, which is the reason why $\psi_{n_1n_2k}(x,y)$ has the subscript $k$. The orthonormal relation is $$\label{72}
\int \Psi^*_{n_1'n_2'k'}({\mathbf x})\Psi_{n_1n_2k}({\mathbf x})\;
d{\mathbf x} =\delta(k-k')\delta_{n_1 n_1'}\delta_{n_2 n_2'}.$$ By the way, the wave functions (\[61\]) satisfy a similar orthonormal relation.
As before, the ground state obtained here, though being of the Gaussian type, is not a minimum wave packet, except when $B=0$ (then ${\mathbf B}=B'{\mathbf e}_x$).
Coherent and squeezed states
============================
Coherent and squeezed states are useful objects in quantum mechanics and quantum optics. These are widely discussed in the literature. In the presence of magnetic fields, some discussions can be found in Ref. [@fan]. There are also examples on other applications of such states [@jing]. However, such states for anisotropic harmonic oscillators in the presence of magnetic fields are not discussed previously, to the best of our knowledge. In this section we will discuss the definition, the properties and the time evolution of such states for the case treated in Sec. III and IV. The discussions can be easily extended to the case of Sec. VI.
Since the motion in the $z$ direction is separable and is essentially that of a usual harmonic oscillator, we only discuss the coherent and squeezed states on the $\tilde x\tilde y$ plane (which is a simple translation of the $xy$ plane), and their time evolution as governed by the Hamiltonian $H$ given by Eq. (\[14a\]) or (\[40\]) (the additional constants in $H^{xy}$ has only trivial consequence for the time evolution). The definitions used below are natural generalizations of those for a single harmonic oscillator [@nieto; @ni; @ka].
We define a unitary displacement operator $D(\alpha_1,\alpha_2)=D_1(\alpha_1) D_2(\alpha_2)$ where $$\label{73}
D_i(\alpha_i)=\exp(\alpha_i a_i^\dagger-\alpha_i^* a_i),\quad i=1,2$$ and the $\alpha_i$’s are complex numbers. For an arbitrary state $|\varphi\rangle$ one can define a corresponding displaced state $$\label{74}
|\varphi,\alpha_1\alpha_2\rangle_D=D(\alpha_1,\alpha_2)|\varphi\rangle=
D_1(\alpha_1) D_2(\alpha_2)|\varphi\rangle.$$ It is easy to show that $$\label{75}
D^\dagger(\alpha_1,\alpha_2)a_i D(\alpha_1,\alpha_2)=a_i +\alpha_i,
\quad i=1,2,$$ so if $|\varphi\rangle=|00\rangle$ is the ground state we have $$\label{76}
a_i|00,\alpha_1\alpha_2\rangle_D=\alpha_i|00,\alpha_1\alpha_2\rangle_D,
\quad i=1,2.$$ Thus the coherent states may be defined as the displaced ground state $|00,\alpha_1\alpha_2\rangle_D$. In terms of the original variables the displacement operator takes the form $$\label{77}
D(\alpha_1,\alpha_2)=\exp\left(-{{\mathrm i}\over
2\hbar}x^D_{i}p^D_{i}\right) \exp\left({{\mathrm i}\over
\hbar}p^D_{i}\tilde x_i\right) \exp\left(-{{\mathrm i}\over
\hbar}x^D_{i}p_{i}\right),$$ where $$\label{78}
x^D_i={\mathrm i}\sqrt\hbar(\alpha_j\eta^*_{ji}-{\mathrm c.c.}),
\quad p^D_i=-{\mathrm i}\sqrt\hbar(\alpha_j\xi^*_{ji}- {\mathrm
c.c.}).$$ If the wave function for the state $|\varphi\rangle$ is $\varphi(\tilde x,\tilde y)$, then the one for the displaced state $|\varphi,\alpha_1\alpha_2\rangle_D$ is $$\label{79}
\varphi_{\alpha_1\alpha_2}^D(\tilde x,\tilde y)=\exp\left(-{{\mathrm
i}\over 2\hbar}x^D_{i}p^D_{i}\right) \exp\left({{\mathrm i}\over
\hbar}p^D_{i}\tilde x_i\right) \varphi(\tilde x-x^D,\tilde y-y^D).$$ Thus we see the reason why the operator $D(\alpha_1,\alpha_2)$ is called a displacement operator. From this and Eq. (\[52\]) it is easy to obtain the wave function for the coherent states. Obviously the displaced state and the original one have the same shape in the configuration space, thus the uncertainty $\Delta X_\alpha$ in $|\varphi,\alpha_1\alpha_2\rangle_D$ is the same as that in $|\varphi\rangle$. This can also be easily shown by using Eqs. (\[37\]) and (\[75\]).
Now consider the time evolution of the displaced states. It is easy to show that $$\label{80}
{\mathrm e}^{-{\mathrm i}Ht/\hbar}a_i {\mathrm e}^{{\mathrm i}Ht
/\hbar}=\exp({\mathrm i}\sigma_i t)a_i, \quad i=1,2.$$ If the state at the initial time $t=0$ is $|\psi(0)\rangle=
|\varphi,\alpha_1\alpha_2 \rangle_D$, then the state at the time $t$ is $$\label{81}
|\psi(t)\rangle=|\varphi(t),\alpha_{1t}\alpha_{2t}\rangle_D
=D_1(\alpha_{1t})D_2(\alpha_{2t}) |\varphi(t)\rangle,$$ where $\alpha_{it}=\exp(-{\mathrm i}\sigma_i t)\alpha_i$ and $|\varphi(t)\rangle= {\mathrm e}^{-{\mathrm i}Ht/\hbar}
|\varphi\rangle.$ Therefore if $|\varphi(t)\rangle$ is known, $|\psi(t)\rangle$ can be obtained by a time-dependent displacement. A simple special case is $|\varphi\rangle=|n_1 n_2\rangle$, that is $$\label{82}
|\psi(0)\rangle=|n_1 n_2,\alpha_1\alpha_2 \rangle_D,$$ a displaced number state. In this case $|\varphi(t)\rangle=\exp[-{\mathrm i}(n_1+\textstyle\frac 12)\sigma_1
t -{\mathrm i}(n_2+\textstyle\frac 12)\sigma_2 t]|n_1 n_2\rangle$, and $$\label{83}
|\psi(t)\rangle=\exp[-{\mathrm i}(n_1+\textstyle\frac 12)\sigma_1 t
-{\mathrm i}(n_2+\textstyle\frac 12)\sigma_2 t] |n_1 n_2,
\alpha_{1t}\alpha_{2t} \rangle_D.$$ This means that $|\psi(t)\rangle$ is also a displaced number state, except that the displacement parameters are time dependent, and a time-dependent phase factor is gained. The position and velocity of the wave packet $\psi(t,\tilde x,\tilde y)=\langle \tilde x\tilde
y|\psi(t)\rangle$ is characterized by the expectation values $x_\alpha^{\mathrm c}(t)= \langle\psi(t)|X_\alpha|\psi(t)\rangle$. Using Eq. (\[37\]) it is easy to show that $$\label{84}
x_\alpha^{\mathrm c}(t)=\sqrt\hbar[\exp(-{\mathrm i}\sigma_j t)
\alpha_j v_{j\alpha}+{\mathrm c.c.}].$$ By straightforward calculations one can show that they satisfy the following equation $$\label{85}
\dot x_\alpha^{\mathrm c}(t)=\Omega_{\alpha\beta} x_\beta^{\mathrm
c}(t),$$ where the eigenvalue equation (\[27\]) has been used, and at $t=0$ they give the same results as those obtained by using Eqs. (\[78\]) and (\[79\]). On the other hand, the equations of motion for $X_\alpha$ governed by the Hamiltonian $H$ according to classical mechanics are $$\label{86}
\dot X_\alpha=\{X_\alpha,H\}_{\mathrm PB}=\Omega_{\alpha\beta}
X_\beta,$$ where the Poisson bracket $\{X_\alpha,H\}_{\mathrm PB}$ turns out to be of the same form as the commutator $[X_\alpha,H]/{\mathrm i}\hbar$ because $H$ is quadratic in $X_\alpha$. The above two equations mean that the center of the wave packet moves like a classical particle. In fact, it can be shown that this holds for any wave packet in any quadratic system. For the displaced number states, some more specific properties should be emphasized. First, the shape of the wave packet keep unchanged with time. Second, the center of the wave packet is oscillating, but the motion is in general not periodic except when $\sigma_2/\sigma_1$ is a rational number.
Next we discuss the squeezed states. We define a unitary squeeze operator $S(\zeta_1,\zeta_2)=S_1(\zeta_1) S_2(\zeta_2)$ where $$\label{87}
S_i(\zeta_i)=\exp(\textstyle\frac12\zeta_i a_i^\dagger a_i^\dagger
-\frac12\zeta_i^* a_i a_i),\quad i=1,2,$$ and the $\zeta_i$’s are complex numbers. For an arbitrary state $|\varphi\rangle$ the corresponding squeezed state may be defined as $$\label{88}
|\varphi,\zeta_1\zeta_2\rangle_S=S(\zeta_1,\zeta_2)|\varphi\rangle=
S_1(\zeta_1) S_2(\zeta_2)|\varphi\rangle.$$ The following equation is useful for subsequent calculations. $$\label{89}
S^\dagger(\zeta_1,\zeta_2)a_i S(\zeta_1,\zeta_2)=a_i\cosh\rho_i
+a_i^\dagger{\mathrm e}^{{\mathrm i}\phi_i}\sinh\rho_i, \quad i=1,2,$$ where $\rho_i=|\zeta_i|$ and $\phi_i=\arg\zeta_i$.
Unlike the displaced state, it is difficult to obtain the explicit wave function for the squeezed state in terms of the wave function for the original one. However, in the squeezed number state $|n_1n_2,
\zeta_1 \zeta_2 \rangle_S$, it can be shown by using Eq. (\[89\]) that $$\label{90}
\langle X_\alpha\rangle=0,$$ $$\label{91}
\Delta X_\alpha=\{\hbar(2n_j+1)[|v_{j\alpha}|^2\cosh2\rho_j+{\mathrm
Re}(v_{j\alpha}^2{\mathrm e}^{{\mathrm i}\phi_j})\sinh2\rho_j]\}
^{1/2}.$$ Compared with the results for $\zeta_1=\zeta_2=0$, we see that the center of the squeezed state is the same as that of the original one, but the uncertainties are different. Thus the states are indeed “squeezed”.
Now consider the time evolution of the squeezed states. If the state at the initial time $t=0$ is $|\psi(0)\rangle=
|\varphi,\zeta_1\zeta_2 \rangle_S$, then the state at the time $t$ is $$\label{92}
|\psi(t)\rangle=|\varphi(t),\zeta_{1t}\zeta_{2t}\rangle_S
=S_1(\zeta_{1t})S_2(\zeta_{2t}) |\varphi(t)\rangle,$$ where $\zeta_{it}=\exp(-{\mathrm i}2\sigma_i t)\zeta_i$ and $|\varphi(t)\rangle= {\mathrm e}^{-{\mathrm i}Ht/\hbar}
|\varphi\rangle.$ Therefore if $|\varphi(t)\rangle$ is known, $|\psi(t)\rangle$ can be obtained by a time-dependent squeeze. A simple special case is $|\varphi\rangle=|n_1 n_2\rangle$, that is $$\label{93}
|\psi(0)\rangle=|n_1 n_2,\zeta_1\zeta_2 \rangle_S,$$ a squeezed number state. In this case $|\varphi(t)\rangle=\exp[-{\mathrm i}(n_1+\textstyle\frac 12)\sigma_1
t -{\mathrm i}(n_2+\textstyle\frac 12)\sigma_2 t]|n_1 n_2\rangle$, and $$\label{94}
|\psi(t)\rangle=\exp[-{\mathrm i}(n_1+\textstyle\frac 12)\sigma_1 t
-{\mathrm i}(n_2+\textstyle\frac 12)\sigma_2 t] |n_1 n_2,
\zeta_{1t}\zeta_{2t} \rangle_S.$$ This means that $|\psi(t)\rangle$ is also a squeezed number state, except that the squeeze parameters are time dependent, and a time-dependent phase factor is gained. Though it is difficult to obtain the wave function $\psi(t,\tilde x,\tilde y)=\langle \tilde
x\tilde y|\psi(t)\rangle$ explicitly, in these states it can be shown that $$\label{95}
\langle X_\alpha\rangle_t=0,$$ and $$\label{96}
\Delta_t
X_\alpha=\{\hbar(2n_j+1)[|v_{j\alpha}|^2\cosh2\rho_j+{\mathrm
Re}(v_{j\alpha}^2{\mathrm e}^{{\mathrm i}\phi_j-{\mathrm i}2\sigma_j
t})\sinh2\rho_j]\} ^{1/2}.$$ These results mean that the center of the wave packet is at rest, but the uncertainties are oscillating. Thus the shape of the wave packet changes with time apparently. As before, the motion is in general not periodic except when $\sigma_2/\sigma_1$ is a rational number.
Summary
=======
In this paper we have studied a charged anisotropic harmonic oscillator moving in a homogeneous static electromagnetic field. Several configurations of the electromagnetic field are considered. One of these configurations has been studied in the literature. However, the formalism used here seems more convenient. We have studied the coherent and squeezed states of these systems in some detail. In terms of these concepts the time evolution of some wave packets can be discussed very conveniently.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported by the National Natural Science Foundation of the People’s Republic of China, and by the Foundation of the Advanced Research Center of Sun Yat-Sen University.
Note added after publication {#note-added-after-publication .unnumbered}
============================
After this paper had been published we became aware of some more previous works dealing with the same problem [@Schuh; @Davies]. In Ref. [@Schuh] the energy spectrum was worked out by an algebraic approach. In Ref. [@Davies] the propagator was calculated by a different method from that of Refs. [@3; @4]. We thank Prof. J. M. Cervero for pointing out these papers to us.
[99]{}
U. Merkt, J. Huser and M. Wagner, Phys. Rev. B [**43**]{}, 7320 (1991).
K.-D. Zhu and S.-W. Gu, Phys. Lett. A [**181**]{}, 465 (1993).
M. Moshinsky and P. Winternitz, J. Math. Phys. [**21**]{}, 1667 (1980).
R. Bogdanovic and M. S. Gopinathan, J. Phys. A [**12**]{}, 1457 (1979).
N. Kokiantonis and D. P. L. Castrigiano, J. Phys. A [**18**]{}, 45 (1985).
M. A. Z. Habeeb, J. Phys. A [**20**]{}, 5549 (1987).
B. Mielnik and D. J. Fernández C., J. Math. Phys. [**30**]{}, 537 (1989).
Q.-G. Lin, Phys. Rev. A [**63**]{}, 012108 (2001).
Q.-G. Lin, J. Phys. A [**34**]{}, 1903 (2001).
L. D. Landau and E. M. Lifshitz, [*Quantum Mechanics*]{}, 3rd ed. (Pergamon, Oxford, 1977).
H.-Y. Fan, Phys. Lett. A [**126**]{}, 150 (1987); H.-Y. Fan, Y.-H. Wang and Y. Fan, Commun. Theor. Phys. [**30**]{}, 125 (1998); H.-Y. Fan and Y. Fan, Commun. Theor. Phys. [**30**]{}, 301 (1998).
H. Jing, Y.-A. Han, J.-L. Chen and Y.-X. Miao, Chin. Phys. Lett. [**12**]{}, 862 (2000).
M. M. Nieto, Phys. Lett. A [**229**]{}, 135 (1997).
G.-J. Ni and S.-Q. Chen, [*Advanced Quantum Mechanics*]{} (Fudan Univ. Press, Shanghai, 2000). (in Chinese)
X.-L. Ka, [*Advanced Quantum Mechanics*]{}, 1st ed. (Higher Education Press, Beijing, 1999); 2nd ed. (Higher Education Press, Beijing, 2001). (in Chinese)
B. Schuh, J. Phys. A [**18**]{}, 803 (1985).
I. M. Davies, J. Phys. A [**18**]{}, 2737 (1985).
[^1]: not for correspondence
[^2]: E-mail: qg\[email protected]
[^3]: published in Commun. Theor. Phys. **38** (2002) 667-674.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we classify Gorenstein stable log surfaces with $(K_X+\Lambda)^2=p_g(X,\Lambda)-1$.'
address: 'Yau Mathematical Sciences Center, Tsinghua University, Beijing 100084'
author:
- Jingshan Chen
title: ' Gorenstein stable log surfaces with $(K_X+\Lambda)^2=p_g(X,\Lambda)-1$'
---
Introduction
============
KSBA stable (log) surfaces are the two-dimensional analogues of stable (pointed) curves. They are the fundamental objects in compactifying the moduli spaces of smooth surfaces of general type.
In general, stable (log) surfaces are difficult to classify. We may first focus on Gorenstein ones, i.e. $K_X$ (resp. $K_X+\Lambda$) being Cartier. In [@LR13], Liu and Rollenske give several inequalities for the invariants of Gorenstein stable log surfaces. One important inequality among them is the stable log Noether inequality $(K_X+\Lambda)^2\ge p_g(X,\Lambda)-2$ (see [@LR13 Thm 4.1]). This can be rephrased as $\Delta(X,K_X+\Lambda)\ge 0 $, where $\Delta$ is Fujita’s $\Delta$-genus.
Gorenstein stable log surfaces with $\Delta(X,K_X+\Lambda)=0$ have been classified in [@LR13]. Normal Gorenstein stable log surfaces with $\Delta(X,K_X+\Lambda)=1$ have been classified in [@Chen18]. Here we continue to classify non-normal Gorenstein stable log surfaces with $\Delta(X,K_X+\Lambda)=1$. The main result is as follows.
\[Main Theorem \] Let $(X,\Lambda)$ be a Gorenstein stable log surface with $\Delta(X,K_X+\Lambda)=1$ which is not normal.
If $X$ is irreducible, let $\pi\colon \bar{X}\to X$ be the normalization map. Then
- either $\Delta(\bar{X},\pi^*(K_X+\Lambda))=1$ and $(X,\Lambda)$ is as in Thm \[delta1,1\],
- or $\Delta(\bar{X},\pi^*(K_X+\Lambda))=0$ and $(\bar{X},\bar{\Lambda}+\bar{D})$ is as in Thm \[delta1,0\].
If $X$ is reducible, write $X=\bigcup X_i$, where $X_i$ is an irreducible component. Then $\Delta(X_i,(K_X+\Lambda)|_{X_i})=1$ or $0$ for each component $X_i$, $X$ has a unique minimal connected component $U$ such that $\Delta(U,(K_X+\Lambda)|_{U})=1$, $X\setminus U$ is composed with several trees $T_j$ with $\Delta(T_j,(K_X+\Lambda)|_{T_j})=0$ and $X$ is glued by $U$ and $T_j$ along lines.
Moreover, $U$ is one of the followings:
- $X=U$ is a string of log surfaces $X_i$ with $\Delta(X_i,(K_X+\Lambda)|_{X_i})=1$ and $|(K_{X}+\Lambda)|_{X_i}|$ composed with a pencil of elliptic curves. The connecting curves are all fibers. Moreover, in this case $|K_X+\Lambda|$ is composed with a pencil of elliptic curves.
- $U$ is a string of surfaces glued along lines. The end surfaces $X_i$ of the string $U$ are non-normal with $\Delta(X_i,(K_X+\Lambda)|_{X_i})=1$.
- $U$ is composed with a single irreducible log surface $X_k$ with $\Delta(X_k,(K_X+\Lambda)|_{X_k})=1$ and $(K_{X}+\Lambda)|_{X_k}$ very ample. Moreover, in this case $K_X+\Lambda$ is very ample.
- $U=X_j\cup X_k $ where $X_j$, $X_k$ are two Gorenstein log surfaces with $\Delta(X_j,(K_X+\Lambda)|_{X_j})=\Delta(X_k,(K_X+\Lambda)|_{X_k})=0$. All the connecting curves of $X$ are lines except $X_j\cap X_k$.
- $U$ is a cycle of log surface $X_i$ with $\Delta(X_i,(K_X+\Lambda)|_{X_i})=0$. All the connecting curves of $X$ are lines.
Now we give a brief account of each section. In \[prelim\], we recall some definitions and facts about Gorenstein stable log surfaces. In \[zero-Delta-genus\], we recall the definition of Fujita’s $\Delta$-genus and include some results about normal Gorenstein stable log surfaces $(X,\Lambda)$ with $\Delta(X,K_X+\Lambda)=0$ or $1$. In \[non-normal\], we include some results about non-normal Gorenstein stable log surfaces. In \[nnorm\], we deal with the case that $X$ is non-normal and irreducible. In \[reducible-stable\], we deal with the case that $X$ is reducible. Finally we describe Gorenstein stable surfaces with $K_X^2= p_g-1$.
Acknowledgements: {#acknowledgements .unnumbered}
-----------------
I am grateful to Prof. Jinxing Cai and Prof. Wenfei Liu for their instructions. I would also thank the anonymous referee for helpful advices and suggestions.
Notations and conventions
-------------------------
We work exclusively with schemes of finite type over the complex numbers $\mathbb{C}$.
- A surface is a connected reduced projective schemes of pure dimension two.
- By abuse of notation, we sometimes do not distinguish a Cartier divisor $D$ and its associated invertible sheaf $\mathcal{O}_X(D)$.
- $\Sigma_d$ denotes a Hirzebruch surface, which admits a $\mathbb{P}^1$ fibration over $\mathbb{P}^1$. We denote $\Gamma$ as a fiber. It has a unique 1-section $\Delta_0$ whose self-intersection is $-d$.
- We use ’$\equiv$’ to denote linear equivalent relation of divisors.
- If $D$ is a Cartier divisor on $X$, then we denote $\Phi_{|D|}\colon X\dashrightarrow \mathbb{P}:=|D|^*$ as the rational map defined by the linear system $|D|$.
- A line $l$ on a variety $X$ with respect to $\mathcal{O}_X(1)$ is a rational curve such that $l\cdot \mathcal{O}_X(1)=1$.
Preliminaries {#prelim}
=============
Let $X$ be a demi-normal surface, i.e. $X$ satisfies $S_2$ and is at worst ordinary double at any generic point of codimension 1. Denote $\pi\colon \bar X \to X$ as the normalisation map of $X$. The conductor ideal $ \mathrm{\mathcal{H}om}_{\mathcal{O}_X}(\pi_*\mathcal{O}_{\bar{X}}, \mathcal{O}_X)$ is an ideal sheaf both on $X$ and $\bar{X}$ and hence defines subschemes $D\subset X \text{ and } \bar D\subset \bar X$, both reduced and of pure codimension 1; we often refer to $D$ as the non-normal locus of $X$.
Let $\Lambda$ be a reduced curve on $X$ whose support does not contain any irreducible component of $D$. Then the strict transform $\bar \Lambda$ in the normalization is well defined. We have $\pi^*(K_X+\Lambda)=K_{\bar{X}}+\bar D+\bar \Lambda$ and $(K_X+\Lambda)^2 = (K_{\bar X}+\bar D+\bar \Lambda)^2$.
\[defin: slc\] We call a pair $(X, \Lambda)$ as above a *log surface*; $\Lambda$ is called the (reduced) boundary.
A log surface $(X,\Lambda)$ is said to have *semi-log-canonical (slc)* singularities if it satisfies the following conditions:
1. $K_X + \Lambda$ is $\mathbb{Q}$-Cartier, that is, $m(K_X+\Lambda)$ is Cartier for some $m\in\mathbb{Z}^{>0}$; the minimal such $m$ is called the (global) index of $(X,\Lambda)$.
2. The pair $(\bar X, \bar \Lambda+\bar D)$ has log-canonical singularities.
The pair $(X,\Lambda)$ is called stable log surface if in addition $K_X+\Lambda$ is ample. A stable surface is a stable log surface with empty boundary.
By abuse of notation we say $(X, \Lambda)$ is a Gorenstein stable log surface if the index is equal to one, i.e. $K_X+\Lambda$ is an ample Cartier divisor.
Gorenstein slc singularities and semi-resolutions
-------------------------------------------------
Normalizing a demi-normal surface looses all information on the gluing in codimension one. Often it is better to work on a simpler but still non-normal surface.
A surface $X$ is called semi-smooth if every singularity of $X$ is either double normal crossing or a pinch point [^1]. The normalization of a semi-smooth surface is smooth.
A morphism of demi-normal surfaces $f\colon Y\rightarrow X$ is called a semi-resolution if the following conditions are satisfied:
1. $Y$ is semi-smooth;
2. $f$ is an isomorphism over the semi-smooth open subscheme of $X$;
3. $f$ maps the singular locus of $Y$ birationally onto the non-normal locus of $X$.
A semi-resolution $f\colon Y\rightarrow X$ is called minimal if no $(-1)$-curve is contracted by $f$, that is, there is no exceptional curve $E$ such that $E^2 =K_Y\cdot E = -1$.
Semi-resolutions always exist and one can also incorporate a boundary [@KollarSMMP 10.5].
\[rem: classification of sings\] Semi-log-canonical surface singularities have been classified in terms of their resolution graphs, at least for reduced boundary [@KSB88].
Let $x\in (X, \Lambda)$ be a Gorenstein slc singularity with minimal log semi-resolution $f\colon Y\to X$. Then it is one of the followings (see [@Kollar-Mori Ch. 4], [@KollarSMMP Sect. 3.3], [@kollar12] and [@LR13]):
Gorenstein lc singularities, $\Lambda=0$
: In this case $x\in X$ is a canonical singularity, or a simple elliptic respectively cusp singularity. For the latter the resolution graph is a smooth elliptic curve, a nodal rational curve, or a cycle of smooth rational curves (see also [@Lau77] and [@Reid97 Ch. 4]).
Gorenstein lc singularities, $\Lambda\neq 0$
: Since the boundary is reduced, $\Lambda$ has at most nodes. If $\Lambda$ is smooth so is $X$ because of the Gorenstein assumption.
If $\Lambda$ has a node at $x$ then $x$ is a smooth point of $X$ or $(X, \Lambda)$ is a general hyperplane section of a cyclic quotient singularity. In the minimal log resolution the dual graph of the exceptional curves is $$\bullet \ {-}\ c_1 \ -\ \cdots \ - \ c_n \ {-}
\ \bullet \qquad (c_i\geq1),$$ where $c_i$ represents a smooth rational curve of self-intersection $-c_i$ and each $\bullet$ represents a (local) component of the strict transform of $\Lambda$. If $c_i=1$ for some $i$ then $n=1$ and $\Lambda$ is a normal crossing divisor in a smooth surface.
non-normal Gorenstein slc singularities, $\Lambda=0$
: We describe the dual graph of the $f$-exceptional divisors over $x$: analytically locally $X$ consists of $k$ irreducible components, on each component we have a resolution graph as in the previous item, and these are glued together where the components intersect. In total we have a cycle of smooth rational curve.
non-normal Gorenstein slc singularities, $\Lambda\neq 0$
: The difference to the previous case is that the local components are now glued in a chain and the ends of the chain intersect the strict transform of the boundary. In this case $X$ itself might not even be $\mathbb{Q}$-Gorenstein.
Normal Gorenstein stable log surfaces with small $\Delta$-genus {#zero-Delta-genus}
===============================================================
Let $X$ be a variety and $\mathcal{L}$ be an ample line bundle on it. Fujita introduced several invariants for such polarized varieties. One important of them is the $\Delta$-genus $\Delta(X,\mathcal{L}):=\mathcal{L}^{\dim X}-h^0(X,\mathcal{L})+\dim X$.
\[gepg-2\] Let $(X,\Lambda)$ be a normal irreducible Gorenstein stable log surface. Then $\Delta(X,K_X+\Lambda)\ge 0$.
Moreover if ’=’ holds, then $(X,\Lambda)$ is one of the followings:
- $X$ is $\mathbb{P}^{2}$, $\mathcal{O}_X(K_X+\Lambda) =\mathcal{O}_{\mathbb{P}^{2}}(1)$ and $\Lambda\in|\mathcal{O}_{\mathbb{P}^{2}}(4)|$;
- $X$ is $\mathbb{P}^{2}$, $\mathcal{O}_X(K_X+\Lambda)=\mathcal{O}_{\mathbb{P}^{2}}(2)$ and $\Lambda\in|\mathcal{O}_{\mathbb{P}^{2}}(5)|$;
- $X$ is $\Sigma_d$, $K_X+\Lambda\equiv \Delta_0+\frac{N+d-1}{2}\Gamma$ and $\Lambda\in|3\Delta_0+\frac{N+3d+3}{2}\Gamma|$; ($N-d-3\ge0$ is an even number);
- $X$ is a singular quadric $C_2$ in $\mathbb{P}^{3}$, $\mathcal{O}_X(K_X+\Lambda)=\mathcal{O}_{C_2}(1)=\mathcal{O}_{\mathbb{P}^{2}}(1)|_{C_2}$ and $\Lambda\in|\mathcal{O}_{C_2}(3)|$;
- $X$ is a cone $C_{N-1}\hookrightarrow \mathbb{P}^{N}$, $\mathcal{O}_X(K_X+\Lambda)=\mathcal{O}_{\mathbb{P}^{N}}(1)|_{C_{N-1}}$ and the proper transformation $\bar{\Lambda}$ in the minimal resolution $\Sigma_{N-1}$ is linearly equivalent to $2\Delta_0+2N\Gamma$. ($N>3$)
\[delta-genus-one\] Let $(X,\Lambda)$ be a normal Gorenstein stable log surface with $(K_X+\Lambda)^2= p_g(X,\Lambda)-1$.
Then $(X,\Lambda)$ is one of the followings:
- $X$ is a double cover of $\mathbb{P}^2$. $\Phi_{|K_X+\Lambda|}\colon X\to \mathbb{P}^2$ is the double covering map. The branch curve $B\in |\mathcal{O}_{\mathbb{P}^2}(2k)|$ is a reduced curve which admits curve singularities of lc double-covering type, and $\Lambda\in |{\Phi_{|K_X+\Lambda|}}^*\mathcal{O}_{\mathbb{P}^2}(4-k)|$. ($p_g(X,\Lambda)=3$, and $k=2,3,4$)
- $X$ is a quadric in $\mathbb{P}^3$, and $\Lambda\in |\mathcal{O}_{\mathbb{P}^3}(4)|_X|$. ($p_g(X,\Lambda)=6$)
- $X$ is $\mathbb{P}^2$ blown up at $k$ points (possible infinitely near), and $\Lambda\in |-2K_X|$. ($p_g(X,\Lambda)=10-k$, and $k=0,1,...,7$)
- $X$ is a cone over an elliptic curve of degree $N$ in $\mathbb{P}^{N-1}$, and $\Lambda\in |-2K_X|$. ($p_g(X,\Lambda)=N$)
- $X$ is obtained from $\tilde{X}$ by contracting a $(-n)$ curve $G$, where $n=p_g(X,\Lambda)-1$ and $\tilde{X}$ is an elliptic surface (possible singular) with $G$ as the rational zero section. Every elliptic fiber of $\tilde{X}$ is irreducible. $\Lambda$ is the image of a sum of two different elliptic fibers which admit at worst $A_n$ type singularities.
- $X$ is a (possibly singular) Del Pezzo surface of degree 1, namely $X$ has at most canonical singularities and elliptic singularities, $-K_X$ is ample and $K_X^2=1$. The curve $\Lambda$ belongs to the system $|-2K_X|$, and $p_a(\Lambda)=2$. $p_g(X,\Lambda)=2$.
- $\Lambda=0$, $|K_X+\Lambda|$ is composed with a pencil of genus $2$ curves. $X$ is canonically embedded as a hypersurface of degree 10 in the smooth locus of $\mathbb{P}(1,1,2,5)$. $p_g(X)=2$.
In Theorem 1.1(1) and Theorem 4.3(i), $k=1,2,3,4$ should be corrected by $k=2,3,4$. The case $k=1$ is excluded as $\Phi_{|K_X+\Lambda|}$ would be an embedding.
Non-normal Gorenstein stable log surfaces {#non-normal}
=========================================
Let $(X,\Lambda)$ be a Gorenstein stable log surface which is non-normal. $D$ is the non-normal locus of $X$.
Let $C\subset D$ be a subcurve of the non-normal locus of a Gorenstein stable log surface $(X,\Lambda)$ and $\nu\colon:\tilde{X}\to X$ be the partial normalization of $X$ along $C$. Denote $\tilde{C}$ and $\tilde{\Lambda}$ as the proper transformation of $C$ and $\Lambda$.
Then $(\tilde{X},\tilde{\Lambda}+\tilde{C})$ is a Gorenstein stable log surface.
We first notice that $\nu^*(K_X+\Lambda)=K_{\tilde{X}}+\tilde{\Lambda}+\tilde{C}$ is ample and Cartier. Second it is easy to verify that $(\tilde{X},\tilde{\Lambda}+\tilde{C})$ still has Gorenstein slc singularities only by the classification of Gorenstein slc singularities. Therefore $(\tilde{X},\tilde{\Lambda}+\tilde{C})$ is a Gorenstein stable log surface as well.
The map $\nu|_{\tilde C}\colon \tilde{C} \to C$ is generically a double cover and thus induces a rational involution $\tau$ on $\tilde{C}$.
\[prop: descend section\] If $(X, \Lambda)$ is a non-normal Gorenstein stable log surface and $\nu$, $C$, $\tilde{C}$, $\tau$ are defined as above, then $\nu^*H^0(X, K_X+\Lambda)\subset H^0(\tilde{X}, \nu^*(K_X+\Lambda)))$ is the subspace of those sections $s$ whose restriction to $\tilde{C}$ is $\tau$-anti-invariant.
\[separateNonnormalCurve\] It is easy to see that $\nu^*H^0(X, K_X+\Lambda)$ can not separate $\tilde{C}$. Actually, $\Phi_{\nu^*H^0(X, K_X+\Lambda)}|_{\tilde{C}}\circ \tau=\Phi_{\nu^*H^0(X, K_X+\Lambda)}|_{\tilde{C}}$ on $\tilde{C}$.
\[fibersection\] Let $(X,\Lambda)$ be a reducible Gorenstein stable log surface such that $X=X_1\cup X_2$ and $C:=X_1\cap X_2$ is the connecting curve. Write $(K_X+\Lambda)|_{X_i}:=(\pi^*(K_X+\Lambda))|_{X_i}=K_{X_i}+C+\Lambda|_{X_i}$.
Let $\mathcal{R}_{X_i\to C}\colon H^0(X_i,(K_X+\Lambda)|_{X_i})\to H^0(C,(K_X+\Lambda)|_C)$ be the restriction map. Then we have the following fiber product diagram of vector spaces $$\xymatrix{
H^0(X, K_X+\Lambda) \ar[r]\ar[d] & H^0(X_1,(K_X+\Lambda)|_{X_1})\ar[d]^{\mathcal{R}_{X_1\to C}}\\
H^0(X_2,(K_X+\Lambda)|_{X_2})\ar[r]^{-\mathcal{R}_{X_2\to C}}& H^0(C, (K_X+\Lambda)|_C)
}.$$
Moreover, denote $r_{X_i\to C}((K_X+\Lambda)|_{X_i}):=\dim\mathcal{R}_{X_i\to C}(H^0(X_i,(K_X+\Lambda)|_{X_i}))$. We have $$\begin{aligned}
h^0(X, K_X+\Lambda)\le& h^0(X_1,(K_X+\Lambda)|_{X_1})+h^0(X_2,(K_X+\Lambda)|_{X_2})\\
&-\max\{r_{X_1\to C}((K_X+\Lambda)|_{X_1}),r_{X_2\to C}((K_X+\Lambda)|_{X_2})\}.\end{aligned}$$
irreducible Non-normal Gorenstein stable surfaces with $\Delta(X,K_X+\Lambda)=1$ {#nnorm}
================================================================================
Let $(X,\Lambda)$ be an irreducible non-normal Gorenstein stable log surface and $\bar{X}$, $\pi$, $D$, $\bar{D}$, $\tau$ defined as before. Since $H^0(X,K_X+\Lambda)\cong \pi^*H^0(X,K_X+\Lambda) \subset H^0(\bar X,\pi^*(K_X+\Lambda))$, we have $\Delta(X,K_X+\Lambda)\ge \Delta(\bar{X},\pi^*(K_X+\Lambda))$, and ’=’ holds if and only if $\pi^*H^0(X,K_X+\Lambda) = H^0(\bar X,\pi^*(K_X+\Lambda))$.
\[nonnorm&irred\] Let $(X,\Lambda)$ be an irreducible non-normal Gorenstein stable log surface.
Then $\Delta(X,K_X+\Lambda)\ge 1$.
We only need to show that the case $\Delta(X,K_X+\Lambda)=\Delta(\bar{X},\pi^*(K_X+\Lambda))=0$ does not occur. First, $h^0(X,K_X+\Lambda)=h^0(\bar{X},\pi^*(K_X+\Lambda))$ implies every section in $H^0(\bar{X},\pi^*(K_X+\Lambda))$ is $\tau$-anti-invariant restricting to $\bar{D}$. Hence $H^0(\bar{X},\pi^*(K_X+\Lambda))$ can not separate points of $\bar{D}$ by Remark \[separateNonnormalCurve\]. However, $\Delta(\bar{X},\pi^*(K_X+\Lambda))=0$ implies $\pi^*(K_X+\Lambda)$ is very ample by Cor \[gepg-2\], therefore it will separate points of $\bar{D}$, a contradiction.
\[delta1,1\]
Let $(X,\Lambda)$ be an irreducible non-normal Gorenstein stable log surface as before. Assume $\Delta(X,K_X+\Lambda)=\Delta(\bar{X},\pi^*(K_X+\Lambda))=1$. Then
- either $X$ is a double cover of $\mathbb{P}^2$ induced by $\Phi_{|K_{X}+\Lambda|}$. The branched curve is $2C+B\in |\mathcal{O}_{\mathbb{P}^2}(2m+2k)|$, where $C,B$ are reduced curves of degree $m,2k$. $\Lambda\in |\Phi_{|K_{X}+\Lambda|}^*\mathcal{O}_{\mathbb{P}^2}(4-k-m)|$. $k=2,3$. $0<m\le 4-k$; ($p_g(X,\Lambda)=3$)
- or $\Lambda=0$, $K_X^2=(K_{\bar{X}}+\bar D)^2=1$. $(\bar{X},\bar{D})$ is a normal Gorenstein stable log surface as in Thm \[delta-genus-one\] (6). $\bar D$ is a 2 section with $p_a(\bar D)=2$. $\tau$ is induced from the double covering $\bar D \to \mathbb{P}^1$. ($p_g(X)=2$)
We see that $\pi^*H^0(X,K_X+\Lambda)=H^0(\bar{X},\pi^*(K_X+\Lambda))$ and $(\bar{X},\bar D+\bar{\Lambda})$ is a stable log surface as in Thm \[delta-genus-one\].
First $\Phi_{|\pi^*(K_X+\Lambda)|}$ is not an embedding by Remark \[separateNonnormalCurve\]. Second $\Phi_{|\pi^*(K_X+\Lambda)|}$ is not composed with a pencil of genus 2 curves since $\bar D+\bar{\Lambda}\not= 0$.
If $\Phi_{|\pi^*(K_X+\Lambda)|}$ is composed with a pencil of elliptic curves, $\bar{D}+\bar{\Lambda}$ is either a 2-section or a sum of two elliptic fibers of the fibration map. For the former case, we see that $\bar{\Lambda}=0$ and $\bar{D}$ is a genus 2 curve by Remark \[separateNonnormalCurve\]. Moreover, $(K_{\bar X}+\bar D)^2=1$. $\tau$ is induced from the double map $\bar{D} \to \mathbb{P}^1$. For the latter case, we see that either $\bar{D}=F_1+F_2$ and $\bar{\Lambda}=0$, or $\bar{D}=F_1$ and $\bar{\Lambda}\not=0$, where $F_i$ is a fiber of the fibration map. Both cases can be excluded since the $F_1\cap F_2$ or $F_1\cap\bar{\Lambda}$ on $\bar{X}$ can not be glued into a Gorenstein slc singularity on $X$.
If $\bar{X}$ is a double cover of $\mathbb{P}^2$, then $\bar{D}+\bar{\Lambda}\in |\Phi_{|K_{\bar{X}}+\bar{D}+\bar{\Lambda}|}^*\mathcal{O}_{\mathbb{P}^2}(4-k)|$, $k=2,3$. Remark \[separateNonnormalCurve\] indicates $\bar{D}=\Phi_{|K_{\bar{X}}+\bar{D}+\bar{\Lambda}|}^{-1}(C)$, where $C$ is a curve on $\mathbb{P}^2$. Denote $m$ as the degree of $C$. Then $\bar{\Lambda}\in |\Phi_{|K_{\bar{X}}+\bar{D}+\bar{\Lambda}|}^*\mathcal{O}_{\mathbb{P}^2}(4-k-m)|$. We have the following commutative diagram: $$\xymatrix{
\bar{D}\ar@{^{(}->}[r]\ar[d] & \bar{X}\ar[d]^{\pi}\ar[rr]^{\Phi_{|K_{\bar{X}}+\bar{D}+\bar{\Lambda}|}} && \mathbb{P}^2\\
D\ar@{^{(}->}[r] & X \ar[rru]_{\Phi_{|K_{X}+\Lambda|}}
}.$$
We see that $X$ is also a double cover of $\mathbb{P}^2$ induced by $\Phi_{|K_{X}+\Lambda|}$. The branch curve is $2C+B\in |\mathcal{O}_{\mathbb{P}^2}(2m+2k)|$. $\Lambda\in |\Phi_{|K_{X}+\Lambda|}^*\mathcal{O}_{\mathbb{P}^2}(4-k-m)|$.
Next we consider those irreducible non-normal Gorenstein stable log surfaces $(X,\Lambda)$ with $\Delta(X,K_X+\Lambda)=1$ and $\Delta(\bar{X},\pi^*(K_X+\Lambda))=0$.
We first notice that $\pi^*H^0(X,K_X+\Lambda)\subset H^0(\bar{X},\pi^*(K_X+\Lambda))$ is of codimension one. Hence it has a base point $c\in \mathbb{P}(H^0(\bar{X},\pi^*(K_X+\Lambda)))$. We have the following commutative diagram: $$\xymatrix{
\bar{X} \ar@{^{(}->}[rrr]^<(0.3){\Phi_{|\pi^*(K_X+\Lambda)|}}\ar[d]^{\pi} & & & \mathbb{P}(H^0(\bar{X},\pi^*(K_X+\Lambda)))\ar@{-->}[d]^{pr_c}\\
X\ar@{-->}[rrr]^<(0.3){\Phi_{|K_X+\Lambda|}}& & & \mathbb{P}(\pi^*H^0(X,K_X+\Lambda))
}.$$
We regard $\Phi_{|\pi^*(K_X+\Lambda)|}$ as an inclusion. To describe $pr_c|_{\bar{X}}$, we use the following theorem (see [@Ber06 Thm 2.1]):
\[linearsectionthm\] Let $X$ be a non-degenerate irreducible subvariety of $\mathbb{P}^r$ and $L$ be a linear subspace of $\mathbb{P}^r$ of dimension $s\le r$ such that $L\cap X$ is a 0-dimensional scheme $\zeta$. Then $\mathrm{length}\, \zeta \le \Delta(X,\mathcal{O}(1))+s+1$.
\[lineIntersecting\] Let $X$ be a non-degenerate irreducible subvariety of $\mathbb{P}^r$ with $\Delta(X,\mathcal{O}(1))=0$ and $L$ is a line intersecting $X$ along a 0-dimensional scheme $\zeta$. Then $\mathrm{length}\, \zeta\le 2$.
Let $X$ be a non-degenerate irreducible subvariety of $\mathbb{P}^N$ with $\Delta(X,\mathcal{O}(1))=0$ and $c\in \mathbb{P}^N$ be a point outside $X$. Denote $pr_c\colon \mathbb{P}^N\dashrightarrow \mathbb{P}^{N-1}$ as the projection from $c$. Assume $pr_c|_X$ is birational and is not an isomorphism. Denote by $W$ the image of $pr_c|_X$. Then $W$ is non-normal and $pr_c|_X\colon X\to W$ is a normalization map. The non-normal locus $D$ of $W$ is a line in $\mathbb{P}^{N-1}$. The pre-image $\bar{D}\colon = pr_c|_X^{-1}(D)$ is described as follows:
- if $X$ is $\mathbb{P}^2$ embedded in $\mathbb{P}^5$, then $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(1)|$.
- if $X$ is a cone $C_{N-1}\hookrightarrow \mathbb{P}^N$, then $\bar{D}$ is a sum of two rulings.
- if $X$ is $\Sigma_d\hookrightarrow \mathbb{P}^N$, either $\bar{D}\in |\Delta_0+\Gamma|$, $N=d+3$, or $\bar{D}=\Delta_0$, $N=d+5$.
$pr_c|_X\colon X\to W$ is a normalization map by Cor \[lineIntersecting\].
We show that $D$ is a line in $\mathbb{P}^{N-1}$ by contradiction hypothesis. Assume $D$ is not a line, then there are two points $p,q\in D$ such that the line $L_{p,q}$ passing $p,q$ is not contained in $D$. Let $H_{p,q}\subset \mathbb{P}^N$ be the pre-image of $L_{p,q}$ with respect to $pr_c$. Then $H_{p,q}\cap X$ has length great than $4$ which contradicts Thm \[linearsectionthm\]. Therefore $D$ is a line in $\mathbb{P}^{N-1}$.
Finally, we describe $\bar{D}$. For the case where $X$ is a cone $C_{N-1}$, $\bar{D}$ must be a sum of two rulings. This can be shown by considering the plane $H_{v, L}\subset \mathbb{P}^{N-1}$ containing $v,L$, where $v$ is the vertex of $C_{N-1}$ and $L$ is a secant line of $X$ passing through $c$.
For the case where $X$ is $\mathbb{P}^2$ embedded in $\mathbb{P}^5$, we have $\bar{D}\cdot \mathcal{O}_{X}(1)=2$ since $D$ is a line in $\mathbb{P}^{N-1}$. Hence $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(1)|$.
For the case where $X$ is $\Sigma_d\hookrightarrow \mathbb{P}^N$, similarly as the former case we have $\bar{D}\cdot \mathcal{O}_{X}(1)=\bar{D}\cdot(\Delta_0+\frac{N+d-1}{2}\Gamma)=2$. Hence we have either $\bar{D}\in |\Delta_0+\Gamma|$, $N=d+3$, or $\bar{D}=\Delta_0$, $N=d+5$.
We first consider the cases where $c\in \bar{X}$.
Case I) $\bar{X}$ is $\mathbb{P}^2$ and $\pi^*(K_X+\Lambda)=\mathcal{O}_{\mathbb{P}^2}(1)$. Then $\bar{\Lambda}+\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(4)|$ and $pr_c|_{\bar{X}}$ is a fibration over $\mathbb{P}^1$. By the classification of Gorenstein slc singularities, we see that $\bar{\Lambda}\cdot\bar{D}$ is even. Hence we have either $\bar{\Lambda},\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(2)|$ or $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(4)|$ and $\bar{\Lambda}=0$. Examples will be given later. We note further that $c\not\in\bar{D}$, otherwise it can not be glued into a Gorenstein slc singularity. Case II) $\bar{X}$ is $C_{N-1}$ and $c$ is the vertex. We see that $\bar{\Lambda}$ is two rulings and $\bar{D}\in \mathcal{O}_{C_{N-1}}(2)$.
Case III) $\bar{X}$ is $C_{N-1}$ and $c$ is not the vertex. We show that this does not occur. By Cor \[lineIntersecting\] $pr_c|_{\bar{X}}$ is an isomorphism outside the ruling $l$ passing $c$. Hence $\bar{D}$ must be $l$. However this is impossible as the vertex of $C_{N-1}$ would not be glued into a Gorenstein slc singularity.
Case IV) $\bar{X}$ is $\Sigma_{d}$. We show that this does not occur as well. Similarly as in Case III), we see that $\bar{D}$ should be a ruling $\Gamma$ passing through $c$. Hence $\bar{D}\cdot \bar{\Lambda}=3$, which is impossible.
Case V) $\bar{X}$ is $\mathbb{P}^2$ embedding in $\mathbb{P}^5$. We show that this does not occur. Since by Cor \[lineIntersecting\], we see that $pr_c|_{\bar{X}}$ is an isomorphism outside $c$, which is impossible.
Next we consider the cases where $c\not\in \bar{X}$. We see $pr_c|_{\bar{X}}$ is a morphism of degree at most two by Cor \[lineIntersecting\] and it contracts no curve on $\bar{X}$.
Case VI) $pr_c|_{\bar{X}}$ is a double cover. In this case $\bar{X}$ is a quadric in $\mathbb{P}^3$. Denote $pr_c|_{\bar{X}}$ as $\delta$. The branch curve $B\in |\mathcal{O}_{\mathbb{P}^2}(2)|$. $K_{\bar{X}}+\bar{D}+\bar{\Lambda}=\mathcal{O}_{\bar{X}}(1)=\delta^*\mathcal{O}_{\mathbb{P}^2}(1)$. Hence $\bar{D}+\bar{\Lambda}\in |\delta^*\mathcal{O}_{\mathbb{P}^2}(3)|=|\mathcal{O}_{\bar{X}}(3)|$. $\bar{D}$ is the pre-image of a curve $C$ on $\mathbb{P}^2$ under $pr_c|_{\bar{X}}$. Denote $m$ as the degree of $C$. $m\le 3$. We have the following commutative diagram: $$\xymatrix{
\bar{D}\ar@{^{(}->}[r] &\bar{X}\ar[d]^{\pi}\ar[drr]^{pr_c|_{\bar{X}}}\ar@{^{(}->}[rr]^{\Phi_{|K_{\bar{X}}+\bar{D}+\bar{\Lambda}|}} && \mathbb{P}^3 \ar@{-->}[d]^{pr_c}\\
& X \ar[rr]_{\Phi_{|K_{X}+\Lambda|}} && \mathbb{P}^2
}.$$ We see that $X$ is also a double cover of $\mathbb{P}^2$ induced by $\Phi_{|K_{X}+\Lambda|}$. The branch curve ia $2C+B$. $\Lambda\in |\Phi_{|K_{X}+\Lambda|}^*\mathcal{O}_{\mathbb{P}^2}(3-m)|$.
If $pr_c|_{\bar{X}}$ is birational, it must be a normalisation map. Its image must be $X$ and $\pi=pr_c|_{\bar{X}}$.
Case VII) $\bar{X}$ is $\mathbb{P}^2$ embedding in $\mathbb{P}^5$. The non-normal locus $D$ is a line in $\mathbb{P}^5$ and $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(1)|$. $\bar{\Lambda}\in |\mathcal{O}_{\mathbb{P}^2}(4)|$.
Case VIII) $\bar{X}$ is $C_{N-1}$, $N>3$. $\bar{D}$ is two rulings. $\bar{\Lambda}$ is linearly equivalent to $2H_{\infty}$ where $H_{\infty}$ is a hyperplane.
Case IX) $\bar{X}$ is $\Sigma_{d}$ embedded in $\mathbb{P}^N$ , $N>3$. Either $N=d+3$, $\bar{D}\in |\Delta_0+\Gamma|$ and $\bar{\Lambda}\in |2\Delta_0+(2d+2)\Gamma|$; or $N=d+5$, $\bar{D}=\Delta_0$ and $\bar{\Lambda}\in |2\Delta_0+(2d+4)\Gamma|$.
We summarize the above results in the following theorem:
\[delta1,0\] Let $(X,\Lambda)$ be an irreducible non-normal Gorenstein stable log surface as before. Assume $\Delta(X,K_X+\Lambda)=1$ and $\Delta(\bar{X},\pi^*(K_X+\Lambda))=0$. Then there are the following possibilities:
- $\bar{X}$ is $\mathbb{P}^2$. $\bar{\Lambda}\in |\mathcal{O}_{\mathbb{P}^2}(2)|$ and $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(2)|$. Moreover, $c\not\in \bar{D}$. ($p_g(X,\Lambda)=2$)
- $\bar{X}$ is $\mathbb{P}^2$. $\bar{\Lambda}=0$ and $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(4)|$. Moreover, $c\not\in \bar{D}$. ($p_g(X,\Lambda)=2$)
- $\bar{X}$ is $C_{N-1}$. $\bar{\Lambda}$ is two rulings and $\bar{D}\in \mathcal{O}_{C_{N-1}}(2)$. $c$ is the vertex of $C_{N-1}$. ($p_g(X,\Lambda)=N$)
- $\bar{X}$ is a quadric in $\mathbb{P}^3$. $\bar{D}\in |\mathcal{O}_{\bar{X}}(m)|$ and $\bar{\Lambda} \in |\mathcal{O}_{\bar{X}}(3-m)|$. $X$ is a double cover of $\mathbb{P}^2$ induced by $\Phi_{|K_{X}+\Lambda|}$. The branch curve is $2C+B$, where $C$,$B$ are reduced curves of degree $m$, $2$. $\Lambda\in |\Phi_{|K_{X}+\Lambda|}^*\mathcal{O}_{\mathbb{P}^2}(3-m)|$. $0<m \le 3$. ($p_g(X,\Lambda)=3$)
- $X$ is the projection image of a Veronese embedding of $\mathbb{P}^2$. $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(1)|$ and $\bar{\Lambda}\in |\mathcal{O}_{\mathbb{P}^2}(4)|$. ($p_g(X,\Lambda)=5$)
- $X$ is the projection image of $C_{N-1}\subset\mathbb{P}^N$, $N>3$. $\bar{D}$ is a sum of two lines passing the vertex, and $\bar{\Lambda}\in \mathcal{O}_{C_{N-1}}(2)$. ($p_g(X,\Lambda)=N$)
- $X$ is the projection image of $\Sigma_d$ embedded in $\mathbb{P}^N$, $N>3$. $D\subset X$ is a line. We have either $N=d+3$, $\bar{D}\in |\Delta_0+\Gamma|$ and $\bar{\Lambda}\in |2\Delta_0+(2d+2)\Gamma|$, or $N=d+5$, $\bar{D}=\Delta_0$ and $\bar{\Lambda}\in |2\Delta_0+(2d+4)\Gamma|$. ($p_g(X,\Lambda)=N$)
Let $\bar{D}$ be a smooth quadric in $\mathbb{P}^2$ defined by $x^2+y^2+z^2=0$. $\tau$ acts on $\bar{D}$ by $x\mapsto -x, y \mapsto -y, z\mapsto z$. Let $\bar{\Lambda}=L_{x=0}+L_{y=0}$, where $L_{x=0}$, $L_{y=0}$ are two lines. Gluing $\mathbb{P}^2$ along $\bar{D}$ by $\tau$ we get a non-normal Gorenstein stable log surface $(X,\Lambda)$ with $(K_X+\Lambda)^2=p_g(X,\Lambda)-1=1$.
Let $\bar{D}$ be a smooth quartic in $\mathbb{P}^2$ defined by $x^4+y^4+z^4=0$. $\tau$ acts on $\bar{D}$ by $x\mapsto -x, y \mapsto -y, z\mapsto z$. Gluing $\mathbb{P}^2$ along $\bar{D}$ by $\tau$ we get a non-normal Gorenstein stable surface $X$ with $K_X^2=p_g-1=1$.
\[irrnonnormalstabledelta1\] Let $X$ be an irreducible non-normal Gorenstein stable surface with $\Delta(X,K_X)=1$. Then $X$ is one of the followings:
- $X$ is a double cover of $\mathbb{P}^2$. The branch curve is $2C+B$, where $C$,$B$ are reduced curves of degree $4-k$, $2k$. $k=2,3$. ($p_g(X)=3$)
- $X$ is obtained from a log surface $(\bar{X},\bar{D})$ by gluing the 2-section $\bar{D}$. $(\bar{X},\bar{D})$ is a normal Gorenstein stable log surface as in Thm \[delta-genus-one\] (6). ($p_g(X)=2$)
- $\bar{X}$ is $\mathbb{P}^2$. $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(4)|$. ($p_g(X)=2$)
- $\bar{X}$ is a quadric in $\mathbb{P}^3$. $\bar{D}\in |\mathcal{O}_{\bar{X}}(3)|$. ($p_g(X)=3$)
reducible Gorenstein stable log surfaces with $\Delta(X,K_X+\Lambda)=1$ {#reducible-stable}
=======================================================================
In this section we consider reducible Gorenstein stable log surfaces. They are glued by irreducible ones along some connecting curves.
\[restsecions\] Let $X$ be a connected $S_2$ scheme of pure dimension, $\mathcal{L}$ be an invertible sheaf such that $\dim \mathrm{Bs} |\mathcal{L}|<\dim X-1$ and $C$ be a subscheme of codimension 1. Then $$\begin{aligned}
r_{X\to C}(\mathcal{L})=\dim <\Phi_{|\mathcal{L}|}(C)>+1,\end{aligned}$$ where $<\Phi_{|\mathcal{L}|}(C)>$ is the projective subspace of $|\mathcal{L}|^*$ spanned by $\Phi_{|\mathcal{L}|}(C)$.
Denote $\mathbb{P}:=|\mathcal{L}|^*$. We have the following commutative diagram: $$\xymatrix{
H^0(\mathbb{P},\mathcal{O}_{\mathbb{P}}(1))\ar[rr]^<(0.2){\mathcal{R}_{\mathbb{P}\to \Phi_{|\mathcal{L}|}(C)}}\ar[d]^{\cong}_{\Phi_{|\mathcal{L}|}^*} & & H^0(\Phi_{|\mathcal{L}|}(C),\mathcal{O}_{\mathbb{P}}(1)|_{\Phi_{|\mathcal{L}|}(C)})\ar@{^{(}->}[d]_{\Phi_{|\mathcal{L}|}^*} \\
H^0(X,\mathcal{L})\ar[rr]^{\mathcal{R}_{X\to C}}& & H^0(C,\mathcal{L}|_C)
}.$$ Therefore $r_{X\to C}(\mathcal{L})=r_{\mathbb{P}\to \Phi_{|\mathcal{L}|}(C)}(\mathcal{O}_{\mathbb{P}}(1))=\dim <\Phi_{|\mathcal{L}|}(C)>+1$.
Still we call a curve $C$ on a demi-normal scheme $X$ a *[line]{} if the proper transformation $\bar{C}$ is a line on the normalization $\bar{X}$ of $X$.*
\[restdim\] Let $(X,\Lambda)$ be an irreducible Gorenstein stable log surface with a reduced curve $C$ on it. Then
- if $X$ is normal and $\Delta(X,K_X+\Lambda)=0$, then $r_{X\to C}(K_X+\Lambda)\ge2$. Moreover ’=’ holds if and only if $C$ is a line on $X$.
- if $X$ is normal and $\Delta(X,K_X+\Lambda)=1$, then $r_{X\to C}(K_X+\Lambda)\ge1$.
Moreover, ’=’ holds if and only if $|K_X+\Lambda|$ is composed with a pencil and $C$ is a fiber on $X$.
If $|K_X+\Lambda|$ is not composed with a pencil, then $r_{X\to C}(K_X+\Lambda)\ge2$ and ’=’ holds if and only if $\Phi_{|K_X+\Lambda|}(C)$ is a line in $|K_X+\Lambda|^*$.
- if $X$ is non-normal and $\Delta(X,K_X+\Lambda)=1$, then $r_{X\to C}(K_X+\Lambda)\ge1$. Moreover, ’=’ holds if and only if $|K_X+\Lambda|$ has a base point, and $C$ is a line passing through the base point.
\(i) and (ii) follows from Thm \[gepg-2\], Thm \[delta-genus-one\] and Lemma \[restsecions\]. To prove (iii), we note that $H^0(X,K_X+\Lambda)\cong \pi^* H^0(X,K_X+\Lambda)$, which corresponds to the space of hyperplane sections passing through the base point $c$ of $\pi^* H^0(X,K_X+\Lambda)$ in $ \mathbb{P}:=|\pi^*(K_X+\Lambda)|^*$. Denote $pr_c\colon \mathbb{P}\dashrightarrow \mathbb{P}':=|K_X+\Lambda|^*$ as the projection from the point $c$.
We have the following commutative diagram: $$\xymatrix{
\bar{C}\ar@{^{(}->}[r]\ar[d] & \bar{X}\ar@{^{(}->}[rr]^{\Phi_{|\pi^*(K_X+\Lambda)|}}\ar@{-->}[rrd]^{\Phi'}\ar[d]_{\pi}& & \mathbb{P}\ar@{-->}[d]^{pr_c}\\
C\ar@{^{(}->}[r] &X\ar@{-->}[rr]_{\Phi_{|K_X+\Lambda|}} &&\mathbb{P}',
}$$ where $\Phi'$ is a map defined by the partial linear system $\pi^*H^0(X,K_X+\Lambda)$. We regard $\Phi_{|\pi^*(K_X+\Lambda)|}$ as an inclusion. By Lemma \[restsecions\], $r_{X\to C}(K_X+\Lambda)\ge1$. If ’=’ holds, $\Phi_{|K_X+\Lambda|}(C)$ is a point. Hence $pr_c(\bar{C})$ is a point, which implies $\bar{C}$ is a line passing $c$. Hence $C$ is a line passing through $\pi(c)$, which is the base point of $|K_X+\Lambda|$.
\[2comps\] Let $(X,\Lambda)$ be a connected Gorenstein stable log surface. Assume $X=X_1\cup X_2$, where $X_1$ is connected and $X_2$ is irreducible. Denote $C:=X_1\cap X_2$ as the connecting curve of $X_1$ and $X_2$.
Then:
- $\Delta(X,K_X+\Lambda)\ge \Delta(X_1,(K_X+\Lambda)|_{X_1})$.
- if ’=’ holds and $\Delta(X_2,(K_X+\Lambda)|_{X_2})\le 1$, then
- either $X_2$ is non-normal and $\Delta(X_2,(K_X+\Lambda)|_{X_2})=1$. $r_{X_2\to C}((K_X+\Lambda)|_{X_2})=1$. $C$ is a line on $X_2$ passing through the base point of $|(K_X+\Lambda)|_{X_2}|$. Moreover, $r_{X_1\to C}((K_X+\Lambda)|_{X_1})\le 1$.
- $X_2$ is normal and $\Delta(X_2,(K_X+\Lambda)|_{X_2})=0$, $|(K_X+\Lambda)|_{X_2}|$ is very ample. $r_{X_2\to C}((K_X+\Lambda)|_{X_2})=2$. $C$ is a line on $X_2$. Moreover, $r_{X_1\to C}((K_X+\Lambda)|_{X_1})\le 2$;
- or $X_2$ is normal and $\Delta(X_2,(K_X+\Lambda)|_{X_2})=1$, $|(K_X+\Lambda)|_{X_2}|$ is composed with a pencil of elliptic curves. $r_{X_2\to C}((K_X+\Lambda)|_{X_2})=1$. $C$ is a fiber on $X_2$. Moreover, $r_{X_1\to C}((K_X+\Lambda)|_{X_1})\le 1$.
$(K_X+\Lambda)^2=(K_X+\Lambda)|_{X_1}^2+(K_X+\Lambda)|_{X_2}^2$ together with Cor \[fibersection\] gives $$\label{deltaIneq}
\begin{split}
\Delta(X,K_X+\Lambda) & \ge \Delta(X_1,(K_X+\Lambda)|_{X_1})+\Delta(X_2,(K_X+\Lambda)|_{X_2})\\
&+\max\{r_{X_1\to C}((K_X+\Lambda)|_{X_1}),r_{X_2\to C}((K_X+\Lambda)|_{X_2})\}-2.
\end{split}$$ By Lemma \[restdim\] we see that $\Delta(X_2,(K_X+\Lambda)|_{X_2})+r_{X_2\to C}((K_X+\Lambda)|_{X_2})\ge 2$. Thus $\Delta(X,K_X+\Lambda) \ge \Delta(X_1,(K_X+\Lambda)|_{X_1})$.
If ’=’ holds, then $$\begin{aligned}
\max\{r_{X_1\to C}((K_X+\Lambda)|_{X_1}),r_{X_2\to C}((K_X+\Lambda)|_{X_2})\}= 2- \Delta(X_2,(K_X+\Lambda)|_{X_2}).
\end{aligned}$$ Therefore, by Lemma \[restdim\], $r_{X_2\to C}((K_X+\Lambda)|_{X_2})=2$, if $\Delta(X_2,(K_X+\Lambda)|_{X_2})=0$. $r_{X_2\to C}((K_X+\Lambda)|_{X_2})=1$, if $\Delta(X_2,(K_X+\Lambda)|_{X_2})=1$. Then other statements of (ii) follow from Lemma \[restdim\].
We then have some corollaries.
\[genus-like\] Let $(X,\Lambda)$ be a connected Gorenstein stable log surface. Assume $Y\subset X$ is a connected subsurface. Then we have $\Delta(X,K_X+\Lambda)\ge \Delta(Y,(K_X+\Lambda)|_{Y})$.
If $X_i\subset X$ is an irreducible surface connected to $Y$, $\Delta(Y,(K_X+\Lambda)|_{Y})\le \Delta(Y\cup X_i,(K_X+\Lambda)|_{Y\cup X_i})$ by Lemma \[2comps\]. Then by induction hypothesis, we have $\Delta(Y,(K_X+\Lambda)|_{Y})\le \Delta(X,K_X+\Lambda)$.
Let $(X,\Lambda)$ be a reducible Gorenstein stable log surface. Write $X=\bigcup X_i$. We say that $\Phi:=\Phi_{K_X+\Lambda}$ separates $X_i$ and $X_j$, if $\Phi(X_i\setminus X_i\cap X_j)\cap \Phi(X_j\setminus X_i\cap X_j)=\emptyset$.
\[globalsection\] Let $(X,\Lambda)$ be a connected reducible Gorenstein stable log surface with $X=X_1\cup X_2$ such that $X_1$ is connected and $X_2$ is irreducible. Then
- We have the following commutative diagram: $$\xymatrix{
0 \ar[r] & \ker \mathcal{R}_{X\to X_1} \ar[r]\ar[d]^{\cong} & H^0(X,K_X+\Lambda) \ar[r]^{\mathcal{R}_{X\to X_1}}\ar[d]^{\mathcal{R}_{X\to X_2}} &H^0(X_1,(K_X+\Lambda)|_{X_1})\ar[d]^{\mathcal{R}_{X_1\to X_1\cap X_2}}\\
0 \ar[r] & \ker\mathcal{R}_{X_2\to X_1\cap X_2} \ar[r] & H^0(X_2,(K_X+\Lambda)|_{X_2}) \ar[r]^{\mathcal{R}_{X_2\to X_1\cap X_2}} &H^0(X_1\cap X_2,(K_X+\Lambda)|_{X_1\cap X_2})
.
}$$
- If $\mathrm{im}\, \mathcal{R}_{X_1\to X_1\cap X_2}\subset \mathrm{im} \, \mathcal{R}_{X_2\to X_1\cap X_2}$, then $\mathcal{R}_{X\to X_1}$ is surjective.
- If $(K_X+\Lambda)|_{X_2}$ is very ample and $X_1\cap X_2$ is a line on $X_2$, then $\mathcal{R}_{X\to X_1}$ is surjective and $\Phi_{|K_X+\Lambda)|}$ separates $X_1$ and $X_2$.
- If each $(K_X+\Lambda)|_{X_i}$ is very ample and $X_1\cap X_2$ is a line on each $X_i$, then $K_X+\Lambda$ is very ample.
\(i) and (ii) is obvious by chasing the diagram.
For (iii), $X_1\cap X_2$ is a line implies $\mathcal{R}_{X_2\to X_1\cap X_2}$ is surjective. Then any section in $H^0(X_1,(K_X+\Lambda)|_{X_1})$ can be extended into a section in $H^0(X,K_X+\Lambda)$. Therefore $\mathcal{R}_{X\to X_1}$ is surjective. Next $\ker \mathcal{R}_{X_2\to X_1\cap X_2}$ is nontrivial and its base part is the line $X_1\cap X_2$. Then $\ker \mathcal{R}_{X\to X_1}$ is nontrivial and its base part is $X_1$. Therefore $\Phi_{|K_X+\Lambda)|}$ separates $X_1$ and $X_2$.
For (iv) we see that $\mathcal{R}_{X\to X_i}$ is surjective and the kernel is nontrivial by (ii). Then we have plenty of sections to separate points and tangents, which implies $K_X+\Lambda$ is very ample.
\[equaldelta\] Let $(X,\Lambda)$ be a log surface as in Lemma \[2comps\]. We assume further $X_1$ is irreducible and $\Delta(X,K_X+\Lambda)= \Delta(X_i,(K_X+\Lambda)|_{X_i})\le 1$ for $i=1,2$. Then
- if $\Delta(X,K_X+\Lambda)=0$, then $X_1\cap X_2$ is a line on each $X_i$. Moreover, $K_X+\Lambda$ is very ample.
- if $\Delta(X,K_X+\Lambda)=1$, then
- either $X_1$ and $X_2$ are both normal. Each $|(K_X+\Lambda)|_{X_i}|$ is composed with a pencil. $X_1\cap X_2$ is a fiber on each $X_i$. $|K_X+\Lambda|$ is composed with a pencil as well.
- or $X_1$ and $X_2$ are both non-normal. The base points of $|(K_X+\Lambda)|_{X_i}|$ coincide into the unique base point of $|K_X+\Lambda|$. $X_1\cap X_2$ is a line passing through the base point of $|(K_X+\Lambda)|_{X_i}|$ on each $X_i$.
\(i) follows from Lemma \[2comps\] and Lemma \[globalsection\].
For (ii), applying Lemma \[2comps\] we see that either $X_i$ is non-normal or $(K_X+\Lambda)|_{X_i}$ is composed with a pencil. It is easy to see that either $X_1$, $X_2$ are both non-normal or $(K_X+\Lambda)|_{X_1}$, $(K_X+\Lambda)|_{X_2}$ are both composed with a pencil of elliptic curve, since the geometric genus of $X_1\cap X_2$ on $X_1$ or $X_2$ should coincide.
If $(K_X+\Lambda)|_{X_1}$, $(K_X+\Lambda)|_{X_2}$ are both composed with a pencil of elliptic curve, then by Lemma \[2comps\] each $(K_X+\Lambda)|_{X_i}$ is composed with a pencil and $X_1\cap X_2$ is a fiber on each $X_i$. Therefore $|K_X+\Lambda|$ is composed with a pencil as well. If $X_1$ and $X_2$ are both non-normal, then by Lemma \[restdim\], $X_1\cap X_2$ is a line passing through the base point of $|(K_X+\Lambda)|_{X_i}|$ on each $X_i$. These two base points must coincide, otherwise no nontrivial section in $H^0(X_i,(K_X+\Lambda)|_{X_i})$ can be glued into a section of $H^0(X,K_X+\Lambda)$.
\[decrease1\] Let $(X,\Lambda)$ be a connected Gorenstein stable log surface which has two irreducible components $X_1$, $X_2$. Assume further $\Delta(X_1,(K_X+\Lambda)|_{X_1})=\Delta(X_2,(K_X+\Lambda)|_{X_2})=0$ and $\Delta(X,K_X+\Lambda)=1$. Then $$r_{X_1\to X_1\cap X_2}((K_X+\Lambda)|_{X_1})=r_{X_2\to X_1\cap X_2}((K_X+\Lambda)|_{X_2})=3.$$
By (\[deltaIneq\]), we have $\max\{r_{X_1\to X_1\cap X_2}((K_X+\Lambda)|_{X_1}),r_{X_2\to X_1\cap X_2}((K_X+\Lambda)|_{X_2}\}\le 3$. Moreover $r_{X_i\to X_1\cap X_2}((K_X+\Lambda)|_{X_i}\ge 2$ by Lemma \[restdim\]. Thus there must be one $r_{X_i\to X_1\cap X_2}((K_X+\Lambda)|_{X_i}=3$.
Now we suppose $r_{X_1}(X_1\cap X_2)=2$ and $r_{X_2}(X_1\cap X_2)=3$ for a contradiction. Then $X_1\cap X_2$ is a line on $X_1$ and not a line on $X_2$. Thus nonzero elements of $\mathrm{im}\mathrm{Res}_{X_i|X_1\cap X_2}$ have different degrees. Hence only those sections in $H^0(X_i,(K_X+\Lambda)|_{X_i})$ vanishing on $X_1\cap X_2$ can be glued together. Therefore $$\begin{aligned}
p_g(X,\Lambda)&=\dim\ker \mathcal{R}_{X_1\to X_1\cap X_2}+\dim \ker \mathcal{R}_{X_2\to X_1\cap X_2}\\
&\le p_g(X_1,(K_X+\Lambda)|_{X_1})-2+p_g(X_2,(K_X+\Lambda)|_{X_2})-2\\
&\le (K_X+\Lambda)^2,
\end{aligned}$$ a contradiction. This completes the proof.
\[diffdelta\] Let $(X,\Lambda)$ be a connected Gorenstein stable log surface. Assume $X=X_1\cup X_2$, where $X_i$ is irreducible. Assume further $\Delta(X,K_X+\Lambda)= \Delta(X_1,(K_X+\Lambda)|_{X_1})=1$ and $\Delta(X_2,(K_X+\Lambda)|_{X_2})=0$.
Then
- either each $(K_X+\Lambda)|_{X_i}$ is very ample and on each $X_i$ the curve $X_1\cap X_2$ is a line on each $X_i$. $K_X+\Lambda$ is very ample;
- or $X_1$ is non-normal. $X_1\cap X_2$ is a line on each $X_i$.
First by Lemma \[2comps\], $X_1\cap X_2$ will be a line on $X_2$ and $r_{X_1\to X_1\cap X_2}((K_X+\Lambda)|_{X_1})=1$, or $2$.
For the case $X_1$ is normal, we claim that $r_{X_1\to X_1\cap X_2}((K_X+\Lambda)|_{X_1})\not =1$. Otherwise, by Lemma \[restdim\], $|(K_X+\Lambda)|_{X_1}|$ is composed with a pencil of elliptic curves and the connecting curve $X_1\cap X_2$ will be an elliptic fiber on $X_1$. While on $X_1$ the connecting curve $X_1\cap X_2$ has geometric genus 0. This is impossible.
Therefore, $r_{X_1\to X_1\cap X_2}((K_X+\Lambda)|_{X_1})=2$. $(K_X+\Lambda)|_{X_1}$ is not composed with a pencil and the log canonical image of $X_1\cap X_2$ is a line. Next we show that $X_1$ is not a double cover of $\mathbb{P}^2$. Otherwise, on $X_1$, $X_1\cap X_2$ is a double covering curve of $\mathbb{P}^1$. Then nonzero sections in the image of $\mathcal{R}_{X_i\to X_1\cap X_2}$ have different degrees. Hence only those sections of $H^0(X_i, (K_X+\Lambda)|_{X_i})$ vanishing on $X_1\cap X_2$ can be glued together. Therefore, $$\begin{aligned}
p_g(X,\Lambda)&=\dim\ker \mathcal{R}_{X_1\to X_1\cap X_2}+\dim \ker \mathcal{R}_{X_2\to X_1\cap X_2}\\
&=p_g(X_1,(K_X+\Lambda)|_{X_1})+p_g(X_2,(K_X+\Lambda)|_{X_2})-4\\
&< (K_X+\Lambda)^2,
\end{aligned}$$ a contradiction. Therefore $(K_X+\Lambda)|_{X_1}$ is very ample. $X_1\cap X_2$ is a line on $X_1$. The other statements follow.
For the non-normal case, we take a partial normalization $\nu\colon \tilde{X}\to X$ along the non-normal locus on $X_1$. $\tilde{X}=\tilde{X_1}\cup X_2$, where $\tilde{X_1}$ is the normalization of $X_1$. It is easy to check that $\Delta(\tilde{X},\nu^*(K_X+\Lambda))=0$. Then we see that $X_1\cap X_2$ is a line on each $X_i$. Hence we complete the proof.
Let $X_1$ be a $\mathbb{P}^2$ blown up at $k$ distinct points as in Thm \[delta-genus-one\] (3). Each $E_i$ is a line w.r.t $-K_{X_1}$. Choosing a general $\Lambda_1=E_1+...+E_k+B\in |-2K_{X_1}|$, then $\Lambda_1$ is nodal and $E_i$ intersects $B$ at 3 distinct points. Let $(X_2,\Lambda_2)$ be a log surface as in Thm \[gepg-2\](iii) such that $\Lambda_2=\Gamma_1+...+\Gamma_s+D$, where $\Gamma_i$ is a ruling and it intersect $D$ at 3 distinct points. Then we can glue $X_1$, $X_2$ along $E_i$, $\Gamma_j$ to obtain a log surface as in Thm \[diffdelta\] (i).
The following theorem and corollary is first obtained in [@LR13].
\[thm: log noether\] Let $(X,\Lambda)$ be a connected Gorenstein stable log surface. Then $$\Delta(X,K_X+\Lambda)\ge 0.$$
This follows directly by Cor \[gepg-2\], Lemma \[nonnorm&irred\] and Lemma \[2comps\].
\[cor: nonnormal equality\] Let $(X, \Lambda)$ be a Gorenstein stable log surface such that $\Delta(X,K_X+\Lambda)=0$. Write $X=\bigcup X_i$, where $X_i$ is an irreducible component.
Then
1. $\Delta(X_i, (K_X+\Lambda)|_{X_i})=0$.
2. $K_X+\Lambda$ is very ample。 It defines an embedding $\phi\colon X\hookrightarrow \mathbb{P}=|K_X+\Lambda|^*$;
3. if $X_{i}\cap X_{j}\not = \emptyset $, then $X_{i}\cap X_{j}$ is a line on both $X_i$ and $X_j$;
4. $X$ is a tree of $X_i$ glued along lines.
In particular, $\Lambda\neq 0$.
Finally we are able to classify Gorenstein stable log surface with $\Delta(X,K_X+\Lambda)=1$.
\[nonnormdelta-1\] Let $(X,\Lambda)$ be a reducible Gorenstein stable log surface with $\Delta(X,K_X+\Lambda)=1$. Write $X=\bigcup X_i$, where $X_i$ is an irreducible component.
Then $X$ has a unique minimal connected component $U$ such that $\Delta(U,(K_X+\Lambda)|_{U})=1$, $X\setminus U$ is composed with several trees $T_j$ with $\Delta(T_j,(K_X+\Lambda)|_{T_j})=0$ and $X$ is glued by $U$ and $T_j$ along lines.
$U$ is one of the followings:
- $X=U$ is a string of log surfaces $X_i$ with $\Delta(X_i,(K_X+\Lambda)|_{X_i})=1$ and $|(K_{X}+\Lambda)|_{X_i}|$ composed with a pencil of elliptic curves. The connecting curves are all fibers. Moreover, in this case $|K_X+\Lambda|$ is composed with a pencil of elliptic curves.
- $U$ is a string of surfaces glued along lines whose end surfaces $X_i$ of the string $U$ are non-normal with $\Delta(X_i,(K_X+\Lambda)|_{X_i})=1$.
Moreover, if $U$ is reducible, $|K_X+\Lambda|$ has a base point $c\in X$ and all the connecting curves of $U$ pass through $c$.
- $U$ is composed with a single irreducible log surface $X_k$ with $\Delta(X_k,(K_X+\Lambda)|_{X_k})=1$ and $|(K_{X}+\Lambda)|_{X_k}|$ very ample. Moreover, in this case $K_X+\Lambda$ is very ample.
- $U=X_j\cup X_k $ where $X_j$, $X_k$ are two Gorenstein log surfaces with $\Delta(X_j,(K_X+\Lambda)|_{X_j})=\Delta(X_k,(K_X+\Lambda)|_{X_k})=0$. All the connecting curves of $X$ are lines except $X_j\cap X_k$.
- $U$ is a cycle of log surface $X_i$ with $\Delta(X_i,(K_X+\Lambda)|_{X_i})=0$. All the connecting curves of $X$ are lines.
First by Cor \[genus-like\] we see each component $X_i$ has $\Delta(X_i,(K_X+\Lambda)|_{X_i})=1$ or $0$. Second, once we obtain a connected component $U\subset X$ such that $\Delta(U,(K_X+\Lambda)|_{U})=1$ and $\Delta(X_i,(K_X+\Lambda)|_{X_i})=0$ for each $X_i\subset X\setminus U$, we see that $X\setminus U$ will be composed with several trees of log surfaces $T_j$ with $\Delta(T_j,(K_X+\Lambda)|_{T_j})=0$ by Lemma \[2comps\] and induction hypothesis. Therefore it remains to describe $U$.
We distinguish between two cases whether there is an irreducible component $X_i$ such that $\Delta(X_i,(K_X+\Lambda)|_{X_i})=1$.
Case 1. There is a component $X_i$ such that $\Delta(X_i,(K_X+\Lambda)|_{X_i})=1$. Then by Cor \[equaldelta\] and Cor \[diffdelta\], either $(K_X+\Lambda)|_{X_i}$ is composed with a pencil of elliptic curves, $X_i$ is non-normal, or $(K_X+\Lambda)|_{X_i}$ is very ample.
If $(K_X+\Lambda)|_{X_i}$ is composed with a pencil of elliptic curves, then by Cor \[equaldelta\] and Cor \[diffdelta\], for each irreducible components $X_j$ connected to $X_i$, $(K_X+\Lambda)|_{X_j}$ is composed with a pencil of elliptic curves as well, $X_i\cap X_j$ is a fiber and $\Delta(X_j,(K_X+\Lambda)|_{X_j})=1$. Therefore, inductively, every $(K_X+\Lambda)|_{X_k}$ is composed with a pencil of elliptic curves. Each $X_i$ is connected to at most two other components, as there are at most two fibers as connecting curves which pass through the base point of $|(K_X+\Lambda)|_{X_i}|$ on $X_i$. Thus $X=U$ is a string of such surfaces glued along fibers.
If $(K_X+\Lambda)|_{X_i}$ is very ample. Then by Cor \[equaldelta\] and Cor \[diffdelta\], every other irreducible component $X_j$ connected to $X_i$ has $\Delta(X_j,(K_X+\Lambda)|_{X_j})=0$ and the connecting curves of them are lines. It is easy to check $U=X_i$. Moreover, $K_X+\Lambda$ is very ample by Lemma \[globalsection\].
If $X_i$ is non-normal, then by Lemma \[2comps\] other component $X_j$ is either non-normal or has $\Delta(X_j,(K_X+\Lambda)|_{X_j})=0$ and $X$ is a tree of such surfaces. Let $\nu\colon \tilde{X}\to X$ be a partial normalization along the non-normal curves on $X_j$. It is easy to check that $\Delta(\bar{X}, \nu^*(K_X+\Lambda))=0$. $\nu^*H^0(X,K_X+\Lambda)\subset H^0(\bar{X}, \nu^*(K_X+\Lambda))$ has a base point $c\in \mathbb{P}:=|\nu^*(K_X+\Lambda)|^*$. We see that $c$ is also the base point of $|(K_X+\Lambda)|_{X_j}|$ for any non-normal $X_j$. If there is only one non-normal surface $X_i$, then $U=X_i$. If there are two non-normal surfaces $X_i$, $X_j$, then $U$ is the unique string of surfaces contained in $X$ with $X_i$, $X_j$ as the ends. If there are more than two non-normal surfaces, we claim that they lies on a unique minimal string of surfaces, which is $U$. Otherwise there are three non-normal surfaces contained in a fork of surfaces $Y\subset X$. By Lemma \[2comps\], we see that the base point $c$ lies on each $X_k\subset Y$. However, in the central surface of $Y$, there are three connecting curve passing through $c$, which is impossible. Hence the claim is true. Case 2. Each irreducible component $X_i$ has $\Delta(X_i,(K_X+\Lambda)|_{X_i})=0$.
If there are two irreducible components $X_j$, $X_{k}$ such that $\Delta(X_j\cup X_k,(K_X+\Lambda)|_{X_j\cup X_k})=1$, then by Lemma \[decrease1\], $X_j\cap X_k$ is neither a line on $X_j$ nor $X_k$, and $r_{X_j\to X_j\cap X_k}((K_X+\Lambda)|_{X_j})=r_{X_k\to X_j\cap X_k}((K_X+\Lambda)|_{X_k})=3$. It is easy to see that $X$ is a tree of log surfaces. Ungluing $X$ along $X_j\cap X_{k}$, there will be two connected tree of surfaces $V_1\supset X_j$, $V_2\supset X_{k}$. We claim $\Delta(V_1,(K_X+\Lambda)|_{V_1})=\Delta(V_2,(K_X+\Lambda)|_{V_2})=0$. Otherwise assume $\Delta(V_1,(K_X+\Lambda)|_{V_1})=1$, then $\Delta(V_1\cup X_{k},(K_X+\Lambda)|_{V_1\cup X_{k}})=1$ implies $ V_1\cap X_{k}=X_j\cup X_{k}$ is a line on $X_k$, a contradiction. Hence $U=X_j\cap X_k$ and $X$ is a tree of log surfaces whose connecting curves are lines except $X_j\cap X_k$.
Finally we consider the case that all the connecting curve are lines. It is easy to see that $X$ is not a tree, otherwise $\Delta(X,K_X+\Lambda)=0$. Hence there is a minimal cycle of surfaces $U$ of log surfaces such that $\Delta(U,(K_X+\Lambda)|_U)=1$.
Let $X$ be a connected reducible Gorenstein stable surface with $K_X^2= p_g-1$. Then $X$ is a union of two $\mathbb{P}^2$ glued along a quartic curve. $X$ is a double cover of $\mathbb{P}^2$ and it can be deformed into a smooth one. ($K_X^2=2$)
If $X\setminus U$ is not empty, then there is a leaf log surface $X_i$ such that $\Delta(X_i,K_X|_{X_i})=0$. While $K_X|_{X_i}=K_{X_i}+\bar{D_i}$ and the connecting curve $\bar{D_i}$ is a line, which is impossible. Hence $X=U$. A similar discussion tells us that $X$ is not a string of $X_i$ such that $|K_X|_{X_i}|$ is composed with a pencil. Therefore $X$ is either a string of surfaces containing non-normal ones or a union of two irreducible components whose connecting curve is not a line. We show that the first case does not occur. Otherwise $X=U$ is a string of surfaces glued along lines passing through $c\in X$ where $c$ is the base point of $|K_X|$. However, in this case $c$ would not be a Gorenstein slc singularity, a contradiction.
Let $X=X_1\cup X_2$ with $\Delta(X_i, K_X|_{X_i})=0$ and $r_{X_i\to X_1\cap X_2}(K_X|_{X_i})=3$. Then $X_i$ is $\mathbb{P}^2$, $\Sigma_d$ or $C_{n-1}$. While in these cases $\ker \mathcal{R}_{X_i\to X_1\cap X_2}=0$ (otherwise $K_X|_{X_i}=K_{X_i}+X_1\cap X_2\ge X_1\cap X_2$ as divisors, which is impossible), thus $h^0(X_i,K_X|_{X_i})=r_{X_i\to X_1\cap X_2}(K_X|_{X_i})=3$. Hence $X_i$ is $\mathbb{P}^2$, and the connecting curve $X_1\cap X_2$ is a curve of degree 4.
Let $X$ be a connected Gorenstein stable surface with $K_X^2= p_g-1$.
If $X$ is irreducible, then it is one of the followings:
- $X$ is a double cover of $\mathbb{P}^2$. The branch curve is $2C+B$, where $C$,$B$ are reduced curves of degree $4-k$, $2k$. $k=2,3$. ($p_g(X)=3$)
- $X$ is obtained from a log surface $(\bar{X},\bar{D})$ by gluing the 2-section $\bar{D}$. $(\bar{X},\bar{D})$ is a normal Gorenstein stable log surface as in Thm \[delta-genus-one\] (6). ($p_g(X)=2$)
- $\bar{X}$ is $\mathbb{P}^2$. $\bar{D}\in |\mathcal{O}_{\mathbb{P}^2}(4)|$. ($p_g(X)=2$)
- $\bar{X}$ is a quadric in $\mathbb{P}^3$. $\bar{D}\in |\mathcal{O}_{\bar{X}}(3)|$. ($p_g(X)=3$)
If it is reducible, then $X$ is a union of two $\mathbb{P}^2$ glued along a quartic curve. ($p_g(X)=3$)
We have confirmed the conjecture in [@LR13] that there are no Gorenstein stable surfaces with $K_X^2= p_g-1\ge 3$.
[plain]{}
Valery Alexeev. Moduli spaces [$M\sb{g,n}(W)$]{} for surfaces. In [*Higher-dimensional complex varieties (Trento, 1994)*]{}, pages 1–22. de Gruyter, August 1996.
Valery Alexeev. Higher-dimensional analogues of stable curves. In [*International [C]{}ongress of [M]{}athematicians. [V]{}ol. [II]{}*]{}, pages 515–536. Eur. Math. Soc., Z[ü]{}rich, 2006.
Wolf P. Barth, Klaus Hulek, Chris A. M. Peters, and Antonius [Van de Ven]{}. , volume 4 of [*Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge.*]{} Springer-Verlag, Berlin, second edition, 2004.
Marieamelie Bertin. On singular varieties having an extremal secant line\[J\]. , 2006, 34(3): 893-909.
Jingshan Chen. Normal Gorenstein stable log surfaces with $(K_X+\Lambda)^2 = p_g(X,\Lambda)- 1$, [*Communications in Algebra*]{} (2018), DOI: 10.1080/00927872.2018.1435794
Marco Franciosi, Rita Pardini, and Sönke Rollenske. Gorenstein stable surfaces with $K_X^2 = 1$ and $p_g>0$. (2015).
Marco Franciosi, Rita Pardini, and Sönke Rollenske. Log-canonical pairs and Gorenstein stable surfaces with $K_X^2=1$. 151.8(2015):1529-1542.
Takao Fujita. On the structure of polarized varieties with $\Delta$-genera zero. J.fac.sci.univ.tokyo Sect.ia Math 22(1975):103-115.
Takao Fujita. Classification theories of polarized varieties. Cambridge University Press, 1990.
La Nave, Gabriele. Explicit stable models of elliptic surfaces with sections. (2002).
J[á]{}nos Koll[á]{}r and Shigefumi [M]{}ori. , volume 134 of [ *Cambridge Tracts in Mathematics*]{}. Cambridge University Press, Cambridge, 1998. With the collaboration of C. H. Clemens and A. Corti, Translated from the 1998 Japanese original.
Janós Kollár. Moduli of varieties of general type. In G. Farkas and I. Morrison, editors, [*Handbook of Moduli: Volume II*]{}, volume 24 of [*Advanced Lectures in Mathematics*]{}, pages 131–158. International Press, 2012.
J[á]{}nos Koll[á]{}r. , volume 200 of [ *Cambridge Tracts in Mathematics*]{}. Cambridge University Press, Cambridge, 2013. With a collaboration of S[á]{}ndor Kov[á]{}cs.
J[á]{}nos Koll[á]{}r. . 2014. book in preparation.
J[á]{}nos Koll[á]{}r and Nick Shepherd-Barron. Threefolds and deformations of surface singularities. , 91(2):299–338, 1988.
Janós Kollár, K. E. Smith, and A. Corti. Rational and nearly rational varieties. Cambridge University Press Cambridge (2004):vi,235.
Henry B. Laufer. On minimally elliptic singularities. , 99(6):1257–1295, 1977.
Wenfei Liu and S[ö]{}nke Rollenske. Geography of Gorenstein stable log surfaces\[J\]. , 2013, 368(4).
Masayoshi Nagata. *On rational surfaces I*. Memoirs of the College of Science University of Kyoto, 33,(1960):351-370.
Miles Reid. Chapters on algebraic surfaces. In [*Complex algebraic geometry*]{}, volume 3 of [*IAS/Park City Math. Ser.*]{}, pages 3–159. Amer. Math. Soc., Providence, RI, 1997.
Fumio Sakai. Semistable curves on algebraic surfaces and logarithmic pluricanonical maps. , 254(2):89–120, 1980.
Shuichiro Tsunoda and De-Qi Zhang. Noether’s inequality for noncomplete algebraic surfaces of general type. , 28(1):21–38, 1992.
[^1]: A local model for the pinch point in $\mathbb{A}^3$ is given by the equation $x^2+yz^2=0$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work, we present a soft capsule for mucus collection in human intestine for diagnostic purpose, with reduced risk of tissue damage compared with other biopsy methods. The capsule implements passive locomotion and the sampling process is triggered by magnetic field using a permanent magnet, which is placed on the skin above the region of interest (ROI). The capsule contains a soft vacuum chamber which is sealed with wax. When magnetic field and mucus are present simultaneously, the circuit inside the capsule will be closed and nichrome wire will start to generate heat to melt the wax, and mucus will be collected into the vacuum chamber due to air pressure. Experiments on heating capability of nichrome wire, mucus collection and reliability of the capsule are conducted to validate this design.'
author:
- 'Xingzhou Du$^{1, 2}$, Kai Fung Chan$^{1, 2}$, Xianfeng Xia$^{2}$, Philip Wai Yan Chiu$^{2, 3, 5}$ and Li Zhang$^{2, 4, 5}$[^1][^2][^3][^4][^5] [^6]'
title: '**A Magnetically-Triggered Soft Capsule for On-Demand Mucus Collection** '
---
Introduction
============
Capsule endoscope is an evolutionary tool for diagnosing diseases in gastrointestinal (GI) tract. Since the first capsule endoscope made by Given Imaging Inc. was approved by FDA in the United States in 2003, millions of patients have benefited from this comfortable endoscopic procedure [@c1]. The clinical application of capsule endoscopes makes it possible to inspect the entire small intestinal areas that was hard to reach using traditional endoscopes and provide painless methods for diagnosing GI tract diseases such as obscure gastrointestinal bleeding (OGIB), polyps, Celiac disease, and Crohn’s disease [@c2], [@c31]. However, there are still limitations for commercialized capsule endoscopes that many practical functions such as drug delivery and biopsy cannot be completed [@c30].
Among the years, many researchers have made various progress in the design of biopsy tools used for capsule endoscopes. In 2013, a biopsy module using magnetic torsion spring mechanism and cylindrical blade was invented by M. Simi et al. [@c4], but the large volume restricted its application in small-sized capsule endoscopes. Then in 2015, V. Le, et al. improved the design and invented a biopsy module using biopsy razer which was actuated by shape memory alloy (SMA) and a pair of small permanent magnet [@c5], and in 2016, V. Le, et al. introduced a compact biopsy structure that based on SMA and cylindrical blade [@c6], however, for these two designs, possible damage to healthy tissue might be caused by inaccurate control of the capsule locomotion. The design has also been improved in 2017, a gripper tool actuated by external magnetic field was developed by V. Le, et al. and the camera was allocated on the lateral side of the capsule, in the same direction with the gripper [@c7]. However, the control method of this capsule will be complicated since two magnets with orthogonal directions were implemented. In 2017, based on the magnetically actuated soft capsule endoscopes (MASCE) design [@c8], a fine needle biopsy tool was invented by D. Son, et al. [@c9], but as the capsule was soft, the needle might stretch out by accident due to the movement of stomach, which is dangerous to patients.
![(a) The conceptual working scenario of the soft capsule for on-demand mucus collection. It consists of a permanent magnet that sticks on patient’s skin and a soft capsule for mucus collection. (b) Schematics of all the components inside the capsule. []{data-label="scenerio"}](Picture1_4.png)
Here we consider implementing biopsy on intestinal mucus instead of tissue to prevent potential damage of the epithelium. Intestinal mucus has significant functions in maintaining normal functions of the digestive system, including receiving pancreatic enzymes (digestion), absorption of nutrients, and protecting epithelium from pathogenic bacteria [@c10]. Sampling and analyzing mucus is useful in diagnosing diseases such as malfunction of digestive glands, bacterial or viral infections (e.g. small intestinal bacterial overgrowth), intestinal parasite infections (e.g. ascaris, amebic dysentery, etc.), inflammatory bowel disease (e.g. Crohn’s disease), and presence of pathogenic bacteria (e.g. Helicobacter pylori) [@c11]-[@c15]. Compared with traditional clinical methods, such as fine-needle aspiration that is uncomfortable and possible to damage tissue, aspiration test that is inaccurate, stool sample test that is difficult to separate component, and trial of treatment that may make patients miss optimal therapeutic period, mucus collection using capsules have various advantages including less tissue damage, increased accuracy and decreased discomfort. Besides, capsules have the capability to reach deeper inside small intestine and collect mucus in required location.
In this paper, we present a soft capsule used for on-demand mucus collection. The capsule consists of a soft vacuum chamber, a reed switch, a nichrome wire, a camera module and a power source. Magnetic field is chosen as a trigger signal for its wide application in miniaturized devices, ranging from millimeter scale [@c8], [@c16] all the way down to nanometer scale [@c25]-[@c27]. The capsule implements passive locomotion that is propelled by natural peristaltic constrains of GI tract. An external permanent magnet attached to the skin above the ROI to mark the sampling position, and when the capsule reaches the location, the magnetic field will close the reed switch and start the mucus collection process. The capsule will then carry the sample outside human body. Further examinations on bacteria, virus, or chemical components could then be conducted based on the sampled mucus.
Design and Operation
====================
Capsule Design
--------------
The working scenario of the mucus collecting capsule and the main components are depicted as Fig. 1. A permanent magnet is attached to patient’s skin above the sampling location in small intestine, to trigger the capsule when it moves to the region of interest (ROI) passively. The capsule contains a soft chamber, which could be pressed to generate a proximate vacuum environment inside. A tube is designed to connect the chamber to outside environment and could be tightly blocked with wax after the chamber is loaded by vacuum. Around the tube twines a nichrome wire that is used to generate heat and melt the wax. A reed switch is used to sense the magnetic field of specific strength. A camera module and a power source are integrated in the capsule for viewing and provide power, respectively. The nichrome wire, reed switch and power source make up a heat generation circuit, where nichrome wire transfers electricity to heat and reed switch is in charge of close the circuit under magnetic field.
![(a) Crane-like structure is designed to prevent capsule sticking to tissue when collecting mucus, and electrodes are used to detect mucus outside the tube, and (b) structure inside tube is designed to prevent leakage and protect collected sample.[]{data-label="detail"}](Picture1_5.png)
![(a) Before triggered by magnetic field, the tube is sealed with wax to keep the vacuum environment inside soft chamber, the heating circuit is not working. (b) When magnetic field and mucus present at the same time, the circuit is closed and nichrome wire melts the wax. Then the mucus will be collected due to air pressure.[]{data-label="Mechanism"}](Picture2.png)
Some detailed features are designed to improve the performance of the capsule. In Fig. 2(a), a crane-like structure is implemented at the end of the tube, to prevent the capsule from sticking to tissue due to the air pressure generated by vacuum chamber when collecting mucus. Pair of electrodes is also used, as a part of the circuit (see Fig. 3(a)), to make sure that the circuit will close only when mucus exist around the tube. What’s more, as in Fig. 2(b) a valve-like structure is also designed to prevent leakage and protect the collected sample.
Operation Mechanism
-------------------
The operation mechanism of the mucus collection capsule could be described as three main steps. First, before swallowed by patient, the soft chamber will be pressed, and the tube should be sealed with wax, to create and maintain a vacuum environment inside the chamber, as shown in Fig. 3(a). Besides, a permanent magnet will be stuck to the skin of the patient, proximally above the ROI. This magnet only serves as a position marker, in a way of generating strong enough magnetic field within the ROI for triggering the mucus collection process, and have no influence on the capsule’s locomotion.
The second step is depicted as Fig. 2(b). The capsule is swallowed by the patient and implements an passive locomotion method propelled by the natural movement of GI tract. After the capsule reaches the desired location, the reed switch will be closed by the magnetic field which is generated by the external magnet on patient’s skin, and when mucus exists around the inlet of the chamber at the same time, the circuit will be connected and nichrome wire will start to generate heat to melt the wax. The wax is amorphous material and will melt at around 60 $^\circ$C. The mucus will then be sucked inside the chamber due to the air pressure generated by the chamber’s tendency to expand to its original shape. The generated heat of nichrome wire will be isolated inside the capsule’s soft encapsulate, which is designed to be made of materials with low heat conductivity, and therefore the intestinal tissue will not be damaged by the heat. If cooperate with camera, the condition of sampling site could also be recorded for analysis.
For the third step, the collected mucus will be protected by capsule and excreted by patient. Further examinations on virus, bacteria or chemical content within the collected mucus could be conducted in laboratory.
Fabrication
-----------
![(a) The inner structure of the mucus collection capsule’s fabricated prototype for function validation. (b) Encapsulated prototype with a ruler. []{data-label="Prototype"}](Picture5_1.png)
The prototype of the soft capsule for function demonstration is shown as Fig. 4(a) and (b). The dimension of the capsule is 14 mm in diameter and 20 mm in length (without camera module and power supply). The soft chamber and encapsulate are fabricated using stereolithography (SLA) 3D printing, with photosensitive silicone as raw material, which has promising heat-resistivity of 70-80 $^\circ$C and satisfying resilience and strength. The hardness of the material is adjustable from 30 to 90 in type A Shore hardness [@c23], and in this design, we choose hardness of 70 to keep a balance between the easiness of deformation and the suction ability. The tube and the chamber are designed as a unitary part to ensure the tightness of the capsule. A hole is opened at the back of the chamber to make it convenient to clean after experiment, and the hole is sealed with reusable adhesive (BluTack, Bostik Inc.) when the capsule is loaded. The encapsulate is designed to cover outside the capsule after all the other parts are assembled.
As for nichrome wire, an alloy with contents of 20% Ni and 80% Cr is used, and heat capability of different wires with different lengths and diameters are tested in the experiment. A wire of 40 mm in length and 0.25 mm in diameter is implemented for its proper heat and resistance, and twined into a spiral shape to increase the contact area with the tube. A reed switch (ORD213, Standex-Meder Electronics) with size of 1.8 mm in diameter and 7 mm in length is connected to the nichrome wire. The camera module and the electrodes are not implemented in this prototype and a 1.5 V DC power supply is used as energy source.
Experimental Results and Discussion
===================================
Heating Capability of Nichrome Wire
-----------------------------------
![The experimental result for heating capability of nichrome wire. With length of nichrome wires unchanged, the relationship between diameter, applied voltage and maximum temperature is shown in (a). With diameter unchanged, the relationship between wire length, applied voltage and maximum temperature is shown in (b).[]{data-label="Heat"}](Picture3_1_.png)
The heat generation capability of nichrome wire is evaluated in this experiment. Nichrome wires with diameters of 0.1 mm, 0.25 mm and 0.5 mm are tested, considering the dimension of the tube, space in capsule and difficulty of fabrication. Wires of different dimensions are twined into spiral shape and connected to a DC regulated power supply, and temperature is tested using a thermocouple that is connected to a multimeter (UT33, Uni-Trend Technology (China) Co., Ltd.). In the experiment, we define the maximum temperature as the highest temperature that remains unchanged over 10 s after applying current. Besides, the resistances of the nichrome wires are also tested.
The experiment result is shown as Fig. 5. According to the result, when wire length is unchanged, with diameter increases, the maximum temperature will accordingly increase. With diameter of the wire remains unchanged, smaller length will lead to higher maximum temperature. The resistance of the nichrome wire is as following: 0.275 $\Omega$/cm for diameter 0.5 mm, 0.45 $\Omega$/cm for diameter 0.25 mm, and 0.175 $\Omega$/cm for diameter 0.1 mm. Based on experiment, the nichrome wire implemented in the prototype has a diameter of 0.25 mm and a length of 4 cm, and work under 1.5 V DC power supply which is the same as the voltage provided by conventional batteries for capsule endoscopes. The chosen wire could reach to a maximum temperature of 55 $^\circ$C, at which the wax starts to melt according to real test and the localized heating could be isolated inside the encapsulate of the capsule.
Mucus Collection Test
---------------------
![(a) The experiment setup for mucus collection test. (b) At the beginning of the experiment, a permanent magnet was placed 40 mm below the capsule to trigger the sampling process. (c) Collection process finished after 100 s; the chamber expanded to original shape. (d) Solution could be collected in the chamber. (e) Collected solution could be extracted from capsule through the hole on the chamber; (f) 0.3 ml solution was collected and wax was easily separated from sample.[]{data-label="Collection"}](Picture8.png)
In this experiment, the capsule’s performance in mucus collection was tested. The experiment setup is shown in Fig. 6(a). A PVC tube was used as the container. An encapsulate that only covers lower half of the chamber was implemented for better view of the soft chamber’s state. The capsule was powered by a 1.5 V DC regulated power supply and been merged in 1.2% alginate solution dyed in blue, which was used to mimic the viscosity of the intestinal mucus while at the same time produce better view of the result. At the start of the experiment, a permanent magnet with surface magnetic field strength of 274 mT was placed at 40 mm below the capsule to trigger the sampling process (see Fig. 6(b)). The 40mm distance is similar to the abdominal subcutaneous fat thickness of obese people [@c24]. The minimum magnetic field required to close the reed switch was 48 mT according to measurement. The results are shown in Fig 6(c)-(f). It took approximate 100 seconds for the sampling procedure. The chamber expanded to its original shape and collected solution could be extracted from the chamber through the rear hole. Approximate 0.3 ml solution was collected. The wax was solidified after sampling process and could be easily separated with the collected sample.
Following optimization methods could be carried out to improve the performance of the capsule. The tube, which is made of photosensitive silicone in this design, could be changed to metal to improve the heat conductivity and therefore decrease the respond time and demanded energy. Also, the nichrome wire could be covered with insulation paint to prevent short circuit. Furthermore, the shape and structure of vacuum chamber could be optimized to improve the utilization of space.
Leakage Test in Physiological Condition
---------------------------------------
The sealing capability of wax and silicone vacuum chamber is evaluated in this experiment, to ensure that the capsule is able to remain sealed until triggered at the required location. In stomach, the pH value of gastric juice is 1.5-3.5 [@c29], [@c17] and the maximum half emptying time is 202 minutes [@c18]. In small intestine, the pH value is from 6 to 7.4 [@c19] and maximum transit time is 210 minutes [@c18]. The leakage test is based on these physiological conditions.
First, to simulate the environment and transit time inside stomach, which the capsule must pass through using passive locomotion, the sealed chamber was placed in 30 ml HCl solution with pH value equals to 2 for 3.5 hours under the temperature of 37 $^\circ$C. Then, the capsule was moved to 30 ml NaOH solution whose pH value is adjusted to 7.4 and placed for 4 hours. The temperature was also regulated to 37 $^\circ$C. There was no leakage happened after the 7.5 hours’ test and no obvious flaw observed on the surface of chamber and encapsulate. Experimental conditions and results are shown in Table 1.
Results
----------- ------------- -------------- ------------
Temperature 37 $^\circ$C
Stomach pH value 2 No leakage
Time 3.5 h
Temperature 37 $^\circ$C
Intestine pH value 7.4 No leakage
Time 4 h
: Conditions and Results of Reliability Test[]{data-label="table_example"}
Conclusions and Future Work
===========================
In this work, a soft capsule for mucus collection is designed and a prototype for function demonstration is fabricated. The capsule implements a soft chamber which is sealed with wax to generate vacuum environment for mucus suction, and moves passively along the GI tract. The mucus collection procedure is triggered by a permanent magnet that sticks on patient’s skin proximally above the ROI. Experiments are conducted to show the validity of the design.
Future work will mainly focus on the following aspects to further complete the results and enhance the efficiency and stability of the capsule.
- Optimization on the capsule design should be made in the next prototypes to shorten the response time and increase the mucus collection capability, including improving the heat conductivity of the tube, changing the hardness of the chamber, etc.
- Mucus collection capability of the capsule should be further tested under various experimental conditions, for example, on soft tissue surface and in the environment with thinner mucus layer.
- Rolling and dragging movements should be added to the leakage test of further prototypes.
- The electrical conductivity of mucus will be tested through ex-vivo experiments in further prototypes.
- In-vivo experiment could be carried out when this technique is more refined.
Furthermore, extension of function will also be considered in the future. When combines with bacteria sensing technologies [@c20], the capsule may achieve in-vivo mucus analysis in further prototypes. Cargo delivery function may also be designed for releasing swarming microrobots in specific location [@c21]-[@c28] for targeted therapy or other biomedical applications.
Acknowledgment {#acknowledgment .unnumbered}
==============
This research is financially supported by Innovation and Technology Fund (project number: ITS/440/17FP), CUHK T Stone Robotics Institute and CUHK-SJTU Joint Research Center on Medical Robotics, The Chinese University of Hong Kong. The authors would like to express their sincere gratitude to all the members in Professor Li Zhang’s research group at CUHK for their valuable suggestions.
[99]{}
S. Yim, “Magnetically actuated soft capsule endoscopes”, Ph.D thesis, Carnegie Mellon University, Pittsburgh, Pennsylvania, United States, 2013. W. El-Matary, “Wireless capsule endoscopy: indications, limitations, and future challenges”, *Journal of Pediatric Gastroenterology and Nutrition,* vol. 46, no. 1, pp. 4-12, 2008. A. Moglia, A. Menciassi, P. Dario and A. Cuschieri, “Capsule endoscopy: progress update and challenges ahead”, *Nature Reviews Gastroenterology & Hepatology*, vol. 6, no. 6, pp. 353-361, 2009. G. Ciuti, A. Menciassi and P. Dario, “Capsule Endoscopy: From Current Achievements to Open Challenges”, *IEEE Reviews in Biomedical Engineering*, vol. 4, pp. 59-72, 2011. P. Valdastri, R. J. Webster, III, C. Quaglia, M. Quirini, A.Menciassi, and P. Dario, “A new mechanism for mesoscale legged locomotion in compliant tubular environments,” *IEEE Transactions on Robotics,* vol. 25, no. 5, pp. 1047–1057, Oct. 2009. M. Simi, G. Gerboni, A. Menciassi and P. Valdastri, “Magnetic Torsion Spring Mechanism for a Wireless Biopsy Capsule”, *Journal of Medical Devices,* vol. 7, no. 4, p. 041009, 2013. V. Le, L. Hernando, C. Lee, H. Choi, Z. Jin, K. Nguyen, G. Go, S. Ko, J. Park and S. Park, “Shape memory alloy–based biopsy device for active locomotive intestinal capsule endoscope”, *Proceedings of the Institution of Mechanical Engineers, Part H: Journal of Engineering in Medicine,* vol. 229, no. 3, pp. 255-263, 2015. V. Le, Z. Jin, H. Leon-Rodriguez, C. Lee, H. Choi, V. Nguyen, G. Go, S. Ko, J. Park and S. Park, “Electromagnetic field intensity triggered micro-biopsy device for active locomotive capsule endoscope”, *Mechatronics,* vol. 36, pp. 112-118, 2016. V. Le, V. Nguyen, C. Lee, G. Go, J. Park and S. Park, “Miniaturized biopsy module using gripper tool for active locomotive capsule endoscope”, *Mechatronics,* vol. 44, pp. 52-59, 2017. S. Yim and M. Sitti, “Design and rolling locomotion of a magnetically actuated soft capsule endoscope”, *IEEE Transactions on Robotics,* vol. 28, no. 1, pp. 183-194, 2012. D. Son, M. Dogan and M. Sitti, “Magnetically Actuated Soft Capsule Endoscope for Fine-Needle Aspiration Biopsy”, In 2017 IEEE International Conference on Robotics and Automation (ICRA), 2017, pp. 1132–1139. J. Forstner, “Intestinal Mucins in Health and Disease", *Digestion*, vol. 17, no. 3, pp. 234-263, 1978. J. Turner, “Intestinal mucosal barrier function in health and disease”, *Nature Reviews Immunology,* vol. 9, no. 11, pp. 799-809, 2009. H. Li, J. Limenitakis, T. Fuhrer, M. Geuking, M. Lawson, M. Wyss, S. Brugiroux, I. Keller, J. Macpherson, S. Rupp, B. Stolp, J. Stein, B. Stecher, U. Sauer, K. McCoy and A. Macpherson, “The outer mucus layer hosts a distinct intestinal microbial niche”, *Nature Communications,* vol. 6, no. 1, 2015. M. Johansson, H. Jakobsson, J. Holmén-Larsson, A. Schütte, A. Ermund, A. Rodríguez-Piñeiro, L. Arike, C. Wising, F. Svensson, F. Bäckhed and G. Hansson, “Normalization of Host Intestinal Mucus Layers Requires Long-Term Microbial Colonization", *Cell Host & Microbe,* vol. 18, no. 5, pp. 582-592, 2015. G. Hansson, “Role of mucus layers in gut infection and inflammation”, *Current Opinion in Microbiology,* vol. 15, no. 1, pp. 57-62, 2012. M. Johansson, H. Sjövall and G. Hansson, “The gastrointestinal mucus system in health and disease”, *Nature Reviews Gastroenterology & Hepatology,* vol. 10, no. 6, pp. 352-361, 2013 L. Yang, Q. Wang, C. Vong and L. Zhang, “A Miniature Flexible-Link Magnetic Swimming Robot With Two Vibration Modes: Design, Modeling and Characterization”, *IEEE Robotics and Automation Letters,* vol. 2, no. 4, pp. 2024-2031, 2017. T. Xu, J. Yu, X. Yan, H. Choi and L. Zhang, “Magnetic Actuation Based Motion Control for Microrobots: An Overview”, *Micromachines*, vol. 6, no. 9, pp. 1346-1364, 2015. F. Qiu, L. Zhang, K. Peyer, M. Casarosa, A. Franco-Obregón, H. Choi and B. Nelson, “Noncytotoxic artificial bacterial flagella fabricated from biocompatible ORMOCOMP and iron coating”, *J. Mater. Chem. B*, vol. 2, no. 4, pp. 357-362, 2014. X. Yan, Q. Zhou, M. Vincent, Y. Deng, J. Yu, J. Xu, T. Xu, T. Tang, L. Bian, Y. Wang, K. Kostarelos and L. Zhang, “Multifunctional biohybrid magnetite microrobots for imaging-guided therapy”, *Science Robotics*, vol. 2, no. 12, p. eaaq1155, 2017. ISO, “868: 2003, Plastics And Ebonite-Determination of Indentation Hardness by Means of A Durometer (Shore Hardness)”, International Organization for Standardization: Geneva, Switzerland, 2003. F. Armellini, M. Zamboni, L. Rigo, T. Todesco, O. Bosello, I. Bergamo-Andreis and C. Procacci, “The contribution of sonography to the measurement of intra-abdominal fat”, *Journal of Clinical Ultrasound,* vol. 18, no. 7, pp. 563-567, 1990. E. Marieb and K. Hoehn, *Human anatomy & physiology*. San Francisco, Calif.: Pearson Education, 2007. J. J. Powell, S. M. Greenfield, R. P. Thompson, “Concentrations of metals in gastric juice in health and peptic ulcer disease”, *Gut,* vol. 33, no. 12, pp. 1617-1620, Dec. 1992. L. P. Degen and S. F. Phillips, “Variability of gastrointestinal transit in healthy women and men”, *Gut,* vol. 39, no. 2, pp. 299-305, 1996. J. Fallingborg, “Intraluminal pH of the human gastrointestinal tract”, *Dan. Med. Bull.,* vol. 46, pp. 183-196, 1999. Y. Zhang, K. Chan, B. Wang, P. Chiu and L. Zhang, “Spore-derived color-tunable multi-doped carbon nanodots as sensitive nanosensors and intracellular imaging agents”, *Sensors and Actuators B: Chemical,* vol. 271, pp. 128-136, 2018. J. Yu, L. Yang and L. Zhang, “Pattern generation and motion control of a vortex-like paramagnetic nanoparticle swarm”, *The International Journal of Robotics Research,* vol. 37, issue 8, pp. 912 - 930, 2018. J. Yu, T. Xu, Z. Lu, C. Vong and L. Zhang, “On-Demand Disassembly of Paramagnetic Nanoparticle Chains for Microrobotic Cargo Delivery”, *IEEE Transactions on Robotics,* vol. 33, no. 5, pp. 1213-1225, 2017. J. Yu, B. Wang, X. Du, Q. Wang and L. Zhang, “Ultra-extensible ribbon-like magnetic microswarm”, *Nature Communications*, vol. 9, no. 3260, 2018.
[^1]: $^{1}$Xingzhou Du and Kai Fung Chan are with the Department of Biomedical Engineering, The Chinese University of Hong Kong, Hong Kong, China.
[^2]: $^{2}$Xingzhou Du, Kai Fung Chan, Xianfeng Xia, Philip Wai Yan Chiu and Li Zhang are with the Chow Yuk Ho Technology Centre for Innovative Medicine, The Chinese University of Hong Kong, Hong Kong, China.
[^3]: $^{3}$Philip Wai Yan Chiu is with the Department of Surgery, The Chinese University of Hong Kong, Hong Kong, China.
[^4]: $^{4}$Li Zhang is with the Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong, China.
[^5]: $^{5}$Philip Wai Yan Chiu and Li Zhang are with the T Stone Robotics Institute, The Chinese University of Hong Kong, Hong Kong, China.
[^6]: Corresponding author to [[email protected]]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A parallel-in-time algorithm based on an augmented Lagrangian approach is proposed to solve four-dimensional variational (4D-Var) data assimilation problems. The assimilation window is divided into multiple sub-intervals that allows to parallelize cost function and gradient computations. Solution continuity equations across interval boundaries are added as constraints. The augmented Lagrangian approach leads to a different formulation of the variational data assimilation problem than weakly constrained 4D-Var. A combination of serial and parallel 4D-Vars to increase performance is also explored. The methodology is illustrated on data assimilation problems with Lorenz-96 and the shallow water models.'
author:
- Vishwas Rao and Adrian Sandu
title: 'A Time-parallel Approach to Strong-constraint Four-dimensional Variational Data Assimilation'
---
Introduction {#sec:Intro}
============
Predicting the behavior of complex dynamical systems, such as the atmosphere, requires using information from observations to decrease the uncertainty in the forecast. Data assimilation combines information from a numerical model, prior knowledge, and observations (all with associated errors) in order to obtain an improved estimate of the true state of the system. Data assimilation is an important application of data-driven application systems (DDDAS [@dddas2], or InfoSymbiotic systems) where measurements of the physical system are used to constrain simulation results.
Two approaches to data assimilation have gained widespread popularity: variational and ensemble-based methods. The ensemble-based methods are rooted in statistical theory, whereas the variational approach is derived from optimal control theory. The variational approach formulates data assimilation as a nonlinear optimization problem constrained by a numerical model. The initial conditions (as well as boundary conditions, forcing, or model parameters) are adjusted to minimize the discrepancy between the model trajectory and a set of time-distributed observations. In real-time operational settings the data assimilation process is performed in cycles: observations within an assimilation window are used to obtain an optimal trajectory, which provides the initial condition for the next time window, and the process is repeated in the subsequent cycles. The variational methodology is widely adopted by most national and international numerical weather forecast centers to provide the initial state for their forecast models. Performing nonlinear optimization in the 4D-Var framework is an inherently sequential process. Computer architectures progressively incorporate more parallelism, while maintaining a constant processor speed. As our understanding of the physics improves the computer models become increasingly more complex. Advanced and scalable parallel algorithms to solve 4D-Var need to be developed to continue to perform data assimilation in real time. This challenge has been partially addressed by exploring parallelism in spatial dimension. Trémolet and Le Dimet [@Tremolet1996657] have shown how variational data assimilation can be used to couple models and to perform parallelization in space for the assimilation process. Rantakokko [@Rantakokko19972017] considers different data distribution strategies to perform parallel variational data assimilation in the spatial dimension. A scalable approach for three dimensional variational (3D-Var) data assimilation is presented in [@Damore:2014] and parallelism is achieved by dividing the global problem into multiple local 3D-Var sub-problems. Multiple copies of modified 3D-Var problem, which ensures feasibility at the boundaries of the sub-domains, are solved across processors and the global 3D-Var minimum is obtained by collecting the local minima.
An important challenge associated with computing solutions to 4D-Var problem is parallelization in the temporal dimension. Fisher [@fisher2013Parallel4DVar] attempts to address this challenge by the saddle point formulation that solves directly the optimality equations. Aguiar et al. [@ICML2011Martins_150] apply the augmented Lagrangian approach to constrained inference problems in the context of graphical models. The approach proposed herein uses the augmented Lagrangian framework in the context of 4D-Var data assimilation. The most computationally expensive components of 4D-Var, namely the cost function and gradient evaluations, are performed in a time-parallel manner.
The remainder of this paper is organized as follows: Section \[sec:DA\] introduces data assimilation and the traditional 4D-Var approach. Section \[sec:Reformulation\] formulates the 4D-Var problem in an augmented Lagrangian framework to expose the time parallelism in the cost function and gradient evaluations. Section \[sec:Algorithm\] gives a detailed description of the parallel assimilation algorithm. Section \[sec:exp\] shows the numerical results with the small, chaotic Lorenz-96 model, and a relatively large shallow water on the sphere model. Concluding remarks and future research directions are discussed in Section \[sec:conc\].
Four-dimensional variational data assimilation {#sec:DA}
==============================================
Data assimilation (DA) is the fusion of information from priors, imperfect model predictions, and noisy data, to obtain a consistent description of the true state ${\mathbf{x}}^{\rm true}$ of a physical system [@daley1993; @kalnay2003; @sandu2011chemical; @sandu2005adjoint]. The best estimate that optimally fuses all these sources of information is called the analysis ${\mathbf{x}}^{\rm a}$.
The prior information encapsulates our current knowledge of the system. Usually the prior information is contained in a background estimate of the state ${\mathbf{x}}^{\rm b}$ and the corresponding background error covariance matrix $\mathbf{B}$.
The model captures our knowledge about the physical laws that govern the evolution of the system. The model evolves an initial state ${\mathbf{x}}_0 \in \mathbb{R}^n$ at the initial time $t_0$ to future states ${\mathbf{x}}_{\rm k} \in \mathbb{R}^n$ at future times $t^{\{ \ell \}}$. A general model equation is represented as follows: $$\label{eqn:genmodel}
{\mathbf{x}}_{\rm k} = \mathcal{M}_{t_0 \rightarrow t^{\{ \ell \}}} \left({\mathbf{x}}_0\right)\,.$$ Observations are noisy snapshots of reality available at discrete time instances. Specifically, measurements ${\mathbf{y}}_{\rm k} \in \mathbb{R}^m$ of the physical state ${\mathbf{x}}^{\rm true}\left(t^{\{ \ell \}}\right)$ are taken at times $t^{\{ \ell \}}$, ${\rm k}=1,\cdots, N$. The model state is related to observations by the following relation: $$\begin{aligned}
\label{eqn:genobs}
{\mathbf{y}}_{\rm k} &=& \mathcal{H}\left({\mathbf{x}}_{\rm k}\right)-\varepsilon_{\rm k}^{\rm obs},\quad {\rm k}=1,\cdots,N, \\
\nonumber
\varepsilon_{\rm k}^{\rm obs} &=& \varepsilon_{\rm k}^{\rm representativeness} + \varepsilon_{\rm k}^{\rm measurement} \,.\end{aligned}$$ The observation operator $\mathcal{H}$ maps the model state space onto the observation space. The observation error term $\left(\varepsilon_{\rm k}^{\rm obs}\right)$ accounts for both measurement and representativeness errors. Measurement errors are due to imperfect sensors. The representativeness errors are due to the inaccuracies of the mathematical and numerical approximations inherent to the model.
Variational methods solve the data assimilation problem in an optimal control framework, where one finds the control variable which minimizes the mismatch between the model forecasts and the observations. Strong-constraint 4D-Var assumes that the model is perfect [@sandu2011chemical; @sandu2005adjoint]. The control parameters are the initial conditions ${\mathbf{x}}_0$, which uniquely determine the state of the system at all future times via the model equation . The background state is the prior best estimate of the initial conditions ${\mathbf{x}}_0^{\rm b}$, and has an associated initial background error covariance matrix $\mathbf{B}_0$. Observations ${\mathbf{y}}_{\rm k}$ at $t^{\{ \ell \}}$ have the corresponding observation error covariance matrices $\mathbf{R}_{\rm k}$, ${\rm k}=1,\cdots,N$. The 4D-Var problem provides the estimate ${\mathbf{x}}_0^{\rm a}$ of the true initial conditions as the solution of the following optimization problem $$\label{eqn:ip}
{\mathbf{x}}_0^{\rm a} = \underset{{\mathbf{x}}_0} {\text{ arg\, min}}~~ {\mathcal{J}}\left({\mathbf{x}}_0\right) \qquad
\text{subject to}~ \text{\eqref{eqn:genmodel}},$$ with the following cost function: $$\begin{aligned}
\label{eqn:fdvar-cf}
\mathcal{J}\left({\mathbf{x}}_0\right) &=& \frac{1}{2} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b} \right)^{\rm T} \mathbf{B}_0^{-1} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b} \right) \\ \nonumber
&&+ \frac{1}{2} \displaystyle \sum_{\rm k=1}^N \left(\mathcal{H}_{\rm k}\left({\mathbf{x}}_{\rm k}\right) - {\mathbf{y}}_{\rm k} \right)^{\rm T} \mathbf{R}_{\rm k}^{-1} \left(\mathcal{H}_{\rm k}\left({\mathbf{x}}_{\rm k}\right) - {\mathbf{y}}_{\rm k} \right).\end{aligned}$$ The first term of the sum quantifies the departure of the solution ${\mathbf{x}}_0$ from the background state ${\mathbf{x}}_0^{\rm b}$ at the initial time $t_0$. The second term measures the mismatch between the forecast trajectory (model solutions ${\mathbf{x}}_{\rm k}$) and observations ${\mathbf{y}}_{\rm k}$ at all times $t^{\{ \ell \}}$ in the assimilation window. The weighting matrices $\mathbf{B}_0$ and $\mathbf{R}_{\rm k}$ need to be predefined, and their quality influences the accuracy of the resulting analysis.
Weak constraint 4D-Var [@sandu2011chemical] removes the perfect model assumption by allowing a model error $\eta_{{\rm k}+1} = {\mathbf{x}}_{\rm k +1} - \mathcal{M}_{\rm k, \rm k+1} \left({\mathbf{x}}_{\rm k}\right)$. Under the assumption that the model errors are normally distributed, $\eta_{\rm k} \in \mathcal{N}(0, \mathbf{Q}_{\rm k})$, the weak constraint 4D-Var solution is the unconstrained minimizer of the cost function $$\begin{aligned}
\label{eqn:weakcf}
\mathcal{J}^{\rm weak}\left({\mathbf{x}}_0,\dots,{\mathbf{x}}_N\right) &=& \mathcal{J}\left({\mathbf{x}}_0\right) + \\\nonumber
&&\frac{1}{2} \displaystyle \sum_{k=0}^{N-1} \left({\mathbf{x}}_{\rm k +1} - \mathcal{M}_{\rm k, \rm k+1} \left({\mathbf{x}}_{\rm k}\right)\right)^{\rm T} \mathbf{Q}_{{\rm k}+1}^{-1} \, \left({\mathbf{x}}_{\rm k +1} - \mathcal{M}_{\rm k, \rm k+1} \left({\mathbf{x}}_{\rm k}\right)\right).\end{aligned}$$ The control variables are the states of the system at all times in the assimilation window.
In this paper we focus on the strong constraint formulation . The minimizer of is computed iteratively using gradient-based numerical optimization methods. First-order adjoint models provide the gradient of the cost function [@cacuci2005sensitivity], while second-order adjoint models provide the Hessian-vector product (e.g., for Newton-type methods). The methodology for building and using various adjoint models for optimization, sensitivity analysis, and uncertainty quantification is discussed in [@cioaca2012second; @sandu2005adjoint]. Various strategies to improve the the 4D-Var data assimilation system are described in [@cioaca2014optimization]. The procedure to estimate the impact of observation and model errors is developed in [@rao2014posterioriJournal; @rao2014posteriori]. A framework to perform derivative free variational data assimilation using the trust-region framework is given in [@ruiz2015derivative].
The iterative solution of is highly sequential: first, one iteration follows the other; next, the forward and adjoint models are run sequentially forward and backward in time, respectively. In order to reveal additional parallelism the solution to 4D-Var problem is is approached using the augmented Lagrangian framework. In this framework, the assimilation window is divided into multiple sub-intervals, and the model constraints are explicitly imposed at the boundaries. This approach bears similarities with the Parareal approach that exploits time parallelism in the solution of ordinary differential equations [@Rao201476].
4D-Var solution by the augmented Lagrangian approach {#sec:Reformulation}
====================================================
The 4D-Var cost function is minimized subject to the generic model constraints . The model equations can also be written as $$\label{eqn:model}
{\mathbf{x}}_{\rm k +1} = \mathcal{M}_{\rm k, \rm k+1} \left({\mathbf{x}}_{\rm k}\right), \quad {\rm k} = 0,1,\cdots,N-1\,,$$ where $\mathcal{M}_{\rm k, \rm k+1}$ represents the model solution operator that propagates the state ${\mathbf{x}}_{k}$ at $t_{k}$ to the state ${\mathbf{x}}_{k+1}$ at $t_{k+1}$. The minimizer of under the constraints is the unconstrained minimizer of the Lagrangian $$\begin{aligned}
\label{eqn:Lagrangian}
\mathbf{L}({\mathbf{x}}_0,\dots,{\mathbf{x}}_N;\lambda_0,\dots,\lambda_N) &=&
\mathcal{J}({\mathbf{x}}_0)-\displaystyle \sum_{\rm k=0}^{\rm N-1} \lambda_{{\rm k}+1}^{\rm T} \cdot \left({\mathbf{x}}_{\rm k +1} - \mathcal{M}_{\rm k, \rm k+1} \left({\mathbf{x}}_{\rm k}\right)\right)\,.\end{aligned}$$ To expose time parallelism the assimilation window is divided into $N$ sub-intervals, namely, $$\label{eqn:interval-partition}
\lbrack t_0, t_N \rbrack = \lbrack t_0, t_1\rbrack \, \cup \,\dots \, \cup \, \lbrack t_{N-1}, t_N\rbrack.$$ The forward model and adjoint model states at the interval boundaries are denoted by $$\label{eqn:solution-partition}
{\mathbf{x}}= [{\mathbf{x}}_0,\, \cdots, \,{\mathbf{x}}_N], \quad \bm{\lambda} = [\lambda_0,\, \cdots, \,\lambda_N],$$ respectively. We denote by ${\mathbf{x}}^{\rm a}$ the optimal solution and by $\bm{\lambda}^{\rm a}$ the optimal value of the Lagrange multiplier in .
The augmented Lagrangian [@wright1999numerical Section 17.3] associated with and reads: $$\begin{aligned}
\label{eqn:AugLagCF}
\mathcal{L}\left({\mathbf{x}},\bm\lambda,\mu\right) &=& \frac{1}{2} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b} \right)^{\rm T} \mathbf{B}_0^{-1} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b} \right) \\ \nonumber
&&+ \frac{1}{2} \sum_{\rm k=1}^{\rm N}\left(\mathcal{H} \left({\mathbf{x}}_{\rm k}\right) - {\mathbf{y}}_{\rm k}\right)^{\rm T} {\mathbf{R}}_{\rm k}^{-1} \left(\mathcal{H} \left({\mathbf{x}}_{\rm k}\right) - {\mathbf{y}}_{\rm k}\right)\\
\nonumber
&& -\sum_{\rm k=0}^{\rm N-1} \lambda_{\rm k+1}^{\rm T} \left({\mathbf{x}}_{\rm k+1} -\mathcal{M}_{\rm k, k+1} \left({\mathbf{x}}_{\rm k}\right)\right) \\\nonumber
&& + \frac{\mu}{2} \sum_{\rm k=0}^{\rm N-1} \left({\mathbf{x}}_{\rm k+1} - \mathcal{M}_{\rm k, k+1} \left({\mathbf{x}}_{\rm k}\right)\right)^{\rm T}\, \mathbf{P}_{\rm k+1}^{-1}\,\left({\mathbf{x}}_{\rm k+1} - \mathcal{M}_{\rm k, k+1} \left({\mathbf{x}}_{\rm k}\right)\right)\,,\end{aligned}$$ where ${\mathbf{P}}_{\rm k}$’s are error scaling matrices. This is the (regular) Lagrangian for the problem that minimizes the cost function $$\begin{aligned}
\label{eqn:augmentedcf}
\mathcal{J}^{\rm augmented}\left({\mathbf{x}}_0,\dots,{\mathbf{x}}_N\right) &=& \mathcal{J}\left({\mathbf{x}}_0\right) + \\ \nonumber
&&\frac{\mu}{2} \displaystyle \sum_{k=0}^{N-1} \left({\mathbf{x}}_{\rm k +1} - \mathcal{M}_{\rm k, \rm k+1} \left({\mathbf{x}}_{\rm k}\right)\right)^{\rm T} \mathbf{P}_{{\rm k}+1}^{-1} \, \left({\mathbf{x}}_{\rm k +1} - \mathcal{M}_{\rm k, \rm k+1} \left({\mathbf{x}}_{\rm k}\right)\right)\end{aligned}$$ subject to the model constraint . The constrained minimization of is equivalent to since the additional term is zero along the constraints. Note that is a constrained minimization problem, unlike , which is unconstrained.
The original 4D-Var problem in is solved in the augmented Lagrangian framework by performing a sequence of unconstrained minimizations $$\label{eqn:ipAug}
{\mathbf{x}}^{\{\ell\}} = \arg\, \min_{{\mathbf{x}}}~~ \mathcal{L}\left({\mathbf{x}}, \widetilde{\bm\lambda}^{\{\ell\}}, \mu^{\{\ell\}} \right), \quad \ell = 0,1,\dots\,.$$ If $\widetilde{\bm\lambda}^{\{\ell\}} \approx \bm\lambda^{\rm a}$ then $\widetilde{{\mathbf{x}}}^{\{\ell\}} \approx {\mathbf{x}}^{\rm a}$, and the solution error decreases with increasing $\mu$ [@wright1999numerical Section 17.3].
The optimization proceeds in cycles of inner and outer iterations. Inner iterations solve the optimization problem for particular values of $\mu^{\{\ell\}}$ and $\widetilde{\bm\lambda}^{\{\ell\}}$. After each solution of the outer iteration $\ell$ is completed by updating the Lagrange multiplier approximation and the penalty parameter, as follows:
\[eqn:update\] $$\begin{aligned}
\label{eqn:updateMu}
\mu^{\rm \{\ell+1\}} &=& \rho\, \mu^{\rm \{\ell\}}\,, \\[3pt]
\label{eqn:updateLambda}
\lambda^{\rm \{\ell+1\}}_{\rm k} &=& \lambda^{\rm \{\ell\}}_{\rm k} - \mu^{\rm \{\ell\}} \left({\mathbf{x}}^{\rm \{\ell\}}_{\rm k} - \mathcal{M}_{\rm k-1, k}({\mathbf{x}}^{\rm \{\ell\}}_{\rm k-1}) \right)\,, \quad \rm{k}=0, \dots , N\,.\end{aligned}$$
The penalty parameter is progressively increased by a constant $\rho > 1$ in order to impose the model constraints . Different other strategies to update $\mu$ and $\bm\lambda$ can be used [@birgin2008improving].
Augmented Lagrangian optimization
---------------------------------
Figure \[fig:Iterations\_Lorenz\] illustrates the convergence process of the augmented Lagrangian 4D-Var. The assimilation window is divided into multiple sub-intervals . The control vector ${\mathbf{x}}$ for the optimization process contains the state vector dat the beginning of each of the sub-intervals . The initial value of the control vector contains the background states ${\mathbf{x}}^{\rm b}_k$ at the beginning of each of the sub-interval, and is therefore a continuous curve. The outer iterations start with a small value of $\mu$ that only imposes the constraints loosely. Consequently, the solutions during the first outer iterations show large discontinuities at the interval boundaries. Subsequently, $\mu$ is increased and as a result the constraints are satisfied accurately. This results in a smooth solution curve resembling closely the serial 4D-Var solution.
The parallel algorithm {#sec:Algorithm}
======================
Most of the computational time required by the 4D-Var solution is claimed by the multiple cost function and gradient evaluations. In the augmented Lagrangian 4D-Var formulation – the forward and adjoint models can be run in parallel over sub-intervals, as explained next.
Parallel-in-time runs of the forward model
------------------------------------------
The value of the augmented Lagrangian cost function can be computed in parallel. Specifically, on each sub-interval $[t^{\{ \ell \}},t^{\{ \ell+1 \}}]$ in a forward solution $\mathcal{M}_{\rm k, k+1} \left({\mathbf{x}}_{\rm k}\right)$ is computed starting from the initial value ${\mathbf{x}}_{\rm k}$ . The forward model runs on each sub-interval can be carried out concurrently. The computational steps are detailed in Algorithm \[alg:cf\].
**Input:** $\lbrack {\mathbf{x}}_0, \dots, {\mathbf{x}}_{\rm N};\, \lambda_1, \dots, \lambda_{\rm N} \rbrack$ **Output:** $\mathcal{L}$ $\mathcal{L} \gets 0$ $\Delta{\mathbf{x}}_{\rm k+1} \gets \,{\mathbf{x}}_{\rm k+1} - \mathcal{M}_{\rm k,k+1}\left({\mathbf{x}}_{\rm k}\right)$ $\Delta {\mathbf{y}}_{\rm k+1} \gets \mathcal{H} \left({\mathbf{x}}_{\rm k+1}\right) - {\mathbf{y}}_{\rm k+1}$ $\mathcal{L} \gets \mathcal{L} + \displaystyle\frac{\mu}{2}\, \Delta {\mathbf{x}}_{\rm k+1}^{\rm T}\, \mathbf{P}^{-1}_{\rm k+1}\, \Delta {\mathbf{x}}_{\rm k+1}
- \lambda_{\rm k+1}^{\rm T}\, \Delta {\mathbf{x}}_{\rm k+1} + \displaystyle\frac{1}{2}\, \Delta {\mathbf{y}}_{\rm k+1}\, {\mathbf{R}}_{\rm k+1}^{-1}\, \Delta {\mathbf{y}}_{\rm k+1}$ $\mathcal{L} \gets \mathcal{L}+\displaystyle\frac{1}{2} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b} \right)^{\rm T} \mathbf{B}_0^{-1} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b} \right)$
Parallel-in-time runs of the adjoint model
------------------------------------------
The gradient of the augmented Lagrangian with respect to the forward model state is given by:
\[eqn:AugLagGrad\] $$\begin{aligned}
\label{eqn:AugLagGrad0}
\nabla_{{\mathbf{x}}_0} \mathcal{L}&=&{\mathbf{B}}_0^{-1} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b}\right) + {\mathbf{M}}^{\rm T}_{0,1} \lambda_1 - {\mu}\, {\mathbf{M}}^{\rm T}_{0,1} \, {\mathbf{P}}_0^{-1} \left({\mathbf{x}}_1 - \mathcal{M}_{0,1}\left({\mathbf{x}}_0\right) \right)\,, \\ [5pt]
\label{eqn:AugLagGrad1}
\nabla_{{\mathbf{x}}_{\rm k}}\mathcal{L} &=&\mathbf{H}_{\rm k}^{\rm T} {\mathbf{R}}_{\rm k}^{-1} \left(\mathcal{H}\left({\mathbf{x}}_{\rm k}\right) -{\mathbf{y}}_{\rm k} \right) - \lambda^{\rm T}_{\rm k} + \mathbf{M}^{\rm T}_{\rm k,\rm k+1} \lambda_{\rm k+1} \\
&& + \mu\, \mathbf{P}^{-1}_{{\rm k}-1}\left({\mathbf{x}}_{\rm k} - \mathcal{M}_{\rm k-1,k}\left({\mathbf{x}}_{\rm k-1}\right)\right) \nonumber \\
&& - \mu\, \mathbf{M}^{\rm T}_{\rm k,\rm k+1}\, \mathbf{P}^{-1}_{\rm k} \left({\mathbf{x}}_{\rm k+1} - \mathcal{M}_{\rm k, k+1} \left({\mathbf{x}}_{\rm k}\right)\right)\,, \nonumber
\nonumber
\quad \ {\rm k} = 1, \dots, N-1, \\ [5pt]
\label{eqn:AugLagGrad2}
\nabla_{{\mathbf{x}}_{\rm N}}\mathcal{L} &=& \mathbf{H}_{\rm N}^{\rm T} {\mathbf{R}}_{\rm N}^{-1}\left(\mathcal{H}\left({\mathbf{x}}_{\rm N}\right) -{\mathbf{y}}_{\rm N} \right) - \lambda_{\rm N}^{\rm T}\\ \nonumber
&&+ \mu\, {\mathbf{P}}_{\rm N -1}^{-1} \left( {\mathbf{x}}_{\rm N} - \mathcal{M}_{\rm{N}-1, \rm N}\left({\mathbf{x}}_{\rm N-1}\right) \right)\,,\end{aligned}$$
where the tangent linear operators of the observation and model solution operators are $$\left.\mathbf{H}_{\rm k} = \frac{\partial \mathcal{H}_{\rm k}}{\partial {\mathbf{x}}_{\rm k}}\right\rvert_{{\mathbf{x}}_{\rm k}},
\qquad \left. \mathbf{M}_{\rm k,k+1} = \frac{\partial \mathcal{M}_{\rm k,k+1}}{\partial {\mathbf{x}}_{\rm k}}\right\rvert_{{\mathbf{x}}_{\rm k}},$$ respectively, and their transposes are the corresponding adjoint operators. Using the notation $$\begin{aligned}
\Delta{\mathbf{x}}_{\rm k} &:=& {\mathbf{x}}_{\rm k} - \mathcal{M}_{\rm k-1, k} \left({\mathbf{x}}_{\rm k-1}\right) \\[5pt]
\Delta{\mathbf{y}}_{\rm k} &:=& \mathcal{H}\left({\mathbf{x}}_{\rm k}\right) -{\mathbf{y}}_{\rm k}\end{aligned}$$ the gradient can be written as $$\begin{aligned}
\nabla_{{\mathbf{x}}_0} \mathcal{L}&=&{\mathbf{B}}_0^{-1} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b}\right) + {\mathbf{M}}^{\rm T}_{0,1} \lambda_1 - {\mu}\, {\mathbf{M}}^{\rm T}_{0,1} \, {\mathbf{P}}_1^{-1} \Delta{\mathbf{x}}_1\,, \\ [5pt]
\nabla_{{\mathbf{x}}_{\rm k}}\mathcal{L} &=&\mathbf{H}_{\rm k}^{\rm T} {\mathbf{R}}_{\rm k}^{-1}\, \Delta{\mathbf{y}}_{\rm k} - \lambda^{\rm T}_{\rm k} + \mathbf{M}^{\rm T}_{\rm k,\rm k+1}\, \left( \lambda_{\rm k+1} - \mu \, \mathbf{P}^{-1}_{\rm k+1} \Delta{\mathbf{x}}_{\rm k+1}\right) \\
&& + \mu\, \mathbf{P}^{-1}_{{\rm k}}\, \Delta{\mathbf{x}}_{\rm k}, \nonumber
\nonumber
\quad \ {\rm k} = 1, \dots, N-1, \\ [5pt]
\nabla_{{\mathbf{x}}_{\rm N}}\mathcal{L} &=& \mathbf{H}_{\rm N}^{\rm T} {\mathbf{R}}_{\rm N}^{-1}\, \Delta{\mathbf{y}}_{\rm N} - \lambda_{\rm N}^{\rm T} + \mu\, {\mathbf{P}}_{\rm N -1}^{-1} \, \Delta{\mathbf{x}}_{\rm N}\,.\end{aligned}$$
The augmented Lagrangian gradient can be evaluated in parallel. On each sub-interval $[t^{\{ \ell \}},t^{\{ \ell+1 \}}]$ in an adjoint solution $ \mathbf{M}^{\rm T}_{\rm k,\rm k+1}\, \lambda_{\rm k+1}$ is computed starting from the terminal value $\lambda_{\rm k+1}$ . The adjoint model runs on each sub-interval can be carried out concurrently. The computational steps are detailed in Algorithm \[alg:grad\].
**Input:** $\lbrack {\mathbf{x}}_0, \dots, {\mathbf{x}}_{\rm N};\, \lambda_1, \dots, \lambda_{\rm N} \rbrack$ **Output:** $\lbrack \nabla_{{\mathbf{x}}_0}\mathcal{L}, \dots, \nabla_{{\mathbf{x}}_{\rm N}}\mathcal{L}\rbrack$ $\Delta{\mathbf{x}}_{\rm k+1} \gets \,{\mathbf{x}}_{\rm k+1} - \mathcal{M}_{\rm k,k+1}\left({\mathbf{x}}_{\rm k}\right)$ $\Delta {\mathbf{y}}_{\rm k+1} \gets \mathcal{H} \left({\mathbf{x}}_{\rm k+1}\right) - {\mathbf{y}}_{\rm k+1}$ $\mathbf{b}_{\rm k+1} \gets \mu\, \mathbf{P}^{-1}_{\rm k+1}\, \Delta{\mathbf{x}}_{\rm k+1} - \lambda_{\rm k+1} $ $\mathbf{d}_{\rm k+1} \gets \mathbf{H}_{\rm N}^{\rm T}\, {\mathbf{R}}_{\rm N}^{-1}\, \Delta {\mathbf{y}}_{\rm k+1} $ ${\mathbf{a}}_{\rm k} \gets \mathbf{M}^{\rm T}_{\rm k,\rm k+1} \, \mathbf{b}_{\rm k+1}$ $\nabla_{{\mathbf{x}}_{\rm k}}\mathcal{L}\gets \mathbf{b}_{\rm k} + \mathbf{d}_{\rm k}-{\mathbf{a}}_{\rm k} $ $\nabla_{{\mathbf{x}}_{\rm 0}}\mathcal{L} \gets {\mathbf{B}}_0^{-1} \left({\mathbf{x}}_0 - {\mathbf{x}}_0^{\rm b}\right) - {\mathbf{a}}_0 $ $\nabla_{{\mathbf{x}}_{\rm N}} \mathcal{L} \gets \mathbf{d}_{\rm N} + \mathbf{b}_{\rm N}$
Initial solution guess
----------------------
The optimization needs to start with some initial guess for ${\mathbf{x}}^{\{0\}}$ and $\bm{\lambda}^{\{0\}}$. The initial guess for the state is obtained by performing a serial forward integration using the background value for initial conditions, ${\mathbf{x}}^{\{0\}}_{\rm k} = {\mathbf{x}}^{\rm b}_{\rm k}$. The initial value for the adjoint variable could be obtained by running once the adjoint model once serially along the background trajectory. In our experiments we choose the simpler, and less expensive, initialization $\lambda^{\{0\}}_{\rm k}=0$.
Updating Lagrange multipliers
-----------------------------
In order to accelerate the convergence of the optimization process we replace the standard updates with the strategy proposed in [@he2010acceleration]. The new update process uses information from the previous two iterations instead of just one iteration. The process takes the following steps:
1. Choose $\bm \lambda^0$ and set $t^{\{1\}} = 1$.
2. Let ${\mathbf{x}}^{\rm \{\ell\}}$ be the solution obtained by solving the optimization problem for particular values of $\mu^{\rm \{\ell\}}$ and $\bm\lambda^{\{\ell\}}$. Apply the classical update to obtain $\mu^{\{\ell+1\}}$ and $\widetilde\lambda^{\{\ell+1\}}$.
3. The updated Lagrange multiplier is obtained as follows: $$\begin{aligned}
\displaystyle t^{\{\ell+1\}} &=&\frac{1}{2}\, \left(1 + \sqrt{1+4 \, (t^{\{\ell\}})^2}\right), \\
\displaystyle \bm\lambda^{\{\ell+1\}} &=& \widetilde {\bm \lambda}^{\{\ell+1\}} + \left(\frac{t^{\{ \ell \}} -1}{t^{\{ \ell+1 \}}}\right) \left( \widetilde{\bm \lambda}^{\{\ell +1\}} - \widetilde {\bm \lambda}^{\rm \{\ell\}} \right) + \left(\frac{t^{\{ \ell \}}}{t^{\{ \ell+1 \}}}\right) \left(\widetilde{\bm \lambda}^{\{\ell +1\}} - \bm \lambda^{\rm \{\ell\}}\right)\,.
\end{aligned}$$
It is important to note that above procedure requires the values of $\widetilde{\bm\lambda}$ from two successive outer iterations, namely $\ell$ and $\ell+1$.
Numerical experiments {#sec:exp}
=====================
We study the performance of the parallel implementation of augmented Lagrangian 4D-Var algorithm using the Lorenz-96 model with 40 variables [@lorenz1996] and a shallow water on the sphere model with $\sim$8,000 variables [@Amik:2007].
Lorenz-96 model
---------------
The Lorenz-96 model [@lorenz1996] is given by $$\label{eqn:Lorenz96}
\frac{\mathrm{d}{\mathbf{x}}_{\rm k}}{\mathrm{d}t} = {\mathbf{x}}_{\rm k-1} \left( {\mathbf{x}}_{\rm k+1} - {\mathbf{x}}_{\rm k-2} \right) - {\mathbf{x}}_{\rm k} + F \,,
\quad k=1,\dots,40,$$ with periodic boundary conditions and the forcing term $F=8$ [@lorenz1996]. We use synthetic observations generated by perturbing the reference trajectory with normal noise with mean zero and standard deviation of $5\%$ of average magnitude of the reference solution. The background uncertainty is set to $8\%$ of average magnitude of the reference solution. The background and observation error covariance matrices are assumed to be diagonal. A vector of equidistant components ranging from $-2$ to $2$ was integrated forward in time for 200 time steps and the final state is taken as a reference initial condition for the experiments.
Shallow water model on the sphere
---------------------------------
The shallow water equations have been used extensively as a simple model of the atmosphere since they contain the essential wave propagation mechanisms found in general circulation models [@Amik:2007]. The shallow water equations in spherical coordinates are:
\[eqn:swe\] $$\begin{aligned}
\frac{\partial u}{\partial t} + \frac{1}{a\cos \theta} \left( u \frac{\partial u}{\partial \lambda} + v \cos \theta \frac{\partial u}{\partial \theta} \right) - \left(f + \frac{u \tan \theta}{a} \right) v + \frac{g}{a \cos \theta} \frac{\partial h} {\partial \lambda} = 0, \\
\frac{\partial v}{\partial t} + \frac{1}{a\cos \theta} \left( u \frac{\partial v}{\partial \lambda} + v \cos \theta \frac{\partial v}{\partial \theta} \right) + \left(f + \frac{u \tan \theta}{a} \right) u + \frac{g}{a} \frac{\partial h} {\partial \theta} = 0, \\
\frac{\partial h}{\partial t} + \frac{1}{a \cos \theta} \left(\frac{\partial\left(hu\right)}{\partial \lambda} + \frac{\partial{\left(hv \cos \theta \right)}}{\partial \theta} \right) = 0.\end{aligned}$$
Here $f$ is the Coriolis parameter given by $f = 2 \Omega \sin \theta$, where $\Omega$ is the angular speed of the rotation of the Earth, $h$ is the height of the homogeneous atmosphere, $u$ and $v$ are the zonal and meridional wind components, respectively, $\theta$ and $\lambda$ are the latitudinal and longitudinal directions, respectively, $a$ is the radius of the earth and $g$ is the gravitational constant. The space discretization is performed using the unstaggered Turkel-Zwas scheme [@Navon19911311; @navon1987application]. The discretized spherical grid has nlon=36 nodes in longitudinal direction and nlat=72 nodes in the latitudinal direction. The semi-discretization in space leads to a discrete model of the form . In the zonal wind, meridional wind and the height variables are combined into the vector ${\mathbf{x}}\in \mathbb{R}^n$ with $n=3\times{\rm nlat}\times{\rm nlon}$. We perform the time integration using an adaptive time-stepping algorithm. For a tolerance of $\displaystyle 10^{-8}$ the average time step size of the time-integrator is 180 seconds. A reference initial condition is used to generate a reference trajectory.
Synthetic observation errors at various times $t_{\rm k}$ are normally distributed with mean zero and a diagonal observation error covariance matrix with entries equal to $(\mathbf{R}_{\rm k})_{\rm i,i}=1$ for $u$ and $v$ components and $(\mathbf{R}_{\rm k})_{\rm i,i}=10^6$ for $h$ components. The $\mathbf{R}_{\rm k}$ values correspond to a standard deviation of $5\%$ for $u$ and $v$ components, and $2\%$ for $h$ component. We construct a flow dependent background error covariance matrix as described in [@Attia:2014; @Attia:2015]. The standard deviation of the background errors for the height component is 2% of the average magnitude of the reference height component in the reference initial condition. The standard deviation of the background errors for the wind components is 15% of the average magnitude of the reference wind component in the reference initial condition.
Experimental setup {#subsec:Tests}
------------------
All numerical experiments are carried out in [Matlab]{}. The parallel implementations are developed using [Matlab]{}’s parallel toolbox. We compare the performance of the proposed parallel implementation with that of the standard 4D-Var. The accuracy of numerical solutions is measured by the root mean square error (RMSE) with respect to a reference solution. The RMSE is given by $$\label{eqn:RMSE}
\textbf{RMSE} = \sqrt{\frac{1}{N}\, \sum_{k=1}^N {\textsc{n}_\textnormal{var}}^{-1} \, \left\Vert {\mathbf{x}}^{\rm a}_{\rm k} -{\mathbf{x}}^{\rm reference}_{\rm k} \right\Vert^2} \,,$$ where the reference solution ${\mathbf{x}}^{\rm reference}$ and the analysis ${\mathbf{x}}^{\rm a}$ are propagated forward in time using the full model, and the difference is measured at all times throughout the assimilation window.
Computing the cost functions and gradients is the most important aspect of the 4D-Var algorithm and hence it is necessary that their computations are scalable. To evaluate the gradient and cost function we carry out numerical integrations of the forward and adjoint models using MATLODE [@adaug:2015]. MATLODE is a [Matlab]{} version of FATODE, which was developed in [Fortran]{} [@zhang2014fatode]. The optimization is carried out using the L-BFGS-B solver [@liu1989limited] implemented in the Poblano optimization toolbox developed at Sandia National Laboratory [@SAND2010-1422].
Results with the Lorenz-96 model
--------------------------------
Figure \[fig:Iterations\_Lorenz\] illustrates the convergence of augmented Lagrangian 4D-Var iterations. The intermediate solutions are discontinuous at sub-interval boundaries. The corresponding errors, with respect to the traditional 4D-Var solution, are shown in Figure \[fig:Errors\_Lorenz\]. They decrease quickly to zero, showing that the augmented Lagrangian 4D-Var solution converges to the solution of traditional 4D-Var.
The errors of the augmented Lagrangian and traditional 4D-Var solutions with respect to the reference solution are shown in Figure \[fig:RMSE\_Lorenz\]. The reference solution is obtained by propagating the reference initial condition using the forward model in . The sequential 4D-Var requires $230$ gradient and $574$ cost function evaluations, where as the parallel 4D-Var requires a total of $100$ gradient and $650$ cost function evaluations for $6$ observations. The Weak scalability results are presented in Figure \[fig:Cputime\_Lorenz\]. As the length of the assimilation window increases, the number of sub-intervals increases, and so does the number of observations (one per sub-interval). The number of cores on which the parallel algorithm is run increases such as to remain equal to the number of sub-intervals. The parallel algorithm is scalable in weak sense: the total computational time increases very slowly with an increased problem size. The most time consuming calculations are those of the cost function and gradient evaluations, which require running the forward and the adjoint models, respectively. The results shown in Figure \[fig:CfGf\_Lorenz\] confirm the good weak scalability of the cost function and gradient computations.
![Lorenz-96 model results. The scaling of overall computational times with increasing number of sub-intervals (which leads to an increase in the length of the assimilation window).[]{data-label="fig:Cputime_Lorenz"}](TimingComparison_Lorenz.pdf){width="0.45\linewidth"}
Results with the the shallow water model
----------------------------------------
Figure \[fig:CfGf\_SWE\] shows the weak scalability of cost function and gradient evaluations with the shallow water model. It can be seen that in both cases the parallel computational time is nearly constant with increasing problem size (number of sub-intervals) and a proportional increase in the number of cores (the number of cores is equal to the number of sub-intervals). Figure \[fig:WorkPerformance\_SWE\] presents the work-precision diagrams, i.e., the evolution of solution accuracy (RMSE) with increasing number of iterations (increasing CPU time). The initial iterates of the parallel 4D-Var solution proceed rapidly, but afterwards the convergence slows down. At this stage the penalty parameter $\mu$ fairly large and the optimization problem becomes more difficult for the LBFGS algorithm. The performance can be improved by replacing LBFGS with algorithms specially tailored to solve optimization problems in the augmented Lagrangian framework [@birgin2014practical]. An alternative strategy is to use a hybrid method: employ parallel 4D-Var for several iterations in the beginning, then continue with traditional serial 4D-Var. Here we perform two outer iterations with small values of $\mu$ and then use this solution as an initial guess for the serial 4D-Var. This strategy improves the performance considerably as seen in Figure \[fig:WorkPerformanceHybrid\]. Table \[tab:timings\] provides the computational times for parallel and serial 4D-Var algorithms. The serial 4D-Var for $9$ observations requires $95$ gradient and $200$ cost function evaluations, whereas the parallel 4D-Var algorithm requires $350$ gradient and $720$ cost function evaluations. In the hybrid methodology, we perform $200$ iterations of parallel 4D-Var and $30$ iterations of serial 4D-Var for $9$ observations. The final RMSE over the assimilation window for both the serial 4D-Var and hybrid methods is $\sim$ 125. We notice a steady increase in speedup as the problem size (number of sub-intervals) is increased. It is possible to further improve the performance of the parallel algorithm by using second derivative information in the form of Hessian-vector products and employ Newton-type methods [@cioaca2012second].
No. of sub-intervals Serial time \[sec.\] Parallel/Hybrid time \[sec.\] Speedup
---------------------- ---------------------- ------------------------------- ---------
5 8,515 10,834 0.786
7 15,745 12,782 1.232
9 22,277 12,971 1.756
: Shallow water equations results. Performance comparison of augmented Lagrangian/parallel and traditional/serial 4D-Var and the corresponding speedups.[]{data-label="tab:timings"}
Conclusions and future work {#sec:conc}
===========================
This work presents an augmented Lagrangian framework to perform strong-constraint 4D-Var data assimilation in parallel. The assimilation window is split in sub-intervals; cost function and gradient evaluations, which are the main components of the algorithm, are performed by running the forward and the adjoint model in parallel across different sub-intervals. To the best of our knowledge this is the first work that uses an augmented Lagrangian approach to data assimilation, and the first one to propose a time-parallel implementation of strong-constraint 4D-var.
Future work will focus on tuning the optimization procedure to improve performance on large scale problems, e.g., data assimilation with the weather research and forecast model [@skamarock2005description]. The size of the control variable in the augmented Lagrangian framework increases with the number of sub-intervals and can become a bottleneck for the optimization. One possible strategy to overcome this is to perform optimization on a coarse grid and use the projected solution as an initial guess for the fine grid. A natural extension to our methodology is to use the augmented Lagrangian framework to expose space-time parallelism in the data assimilation problem such as to create more parallel tasks and improve the overall scalability. Space parallelism in a penalty formulation has been recently discussed in [@Damore:2014]. We will consider the use of optimization algorithms that are specifically tuned to work well in the augmented Lagrangian framework [@birgin2014practical]. Next, Hessian information can be used to accelerate the convergence significantly when the iterates are close to minima. In order to implement this it is useful to explore the construction of second order adjoint models that compute the Hessian-vector products in parallel.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported in part by the awards AFOSR FA9550–12–1–0293–DEF, AFOSR 12-2640-06, NSF DMS–1419003, NSF CCF–1218454, and by the Computational Science Laboratory at Virginia Tech.
References {#references .unnumbered}
==========
[10]{}
P. Aguiar, E. P. Xing, M. Figueiredo, N. A. Smith, and A. Martins. An augmented [L]{}agrangian approach to constrained [MAP]{} inference. In [*Proceedings of the 28th International Conference on Machine Learning (ICML-11)*]{}, pages 169–176. ACM, 2011.
A. Attia, V. Rao, and A. Sandu. A hybrid [Monte-Carlo]{} sampling smoother for four dimensional data assimilation. , Submitted, 2014.
A. Attia, V. Rao, and A. Sandu. A sampling approach for four dimensional data assimilation. In [*Proceedings of the Dynamic Data Driven environmental System Science Conference*]{}, 2014.
A. Augustine and A. Sandu. solver and sensitivity analysis toolbox. , Submitted, 2015.
A. Aved, F. Darema, and E. Blasch. . [www.1dddas.org](www.1dddas.org), 2014.
E. G. Birgin and J. M. Mart[í]{}nez. Improving ultimate convergence of an augmented [Lagrangian]{} method. , 23(2):177–195, 2008.
E. G. Birgin and J. M. Mart[í]{}nez. , volume 10. SIAM, 2014.
D. G. Cacuci, M. Ionescu-[B]{}ujor, and I. M. Navon. , volume 2. CRC Press, 2005.
A. Cioaca, M. Alexe, and A. Sandu. Second-order adjoints for solving [PDE]{}-constrained optimization problems. , 27(4-5):625–653, 2012.
A. Cioaca and A. Sandu. An optimization framework to improve 4[D-Var]{} data assimilation system performance. , 275:377–389, 2014.
R. Daley. , volume 2. Cambridge University Press, 1993.
L. D’Amore, R. Arcucci, L. Carracciuolo, and A. Murli. A scalable approach for variational data assimilation. , 61(2):239–257, 2014.
D. M. Dunlavy, T. G. Kolda, and E. Acar. Poblano v1.0: A [MATLAB]{} toolbox for gradient-based optimization. Technical report, Sandia National Laboratories, 2010.
M. Fisher. Parallelization of 4[D-Var]{} in the time dimension using a saddlepoint algorithm. Slides., 2013.
B. He and X. Yuan. On the acceleration of augmented [Lagrangian]{} method for linearly constrained optimization. , 2010.
M. Navon I and R. De Villiers. The application of the [Turkel-Zwas]{} explicit large time-step scheme to a hemispheric barotropic model with constraint restoration. , 115(5):1036–1052, 1987.
E. Kalnay. . Cambridge University Press, 2003.
D. C. Liu and J. Nocedal. On the limited memory [BFGS]{} method for large scale optimization. , 45(1-3):503–528, 1989.
E. N. Lorenz. Predictabilty: [A]{} problem partly solved. In [*Proceedings of Seminar on Predictability*]{}, pages 40–58, 1996.
I. M. Navon and J. Yu. Exshall: A [Turkel-Zwas]{} explicit large time-step [FORTRAN]{} program for solving the shallow-water equations in spherical coordinates. , 17(9):1311 – 1343, 1991.
J. Rantakokko. Strategies for parallel variational data assimilation. , 23(13):2017 – 2039, 1997.
V. Rao and A. Sandu. A-posteriori error estimates for inverse problems. , Submitted, 2014.
V. Rao and A. Sandu. An adjoint-based scalable algorithm for time-parallel integration. , 5(2):76 – 84, 2014. Empowering Science through Computing + BioInspired Computing.
V. Rao and A. Sandu. A posteriori error estimates for dddas inference problems. , 29:1256–1265, 2014.
E. D. N. Ruiz and A. Sandu. A derivative-free trust region framework for variational data assimilation. , 2015.
A. Sandu, D. N. Daescu, G. R. Carmichael, and T. Chai. Adjoint sensitivity analysis of regional air quality models. , 204(1):222–252, 2005.
Adrian Sandu and Tianfeng Chai. Chemical data assimilation—an overview. , 2(3):426–463, 2011.
William C Skamarock, Joseph B Klemp, Jimy Dudhia, David O Gill, Dale M Barker, Wei Wang, and Jordan G Powers. A description of the advanced research wrf version 2. Technical report, DTIC Document, 2005.
A. St-Cyr, C. Jablonowski, J. M. Dennis, H. M. Tufo, and S. J. Thomas. A comparison of two shallow water models with nonconforming adaptive grids. , 136:1898–1922, 2008.
Y. Trémolet and F.-X. Le Dimet. Parallel algorithms for variational data assimilation and coupling models. , 22(5):657 – 674, 1996.
S. J. Wright and J. Nocedal. , volume 2. Springer New York, 1999.
H. Zhang and A. Sandu. : [A]{} library for forward, adjoint, and tangent linear integration of [ODE]{}s. , 36(5):C504–C523, 2014.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove that the Local Fundamental Group Scheme satisfies the Lefschetz- Bott theorems in characteristic $p$. The proofs are standard applications of the E-S-Z-S vanishing theorems and are routine.'
address: 'School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Mumbai-400005, India'
author:
- 'V. B. Mehta'
title: ' A Simple Application of Enriques-Severi-Zariski-Serre to the Local Fundamental Group Scheme'
---
Introduction
============
The Fundamental Group Scheme was introduced by Madhav Nori in \[N1,N2\]. In order to prove the conjectures made in \[loc.cit\], S.Subramanian and the present author had introduced the “Local Fundamental Group Scheme ”, denoted by $\pi^{loc}(X)$\[MS1\], which is the infinitesimal part of $\pi(X)$,\[MS2\]. One may ask whether the Lefschetz-Bott theorems hold for $\pi^{loc}(X)$. In other words, let $X$ be a smooth projective variety over an algebraically closed filed of characteristic $p$,and let $H$ be a very ample line bundle on $X$. The question is if $\pi^{loc}(Y)
\to \pi^{loc}(X)$ is a [*surjection*]{} if dim $X \geq 2$ and degree $Y
\geq n_0$, where $n_0$ is an integer depending [*only*]{} on $X$. Similiarly , if dim $X \geq 3$,the question is if $\pi^{loc}(Y) \to
\pi^{loc}(X)$ is an [*isomorphism*]{} if deg $Y \geq n_1$, where $n_1$ is an integer depending [*only*]{} on $X$.
In the present note, we give a positive answer to both these questions. The methods only involve applying the lemma of Enriques-Severi-Zariski-Serre ( E-S-Z-S) over and over again. Hence most of the proofs are only sketched, and some are completely omitted. The notation is the same as in \[MS2\],which is briefly recalled.
The first theorem
==================
\[t1\]Let $X$ be a non-singular projective variety over an algebraically closed field of char $p$,of dimension $\geq 2$. Let $H$ be a very ample line bundle on $X$. Then there exists an integer $n_0$, depending [*only*]{} on $X$, such that for any smooth $Y \in |nH|$, of deg $n \geq n_0$, the canonical map : $\pi^{loc}Y \to \pi^{loc}(X)$ is a surjection.
Before we begin the proof, we recall some facts from \[MS2\]. Let $F: X
\to X$ be the Frobenius map. For any integer $t \geq 1$, denote by $C_t$ the category of all $V \in Vect(X)$ such that $F^{t*}(V)$ is the [ *trivial*]{} vector bundle on $X$.Fix a base point $x_0 \in X$, Consider the functor $S : C_t \to Vect (X)$ given by $V \to V_{x_0}$ It is seen that $C_t$ , with the fibre functor $S$, is a Tannaka category \[MS1,MS2\]. denote the corresponding Tannaka group by $G_t$. It is also easily seen that $$\pi^{loc}(X) \simeq \lim_{\leftarrow t}
G_t$$, where $\pi^{loc}(X)$ is the Tannaka group associated to the category $(V \in Vect (X)\ F^t(V)$ is trivial for some $t$.
Now let $t = 1$, and consider $G_1(Y)$ and $G_1(X)$. Let $V \in C_1(X)$. with $V$ stable.$V$ corresponds to a principal H bundle $E \rightarrow X$ ,where $E$ is [*reduced*]{} in the sense of Nori \[N1,page 87,Prop.3\], or just N-reduced. To prove that $G_1(Y) \to G_1(X)$ is a surjection , it is enough to prove that $V/Y$ is [*stable*]{}, or better still that $E/Y$ is [*N-reduced*]{}.
\[2\] If X and E are as above , then there exists an integer $n_0$, depending only on $X$ ,such that for all $n \geq n_0$, and for all smooth $Y \in |nH|$, $E/Y$ is N-reduced.
Let $f: E \rightarrow X$ be given.Then $f_*{{{\mathcal O}}_E}$ belongs to $C_1$. Denote it by $W$ for simplicity. Now consider $$0 \rightarrow {{{\mathcal O}}_X}
\rightarrow F_*{{{\mathcal O}}_X} \rightarrow B^1 \rightarrow 0\eqno{(1)}$$, Tensor (1) with $W(n)$. So there is an $n_0$ such that for $n \geq n_0$, we have $Hom(W(n), B^1) = 0$. So the canonical map $H^1(W(n) \rightarrow H^1(F^*(W(n))$ is injective. But the last space is just $H^1({{{\mathcal O}}_X}(np))$ which is $0$ as soon as $n \geq n_1$, independent of $V \in C_1(X)$. Hence $H^0(Y,W/Y) =1$, which proves that $E$ restricted to $Y$ is also N-reduced.
Now assume $t \geq 2$. We shall assume in fact that $t = 2$, because an almost identical proof works for $t \geq 3$.Consider the map $F^2 : X
\rightarrow X$. Let $B_2^1$ be the cokernel.We have the exact sequence$$0 \rightarrow B^1 \rightarrow B_2^1 \rightarrow F_*B_1 \rightarrow
0\eqno{(2)}$$ Let $W \in C_2(X)$ and tensor (2) with $W(n)$. One sees immediately that if $H^1(W(n){{\otimes}}B^1) = 0$ for all $n \geq n_0$, then also $H^1((W(n){{\otimes}}B_2^1) = 0$ for all $n \geq n_0$, for the [*same*]{} integer $n_0$.So if E is N-reduced on $X$, then E remains N-reduced on $Y$, if deg $Y \geq n_0$ , for the same integer $n_0$, which worked for $C_1(X)$.The proof for bigger $t$ goes the same way , by taking the cokernel of $F^t$, where $t \geq3$. This concludes the proof of Theorem 2.1.
the second theorem
==================
\[t2\]Let $X$ be smooth and projective of dim $\geq 3$. Then there exists an integer $n_0$, depending only on $X$ ,such that for any smooth $Y \in |nH|,n \geq n_0$, the canonical map $\pi^{loc}(Y)
\to \pi^{loc}(X)$ is an isomorphism.
We assume that dim $X \geq 3$, and we pick an arbitrary smooth $Y \in
|nH|$.We first consider a $W \in C_1(Y)$ and show that it lifts uniquely to $V \in C_1(X)$ . Such a $W$ is trivialized by the Frobenius,so there exists $M$, an $r \times r$ matrix of 1- forms on $Y$,which gives an integrable connection $\nabla$ on ${{\mathcal O}}^r_Y$, with $p$- curvature $0$. Consider $$0 \rightarrow \frac{I}{I^2} \rightarrow \Omega^1_X/Y
\rightarrow
\Omega^1_Y \rightarrow 0\eqno{(1)}$$ and $$0 \rightarrow \Omega_X^1(-n)
\rightarrow \Omega^1_X \rightarrow \Omega^1_X/Y \rightarrow 0\eqno{(2)}$$
Here $I$ is the ideal sheaf of $Y, I = {{\mathcal O}}_X(-n)$. We get an integer $n_0$, such that for all $n \geq n_0, H^0(X,\Omega^1_X) \to H^0(Y,\Omega^1_Y)$ is an isomorphism. So $M$ lifts to $\ M_1$, a $r \times r$ matrix of $1$ forms on $X$.Now $\nabla$ has $p$-curvature $0$ on $Y$.We now show that $\nabla_1$ defined by $M_1$ has $p$-curvature $0$ on $X$. The curavture of $\nabla$ is an element of $H^0(End {{\mathcal O}}^r_Y \otimes \Omega^2_Y)$, Again by E-S-Z-S, the curvature of $\nabla_1$ is $0$ if deg $Y \geq n_1$, for some integer $n_1$, depending only on $X$. The $p$-curvature of $\nabla_1$ is an element of $H^0(F^*\Omega^1_X
\otimes End {{\mathcal O}}^r_X)$, which again vanishes if deg $Y \geq n_2$. So any element $W \in C_1(Y)$ lifts uniquely to an element $V \in C_1(X)$.
With X as above, then there exists an integer $n_3$ ,depending only on $X$, such that for any $V \in C_t(X)$, any $t$, we have $H^1(\Omega^1_X( -n) \otimes V) = 0$, for all $n \geq n_3$.
Let $T_X$ be the tangent bundle of $X$. For some $s$ , depending only on $X$, $T_X(s)$ is generated by global sections. So we have $$0 \rightarrow
S^* \rightarrow {{\mathcal O}}_X^N \rightarrow T_X(s) \rightarrow 0.$$ Dualizing,we get $$0 \rightarrow \Omega^1(-s) \rightarrow {{\mathcal O}}_X^N \rightarrow S
\rightarrow 0 \eqno{(1)}$$.Tensor $(1)$ with $V(-t)$,to get $$0
\rightarrow V(-t) \otimes \Omega^1(-s) \rightarrow V(-t)\otimes {{\mathcal O}}_X^N
\rightarrow V(-t)\otimes S \rightarrow 0\eqno{(2)}$$. Applying $F$ to $(2)$ we get $$0 \rightarrow [V(-t) \otimes \Omega^1(-s)]^p \rightarrow
[V(-t)\otimes {{\mathcal O}}_X^N]^p \rightarrow [V(-t) \otimes S]^p \rightarrow
0\eqno{(3)}$$. It is clear that $H^0(V(-t) \otimes S) = 0$ for $t >> 0$, *independent* of $V$. Now look at $$H^1(V(-t) \otimes \Omega^1(-s))
\rightarrow H^1(V(-t) \otimes {{\mathcal O}}_X^N) \rightarrow H^1(V(-t)
\otimes {{\mathcal O}}_X^N)^p \rightarrow 0\eqno{(4)}$$,where the right arrow is the map induced by F.But one knows that $H^1(V(-t) \otimes {{\mathcal O}}_X^N)^p$ vanishes for $t >> 0$, independently of $V \in C_n(X)$ for any $n$. Therefore $H^1(V(-t) \otimes \Omega^1)$ vanishes for all $t >> n$, for an integer $n$ *independent* of $X$.This concludes the proof of Lemma 3.1
Proceeding with the proof of Theorem $3.1$ by the discussion preceeding the Lemma 3.1,it follows that if $W \in C_1(Y)$ lifts to $V$ on $X$,such that $V \in C_1(X)$ and $W$ has a $p$-flat connection ,then $V$ also has a connection.The integrability of this connection on $V$ is also defined by the vanishing of a tensor in $H^0(EndV \otimes \Omega^2_X)$. It is again clear that this tensor vanishes if its restriction to $Y$ vanishes, for $Y$ of degree say $n_4$ ,again independent of $V \in C_t(X)$, for all $t$. Again the $p$-flatness of this connection on $V$ also vanishes ,if the corresponding $p$-curvature vanishes for $V$ restricted to $Y$, for all smooth $Y \in |nH|$, for $n$ independent of $V$. This means that on $X$, if $V$ restricts to $W$ on $Y$, and $W$ has a $p$-flat connection, then $V$ also has a $p$-flat connection. So if $W$ on $Y$ descends to $W_1$ , then $V$ also descends under $F$ to $V_1$, and that $V_1$ restricts to $W_1$ on $Y$. Clearly ,this continues to hold for $V \in C_t(X)$, restricting to $W \in C_t(Y)$, any $t$,for $Y$ of a fixed degree ,depending *only* on $X$. Hence the canonical map of Tannaka Categories : $C_t(X) \rightarrow
C_t(Y)$ , given by restriction from $X$ to $Y$, induces an *isomorphism*:$$G_t(Y) \rightarrow G_t(X),$$ for all $t$.But $$\pi^{loc}(X) \simeq \lim_{\leftarrow t}G_t(X)$$ and similiarly for $Y$. This finishes the proof of Theorem $3.1$
It would be of great interest to prove the above theorems when X is only assumed to be *normal*.
It is also interesting to determine if the category of F-trivial vector bundles ,on a smooth $X$,is $m$- *regular*. This is the case if dim $X \leq 3$.
, Invent. Math. [**148**]{},(2002),143-150. submitted to the Proc. of the Ind. Acad. of Sciences, (Math Sci). ,Proc. of the Ind. Acad. of Sci.(math Sci.),[**91**]{},1982. ,Compositio Math.[**33**]{}, 1976.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Individually trapped $^{137}$Ba$^+$ in an RF Paul trap is proposed as a qubit candidate, and its various benefits are compared to other ionic qubits. We report the current experimental status of using this ion for quantum computation. Future plans and prospects are discussed.'
author:
- 'M. R. Dietrich, A. Avril, R. Bowler, N. Kurz, J. S. Salacka, G. Shu and B. B. Blinov'
bibliography:
- 'NNP2008Proceedings.bib'
- 'books.bib'
- 'other.bib'
title: Barium Ions for Quantum Computation
---
[address=[University of Washington Department of Physics, Seattle, Washington, 98195]{}]{}
Introduction
============
At present, trapped ions are the leading contender for use as a qubit in quantum computation schemes. This is a result of the high degree of motional control possible over the ion, the availability of long established techniques for quantum manipulation of trapped ions, and the success of shelving schemes as a highly efficient readout mechanism. To date, Be$^+$ [@Turchette1998], Ca$^+$ [@Blatt2004; @Lucas2007], Cd$^+$ [@Lee2003], Mg$^+$ [@Schaetz2007], Sr$^+$ [@Letchumanan2007], and Yb$^+$ [@Olmschenk2007] have all been demonstrated as possible ionic qubits. It is possible to create ionic qubits using either two hyperfine levels [@Blinov2004b] or, in some species, two levels separated by an optical transition [@Blatt2004] as the computational basis. We propose to use the ground state hyperfine levels of $^{137}$Ba$^+$ as a qubit. Although Ba$^+$ was the first ion trapped in isolation [@Dehmelt1980], $^{137}$Ba$^+$ was not trapped for another 20 years [@DeVoe2002]. The spectroscopic properties of Ba$^+$ have been carefully studied since then because of its potential applications as an optical frequency standard [@Koerber2002] and in a test of parity non-conservation [@Fortson1993]. It has several desirable qubit properties, including visible wavelength transitions, high natural abundance of the $^{137}$Ba isotope, and a long lived shelving state. Here we demonstrate single qubit initialization, rotations and readout on this new qubit, and discuss the future directions of our work.
Experimental Setup
==================
Ba$^+$, like some of its qubit competitors such as Ca$^+$ and Sr$^+$, has an energy level structure that includes two low lying, long lived D states - see Fig. 1(a). In contrast with these ions, however, all of the Ba$^+$ dipole transitions lie in the visible spectrum, which greatly simplifies laser alignment. The doppler cooling consists of a blue 493 nm transition from the ground state to the $P_{1/2}$ state, which has a branching ratio of 0.244 [@Davidson1991] to $D_{3/2}$. Because this state is long lived, a repump laser at 650 nm is necessary for continuous cooling. However, the upper D state, $D_{5/2}$, is isolated from the cooling cycle and so constitutes a “dark” state which can be used for high fidelity readout. Its lifetime of 35 s helps reduce the overall error rate during readout, compared to the relatively short lifetime of, for example, Ca$^+$ at 1 s. The predominant isotope of Ba is 138 (72% abundance), which has no nuclear spin and so no hyperfine structure to use as a qubit. When trapping the odd isotope of Ba$^+$ (11% abundance), all the various hyperfine levels must be addressed by introducing sidebands onto the two cooling lasers. For the red 650 nm laser, the $D_{3/2}$ levels lie close enough [@Silverans1986] that not all possible transitions need to be covered and so only three modulation frequencies are used - 614, 539, and 394 MHz. For the blue, modulation is introduced with frequency equal to the ground state hyperfine splitting, and the carrier is set such that all ground states are excited only into the $P_{1/2}$ (F=2) manifold, see Fig. 1(b). This is advantageous not only for the optical pumping reasons discussed below, but also because one avoids the $S_{1/2}$ (F=1) to $P_{1/2}$ (F=1) transition which is extremely weak as a result of small, destructively interfering geometric factors [@Metcalf].
The ion can be excited directly from the F=2 ground state to the $D_{5/2}$ level using a 1762 nm fiber laser (Koheras Adjustik). While in this shelved state, the ion will not flouresce when illuminated with the cooling lasers. A bright ion can be distinguished from a dark one with nearly perfect fidelity after a couple ms of observation time, resulting in highly reliable readout. Because the transition is so weak (E2 transition), resonant excitation would require a carefully stabilized laser. However, using adiabatic passage, it is possible to perform highly efficient population transfers using only a poorly stabilized laser [@Wunderlich2005] . Using the 10 mW of available power, simulation indicates that this should be possible on the ms time scale. Because of the many magnetic sublevels of the $D_{5/2}$ state, however, it is necessary to perform some stabilization to prevent accidental excitation into an undesired state.
{width="6cm"}\
(a) \[leveldiagram1\]
{width="6cm"}\
(b) \[leveldiagram2\]
The ground state of $^{137}$Ba$^+$ is split due to the hyperfine interaction by about 8.037 GHz [@Blatt1981], which makes this isotope usable as a hyperfine qubit. We optically pump into the upper $\textrm{m}_F$=0 state with $\pi$ polarized 493 nm light, since parity symmetry forbids the $S_{1/2}$ (F=2,$\textrm{m}_F$=0) to $P_{1/2}$ (F=2,$\textrm{m}_F$=0) transition. After state preparation, we can cause direct Rabi flops between the upper and lower magnetically insensitive $\textrm{m}_F$=0 states using microwaves at the known hyperfine frequency, exposed for varying periods of time.
![A schematic of the experimental setup. The cooling lasers are locked using a commercial wavemeter, which feeds back to the laser systems. Acousto-optic modulators (AOMs) are used for high speed shuttering and, in the case of the 1762, frequency modulation and shifting (Double Pass AOM). Barium ionization is achieved through photoionization with a Xe flash lamp. A mirror on a motorized mount changes the detection from a CCD camera to a PMT.[]{data-label="apparatus"}](minisetup3)
The cooling light is provided by two external cavity diode lasers (ECDL), one at 650 nm (Toptica DL-100) and the other at 493 nm (Toptica SHG-110), the latter of which is a 986 nm diode frequency doubled in a bow-tie enhancement cavity. The available power is 10 mW and 20 mW respectively. Both lasers are stabilized to within about 3 MHz using a high-precision wavelength meter (HighFinesse WS-7). The blue is then modulated using a resonant EOM at 8.037 GHz (New Focus model 4851), while the red has its drive current modulated directly with a bias-T. The two are then combined using a dichroic mirror and coupled into a single mode fiber, which provides colinearization and mode cleaning. The fiber also allows us to send the cooling laser light into another room where the active ion trap presently resides. A small fraction, about 10 $\mu$W, of 493 nm light is split off after the EOM for the optical pumping. The 1762 nm fiber laser is output onto the second table, where it is stabilized using a high finesse Zerodur cavity with 500 MHz free spectral range suspended in a vacuum chamber which is temperature stabilized to within about 10 mK. The transmitted laser intensity is monitored to maintain the lock. The adiabatic passage is achieved by driving an AOM with a linear analog ramp, ensuring maximum adiabaticity. To load Ba into the trap, we heat a sample of metallic barium contained in an alumina cylinder to several hundred degree Celsius to create an atomic beam and then use a Xe flash lamp to photoionize. The abundance of $^{137}$Ba is sufficiently high that one or two ions can be reliably trapped without the need for isotope selective photoionization.
Current Status
==============
After qubit state preparation by optical pumping we drive Rabi flops by applying microwaves and detect the final state with the 1762 nm laser as described above. The resulting detected Rabi flops are shown in Fig. \[rabiflops\]. At the time of publication, the 1762 nm laser stabilization was not complete, so that the 1762 frequency sweep crossed several transitions, resulting in only 60% efficient population transfer from the $S_{1/2}$ (F=2, $\textrm{m}_F$=0) to the $D_{5/2}$ (F=3) and (F=4) manifold. It should be noted that without optical pumping the lowest possible shelving probability would be 12.5% since all ground state $\textrm{m}_F$ levels would be populated evenly, and so this data simultaneously illustrates optical pumping.
We have also recently demonstrated a full optical Rabi flop on a single $^{138}$Ba$^+$ using a single ultrafast laser pulse of 400 fs duration on the $S_{1/2}$ to $P_{3/2}$ transition at 455 nm [@Kurz2008]. This allowed the measurement of the branching ratios of the $P_{3/2}$ state with high precision. Branching ratio measurements provide an experimental test of computational models of Ba$^+$ atomic structure, which are important for the parity non-conservation test [@Fortson1993]. Also, when compared against astronomical measurements of the branching ratio, it provides a bound on prehistoric variations in $\alpha$. The ultrafast excitation of single Ba ions paves the way to the ion-photon [@Blinov2004] and remote ion-ion [@Moehring2007] entanglement.
![Probability of shelving as a function of microwave exposure time. The Rabi frequency here is 10.1 kHz. The maximum shelving probability is limited by the stability of the shelving laser, and the minimum by optical pumping efficiency and microwave frequency detuning. The minimum shelving observed was 4.8%, and the contrast is 54%. $\chi^2_\nu=.55$ for this fit, with $\nu=13$. The error bars are statistical. The dashed line indicates the 12.5% minimum which would be observed in the absence of optical pumping.[]{data-label="rabiflops"}](rabiflops)
Future Plans
============
If tuned to the $P_{1/2}$ transition, ultrafast Rabi flops could be made to apply a state dependent impulse on the ion, which would allow us to perform ultrafast gates, such as the García-Ripoll phase gate [@GarciaRipoll2003], on two ions in the same trap. The benefits of such a gate are its intrinsic speed and that it does not necessitate cooling to the ground state of motion. This, combined with a laser for stimulated Raman transitions between the hyperfine levels, will form the basis of quantum computing with barium.
A more immediate objective with Ba$^+$ is the remote entanglement of two ions seperated by a kilometer or more. Remote entanglement of ions is accomplished by first exciting each ion with the ultrafast laser and allowing it to spontaneously decay, resulting in a photon whose frequency or polarization state is entangled with the final spin state of the relaxed ion. Once this is done with two separate ions, those photons can be jointly measured in the appropriate parity basis, resulting in an entangled state between the two ions. Entanglement between a photon and ion was originally seen in Cd$^+$ [@Blinov2004] and between two ions just recently in Yb$^+$ [@Moehring2007]. At present, the distance of entanglement is partly limited by the length that short wavelengths of light travel through an optical fiber, another potential advantage of Ba$^+$, since its cooling wavelength is longer than these other examples. The ability to remotely entangle qubits has application not only for quantum repeaters [@Briegel1998; @Duan2004b] but also for loophole free Bell inequality tests [@Simon2003]. Necessary to the success of such an experiment is the ability to perform every operation very quickly, on the $\mu$s time scale. In order to achieve such time scales, the power of the 1762 nm laser will have to be amplified, and the collection efficiency of the ion florescence will have to be increased. The former task can be accomplished using a Tm based fiber amplifier presently under development. The latter task will involve a high numerical aperture light collecting mirror placed inside the vacuum, near the trap itself. This will greatly increase the light collection solid angle, and thereby decrease the bright/dark discrimination time, which is currently limited to a ms. This modified trap design is presently in testing.
Acknowledgements
================
We would like to give special thanks to Sanghoon Chong, Tom Chartrand, Adam Kleczewski, Viki Mirgon, Joseph Pirtle and Edan Shahar for their various contributions to the experiment. This work was supported by NSF AMO program, the ARO DURIP grant, and the University of Washington Royalty Research Fund.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Recognition of identity documents using mobile devices has become a topic of a wide range of computer vision research. The portfolio of methods and algorithms for solving such tasks as face detection, document detection and rectification, text field recognition, and other, is growing, and the scarcity of datasets has become an important issue. One of the openly accessible datasets for evaluating such methods is MIDV-500, containing video clips of 50 identity document types in various conditions. However, the variability of capturing conditions in MIDV-500 did not address some of the key issues, mainly significant projective distortions and different lighting conditions. In this paper we present a MIDV-2019 dataset, containing video clips shot with modern high-resolution mobile cameras, with strong projective distortions and with low lighting conditions. The description of the added data is presented, and experimental baselines for text field recognition in different conditions.
The dataset is available for download at <ftp://smartengines.com/midv-500/extra/midv-2019/>.
author:
- |
Konstantin Bulatov, Daniil Matalov, Vladimir V. Arlazarov Federal Research Center “Computer Science and Control” of Russian Academy of Sciences, Moscow, Russia;\
Smart Engines Service LLC, Moscow, Russia
bibliography:
- 'bibliography.bib'
title: 'MIDV-2019: Challenges of the modern mobile-based document OCR'
---
Introduction
============
Usage of smartphones and tablet computers for solving business process optimization problems in enterprise systems, as well as processes in government systems, lead to a new development turn for computer vision systems operating on mobile devices. The increased interest in implementing corporate workflow management using mobile documents processing, and the necessity of entering document data in uncontrolled conditions elevate the requirements for document recognition, entry, and analysis systems which use mobile devices [@mollah_design; @google-doc-10].
The images obtained using mobile cameras have a range of specific properties and distortions, such as low resolution (especially for low-end smartphones and tablet computers), insufficient or inconsistent lighting, blur, defocus, highlights on reflective surfaces of the objects of interest, and others [@arl-small-scale-cameras]. Such properties increase the requirements for mobile optical recognition systems and brings necessity to the development of new methods and algorithms which would be more robust against such distortions. This particularly concerns the models and methods of optical recognition of objects in camera-based environments, autonomous methods which can work in isolated mobile computational systems (and thus dealing with constrained computational power) [@7994570; @Takhirov:2016:EAC:2934583.2934615; @Yanai:2016:EMI:2964284.2967243], and methods for analyzing video stream input in real time [@8270252; @vestnik_integration]. To facilitate the research on these topics adequate open datasets should be created and maintained.
A particular interest in the field of mobile computer vision systems is given to the task of identity document recognition [@8283074; @SMART_IDREADER_ICDAR]. Automatic entry of data from identity documents is used in such industries as fintech, banking, insurance, travel, e-government, and in such processes as user identification and authentication, KYC/AML (Known Your Customer / Anti-Money Laundering) procedures and others. Computer vision problems which are associated with automatic identity document entry using mobile devices, include:
1. Determining the document class, type, subtype, or country of issue;
2. Document boundaries detection in an image, or document page segmentation from the background;
3. Per-field document segmentation and layout analysis;
4. Personal photo detection or facial features extraction;
5. Optical character recognition, capturing and recognition of text fields and properties of the document;
6. Video stream analysis in real time;
7. Image quality estimation;
8. Security features detection, optical variable devices analysis (holograms, dynamic color embossing, etc.);
9. Other related tasks.
An important issue which comes up in relation to research and scientific publications on the topic of identity document processing is the availability of datasets. Identity documents contain sensitive personal information, so storing, transmitting or otherwise make this data public is impossible. In order to facilitate research in some of the topics mentioned above, the MIDV-500 dataset was introduced [@midv500-arxiv]. The dataset contained video clips of 50 identity documents with different types. Since it’s impossible to create a public dataset of valid and authentic identity documents, the dataset contained mostly “sample” or “specimen” documents which could be found in WikiMedia and which were distributed under public copyright licenses. And thus, although the variability of the documents used in the dataset is comparatively low, the target objects featured in this dataset shared the common identity document features.
The MIDV-500 dataset contained 500 video clips, 10 clips per document type. The clips were captured using Apple iPhone 5 and Samsung Galaxy S3 (GT-I9300), the smartphone models which could already be considered obsolete by the time the dataset was published. However, the increasingly common usage of identity document recognition in various business or government processes implies the need to support a wide range of devices, from cheap low-end devices to the “flagship” models. Using each of the two smartphone models, each document was shot in five distinct conditions: “Table”, “Keyboard”, “Hand”, “Partial”, and “Clutter”. The “Table” condition represented the simplest case with the document laying on the table with homogeneous surface texture. The “Keyboard” represented the case when the document lays on various keyboards thus making it harder to utilize conventional edge detection methods, because of the background cluttered with straight edges and text. The “Hand” condition represented the case of a hand-held document. The “Partial” condition had some frames when the document is partially or completely hidden outside the camera frame. Finally, the “Clutter” condition had the background intentionally cluttered with random objects. The conditions represented in MIDV-500 dataset had some diversity regarding the background (“Table”, “Keyboard”, “Hand”, and “Clutter”) and the positioning of the document in relation to the capturing process (“Partial”), however they did not include variation in lighting conditions, significant projective distortions, or a variation in the quality of the camera. Example images of every condition presented in MIDV-500 are illustrated in Figure \[fig:midv500\_samples\].
![Conditions of the original MIDV-500 dataset [@midv500-arxiv]. From left to right: “Table”, “Keyboard”, “Hand”, “Partial”, and “Clutter”[]{data-label="fig:midv500_samples"}](fig1.jpg){width="\textwidth"}
The MIDV-500 dataset has been used to perform research of methods for document type recognition using similarity metric aimed at high classification precision and robustness against projective distortions [@lynchenko2019document], for evaluation of image quality assessment methods and their impact on the document recognition system performance [@chernov2019application], for analyzing the per-frame text field recognition results combination in a video stream [@vestnik_integration] and construction a stopping rule for text string recognition in a video stream [@Bulatov2019].
Even though modern smartphones and tablet computers have camera modules with substantially higher quality, and their computational power have increased drastically, recognition of identity documents in images or a video stream in uncontrolled capturing conditions remains a scientific and technological challenge. In many use cases the documents are captured not by trained personnel, but remotely by document holders who are likely to be performing such capture very rarely and do not have any information about the processing algorithms involved. Thus, such conditions as low scene lighting, high projective distortions, and other complications, lead to a demand of sophisticated processing techniques and methods, even if input images are captured with high-end mobile devices.
In this paper we present an extension to the MIDV-500 dataset, called MIDV-2019, which consists of video clips of the 50 original identity documents, but shot with more complex conditions and using high-end smartphone cameras. The two complex conditions targeted in this dataset extension are low lighting and strong projective distortions. The prepared extension is aimed to provide a platform for creating and evaluation of new methods and algorithms, designed to operate in challenging environments.
Dataset description {#sec:dataset}
===================
As with the MIDV-500 dataset, the new dataset MIDV-2019, presented in this paper, contains video clips of 50 different identity document types, which includes 17 ID cards, 14 passports, 13 driving licences and 6 other identity documents of different countries. The same printed samples which were used as a source of MIDV-500 dataset were also used to prepare MIDV-2019. For each printed document video clips were recorded under two different capturing conditions and using two mobile devices, thus obtaining 4 new video clips per document (200 new video clips in total). New clip identifiers are described in Table \[tbl:new\_conditions\]. Sample images of the added conditions are presented in Figure \[fig:midv2019\_samples\].
Identifier Description
------------ ---------------------------------------------------------------------------------------------------------------------------------
DG “istorted” – documents were shot with higher projective distortions, videos were captured using Samsung alaxy S10 (SM-G973F/DS)
DX “istorted” – documents were shot with higher projective distortions, videos were captured using Apple iPhone S Max
LG “ow-lighting” – documents were shot in very low lighting conditions, video were capturing using Samsung alaxy S10 (SM-G973F/DS)
LX “ow-lighting” – documents were shot in very low lighting conditions, video were capturing using Apple iPhone S Max
: Clip types added in MIDV-2019[]{data-label="tbl:new_conditions"}
![Conditions of the MIDV-2019 dataset. “Distorted” condition (left) and “Low-lighting” condition (right)[]{data-label="fig:midv2019_samples"}](fig2.jpg){width="\textwidth"}
The first new capturing condition introduced in the MIDV-2019 dataset is the “Distorted” condition (clips “DG” and “DX”), in which the documents were shot with strong projective distortions. The requirement for the document recognition systems to be able to operate in uncontrolled condition sometimes lead to the users capturing the documents with high projective distortions – for example, to avoid highlights on the reflective surfaces of the document. The methods which perform preliminary document detection and localization try to rectify the document image prior to processing, however it is important to have a highly distorted dataset of samples in order to assess the limits to the applicability of such methods. For methods which perform text segmentation and recognition without prior rectification [@HE2019107026; @ElBahi2019], and, specifically, the text components of identity documents, such as machine-readable zones [@SkoriukinaMRZ], such capturing conditions may provide a valuable reference.
Perhaps the most significant challenge added in the MIDV-2019 dataset is the clips shot in a low-lighting conditions without flash. Such use cases as checking the identity documents in a long-distance travel, using mobile systems to enter identity document data by law enforcement officials, and others, sometimes require the ability to recognize documents with very low ambient light. In the images thus obtained the text is still visible and could be discerned by a human, however modern OCR systems struggle with this task. This is the primary reason for adding the clips “LG” and “LX”, which represents the “Low-lighting” condition, to the MIDV-2019 dataset. The examples of text field images cropped from frames captured in low-lighting condition are presented in Figure \[fig:low\_lighting\_fields\].
![Examples of text fields cropped from “LG” and “LX” clips[]{data-label="fig:low_lighting_fields"}](fig3.png){width="50.00000%"}
All clips were shot in Ultra HD resolution (2160x3840). Each video were at least 3 seconds in duration and the first 3 seconds of each video were split with 10 frames per second. As in the MIDV-500 dataset, for each frame the ideal coordinates of the document’s boundaries were annotated by hand, and if the corners of the document were not visible on the frame, the corresponding coordinate points were extrapolated outside the frame. The provided document coordinates combined with an ideal template segmentation ground truth, which is provided with the original MIDV-500, allow to crop document elements, such as text fields, from each frame, and perform evaluation of algorithms for face detection, text field recognition, as well as full document detection, classification, and segmentation.
Evaluation baselines
====================
In order to provide basic baselines for future experiments based on the presented MIDV-2019 dataset we performed text field recognition evaluation using an open-source recognition system Tesseract v4.1.0 [@tesseract-paper]. As in the original paper presenting MIDV-500, four field groups were analyzed: numeric dates, document numbers, machine-readable zone lines, and Latin name components (which contain only Latin characters with no diacritical marks). Only the frames on which the document boundaries laid fully inside the frame were considered. To be consistent with MIDV-500, all fields were cropped with the resolution of 300 DPI (achieved using known physical dimensions of all document types present in the dataset), and for each field a margin was allowed with width equal to 10% of the minimal dimension of the field’s bounding box.
Table \[tbl:fields\_total\] lists the number of field images thus extracted from both MIDV-500 clips and from MIDV-2019 clips, grouped by the capturing condition. The first five rows of the table represent conditions in MIDV-500 dataset (“Table”, “Keyboard”, “Hand”, “Partial”, and “Clutter”). The last two rows represent conditions from the new MIDV-2019 dataset (“Distorted” and “Low-lighting”).
In Table \[tbl:fields\_accuracy\] the text field recognition accuracy is presented for all aforementioned conditions and grouped by the type of the text field. The comparison of the recognized and correct values was case-insensitive and there were no distinction between the Latin letter “O” and the digit “0”. While given the fixed recognition system the absolute values of the recognition accuracy is of a lesser interest, the main distinction which should be noted is between the capturing conditions. From the seven analyzed conditions the highest quality is achieved in the simplest case – the one with document laying on the table (“TA”, “TS”) with homogeneous background and with the smallest projective distortion. Even though the clips “TA” and “TS” were shot with older smartphone models and with Full HD resolution (1080x1920), the recognition precision for images taken in those conditions turned out to be higher than that of clips “DG” and “DX”, which were shot with higher projective distortions but with Ultra HD resolution (2160x3840).
By far the lowest text field recognition accuracy can be seen on the “Low-lighting” clips “LG” and “LX”. Even if shot with high resolution and on modern smartphones, the task of document recognition in such conditions is still a clear challenge and should be addressed by the community. It should be noted that while the recognition accuracy on the “Low-lighting” clips is very low, the accuracy ordering by text field group is mostly the same as for the other conditions.
Conclusion
==========
In this paper we presented the dataset MIDV-2019 containing video clips of identity documents captured using modern smartphones in low lighting conditions and with higher projective distortions. The paper presents experimental baselines for text field recognition for different capturing conditions and for different field groups presented in the dataset, and the reported result show that the text field recognition in low lighting is still a very challenging problem for modern mobile recognition systems. With the added data the MIDV-500 dataset is expanded by 40%.
Authors believe that the provided dataset will serve as a valuable resource for document recognition research community and lead to more high-quality scientific publications in the field of identity documents analysis, as well as in the general field of computer vision.
This work is partially financially supported by Russian Foundation for Basic Research (projects 17-29-03170 and 17-29-03236). Source images for MIDV-2019 dataset are obtained from Wikimedia Commons. Author attributions for each source image are listed in the description table at <ftp://smartengines.com/midv-500/documents.pdf>.
| {
"pile_set_name": "ArXiv"
} |
---
address: |
$^{\star}$Singapore University of Technology and Design (SUTD)\
$^{\dagger}$The University of Adelaide
bibliography:
- 'hash.bib'
title: 'Supervised Hashing with End-to-End Binary Deep Neural Network'
---
deep neural network, learning to hash, optimization, image retrieval
Acknowledgements {#acknowledgements .unnumbered}
================
This research is supported by the National Research Foundation Singapore under its AI Singapore Programme (Award number: AISG-100E-2018-005)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper we study spaces of holomorphic functions on the right half-plane $\cR$, that we denote by $\cM^p_\omega$, whose growth conditions are given in terms of a translation invariant measure $\omega$ on the closed half-plane $\ov\cR$. Such a measure has the form $\omega=\nu\otimes m$, where $m$ is the Lebesgue measure on ${{\mathbb R}}$ and $\nu$ is a regular Borel measure on $[0,+\infty)$. We call these spaces [*generalized Hardy–Bergman*]{} spaces on the half-plane $\cR$.
We study in particular the case of $\nu$ purely atomic, with point masses on an arithmetic progression on $[0,+\infty)$. We obtain a Paley–Wiener theorem for $\cM^2_\omega$, and consequentely the expression for its reproducing kernel. We study the growth of functions in such space and in particular show that $\cM^p_\omega$ contains functions of order 1. Moreover, we prove that the orthogonal projection from $L^p(\cR,d\omega)$ into $\cM^p_\omega$ is unbounded for $p\neq2$.
Furthermore, we compare the spaces $\cM^p_\omega$ with the classical Hardy and Bergman spaces, and some other Hardy–Bergman-type spaces introduced more recently.
address: 'Dipartimento di Matematica, Università degli Studi di Milano, Via C. Saldini 50, 20133 Milano, Italy'
author:
- 'Marco M. Peloso'
- Maura Salvatori
bibliography:
- 'm2-mathscinet-1.bib'
title: 'On some spaces of holomorphic functions of exponential growth on a half-plane'
---
[^1] [^2]
Introduction
============
This paper is concerned with spaces of holomorphic functions on the right half-plane $\cR$ whose growth condition is given in terms of a translation invariant regular measure on $\ov\cR$, and that can be defined as generalized Hardy–Bergman spaces. It is easy to see that a measure $\omega$ is translation invariant on $\ov\cR$ if and only if it has the form $\omega= \nu\otimes m$, where $m$ denotes the Lebesgue measure on ${{\mathbb R}}$ and $\nu$ is a measure on $[0,+\infty)$. We will simply write $dm(y)=dy$. These measures play the same role as the radial measures on the unit disk ${{\mathbb D}}$.
Holomorphic function spaces on $\cR$ with integrability conditions given in terms of this type of measures have been studied by several authors, and here we mention in particular Z. Harper [@Harper-Laplace; @Harper-boundedness], B. Jacob, S. Pott and J. Partington [@Jacob-Partington-Pott1], and I. Chalendar and J. Partington [@Chalendar-Partington], and our recent paper [@PS]. We will come back to the spaces these authors considered in Section \[comparison\].
For $0<a<b<\infty$, denote by $S_{a,b}$ the vertical strip $\{z=x+iy: a<x<b\}$ and by $H^p(S_{a,b})$ the classical Hardy space $$H^p(S_{a,b}) =\big\{ f\text{ holomorphic in } S_{a,b}:\,
\sup_{a<x<b} \int_{-\infty}^{+\infty} |f(x+iy)|^p\, dy
<\infty\big\}\,.$$
We simply write $S_b$ to denote $S_{0,b}$.
\[mix-spaces-def\]
Let $1\le p<\infty$. Let $\omega=\nu\otimes dy$ be a translation invariant measure on $\ov\cR$.
When $\nu(\{0\})=0$, we define $$\label{Mp-omega-def1}
\cM^p_\omega(\cR)
= \big\{ f\in \Hol(\cR):\, f\in H^p(S_{a,b}) \text{ for all }
0<a<b<\infty,\ \text{and } f \in L^p(\cR, d\omega)\big\}\, ,$$ and if $\nu(\{0\})>0$, we define $$\label{Mp-omega-def2}
\cM^p_\omega(\cR)
= \big\{ f\in \Hol(\cR):\, f\in H^p(S_b) \text{ for all }
0<b<\infty,\ \text{and } f \in L^p(\ov\cR, d\omega)\big\} \, .$$ In both cases we endow $\cM^p_\omega(\cR)$ with the norm $$\| f\|_{\cM^p_\omega(\cR)}^p =
\int_0^{+\infty} \int_{{\mathbb R}}|f(x+iy)|^p\, dyd\nu(x) \, .$$ We call the spaces $\cM^p_\omega(\cR)$ [*generalized Hardy–Bergman spaces*]{} on the half-plane $\cR$.
We point out that the definition implies that in the case $\nu(\{0\})>0$, a function in $\cM^p_{\omega}$, although not initially defined on the imaginary line, admits a boundary value function that in particular is in $L^p(dy)$.
In this paper we consider in particular the measures of the form $$\label{omega-arho-def}
\omega_{a,\rho} =\nu_{a,\rho}(x) \otimes dy
=\sum_{k=0}^{+\infty} \frac{a^n}{n!} \delta_{\rho\frac{n}{2}}(x) \otimes dy\, ,$$ $a,\rho>0$ are fixed parameters. Thus, the measure $\omega_{a,\rho}$ is translation invariant in $\ov\cR$, has purely atomic part $\nu_{a,\rho}$ in the $\Re z$-component and moreover, such atomic measure $\nu_{a,\rho}$ has support on an arithmetic progression $\{ \frac\rho2
n\}$ with weight $\frac{a^n}{n!}$ at the point-mass $\frac\rho2 n$.
Explicitly, $\cM^p_{\omega_{a,\rho}}(\cR)$ is the space of holomorphic functions on $\cR$ that belong to $L^p(\ov\cR,d\omega_{a,\rho}) \cap \big(
\bigcap_{b>0} H^p(S_b))$ with norm $$\label{L2-omega-norm}
\|f\|_{\cM^p_{\omega_{a,\rho}}(\cR)}^p =
\sum_{n=0}^{+\infty} \frac{a^n}{n!} \int_{-\infty}^{+\infty}
|f\big(\rho\tnt +iy\big)|^p\, dy <\infty \,,$$ where $f$ is defined on the imaginary axis as its boundary values as function in $H^p(S_b)$.
We observe that we equivalently could define $\cM^p_{\omega_{a,\rho}}(\cR)$ as the closure in $L^p(\ov\cR,d\omega_{a,\rho})$ of the space $\Hol(\ov\cR)\cap \big( \bigcap_{b>0} H^p(S_b))$ with norm given by . Moreover, it suffices to require that $f\in \bigcap_{n>0}
H^p(S_{\rho\frac{n}{2}})$.
From simplicity of notation, we write $\cM^p_{a,\rho}$ in place of $\cM^p_{\omega_{a,\rho}}(\cR)$.
In [@PS] we introduced and studied the space $\cM^2_{\omega_{2,1}}=\cM^2_{2,1}$ to give some necessary and sufficient conditions for the solutions of the problem for the Bergman space (on the disk $\Delta=D(1,1)$). Such problem was first stated and studied by S. Krantz, C. Stoppato and the first named author [@KPS2], in connection with the question of completeness in the Bergman space on some special domain in ${{\mathbb C}}^2$, the so-called [*Diedriech–Fornæss worm domain*]{}.
In Section \[Muntz-Szasz\] we recall the main results of [@PS] (without proofs) to motivated the analysis of the slightly more general spaces $\cM^2_{a,\rho}$. In Section \[M2-arho-sec\] we prove some of these extensions and, at the same time, show that some of main properties of $\cM^2_{2,1}$ can not so easily generalized to the larger class, thus raising some natural questions.
In Section \[comparison\] we take a look at other Hardy–Bergman type spaces studied in particular in [@Harper-Laplace; @Harper-boundedness], [@Jacob-Partington-Pott1], and [@Chalendar-Partington] and sometimes called [*Zen spaces*]{}. These spaces are denoted by $A^p_\omega$. We prove a Paley–Wiener theorem for $A^2_\omega$ and, as a consequence, give a description of the reproducing kernel of $A^2_\omega$. We also clarify some properties of $A^p_\omega$, and in particular discuss how functions in such spaces lie in the Hardy space $H^p(\cR_a)$, for every $a>0$, where $\cR_a$ is the half-plane $\{\Re z>a\}$. In this section we also present some examples of holomorphic functions and norms on $\cR$ to illustrate some of the possible behaviors of the average function $a_{f,p}(x)=\int_{{\mathbb R}}|f(x+iy)|^p\, dy$, for $f\in \Hol(\cR)$.
We conclude the paper with some remarks and open questions.
The Müntz–Szász problem for the Bergman space {#Muntz-Szasz}
=============================================
Goal of this section is to motivate the study of the spaces $\cM^p_{a,\rho}$ by recalling the main results from [@PS], concerning the special case $\cM^2_{2,1}$.
Let $\Delta$ be the disk $\{\z:\, |\z-1|<1\}$, denote by $dA$ the Lebesgue measure in ${{\mathbb C}}$ and consider the (unweighted) Bergman space $A^2(\Delta)$. Then the complex powers $\{\zeta^{\lambda-1}\}$ with $\Re\lambda>0$ are well defined and in $A^2(\Delta)$.
Following [@KPS2], the problem for the Bergman space is the question of characterizing the sequences $\{\lambda_j\}$ in $\cR$ such that $\{\z^{\lambda_j-1}\}$ is a complete set in $A^2(\Delta)$, that is, $\operatorname{span}\{\z^{\lambda_j-1}\}$ is dense in $A^2(\Delta)$.
The classical theorem concerns with the completeness of a set of powers $\{t^{\lambda_j-\frac12}\}$ in $L^2\big([0,1]\big)$, where $\Re\lambda_j>0$. The solution was provided in two papers separate by Müntz [@Muntz] and by Szász [@Szasz] where they show that the set $\{t^{\lambda_j-\frac12}\}$ is complete $L^2\big([0,1]\big)$ if and only if the sequence $\{\lambda_j\}$ is a set of uniqueness for the Hardy space of the right half-plane $H^2(\cR)$, that is, if $f\in H^2(\cR)$ and $f(\lambda_j)=0$ for every $j$, then $f$ is identically $0$.
As in the classical case, in order to study the problem for the Bergman space we transformed the question into characterizing the sets of uniqueness for some (Hilbert) space of holomorphic functions. In [@PS] we showed that $\{\z^{\lambda_j-1}\}$ is complete in $A^2(\Delta)$ if and only if $\{\z^{\lambda_j-1}\}$ is a set of uniqueness for $\cM^2_{2,1}$ and found some sufficient and some necessary conditions. We now outline the most relevant results of [@PS].
For $f\in A^2(\Delta)$ and $z\in\cR$ we define the Mellin–Bergman transform $$\label{M-Bt}
M_\Delta f(z) = \frac{1}{\pi} \int_\Delta f(\z) \ov \z^{z-1}\, dA(\z)\, .$$ The function $\z^{\ov z-1}$ is well defined and belongs to $A^2(\Delta)$. Then a set $\{\z^{\lambda_j-1}\}$ is complete in $A^2(\Delta)$ if and only if $f\in A^2(\Delta)$ and $M_\Delta f(\ov\lambda_j)=0$ for all $j$ implies that $f$ vanishes identically. In order to describe the image of $A^2(\Delta)$ under the Mellin–Bergman transform $M_\Delta$ consider the space $$\label{H}
\cH = \big\{ g\in \Hol(\cR):\,
\textstyle{\frac{\Gamma(1+z)}{2^z}} g(z) \in \cM^2_{2,1} \big\} \, ,$$ with norm $$\begin{aligned}
\|g\|_\cH^2 & = \Big\| \frac{\Gamma(1+z)}{2^z} g
\Big\|_{\cM^2_{2,1}}^2
=
\sum_{n=0}^{+\infty} \int_{-\infty}^{+\infty} |g({\tnt}+iy)|^2
\frac{|\Gamma({\tnt}+1+iy)|^2}{\Gamma(n+1)} \, dy \,. \end{aligned}$$
\[new\] The Mellin–Bergman transform $$M_\Delta: A^2(\Delta)\to \cH$$ is a surjective isomorphism. The space $\cH$ consists of holomorphic functions on $\cR$ that are of exponential type at most $\pi/2$ and the polynomials are dense in $\cH$. Moreover, it is a Hilbert space with reproducing kernel $$H(z,w) = \frac{1}{2\pi} \frac{\Gamma(z+\ov w)}{\Gamma(1+z)
\Gamma(1+\ov w)} \, .$$
We point out that, as corollary of the proof, we obtain a remarkable factorization theorem for functions in $\cM_{2,1}^2$. It would be interesting to prove, if it exists, a similar factorization theorem for $\cM_{a,\rho}^2$.
\[M2-fact\] We have that $$\cM_{2,1}^2 =\frac{\Gamma(1+z)}{2^z} \cH\, .$$
We remark that $\cH$ consists of functions that of exponential type $\pi/2$ on $\cR$. Using this factorization we obtained a formula of Carleman type for functions in $\cM_{2,1}^2$.
Recall that the exponent of convergence of a sequence $\{z_j\}$, with $|z_j|\to +\infty$, is $\rho_1= \inf \{\rho>0:\, \sum_{j=1}^{+\infty} 1/|z_j|^\rho
<\infty\}$, the counting function is $n(r) = \# \{ z_j:\, |z_j|\le r\}$ and the upper and lower densities $d^\pm=d_{\{z_j\}}^\pm$ are $$d^+ = \limsup_{r\to+\infty} \frac{n(r)}{r^{\rho_1}}\,, \quad
d^- = \liminf_{r\to+\infty} \frac{n(r)}{r^{\rho_1}}\,.$$
We are now in the position to state a necessary and a sufficient condition for zero-sets of $\cM^2_{2,1}$.
\[zero-set-thm\] Let $\{z_j\}\subseteq\cR$, $1\le|z_j|\to+\infty$. The following properties hold.
- If $\{z_j\}$ has exponent of convergence $1$ and upper density $d^+<\frac12$, then $\{z_j\}$ is a zero-set for $\cM^2_{2,1}\cap\Hol(\ov\cR)$.
- If $\{z_j\}$ is a zero-set for $\cM^2_{2,1}\cap\Hol(\ov\cR)$, then $$\label{our-Carleman-cond}
\limsup_{R\to+\infty} \frac{1}{\log R} \sum_{|z_j|\le R} \Re\big(1/z_j\big) \le
\frac2\pi \,.$$
The next result gives a partial solution to the problem for the Bergman space.
\[MS-thm\] A sequence $\{z_j\}$ of points in $\cR$ such that $\Re z_j\ge \eps_0$, for some $\eps_0>0$ and that violates condition , is a set of uniqueness for $\cM^2_{2,1}(\cR)$.
As a consequence, if $\{z_j\}$ is a sequence as above, the set of powers $\{ \z^{z_j-1}\}$ is a complete set in $A^2(\Delta)$.
The spaces $\cM^2_{a,\rho}$ {#M2-arho-sec}
===========================
In this section we study the basic properties of $\cM^2_{a,\rho}$. In particular we prove a Paley–Wiener type theorem that allows us to compute its reproducing kernel. We also prove that the Mellin transform is a a surjective isometry between a suitable $L^2$ space on the positive half-line and $\cM^2_{a,\rho}$.
Notice that trivially $H^p(\cR)$ is a closed subset of $ \cM^p_{a,\rho}$ and that $\cM^p_{a,\rho}$ is closed in $L^p(\ov\cR, d\omega)$.
The Paley–Wiener type theorem and its consequences {#PW-sec}
--------------------------------------------------
We being by proving a characterization of $\cM^2_{a,\rho}$ in terms of the Fourier transform of its boundary values, as in the spirit of the classical Paley–Wiener theorem.
The Fourier transform of a function $\psi\in L^1({{\mathbb R}})$ is $$\widehat \psi (\xi) = \frac{1}{\sqrt{2\pi}} \int_{-\infty}^{+\infty}
\psi(x) e^{-ix\xi}\, dx\,.$$
For $f\in \cM^2_{a,\rho} $, we write $f_0=f(0+i\cdot)$ to denote its boundary values on the imaginary axis. We recall that the classical Paley–Wiener theorem for $\cH^2(\cR)$ establishes a surjective isomorphisms between $\cH^2(\cR)$ and $L^2((-\infty,0),d\xi))$.
\[PW-thm\] Let $f\in \cM^2_{a,\rho}$. Then $\widehat f_0 \in L^2({{\mathbb R}},e^{ae^{\rho\xi}}d\xi)$ and $$\label{PW-cA2-equality}
\|f\|_{\cM^2_{a,\rho}} = \|\widehat f_0
\|_{L^2({{\mathbb R}},e^{ae^{\rho\xi}})} \,.$$
Conversely, if $\psi\in L^2({{\mathbb R}},e^{ae^{\rho\xi}}d\xi)$ and for $z\in\cR$ we set $$\label{PW-cA2-def}
f(z) =\frac{1}{\sqrt{2\pi}} \int_{-\infty}^{+\infty} \psi(\xi)
e^{z\xi}\, d\xi\, ,$$ then $f\in \cM^2_{a,\rho}$, equality holds and $\psi=\widehat f_0$.
Since $f\in H^2(S_{\rho\frac{n}{2}})$ for every $n$, by the classical Paley–Wiener theorem on a strip ([@PW], see also [@PS Thm. 1.1]), we see that $$\begin{aligned}
\|f\|_{\cM^2_{a,\rho}}^2
& = \sum_{n=0}^{+\infty} \frac{a^n}{n!} \|
f(\rho\tnt+i\cdot)\|_{L^2({{\mathbb R}})}^2 \notag \\
& = \sum_{n=0}^{+\infty} \frac{a^n}{n!} \big\|
\big[f(\rho\tnt+i\cdot)\big]\widehat{\ }\,\big\|_{L^2({{\mathbb R}})}^2 \notag \\
& = \sum_{n=0}^{+\infty} \frac{a^n }{n!} \int_{-\infty}^{+\infty}
e^{n\rho\xi}|\widehat f_0(\xi)|^2 \, d\xi \notag \\
& = \int_{-\infty}^{+\infty}
|\widehat f_0(\xi)|^2 e^{ae^{\rho\xi}}\, d\xi \label{one-more-PW}
\,.\end{aligned}$$
Conversely, given $\psi\in L^2({{\mathbb R}},e^{ae^{\rho\xi}}d\xi)$ observe that the integral in is absolutely convergent for $z\in\cR$: $$\label{pt-eval}
\int_{-\infty}^{+\infty} |\psi(\xi) e^{z\xi}|\, d\xi
\le \|\psi\|_{L^2({{\mathbb R}},e^{ae^{\rho\xi}}d\xi)} \bigg(\int_{-\infty}^{+\infty}
e^{2x\xi} e^{-ae^{\rho\xi}}\, d\xi \bigg)^{1/2}<\infty\,.$$ Therefore, if $f$ is given by it is holomorphic in $\cR$ and $f\in H^2(S_{\rho\frac{n}{2}})$ for every $n$, since $e^{\rho\frac{n}{2}(\cdot)}\psi\in L^2({{\mathbb R}})$. It is also clear that $\widehat f_0=\psi$ and arguing as for we obtain .
The space $\cM^2_{a,\rho}$ is a reproducing kernel Hilbert space and its reproducing kernel is $$K(z,w) = \frac{1}{2\pi\rho} a^{-\frac{z+\ov w}{\rho}}
\, \Gamma\Big(\frac{z+\ov w}{\rho} \Big)
\,.$$
By it follows that point evaluations are continuous in $\cM^2_{a,\rho}$ and it is elementary to see that it is a Hilbert space.
Let $K_z\in \cM^2_{a,\rho}$ be such that $\la f,\,
K_z\ra_{\cM^2_{a,\rho}}=f(z)$ for every $z\in\cR$ and every $f\in
\cM^2_{a,\rho}$. By Theorem \[PW-thm\] and we obtain $$\begin{aligned}
f(z)
& = \sum_{n=0}^{+\infty} \frac{a^n}{n!} \big\la f(\rho\tnt+i\cdot),\,
K_z(\rho\tnt+i\cdot)\big\ra_{L^2} \\
& = \sum_{n=0}^{+\infty} \frac{a^n}{n!} \big\la
\big[f(\rho\tnt+i\cdot)\big]\widehat{\ },\,
\big[K_z(\rho\tnt+i\cdot)\big]\widehat{\ }\big\ra_{L^2} \\
& = \sum_{n=0}^{+\infty} \frac{a^n }{n!} \big\la e^{\rho\frac{n}{2}\xi}\widehat f_0,\,
e^{\rho\frac{n}{2}\xi} \widehat K_{z,0} \big\ra_{L^2} \\
& = \int_{-\infty}^{+\infty} \widehat f_0(\xi)
\ov{\widehat {K_{z,0}}(\xi)}\, e^{ae^{\rho\xi}}\, d\xi\,,\end{aligned}$$ where switching the integral with the sum is justisfied since the last integral converges absolutely.
On the other hand, $$f(z) =\frac{1}{\sqrt{2\pi}} \int_{-\infty}^{+\infty} e^{z\xi}\widehat f_0(\xi)\, d\xi\,,$$ so that $$\widehat {K_{z,0}}(\xi)
= \frac{1}{\sqrt{2\pi}} e^{-ae^{\rho\xi}} e^{\ov z\xi} \,,$$ and $$\begin{aligned}
K_z(w)
&= \frac{1}{2\pi} \int_{{\mathbb R}}e^{w\xi} e^{-ae^{\rho\xi}} e^{\ov z\xi}\, d\xi\\
&= \frac{1}{2\pi\rho} a^{-\frac{w+\ov z}{\rho}} \int_0^{+\infty} t^{w+\ov z-1}
e^{-t} \, dt \\
& = \frac{1}{2\pi\rho} a^{-\frac{w+\ov z}{\rho}}
\, \Gamma\Big(\frac{w+\ov z}{\rho} \Big) \,. \qed \end{aligned}$$
It is now clear that $\cM^2_{a,\rho}$ contains functions of order 1 in the right half-plane, namely, for $w\in\cR$ $$K_w(z) = \frac{1}{2\pi\rho} a^{-\frac{z+\ov w}{\rho}}
\, \Gamma\Big(\frac{z+\ov w}{\rho} \Big) \,.$$ The next result shows that the growth of functions in $\cM^2_{a,\rho}$ is at most of order 1. Precisely,
Let be $a,\rho>0$ given. The functions in $\cM^2_{a,\rho}$ satisfy the growth condition $$|f(z)| \le C (\Re z)^{1/4} \Big( \frac{2}{a}\Big) ^ {(\Re z)/\rho} \Gamma\Big( \frac{\Re z}{\rho}\Big)\,.$$
We have $$\begin{aligned}
|f(z)|
& = \big|
\la \widehat{f_0},\, \widehat{K_{z,0}} \ra_{L^2({{\mathbb R}},e^{ae^{\rho\xi}})}
\big| \le C \| \widehat{K_{z,0}} \|_{L^2({{\mathbb R}},e^{ae^{\rho\xi}})} \\
& \le C \bigg( \int_{{\mathbb R}}e^{2
\Re z \xi} e^{-ae^{\rho\xi}} \, d\xi\bigg) ^{1/2} =C \frac{1}{a^{(\Re z)/\rho}}
\Gamma \Big( \frac{2\Re z}{\rho}\Big)^{1/2}
\\
& \le C (\Re z)^{1/4} \Big( \frac{2}{a}\Big) ^ {(\Re z)/\rho} \Gamma \Big( \frac{\Re z}{\rho}\Big)
\, , \end{aligned}$$ using the standard asymptotics of the Gamma function.
The next result is obvious.
Given $a,a'>0$ and $\rho,\rho'>0$ we have the continuous embedding $\cM^2_{a,\rho}\subseteq \cM^2_{a',\rho'}$ if and only if $\rho>\rho'$ for any $a,a'$ or if $\rho=\rho'$ and $a> a'$.
There is one more consequence of the Payley–Wiener type theorem about a density result in $\cM^2_{a,\rho}$. To state it we first need to introduce some further notation.
For every $\eps>0$ we denote by $\cM^p_{a,\rho}(\cR_{-\eps})$ the subspace of $\cM^p_{a,\rho}$ of functions that are holomorphic for $\Re z>-\eps$ and that are in $H^p(S_{(-\eps,b)})$ for every $b>0$.
The following proposition is proved as [@PS Prop. 3.4] and it is used in the next section.
For $1\le p,q<\infty$, the space $\bigcap_{\eps>0}\cM^p_{a,\rho}({\cR_{-\eps}})\cap \cM^q_{a,\rho}(\cR_{-\eps})$ is dense in $\cM^p_{a,\rho}$.
The Mellin transform {#Mellin-subsec}
---------------------
We want to show that the Mellin transform is a surjective isometry between $\cM^2_{a,\rho}$ and some suitable $L^2$ space of the positive half-line.
More precisely, if $\vp$ is a function defined on $(0,+\infty)$ we consider the (re-normalized) Mellin transform, that is $$\label{Mellin-def}
M\vp (z) = \frac{1}{\sqrt{2\pi}} \int_0^{+\infty} \vp(t) t^{z-1}\, dt\,.$$
\[isometry-thm\] The mapping $$M: L^2 \Big( (0,+\infty),\, e^{a\xi^{\rho}}\,
\frac{d\xi}{\xi}\, \Big) \to
\cM^2_{a,\rho}$$ is a surjective isometry.
Clearly, the main point of this result is the fact that $M$ is an isometry that is also surjective.
Suppose that $\vp \in L^2 \Big( (0,+\infty),\, e^{a\xi^{\rho}}\,
\frac{d\xi}{\xi}\, \Big) $, then for every $z=x+iy\in\cR$ $$\begin{aligned}
|M\vp(z)|
& \le \frac{1}{\sqrt{2\pi}} \int_0^{+\infty} |\vp(\xi)| \xi^{x-1}\, d\xi\\
& \le \frac{1}{\sqrt{2\pi}} \| \vp \|_{L^2 ( (0,+\infty), e^{a\xi^{\rho}}\,
\frac{d\xi}{\xi})} \Big(
\int_0^{+\infty} e^{-a\xi^{\rho}} \xi^{2x-1}\, d\xi \Big)^{1/2} \\
& = \frac{1}{\sqrt{2\pi \rho}} \| \vp \|_{L^2 ( (0,+\infty), e^{a\xi^{\rho}}\,
\frac{d\xi}{\xi})}a^{-\frac{x}{\rho}} \,\Gamma\big(\textstyle{ \frac{2x}{\rho}}\big)^{1/2}
\,. \end{aligned}$$ This shows that $M\vp$ is a well defined holomorphic function in $\cR$. Next, we observe that $$\begin{aligned}
M\vp(z)
& = \frac{1}{\sqrt{2\pi}} \int_{-\infty}^{+\infty} \big(\vp\circ\exp)(s) e^{zs}\, ds\\
& = \cF^{-1} \big( (\vp\circ\exp) e^{x(\cdot)}\big)
(y)\,,\end{aligned}$$ so that we obtain $$\begin{aligned}
\big\| M\vp(x+i\cdot) \big\|_{L^2({{\mathbb R}})}^2
& = \| \vp \|_{L^2( (0,+\infty), \xi^{2x-1}d\xi)}^2 \\
& \le C_x \| \vp \|_{L^2 ( (0,+\infty), e^{a\xi^{\rho}}\,
\frac{d\xi}{\xi})}^2
\, ,\end{aligned}$$ uniformly in $x\in (0,b]$. Hence, $M\vp \in H^2(S_b)$ for every $b>0$ if $\vp\in L^2 ( (0,+\infty), e^{a\xi^{\rho}}\,
\frac{d\xi}{\xi})$.
Next we consider $\vp,\psi\in L^2 ( (0,+\infty), e^{a\xi^{\rho}}\,
\frac{d\xi}{\xi})$ and we first assume that both have compact support. Then, $$\begin{aligned}
\la M\vp,\, M\psi\ra_{\cM^2_{a,\rho}}
& = \frac{1}{2\pi} \sum_{n=0}^{+\infty} \frac{a^n}{n!} \int_{-\infty}^{+\infty}
M\vp (\tnt +iy) \ov{M\psi (\tnt +iy) }\, dy \\
& = \sum_{n=0}^{+\infty} \frac{a^n}{n!} \int_{-\infty}^{+\infty}
\cF^{-1} \Big( \big(\vp\circ\exp) e^{\rho \frac{n}{2}(\cdot)}\Big) (y)
\ov{ \cF^{-1} \Big( \big(\psi\circ\exp) e^{\rho \frac{n}{2}(\cdot)}\Big) (y)} \, dy \\
& = \sum_{n=0}^{+\infty} \frac{a^n}{n!} \int_{-\infty}^{+\infty}
\vp \big(e^y\big) e^{\rho\frac{n}{2}y}
\ov{\psi \big(e^y\big) } e^{\rho\frac{n}{2}y} \, dy \\
& = \int_{-\infty}^{+\infty}
\vp \big(e^y\big) \ov{\psi \big(e^y\big) } e^{2e^y}\, dy \\
& = \int_0^{+\infty} \vp (t)
\ov{\psi (t) } \, e^{at^{\rho}} \frac{dt}{t}\\
&= \la \vp,\,\psi\ra_{ L^2 ( (0,+\infty), e^{a\xi^{\rho}}\,
\frac{d\xi}{\xi})}
\,. \end{aligned}$$ Therefore, the Mellin transform $M$ is a partial isometry. In order to prove that $M$ is onto we need to recall some well-known facts about the inversion of the Mellin. If, for every $c$ in some interval $I$ it is $g(c+i\cdot)\in L^2({{\mathbb R}})\cap L^1({{\mathbb R}})$ and $g(c+iy)\to 0$ as $|y|\to+\infty$ uniformly in $c\in I$, then $$\label{M-inv-c}
M_c^{-1} g(\xi) =\frac{1}{\sqrt{2\pi}}
\int_{-\infty}^{+\infty} g(c+it) \xi^{-c-it} \, dt$$ is well defined for every $\xi>0$, independent of $c\in I$ and $M M_c^{-1}g =g$. Fix now $\eps>0$ and suppose $f\in \cM^2_{a,\rho}(\cR_{-\eps})\cap \cM^1_{a,\rho}(\cR_{-\eps})$, satisfying $f(c+iy)\to 0$ as $|y|\to+\infty$ uniformly in $c\in I$. Therefore, $M_{\rho \frac{n}{2}}^{-1} f$ is independent of $n\ge 0$ and $M
M_{\rho \frac{n}{2}}^{-1} f=f$. For every such function $f$, we define $$\label{def-cM-inv}
M^{-1} f = M_{\frac{n}{2}}^{-1} f\,$$ and, by density, it is enough to show that $$\label{second-isometry}
\|M^{-1} f \|_{L^2( (0,+\infty),\,e^{a\xi^{\rho}}\frac{d\xi}{\xi}\,)} = \| f\|_{\cM^2_{a,\rho}} \,.$$
If $f(c+i\cdot)\in L^2\big({{\mathbb R}}\big)$ we put $\vp_c(t)=
t^{-c} f\big( c-i\log t)$ and claim that if $\xi>0$ we have $$\label{M-inv-c-f-equals-M-phi-c}
\xi^c M_c^{-1} f(\xi) =M \vp_c (c+i\log \xi)\,.$$ For, $$\begin{aligned}
\big(M \vp_c \big)(c+i\log \xi)
&=\frac{1}{\sqrt{2\pi}} \int_0^{+\infty} t^{-c} f(c-i\log t)t^{c+i\xi-1}\,dt\\
&=\frac{1}{\sqrt{2\pi}} \int_{-\infty}^{+\infty} f(c+is) \xi^{-is}\,ds\\
&=\xi^{c} \big(M_c^{-1} f\big) (\xi)
\end{aligned}$$ as we claimed. Now, [@BuJa Lemma 2.3] shows that, for every function $g$ such that $g(t) t^c \in L^2\big((0,+\infty), \frac {dt}t\big)$ it is $$\int_0^{+\infty} t^{2c} |g(t)|^2 \, \frac{dt}t= \int_{-\infty}^{+\infty} |Mg(c+iy)|^2 \,dy\,.$$ Using all the above, arguing as in [@PS Lemma 3.2] we obtain $$\begin{aligned}
\|f\|_{\cM^2_{a,\rho}}^2
& =\| M^{-1} f\|_{L^2( (0,+\infty),\,e^{a\xi^{\rho}}\, \frac{d\xi}{\xi})}^2\end{aligned}$$ that is, holds, and we are done.
Unboundedness of the orthogonal projection
------------------------------------------
Next we consider the question of the boundedness of the orthogonal projection. We find it quite interesting that the orthogonal projection does not extend to a bounded operator on $\cM_{a,\rho}^p$ for any $p\neq2$.
\[proj-thm\] The orthogonal projection operator $P: L^2(\ov\cR,d\omega_{a,\rho} ) \to \cM_{a,\rho}^2$, is unbounded as operator $$P: L^p(\ov\cR,d\omega_{a,\rho})\cap L^2(\ov\cR,d\omega_{a,\rho}) \to \cM_{a,\rho}^p$$ for every $p\neq2$.
The proof follows the same lines of the one for the spaces $\cM_{2,1}^p$ (see [@PS Thm. 8]) so that we simply sketch it. A necessary condition for the projections to be bounded on $L^p$ is that the kernels $K_w$ belong to $\cM^{p'}(\ov\cR, d\omega_{a,\rho})$ (with $1/p+1/{p'}=1$). By duality and since $P$ is self-adjoint it is enough to show that $K_w \notin L^p(\ov\cR,d\omega_{a,\rho})$ for any $p>2$.
First we observe that, there exists $C>0$ such that $$\big| \Gamma(\textstyle{\tnt+\frac{u}{\rho}+iy})\big|^p
\ge Ce^{-p\frac{u}{ \rho}} e^{-p\frac{\pi}{2}|y|}
\exp\Big\{ \textstyle{ \frac{p(n-1)}{2}\log (\frac{n}{2})
-\frac{pn}{2} } \Big\} \, .$$
Therefore, $$\begin{aligned}
\|K_w\|_{\cM^p}^p
&=C \frac{1}{a^{\frac{pu}{\rho}} }
\sum_{n=0}^{+\infty} \frac{a^{n(1-\frac{p}{2})}}{n!} \int_{-\infty}^{+\infty}
\big| \Gamma(\textstyle{\tnt+\frac{u}{\rho}+iy})\big|^p \, dy
\\
& \ge C_u \sum_{n=0}^{+\infty} \frac{e^
{n[(1-\frac{p}{2}) \log a-\frac{p}{2}(1+\log2)] }}{n! \, n^{\frac{p}{2}}} e^{\frac{p}{2}n\log n} \, ,\end{aligned}$$ which clearly diverges when $p>2$.
Some open questions
-------------------
We collect here some properties of $\cM^2_{2,1}$ that did not easily carry over to the more general case of $\cM^2_{a,\rho}$.
1. Corollary \[M2-fact\] shows that $\cM_{2,1}^2 =\frac{\Gamma(1+z)}{2^z} \cH$, where $\cH$ consists of functions of exponential type $\pi/2$ in $\cR$. Thus, functions in $\cM_{2,1}^2$ are of order 1 in $\cR$, but can be factored as product of a non-vanishing term times a function in the space $\cH$. It is interesting to notice that the space $\cH$, that appears in [@PS] as $M_\Delta(A^2(\Delta))$, had already appeared in the literature, in a different context [@KT1; @KT2].
It would certainly be of interest to extend this factorization result to the case $\cM^2_{a,\rho}$, and possibly also the description of the analogous of the factor $\cH=M_\Delta(A^2(\Delta))$.
2. Along the same lines as above, we mention that we do not know whether the Carleman formula, Theorem \[zero-set-thm\] (ii), holds in the general case of $\cM^2_{a,\rho}$. Again, proving a factorization theorem for $\cM_{a,\rho}^2$ would give us a tool for describing its zero-sets, as we obtained in the case of $\cM_{2,1}^2$.
3. It would also be of interest to obtain a description of the problem for [*weighted*]{} Bergman spaces on $\Delta$. Here we have two most natural choices of weights, namely $v(\z)= (1-|\z-1|^2)^\alpha$, and $\tilde v(\z)=|\z|^\alpha$, $\alpha>-1$. The weight $v$ is radially symmetric in the disk $\Delta$, while $\tilde v$ is radial in ${{\mathbb C}}$.
Comparison with other function spaces {#comparison}
=====================================
In this section we compare the space $\cM_{a,\rho}^p$ with other Hilbert spaces of holomorphic functions on the right half-plane.
Other spaces of Hardy–Bergman type: Zen spaces {#gen-H-B-spaces-subsec}
----------------------------------------------
Let $$\label{gen-HB-def}
A^p_\omega(\cR)
= \Big\{ f\in \Hol(\cR): \, \sup_{r>0} \int_0^{+\infty} \int_{{\mathbb R}}|f(x+r+iy)|^p\, dy\, d\nu(x) <+\infty
\Big\}\, ,$$ where $\omega=\nu(x)\otimes dy$ is a regular Borel measure on $\cR$ and $\nu$ is such that there exists $R>0$ such that $$\label{Delta-2-cond}
\sup_{t>0} \frac{\nu([0,2t))}{\nu([0,t))} \le R\, ,$$ which is a doubling-type condition at the origin.
Such spaces have been studied by several authors; here we mention Z. Harper [@Harper-Laplace; @Harper-boundedness], B. Jacob, S. Pott and J. Partington [@Jacob-Partington-Pott1], and by I. Chalendar and J. Partington [@Chalendar-Partington], and they are sometimes called [*Zen spaces*]{}. We mention here that previously, this kind of spaces had been considered in [@Gustavo] by G. Garrigós, in the much more general case of tube domains over cones, although only for quasi-invariant measures $d\nu$.
In [@Jacob-Partington-Pott1] two facts about the spaces $A^p_\omega$ are stated without proof, namely:
- $A^2_\omega$ is a Hilbert space;
- $f\in A^p_\omega$ implies that $f\in H^p(\cR_a)$ for every $a>0$, where $H^p(\cR_a)$ denotes the Hardy space of the half-plane $\{\Re z>a\}$.
We believe that both statements require some proof. The reason being that, from the definition it follows that for $f\in
A^p_\omega$, the avarages $a_{f_r,p}=\int_{{\mathbb R}}|f(x+iy)|^p\, dy$ are $\nu$-a.e. finite, for every $r>0$, where $f_r=f(\cdot+r)$. However, this condition is not easily exploited.
We remark that, for $1\le p<\infty$, the requirement $f\in H^p(S_{a,b})$ for all $0<a<b<\infty$, implies that the average function $a_{f,p}(x)$ is finite everywhere in $(0,+\infty)$ and it is convex. For, if $x=\delta a+(1-\delta)b$, for some $0<\delta<1$, then it is well known that $$a_{f,p}(x) \le a_{f,p}(a)^{\delta} a_{f,p}(b)^{1-\delta}
\le \delta a_{f,p}(a) +(1-\delta) a_{f,p}(b)
\,.$$ This implies that $a_{f,p}$ must attains its supremum value as either limit as $x\to 0^+$ or as $x\to+\infty$.
Hence, once we know that $f\in A^p_\omega$ implies that $f\in H^p(S_{a,b})$, then the “sup”-condition in the definition of the norm of $A^p_\omega$ forces $a_{f,p}(x)\to0^+$ as $x\to+\infty$. Thus, in particular $f\in H^p(\cR_a)$, for every $a>0$.
However, if $f\in\Hol(\cR)$ the average function $a_{f,p}$ can have some extreme behaviors, as one could see from [@Harper-Laplace]. For example, it follows easily from [@Harper-Laplace Sec. 4] that given any sequence of separated intervals $I_n\subset(0,+\infty)$, that is such that $\ov I_n\cap \ov I_m
=\emptyset$ for $n\neq m$, there exists $f\in\Hol(\cR)$ such that $f\in H^p(S_{I_n})$, for all $n$, but $f$ is not in $H^p$ of any larger strip.
In the remainder of this section we are going to prove that functions in $ A^p_\omega$ can be bounded uniformly on compact subsets of $\cR$ only in terms of the $ A^p_\omega$-norm and of some constant depending only the compact set. This property will (easily) imply that $A^2_\omega$ is a Hilbert space, and also provide a tool to prove statement (ii) above.
\[sub-mean-val-prop\] Let $1\le p<\infty$ and $ A^p_\omega$ and $\nu$ be as in and , resp. Then, for each compact set $E\subset\cR$ there exists $C_E>0$ such that $$\sup_{\lambda\in E} |F(\lambda)|\le C_E \| F\|_{A^p_\omega} \,.$$
Assume first that the measure $\nu$ has an atomic part $\nu_{\rm at}$. If $\nu(\{0\})>0$ then $A^p_\omega$ embeds continuously in $H^p(\cR)$, and the statement follows. If $\nu(\{0\})=0$, but $\nu_{\rm at}$ has support $\{a_n\}_{n\in{{\mathbb Z}}}$, with $a_n\to 0^+$ as $n\to-\infty$. Arguing as before, we obtain that $f\in A^p_\omega$ belongs to $H^p({{\mathbb C}}_{a_n})$ for every $n$, and the conclusion follows again.
Suppose now that $\nu$ has no atomic part that accumulates at the orgin. Let $a>0$. By condition we can find $\eps<\frac{a}{4}$ such that $\nu([\frac34 \eps, \frac54\eps))>0$. Consider the function $g(t)
=\nu([t-\frac\eps2, t+\frac\eps2))$. Notice that $g(\eps)>0$, and since $\nu$ has no atomic part that accumulates at the origin, $g$ is continuous, for $\eps$ suitably small. Then, $g$ has a positive minimum in an interval $[\eps-\delta,\eps+\delta]$.
For $\lambda \in \ov S_{a-\delta,a+\delta}$, let $\lambda_0 = \lambda
-(a-\eps)$, so that $\Re\lambda_0\in [\eps-\delta,\eps+\delta]$. Denote also $Q(\lambda_0,\eps)$ the square centered in $\lambda_0$ with sides of lenght $2\eps$. We then have, $$\begin{aligned}
|F(\lambda)|^p
& = \frac{1}{\omega(D(\lambda_0,\eps))} \int_{ D(\lambda_0,\eps)}
|F(\lambda_0+a-\eps )|^p\, d\omega(z)\notag \\
& \le \frac{C}{\eps^2 \omega(D(\lambda_0,\eps))} \int_{ D(\lambda_0,\eps)}
\int_{ D(\lambda_0,\eps)}
|F(w+a-\eps)|^p\, dA(w) \, d\omega(z)\notag \\
& \le \frac{C}{\eps^2 \omega(D(\lambda_0,\eps))}
\int_{ D(\lambda_0,\eps)} \int_{ D(z,2\eps)}
|F(w+a-\eps)|^p\, dA(w) \, d\omega(z)\notag \\
& = \frac{C}{\eps^2 \omega(D(\lambda_0,\eps))}
\int_{ D(\lambda_0,\eps)} \int_{ D(0,2\eps)}
|F(z+\z+a-\eps)|^p\, dA(\z) \, d\omega(z)\notag \\
& = \frac{C}{\eps^2 \omega(D(\lambda_0,\eps))}
\int_{ D(0,2\eps)}
\int_{ D(\lambda_0,\eps)} |F(z+\z+a-\eps)|^p\, d\omega(z)\, dA(\z) \notag \\
& \le \frac{C}{\eps^2 \omega(D(\lambda_0,\eps))}
\int_{ D(0,2\eps)}
\int_{Q(\lambda_0,\eps)}
|F(x+\xi+a-\eps+i(y+\eta))|^p\, d\nu(x)dy\, dA(\z) \notag \\
& \le C_\eps
\frac{1}{\nu([\Re\lambda_0-\frac\eps2, \Re\lambda_0+\frac\eps2))}
\|F\|_{A^p_\omega}^p
\notag \\
& \le C_\eps \,\|F\|_{A^p_\omega}^p \, .
\label{last-disp-submean}\end{aligned}$$ because of our construction of the function $g$.
An easy argument now shows that $A^2_\omega$ is a Hilbert space. We do not know of any proof of such fact that does not use the fact that norm convergence implies uniform convergence on compact subsets.
It also follows that $f\in A^p_\omega$ is bounded on every closed strip $\ov{S_{a,b}}\subset\cR$. Therefore, if $f(x+i\cdot)\in
L^p({{\mathbb R}})$ for $x=a,b$, it follows that $f\in
H^p(S_{a,b})$. Hence we have,
\[Hp-strip-cor\] [ (1)]{} If $f\in A^p_\omega$ then $f\in
H^p(\cR_a)$ for any $a>0$ and $$\|f\|_{ A^p_\omega}^p
= \int_0^{+\infty} \int_{{\mathbb R}}|f(x+iy)|^p \, dyd\nu(x)\, .$$
[(2)]{} When $\nu(\{0\})=0$, we have the equality $$A^p_\omega
= \big\{ f\in \Hol(\cR):\, f\in H^p(\cR_a) \text{ for all }
a>0,\ \text{and } f \in L^p(\cR, d\omega)\big\}\, ,$$ and if $\nu(\{0\})>0$ we have the equality $$A^p_\omega
= \big\{ f\in \Hol(\cR):\, f\in H^p(\cR) \text{ and } f \in L^p(\ov\cR, d\omega)\big\} \, .$$ In both cases we have $$\| f\|_{A^p_\omega}^p =
\int_0^{+\infty} \int_{{\mathbb R}}|f(x+iy)|^p\, dyd\nu(x) \, .$$
It is now clear that $A^p_\omega$ is a closed subspace of $\cM^p_\omega$. It is worth to notice that while functions in $A^p_\omega$ are bounded in $\cR_a$ for every $a>0$, functions in $\cM^p_\omega$ are, in general, allow to grow at infinity, as the case of $\cM_{a,\rho}^p$ shows.
A significant consequence of Prop. \[sub-mean-val-prop\] is a Paley–Wiener theorem type theorem for the space and in $f\in A^2_\omega$. In particular it proves that isometry considered in [@Jacob-Partington-Pott1 Prop. 2.3], while its proof is inspired by the one of [@Harper-Laplace Thm. 2.1].
We need a couple of definitions. For $\xi<0$ we set let $$v(\xi) =
\int_0^{+\infty} e^{2\xi x}\, d\nu(x)\, .$$ It was already observed in [@Jacob-Partington-Pott1 Prop. 2.3] that the condition implies that the integral above converges. For $\vp\in L^2((-\infty,0), v(\xi)d\xi)$ and $z\in\cR$ define $$T\vp (z)
= \frac{1}{\sqrt{2\pi}} \int_{-\infty}^0 e^{z\xi} \vp(\xi)\, d\xi
= \cF^{-1}\big(e^{x\xi}\vp(\xi)\big)(y)
\, .$$
\[PW-thm-Berg-ab\] [(1)]{} If $\vp\in L^2((-\infty,0), v(\xi)d\xi)$ then $T\vp\in
A^2_\omega$ and $$\|T\vp\|_{A^2_\omega}= \|\vp\|_{L^2((-\infty,0), v(\xi)d\xi)}
\,.$$ [(2)]{} Conversely, if $F\in A^2_\omega$, then there exists $\vp\in L^2((-\infty,0), v(\xi)d\xi)$ such that $F=T\vp$ and $\| F\|_{A^2_\omega}=
\|\vp\|_{L^2((-\infty,0), v(\xi)d\xi)}$.
We remark again that (1) is just [@Jacob-Partington-Pott1 Prop. 2.3].
For (2), let $0<a<b<\infty$. We know that $\int_{{\mathbb R}}|F(x+iy)|^2\, dy<+\infty$, for all $x>0$. Let $Y>0$ be fixed and let $R_Y$ be the rectangle of vertices $(a,-Y)$, $(b,-Y)$, $(b,Y)$, and $(a,Y)$. By Cauchy’s integral formula, for $\lambda \in
R_Y$ we have $$F(\lambda)= \frac{1}{2\pi i} \int_{\p R_Y} \frac{F(z)}{z-\lambda}\,
dz\, .$$ Letting $Y\to+\infty$, using the fact that $F$ is bounded in every closed strip, for $a< \Re\lambda< b$ we obtain $$\begin{aligned}
F(\lambda)
& = \frac{1}{2\pi i} \Big(
\int_{b+i{{\mathbb R}}} \frac{F(z)}{z-\lambda}\, dz - \int_{a+i{{\mathbb R}}} \frac{F(z)}{z-\lambda}\, dz
\Big) \\
& = \frac{1}{2\pi } \Big(
\int_{{\mathbb R}}\frac{F(b+iy)}{iy-(\lambda-b)}\, dy
-\int_{{\mathbb R}}\frac{F(a+iy)}{iy-(\lambda-a)}\, dy \Big)\, .\end{aligned}$$ Now notice that $$\big| \int_{{\mathbb R}}\frac{F(b+iy)}{iy-(\lambda-b)}\, dy \big|^2
\le \| F_b\|_{L^2({{\mathbb R}})}^2 \int_{{\mathbb R}}\frac{1}{y^2+|\Re\lambda-b|^2}\, dy
\to +\infty$$ as $b\to+\infty$, since $F\in H^2(\cR_{-\eps})$ for every $\eps>0$. Therefore, observing that $\lambda-a\in \cR$, $$\label{who}
F(\lambda)=
\frac{1}{2\pi } \int_{{\mathbb R}}\frac{F(a+iy)}{(\lambda-a)-iy}\, dy =
\cS(F_{a})(\lambda-a)\, ,$$ where $\cS$ denotes the Szegö projection on $H^2(\cR)$. By the classical Paley–Wiener theorem there exists $g_a \in L^2((-\infty,0),d\xi)$ such that $$F(\lambda)= \cS(F_{a})(\lambda-a)
= \frac{1}{\sqrt{2\pi}} \int_{-\infty}^0 e^{(\lambda-a)\xi} g_a(\xi)\, d\xi\, ,$$ for $\Re\lambda>a$. By the uniqueness of the Fourier transform, $e^{-a\xi} g_a = e^{-a'\xi} g_{a'}$ for every $a,a'>0$.
Hence, $$\label{F-from-vp}
F(\lambda)
= \frac{1}{\sqrt{2\pi}} \int_{-\infty}^0 e^{\lambda\xi} \vp(\xi)\, d\xi
\, ,$$ where, for $\xi<0$, $\vp(\xi) =
e^{-a\xi} g_a(\xi)$, any $a>0$, so that $e^{a\xi}\vp \in L^2((-\infty,0),d\xi)$ for every $a>0$. Thus the integral in is indeed absolutely convergent and by Plancherel theorem again, $$\begin{aligned}
\| F\|_{A^2_\omega}^2
& = \sup_{r>0} \int_0^{+\infty} \int_{{\mathbb R}}|\widehat{F_{x+r}}(\xi)|^2\, d\xi\, d\nu(x)\\
& = \sup_{r>0} \int_0^{+\infty} \int_{-\infty}^0 e^{2(x+r)\xi}
|\vp(\xi)|^2\, d\xi \,
d\nu(x)\\
& = \sup_{r>0} \int_{-\infty}^0 \Big( \int_0^{+\infty}
e^{2x\xi}\, d\nu(x)\Big) e^{2r\xi} |\vp(\xi)|^2
\, d\xi \\
& = \int_{-\infty}^0 |\vp(\xi)|^2 v(\xi)\, d\xi\, ,\end{aligned}$$ as we wished to show.
As usual, given a Paley–Wiener theorem for $A^2_\omega$, it is possible to obtain its reproducing kernel. Indeed,
\[RK-thm-Berg-ab\] The reproducing kernel for $A^2_\omega$ is $$K(z,w)=\frac 1 {2\pi} \int_{-\infty}^0 e^{(z+\ov w)\xi} \frac {d\xi}{v(\xi)}.$$
With the notation as in Theorem \[PW-thm-Berg-ab\], to every function $G$ in $A^2_\omega$ we associate the function $\vp_G$ in $L^2((-\infty,0), v(\xi)d\xi)$ such that $T\vp_G=G$.
From Proposition \[sub-mean-val-prop\] we know that $A^2_\omega$ is a reproducing kernel Hilbert space. Let $K_z$ be the reproducing kernel in $A^2_\omega$, that is, for $F\in A^2_\omega$ and $z\in \cR$ it holds that $F(z)=\la F, K_z\ra_{A^2_\omega}$.
Then, $$\begin{aligned}
F(z)
& = \int_0^{+\infty} \la F(u+i\cdot), K_z(u+i\cdot)\ra_{L^2({{\mathbb R}})} d\nu(u)\\
& = \int_0^{+\infty} \big\la \big[F(u+i\cdot)\big]\widehat{\ },\,
\big[K_z(u+i\cdot)\big]\widehat{\ }\big\ra_{L^2({{\mathbb R}})} d\nu(u)\\
& = \int_0^{+\infty} \big\la e^{u\xi}\vp_F,\,
e^{u\xi} \vp_{K_{z} }\big\ra_{L^2({{\mathbb R}})} d\nu(u)\\
& = \int_{-\infty}^0 \vp_F(\xi) \ov{ \vp_{K_{z}} (\xi) }v(\xi)\,d\xi,\end{aligned}$$ where switching the integral with the sum is justisfied since the last integral converges absolutely.
On the other hand, $$F(z) =\frac{1}{\sqrt{2\pi}} \int_{-\infty}^0 e^{z\xi} \vp_F(\xi)\, d\xi\,,$$ so that $$\vp_{K_{z} }(\xi)
= \frac{1}{\sqrt{2\pi}} \frac{1}{v(\xi)} e^{\ov z\xi} \,,$$ and $$K_z(w)
= \frac{1}{2\pi} \int_ {-\infty}^0 e^{w\xi} e^{\ov z\xi} \frac{1}{v(\xi)}\, d\xi$$ as we wished to show.
The space $L^2(\ov\cR,d\omega)\cap \Hol(\ov\cR)$
------------------------------------------------
We take the opportunity to discuss the naive definition of Bergman space with respect to a general measure such as $\omega_{a,\rho}$ . Define $\cX_{a,\rho}^2$ as the closure of the holomorphic functions on $\cR$ that extend continuously to the boundary and in the norm of $L^2(\ov\cR,d\omega_{a,\rho})$.
We show that such definition is quite inadequate from the point of view of complex analysis. For, we prove the following
\[unb-eval-pts\] The following properties hold:
- the set of points $z\in\cR$ such that $\cX_{a,\rho}^2\ni f\mapsto f(z)$ is unbounded is dense in $\ov\cR$;
- $\cX_{a,\rho}^2$ contains functions that are not holomorphic in $\cR$.
We prove the statement for the space $\cX_{2,1}^2$, the proof in the general case being completly analogous.
\(1) Let $h(z)= (1+z)^{-1} \exp\{i e^{2\pi iz}\}$ and set $f_k(z)=h(kz)$. Observe that $|\exp\{i e^{2\pi ik(\frac n2 +iy)}\}|=
|\exp\{ ie^{-2k\pi y}\cos (kn\pi) \}|=1$. Then
- $f_k(\tnt+iy)\to 0$ a.e. $dy$ as $k\to +\infty$;
- $\| f_k\|_{L^2(\cR,d\omega_{2,1})}^2
= \pi \sum_{n=0}^{+\infty} \frac{2^n}{n!} \big(1+k\frac n2\big)^{-2} \to 0$, as $k\to +\infty$,
as an easy calculation shows. Now let $z\in\cR$ with $\Re z>0$ rational, equal to $p/q$ with $p,q$ relatively prime and $q\neq2$. Such points are dense in $\ov\cR$ and $$f_k (z)= \frac{1}{1+\frac{kp}{q} +iky} \exp \{ ie^{-2k\pi y} \big(
\cos(kp\pi/q) +i \sin(kp\pi/q) \big)\}\, .$$ If we choose $k=\ell_0 +2q\ell$, $\ell=1,2,\dots$ we see that $$|f_k(z)| \approx \ell^{-1} e^{-2(\ell_0+2q\ell)y}\sin (\ell_0 p\pi/q)\, .$$ Choosing $\ell_0$ such that $\sin (\ell_0 p\pi/q)$ we see that $|f_k(z)| \to +\infty$ if $y<0$. Thus, the point evalution at $z=\frac pq+iy$ with $y<0$ are not bounded on $\cX_{2,1}^2$.
To deal with the case $y>0$ it suffices to replace $\exp\{i e^{2\pi iz}\}$ with $\exp\{i e^{-2\pi iz}\}$ in the definition of $f_k$.
\(2) Let $$f_k(z) =\frac{1}{1+z} \frac{\exp\{ ie^{4k\pi iz}-1\} }{e^{4k\pi iz}} \,.$$ Since $$f_k({\tnt}+iy)
= \frac{1}{1+\frac n2 +iy} \frac{\exp\{ ie^{-4k\pi y}-1\} }{e^{-4k\pi y}}$$ and for $t>0$, $\big| \frac{e^{it}-t}{t}\big|\le C$, we have $$\begin{aligned}
\| f_k\|_{L^2(\cR,d\omega{2,1})}^2
& =\sum_{n=0}^{+\infty} \frac{2^n}{n!} \int_{{\mathbb R}}|f_k(\tnt+iy)|^2\, dy \\
& \le C \sum_{n=0}^{+\infty} \frac{2^n}{n!} \int_{{\mathbb R}}\frac{1}{\big(1+\frac n2\big)^2 +y^2} \, dy\\
& \le C
\,.\end{aligned}$$ Then, $f_k\in \cH$ for $k=1,2,\dots$. Moreover, $f_k\to g$ $\omega$-a.e. where, $$g(\tnt+iy)= \begin{cases}
i(1+\tnt+iy)^{-1} \quad & y>0 \cr
0 & y<0\, .
\end{cases}$$ It is now easy to see that $f_k\to g$ in $L^2({{\mathbb R}},d\omega)$. Since $g$ cannot be extended to a holomorphic function on $\cR$, this concludes the proof.
Final remarks
=============
It is interesting to notice that the space $\cH$ defined in , and that in [@PS] we showed to be equal to $M_\Delta(A^2\Delta)$, had already appeared in the literature, in a different context [@KT1; @KT2]. However, $\cH$ can be described as the closure of polynomials in the $L^2(\cR,
d\mu)$-norm where $d\mu(z)=2^{-x}|\Gamma(1+z)|^2 dA(z)$, which is not translation invariant in $\cR$. In [@KT1] the authors also discussed a -type question, concerning the completeness of the powers $\{(1-z)^{\lambda_n}\}$ in $H^2({{\mathbb D}})$, for $\lambda_n>0$ and $\lambda_{n+1}-\lambda_n>\delta$; their results however have no (obvious) connection with the problem for the Bergman space.
Of course, there exist other papers dealing with -type questions. In [@Sed] A. Sedletkskii studied the completeness of sets of exponentials in weighted $L^p$ spaces on $(0,+\infty)$ in terms of zeros of functions in the classical Bergman space on a half-plane.
In [@Lukacs57; @Lukacs] E. Lukacs studied positive measures $\nu_L$ on the real line that are Fourier transform of restriction of entire functions. Then, we can consider the measures on $\ov\cR$ of the form $\nu_L\otimes dy$, (with $\nu_L$ restricted to $[0,+\infty)$). We believe these measures constitute an interesting class of measures for which studying the properties of the function spaces $\cM^p_\omega$. We wish to come back to this, and the other open problems we mentioned, in a future work.
[^1]: [*Math Subject Classification*]{} 30H99, 46E22, 30C15, 30C40.
[^2]: Authors partially supported by the grant Prin 2010-11 [ *Real and Complex Manifolds: Geometry, Topology and Harmonic Analysis*]{}.
| {
"pile_set_name": "ArXiv"
} |
Declaration of Authorship {#declaration-of-authorship .unnumbered}
=========================
The work described in this thesis was carried out in collaboration with Professor Nick Evans and Tom Waterson. The following list details our original work and gives references for the material.
- Chapter \[ch:IRimp\]: Nick Evans, Andrew Tedder and Tom Waterson, JHEP v.01 2007, p.058, arXiv:hep-ph/0603249
- Chapter \[ch:UVimp\]: Nick Evans and Andrew Tedder, Phys. Lett. B642 2006, p.546-550, arXiv:hep-ph/0609112
- Chapter \[ch:Hadronisation\]: Nick Evans and Andrew Tedder, 2007, arXiv:0711.0300 (hep-ph), accepted for publication in Physics Review Letters
There is also some original work of ours in chapter \[ch:CM\] (the quantitative glueball spectrum, and the $m_{\rho}:m_{\pi}^2$ graph) and chapter \[ch:AdS-QCD\] (the possible effect of gluonic contributions to $g_5$). However no claims to originality are made for the rest of chapters \[ch:CM\] and \[ch:AdS-QCD\], and all of chapters \[ch:intro\] to \[ch:flavour\], whose content has been complied from a variety of sources.
Acknowledgements {#acknowledgements .unnumbered}
================
I would like to thank my supervisor, Professor Nick Evans, for coming up with such interesting research topics, guiding me towards publication and for humouring my more fanciful ideas. The students, post-doctorates and staff at Southampton have all been very friendly and willing to help me. I like to think I made them consider the simple questions that they thought were beneath them! In particular I single out Ed Threlfall, Andreas Juettner, James Ettle and Michael ‘one hour’ Donnellan for their general help.\
Jonathan Shock and Tom Waterson guided my education in AdS/CFT matters for which I am grateful. I have also had great fun doing outreach activities with Pearl John, who is always quick to remind me that not everyone knows what a quark is.\
I am very grateful to my parents who have been very supportive throughout my life, and taught me the value of a balanced education from a young age.\
Financially I thank the University of Southampton for funding me through my PhD studies.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[Very recently, field-effect transistors based on few-layer phosphorene crystals with thickness down to a few nanometres have been successfully fabricated, triggering interest in this new functional two-dimensional material. In this work, we apply the first-principles calculations to studying the evolution of electronic structures and lattice dynamics with vertical strain for monolayer and bilayer phosphorenes. It is found that, by changing the thickness of phosphorene or the strain applied on it, its band gap width can be well tuned, and there will appear a transition from semiconductor to metal. In particular, the bilayer phosphorene may become a good BCS superconductor by adjusting the interlayer distance, in which the interlayer van der Waals coupling is favorable to the dynamical stability against strain. ]{}'
author:
- 'G. Q. Huang$^{1,2}$, Z. W. Xing$^{2,3}$'
title: ' Semiconductor to Metal Transition, Dynamical Stability and Superconductivity of Strained Phosphorene '
---
Introduction
============
Black phosphorus (BP) is the most stable allotrope of the element phosphorus and has many interesting physical properties. BP can server as the electrode material for Lithium-Ion batteries.[@1; @2] Several structural phase transitions were found under pressure accompanied by semiconductor-semimetal-metal transition.[@3; @4] Furthermore, it was reported that the BP single crystal shows superconductivity with $T_{C}$ higher than 10K under high pressure.[@5; @6] The structure of BP consists of puckered double layers, which are held together by weak van der Waals (vdW) force. This peculiar layer structure permits to employ mechanical exfoliation to extract thin black phosphorus from a bulk crystal. Recently, the few-layer black phosphorus (phosphorene) had been successfully exfoliated and used them to create field-effect transistors, which exhibit large current on-off ratios and high mobilities.[@7; @8; @9] The successful fabrication of this novel two-dimensional (2D) semiconducting material was soon paid amounts of attention and predicted to have a great potential for practical applications.[@10; @11]
Bulk BP is a direct-gap semiconductor with a 0.33 eV band gap.[@12] However, the band gap of phosphorene is predicted to be highly sensitive to its thickness and strain. Previous first-principles calculations show that the band gap ranges from 1.5 eV for a monolayer to 0.6 eV for a 5-layer.[@13] A 3$\%$ in-plane strain can change phosphorene from a direct-gap to an indirect-gap semiconductor,[@8] while a vertical compression can induce the semiconductor to metal transition.[@14] Strain is an effective method to tailor electronic properties for 2D semiconductors. For monolayer phosphorene (MLP) and bilayer phosphorene (BLP), whether there are different strain effects on electronic properties is one of the motivation of this study. Furthermore, due to the interlayer coupling bonded by weak van der Waals force, it is also interesting to see how bilayer phosphorene responses to the change of interlayer distance.
While there have been substantial works on the electronic properties of phosphorene, investigations on the vibrational properties of phosphorene are scarce. The Raman spectra was measured for the BP flake with thicknesses ranging from 1.6 nm to 9 nm.[@15] The phonon dispersion of MLP was calculated by first-principles calculations.[@10; @16] The lattice vibrational modes of MLP with in-plane strain were also reported by Fei *et al.* through first-principles simulations.[@16] However, phosphorene has an anisotropic structure, whether it has different dynamical behavior for phosphorene with out-of-plane strain? This problem is our concern. We will pay particular attention to the dynamical stability of strained phosphorene in order to better use them for practical application.
In this work, geometric structure, electronic structure and lattice dynamics of MLP and BLP under vertical strain are studied and compared by first-principles calculations. The calculated results show that the evolution behavior of band structure for MLP and BLP under vertical strain is different. The dynamical stable range of the strained BLP is wider than that of the strained MLP. Furthermore, we also find that BLP may become a good BCS superconductor by adjusting the interlayer distance.
Computational Details
======================
The first-principles calculations have been performed within the density functional theory through the PWSCF program of the Quantum-ESPRESSO distribution.[@17] The ultrasoft pseudo-potential and general gradient approximation (GGA-PBE) for the exchange and correlation energy functional are used with a cutoff of 30 Ry for the expansion of the electronic wave function in plane waves. MLP and BLP are modeled using the slabs, which are separated by vacuum layer with the thickness of about 35 $\AA$. For the electronic structure calculations, the Brillouin zone integrations are performed with a (8,10,1) wavevector $k$-space grid by using the first-order Hermite-Gaussian smearing technique. Within the framework of the linear response theory, the dynamical matrixes are calculated for special $\bf{q}$ points in the two dimensional irreducible Brillouin zone and are Fourier interpolated throughout the full Brillouin zone. The dense (24,30,1) grid are used in the Brillouin zone integrations in order to produce the accurate electron-phonon (EP) interaction matrices. As BP presents a laminar crystal structure, vdW correction proposed by Grimme (DFT-D2) is included in our calculations.[@18]
Results and discussion
=======================
Geometric structure under vertical strain
------------------------------------------
BP has a layered orthorhombic structure. Each sheet consists of an atomic double layers known as a puckered layer. The stacking of the puckered layer for BP is *ABAB* $\ldots$ sequence with weak van der Waals bonding between layers. The edge of the puckered hexagon of A layer is located in the center of the puckered hexagon of B layer and vice versa. The top and side views of the atomic structure of a puckered layer are presented in Figs. 1 (a) and (b), respectively. Each layer may be viewed as consisting of armchair and zigzag chains along $x$ and $y$, respectively. Each phosphorus atom is covalently bonded to three neighbors within the puckered layer. Our optimized bond lengthes, bond angles (as labeled in Fig. 1) and lattice constants for BP are listed in Table 1. For comparison, the results without vdW correction for bulk BP are also listed in Table 1. It can be clearly seen that the inclusion of vdW interaction changes $a$ and $b$ parameters slightly, but reduces the c parameter substantially (about 0.6${\AA}$). The overall best agreement with experimental[@19] and previous computational results[@20] is achieved with inclusion of vdW correction.
![The top (a) and side views (b) of the atomic structure of phosphorene. (c) Vibrational pattern of $B_{3g}$ and $B_{1g}$ modes.[]{data-label="fig.1"}](fig1.pdf){width="4.in"}
For the free MLP and BLP (see Table 2), lattice constant $b$ and bond lengthes are very close to those in the bulk BP, while $a$ and bond angles $\theta_{2}$ are larger than those for the bulk BP. This fact shows that the increase of lattice constants $a$ is not due to the change of bond lengthes, but is due to the flattening of puckered layer. Next we concentrate mainly on the strained phosphorene. The vertical strain is modeled by constraint $z$ for the outmost layer atoms, the unit cell and other atomic positions are then relaxed. The strain is defined as $\sigma=\frac{h-h_{0}}{h_{0}}$, where $h$ and $h_{0}$ are the thickness of the strained and the free phosphorene, respectively. The positive (negative) of $\sigma$ corresponds to the tensile (compressive) strain. The obtained geometric parameters for MLP and BLP under vertical strain are listed in Table 2.
For the MLP, when tensile (compressive) strain is applied in the $z$ direction, the lattice constant $a$ decreases (increases) substantially. While the lattice constant $b$ and the bond length $d_{1}$ show only a weak variation under strain, reflecting the rigidity of the strong covalent bonding along the zigzag direction. So the anisotropic geometric structure along armchair ($x$ ) and zigzag ($y$ ) directions result in their different responses to the vertical strain. With the decrease of thickness ($h$) of phosphorene, the bond length $d_{2}$ decreases and the bond angle $\theta_{2}$ increases, reflecting the flattening of the puckered layer. It is need to mention that the range of strain studied here is not too large, the puckered character of structure still remains. The previous calculations by Rodin *et al.*[@14] showed that the monolayer approaches a plane square lattice configuration under severe compression.
For the BLP, when tensile strain is applied, the distance between puckered layers increases, while the covalent bonds within the puckered layer change slightly. And we also find that the geometric parameters of BLP are tend to those in the MLP when tensile strain is increased to $40\%$. This result is reasonable, since the coupling between puckered layers may be ignored at so large interlayer distance (about 6 ${\AA}$). With increase of compressive strain applied, the change of geometric parameters of BLP is qualitatively same as in the case of MLP but with smaller magnitude. Unlike to that for the MLP, the lattice constant $a$ of the BLP always increases whatever compressive or tensile strain is applied along z direction. This peculiar property is same to that observed in the bulk BP.[@21] This result implies that BLP has the same unusual mechanical response as bulk BP.
Electronic structures under vertical strain
--------------------------------------------
The calculated band structures of the strained MLP and BLP are shown in Fig. 2. For the free MLP (Fig. 2c), the band dispersion along the major high symmetry directions compares very well with the previous calculations by Rodin *et al.*.[@14] It is a direct band gap semiconductor with the band gap energy about 0.83 eV. The top of the valence band and the bottom of the conduction band have predominantly $p_{z}$ character mixed with $p_{x}$ and $s$ character. The band structure of the free BLP (Fig. 2h) has similar shape to that of MLP but with smaller gap energy (0.40 eV). The photoluminescence measurements[@15] and previous calculations[@13] indicated a thickness dependent band structure of the phosphorene.
\
\
\
The vertical strain effect on the electronic structure of MLP can be seen from (a) to (e) of Fig. 2. Both the compressive and tensile strains applied for the MLP can result in the transition from semiconductor to semimetal or metal. The band gap $E_{g}$ located at $\Gamma$ point keeps direct and decreases with the tensile strain, turning into zero when the tensile strain reaches $8\%$ (Fig. 2d), and then converting into semimetal or metal (Fig. 2e) with increasing the tensile strain continuously. When compressive strain is applied, the band gap firstly keeps direct and increases slightly, but turns into indirect when the compressive strain reaches about $4\%$. Then indirect band gap decreases and turns into zero at about $25\%$ compressive strain (Fig. 2b), and finally converting into semimetal or metal when the compressive strain continues to increase(Fig. 2a)
For the BLP, when tensile strain is applied, the band gap $E_{g}$ keeps direct and increases until tends to the value in the free MLP. The corresponding band structure (Fig. 2i) is also similar to that of the free MLP (see Fig. 2c). Under small compressive strain ($1\%$), the band gap soon turns into indirect and decreases with increasing the compressive strain. It turns into zero when the compressive strain reaches about $8\%$ (Fig. 2g), and finally converts into metal (Fig. 2f).
The evolution of the band gap $E_{g}$ with the strain for the MLP and BLP is summarized in Fig. 2j. The change of band gap is very extensive. So we can find that the electronic structure of thin BP is very sensitive to its thickness and the strain applied. This properties help to tailor materials in electronics. In order to better understand this novel 2D materials and use them for potential applications, next we turn to discuss the dynamical stability of the strained phosphorene.
Lattice vibration under vertical strain
----------------------------------------
The symmetry of phosphorene is described by the $D_{2h}$ point group. For the MLP, the primitive cell contains four atoms, leading to 12 vibrational modes, i.e. nine optical and three acoustic phonons branches. The $\Gamma$ point modes can be decomposed as $$\Gamma_{acoustic}=B_{1u}+B_{2u}+B_{3u}$$ and $$\Gamma_{optical}=B_{1u}+B_{3u}+A_{u}+B_{1g}+B_{3g}+2B_{2g}+2A_{g}$$ Among optical modes, the $A_{u}$ mode is silent, $B_{1u}$ and $B_{3u}$ are infrared active, while all others are Raman active. The structure of puckered layer remains the reflection symmetry in the $y$ (zigzag) direction but breaks the reflection symmetry in the $z$ and $x$ directions. So the vibrations along the $y$ direction are strict, while those along the $z$ and $x$ directions are mixed. $B_{3g}$ and $B_{1g}$ Raman modes correspond to the vibrations along $y$ direction and their vibrational patterns are illustrated in Fig. 1c. For $B_{3g}$ mode, the vibration of atom in the different layer is out-of-phase, but is in-phase in the same layer, i.e. two layers beating against each other. $B_{1g}$ mode corresponds to the bond stretching modes from $d_{1}$ and $d_{2}$ covalent bonds. The other four Raman active modes have similar vibrational patterns along $x$ and $z$ directions, but each mode having mixed $x$ and $z$ characters. The Raman frequencies of free MLP are listed in Table 3. Our calculated frequencies are slightly smaller than those in Ref. \[\]. For the free BLP, the frequencies of corresponding Raman active modes are also listed in Table 3. We find that the frequency shift between MLP and BLP is small except for $A_{g}^{1}$ and $B_{2g}^{2}$ modes, which vibrate mainly along the $z$ direction. The frequencies of this two modes in BLP is smaller that those in MLP probably due to the attractive vdW interaction existed in BLP.
The phonon dispersion of free MLP calculated by the density functional perturbation theory is presented in Fig. 3 (a), which compares very well with previous calculations.[@10; @16] Near the $\Gamma$ point, two in-plane acoustic modes exhibit linear dispersions, while the off-plane acoustic mode exhibits a parabolic dispersion due to the 2D character of phosphorene. The vibrations of the bond-breathing modes occupy the high-frequency region, while the vibrations of layer-breathing modes and the acoustic modes occupy the low-frequency region. Fig. 3 (b) shows the calculated phonon dispersion of free BLP which has 24 phonon branches. The optical branches have small splitting and are nearly double degenerate due to the weak interlayer interaction. Near the $\Gamma$ point, the splitting has an appreciable magnitude. Three branches from the opposite vibrations of two puckered layers (i.e. layer-breathing modes) deviate from three acoustic branches. From Fig. 3a and 3b, we can see that there is no imaginary frequency in the full phonon spectra, indicating the dynamical stability for both free-standing MLP and BLP. This conclusion is in good agreement with the experiment. Raman spectroscopy and transmission electron microscopy measurements show that the exfoliated flakes of BP are stable even in free-standing form.[@15]
![The phonon dispersion of free MLP (a) and BLP (b).[]{data-label="fig.3"}](fig3.pdf){width="4.in"}
When vertical strain is increased to a certain value, we find phosphorene becomes unstable. For MLP, there appears imaginary frequency in the phonon dispersions when tensile strain reaches 15$\%$ and compressive strain reaches 4$\%$. The appearance of imaginary frequency of phonon modes indicates the dynamical instability under such strain. The dynamical stable range of MLP for strain $\sigma$ is about ($-3\%,12\%$). Within this range, the frequencies of Raman active modes under several strains are also listed in Table 3. The response of each Raman mode to strain is not the same. Four modes ( $B_{1g}, A_{g}^{1}, B_{2g}^{1},B_{2g}^{2}$) exhibit monotonic change: their frequencies are red shifted under tensile strain and blue shifted under compressive strain. Among them, $A_{g}^{1}, B_{2g}^{2}$ modes which are vibrated mainly along the z direction have prominent shifts. The other two Raman modes exhibit nonmonotonic change. $B_{3g}$ mode exhibits an abnormal red shift under small compressive strain. This is possibly due to the fact that the flattening of the puckered layer under compressive strain is in favor of the opposite vibrations of two layers along the zigzag chain. $A_{g}^{2}$ mode exhibits a nonmonotonic shift under tensile strain. Its frequency decreases at small tensile strain and then increases with increasing tensile strain. This abnormal behavior can be understood from the following: $A_{g}^{2}$ mode corresponds to the vibration of $d_{1}$ bond-breathing mainly along $x$ direction. The lattice constant $a$ is contracted when vertical tensile strain is applied. When the effect of this contraction of $a$ dominates over the effect of increase of thickness of phosphorene, it will enhance the interatomic interactions, resulting in the increase of frequency of $A_{g}^{2}$ mode.
For BLP, we find its dynamical stable range for strain $\sigma$ is wider than that of MLP. When compressive strain reaches 8$\%$, imaginary frequency appears in the phonon dispersion along $\Gamma$-X direction. Within the stable dynamical range for strain $\sigma$, the frequencies of Raman active modes are also listed in Table 3. Under compressive strain, $B_{3g}$ mode exhibits an abnormal red shift as in the case of MLP, while the frequencies of other Raman modes are blue shifted as expected as usual. BLP is still stable when tensile strain reaches 40$\%$. As discussed above, large tensile strain only results in large interlayer distance with other geometric parameters in close to those in the MLP. Then BLP may be regarded as two nearly independent MLPs, which are stable in free-standing form as discussed above. Under the wide range of tensile strain, the changes of frequencies of four modes ( $B_{3g},B_{1g}, A_{g}^{2}, B_{2g}^{1}$) are all very small. This shows that the vertical tensile strain has little effect on the in-plane vibrational modes. With the increase of tensile strain, $A_{g}^{1}, B_{2g}^{2}$ modes vibrated mainly along the $z$ direction are red shifted at first as expected, but they are then blue shifted. This abnormal blue shift may be due to the decrease of attractive vdW interaction at large interlayer distance.
Superconductivity by adjusting the interlayer distance
------------------------------------------------------
The above phonon calculations show that BLP is dynamical unstable when it achieves the transition from semiconductor to metal. Here we propose that stable metal phase or even a BCS superconductor can be achieved by adjusting the interlayer distance of BLP. We fix the interlayer distance at different given values, the unit cell and other atomic positions are then relaxed. Since the results with increasing the interlayer distance are very similar to those in the tensile strained BLP, next we only discuss the opposite case.
When the interlayer distance $0.6d_{0} < d < d_{0}$ ($d_{0}$ is the interlayer distance of free BLP), our calculations show that the lattice constant $a$ increases slightly, $d_{2}$ and $\theta_{2}$ show only a weak variation, while the lattice constant $b$, $d_{1}$ and $\theta_{1}$ hardly change. This means that the puckered character of BLP remains and changes little at this range of interlayer distance. When $d < 0.6d_{0}$, the structure of BLP starts to change substantially. The calculations of band structure show that the band gap $E_{g}$ goes to zero when $d = 0.85d_{0}$. Keeping on decreasing the interlayer distance will result in the transition from semiconductor to metal. The calculations of phonon show that BLP is dynamical unstable when $d$ is $0.65d_{0}$ or less. So the range of $d$ for the stable metal phase of BLP is about $0.65d_{0} < d < 0.85d_{0}$.
The Eliashberg spectral function $\alpha^{2}F(\omega)$ depends directly on the EP matrix element, $g_{{\bf{k}+\bf{q}}j^{'}{\bf{k}}j}^{\bf{q}\nu}$, which can be determined self-consistently using linear response theory. By plotting $\alpha^{2}F(\omega)$ we can estimate the relative strength of the EP coupling. The phonon density of states $F(\omega)$ and $a^{2}F(\omega)$ are plotted in Fig. 4 for three different interlayer distances, which are within the range of stable metal phase of BLP. When $d = 0.8d_{0}$, $d = 0.75d_{0}$ and $d =
0.7d_{0}$, $F(\omega)$ is very similar, while $a^{2}F(\omega)$ is different. With smaller interlayer distance, the Eliashberg spectral function $a^{2}F(\omega)$ has a significantly enhanced peak at the low-frequency side. The characteristic phonon modes which dominate the EP coupling are just from the opposite vibrations of two puckered double layers. So we speculate that the interlayer vdW attractive interaction may play an important role in enhancing the EP coupling of BLP with smaller interlayer distance.
![The phonon density of states $F(\omega)$ and the Eliashberg spectral function $\alpha^{2}F(\omega)$ for BLP at three different interlayer distances.[]{data-label="fig.4"}](fig4.pdf){width="4.in"}
The superconducting temperature $T_{c}$ can be estimated from the Allen-Dynes modified McMillan equation[@22]: $$T_{C}=
\frac{\omega_{ln}}{1.2}\exp
\left(-\frac{1.04(1+\lambda)}{\lambda-\mu^{*}(1+0.62\lambda)}\right),$$ where $\omega_{ln}$ is the logarithmically averaged frequency, $\mu^{*}$ is the Coulomb repulsion parameter. The calculated electronic densities of states at Fermi level $N(E_{F})$, EP coupling constant $\lambda$, $\omega_{ln}$ and the estimated $T_{C}$ when taking $\mu^{*}=0.1$ are summarized in Table 4. With the decrease of interlayer distance, BLP apparently becomes more metallic and results in the increase of $N(E_{F})$. Most remarkably, we see $\lambda$ increases dramatically. The increase of both $N(E_{F})$ and $\lambda$ is favorable to enhance $T_{C}$. When $d = 0.7d_{0}$, $\lambda$ reaches as high as 1.45 and the estimated $T_{C}$ is about 10K. The superconductivity of BP was reported early under high pressure.[@5; @6] Our results suggest that BLP may become a good BCS superconductor by adjusting the interlayer distance.
Summary
=======
In conclusion, MLP and BLP under vertical strain are studied by density functional and density-functional perturbation theory. The results show that the electronic structure of thin BP is very sensitive to its thickness and the strain applied. The change is extensive. It can increase direct band gap, decrease indirect band gap or realize the transition from semiconductor to metal, which help to tailor materials in a variety of settings, from infrared optoelectronics to high-mobility quantum transport. The shift of zone-center Raman active modes for strained phosphorenes is analyzed and explained combining the relaxation of geometric structure. Our study clearly show that both MLP and BLP in free-standing form are dynamical stable. However, material becomes dynamical unstable with large vertical strain. We find that the dynamical stable range for strain $\sigma$ in BLP is wider than that in MLP due to add the interlayer weak vdW coupling, this additional freedom. This information is essential for future device fabrication and potential applications of phosphorene. Furthermore we also find that BLP may become a good BCS superconductor by adjusting the interlayer distance.
Acknowledgments
===============
The authors acknowledge the support of the Natural Science Foundation of Jiangsu Province in China under Grant No. BK20141441,the State Key Program for Basic Researches of China (2014CB921103 and 2010CB923404), the National “Climbing” Program of China (91021003), and the National Science Foundation of Jiangsu Province (BK2010012).
[0]{} C. M. Park and H. J. Sohn, Adv. Mater. 19, 2465 (2007). L. Q. Sun, M. J. Li, K. Sun, S. H. Yu, R. S. Wang, and H. M. Xie, J. Phys. Chem. C, 116,14772 (2012). Y. Ktayama, T. Mizutani, W. Utsumi, O. Shimomura, M. Yamakata, K. Funakoshi, Nature, 403,170 (2000). C. A. Vanderborgh and D. Schifer, Phys. Rev. B, 49, 9595 (1989). H. Kawamura, I. Shirotani, K. Tachikawa, Solid State Commun, 49, 879, (1984). J. Wittig, B. T. Matthias, Science 160, 994, (1968). L. Li, Y. Yu, G. J. Ye, Q. Ge, X. Ou, H. Wu, D. Feng, X. H. Chen and Y. Zhang, Nature Nanotech. 9, 372 (2014). H. Liu, A. T. Neal, Z. Zhu, D. Tom$\acute{a}$nek and P. D. Ye, ACS Nano 8, 4033 (2014). S. P. Koenig, R. A. Doganov, H. Schmidt, A. H. Castro Neto and O. Barbaros, Applied Physics Letters 104, 103106 (2014). Ruixiang Fei, Alireza Faghaninia, Ryan Soklaski, Jia-An Yan, Cynthia Lo, Li Yang, Preprint at http://arxiv.org/abs/1405.2836 (2014). M. S. Scheurer, J. Schmalian, Preprint at http://arxiv.org/abs/1404.4030 (2014). R. W. Keyes, Phys. Rev. 92, 580 (1953). J. S. Qiao, X. H. Kong, Z. X. Hu, F. Yang, W. Ji, Nature Communications 5, 4475 (2014) A. S. Rodin, A. Carvalho, and A. H. Castro Neto, Phys. Rev. Lett. 112, 176801 (2014). A. Castellanos-Gomez *et al.*, 2D Matererials 1, 025001 (2014). R. X. Fei and L. Yang, Applied Physics Letters 105, 083120 (2014). P. Giannozzi *et al.*, *J. Phys.Condens. Matter* [**21**]{} (2009) 395502. and http://www.quantum-espresso.org S. Grimme, J. Comput. Chem. 27, 1787,(2006). Y. Takao, Physica (Amsterdam) 105B, 93 (1981). Y. L. Du, C. Y. Ouyang, S. Q. Shi, and M. S. Lei, J. Appl. Phys. 107, 093718 (2010). G. Z. Qin, Z. Z. Qin, S. Y. Yue, H. J. Cui, Q. R. Zheng, Q. B. Yan, and G. Su, Preprint at http://arxiv.org/abs/1406.0261 (2014). P. B. Allen and R. C. Dynes, Phys. Rev. B [**12**]{}, 905 (1975).
[**Figure Caption**]{}
Figure 1. The top (a) and side views (b) of the atomic structure of phosphorene. (c) Vibrational pattern of $B_{3g}$ and $B_{1g}$ modes.\
\
Figure 2. The band structures of the strained phosphorene, (a)-(e) are for MLP; (f)-(i)are for BLP. (j) The evolution of the band gap $E_{g}$ with the strain.\
\
Figure 3. The phonon dispersion of free MLP (a) and BLP (b).\
\
Figure 4. The phonon density of states $F(\omega)$ and the Eliashberg spectral function $\alpha^{2}F(\omega)$ for BLP at three different interlayer distances.
------------------------- ------------ ------------ ------------ ------------------ ------------------ ----------------- ------------------
a(${\AA}$) b(${\AA}$) c(${\AA}$) $d_{1}$(${\AA}$) $d_{2}$(${\AA}$) $\theta_{1}$ $\theta_{2}$
Exp. (Ref.) 4.376 3.314 10.478 2.224 2.244 96.34$^{\circ}$ 102.09$^{\circ}$
Theo.(Ref.) 4.422 3.348 10.587 2.238 2.261 96.85$^{\circ}$ 102.31$^{\circ}$
This work (with vdW) 4.398 3.325 10.429 2.227 2.257 96.56$^{\circ}$ 102.20$^{\circ}$
This work (without vdW) 4.503 3.311 11.033 2.227 2.261 96.07$^{\circ}$ 103.03$^{\circ}$
------------------------- ------------ ------------ ------------ ------------------ ------------------ ----------------- ------------------
: Structural parameters of bulk BP.[]{data-label="tab.1"}
---------- ------------ ------------ ------------------ ------------------ ----------------- ------------------
$\sigma$ a(${\AA}$) b(${\AA}$) $d_{1}$(${\AA}$) $d_{2}$(${\AA}$) $\theta_{1}$ $\theta_{2}$
MLP
0.15 4.281 3.293 2.198 2.545 97.04$^{\circ}$ 100.27$^{\circ}$
0.10 4.379 3.294 2.210 2.451 96.38$^{\circ}$ 101.24$^{\circ}$
0.05 4.441 3.311 2.224 2.355 96.22$^{\circ}$ 102.04$^{\circ}$
0.00 4.525 3.311 2.222 2.255 96.35$^{\circ}$ 103.35$^{\circ}$
-0.05 4.671 3.302 2.215 2.200 96.40$^{\circ}$ 105.10$^{\circ}$
-0.10 4.904 3.292 2.201 2.159 96.81$^{\circ}$ 107.74$^{\circ}$
-0.25 5.729 3.287 2.176 2.151 98.14$^{\circ}$ 116.00$^{\circ}$
BLP
0.40 4.527 3.308 2.221 2.256 96.27$^{\circ}$ 103.33$^{\circ}$
0.15 4.523 3.308 2.222 2.261 96.21$^{\circ}$ 103.31$^{\circ}$
0.10 4.513 3.310 2.223 2.265 96.20$^{\circ}$ 103.16$^{\circ}$
0.05 4.486 3.311 2.225 2.267 96.19$^{\circ}$ 102.92$^{\circ}$
0.00 4.485 3.314 2.225 2.259 96.24$^{\circ}$ 102.98$^{\circ}$
-0.05 4.505 3.320 2.226 2.233 96.46$^{\circ}$ 103.33$^{\circ}$
-0.10 4.607 3.324 2.225 2.208 96.67$^{\circ}$ 104.41$^{\circ}$
-0.20 4.994 3.390 2.220 2.199 99.55$^{\circ}$ 107.91$^{\circ}$
---------- ------------ ------------ ------------------ ------------------ ----------------- ------------------
: Structural parameters of strained MLP and BLP. []{data-label="tab.2"}
---------- ---------- ---------- -------------- -------------- ------------- -------------
$\sigma$ $B_{3g}$ $B_{1g}$ $B_{2g}^{1}$ $B_{2g}^{2}$ $A_{g}^{1}$ $A_{g}^{2}$
MLP
0.00 196 433 226 427 368 456
MLP
0.12 170.1 417.4 152.5 276.6 147.8 455.6
0.10 175.2 415.8 162.8 297.1 152.0 444.0
0.08 180.9 416.2 175.9 320.1 170.1 433.5
0.05 187.2 418.1 196.0 357.8 266.9 431.4
0.00 188.2 421.9 222.9 424.5 349.7 449.9
-0.02 187.8 424.3 226.8 437.5 354.5 459.2
-0.03 185.4 425.5 230.1 449.0 359.2 466.6
BLP
0.40 188.9 422.3 221.9 423.4 349.5 449.5
0.15 189.0 421.7 221.4 420.3 345.4 448.2
0.10 189.1 421.2 221.1 418.0 344.9 448.5
0.05 188.9 421.2 220.0 414.7 341.5 446.2
0.00 188.1 422.6 221.2 419.9 346.7 450.3
-0.03 186.2 424.2 226.3 431.0 353.6 452.6
-0.05 185.8 426.0 228.1 436.7 357.8 463.0
---------- ---------- ---------- -------------- -------------- ------------- -------------
: Frequencies ($cm^{-1}$)of Raman modes of strained phosphorene. []{data-label="tab.3"}
----------- ----------------------- ----------- ------------------ -------------
d/$d_{0}$ $N(E_{F})(States/eV)$ $\lambda$ $\omega_{ln}(K)$ $T_{C} (K)$
0.80 0.68 0.32 183.2 0.2
0.75 0.86 0.85 107.9 5.7
0.70 1.08 1.45 96.1 10.6
----------- ----------------------- ----------- ------------------ -------------
: The calculated electronic densities of states at Fermi level $N(E_{F})$, EP coupling constant $\lambda$, the logarithmically averaged frequency $\omega_{ln}$ and the estimated $T_{C}$ for BLP at three different interlayer distances.[]{data-label="tab.4"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Tübingen NLTE Model-Atmosphere Package (*TMAP*) calculates fully metal-line blanketed white dwarf model atmospheres and spectral energy distributions (SEDs) at a high level of sophistication. Such SEDs are easily accessible via the German Astrophysical Virtual Observatory (*GAVO*) service *TheoSSA*. We discuss applications of *TMAP* models to (pre) white dwarfs during the hottest stages of their stellar evolution, e.g. in the parameter range of novae and super soft sources.'
---
Introduction {#sect:introduction}
============
Novae in outburst may become the brightest sources of soft X-ray emission in the sky. During this so-called super soft source (SSS) phase (some weeks or months after the outburst), it is possible to perform high-resolution and high-S/N spectroscopy with X-ray satellites like *Chandra* or XMM-*Newton*. These spectra require adequate NLTE model atmospheres for a reliable analysis.
A detailed analysis of such a nova, V4743Sgr, was recently presented by [@Rauch_etal_2010]. They used plane-parallel, hydrostatic models calculated with the Tübingen NLTE Model-Atmosphere Package (). Although the velocity field and the expansion of the nova’s atmosphere was neglected, the overall slope of the continuum flux was well reproduced. Moreover, the strengths of prominent photospheric absorption lines (C[v]{}, C[vi]{}, N[vi]{}, N[vii]{}, O[vii]{}) as well as the strengths of absorption edges were in very good agreement with the observation (Fig.\[fig:cn\], \[fig:final\]). [@Rauch_etal_2010] could e.g. show that the C/N abundance ratio was increasing from March to September 2003 (Fig.\[fig:cn\]).
![Comparison of flux-calibrated *Chandra* observations of the nova V4743Sgr with synthetic spectra (H+He+C+N+O+Ne+Mg+Si+S + (Ca$-$Ni) opacities considered in modeling, $T_\mathrm{eff}\hspace{-0.5mm}=\hspace{-0.5mm}720\,000\,\mathrm{K}$, $\log g = 9$, cf. [@Rauch_etal_2010 Rauch 2010]). Positions of line transitions are marked at top. $v_\mathrm{rad} = -2300\,\mathrm{km/sec}$ was applied to the models to match the C[vi]{} and N[vi]{} resonance lines. \[X\] denotes log(abundance / solar abundance). For clarity, the July and September observations are artificially shifted in flux.[]{data-label="fig:cn"}](rauch_01.eps){width="\textwidth"}
Model Atmospheres for SSSs {#sect:model}
==========================
*TMAP* model atmospheres were successfully employed for spectral analyses of hot, compact stars. See [@RauchWerner_2010] and [@Rauch_2011] for a brief summary of this application to a variety of objects from post-AGB stars ($T_\mathrm{eff}\hspace{-0.5mm}\approx\hspace{-0.5mm}100\,000\,\mathrm{K}$) to neutron stars in low-mass X-ray binaries ($T_\mathrm{eff}\hspace{-0.5mm}\approx\hspace{-0.5mm}10\,000\,000\,\mathrm{K}$).
[*What do we need to analyze X-ray spectra taken during the SSS phase following a nova outburst reliably?*]{} The evolution of both stars has to be modeled, including e.g. stellar rotation, mass-loss history, common-envelope phase, and a long series of outbursts. Thus, 3-dimensional magneto-hydrodynamical calculations that follow the stars for a long time are definitely the best choice. In addition, 3D NLTE radiation transfer, accounting for the expanding atmosphere and mass loss not only during the outburst, clumping in the ejected matter, etc. is necessary. This should be combined with 3D photoionization modeling to describe the surrounding gas and dust. Although all these state-of-the-art ingredients are already available, their combination and application to an individual nova system is still a dream.
In the case of V4743 Sgr, [@Ness_etal_2003] measured a blueshift ($v_\mathrm{rad} = - 2000\,\mathrm{km/sec}$) in the *Chandra* spectrum. This strong evidence for expansion was found in the SSS spectra of other novae as well ([@Ness_2010 Ness 2010]). A new version of the multi-purpose model-atmosphere code *PHOENIX* ([@HauschildtBaron_1999 Hauschildt & Baron 1999]) accounts for this expansion ([@vanRossumNess_2010 van Rossum & Ness 2010]). These models predict typical wind effects that are already known from the UV wavelength range in cooler post-AGB stars. In the relevant parameter range for SSSs, the model flux for $\lambda < 30\,\mathrm{\AA}$ ($E > 0.41\,\mathrm{keV}$) is increasing with the mass-loss rate and absorption edges of C[vi]{} (25.30Å), N[vii]{} (18.59Å), and O[vii]{} (16.77Å) become weaker. From these results, one has to conclude that, of course only in case of higher mass-loss rates ($> 10^{-8}\,\mathrm{M_\odot / yr}$), spectral analyses of SSSs based on hydrostatic model atmospheres tend to overestimate $T_\mathrm{eff}$ and metal abundances.
The question remains, [*why do the static TMAP models reproduce the observations that well*]{} (Fig.\[fig:final\])[*?*]{} The answer may be very simple – the strong mass-loss phase of V4743 Sgr is over. Obviously, there are no PCygni line profiles observed and thus, the “wind-blown” material around V4743 Sgr is not dense enough to provide the emission. The matter in the line of sight is still moving towards us and thus, we see blueshifted, pure absorption lines. An approach like presented by [@Rauch_etal_2010] appears well justified. It is possible that accretion already takes place and changes the C/N abundance ratio (Sect.\[sect:introduction\]).
![Comparison of flux-calibrated XMM-*Newton* observations of V4743Sgr (April 2003) with the final model (H+He+C+N+O+Ne+Mg+Si+S + (Ca-Ni), $T_\mathrm{eff}\hspace{-0.5mm}=\hspace{-0.5mm}740\,000\,\mathrm{K}$, $\log g = 9$) of [@Rauch_etal_2010]. Positions of ground-state thresholds are marked at top (blue), those of line transitions just below that marks (red).[]{data-label="fig:final"}](rauch_02.eps){width="\textwidth"}
*TheoSSA* - SSSs Model Spectra on Demand {#sect:theossa}
========================================
*TheoSSA*[^1] is a registered Virtual Observatory[^2] (*VO*) service developed by the German Astrophysical Virtual Observatory *GAVO*[^3]. It provides access to SEDs at three levels: 1) pre-calculated SED grids for a fast comparison with observations can be downloaded, 2) SEDs with individual parameters (and provided standard model atoms from the Tübingen Model-Atom Database, *TMAD*[^4]) can be calculated[^5] without knowledge of the model-atmosphere code, 3) SEDs with individual parameters and self-constructed model atoms can be computed by experienced users. *TheoSSA* was firstly based on *TMAP* models only but it is prepared for SEDs from any other stellar model-atmosphere code.
Pre-calculated grids of SEDs for SSSs ([@RauchWerner_2010 cf. Rauch & Werner 2010]) are also easily accessible in *VO*-compliant form via *TheoSSA*’s WWW interface[^6]. In case of any request, we will calculate new grids with different abundances.
In addition, some SED grids for SSSs ($T_\mathrm{eff}\hspace{-0.5mm}=\hspace{-0.5mm}0.55 - 1.05\,\mathrm{MK}$, [@RauchWerner_2010 Rauch & Werner 2010]) are available converted into atables[^7] for the direct use within *XSPEC*[^8].
Conclusions {#sect:conclusions}
===========
Spectral energy distributions for SSSs (calculated with *TMAP*) are easily available via *TheoSSA* – use them! Blackbody energy distributions generally have – at the same temperature – a much lower peak intensity (about a factor of three in the parameter range of SSSs, [@Rauch_etal_2010 Rauch 2010]) and the flux maximum is located towards lower energies (about a factor of two). Thus, “spectral analyses” based on blackbodies yield wrong results.
There are novae that are observed in their SSS phase where the impact of mass-loss is not significant. For these objects, analyses based on static model atmospheres like *TMAP* provide reliable results within typical error ranges ([@Rauch_etal_2010 Rauch 2010]). In case of prominent PCygni line profiles, however, a code like *PHOENIX* has to be employed.
A time series (about all three months) of V4743 Sgr X-ray observations has shown that the white dwarf’s surface is extremely hot ($T_\mathrm{eff}$ higher that 500000K) for at least half a year ([@Rauch_etal_2010 Rauch 2010]). The phase of increasing $T_\mathrm{eff}$ appears important because the mass-loss rate will be presumably higher and, hence, expanding model atmospheres can be used to study wind properties.
A time series of X-ray spectra on a short time scale (about weekly) starting at an early outburst phase and covering the SSS phase like taken e.g. for Nova V1974 Cyg in the ultraviolet with IUE[^9] (cf. [@GonzalesKrautter_1998 González-Riestra & Krautter 1998]) is highly desirable. Since *Chandra* and XMM-*Newton* will unfortunately not work forever, such spectra have to be taken in the very near future. The development of appropriate, reliable model atmospheres is already on the way.
This work is supported by the German Aerospace Center (DLR, grant 05OR0806), the German Research Foundation (DFG, grant WE1312/41$-$1), and the Federal Ministry of Education and Research (BMBF, grant 05A11VTB).
González-Riestra, R., & Krautter, J. 1998, in: W. Wamsteker, R. González-Riestra (eds.), *Ultraviolet Astrophysics Beyond the IUE Final Archive*, ESA SP, 413, 367
Hauschildt, P. H., & Baron, E. 1999, *J. of Computational and Applied Mathematics*, 109, 41
Ness, J.-U., Starrfield, S., Burwitz, V., 2003, *ApJ*, 594, L127
Ness, J.-U. 2010, *AN*, 331, 179
Rauch, T. 2011, in: *The X-ray Universe 2011*, available online at: http://xmm.esac.esa.int/ external/xmm\_science/workshops/2011symposium/, 61
Rauch, T., Orio, M., Gonzáles-Riestra, R., Nelson, T., 2010, *ApJ*, 717, 363
Rauch, T., & Deetjen, J. L. 2003, in: I. Hubeny, D. Mihalas, K. Werner (eds.), *Stellar Atmosphere Modeling*, The ASP Conference Series (San Francisco: ASP), 288, 103
Rauch, T., & Werner, K. 2010, *AN*, 331, 146
Rauch, T., Ringat, E., & Werner, K. 2011, *Universal Academy Press*, in press, arXiv:1011.3628
van Rossum, D. R., & Ness, J.-U. 2010, *AN*, 331, 175
Werner, K., Dreizler, S., Deetjen, J. L., Nagel, T., Rauch, T., & Schuh, S. L. 2003, in: I. Hubeny, D. Mihalas, K. Werner (eds.), *Stellar Atmosphere Modeling*, The ASP Conference Series (San Francisco: ASP), 288, 31
[^1]: Theoretical Stellar Spectra Access
[^2]: http://www.ivoa.net
[^3]: http://www.g-vo.org
[^4]: http://astro.uni-tuebingen.de/TMAD/TMAD.html
[^5]: http://astro.uni-tuebingen.de/TMAW/TMAW.shtml
[^6]: http://dc.g-vo.org/theossa
[^7]: http://astro.uni-tuebingen.de/rauch/TMAF/flux\_HHeCNONeMgSiS\_gen.html
[^8]: http://heasarc.gsfc.nasa.gov/docs/xanadu/xspec
[^9]: International Ultraviolet Explorer
| {
"pile_set_name": "ArXiv"
} |
---
date: 'January, 1994'
---
phyzzx.tex \#1[[*Lett. Nuovo Cim.*]{} [**[\#1]{}**]{}]{} \#1[[*Riv. Nuovo Cim.*]{} [**[\#1]{}**]{}]{} \#1[[*Phys. Rev.*]{} [**D[\#1]{}**]{}]{} \#1[[*Nucl. Phys.*]{} [**B[\#1]{}**]{}]{} \#1[[*Phys. Rev. Lett.*]{} [**[\#1]{}**]{}]{} \#1[[*Phys. Lett.*]{} [**B[\#1]{}**]{}]{} \#1[[*Z. Phys.*]{} [**C[\#1]{}**]{}]{} \#1[[*Rev. Mod. Phys.*]{} [**[\#1]{}**]{}]{} \#1[[*Phys. Rep.*]{} [**[\#1]{}**]{}]{}
\#1\#2 \#1[\#1|]{} \#1[| \#1]{} \#1[\#1]{} ‘=11 \#1[$\bf#1$]{} \#1[$\bf\overline{#1}$]{} 1[[1]{}]{} \#1\#2 /\#1[\#1-6pt/]{} \#1\#2\#3[Nucl. Phys. B [**\#1**]{} (19\#2) \#3]{} \#1\#2\#3[Phys. Lett. B [**\#1**]{} (19\#2) \#3]{} \#1\#2\#3[B [**\#1**]{} (19\#2) \#3]{} \#1\#2\#3[Phys. Rev. D [**\#1**]{} (19\#2) \#3]{} \#1\#2\#3[Phys. Rev. Lett. [**\#1**]{} (19\#2) \#3]{} \#1\#2\#3[Phys. Rep. [**\#1**]{} (19\#2) \#3]{} \#1\#2\#3[Mod. Phys. Lett. A [**\#1**]{} (19\#2) \#3]{} \#1\#2\#3[Int. J. Mod. Phys. A [**\#1**]{} (19\#2) \#3]{} \#1[Texas A & M University preprint CTP-TAMU-\#1]{} \#1\#2\#3[Ann. Rev. Astron. Astrophys. [**\#1**]{} (19\#2) \#3]{} \#1\#2\#3[Ann. Rev. Nucl. Part. Sci. [**\#1**]{} (19\#2) \#3]{}
=6.5in =8.5in
**EXPLORING THE YUKAWA UNIFIED MINIMAL**
.05in
**SUPERGRAVITY MODEL AT THE TEVATRON, LEP II AND THE LHC**
J. F. GUNION and H. POIS
*Davis Institute for High Energy Physics,*
*Department of Physics, U. C. Davis, Davis, CA 95616*
.075in
**Abstract**
.075in
.15in .075in Incorporation of the Standard Model (SM) into a more fundamental theory that includes supersymmetry (SUSY) is both aesthetically attractive and a theoretically compelling solution to the naturalness and hierarchy problems. Although sparticles have not been directly detected to date, and we still lack a realistic mechanism for SUSY breaking, the success of SUSY gauge coupling unification and the accurate prediction of $\sin\theta_W$, along with an elegant radiative electroweak symmetry breaking (EWSB) scenario(which requires a heavy top quark) all provide dramatic indirect support for SUSY.
Squark, gluino, slepton, chargino and neutralino production, decay and detection have been studied in a generic fashion at LEP-200, Fermilab and hadron supercolliders,as well as in supergravity GUT models at LEP-200, Fermilab and HERA.
The phenomenology of the five physical Higgs bosons (the $\hl,\hh,\ha,\hpm$) of the Minimal Supersymmetric Model (MSSM), has also been examined at $\epem$ and hadron supercolliders, including one-loop radiative corrections. For recent reviews of the hadron collider phenomenology, and further references, see Refs. \[,\]. If we assume that SM final state modes dominate Higgs decays, the discovery channels of most obvious utility at hadron colliders are: i) $\hl,\hh,\ha\rta \gamma \gamma$; ii) $l(\hl,\hh\rta \gamma \gamma)$; iii) $\hl,\hh\rta 4l$; iv) $t\rta \hp b$; and possibly v) $\hh,\ha\rta \tau^+\tau^-$. Further, $\sim 30\%$ efficiency and $\sim 1\%$ purity for $b$-tagging will allow neutral and charged Higgs detection in the production/decay modes $t\bar t (\hl,\hh\rta b\bar b)$and $t \bar b (\hp\rta t\bar b)$ for some parameter choices. In particular, the $t\anti t(\hl,\hh\rta b\anti b)$ mode combined with i)-iv) above implies that there is a no-lose theorem, according to which detection of at least one of the MSSM Higgs bosons will be possible at the LHC. However, over much of parameter space, in particular for $\mha\gsim 100-200\gev$, only the $\hl$ will be easily detected, and it will be relatively SM-like.
However, if the branching ratios for $\hl,\hh,\ha,\hpm \rta {\rm sparticles}$ are large, Higgs detection in direct SM final state decay modes will become more difficult, while new opportunities in the sparticle channels will arise. Supergravity theories provide a very attractive framework for evaluating the effects of sparticle pair channels, while simultaneously allowing an analysis of sparticle decays as they affect not only the detection of such indirectly produced sparticles but also direct sparticle searches. Assuming coupling constant unification and universal soft-SUSY breaking at the unification scale $\mgut$, just [*five*]{} independent parameters are needed to specify the models. One possible set comprises $\lambda_t(\mgut),\mhalf,A,m_0,\mu_0=\mu(\mgut)$. Here, $m_0,A,\mhalf$ are the (common) scalar, $A$-term and gaugino masses at $\mgut$, respectively. Alternatively, a bottom-up approach can be employed in which all (five) input parameters are specified at the weak scale. In this work, we adopt a bottom-up approach but simplify the parameter space further by exploring the ‘yukawa unified’ ($\lambda_b(\mgut)=\lambda_\tau(\mgut)$) minimal supergravity (YUMS) model in which the soft-SUSY-breaking parameters $m_0$ and $A$ are taken to be zero. The particle content is that of the MSSM: SM fields; two Higgs doublets; and their superpartners. We require gauge coupling unification, but do not insist on any particular GUT embedding (thus, proton decay need not be a difficulty). The combination of gauge and yukawa unification implies that a choice for $\tanb$ determines a value for $\mt$, independently of the soft-SUSY breaking parameters. (Higgs phenomenology could be pursued at this point, for generic choices of the remaining parameters; see Ref. \[\].) If in addition we require $m_0=A=0$, specification of just $\tanb$ and $\mha$ (at low-energy), along with the sign of $\mu$, completely determines the supergravity model. Although we allow for $B_0\equiv B(\mgut)\neq0$, the requirement $B_0=0$ yields an interesting special case of the YUMS model. (Indeed, if one wants to assume for simplicity that only one type of soft SUSY breaking is dominant, the only possibility not ruled out by theoretical and/or phenomenological constraints is to have $\mhalf\gg m_0,A,B_0$.) In this case, specifying $\tanb$ alone (and the sign of $\mu$) is sufficient to parameterize the model, since there is a unique value for $\mha$ such that $B_0=0$. Full details regarding our bottom-up approach to the renormalization group equations will be presented in a future paper. The YUMS boundary conditions have significant motivation. Yukawa unification is especially motivated in that: i) the constraint arises naturally in GUT models where the $b,\tau$ fermions both reside in the same representations; and ii) the $\lambda_b(M_U)=\lambda_{\tau}(M_U)$ condition constrains the model to be close to the RGE-fixed point region where $\lambda_t(M_U)\gsim 1$,and naturally predicts a large top quark mass.
As for the $m_0=A=0$ assumption, Ref. \[\] notes that soft mass and tri-linear terms are generally not flavor or generation universal in generic supergravity models, and therefore may lead to large flavor-changing neutral currents.If the scalar masses are required to be [*zero*]{} at $\mgut$, then the RGE’s automatically maintain sufficient degeneracy among the squarks to suppress FCNC’s. In the context of supergravity, the $m_0=A=0$ boundary condition arises naturally in the ‘no-scale’ theories. However, we must note that for the [*known*]{} explicit realizations of this boundary condition at the string scale in superstring theories, threshold corrections are apparently not large enough to allow gauge coupling unification and $m_0=A=0$ at the lower $\mgut$ scale. However, employing these boundary conditions at $\mgut$ could still be a reasonable first approximation.
Given the boundary condition $m_0=0$, comparatively low masses ( $m_{\tilde l_R}\simeq 0.15-0.19 \mgl$, and $\msq\simeq 0.86-0.92 \mgl$) are predicted for the $\tilde l,\tilde q$ SUSY partners. This has crucial implications for Higgs/sparticle phenomenology. As an important example, in addition to the invisible decay $\hl\rta \cnone \cnone$ (discussed, for instance, in Refs. \[,,\]) we find that the (invisible) $\hl\rta\snu\snubar$ decay is also kinematically accessible in a portion of the parameter space and can even dominate over the $\hl\rta b \anti b$ mode. These, and other sparticle decay modes of the Higgs bosons, must be included in assessing Higgs detection in the context of $m_0\sim 0$ models.
We note that the possibility of detecting a generic invisibly-decaying Higgs boson at a hadron supercollider, $\h\rta I$ decay (where $I$ is an unspecified invisible channel), has been explored. The two possible detection modes at a hadron supercollider are $t\bar t+\h\rta \ell+jets+\ptmiss$and $W+\h\rta \ell+\ptmiss$,with the optimistic conclusion that even an invisibly decaying Higgs can be detected so long as its $\h t\bar t$ and $\h \wp\wm$ couplings, respectively, are not too suppressed compared to SM strength. Variation of these couplings can be determined in our constrained YUMS model, and the invisible mode procedures of Ref. \[\] are incorporated in our analysis.
.15in .075in
In this section, we present the basic sparticle and Higgs boson phenomenology of the YUMS model at the Tevatron, LEP-II and the LHC. We begin by determining the region of $\tanb,\mha$ parameter space that is allowed by current experimental limits on the masses of the various sparticles and Higgs bosons in combination with the requirements of a neutral LSP, correct symmetry breaking (including the demand that $V<0$ at the minimum) and a bounded Higgs potential at $\mgut$. A precise delineation of the input limits and experimental constraints will be given in Ref. \[\]. Considerations of dark matter/cosmology are not included in our analysis; however, we do not expect this to be at all limiting.Constraints from $BR(b\rta s\gamma)$ do not further restrict our parameter space if full account is taken of both the experimental errors and the significant uncertainties in the theoretical computation (mainly arising from strong dependence on the undetermined renormalization scale $\mu_{QCD}$). Finally, we note that the proton lifetime can be made acceptably large in the YUMS model. For example, non-minimal $SU(5)$ models can be constructed in which dangerous dimension-5 graphs do not arise. However, minimal $SU(5)$ embedding would be problematical because of the large super-heavy triplet masses required and the resulting mass hierarchy problems. For our figures, we present the $\mu>0$ case. The allowed region of yukawa-unified minimal supergravity (YUMS) parameter space is displayed in Fig. . The left and right sides of the solid line boundary are essentially obtained by the requirements $\mslep>45\gev$ and $\mstauone>\mcnone$ ( a neutral LSP), respectively. The bottom edge of the allowed region is a result of the LEP-I Higgs constraint $\mhl\gsim 60\gev$. We note that for $\tanb\simeq 1.5$ our $\mha$ limits agree with the third reference in Ref. \[\].
Note that the combination of the yukawa unification and $m_0=A=0$ boundary conditions at $\mgut$ and the general phenomenological constraints has severely limited the range of $\tanb,\mha$ values allowed. The precise region is somewhat sensitive to the values taken for $m_b(m_b)$ and $\alpha_s(\mz)$. However, for our choices of $m_b(m_b)=4.25\gev$ and $\alpha_s(\mz)=0.12$, we find $1.5\lsim\tanb\lsim 10$, and $185 \gev\lsim \mha \lsim 690 \gev$. Due to the yukawa unification condition, this immediately implies that $\mt\gsim 155\gev$. As $m_b(m_b)$ is increased, or $\alpha_s(\mz)$ is decreased, slightly higher $\tanb$ values are allowed. Although not shown, for the $\mu<0$ case the right boundary shifts considerably to the left since $\mcnone$ increases in the $\mu<0$ case. This makes the neutral LSP restriction even more severe so that $\mha$ must be $\lsim
380\gev$. Finally, adding the $B_0=0$ constraint to the YUMS boundary conditions yields a relation between $\tanb$ and $\mha$ (see Fig. 2) which is very much within the allowed parameter space.
Limitations on the parameter space from $\rho\equiv\mw^2/(\mz^2\cos^2\theta_W)$ can also be significant in yukawa-unified approaches that lead to parameter space regions with large $\mt$ values. In the YUMS model, our particular choices of $\mb(\mb)=4.25\gev$ and $\alpha_s(\mz)=0.12$ imply that part of the otherwise-allowed parameter space corresponds to $\mt$ values sufficiently large as to be disfavored by $\Delta\rho$. In Fig. , we show the contour for $\Delta\rho/\rho=0.01$. At large $\mha$, contributions to $\Delta \rho$ are dominated by the top/bottom loop and $\mt\lsim 180\gev$ ($\tanb\lsim 3$) is favored (as in the SM). As $\mha$ becomes smaller, stop/sbottom loop corrections also enter and the favored upper limit on $\mt$ (and $\tanb$) decreases. However, if we adopt $\alpha_s(\mz)=0.11$, the maximum $\mt$ allowed by all other constraints is $\lsim 180\gev$, and no additional restriction on allowed parameter space would result from requiring $\Delta\rho/\rho\lsim 0.01$.
A significant feature of the YUMS model, which sets it apart from models with $m_0\gsim \mhalf$, is that squark masses are driven by the RGE’s to much higher values than are the slepton masses — all start from $\sim 0$ at $\mgut$, but only the squark masses have the strong $SU(3)$ group $g_s^2\mgl^2$ driving terms. The result, is a mass ordering roughly described by $\mcnone\simeq 0.08-0.14 \mgl$, $m_{\tilde l_R}\simeq 0.15-0.19 \mgl$, $\mcpone\simeq 0.16-0.26\mgl$, and $\msq\simeq 0.86-0.92 \mgl$. This means that the decays of sparticles and Higgs bosons will be dominated by unexpected channels. For instance, since charged slepton masses are almost always lower than the mass of the lightest chargino, the two-body decay $\cpone\rta \slep^+ \nu$ will almost always be present and will dominate over the three-body (mainly hadronic) modes. Over most of the parameter space the sneutrino is also lighter than the chargino; the relevant contour is shown in Fig. . This implies that $\cpone\rta \ell^+\snu$ decays will also be important. In this same region, the sneutrino decays entirely invisibly ($\snu\rta\cnone\nu$). Note that throughout the allowed parameter space, the second neutralino, $\cntwo$, with mass similar to that of the $\cpone$, has [*both*]{} $\snu \nu$ and $\slep \ell$ decays, whereas $\cnone\hl$ and $\cpmone \wmp$ are never two-body allowed. Meanwhile, the sleptons almost always decay via $\slep^+\rta \ell^+ \cnone$, the only exception being the above-noted border region where $\slep^+\rta\nu\cpone$ is allowed (but with small phase space). Thus, $\slep\,$’s are easily visible. Further, a decent fraction of $\cntwo$ decays are to $\slep\ell$, which thus yields [*several*]{} reasonably energetic ( visible) charged leptons.
Squarks, being lighter than the gluino, will play a prominent role, and will decay primarily via $\sq\rta q\wt\chi$, where $\wt\chi=\cpone$ has a large share of the branching ratio and subsequently decays as described above. Gluino searches will focus on the two-body channel $\gl\rta q \sq$, followed by $\sq$ decay.
Higgs decays also have unexpected features. For $\mha\lsim 250\gev$, the invisible channels $\hl\rta \cnone\cnone,\snu\snubar$ can be significant, depending on the value of $\tanb$. Fig. shows the region where $BR(\hl\rta \cnone\cnone+\snu\snubar)>0.2$. Indeed, for $\tanb\gsim 6$ and $\mha\lsim 200\gev$, we find that $BR(\hl\rta\snu\snubar)\gsim 0.9$ so that invisible modes overwhelm all other decay modes. (However, the only allowed invisible mode in the $\Delta\rho/\rho<0.01$ region is $\hl\rta\cnone\cnone$.) Sparticle-pair channels, $\wt\chi\wt\chi$ and $\slep\slep$, are even more important for the $\hh,\ha,\hpm$ Higgs bosons. (Squark pair modes are not present in the allowed parameter space region due to the relatively large $\msq$ values.) For the $\hh$, the visible $\wt\chi\wt\chi$ modes have a branching ratio of order 40% for much of parameter space, while the visible $\slep\slep$ modes are generally below 5%. For the $\ha$, visible $\slep\slep$ modes are always insignificant since they are determined by mixing proportional to fermion Yukawa couplings, but visible $\wt\chi\wt\chi$ channel branching ratios between 80% and 90% are quite common. In contrast, the branching ratio for $\hpm$ decays to (visible) $\slep^{\pm}\snu$ modes is commonly in the 50% to 90% range, while visible $\wt\chi\wt\chi$ channels seldom exceed 5%. More details will appear in Ref. \[\].
At LEP-II, the most promising sparticle searches will involve $\cpone\cmone$ and $\slep^+\slep^-$ pair production. In Fig. we show the contour for $\mcpone=100\gev$, the absolute upper limit for $\cpmone$ detection for a LEP-II energy of $200\gev$. The fact that $\msnu<\mcpone$ over most of the allowed parameter space (see the contour in Fig. ) implies that the $\epem\rta\cpone\cmone$ production rate will be lowered due to the destructive interference of the $\snu$ exchange diagram. However, the $\cpone$ decays are (as noted above) two-body and yield energetic leptons. This means $\cpone\cmone$ pair production will be relatively easily seen via the di-lepton signal. Nonetheless, sleptons provide more coverage, since the $\cpone$ (and other ino’s) are too heavy to be detectable over about half of the allowed parameter space. The lightest slepton, $\stauone$, is likely to be detected at LEP-200 for $\mstauone\lsim 90-95\gev$. The contour corresponding to 100 $\stauone^+\stauone^-$ events, shown in Fig. , indicates that detection will be possible for all but a small section of parameter space if $\tau^+\tau^-+\ptmiss$ final states can be efficiently employed. $\slep^+\slep^-$ ($\ell=e,\mu$) pairs have only slightly higher thresholds and would be easily detected in the $\ell^+\ell^-+\ptmiss$ final state. Since slepton-pair production can indirectly probe $\mcpone$ values as large as $170\gev$, depending on $\tanb$, it provides a deeper probe of the model than chargino production. This is also discussed in Ref. \[\].
What about the Higgs bosons? First, we note that $\mhl<110 (105)\gev$ for $\mu>0 (\mu<0)$. Fig. shows the contour for 100 $Z\hl$ events (the internal solid line) at LEP-200 (before cuts or branching ratios), as being an upper limit for $\hl$ discovery. Note that $Z\hl$ events will be observable for almost all of parameter space, in particular in that part of parameter space where $\stauone$ detection becomes questionable. Because $\mha\gsim 185\gev$ for the allowed parameter region (for both signs of $\mu$), none of the other Higgs bosons will be detectable at LEP-200. As already noted (see Fig. ) invisible decays of the $\hl$ will be important, and can even be dominant, at smaller $\mha$ values. However, one can still obtain a Higgs signature via mass reconstruction using the recoiling $Z$ in $Z\hl$ events. Fig. shows that the combined $\hl$ and $\slep^\pm$ searches should completely explore the YUMS parameter space for $\mu>0$ except for the small window outlined by the $\hl Z$ and $\stauone^+\stauone^-$ contours in the vicinity of $3.5\lsim \tanb\lsim 4.5$, $450\gev \lsim \mha \lsim 500\gev$, where no new particle could be detected. For the $\mu<0$ case, there is no such window, and complete exploration of the parameter space is possible at LEP-200. Even for $\mu>0$, if $\Delta\rho<0.01$ is imposed there will be no window, $\hl$ discovery being guaranteed.
The strongly interacting $\gl$ and $\sq\,$’s are most appropriately sought at a hadron collider. Since $\msq<\mgl$ throughout all of parameter space, the $\sq\,$’s should provide the most direct signal. The decays $\sq\rta q^\prime \cpmone,q\cnone$ are always allowed, with (as noted above) $\cpone\rta \ell^+\snu+\slep^+\nu$, as opposed to the more standard three-body modes. Generally, $\mcpone-\msnu$ and $\mslep-\mcnone$ (which determine the lepton spectra in these two respective modes ) while not large, are adequate to yield a reasonably hard $\ell$. Thus, generally speaking, isolated energetic leptons can provide an excellent probe for both the squarks and the gluino. (For the latter, the like-sign dilepton signature can have a high rate.) However, the YUMS $\gl$ and $\sq$ mass scales are probably too large for discovery at the Tevatron. In Fig. , the contour for $\mgl=300\gev$ (dotted line) represents our estimate of the absolute maximum value that could possibly be probed by an upgraded Tevatron (with $L\sim 1\fbi$), even given the fact that the leptons for the like-sign dilepton signature will (as noted above) be reasonably energetic. Detection of the $\gl$ in the YUMS model seems unlikely prior to the operation of the LHC. The lightest squark is always the $\stop_1$ (the lightest of the two stops), and we find $m_{\stop_1}\gsim 150\gev$. The $\stop_1$ decays are fairly standard, with $\cpone b$ being always allowed, while $\cnone t$ is allowed at higher $\mha$ values. The other squarks are even heavier ($\msq\gsim
0.9 \mgl>250\gev$). These mass values are sufficiently large that the Tevatron will also have considerable difficulty in discovering the squarks until very high luminosity is available, despite the favorable energetic lepton signature for the $\stop_1$ and the other, heavier, $\sq\,$’s.Of course, the inos and sleptons are much lighter; however, they have lower production rates. Existing studies of ino detection at a high luminosity Tevatron suggest that extracting a tri-lepton signal (from $\cpone\rta \ell^+\snu,
\slep^+\nu$ and $\cntwo\rta \slep\ell$, with $\slep \rta \ell\cnone$) should certainly be possible for $\mcpone\lsim 100\gev$, possibly extending to $\mcpone\lsim 165\gev$ (see the dotdash curve in Fig. ).
At the LHC, there should be no difficulty in detecting the strongly interacting gluino and squarks of the YUMS. For instance, gluinos heavier than $700\gev$ are not allowed, and, as discussed, these will be easily discovered in either the missing energy or like-sign dilepton modes. (See Ref. \[\] and Refs. \[,,\].) Similarly, squarks much heavier than $\sim 540\gev$ are not allowed by the general model constraints, and multi-lepton signals should provide an excellent probe for $\sq\sq$ and $\sq\gl$ production. (See Refs. \[,\].) Direct slepton and ino detection will be more difficult, but may prove possible. Direct slepton production at the LHC has been considered in Refs. \[\] and \[\]. Both conclude that $\mslep\lsim200-250\gev$ should be explored via the di-lepton signal within one LHC year. However, the generic decay chain masses assumed in these studies differ in detail from the predictions of the YUMS, and Ref. \[\] assumes that $\snu$’s are visible through $\snu\rta \ell^\mp\cpmone$ decays, a mode not generally present in the YUMS model. $ep$ collisions at the LHC would provide a cleaner environment for slepton searches. Detection of ino pair production at the LHC was studied in Refs. \[,\], which focused on 3 and 5 lepton final states. The conclusion was that for generically chosen masses and decay chains pair production of inos (which had to include at least one of the heavier inos) should be detectable for an ino spectrum characterized by modest $M$ and $\mu$ values (up to $\sim 200\gev$). A reanalysis of all the above results in the YUMS context would be quite useful, given the fact that the lighter inos can have two-body decays to leptons, thereby typically yielding more energetic leptons, that would more easily pass appropriate cuts.
Turning to Higgs boson detection at the LHC, we display in Fig. the contours for $\hl$ detection in the $\gamma\gamma,\ell\gamma\gamma,t\bar t b\bar b$ and invisible $(\hl\rta\cnone\cnone,\snu\snubar)$ modes; each is viable for some region of the allowed parameter space. We see that detection of the $\hl$ will be possible over all of parameter space in at least one mode (most often several modes). However, at low $\mha$ ($\mha\sim 200-240\gev$) the $\hl$ can only be detected at the LHC via the invisible decay procedures. Since $\tanb\lsim 10$ and $\mha\gsim 185\gev$, detection of the $\hh,\ha,\hpm$ in SM particle final states will be very difficult at the LHC. The proposed modes, such as $\hh\rta 4\ell$, $\ha\rta\gam\gam$, $\ha\rta Z\hl\rta Z\tau\tau$ for small to moderate $\tanb$, and $\hh,\ha\rta\tau^+\tau^-$ at large $\tanb$, all fail due to the fact that the $\hh,\ha\rta \wt\chi\wt\chi,\slep\slep$ decay modes have large branching ratios (see earlier discussion) and therefore deplete significantly the SM final state branching ratios assumed in the detector studies. In our analysis, the possibility of detecting the $\hh$ and $\ha$ in the invisible $\cnone\cnone,\snu\snubar$ channels was examined; $4\sigma$ signals are not achievable in the allowed parameter space. Assessment of the prospects for heavy Higgs detection in visible $\wt\chi\wt\chi$ and $\slep\slep$ modes at a hadron supercollider has only just begun, but some optimism seems warranted. For $\mu<0$, since $\mhl<105\gev$, $\mgl\lsim 500\gev$ (and therefore $m_{\slep_R}\lsim 75\gev$), $\slep^+\slep^-$ and $\hl Z$ will be observed at LEP-200 for the whole of the allowed parameter space. At the LHC, the no-lose theorem for $\hl$ Higgs detection still holds; however due to the positive shift in $\mcnone$ for $\mu<0$, the kinematically allowed region for $\hl\rta\cnone\cnone$ disappears. Instead, the $b\anti b$ mode covers the small-$\tanb$, small-$\mha$ corner of allowed parameter space, while for $\tanb\gsim 6, \mha\lsim 200\gev$ the $\hl\rta \snu\snubar$ decays are dominant and the invisible decay detection procedures would be necessary.
.15in .075in
We have determined the allowed parameter space and explored the prospects for sparticle and Higgs detection (at LEP-200, the Tevatron and the LHC) in the yukawa-unified ($\lambda_b(M_U)=\lambda_{\tau}(M_U)$, $m_0=A=0$) minimal supergravity model. The model is completely specified by the low energy parameters $\tanb,\mha$ (and the sign of $\mu$). The result of requiring yukawa unification in addition to the usual demand that the LSP be neutral is the upper limit $\tanb\lsim 10$. This is [*much*]{} more restrictive than obtained in $m_0=A=0$ supergravity scenarios that do not require yukawa unification. The more restrictive $B_0=0$ model falls within the YUMS scenario, yielding $\tanb\sim 4$, implying $\mt\sim 185\gev$ (for $\mb(\mb)=4.25\gev$). Exploration of SUSY breaking scenarios that are more general than the YUMS model and a thorough discussion of the bottom-up approach will be presented in a future publication. With regard to experimental probes, we find that for $\mu<0$ the combined $\cpmone,\stauone^\pm,\slep^\pm,\hl$ searches at LEP-200 should completely explore the presently allowed parameter space, yielding a definitive test of the model. For $\mu>0$, Fig. shows a small remaining window which should be well-explored by the LHC.
We also emphasize that the ’no-lose’ theorem for $\hl$ Higgs detection at the LHC is maintained, although the techniques developed for detecting the invisible $\hl\rta \cnone\cnone,\snu\snubar$ decays could be crucial if $\mha\lsim 275\gev$. Although the $\hpm,\ha,\hh$ Higgs bosons are produced in significant numbers at the LHC, the experimental and consistency constraints on the parameter space imply that none of their SM decay modes will yield viable signals. However, their detection at the LHC may prove possible in the generally dominant $\wt\chi\wt\chi,\slep\slep$ modes. If not, the $\hl$ would be the [*only*]{} detectable Higgs boson before construction of the NLC. At that, a rather high energy NLC may be required to see the $\hh,\ha,\hpm$. Given that much of the allowed parameter space (at least for $\mu>0$) lies in the $\mha\sim\mhh\sim\mhpm\gsim 250\gev$ domain for which $\hh\ha$, $\hp\hm$ pair production are the only large cross section processes in $\epem$ collisions, it is clear that $\sqrt s=500\gev$ would be far from certain to allow us to see these heavier scalars. If the NLC is run in the $\gam\gam$ collider mode, for which single $\ha,\hh$ production is possible, the reach could be extended somewhat. At Fermilab, the gluino and squarks of the YUMS models are essentially beyond detection for $L\lsim 1\fbi$ (with the possible exception of the $\stop_1$), whereas the inos might be detectable. All should be within easy reach of the LHC given that $\mgl\lsim 700(500)\gev$, $\msq\lsim 630(450)\gev$ for $\mu>0(\mu<0)$ and the inos are relatively light. A linear $\epem$ collider with sufficient energy would also allow discovery of a large selection of the super partners.
**Acknowledgments**
This work has been supported in part by Department of Energy grant \#DE-FG03-91ER40674 and by Texas National Research Laboratory grant \#RGFY93-330.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a parametric formulation for learning generative models for grasp synthesis from a demonstration. We cast new light on this family of approaches, proposing a parametric formulation for grasp synthesis that is computationally faster compared to related work and indicates better grasp success rate performance in simulated experiments, showing a gain of at least 10% success rate $(p < 0.05)$ in all the tested conditions. The proposed implementation is also able to incorporate arbitrary constraints for grasp ranking that may include task-specific constraints. Results are reported followed by a brief discussion on the merits of the proposed methods noted so far.'
author:
- 'Ermano Arruda$^{1}$, Claudio Zito$^{1}$, Mohan Sridharan$^{1}$, Marek Kopicki$^{1}$ and Jeremy L. Wyatt$^{1}$[^1][^2]'
bibliography:
- 'grasp\_synthesis.bib'
- 'general\_ml.bib'
- 'simulation.bib'
title: '**Generative grasp synthesis from demonstration using parametric mixtures** '
---
Introduction
============
We address the problem of modelling generative models for grasp synthesis using parametric representations for probability densities. In recent decades, considerable effort has been put towards tackling the problem of grasping [@shimoga1996robot; @bicchi2000a; @curtis2008a; @ben-amor2012a; @rietzler2013a; @detry2013c; @hjelm2014a; @saxena2008b; @bohg2011b; @kopicki2015; @Pas2015; @Gualtieri2016; @Levine2016]. Generative approaches for grasp synthesis have demonstrated to be more versatile since they can be used to generate complete hand configurations in novel contexts, e.g. [@Montesano2008; @Song2011; @Morrison2018]. Kopicki et al [@kopicki2015] have proposed a non-parametric technique to learn generative models from demonstration for dexterous grasping via Kernel density estimation (KDE). Such approach makes use of KDEs to approximate probability densities over the special Euclidian group - $SE(3)$, as well as hand configuration models encoding the shape of the hand for a given demonstration. The original approach has been shown to be able to generalise demonstrated grasps to novel query object point clouds, without requiring prior knowledge about object pose, shape (such as CAD mesh models) or dynamic properties such as friction coefficients, mass distribution, among others.
Although KDEs are very appealing, one drawback of learning such non-parametric models is the computational time when evaluating the likelihood of samples. If a KDE has $K$ kernels and one wishes to evaluate the likelihood of $N$ data points under the model, the time complexity for computing the data likelihood is $O(K N)$. Noting that for KDEs $K$ is typically large, in the order of hundreds, since commonly a kernel is placed on every training data point. This large number of kernels makes the final time complexity dependent on the size of the training data set used to approximate a desired density. In contrast, parametric mixtures usually need a smaller number for $K$ to effectively approximate a probability density. Once the model parameters are learnt via Expectation Maximisation (EM) [@Bishop2007PatternRA], the runtime for evaluating the likelihood of query data points is not dependent on the size of the data set used for training, but only on the fixed number chosen for $K$. Thus, although the complexity for data likelihood computation is the same in principle, in practice parametric mixtures are orders of magnitude faster than KDEs since $K$ is rarely greater than 10 for most scenarios.
This work re-interprets the approach in [@kopicki2015] for grasp synthesis using a parametric mixture formulation for density modelling. We show the benefits for grasp generation time and performance in simulation under different sensor noise conditions, also showing deployment on a robot platform.
First, we proceed to describe the basic representations utilised throughout this paper for modelling probability densities, followed by the description of our approach. Finally, we present our experimental results in simulation and show preliminary deployment of the approach on a robot platform.
Representations
===============
We are interested in modelling joint probability densities over rigid body transformations belonging to the $SE(3)$ and arbitrary feature vectors in $\mathcal{R}^n$. Thus, realisations of the random variable we intend to model lie in the space $\mathbf{x} \in SE(3) \times \mathcal{R}^n$. Concretely, the special Euclidean group is the cartesian product of members of the special orthogonal group $SO(3)$ (the group of rotations) and members of $\mathcal{R}^3$ representing translations, thus $SE(3) = \mathcal{R}^3 \times SO(3)$. If we choose to represent rotations with unit quaternions, then $\mathbf{x}=(\mathbf{p}, \mathbf{q}, \mathbf{r})$ is a $d=7 + n$ dimensional vector, where $\mathbf{p} \in \mathcal{R}^3$, $\mathbf{q} \in \mathcal{R}^4$ and $\mathbf{r} \in \mathcal{R}^n$. In this work, the same type of feature used by [@kopicki2015] will be utilised, which are the eigen-values $k_1$ and $k_2$ of the principal curvatures for a given point $\mathbf{p}$ from an object point cloud. In Fig. \[fig:point\_cloud\_and\_feat\] this type of feature is illustrated. Thus, concretely, the feature vector is represented as $\mathbf{r}=[k_1,k_2] \in \mathcal{R}^2$.
We will see that we are able to learn probability densities over this manifold using data acquired from demonstration where a data set $\mathcal{D}$ is acquired, such that
$$\label{eq:data}
\mathcal{D} = \{ \mathbf{x_j} | \mathbf{x_j} \in \mathcal{R}^3 \times \mathcal{SO}(3) \times \mathcal{R}^n \}_{j=1}^{J_D}$$
where $\mathbf{x}_j \sim \mathbf{pdf}(\mathbf{p}, \mathbf{q}, \mathbf{r})$.
In particular, we will represent quaternions noting their relationship with angular velocities $\bm{w} \in \mathcal{R}^3$ through the logaritimic and exponential maps defined by Eq. \[eq:log\_map\] and \[eq:exp\_map\], as similarly noted by [@Ude2014]. Given a unit quaternion $\mathbf{q} = [\mathbf{q}_v, q_w]^T$, we define its logarithmic map as:
$$\label{eq:log_map}
\bm{\omega} = \log(\mathbf{q}) = \left\{
\begin{array}{ll}
\arccos(q_w) \frac{\mathbf{q}_v}{ \| \mathbf{q}_v \| } & \quad \mathbf{q}_v \neq \mathbf{0} \\
{[0,0,0]^T} & \quad otherwise
\end{array}
\right.$$
\
And conversely, the exponential map is given by:
$$\label{eq:exp_map}
\mathbf{q} = \exp(\bm{\omega}) = \left\{
\begin{array}{ll}
{[\sin(\| \bm{\omega}\| ) \bm{\omega}, \cos(\| \bm{\omega}\| )]^T} & \quad \bm{\omega} \neq \mathbf{0} \\
{[0,0,0,1]^T} & \quad otherwise
\end{array}
\right.$$
Using these mappings, we will fit parametric mixtures to model joint distributions over $\mathbf{x} = (\mathbf{p}, \bm{\omega}, \mathbf{r})$, where the quaternion representation for $\bm{\omega}$ is readily given by the exponential map $\mathbf{q} = \exp(\bm{\omega})$.
Probability density approximation using parametric mixtures
-----------------------------------------------------------
The density $\mathbf{pdf}(\mathbf{x})$ is approximated using a parametric Gaussian mixture model (GMM) modified so as to internally take care of the appropriate exponential and logarithmic mappings for quaternions. It is then defined as follows:
$$\label{eq:density_mixture}
\mathbf{pdf}(\mathbf{x}) = \sum_{j=1}^{K} w^j \mathcal{N}(\mathbf{x} | \bm{\mu_x}^j, \bm{\Sigma_x}^j)$$
For convenience, let $\mathbf{u}=(\mathbf{p},\mathbf{q})$, i.e. a rigid body transformation. Thus, with ${\bm{\Sigma_x}^j}^{-1} = \bm{\Lambda_x}^j$, we denote:
$$\begin{aligned}
%\mathbf{x} &= (\mathbf{p}, \mathbf{q}, \mathbf{r}) , \label{eq:x_var} \\
\bm{\mu_x}^j &= (\bm{\mu_u}^j, \bm{\mu}_r^j), \label{eq:mu_var} \\
\bm{\Sigma_x} &= \begin{pmatrix}
\begin{matrix}
\bm{\Sigma_{uu}}^j & \bm{\Sigma_{ur}}^j \\
\bm{\Sigma_{ur}}^j & \bm{\Sigma_{rr}}^j
\end{matrix}
\end{pmatrix} , \bm{\Lambda_x}^j = \begin{pmatrix}
\begin{matrix}
\bm{\Lambda_{uu}}^j & \bm{\Lambda_{ur}}^j \\
\bm{\Lambda_{ur}}^j & \bm{\Lambda_{rr}}^j
\end{matrix}
\end{pmatrix} \end{aligned}$$
In this work the density $p( \mathbf{u} | \mathbf{r})$ is a conditional probability density of rigid body transformations given a contact feature $\mathbf{r}$, and is also modelled as a Gaussian mixture model. The parameters of this conditional mixture are obtained in closed form from Eq. \[eq:density\_mixture\] as (see [@Bishop2007PatternRA] for an overview):
$$\begin{aligned}
\bm{\mu_{u|r}}^j &= \bm{\mu_u}^j - \bm{\Sigma_{uu}}^j \bm{\Lambda_{ur}}^j (\mathbf{r} - \bm{\mu_r}^j), \label{eq:mu_var_ur} \\
\quad \nonumber \\
\bm{\Sigma_{u|r}}^j &= \bm{\Sigma_{uu}}^j \label{eq:sigma_var_ur} \\
\quad \nonumber \\
\qquad p^j(\mathbf{u} | \mathbf{r}) &= \mathcal{N}(\mathbf{x} | \bm{\mu_{u|r}}^j , \bm{\Sigma_{u|r}^j}) \\
\quad \nonumber \\
\pi^j &= \frac{w^j \mathcal{N}(\mathbf{r} | \bm{\mu_r}^j, \bm{\Sigma_{rr}}^j)}{\sum_{j=1}^{K} \mathcal{N}(\mathbf{r} | \bm{\mu_r}^j, \bm{\Sigma_{rr}}^j)}\end{aligned}$$
Therefore, $p( \mathbf{u} | \mathbf{r})$ is expressed as
$$\label{eq:density_mixture_rigid_body}
p( \mathbf{u} | \mathbf{r}) = \sum_{j=1}^{K} \pi^j p^j( \mathbf{u} | \mathbf{r})
% p( \mathbf{p}, \mathbf{q}; \bm{\theta}_{pq}) = p( \mathbf{p}; \bm{\theta}_{p}) p( \mathbf{q} ; \bm{\theta}_{q}),$$
The final density over features $\mathbf{r}$, i.e. $p(\mathbf{r})$, is also modelled as a Gaussian mixture and is obtained via marginalisation of Eq. \[eq:density\_mixture\] with respect to $\mathbf{u}$ such that:
$$\label{eq:density_mixture_features_marg}
p^j( \mathbf{r}) = \int \mathcal{N}(\mathbf{x} | \bm{\mu_x}^j, \bm{\Sigma_x}^j) du = \mathcal{N}(\mathbf{r} | \bm{\mu_r}^j, \bm{\Sigma_r}^j)$$
Thus, it follows that
$$\label{eq:density_mixture_features}
p(\mathbf{r}) = \sum_{j=1}^{K} w^j p^j( \mathbf{r}).$$
The mixture in Eq. \[eq:density\_mixture\] is learned using the expectation maximisation algorithm (EM) over contact data acquired from a grasp demonstration $\mathcal{D}= \{ \mathbf{x_j} | \mathbf{x_j} \in \mathcal{R}^3 \times \mathcal{SO}(3) \times \mathcal{R}^n \}$. The mixture formulation is more compact and less memory hungry. Once the density is learnt, we only need to keep its learnt parameters in memory and compute in closed form the additional probability densities in Eq. \[eq:density\_mixture\_features\] and Eq. \[eq:density\_mixture\_rigid\_body\]. In addition, it needs fewer kernels to approximate the densities, therefore $K$ is typically smaller than 10. In contrast to the KDE approach, each data point becomes a the center of a kernel, which is of the order of hundreds per contact model learnt.
![An example point cloud of a mug and feature representation. For a given point $\mathbf{p}$, the axis in blue is the surface normal pointing outwards the object, the direction of the first principal curvature $k_1$ is depicted as the horizontal red axis, and the direction of the second principal curvature $k_2$ is depicted as the green vertical axis. This right-handed frame is constructed by taking the cross-product between the normal and the first principal curvature direction. Together with the point $\mathbf{p}$, these three axis define a rigid body pose, whose position and quaternion representation are given by $\mathbf{v}=(\mathbf{p},\mathbf{q})$. Therefore, each point of the object has a rigid frame attached in similar fashion, although in the picture we chose to highlight only one to avoid clutter.[]{data-label="fig:point_cloud_and_feat"}](assets/pc_features.png){width="0.45\columnwidth"}
Grasp synthesis using parametric contact models
===============================================
Having defined the basic representations for learning distributions using parametric mixtures, we proceed to describe the main components of what we refer to as the parametric grasp synthesis approach.
We will be learning joint probability densities over $\mathbf{x}=(\mathbf{p}, \mathbf{q}, \mathbf{r})$. For compactness we will refer to the position and orientation (pose) of a feature $\mathbf{r}$ as $\mathbf{v}=(\mathbf{p}, \mathbf{q})$, and therefore $\mathbf{x}$ can be written as
$$\label{eq:compact_x}
\mathbf{x} = (\mathbf{v},\mathbf{r}),$$
where $\mathbf{v} \in SE(3)$ and $\mathbf{r} \in \mathcal{R}^n$ is a feature vector.
The object representation
-------------------------
At all times, the object intended to be grasped is represented by $O(\mathbf{v}, \mathbf{r})$. Concretely, it consists of the current object point cloud augmented with features $\mathbf{r} = [k_1, k_2]$ representing the principal curvatures at each point of the object. Finally, for a given point $\mathbf{p}$ from the point cloud, using corresponding eigenvectors of the principal curvatures and estimated normal at this given point, one can construct a frame $\mathbf{v} = (\mathbf{p},\mathbf{q})$. Hence, the object point cloud is then augmented to be a cloud of rigid body transformations associated with a features $\mathbf{r}$ as depicted by Fig. \[fig:point\_cloud\_and\_feat\].
Contact model
-------------
For a given hand link $L_i$, the conditional contact model density $M_i(\mathbf{u}|\mathbf{r})$ is modelled as a conditional probability density over finger link poses $\mathbf{u} \in SE(3)$ given contact surface features $\mathbf{r} \in \mathcal{R}^n$. This density is proportional to the likelihood of finding a finger link located at $\mathbf{u}$ with respect to the location $\mathbf{v}$ of a contact point with feature vector $\mathbf{r}$. Such model is constructed from the dataset $\mathcal{D}_O = \{\mathbf{x}_j\}_{j=1}^{K_O}$ as follows. Let $\mathbf{s}_i = (\mathbf{p}_i, \mathbf{q}_i) \in SE(3)$ be the pose of the link $L_i$ in world frame. The relative pose of link $L_i$ with respect to a feature pose $\mathbf{v}_j$ is given by
$$\label{eq:relative_pose_u_mixture}
\mathbf{u}_{ij} = (\mathbf{p_{ij}},\mathbf{q_{ij}}) = \mathbf{v}_j^{-1} \circ \mathbf{s}_i$$
Thus we are able to construct the contact model representing a probability distribution of relative poses with respect to contact features $\mathbf{r}_j$ as
$$\label{eq:contact_model_mixture}
M_i(\mathbf{u}|\mathbf{r}) \approx p( \mathbf{u} | \mathbf{r})
% M_i(\mathbf{u}) \approx p( \mathbf{u} ; \bm{\theta}_{u})$$
where $p( \mathbf{u} | \mathbf{r})$ is defined in Eq. \[eq:density\_mixture\_rigid\_body\] and is obtained via conditioning Eq. \[eq:density\_mixture\] learnt using the demonstration data set $\mathcal{D}_O$.
Furthermore, it is worth noting that $M_i$ is the contact model for link $L_i$ in proximity during demonstration to a set of features modelled by
$$\label{eq:contact_model_mixture_feature}
% M_i(\mathbf{u}|\mathbf{r}) \approx p( \mathbf{u} | \mathbf{r}; \bm{\theta}_{u})
M_i(\mathbf{r}) \approx p( \mathbf{r})$$
where $M_i(\mathbf{r})$ is obtained via marginalisation of Eq. \[eq:density\_mixture\] as described by Eq. \[eq:density\_mixture\_features\_marg\]. The mixture in Eq. \[eq:density\_mixture\] is learnt with EM using the demonstration data set $\mathcal{D}_O = \{\mathbf{x}_j\}_{j=1}^{K_O}$, a data set consisting of data points $\mathbf{x}_j = (\mathbf{v}_j,\mathbf{r}_j)$ in proximity to finger link $L_i$.
Let $(\mathbf{\tilde{v}}, \mathbf{\tilde{r}}) \sim O(\mathbf{v}, \mathbf{r})$ be a sample from the object intended to be grasped represented by $O(\mathbf{v}, \mathbf{r})$. It follows that if were to compute the data likelihood of features $\mathbf{\tilde{r}}$ over $O(\mathbf{v}, \mathbf{r})$, then features with high data likelihood reflect higher affinity that this finger link $L_i$ could be positioned over $\mathbf{\tilde{r}}$ located at $\mathbf{\tilde{v}}$. The location of this finger link in world frame is given by $\mathbf{\tilde{s}} = \mathbf{\tilde{v}} \circ \mathbf{\tilde{u}}$, where $\mathbf{\tilde{u}} \sim M_i(\mathbf{u}|\mathbf{\tilde{r}})$, together with its weight representing this affinity $\tilde{w} = M_i(\mathbf{\tilde{r}})$, forming the n-uple $(\mathbf{\tilde{s}},\tilde{w})$.
Depending on the probability distribution of features during demonstration, different finger links show different affinities, thus modelling different contact distributions. Figure \[fig:cd\_feat\] depicts this property for two different links of a WSG 50 Schunk gripper. For instance, in Fig. \[fig:cd\_feat\] (left), the link number $L_2$ is closer to flat patches of surface. By evaluating $M_2(\mathbf{\tilde{r}})$ over $(\mathbf{\tilde{v}}, \mathbf{\tilde{r}}) \sim O(\mathbf{v}, \mathbf{r})$, one is able to determine the regions of high and low data likelihood/affinity over the object surface represented by $O(\mathbf{v}, \mathbf{r})$, which in this case is a soup box. In Fig. \[fig:cd\_feat\] (right), the same process is repeated this time for $L_1$ . Note that $L_2$ prefers flat surfaces, whereas $L_1$ prefers corners, due to its closer proximity to features describing the edges of the object during demonstration. The features we chose to utilise are the same features described by $\cite{kopicki2015}$, which are the principal axis of curvature, computed using a PCA method. Thus, each feature $\mathbf{r} \in \mathcal{R}^2$ is a two dimensional vector containing the two largest eigen values approximating the curvatures of a given point of the object point cloud.
Hand configuration model
------------------------
A probability density constructed from a demonstrated grasp trajectory containing starting and final joint configurations of the robot hand. Here, this model is constructed as originally proposed by [@kopicki2015]. Denoting $\mathbf{h}^t_e$ as the joint angles at the start of a given demonstrated grasp (pre-grasp configuration), and $\mathbf{h}^t_g$ the joint angles at contact with the object. Finally, let the set of hand configurations be $\mathcal{H}_c = \{ \mathbf{h}(\gamma) : \gamma \in [-\beta, \beta], \beta \in \mathcal{R}^+\}$, where $\mathbf{h}(\gamma) = (1 - \gamma) \mathbf{h}^t_g + \gamma \mathbf{h}^t_e$, the probability density $C(\mathbf{h}_c)$ is then non-parametrically approximated from this set via a KDE as
$$\label{eq:manipulator_model}
C(\mathbf{h}_c) = \sum_{\gamma \in [-\beta, \beta]} w(\mathbf{h}_c(\gamma)) \mathcal{N}_D(\mathbf{h}_c|\mathbf{h}_c(\gamma),\sigma_c)$$
where $w(\mathbf{h}_c(\gamma)) = \exp(-\alpha \| \mathbf{h}_c(\gamma) - \mathbf{h}^t_g \|)$ and $\alpha \in \mathcal{R}^+$.
Grasp synthesis for novel objects
---------------------------------
### Contact query density
From the learnt densities described above, given a query point cloud for a new object in its augmented representation $O(\mathbf{v}, \mathbf{r})$, a so called contact query density $Q_i(\mathbf{s})$ can be constructed for each link $L_i$. This probability density is constructed by first sampling $(\mathbf{\tilde{v}}, \mathbf{\tilde{r}}) \sim O(\mathbf{v}, \mathbf{r})$ from the object intended to be grasped. Next, $\tilde{\mathbf{u}}_i \sim M_i(\mathbf{u}|\mathbf{r})$ is sampled from our learnt contact model for finger link $L_i$ in Eq. \[eq:contact\_model\_mixture\], such that a set of $N_{Q_i}$ samples $\tilde{\mathbf{s}}_{ij} = \tilde{\mathbf{v}}_j \circ \tilde{\mathbf{u}}_i = (\tilde{\mathbf{p}}_{ij},\tilde{\mathbf{q}}_{ij})$ is constructed $\mathcal{D}_{Q_i} = \{\tilde{\mathbf{s}}_{ij}\}_{j=1}^{N_{Q_i}}$ and used to fit a query density defined as
$$\label{eq:query_density_approx}
Q_i(\mathbf{s}) \approx \sum_{k=1}^{K_{Q_i}} w_{k} \mathcal{N}(\mathbf{s} | \bm{\mu_s}^k,\bm{\Sigma_s}^k)$$
This is density is once again learnt as a GMM using the EM algorithm.
In this fashion, the set of contact query models, one for each finger link and for a demonstrated grasp $g$, is denoted by $\mathcal{Q}_g=\{Q_i(\mathbf{s})\} \forall_i$.
### Grasp sampling
The query density described above allows one to sample finger link poses over the given query point cloud surface. To construct a full hand posture, first a link pose is sampled using Eq. \[eq:query\_density\_approx\], i.e. $\tilde{\mathbf{s}} \sim Q_i(\mathbf{s})$. Finally, the full hand posture of a hand is completed using forward kinematics by sampling the remaining joint angles for the remaining links from the hand configuration model $C(\mathbf{h}_c)$. Many grasp candidates can be generated in this fashion.
### Grasp optimisation
The generated grasp samples can be later optimised via a derivative-free method using the product of experts likelihood criterion as energy function, such as simulated annealing as done in [@kopicki2015].
We denote a grasp solution as the tuple $\mathbf{h} = (\mathbf{h}_w, \mathbf{h}_c)$ containing respectively the hand wrist pose $\mathbf{h}_w \in SE(3)$ and hand joint configuration $\mathbf{h}_c \in \mathcal{R}^D$, such that if $fk(\cdot)$ is the forward kinematic function of the hand, then the hand links in workspace are given by $s_{1:N_L}$, where $N_L$ is the number of links of a robotic hand. Thus, we have
$$\label{eq:fk_function}
s_{1:N_L} = fk(\mathbf{h}), s_l = fk_l(\mathbf{h}).$$
The basic optimisation criterion is given by:
$$\label{eq:opt_criterion_std}
\mathcal{L}(\mathbf{h}) = C(\mathbf{h}) \prod_{Q_i \in \mathcal{Q}} Q_i(fk_i(\mathbf{h}))$$
$$\label{eq:opt_std}
\mathbf{\overset{*}{\mathbf{h}}} = \operatorname*{arg\,max}_{\mathbf{h}=(\mathbf{h}_w,\mathbf{h}_c)} = \mathcal{L}(\mathbf{h}).$$
In [@kopicki2015] an additional expert is defined $W(\mathbf{h})$ so as to penalise collisions in a soft manner, as it was found to prune the solution space to contain better solutions. Collisions are penalised exponentially by the degree of penetration through the object point cloud by any of the hand links. We have also employed a collision expert in the implementation of this work.
The objective function with the additional collision expert is then given by
$$\label{eq:opt_criterion_std_col}
\mathcal{L}(\mathbf{h}) = W(\mathbf{h}) C(\mathbf{h}) \prod_{Q_i \in \mathcal{Q}} Q_i(fk_i(\mathbf{h})).$$
And again, the best grasp is found by optimising
$$\label{eq:opt_std_col}
\mathbf{\overset{*}{\mathbf{h}}} = \operatorname*{arg\,max}_{\mathbf{h}=(\mathbf{h}_w,\mathbf{h}_c)} \mathcal{L}(\mathbf{h})$$
where $W(\cdot)$ is the collision expert, $C(\cdot)$ is the hand configuration expert defined in Eq. \[eq:manipulator\_model\], and $Q_i(\cdot)$ is the contact query expert for link $L_i$ defined by Eq. \[eq:query\_density\_approx\]. For a robot hand and a given grasp demonstration $g$, there is a set of contact query experts $\mathcal{Q}_g$. This optimisation criterion, therefore, tries to maximise the product of experts [@Hinton99], where each expert is a probability density. Each individual expert is responsible for assigning high likelihood to candidate grasps that satisfy just one of the constraints.
In this work we will explore the mixture of experts approach even further and allow our optimisation criterion to incorporate multiple experts to represent arbitrary task constraints. This will give the potential to generate grasps specifically tailored for a given task. For this purpose let $\mathcal{E} = {E_k(\cdot)}_{k=1}^{N_k}$, such that each $E_i(\cdot)$ assigns a probability to a grasp solution $\mathbf{h}$ that is close to one if the grasp satisfies the constraint represented by $E_i(\cdot)$ and close to zero if the constraint is not satisfied. This set of constraints can be a set of constraints defined a priori by a human for a given task or learnt autonomously by the robot manipulator. We augment our optimisation criterion as follows:
$$\label{eq:opt_criterion_extended}
\mathcal{L}(\mathbf{h}) = W(\mathbf{h}) C(\mathbf{h}) \prod_{Q_i \in \mathcal{Q}} Q_i(fk_i(\mathbf{h})) \prod_{E_i \in \mathcal{E}} E_i(\mathbf{h})$$
Furthermore, we define a hard constraint as an expert that returns binary values of one or zero, accepting or rejecting instantly a grasp, thus implementing rejection sampling. Continuous experts whose image are on the continuous interval $[0,1]$ are referred to as soft constraints in this paper.
![\[fig:objects\] ](assets/object_dataset60.png){width="0.75\columnwidth"}
With this extension, we have then created a framework that allows a great variety of constraints to be incorporated into the grasp generation pipeline. For instance, one can think of experts that imbue dynamic constraints to grasps or specific kinematic constraints such as manipulability, thus paving the way towards task oriented grasp synthesis.
{width=".5\linewidth"} \[fig:conditionANC\]
{width=".5\linewidth"} \[fig:conditionBNC\]
{width=".5\linewidth"} \[fig:conditionCNC\]
{width=".5\linewidth"} \[fig:conditionDNC\]
Thus far in this work, the following constraint experts have been defined: i) hard kinematic constraint and ii) soft principal axis alignment constraint. The former expert returns zero if a given grasp solution $\mathbf{h}$ is not kinematically feasible by the robot platform, and one otherwise. It implements a hard constraint over the grasp generation process, rejecting grasp samples that are not kinematically feasible. The latter expert implements a soft constraint that returns a probability close to one if the wrist frame of the generated grasp $\mathbf{h}$ is aligned with a given axis provided by the user. This is useful to encourage grasps that always pick an object from the top, for instance.
Experiments
===========
We perform two separate experiments. The first experiment compares the success rate between the KDE and GMM-based success rate in simulation (we used Bullet physics [@Coumans2019]) in four conditions: A) without depth sensing noise and with optimisation, B) with simulated depth sensing noise and with optimisation, C) without noise and without optimisation and D) with noise and without optimisation. The second experiment shows the runtime performance for each respective method when generating 200 hundred grasps for an object, results are shown in Fig. \[fig:runtime\].
For all experiments and grasp synthesis methods, the optimisation utilised was the derivative-free simulated annealing method with 100 iterations and maximum initial temperature of 1.0 which decays linearly over the number of iterations.
We adopted the same depth sensor noise model as in [@Johns2016]. The noise model combines additive and lateral shift Gaussian noise applied to every pixel in a simulated depth image:
$$\label{eq:noise_model}
\tilde{z}(u,v) = z(u + \mathcal{N}(0, \sigma_p^2), v + \mathcal{N}(0, \sigma_p^2)) + \mathcal{N}(0, \sigma_d^2),$$
where $\sigma_p$ is the standard deviation for the noisy lateral shift of pixels, which has been set to 1 for these experiments. And $\sigma_d$ is the standard deviation for depth measurements, which was set to 1mm.
Performance comparison
----------------------
For the comparison, we use a parallel gripper (WSG 50 Shunk Gripper) and a single demonstrated grasp on a box, from which object the contact models are learnt for both KDE and GMM-based methods. The KDE-based method has utilised one kernel per data point to approximate the densities. The GMM-based contact model has approximated Gaussian mixture densities with $K=2$, given the bi-modality of the demonstration as depicted in Fig. \[fig:cd\_feat\] and with the number of kernels $K_{Q}=5$ for GMM-based query density approximation from Eq. \[eq:query\_density\_approx\].
A total of 60 objects were used for testing the performance of each respective method. Objects are varied, some objects have box-like features, others are curved, cylindrical, convex, non-convex, and of combinations of sizes and shapes as shown in Fig. \[fig:objects\]. Objects are always placed in the same pose for both approaches, and simulation parameters are kept fixed. For grasp generation, simulated point clouds are acquired from 4 fixed camera poses around the object, thus obtaining a full 3D point cloud of the object. The acquired point clouds are stitched together and post-processed identically for both approaches, including principal curvature feature computation with search radius set to $1cm$.
For each respective approach (KDE and GMM-based), a total number of 200 grasps are generated per object. Sampled grasp solutions are arranged in decreasing order of value as given by Eq. \[eq:opt\_criterion\_std\], in which the highest value represents the best grasp in the sequence. For each approach, the best grasp is executed on each object. Before every execution attempt the object is restarted to the same initial pose. This process is repeated 10 times over the set of 60 object and results are averaged over these 10 trials. We measure success rate as the proportion of successful grasps over the total number of objects.
Finally, a two-tailed paired t-test is performed to assert the significance of the results between KDE and GMM-based methods over the 10 repeated trials.
![[]{data-label="fig:runtime"}](assets/01_opt_nonoise/runtime_stacked.png){width="0.90\columnwidth"}
Preliminary tests on robot platform
-----------------------------------
We have preliminarily deployed the proposed approach on a real robot platform as shown by Fig. \[fig:cylinder\_grasp\]. For the robot platform, we made use task-specific kinematic constraints, such as kinematic feasibility and axis alignment that can be straightforwardly added in our implementation using Eq. \[eq:opt\_criterion\_extended\] to prune the grasp synthesis process. A summary video on the Boris platform is available on <https://youtu.be/9RqZaTAH4Fs>.
Results
-------
The results in all four different conditions indicate that the GMM-based method shows superior performance over the KDE-based method ($p < 0.05$) in this setup. We see that in the absence of noise, the final optimisation step has little effect in grasp performance as shown by Fig. \[fig:conditionANC\] and \[fig:conditionCNC\]. In contrast, optimisation seems to become crucial under noisy conditions as depicted by Fig. \[fig:conditionBNC\] and \[fig:conditionDNC\], which seems natural. Under the influence of noise, local optimisation with simulated annealing is expected to help to compensate from imperfect sensing, hence still converging to a local or global minimum under the product of experts objective function in Eq. \[eq:opt\_criterion\_std\].
Perhaps the biggest impact is shown in terms of the total time required for grasp synthesis. The stacked bar chart in Fig. \[fig:runtime\] shows the total runtime for generating 200 grasps for a given object, highlighting the individual contributions from different steps involved in the grasp synthesis process: i) query density estimation, ii) sampling and ranking and finally iii) optimising top 10 grasps (out of the 200 samples). Due to the fact that the GMM-based method requires fewer kernels to approximate the query density in Eq. \[eq:query\_density\_approx\], it directly affects the total runtime which is largely dominated by the number of query likelihood evaluations (Eq. \[eq:opt\_criterion\_std\]).
Conclusion & Future Work
------------------------
We proposed an alternative formulation for learning generative models for grasp synthesis using parametric mixtures, such as GMMs. In our experimental setup, the GMM-based method has indicated better performance than related non-parametric approach using KDEs [@kopicki2015]. In addition to that, we extended the product of experts criterion, allowing us to accommodate task-specific constraints as needed for different deployment contexts. We gave a brief instance of that by including kinematic constraints directly in the evaluation step of the grasp synthesis pipeline on the Boris platform. Finally, we showed the GMM-based approach also offers faster execution times. This is a promising prospect that may allow us to employ the proposed method to real-time grasp synthesis as future work. Subsequent work in this direction shall be conducted on a real robot platform.
[^1]: $^{1}$School of Computer Science, University of Birmingham, B15 2TT, UK (exa371,c.zito,M.Sridharan,msk,jlw)@cs.bham.ac.uk
[^2]: Accepted for publication in the RSS workshop on Task-Informed Grasping (TIG-II), 2019.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In fruit production, critical crop management decisions are guided by bloom intensity, i.e., the number of flowers present in an orchard. Despite its importance, bloom intensity is still typically estimated by means of human visual inspection. Existing automated computer vision systems for flower identification are based on hand-engineered techniques that work only under specific conditions and with limited performance. This work proposes an automated technique for flower identification that is robust to uncontrolled environments and applicable to different flower species. Our method relies on an end-to-end residual convolutional neural network (CNN) that represents the state-of-the-art in semantic segmentation. To enhance its sensitivity to flowers, we fine-tune this network using a single dataset of apple flower images. Since CNNs tend to produce coarse segmentations, we employ a refinement method to better distinguish between individual flower instances. Without any pre-processing or dataset-specific training, experimental results on images of apple, peach and pear flowers, acquired under different conditions demonstrate the robustness and broad applicability of our method.'
author:
- 'Philipe A. Dias$^{1}$, Amy Tabb$^{2}$, and Henry Medeiros$^{1}$[^1][^2] [^3] [^4][^5][^6][^7] [^8]'
bibliography:
- 'refs.bib'
title: Multispecies fruit flower detection using a refined semantic segmentation network
---
Bloom intensity estimation, flower detection, semantic segmentation networks, precision agriculture
INTRODUCTION
============
loom intensity corresponds to the number of flowers present in orchards during the early growing season. Climate and bloom intensity information are crucial to guide the processes of pruning and thinning, which directly impact fruit load, size, coloration, and taste [@Forshey1986; @Link2000]. Accurate estimates of bloom intensity can also benefit packing houses, since early crop-load estimation greatly contributes to optimizing postharvest handling and storage processes.
Visual inspection is still the dominant approach for bloom intensity estimation in orchards, a technique which is time-consuming, labor-intensive and prone to errors [@Gongal2016]. Since only a limited sample of trees is inspected, the extrapolation to the entire orchard relies heavily on the grower’s experience. Moreover, it does not provide information about the spatial variability in the orchard, although the benefits of precision agriculture practices are well known [@Zhang2002precision].
These limitations added to the short-term nature of flower appearance until petal fall make an automated method highly desirable. Multiple automated computer vision systems have been proposed to solve this problem, but most of these methods rely on hand-engineered features [@Kapach2012], making their overall performance acceptable only under relatively controlled environments (e.g. at night with artificial illumination). Their applicability is in most cases species-specific and highly vulnerable to variations in lightning conditions, occlusions by leaves, stems or other flowers [@Gongal2015].
In the last decade, deep learning approaches based on convolutional neural networks (CNNs) led to substantial improvements in the state-of-the-art of many computer vision tasks [@Guo2016cnn]. Recent works have adapted CNN architectures to agricultural applications such as fruit quantification [@Chen2017], classification of crops [@Dyrmann2016crops], and plant identification from leaf vein patterns [@Grinblat2016vein]. To the best of our knowledge, our work in [@Dias2018] was the first to employ CNNs for flower detection. In that work, we combined superpixel-based region proposals with a classification network to detect apple flowers. Limitations of that approach are intrinsic to the inaccuracies of superpixel segmentation and the network architecture.
In the present work, we provide the following contributions for automated flower segmentation:
- A novel technique for flower identification that is i) automated, ii) robust to clutter and changes in illumination; and, iii) generalizable to multiple species. Using as starting point a fully convolutional network (FCN) [@Chen2016] pre-trained on a large multi-class dataset, we describe an effective fine-tuning procedure that adapts this model for fine pixel-wise flower segmentation. Our final method evaluates in less than $50$ seconds high-resolution images covering each a full tree. Although the task comparison is not one-to-one, human workers may need on average up to $50$ minutes to count the number of flowers per tree.
- A feasible procedure for evaluating high-resolution images with deep FCNs on commercial GPUs. Fully convolutional computations require GPU memory space that exponentially increases according to image resolution. We employ an image partitioning mechanism with partially overlapping windows, which reduces artifacts introduced by artificial boundaries when evaluating disjoint image regions.
- Release of an annotated dataset with pixel-accurate labels for flower segmentation on high resolution images [@Tabb2018Datasets]. We believe this can greatly benefit the community, since this is a very time consuming yet critical task for both training and evaluation of segmentation models.
Previous attempts at automating bloom intensity estimation were mostly based on color thresholding, such as the works described in [@Aggelopoulou2011color; @Hocevar2014color] and [@Thorp2011]. Despite differences in terms of color-space used for analysis (e.g. HSL and RGB), all these methods fail when applied in uncontrolled environments. Apart from size filtering, no morphological feature is taken into account, such that thresholding parameters have to be adjusted in case of changes in illumination, camera position or flowering density. Even strategies using aerial multispectral images such as [@Horton2017PeachAir] also rely solely on color information for image processing.
Our previous work in [@Dias2018] introduced a novel approach for apple flower detection that relies on a fine-tuned Clarifai CNN [@Zeiler2014clarifai] to classify individual superpixels composing an image. That method highly outperformed color-based approaches, especially in terms of generalization to datasets composed of different flower species and acquired in uncontrolled environments. However, existing superpixel algorithms rely solely on local context information, representing the main source of imprecisions in scenarios where flowers and the surrounding background present similar colors.
While early attempts for autonomous fruit detection also relied on hand-engineered features (e.g. color, texture, shape) [@Gongal2015], recent works have been exploring more advanced computer vision techniques. One example is the work of Hung et al. [@Hung2013Fruit], which combines sparse autoencoders [@Guo2016cnn] and support vector machines (SVM) for segmenting leaves, almonds, trunks, ground and sky. The approaches described by Bargoti and Underwood in [@Bargoti2016Fruit] and Chen et al. in [@Chen2017] for fruit detection share some similarities with our method for flower segmentation. In [@Bargoti2016Fruit], the authors introduce a Faster R-CNN trained for the detection of mangoes, almonds and apple fruits on trees. The method introduced in [@Chen2017] for counting apples and oranges employs a fully convolutional network (FCN) to perform fruit segmentation and a convolutional network to estimate fruit count.
End-to-end fully convolutional networks [@Long2015fcn] have been replacing traditional fully connected architectures for image segmentation tasks [@Garcia2017rev]. Conventional architectures such as the *Alexnet* [@Krizhevsky2012alexnet] and VGG [@Simonyan2014vgg] networks are very effective for image classification but provide coarse outputs for image segmentation tasks. This is a consequence of the image downsampling introduced by the *max-pooling* and *striding* operations performed by these networks, which allow the extraction of learned hierarchical features at the cost of pixel-level precision [@Chen2016].
Different strategies have been proposed to alleviate the effects of downsampling [@Garcia2017rev], including the use of deconvolution layers [@Long2015fcn; @Noh2015deconv], and encoder-decoder architectures with skip layer connections [@Kendall2015segnet; @Hariharan2015hyper]. The DeepLab model introduced in [@Chen2016] is one of the most successful approaches for semantic image segmentation using deep learning. By combining the ResNet-101 [@He2016resnet] model with *atrous* convolutions and spatial pyramid pooling, it significantly reduces the downsampling rate and achieves state-of-the-art performance in challenging semantic segmentation datasets such as the PASCAL VOC [@Everingham2015pascal] and COCO [@Lin2014coco].
In addition to the changes in CNN architecture, the authors of DeepLab also employ the dense CRF model described in [@Krahenbuhl2012] to produce fine-grained segmentations. Although providing visually appealing segmentations, this refinement model relies on parameters that have to be optimized by means of supervised grid-search. In [@Dias2018rgr], we introduced a generic post-processing module that can be coupled to the output of any CNN to refine segmentations without the need for dataset-specific tuning. Called region growing refinement (RGR), this algorithm uses the score maps available from the CNN to divide the image into regions of high confidence background, high confidence object and uncertainty region. By means of appearance-based region growing, pixels within the uncertainty region are classified based on initial seeds randomly sampled from the high confidence regions.
In this section, we first describe the pre-training and fine-tuning procedures carried out to obtain a CNN highly sensitive to flowers. Subsequently, we describe the sequence of operations that our pipeline performs to segment flowers in an image.
Network training
----------------
One of the largest datasets available for semantic segmentation, the COCO dataset [@Lin2014coco] was recently augmented by Caesar et al. [@Caesar2017stuff] into the COCO-Stuff dataset. This dataset includes pixel-level annotations of classes such as *grass, leaves, tree* and *flowers*, which are relevant for our application. In the same work, the authors also discuss the performance of modern semantic segmentation methods on COCO-Stuff, with a DeepLab-based model outperforming the standard FCN. Thus, we opted for the publicly available DeepLab-ResNet model pre-trained on the COCO-Stuff dataset as the starting point for our pipeline. Rather than fine-tuning the dense CRF model used in the original DeepLab work, we opt for the generic RGR algorithm as a post-processing module to obtain fine-grained segmentations.
The base model was originally designed for segmentation within the $172$ COCO-Stuff classes. To adapt its architecture for our binary flower segmentation task, we perform procedures known as *network surgery* and *fine-tuning* [@Girshick2014rcnn]. The surgery procedure is analogous to the pruning of undesired branches in trees: out of the original $172$ classification branches, we preserve only the weights and connections responsible for the segmentation of classes of interest.
We considered first an architecture preserving only the *flower* classification branch, followed by a sigmoid classification unit. However, without the normalization induced by the model’s original softmax layer, the scores generated by the transferred *flower* branch are unbounded and the final sigmoid easily saturates. To alleviate the learning difficulties caused by such a poor initialization, we opted for tuning a model with two-branches, under the hypothesis that a second branch would allow the network to learn a background representation that properly normalizes the predictions generated by the foreground (*flower*) branch.
We have observed experimentally that nearby leaves represent one of the main sources of misclassification for flower segmentation. Moreover, predictions for the class *leaf* presented the highest activations when applying the pre-trained model to our training dataset. For these reasons, we opt for this branch together with the one associated with *flowers* to initialize our two-branch flower segmentation network.
The adapted architecture was then fine-tuned using the training set described in Section \[sec:datasets\], which contains $100$ images of apple trees. For our experiments, the procedure was carried out for $10,000$ iterations using the Caffe framework [@Jia2014a], with an initial learning rate of $10^{-4}$ that polynomially decays according to $10^{-4}\times(1 - i/10000)^{0.9}$, where $i$ is the iteration number. Aiming at scale robustness, our fine-tuning procedure employs the same strategy used for model pre-training, where each training portrait is evaluated at ($0.5$, $0.75$, $1.0$, $1.25$, $1.5$) times its original resolution.
While the validation set has pixel-accurate annotations obtained using the procedure described in Section \[sec:datasets\], the training set was annotated using the less precise but quicker superpixel-based procedure described in our previous work [@Dias2018]. Less than $5\%$ of the total image areas in this dataset contain flowers. To compensate for this imbalance, we augmented portraits containing flowers by mirroring them with respect to vertical and horizontal axes. Following the original network parameterization, we split the $100$ training images into portraits of $321\times321$ pixels, corresponding to a total of $52,644$ training portraits after augmentation.
Sliding window: divide $I$ into a set of $n$ portraits $P$. \[ln:sw\] Obtain $M_B$ and $M_F$ by fusing $m_B^{(i)}$ and $m_F^{(i)}$ ($i=1,\ldots,n$), respectively according to Eq. \[eq:fusion\]. \[ln:fuse\] Normalize $M_B$ and $M_F$ into $\Tilde{M}_B$ and $\Tilde{M}_F$, respectively according to Eq. \[eq:softmax\]. \[ln:soft\] Generate $\hat{Y}$ by applying RGR to $\Tilde{M}_B$ and $\Tilde{M}_F$.
Segmentation pipeline
---------------------
The method we propose for fruit flower segmentation consists of three main operations: 1) divide a high resolution image into smaller patches, in a sliding window manner; 2) evaluate each patch using our fine-tuned CNN; 3) apply the refinement algorithm on the obtained scoremaps to compute the final segmentation mask. These steps are described in detail below. In our description, we make reference to Algorithm \[alg:pseudocode\] and Figure \[fig:diagram\].
*1) Step 1 - Sliding window:* As mentioned above, the adopted CNN architecture either crops or resizes input images to $321\times321$ portraits. Since our datasets are composed of images with resolution ranging from $2704\times1520$ to $5184\times3456$ pixels (see Section \[sec:datasets\]), we emulate a sliding window approach to avoid resampling artifacts. More specifically, we split each input image $I$ into a set $P$ of $n$ portraits $p^{(i)} \in P$. Each portrait is $321\times321$ pixels large, i.e. $p^{(i)} \in \mathbb{R}^{r\times r}$ with $r = 321$. Cropping non-overlapping portraits from the original image introduces artificial boundaries that compromise the detection quality. For this reason, in our approach each portrait overlaps a percentage $s$ of the area of each immediate neighbor. For our experiments, we adopted $s = 10\%$. When the scoremaps are fused, the results corresponding to the overlapping pixels are discarded. Figure \[fig:pad\] illustrates this process for a pair of subsequent portraits. The scores obtained for each portrait are depicted as a heatmap, where blue is associated with lower scores and higher scores are illustrated with red.
[lcccccc]{} & & **Weather** & & & **Resolution** &\
*AppleA* &
----------------------------
$100$ (train) + $30$ (val)
----------------------------
& Sunny & No & Canon EOS 60D & $5184\times3456$ & Hand-held\
*AppleB* & $18$ & Sunny & Yes & GoPro HERO5 & $2704\times1520$ & Utility vehicle\
*Peach* & $24$ & Overcast & No & GoPro HERO5 & $2704\times1520$ & Hand-held\
*Pear* & $18$ & Overcast & No & GoPro HERO5 & $2704\times1520$ & Hand-held\
---------- ------------- ------------- ------------- ------------- ------------- ------------- --
Dataset *$\mu_{H}$* *$IQR_{H}$* *$\mu_{S}$* *$IQR_{S}$* *$\mu_{V}$* *$IQR_{V}$*
*AppleA* $74.6$ $49.3$ $32.9$ $24.3$ $53.7$ $30.2$
*AppleB* $219.6$ $21.1$ $88.6$ $44.3$ $47.1$ $16.9$
*Peach* $223.8$ $199.9$ $11.8$ $20.7$ $42.3$ $46.6$
*Pear* $85.9$ $178.8$ $16.4$ $23.4$ $42.4$ $20.8$
---------- ------------- ------------- ------------- ------------- ------------- ------------- --
: HSV statistics of images composing each dataset.[]{data-label="tab:imghsv"}
---------- ------------- ------------- ------------- ------------- ------------- -------------
Dataset *$\mu_{H}$* *$IQR_{H}$* *$\mu_{S}$* *$IQR_{S}$* *$\mu_{V}$* *$IQR_{V}$*
*AppleA* $136.6$ $205.5$ $6.3$ $9.8$ $77.3$ $24.3$
*AppleB* $56.3$ $80.2$ $7.5$ $9.8$ $86.7$ $23.1$
*Peach* $325.2$ $26.7$ $21.2$ $13.3$ $50.2$ $13.7$
*Pear* $215.4$ $173.2$ $5.9$ $5.9$ $84.7$ $22.4$
---------- ------------- ------------- ------------- ------------- ------------- -------------
: HSV statistics of flowers composing each dataset.[]{data-label="tab:flhsv"}
![**Best viewed in color.** Illustration of the sliding window and subsequent fusion process that comprise our segmentation pipeline. Each portrait overlaps a certain area of its neighbors, which is discarded during fusion to avoid artifacts caused by artificial boundaries.[]{data-label="fig:pad"}](./Figures/Fig2-eps-converted-to.pdf){width="48.00000%"}
*2) Step 2 - CNN prediction:* We evaluate in parallel each portrait $p^{(i)}$ with our fine-tuned network for flower identification. The CNN is equivalent to a function $f$ $$f: p^{(i)} \rightarrow \{m_F^{(i)},m_B^{(i)}\},$$ which maps each input $p^{(i)}$ into two pixel-dense scoremaps: $m_F^{(i)} \in \mathbb{R}^{r\times r}$ represents the pixel-wise likelihood that pixels in $p^{(i)}$ belong to the foreground (i.e., flower), while $m_B^{(i)} \in \mathbb{R}^{r\times r}$ corresponds to the pixel-wise background likelihood. The heatmaps in Figures \[fig:refine\](a) and (b) are examples of scoremaps computed for a given portrait.
*3) Step 3 - Fusion and refinement:* After evaluating each portrait, we generate two global scoremaps $M_B$ and $M_F$ by combining the predictions obtained for all $p^{(i)} \in P$. Let $c^{(i)}$ represent the pixel-coordinates of $p^{(i)}$ in $I$ after discarding the padding pixels. The fusion procedure is defined as $$\label{eq:fusion}
\forall p^{(i)} \in P, \quad M_{F,B}(c^{(i)})=m_{F,B}^{(i)},$$ such that both scoremaps $M_B$ and $M_F$ have the same resolution as $I$. As illustrated in Figure \[fig:pad\], the padded areas of $m_{F,B}^{(i)}$ (outside the red box) are discarded during fusion. For every pixel in the image, a single prediction score is obtained from exactly one portrait, such that artifacts introduced by artificial boundaries are avoided.
After fusion, the scoremaps $M_B$ and $M_F$ are normalized into scoremaps $\Tilde{M}_B$ and $\Tilde{M}_F$ using a softmax function $$\label{eq:softmax}
\Tilde{M}_{F,B}(q_j)=\frac{\exp(M_{F,B}(q_j))}{\exp(M_{B}(q_j))+\exp(M_{F}(q_j))},$$ where $q_j$ is the $j$-th pixel in the input image $I$. With this formulation, for each pixel $q_j$ the scores $\Tilde{M}_B(q_j)$ and $\Tilde{M}_F(q_j)$ add to one, i.e. they correspond to the probability that $q_j$ belongs to the corresponding class.
As Figure \[fig:refine\](c) shows, the predictions obtained directly from the CNN are coarse in terms of adherence to actual flower boundaries. Therefore, rather than directly thresholding $\Tilde{M}_F$, this scoremap and the image $I$ are fed to the RGR refinement module described in [@Dias2018rgr]. For our application, the refinement algorithm relies on two high-confidence classification regions $R_F$ and $R_B$ defined according to $$\begin{aligned}
{3}
R_{F,B}&=\left\{ q_{j}|\Tilde{M}_{F,B}(q_{j})>\tau_{F,B}\right\}, \label{eq:ru}\end{aligned}$$ where $\tau_B$ and $\tau_F$ are the high-confidence background and foreground thresholds. Using the high-confidence regions as starting points, the RGR algorithm performs multiple Monte Carlo region growing steps that groups similar pixels into clusters. Afterwards, it performs majority voting to classify each cluster according to the presence of flowers. Each pixel $q_j$ within a cluster contributes with a positive vote if its score $\Tilde{M}_F(q_j)$ is larger than a threshold $\tau_0$. As detailed in Section \[sec:results\], this parameter can be empirically tuned according to the dataset under consideration. Based on a grid-search optimization on our training dataset, we selected $\tau_0=0.3$ for all our experiments and fixed $\tau_B=0.1$ and $\tau_F=1.25\times\tau_0$.
{#sec:datasets}
We evaluate our method on four datasets that we created and made publicly available: *AppleA*, *AppleB*, *Peach*, *Pear* [@Tabb2018Datasets]. As summarized in Table \[tab:datasets\], images from different fruit flower species were collected in diverse uncontrolled environments and under different angles of capture.
Both datasets *AppleA* and *AppleB* are composed of images of apple trees, which were collected in a USDA orchard on a sunny day. In both datasets, the trees are supported with trellises and planted in rows. *AppleA* is a collection of $147$ images acquired using a hand-held camera. From this total, we randomly selected $100$ images to build the training set used to train the CNN. Out of the remaining $47$ images, $30$ were randomly selected to compose the testing set for which we report results in Section \[sec:results\].
This dataset contains flowers that greatly vary in terms of size, cluttering, occlusion by leaves and branches. Flowers composing its images have an average area of $10,730$ pixels, but with a standard deviation of $17,150$ pixels. On average, flowers compose only $2.5\%$ of the total image area within this dataset, which is otherwise vastly occupied by leaves.
Differently from *AppleA*, for the *AppleB* dataset, a utility vehicle equipped with a background unit was used for imaging, such that trees in other rows are not visible in the images. Figure \[fig:gator\] illustrates the utility vehicle used for image acquisition, and Figures \[fig:appleA\] and \[fig:appleB\] illustrate the differences between datasets *AppleA* and *AppleB*.
The *Peach* and *Pear* datasets differ both in terms of species and acquisition conditions, therefore representing adequate scenarios for evaluating the generalization capabilities of the proposed method. Both datasets contain images acquired on an overcast day and without a background unit. Compared to the *AppleA* dataset, images composing these datasets present significantly lower saturation and value means. Tables \[tab:imghsv\] and \[tab:flhsv\] summarize the differences among datasets in terms of the statistics of the HSV color components, where $\mu$ stands for mean values and $IQR$ for interquartile ranges.
Regarding the flower characteristics, apple blossoms are typically white, with hue components spread in the whole spectrum (high $IQR_H$) and low saturation mean. Flowers composing the *AppleB* dataset present higher brightness ($\mu_V$), while peach flowers show a pink hue centered on $\mu_H=325^\circ$, with higher saturation and lower value means. Moreover, pear flowers are slightly different in terms of color (greener) and morphology, as illustrated in Figure \[fig:pear\].
Labeling
--------
Image annotation for segmentation tasks is a laborious and time-consuming activity. Labels must be accurate at pixel-level, otherwise both supervised training and the evaluation of segmentation techniques are compromised. Most existing annotation tools rely on approximating segmentations as polygons, which provide ground truth images that frequently lack accurate adherence to real object boundaries [@Dias2018rgr].
We opted for a labeling procedure that combines freehand annotations and RGR refinement [@Dias2018rgr]. Using a tablet, the user draws traces on regions of the image that contain flowers, indicating as well hard negative examples when necessary. These traces indicate high-confidence segmentation points, which are used as reference by RGR to segment the remaining parts of the image. Figure \[fig:label\] shows an example of a ground truth segmentation obtained using this procedure[^9].
{#sec:results}
We aim at a method capable of accurate multi-species flower detection, regardless of image acquisition conditions and without the need for dataset-specific training or pre-processing. To verify that our method satisfies all these requirements, we performed experiments on the four different datasets described in Section \[sec:datasets\] while only using the *AppleA* dataset for training.
We adopt as the main baseline our previous model described in [@Dias2018], which highly outperformed existing methods by employing the Clarifai CNN architecture to classify individual superpixels. We therefore refer to that model as <span style="font-variant:small-caps;">Sppx+Clarifai</span> and to our new method as <span style="font-variant:small-caps;">DeepLab+RGR</span>. We also compare our results against a <span style="font-variant:small-caps;">HSV-based</span> method [@Hocevar2014color] that segments images based only on HSV color information and size filtering according to threshold values optimized using grid-search.
(0,0) – (.5,0) node \[right,color=black\] (text1) ;; (text1.east) – (\[xshift=5mm\]text1.east) node \[right,color=black\] (text2) ;; (text2.east) – (\[xshift=5mm\]text2.east) node \[right,color=black\] (text3) ;;
![**Best viewed in color.** Examples of flower detection in one image composing the *AppleB* dataset.[]{data-label="fig:appleB"}](./Figures/Fig7.png){width="48.00000%"}
(0,0) – (.5,0) node \[right,color=black\] (text1) ;; (text1.east) – (\[xshift=5mm\]text1.east) node \[right,color=black\] (text2) ;; (text2.east) – (\[xshift=5mm\]text2.east) node \[right,color=black\] (text3) ;;
All three methods were tuned using the *AppleA* training dataset, with differences in the pipeline for transfer learning. For the three unseen datasets, the <span style="font-variant:small-caps;">Sppx+Clarifai</span> relies on a pre-processing step that enhances contrast and removes the different backgrounds present in the images. Our new method <span style="font-variant:small-caps;">DeepLab+RGR</span> does not require any pre-processing. Instead, it employs the same pipeline regardless of the dataset, requiring only adjustments in portrait size. As summarized in Table \[tab:datasets\], images composing the *AppleA* dataset have resolution $4.3\times$ larger than images in the other three datasets. Thus, we split images in these datasets into portraits of $155\times155$ pixels, rather than the $321\times321$ pixels portraits used for *AppleA*.
{width="49.00000%"} {width="49.00000%"}
(0,0) – (.5,0) node \[right,color=black\] (text1) ;; (text1.east) – (\[xshift=5mm\]text1.east) node \[right,color=black\] (text2) ;; (text2.east) – (\[xshift=5mm\]text2.east) node \[right,color=black\] (text3) ;;
The quantitative analysis of segmentation accuracy relies on precision, recall, $F_1$ and intersection-over-union (IoU) metrics [@Everingham2015pascal] computed at pixel-level, instead of the superpixel-wise metrics used in our previous work. Table \[tab:results\] summarizes the results obtained by each method on the different datasets.
** **IoU** [$\mathbf{F_1}$]{} **Recall** **Precision**
---- ------------------------------------------------------------- ------------------- -------------------- ------------------- -------------------
<span style="font-variant:small-caps;">HSV-based</span> $28.0\%$ $43.7\%$ $56.5\%$ $35.7\%$
<span style="font-variant:small-caps;">Sppx+Clarifai</span> $51.3\%$ $67.8\%$ $73.2\%$ $63.1\%$
<span style="font-variant:small-caps;">DeepLab+RGR</span> $\mathbf{71.4\%}$ $\mathbf{83.3\%}$ $\mathbf{87.7\%}$ $\mathbf{79.4\%}$
<span style="font-variant:small-caps;">HSV-based</span> $49.3\%$ $66.0\%$ $58.9\%$ $75.1\%$
<span style="font-variant:small-caps;">Sppx+Clarifai</span> $50.6\%$ $67.2\%$ $68.4\%$ $66.1\%$
<span style="font-variant:small-caps;">DeepLab+RGR</span> $\mathbf{63.0\%}$ $\mathbf{77.3\%}$ $\mathbf{91.2\%}$ $\mathbf{67.1\%}$
<span style="font-variant:small-caps;">HSV-based</span> $0.1\%$ $1.4\%$ $1.4\%$ $1.6\%$
<span style="font-variant:small-caps;">Sppx+Clarifai</span> $49.1\%$ $67.2\%$ $71.3\%$ $61.2\%$
<span style="font-variant:small-caps;">DeepLab+RGR</span> $\mathbf{59.0\%}$ $\mathbf{74.2\%}$ $\mathbf{64.8\%}$ $\mathbf{86.8\%}$
<span style="font-variant:small-caps;">HSV-based</span> $39.7\%$ $56.8\%$ $65.6\%$ $50.1\%$
<span style="font-variant:small-caps;">Sppx+Clarifai</span> $40.5\%$ $57.6\%$ $49.6\%$ $68.7\%$
<span style="font-variant:small-caps;">DeepLab+RGR</span> $\mathbf{75.4\%}$ $\mathbf{86.0\%}$ $\mathbf{79.2\%}$ $\mathbf{94.1\%}$
: Summary of results obtained for each method.[]{data-label="tab:results"}
Our new model outperforms the baseline methods for all datasets evaluated, especially in terms of generalization to unseen datasets. By combining a deeper CNN architecture and the RGR refinement module, <span style="font-variant:small-caps;">DeepLab+RGR</span> improves both prediction and recall rates in the validation *AppleA* set by more than $15\%$. Figure \[fig:appleA\] provides a qualitative example of flower detection accuracy in this dataset.
As Figure \[fig:appleB\] illustrates, images composing the *AppleB* dataset present a higher number of flower buds and illumination changes, especially in terms of sunlight reflection by leaves. Despite the larger variance in comparison to the previous dataset, the performance obtained by <span style="font-variant:small-caps;">DeepLab+RGR</span> surpasses $77\%$ in terms of $F_1$.
Results obtained for the *Peach* dataset demonstrate the limitation of color-based methods and two important generalization characteristics of our model. The <span style="font-variant:small-caps;">HSV-based</span> method is incapable of detecting peach flowers, since their pink color is very different from the white apple blossoms used for training. On the other hand, our method presents $F_1$ near $75\%$, indicating that it can properly detect even flowers that differ to a great extent from apple flowers in terms of color. Moreover, images composing this dataset are characterized by a cloudy sky and hence poorer illumination. Most cases of false negatives correspond to flower buds, due to the lack of such examples in the training dataset. As illustrated in Figure \[fig:peach\], poor superpixel segmentation leads the <span style="font-variant:small-caps;">Sppx+Clarifai</span> approach to incorrectly classify parts of the sky as flowers. This problem is overcome by our new model, which greatly increases precision rates to above $80\%$.
(0,0) – (.5,0) node \[right,color=black\] (text1) ;; (text1.east) – (\[xshift=5mm\]text1.east) node \[right,color=black\] (text2) ;; (text2.east) – (\[xshift=5mm\]text2.east) node \[right,color=black\] (text3) ;;
Furthermore, the high recall rate provided by <span style="font-variant:small-caps;">DeepLab+RGR</span> in the *Pear* dataset demonstrates its robustness to slight variations in both flower morphology and color. As shown in Figure \[fig:pear\], similar to the *Peach* dataset, these images also present a cloudy background. In addition to that, their background is characterized by a high level of clutter caused by the presence of a large number of branches. These high texture components compromise the background removal model used by <span style="font-variant:small-caps;">Sppx+Clarifai</span>. Still, the <span style="font-variant:small-caps;">DeepLab+RGR</span> method provides a very accurate detection of flowers, with precision above $90\%$.
The results obtained by our method for *AppleB*, *Peach* and *Pear* datasets can be further improved by adjusting the parameter $\tau_0$ used for final classification and refinement. As summarized in Figure \[fig:tauF\], increasing $\tau_0$ from $0.3$ to $0.5$ increases in $3\%$ the $F_1$ performance on *AppleB*, reaching both recall and precision levels around $80\%$. For the *Peach* dataset, decreasing $\tau_0$ to $0.2$ increases the recall rate to above $70\%$. Such adjustment can be carried out quickly through a simple interactive procedure, where $\tau_0$ is chosen according to its visual impact on the segmentation of a single image.
![Segmentation performance in terms of $F_1$ measure on each dataset according to the parameter $\tau_0$.[]{data-label="fig:tauF"}](./Figures/Fig10-eps-converted-to.pdf){width="49.00000%"}
In terms of inference time, the current implementation of our algorithm on an Intel XeonCPU E5-2620 v3 @ 2.40GHz (62GB) with a Quadro P6000 GPU requires on average $50$ seconds to evaluate each high-resolution image composing our datasets. Around $5$ seconds are required to save portraits as individual files and load their corresponding prediction scores, a process that can be simplified by generating portraits directly within the neural network framework.
{#section-2}
We have presented a novel automated approach for flower detection, which exploits state-of-the-art deep learning techniques for semantic image segmentation. The applicability of our method was demonstrated by its high flower segmentation accuracy across datasets that vary in terms of illumination conditions, background composition, image resolution, flower density and flower species. Without any supervised fine-tuning or image pre-processing, our model trained using only images of apple flowers succeeded in generalizing for peach and pear flowers, which are noticeably different in terms of color and morphology.
In the future, we intend to further improve the generalization capabilities of our model by training and evaluating it on multi-species flower datasets. We ultimately aim at a completely autonomous system capable of online bloom intensity estimation. The current implementation of our model can evaluate high-resolution images of complete trees an order of magnitude faster than human workers. While in this work we are not creating maps of flowers at the block level, this method will scale well for precision agricultural applications such as predicting thinning spray treatments and timing.
[^1]: Manuscript received: February, 24, 2018; Revised May, 17, 2018; Accepted June, 15, 2018.
[^2]: This paper was recommended for publication by Cyrill Stachniss upon evaluation of the Associate Editor and Reviewers’ comments.
[^3]: We acknowledge the support of USDA-ARS agreement \#584080-5-020, and of NVIDIA Corporation with the donation of the GPU used for this research.
[^4]: \*Mention of trade names or commercial products in this publication is solely for the purpose of providing specific information and does not imply recommendation or endorsement by the U.S. Department of Agriculture. USDA is an equal opportunity provider and employer.
[^5]: $^{1}$P.A. Dias and H. Medeiros are with the Department of Electrical and Computer Engineering, Marquette University, Milwaukee, WI, USA, [[email protected] ; [email protected]]{}
[^6]: $^{2}$A. Tabb is with the U.S. Department of Agriculture (USDA), Kearneysville, WV, USA [[email protected]]{}
[^7]: Digital Object Identifier (DOI): 10.1109/LRA.2018.2849498
[^8]: The citation information for this paper is: P. A. Dias, A. Tabb and H. Medeiros, “Multispecies Fruit Flower Detection Using a Refined Semantic Segmentation Network,” in IEEE Robotics and Automation Letters, vol. 3, no. 4, pp. 3003-3010, Oct. 2018.
[^9]: We will make the annotation tool publicly available as future work.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Understanding micro-architectural behavior is profound in efficiently using hardware resources. Recent work has shown that, despite being aggressively optimized for modern hardware, in-memory online transaction processing (OLTP) systems severely underutilize their core micro-architecture resources [@Sirin:2016]. Online analytical processing (OLAP) workloads, on the other hand, exhibit a completely different computing pattern. OLAP workloads are read-only, bandwidth-intensive and include various data access patterns including both sequential and random data accesses. In addition, with the rise of column-stores, they run on high performance engines that are tightly optimized for the efficient use of modern hardware. Hence, the micro-architectural behavior of modern OLAP systems remains unclear.
This work presents the micro-architectural analysis of a breadth of OLAP systems. We examine CPU cycles and memory bandwidth utilization. The results show that, unlike the traditional, commercial OLTP systems, traditional, commercial OLAP systems do not suffer from instruction cache misses. Nevertheless, they suffer from their large instruction footprint resulting in slow response times. High performance OLAP engines execute tight instruction streams; however, they spend 25 to 82% of the CPU cycles on stalls regardless the workload being sequential- or random-access-heavy. In addition, high performance OLAP engines underutilize the multi-core CPU or memory bandwidth resources due to their disproportional compute and memory demands. Hence, analytical processing engines should carefully assign their compute and memory resources for efficient multi-core micro-architectural utilization.
author:
- |
Utku Sirin\
\
Anastasia Ailamaki\
\
bibliography:
- 'vldb\_sample.bib'
title: 'Micro-architectural Analysis of OLAP: Limitations and Opportunities \[Experiment and Analysis\]'
---
Introduction {#section:introduction}
============
Online analytical processing (OLAP) is an ever-growing, multi-billion dollar industry. Many industrial and community organizations rely on fast and efficient analytical processing to extract valuable information from their data. Understanding micro-architectural behavior of OLAP systems, on the other hand, is profound in providing high performance. Micro-architectural behavior reveals the limitations and opportunities in efficiently using modern hardware resources, and hence allows delivering high performance. Research has shown that OLAP systems can improve performance orders of magnitude by more efficiently using the modern hardware resources [@Manegold:2002].
Micro-architectural behavior of online transaction processing (OLTP) systems has been investigated extensively. Recent work has shown that, despite being aggressively optimized for modern hardware, in-memory OLTP systems severely underutilize their core micro-architecture resources [@Sirin:2016]. OLAP workloads, on the other hand, exhibit a completely different computing pattern. Unlike the update-heavy OLTP workloads, OLAP workloads are read-only. Therefore, they require neither a concurrency control and logging mechanism nor a complex buffer pool for synchronizing the modified pages with disk. Moreover, OLAP workloads are arithmetic-operation- and bandwidth-intensive. Th- ey process large amounts of data with various data access patterns including both sequential and random data accesses.
In addition, with the rise of column-stores [@Abadi:2013; @Idreos:2012], a diverse set of OLAP execution models, e.g., vectorized [@Boncz:2006] and compiled execution [@Kemper:2011], and system prototypes, e.g., Proteus [@Karp:2016] and Typer and Tectorwise [@Kersten:2018] have been proposed. Most major database systems such as SQL Server, Oracle, and DB2 now support a column-store extension [@Lahiri:2015; @Larson:2011; @Raman:2013]. Column-stores operate only on the columns that are necessary for the query, and hence utilize memory bandwidth more efficiently. Moreover, they process columns in tight, hardware-friendly execution loops optimized for the efficient use of the CPU cycles.
Following a read-only, arithmetic-operation- and bandwid- th-intensive computing pattern, and running on systems using various execution models, micro-architectural behavior of modern OLAP systems remains unclear. In this paper, we examine the hardware behavior of a breadth of OLAP systems from different categories of systems and execution models. We profile a traditional, commercial row-store, DBMS R, a column-store extension of a traditional, commercial row-store, DBMS C, an open source, high performance OLAP engine following a compiled execution model, Typer [@Kersten:2018], and an open source, high performance OLAP engine following a vectorized execution model, Tectorwise [@Kersten:2018]. We examine both CPU cycles and memory bandwidth utilization. We examine how well each OLAP system utilizes the hardware resources and what the limitations and opportunities are. Our paper demonstrates the following:
- Unlike the traditional, commercial OLTP systems, the traditional, commercial OLAP system and its column-store extension do not suffer from instruction cache misses. Nevertheless, they suffer from their large instruction footprints, which results in orders of magnitude lower performance than high performance OLAP engines.
- High performance OLAP engines execute a tight instruction stream; however, they dramatically suffer from high stall cycles ratio regardless the workload being sequential- or random-data-access-heavy. Workloads with large sequential scans stress the memory subsystem. Hence, the hardware prefetchers fall behind, and 50 to 75% of the CPU cycles are spent on stalls. Workloads with many random data accesses suffer from long-latency data stalls, and hence spend 25 to 82% of the CPU cycles on stalls.
- High performance OLAP engines underutilize multi-core CPU or memory bandwidth resources due to their disproportional compute and memory demands. Workloads with large sequential scans quickly saturate the multi-core memory bandwidth underutilizing multi-core CPU resources. Workloads with many random accesses saturate the multi-core CPU resources before saturating the multi-core memory resources, leaving the multi-core memory bandwidth underutilized. Hence, analytical processing engines should carefully assign their compute and memory resources for efficient multi-core micro-architectural utilization.
The rest of the paper is organized as follows. Section \[section:methodology\] presents the experimental setup and methodology. Section \[section:projection\], \[section:selection\] and \[section:join\] present the projection, selection and join micro-benchmark analyses. Section \[section:tpch\] presents the analysis of TPC-H queries. Section \[section:predication\], \[section:simd\], \[section:prefetchers\] and \[section:multi-core\] present the analyses of predication, SIMD, hardware prefetchers, and multi-core execution. Lastly, Section \[section:relatedwork\] and \[section:conclusions\] present the related work and conclusions.
Setup & Methodology {#section:methodology}
===================
This section presents our experimental setup and methodology.
**Benchmarks:** We use micro-benchmarks and a subset of TPC-H queries [@tpc]. We use projection, selection and join micro-benchmarks as they constitute the basic SQL operators. All the systems use hash join algorithm when running the join micro-benchmark. We also performed a group by micro-benchmark, however, we observe that it behaves similarly to the join at the micro-architectural level, and hence, omitted the discussion on them.
All the micro-benchmarks use the TPC-H schema. The projection micro-benchmark does a single SUM() over a set of columns from the lineitem table. We vary the number of columns from one to four. We use l\_extendedprice, l\_discount, l\_tax and l\_quantity columns. We add the projected columns inside the SUM(). We call the projection micro-benchmark doing a SUM() over $n$ columns as a projection query with degree of $n$.
The selection micro-benchmark extends the projection qu- ery with degree of four with a WHERE clause of three predicates over three columns of the lineitem table: l\_shipdate, l\_commitdate and l\_receiptdate. It varies the selectivity of each individual predicate from 10% to 50% and 90%. The join micro-benchmark does a join over two tables followed by a projection. The small-sized join micro-benchmark joins supplier and nation tables over the nationkey attribute, and does a SUM() over the addition of s\_acctbal and s\_suppkey. The medium-sized join joins partsupplier and supplier tables over the supplierkey attribute, and does a SUM() over the addition of ps\_availqty and ps\_supplycost. The large-sized join joins lineitem and orders table over the orderkey attribute, and does a SUM() over the addition of the four columns that the projection query with degree of four uses.
We also profile four TPC-H queries: Q1, Q6, Q9 and Q18. We chose these queries as each represents a particular category. Q1 is a low-cardinality group by (4 groups), Q6 is a highly selective filter, Q9 is a join-intensive query and Q18 is high-cardinality group by (1.5 million groups).
[|c|c|]{}Processor &
------------------------
Intel(R) Xeon(R) CPU
E5-2680 v4 (Broadwell)
------------------------
: Broadwell server parameters.[]{data-label="table:broadwell"}
\
\#sockets & 2\
\#cores per socket & 14\
Hyper-threading & Off\
Turbo-boost & Off\
Clock speed & 2.40GHz\
Per-core bandwidth &
---------------------
12GB/s (sequential)
7GB/s (random)
---------------------
: Broadwell server parameters.[]{data-label="table:broadwell"}
\
Per-socket bandwidth &
---------------------
66GB/s (sequential)
60GB/s (random)
---------------------
: Broadwell server parameters.[]{data-label="table:broadwell"}
\
L1I / L1D (per core) &
-----------------------
32KB / 32KB
16-cycle miss latency
-----------------------
: Broadwell server parameters.[]{data-label="table:broadwell"}
\
L2 (per core) &
-----------------------
256KB
26-cycle miss latency
-----------------------
: Broadwell server parameters.[]{data-label="table:broadwell"}
\
L3 (shared) &
------------------------
(inclusive) 35MB
160-cycle miss latency
------------------------
: Broadwell server parameters.[]{data-label="table:broadwell"}
\
Memory & 256GB\
**Hardware:** We conduct our experiments on an Intel Broadwell server. Table \[table:broadwell\] presents the server parameters. As the Broadwell micro-architecture does not support AVX-512 instructions, we conduct the SIMD experiments on a separate Skylake server. The Skylake server has a similar execution engine but a different memory hierarchy from the Broadwell server. The Skylake server has a significantly larger L2 cache (1 MB), a smaller non-inclusive L3 cache (16MB), a smaller per-core (10 GB/s) and a larger per-socket (87 GB/s) sequential access bandwidth. It has a similar per-core and per-socket random access bandwidth.
We use Intel’s Memory Latency Checker (MLC) [@mlc] to measure cache access latencies and maximum single- and multi-core bandwidths.
**OLAP systems:** We examine a commercial row-store, DBMS R, the column-store extension of the row-store, DBMS C, an open-source OLAP engine implementing a compiled execution engine, Typer [@Kersten:2018], and an open-source OLAP engine implementing a vectorized execution engine, Tectorwise [@Kersten:2018]. We chose these systems as each represents a different category of a system and execution model. As commercial systems are closed-source we do a best-effort explanation of their behavior.
**OS & Compiler:** We use Ubuntu 16.04.6 LTS and gcc 5.4.0 on the Broadwell server, and Ubuntu 18.04.2 LTS and gcc 7.4.0 on the Skylake server.
**VTune:** We use Intel VTune 2018 on the Broadwell server, and VTune 2019 on the Skylake server. We use VTune’s built-in general-exploration (uarch-exploration on VTune 20- 19) analysis type for CPU cycles breakdown. VTune’s general-exploration provide full CPU cycles breakdown [@Sirin:2017; @Yasin:2014]. We examine CPU cycles at two-levels. We firstly break down the CPU cycles into Retiring and Stall cycles. Retiring cycles represent the percentage of the useful cycles spent on retiring instructions. Stall cycles represent the percentage of the wasted cycles spent on stalls. Secondly, we zoom into the Stall cycles breakdown composed of five components: (i) Branch mispredictions, (ii) Icache, (iii) Decoding, (iv) Dcache and (v) Execution. Branch mispredictions represent stalls due to the branch mispredictions. Icache represents Icache misses stalls. Decoding represents the stalls due to the inefficiencies in the instruction decoding micro-architecture. Dcache represents the stalls due to memory hierarchy. Lastly, Execution represents the stalls due to the saturation of the core execution resources.
We use VTune’s built-in memory-access analysis type to measure the used memory bandwidth. As we numa-localize our experiments on a single socket, we report average bandwidth per-socket values.
**Measurements:** For every experiment, we firstly populate the database. We use one minute of warmup period followed by three minutes of VTune profiling period. We disable Hyper-threading and Turbo-boost as they jeopardize VTune counter values [@Intel:2018].
We numa-localize every experiment by using Linux’s numactl command. Except the multi-core execution experiments presented in Section \[section:multi-core\], we run all the experiments on a single core over a TPC-H database with scaling factor of five, i.e., a database of 5GB. We chose scaling factor of five as it is large enough to run out-of-cache experiments. We run the multi-core experiments on a single socket, over a TPC-H database with scaling factor of 70, i.e., database of 70GB.
Hardware prefetchers experiments in Section \[section:prefetchers\] are done by modifying the relevant model-specific register (msr) of the processor [@hwpfers].
Projection {#section:projection}
==========
![CPU cycles breakdown for projection as projectivity increases for DBMS R and DBMS C.[]{data-label="fig:proj_cpu1"}](figures/proj_cpu1)
![Stall cycles breakdown for projection as projectivity increases for DBMS R.[]{data-label="fig:proj_stall1"}](figures/proj_stall1)
This section presents the micro-architectural analysis of the projection micro-benchmark. Our goal is to observe how the micro-architectural behavior changes as the projectivity increases. Figure \[fig:proj\_cpu1\] shows the CPU cycles breakdown for DBMS R and C. The figure shows that while DBMS R spends about half of the CPU cycles for Retiring, DBMS C spends almost 90% of the CPU cycles for Retiring.
Figure \[fig:proj\_stall1\] shows the stall cycles breakdown. As the figure shows, DBMS R spends the majority of the stall cycles on Dcache and Execution stalls. This shows that, unlike the traditional, commercial OLTP systems spending the majority of the CPU cycles to Icache stalls, traditional, commercial OLAP systems do not suffer significantly from Icache stalls. On the other hand, DBMS C spends the majority of the stall cycles on Branch mispredictions and Icache stalls. However, its overall stall cycles ratio is less than 10%. Hence, none of the systems suffer significantly from the Icache stalls.
![CPU cycles breakdown for projection as projectivity increases for Typer and Tectorwise.[]{data-label="fig:proj_cpu2"}](figures/proj_cpu2)
![Stall cycles breakdown for projection as projectivity increases for Typer and Tectorwise.[]{data-label="fig:proj_stall2"}](figures/proj_stall2)
Figure \[fig:proj\_cpu2\] shows the CPU cycles breakdown for Typer and Tectorwise. As can be seen, both Typer and Tectorwise spend the majority of CPU cycles on stalls. While Typer’s stall cycles ratio increases from 60% to 75% as projectivity increases, Tectorwise’s stall cycles ratio remains 60% as the projectivity increases.
Figure \[fig:proj\_stall2\] shows the stall cycles breakdown for Typer and Tectorwise. We observe that the stall cycles of Typer are mainly due to Dcache stalls. Moreover, as the projectivity increases, Dcache stall cycles ratio also increases. On the other hand, the stall cycles breakdown of Tectorwise remains the same as the projectivity increases: Dcache and Execution stalls contribute equally to the stall cycles.
The reason for the increasing Dcache stalls of Typer is the saturation of the single-core sequential memory bandwidth, which results in a super-linear increase in the Dcache stalls. The reason for the stable stall cycles breakdown of Tectorwise is the vectorized execution model. Vectorized execution model adds two vectors and materializes the result in another intermediate vector, which will later be used to add the other columns in the aggregation. Hence, at the degree of two and onwards, the processor is subject to the same computing pattern: two vectors being added and the result being written to a third vector. As a result, the stall cycles breakdown remains stable despite the increasing projectivity.
![Single-core sequential access bandwidth utilization for Typer and Tectorwise when running projection as projectivity increases.[]{data-label="fig:proj_bw"}](figures/proj_bw)
![Normalized (with respect to Typer) response time breakdown for the projection query with degree of 4 across DBMSs.[]{data-label="fig:proj_exec_time_norm"}](figures/proj_exec_time_norm)
Figure \[fig:proj\_bw\] presents the memory bandwidth utilization of Typer and Tectorwise. As the figure shows, Typer almost saturates the single-core memory bandwidth at the projectivity of degree two and onwards due to the high pressure on the memory subsystem. Tectorwise, on the other hand, has low memory bandwidth utilization due to materialization overheads. The materialization overheads cut the memory pressure of Tectorwise.
The projection query is a simple sequential sum over a set of columns with a highly predictable data access pattern. As a result, hardware prefetchers are highly likely to predict the future data blocks. Despite that, Dcache stalls constitute a large portion of the execution time even when the memory bandwidth is not saturated. This shows that today’s hardware prefetchers are not fast enough for sequential-scan-heavy analytical workloads. We examine hardware prefetchers behavior in more detail in Section \[section:prefetchers\].
In addition, today’s power-hungry server processors provide wide execution engines. Our Broadwell server has eight execution ports, four of them including an ALU unit [@intel:2019]. Despite that, Typer at the projectivity of degree 1, and Tectorwise at all the projectivities significantly suffer from Execution stalls. This shows that, despite being wide, todays processors fall short on providing enough execution units for the arithmetic-heavy analytical workloads.
Figure \[fig:proj\_exec\_time\_norm\] presents the normalized response time breakdown for the four systems when running the projection micro-benchmark with degree of four. As the figure shows, DBMS R is two orders of magnitude slower than Typer and Tectorwise. While DBMS C’s optimizations make it an order of magnitude faster than DBMS R, it is nevertheless an order of magnitude slower than Typer and Tectorwise. Moreover, both DBMS R and C have high Retiring cycles ratio. As the number of Retiring cycles is correlated to the number of retired instructions, high Retiring cycles show that DBMS R and C severely suffer from their large instruction footprints.
Note that, despite that Typer and Tectorwise are open-source OLAP prototype engines, their performance is representative for fully implemented high-performant systems such as HyPer and VectorWise. Kersten et al. [@Kersten:2018] have compared the performance of Typer and Tectorwise with real-life systems HyPer and VectorWise, and have shown that Typer and Tectorwise are only slightly faster than HyPer and VectorWise.
![CPU cycles breakdown for selection as selectivity increases for DBMS R and DBMS C.[]{data-label="fig:sel_cpu1"}](figures/sel_cpu1)
![Stall cycles breakdown for selection as selectivity increases for DBMS R and DBMS C.[]{data-label="fig:sel_stall1"}](figures/sel_stall1)
Selection {#section:selection}
=========
Having examined the projection, we now move to examining the selection micro-benchmark. Our goal is to examine how influential the branch mispredictions stalls are on the micro-architectural behavior. Figure \[fig:sel\_cpu1\] shows the CPU cycles breakdown for DBMS R and C. We observe that the Retiring cycles ratio increases as the selectivity increases both for DBMS R and C. It is because the higher the amount of computation is (due to the higher selectivity), the more the instruction overheads consume the CPU cycles.
Figure \[fig:sel\_stall1\] shows the stall cycles breakdown. We observe that DBMS R does not suffer from Icache stalls significantly. Moreover, DBMS C suffers from the Decoding stalls only at the higher selectivities, where the overall stall cycles ratio is low. Hence, none of the commercial systems majorly suffer from the instruction-related stalls. Nevertheless, DBMS R and C are 1.6x to 40x slower than the high performance OLAP engines proving their large instruction footprints (graph not shown).
![CPU cycles breakdown for selection as selectivity increases for Typer and Tectorwise.[]{data-label="fig:sel_cpu2"}](figures/sel_cpu2)
![Stall cycles breakdown for selection as selectivity increases for Typer and Tectorwise.[]{data-label="fig:sel_stall2"}](figures/sel_stall2)
Figure \[fig:sel\_cpu2\] shows the CPU cycles breakdown for Typer and Tectorwise. Unlike DBMS R and C, Typer and Tectorwise have the highest stall cycles ratio at 50%. Figure \[fig:sel\_stall2\] shows the stall cycles breakdown. The figure shows that branch mispredictions dominate the stall cycles. In addition, the highest branch misprediction stalls are at the 50% selectivity. The reason is that the prediction task is the hardest at the 50% selectivity, corroborating the existing work by Sompolski et al. [@Sompolski:2011].
Typer suffers less from the branch mispredictions at the 10% selectivity compared to Tectorwise. This is because Tectorwise is a vectorized engine, and thus individually evaluates every predicate. As a result, the branch predictor always faces the 10% individual selectivity of each predicate. Typer, however, is a compiled engine, and thus evaluates all the predicates at once. Hence, the branch predictor faces a much lower overall selectivity (10% $\times$ 10% $\times$ 10% = 0.01%), making the prediction task easier for the branch predictor.
We also examine the memory bandwidth utilization for the selection micro-benchmark (graph not shown). We observe that memory bandwidth utilization is well below the maximum bandwidth both for Typer and Tectorwise. Typer uses, 3, 5 and 5 GB/s, whereas Tectorwise uses 2.5, 3 and 3 GB/s of memory bandwidth for selectivities of 10%, 50% and 90%, respectively. This is because of the branch misprediction stalls preventing the core creating enough memory traffic to saturate bandwidth.
Overall, the experiments in this section show that branch mispredictions constitute a significant performance bottleneck for high performance OLAP systems. We examine predication in Section \[section:predication\] to see how micro-architectural behavior changes when branch mispredictions are eliminated.
![CPU cycles breakdown for join for DBMS R and C. Sm., Md. and Lr. stand for Small, Medium and Large, respectively.[]{data-label="fig:join_cpu1"}](figures/join_cpu1)
![CPU cycles breakdown for join for Typer and Tectorwise. Sm., Md. and Lr. stand for Small, Medium and Large, respectively[]{data-label="fig:join_cpu2"}](figures/join_cpu2)
Join {#section:join}
====
In this section, we examine the join micro-benchmark. We force all the systems to use hash join algorithm. Unlike the selection and projection micro-benchmarks with a sequential data access pattern, hash join includes many random data accesses. Our goal is to understand the effect of random data accesses in the overall micro-architectural behavior.
Figure \[fig:join\_cpu1\] shows the CPU cycles breakdown for DBMS R and C. As can be seen, DBMS R and C spend 52 to 72% of the CPU cycles for retiring instructions for all the join sizes. Observing a similar CPU cycles breakdown despite the orders of magnitude increase in the joined table sizes indicates the large instruction footprint of DBMS R and C that overshadows the micro-architectural behavior.
![Stall cycles breakdown for join for Typer and Tectorwise. Sm., Md. and Lr. stand for Small, Medium and Large, respectively[]{data-label="fig:join_stall2"}](figures/join_stall2)
![Left: Single-core random access bandwidth utilization when running the large-sized join. Right: Normalized (with respect to Typer) response time breakdown for the large-sized join. R, C, Ty. and Tw. stand for DBMS R, DBMS C, Typer and Tectorwise.[]{data-label="fig:join_bw_exec_time_norm"}](figures/join_bw_exec_time_norm)
Figure \[fig:join\_cpu2\] and \[fig:join\_stall2\] show the CPU and stall cycles breakdown for Typer and Tectorwise. The CPU cycles breakdown shows that the stall cycles ratio increases as the join size increases. The Retiring cycles ratio can get as low as 18% for the large-sized join operation. The stall cycles breakdown shows that Dcache stalls become more and more dominant as the join size increases. As expected, the larger the joined table sizes are, the larger the number of random data accesses are, which results in an increased stall cycles ratio. On the other hand, for the small- and medium-sized joins, the Execution stalls constitute a significant portion of the stall cycles highlighting costly hash computations.
Figure \[fig:join\_bw\_exec\_time\_norm\] (left) shows the memory bandwidth utilization when running the large-sized join micro-benchmark. As the figure shows, the bandwidth utilizations are well-below the maximum random access bandwidth. This shows that Typer and Tectorwise do not create enough memory traffic to use the available single-core random access bandwidth. This finding corroborates with the existing work on using coroutines to improve hash join performance [@Jonathan:2018; @Yorgos:2017]. Coroutines allow overlapping long-latency memory stalls with computation enabling a more efficient utilization of the memory bandwidth. Psaropoulos et al. [@Yorgos:2018; @Yorgos:2019] has shown that memory bandwidth starts getting saturated with 28 cores (per-socket) and upwards.
Figure \[fig:join\_bw\_exec\_time\_norm\] (right) presents the normalized response time breakdowns. As the figure shows, DBMS R and C are 4.5x and 6.3x slower than Typer and Tectorwise by spending orders of magnitude more time on retiring instructions. Hence, DBMS R and C suffers from orders of magnitude larger instruction footprints compared to Typer and Tectorwise.
TPC-H {#section:tpch}
=====
Up to now, we have examined simple micro-benchmarks. In this section, we analyze four TPC-H queries: Q1, Q6 and Q9, Q18, each of which represents a particular class of queres: (i) Q1 is a low-cardinality group by (4 groups), (ii) Q6 is a highly selective filter, (iii) Q9 is a join-intensive query and (iv) Q18 is a high-cardinality group by (1.5 million groups). We, once again, observed orders of magnitude difference in the response times of the commercial and high performance systems. Hence, we omit the discussion on the commercial systems, and focus on the two high performance systems we profile.
Figure \[fig:tpch\_cpu2\] shows the CPU cycles breakdown. As can be seen, both Typer and Tectorwise have the highest Retiring cycles ratio when running Q1. Whereas Typer has the lowest Retiring cycles ratio when running Q9, and Tectorwise has the lowest Retiring cycles ratio when running Q6.
Figure \[fig:tpch\_stall2\] shows the stall cycles breakdown. The figure shows that the stall cycles breakdowns are dominated by the Execution stalls when running Q1. This is similar to the small-sized hash join micro-benchmark. Q1 has a small hash table keeping the small number of group aggregations. Hence, its main working set fits into the cache, and Dcache stalls are mostly eliminated. This time, however, the Execution stalls surface up causing the processor to stall $\sim$40%.
Figure \[fig:tpch\_stall2\] shows that Q6 is dominated by Dcache stalls for Typer, but Branch mispredictions for Tectorwise. The overall selectivity of Q6 is $\sim$2%. However, it includes five individual predicates with varying selectivities. As being a compiled engine, Typer only experiences the 2% overall selectivity, whereas Tectorwise, as being a vectorized engine, evaluates every predicate individually. Hence, the processor experiences the selectivities of the individual predicates that vary from 16% to 48%, causing much more branch mispredictions.
Figure \[fig:tpch\_stall2\] shows that Dcache stalls are the dominant factor in the overall stall cycles both for Typer and Tectorwise when running Q9. This is expected as Q9 is join-intensive, and hence makes many random data accesses. However, unlike the simple join micro-benchmark, Q9 also incurs a significant number of Branch misprediction stalls. This is because Q9’s successive hash joins require successive hash table probings whose outcome is less and less predictable as the processing moves upwards in the query plan. As a result, the branch predictor fails significant number of times resulting in as high Branch misprediction stalls as $\sim$38%.
![CPU cycles breakdown for TPC-H queries for Typer and Tectorwise.[]{data-label="fig:tpch_cpu2"}](figures/tpch_cpu2)
![Stall cycles breakdown for TPC-H queries for Typer and Tectorwise.[]{data-label="fig:tpch_stall2"}](figures/tpch_stall2)
Figure \[fig:tpch\_stall2\] shows that the stall cycles breakdown of Q18 is similar to that of Q9, with less Dcache stalls. This is because Q18 is a high-cardinality group by (1.5 million groups), which requires successive hashing of multiple attributes. As the number of groups is large, the hash table keeping the group aggregations does not fit into the cache. Hence, the group aggregation updates result in Dcache stalls. Q18 also significantly suffers from Branch mispredictions and Execution stalls. The Branch mispredictions stalls are due to the frequent hash collusions during the aggregation updates. The Execution stalls are due to the successive hash computations for grouping the group by attributes.
Hash tables built for group by operator cause more hash collusions than the hash tables built for hash join. The reason is that different groups that share a common group attribute have a higher chance of colluding than the evenly distributed primary/foreign keys. Our analysis of a group by and hash join micro-benchmark showed that the lengths of the hash table chains vary from 0 to 7 with an average chain length of 0.23 and standard deviation of 0.5 for the group by, whereas vary from 0 to 1 with an average chain length of 0.44 and standard deviation of 0.49 for the hash join micro-benchmark. Hence, the hash table built for the group by micro-benchmark is much more irregular than that of the hash join, causing more hash collusions.
We also examined the memory bandwidth utilization for TPC-H queries (graph not shown). We observed that the used bandwidth is always less than 1 GB/s for all the queries and for both Typer and Tectorwise, except Typer when running Q6. Q6 running on Typer has 4.7 GB/s bandwidth utilization. The reason is that Q6 is a low-selectivity selection query making sequential scans to evaluate the predicates. All the other queries suffer from costly hash computations preventing the system creating high memory pressure.
Overall, the experiments in this section show that the major micro-architectural bottlenecks that the micro-benchma- rks identified generalizes to complex TPC-H queries. Hence, we can evaluate micro-architectural behavior of a given query by examining its individual operators, and based on that, we can identify the limitations and opportunities in using the micro-architectural resources more efficiently.
![Response time breakdown for Typer when running branched and branch-free selection query.[]{data-label="fig:predication_typer_cpu"}](figures/predication_typer_cpu)
![Stall time breakdown for Typer when running branched and branch-free selection query.[]{data-label="fig:predication_typer_stall"}](figures/predication_typer_stall)
Predication {#section:predication}
===========
In this section, we examine the predication optimization. Predication is used to eliminate branches. Its idea is to convert control dependencies to data dependencies by computing the predicate as an arithmetic expression, and using it to increment the index/aggregation. The trade-off is doing more computation but avoid branches. Our goal is to examine how predication changes the micro-architectural behavior. We examine predication for Typer and Tectorwise.
![Response time breakdown for Tectorwise when running branched and branch-free selection query.[]{data-label="fig:predication_twise_cpu"}](figures/predication_twise_cpu)
![Stall time breakdown for Tectorwise when running branched and branch-free selection query.[]{data-label="fig:predication_twise_stall"}](figures/predication_twise_stall)
Figure \[fig:predication\_typer\_cpu\] shows the response time breakdown for Typer. While Br. on the x-axis refers to the branched version, Br.-free refers to the predicated, branch-free version. The figure shows that predication increases the response time for 10% selectivity, whereas decreases the response time for 50% and 90% selectivities. Figure \[fig:predication\_typer\_stall\] shows the stall time breakdown for Typer. The figure shows that predication successfully eliminates the branch mispredictions, and makes the selection query Dcache- and Execution-stalls-bound similar to the projection query.
Figure \[fig:predication\_twise\_cpu\] and \[fig:predication\_twise\_stall\] shows the response and stall time breakdown for Tectorwise. As the figures show predication always decreases the response time for all the selectivities, and eliminates the branch misprediction stalls. In addition, predication makes the selection query Dcache- and Execution-stalls-bound similar to the projection query.
The reason for predication to be more successful on Tectorwise than Typer is that Tectorwise relies on intermediate selection vectors to implement the selection operator. As a result, the only additional computation that predication brings is to compute more during the computation of the selection vectors, whereas the bulk of the projection computation remains the same. Whereas for Typer, predication requires computing the projection for the entire table for all the selectivities, which pays off for 50% and 90%, but does not pay off for 10% selectivity.
We also profiled predicated TPC-H, Q6 on Typer and Tectorwise, and we reached similar conclusions. While Typer’s execution time is decreased by 11%, Tectoriwise’s execution time is decreased by 52% thanks to predication.
![Single-core bandwidth utilization for Typer and Tectorwise when running the predicated selection queries.[]{data-label="fig:predication_bw"}](figures/predication_bw)
Figure \[fig:predication\_bw\] shows the memory bandwidth utilization. As can be seen, Typer’s memory bandwidth utilization is high similar to the projection query, and remains stable across the selectivities. This is because Typer’s predicated selection queries only involve scanning the predicated and the projected columns and making arithmetic operations on them.
On the other hand, Tectorwise’s bandwidth utilization is lower than that of Typer. This is because of the materialization overheads. In addition, the bandwidth utilization of Tectorwise varies across the selectivities. The reason for the highest bandwidth utilization at 50% is that the data access pattern is the most confusing for the hardware prefetchers at the 50% selectivity. As a result, hardware prefetchers create unnecessary memory traffic resulting in higher bandwidth consumption. The bandwidth utilization of 90% selectivity is higher than that of 10% selectivity as more data is processed at the 90% selectivity. When not using predication, the bandwidth utilization is 3, 5 and 5 GB/s for Typer and 2.5, 3 and 3 GB/s for Tectorwise for 10%, 50% and 90% selectivities, respectively (graph not shown). Hence, predication significantly improves the bandwidth utilization both for Typer and Tectorwise, though less for Tectorwise due to the materialization overheads.
We also profiled the bandwidth utilization of the predicated Q6. The results showed that the bandwidth utilization of Typer increases from 4.7 to 6.9 GB/s, whereas the bandwidth utilization of Tectorwise increases from 1 to 4.7 GB/s. Hence, predication significantly increases the memory bandwidth utilization for TPC-H, Q6, too.
Overall, the experiments in this section show that predicated selection queries behave similarly to the projection queries at the micro-architectural level. While Dcache and Execution stalls are the main stall cycle bottlenecks, Typer stresses the memory bandwidth close to the maximum, where- as Tectorwise underutilizes the memory bandwidth due to the materialization overheads.
SIMD {#section:simd}
====
The second optimization we examine is SIMD. SIMD instructions are used to reduce the number of instructions required to perform arithmetic operations. We test Tectorwise when running the projection, selection and join micro-benchmarks with and without using the SIMD instructions. As our Broadwell server does not support AVX-512 instructions, we do all the SIMD experiments on a Skylake server supporting AVX-512 instructions.
Note that the Skylake server that we profile in this section has a different memory hierarchy with a different maximum memory bandwidth than that of the Broadwell server. As a result, the reported values that do not use SIMD do not exactly match with the values reported earlier in the paper (see Section \[section:methodology\], Hardware subsection for more details).
Projection & Selection
----------------------
In this section, we examine the projection and selection micro-benchmarks. We use the predicated, branch-free versions of the selection queries as SIMD is more effective when branch mispredictions are eliminated. Figure \[fig:simd\_proj\_sel\_cpu\] shows the normalized response time breakdown, where the response time without SIMD is taken as the base.
![Normalized response time breakdown for Tectorwise when running projection and selection queries with and without SIMD.[]{data-label="fig:simd_proj_sel_cpu"}](figures/simd_proj_sel_cpu)
![Normalized stall time breakdown for Tectorwise when running projection and selection queries with and without SIMD.[]{data-label="fig:simd_proj_sel_stall"}](figures/simd_proj_sel_stall)
Figure \[fig:simd\_proj\_sel\_cpu\] shows that SIMD reduces the response time by 22% for projection, and 42%, 23% and 21% for selection for 10%, 50% and 90% selectivities. For all the four cases, the figure shows that there is 70% to 87% decrease in the amount of time spent for Retiring cycles. As Retiring cycles are correlated to the number of retired instructions, reduced Retiring cycles shows that SIMD successfully reduces the number of retired instructions.
Figure \[fig:simd\_proj\_sel\_stall\] shows the normalized stall time breakdown, where the stall time without using SIMD is taken as the base. The figure shows that SIMD increases Dcache stalls while reducing the Execution stalls. The decrease in the Execution stalls is due to using the SIMD execution engine to perform the arithmetic operations rather than the regular arithmetic logic units. The increase in the Dcache stalls is because SIMD instructions stress the memory bandwidth more, which results in waiting more on Dcache stalls. Therefore, SIMD instructions make the projection and predicated selection queries more Dcache-stalls-bound than Execution-stalls-bound.
![Single-core bandwidth utilization for Tectorwise when running projection and selection queries with and without SIMD.[]{data-label="fig:simd_proj_sel_bw"}](figures/simd_proj_sel_bw)
Figure \[fig:simd\_proj\_sel\_bw\] shows the single-core bandwidth utilization for the projection and selection micro-benchmarks. We observe that SIMD significantly increases the bandwidth utilization for almost all the cases. This shows that SIMD effectively exploits the underutilized bandwidth.
SIMD does not significantly improve the bandwidth utilization at 50% selectivity. This is because the data access pattern is the most confusing for the hardware prefetchers at the 50% selectivity. As a result, the consumed bandwidth is already high when not using SIMD leaving only a small room for SIMD to exploit. Overall, the experiments in this section show that a vectorized engine, despite suffering from vectorization overheads, can significantly stress the memory bandwidth by using SIMD instructions.
Join {#join}
----
In this section, we examine the large-sized join micro-benchmark. We compare only the hash table probing phases. Figure \[fig:simd\_join\_cpu\_stall\_bw\] (left) shows the normalized response time breakdown, where the response time without SIMD is taken as the base. Note that the response time breakdown includes the stall time breakdown inside on the same graph.
Figure \[fig:simd\_join\_cpu\_stall\_bw\] (left) shows that SIMD reduces the response time by 27%. The reason is both the reduced number of retired instructions and the reduced Dcache stalls. We examine the memory bandwidth utilization in Figure \[fig:simd\_join\_cpu\_stall\_bw\] (right). The figure shows that SIMD improves the utilization by 50%. The reduced data stalls and increased bandwidth utilization show that SIMD effectively parallelizes the random accesses of hash table probings.
![Left: Normalized response time breakdown for Tectorwise when running the large-sized join query. Right: Single-core bandwidth utilization for Tectorwise when running the large-sized join query.[]{data-label="fig:simd_join_cpu_stall_bw"}](figures/simd_join_cpu_stall_bw)
Prefetchers {#section:prefetchers}
===========
Section \[section:projection\] and \[section:predication\] have shown that the projection and predicated selection queries suffer from Dcache stalls. Both the projection and predicated selection queries are essentially sequential scans of the relevant columns with a highly predictable data access pattern. Despite that, large Dcache stalls raise the question how useful hardware prefetchers are.
In this section, we study the four hardware prefetchers that today’s server processors provide: L1 next line (L1 NL), L1 streamer (L1 Str.), L2 next line (L2 NL) and L2 streamer (L2 Str.) prefetchers. We turn on and off the four hardware prefetchers and examine their effects on the micro-architectural behavior. We examine the following six configurations: (i) all hardware prefetchers are disabled, (ii) only L1 NL enabled, (iii) only L1 Str. enabled, (iv) only L2 NL enabled, (v) only L2 Str. enabled, and (vi) all hardware prefetchers are enabled (which is the default case for all the experiments in the rest of the paper).
![Response time breakdown for the six prefetcher configuration for Typer when running the projection micro-benchmark with degree of four.[]{data-label="fig:prefetcher_proj_cpu_stall"}](figures/prefetcher_proj_cpu_stall)
Figure \[fig:prefetcher\_proj\_cpu\_stall\] shows the response time breakdown for the six prefetcher configurations we examine for Typer when running the projection query with degree of four. Observe that the response time breakdown includes the stall time breakdown inside on the same graph.
The figure shows that prefetchers reduce the Dcache stalls by 85%, and the response time by 73%. Moreover, L2 streamer is the most effective hardware prefetcher as it alone reduces the response time and the stalls as much as all the four prefetchers together do.
We also examined the projection query on Tectoriwse, and the branched and branch-free selection queries on Typer and Tectorwise. The results agree with our findings for the projection query on Typer. Hence, we omit the graphs for these experiments. We also examined the join micro-benchmark. As expected, prefetchers are in general not so useful for the join queries due to the large number of random data accesses. Prefetchers reduce the response time by $\sim$ 20% for the large-sized join both for Typer and Tectorwise. We omit the graphs for these queries.
Overall, the experiments in this section show that hardware prefetchers are indeed effective in reducing Dcache stalls for large-sequential-scan workloads. However, they are nevertheless not fast enough, which results in 50 to 75% of the CPU cycles spent on stalls.
Multi-core Execution {#section:multi-core}
====================
We lastly examine the hardware utilization for multi-core execution. Most OLAP operations scale well across multi-cores. As a result, we do not expect a big difference in the micro-architectural behavior of the multi-core execution compared to the single-core execution. We use the four TPC-H queries as they are more complex than the micro-benchmark queries, and hence harder to scale. We profile the systems at fourteen threads, i.e., at the number of cores per socket, as they all have the highest performance at fourteen threads.
Figure \[fig:mt\_tpch\_cpu\] and \[fig:mt\_tpch\_stall\] show the CPU and stall cycles breakdowns. As can be seen, both CPU and stall cycles breakdowns are similar to the single-core breakdowns. While the low-cardinality group by Q1 has the highest Retiring cycles ratio both for Typer and Tectorwise, join-intensive Q9 has the lowest Retiring cycles ratio for Typer, and highly selective filter Q6 has the lowest Retiring cycles ratio for Tectorwise. While Execution stalls dominate Q1, Dcache stalls dominate the rest of the queries for Typer and Tectorwise, except that Q6 is Branch misprediction dominated for Tectorwise.
While multi-core execution does not create a significant difference in the core micro-architectural behavior, it creates an increasing pressure on the multi-core memory bandwidth. In Figure \[fig:mt\_proj\_bw\] and \[fig:mt\_join\_bw\], we examine the bandwidth utilization for the projection query with degree of four and for the large-sized join query. We measure per-socket average memory bandwidth utilization. We choose the micro-benchmarks as they are easier to scale, and hence, stress the memory bandwidth more.
![CPU cycles breakdown for multi-core execution when running TPC-H queries on Typer and Tectorwise.[]{data-label="fig:mt_tpch_cpu"}](figures/mt_tpch_cpu)
![Stall cycles breakdown for multi-core execution when running TPC-H queries on Typer and Tectorwise.[]{data-label="fig:mt_tpch_stall"}](figures/mt_tpch_stall)
![Multi-core bandwidth utilization for Typer and Tectorwise when running the projection query with degree of four.[]{data-label="fig:mt_proj_bw"}](figures/mt_proj_bw)
Figure \[fig:mt\_proj\_bw\] shows that Typer saturates the multi-core bandwidth at eight cores. This shows that Typer’s performance is bandwidth-limited after eight cores. Hence, using more than eight cores for Typer when running the projection query would waste the cores. Tectorwise saturates the bandwidth at twelve cores. This is because Tectorwise underutilizes the per-core bandwidth due to the materialization overheads. At twelve cores, however, it reaches the maximum per-socket memory bandwidth. Hence, using more than twelve cores would waste the cores when Tectorwise running the projection query.
Figure \[fig:mt\_join\_bw\] presents the bandwidth utilization when running the large-sized join query. As the figure shows, both Typer and Tectorwise largely underutilize the memory bandwidth. This is because costly hash computations are preventing the system to create enough memory traffic. As a result, Typer and Tectorwise saturates the multi-core CPU resources before saturating the multi-core memory resources, leaving the multi-core memory bandwidth underutilized.
We also profiled the bandwidth utilization of the TPC-H queries (graph not shown). The results showed that the bandwidth utilization varies between the high utilization of the projection and the low utilization of the join micro-benchmarks. While the predicated Q6 comes close to the maximum sequential bandwidth (56 GB/s both for Typer and Tectorwise), Q1, Q9 and Q18 exhibits a similar bandwidth utilization to the join micro-benchmark due to their low pressure on the memory bandwidth.
Overall, the experiments in this section show that the disproportional compute and memory demands of the OLAP systems result in underutilization of either the compute or memory resources. Hence, analytical processing systems should carefully schedule their compute and memory resources to efficiently use the multi-core micro-architectural resources.
The memory pressure of Tectorwise can be increased by using SIMD when running the join query. If we apply the SIMD improvement in Section \[section:simd\], the multi-core bandwidth utilization of Tectorwise would increase from 21 to 31.5 GB/s. Similarly, the bandwidth utilization of Typer and Tectorwise can be improved by using hyper-threading. Our analysis with hyper-threading showed that the bandwidth utilization is improved by 1.3x both for Typer and Tectorwise. Hence, Tectorwise’s (together with SIMD) and Typer’s bandwidth utilizations would raise up to 40 GB/s and 27 GB/s when hyper-threading is enabled. While the improvements are substantial, they are nevertheless below the maximum random access bandwidth. Moreover, our main finding of underutilization of compute or memory resources in the face of disproportional demands still holds for many scenarios.
Related Work {#section:relatedwork}
============
There is a large body of work on the micro-architectural analysis of database workloads. Ailamaki et al. [@Ailamaki:1999] and Hardavellas et al. [@Hardavellas:2007] present database workload characterization both for analytical and transactional workloads. Tozun et al. [@Tozun:2013b; @Tozun:2013a] presents micro-architectural analysis of disk-based OLTP systems. Sirin et al. [@Sirin:2016] presents micro-architectural analysis of a breadth of OLTP systems. Our work complements all these studies by presenting an analysis of modern analytical processing systems on a modern processor.
Kersten et al. [@Kersten:2018] presents an analysis of vectorized and compiled OLAP engines without getting deep into micro-architectural analysis. Sompolski et al. [@Sompolski:2011] presents a comparison between vectorized and compiled engines in terms of particular optimizations such as predication and SIMD. Our work comprehensively extends and complements these works focusing on a deep analysis of single- and multi-core micro-architectural behavior of a breadth of OLAP systems.
![Multi-core bandwidth utilization for Typer and Tectorwise when running the large-sized join query.[]{data-label="fig:mt_join_bw"}](figures/mt_join_bw)
Ferdman et al. [@Ferdman:2012] presents micro-architectural analysis of a suite of cloud workloads, concluding that there is a fundamental mismatch among what today’s server processors provide and what cloud workloads demand. Our work agrees with this work, and extends its conclusions to modern analytical processing systems. Yasin et al. [@Yasin:2014] introduces the Top-Down Micro-architec- ture Analysis Methodology (TMAM), that is adopted by Intel VTune as general exploration. Sirin et al. [@Sirin:2017] presents an improvement on Yasin et al. [@Yasin:2014]’s methodology, which is adopted by Intel VTune in version 2018 and onwards.
Yasin et al. [@Yasin:2014b] analyzes cloud workloads. Sridharan and Patel [@Sridharan:2014] examines the evaluation of workloads on the popular data analysis language R over a commodity processor. Awan et al. [@Awan:2015; @Awan:2016] presents a micro-architectural analysis of Spark at the micro-architectural level. Kanev et. al. [@Kanev:2015] presents a profiling study of scale out workloads at the micro-architectural level. Our work complements these studies by presenting an analysis of modern analytical processing systems.
Conclusions {#section:conclusions}
===========
In this work, we evaluate the micro-architectural behavior of a breadth of OLAP systems from different categories of systems and execution models. We examine CPU cycles and memory bandwidth utilizations. The results show that, unlike traditional, commercial OLTP systems, traditional, commercial OLAP systems do not suffer from instruction cache misses. Nevertheless, they suffer from their large instruction footprint making them orders of magnitude slower than high performance OLAP engines.
High performance engines execute a tight instructions str- eam; however, they spend 25 to 82% of the CPU cycles on stalls regardless the workload being sequential- or random-access-heavy. Sequential-access-heavy workloads stress the memory bandwidth so high that hardware prefetchers fall behind resulting in high data cache stalls. Random access workloads suffer from long-latency data stalls consuming the majority of CPU cycles. Lastly, high performance OLAP engines underutilize the multi-core CPU or memory resources due to their disproportional compute and memory demands, showing that analytical processing engines should carefully schedule their compute and memory resources for efficient multi-core micro-architectural utilization.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using recent collider data, especially on the hadronic width the $Z^0$, we exclude axigluons in the currently allowed low-mass window, namely axigluons in the mass range $50\,GeV < M_A < 120\,GeV$. Combined with hadron collider data from di-jet production, axigluons with masses below roughly $1\,TeV$ are now completely excluded.'
---
by 60pt by -40pt by 50pt by -25pt by -25pt \_space[ plus 1pt minus 1pt]{} \_and\_a\_half\_space[ plus 1pt minus 1pt]{} \_spacesp[ plus 2pt minus 2pt]{}
[**PSU/TH/197\
April 1998** ]{}
1.5cm 1.0cm
M. A. Doncheski\
Department of Physics\
The Pennsylvania State University\
Mont Alto, PA 17237 USA\
0.1cm and\
0.1cm R. W. Robinett\
Department of Physics\
The Pennsylvania State University\
University Park, PA 16802 USA\
1.0cm
\_spacesp
Models which extend the standard color gauge group to $SU_3(L) \times SU_3(R)$ at high energies, so-called chiral color theories, include a wide variety of new particles beyond the standard model with the exact spectrum depending on the details of the theory. All such models, however, necessarily predict the existence of a massive, color-octet gauge boson, the axigluon $A$, which couples to quarks with an axial vector structure and the same strong interaction coupling strength as QCD. The original models [@glashow] explicitly connected the scale of the breakdown of chiral color to ordinary QCD with the electroweak scale and so even more specifically predicted that axigluons should have masses no larger than $M_A {\stackrel{<}{\!\ _{\sim}}}300\,GeV$. Early bounds from $\Upsilon$ decays [@cuypers_1] quickly found the limit $M_A {\stackrel{>}{\!\ _{\sim}}}25\,GeV$, while analyses of limits arising from axigluon contributions to the hadronic cross-section in $e^+e^-$ reactions (the $R$ ratio) [@cuypers_2] gave the limit $M_A > 50\,GeV$ (at $95\%$ confidence level) using then current PEP/PETRA data. Early suggestions [@glashow] that axigluons might be visible as an enhancement in the di-jet cross-section at hadron colliders were first used by Bagger, Schmidt, and King [@bagger] and then by the UA1 collaboration [@ua1] to exclude axigluons in the mass range $125 \, GeV < M_A < 310 \,GeV$. More recent searches for structure in the jet-jet invariant mass at the TEVATRON have led to dramatically enhanced limits, especially for heavy axigluon masses, with CDF data [@cdf] now excluding axigluons in the range $120\,GeV < M_A < 980\,GeV$. These are the limits which appear in the Particle Data Group discussion of bounds on the axigluon mass [@pdg] and we note that there is still a window of allowed masses in the range $50\,GeV < M_A < 120\,GeV$ which has not yet been excluded. This allowed window constitutes a large fraction of the range in masses (namely, up to $\sim 300\,GeV$) predicted by the original models which motivated the searches for axigluons and improved limits in this region would be useful in testing chiral color theories. In this note, we will use several rather different types of recent collider data to exclude axigluons in this mass region for the first time. Combined with the di-jet limits, this will imply that axigluons with masses in the entire range below roughly $1\,TeV$ will be definitively excluded. To the extent that chiral color models are constructed with the scale of color breaking directly tied to the electroweak scale, all such theories are also excluded.
The present authors have recently considered the effects of axigluons on the dominant ($90\%$ of the production cross-section) $q\overline{q} \rightarrow t\overline{t}$ subprocess contributing to the top quark production cross-section [@top_quark] and noted that the inclusion of axigluons with masses in the low-mass window more than doubles the tree-level cross-section. Even with the good agreement of the NLO QCD predictions with the CDF and D0 data, the inclusion of axigluons in the low-mass window is not yet definitively excluded due to the rather large experimental errors, but is definitely disfavored at the $(1\!-\!1.5)\sigma$ level. Given the large NLO QCD corrections to the tree-level $q\overline{q}$ process, one might well imagine that a complete NLO analysis, including the effects of low-mass axigluons, would make an unacceptably large contribution to the $t\overline{t}$ cross-section. It was pointed out some time ago, however, that relatively light axigluons can disturb the perturbative calculability of tree-level partial wave amplitudes [@partial_wave] for processes involving heavy quarks, so that top quarks coupled to sufficiently light axigluons would be strongly interacting. Extending work by Chanowitz, Furman, and Hinchliffe [@ultra_heavy] on the interactions of ultra-heavy fermions, one of the present authors [@partial_wave] found that the $J=0$ tree-level partial-wave amplitude for $Q\overline{Q} \rightarrow Q \overline{Q}$ (via $s$- and $t$-channel axigluon exchange) would become non-perturbative (i.e. $|a_0| > 1$) unless the axigluon mass satisfied the inequality $M_A > \sqrt{5\alpha_s/3} M_Q$. Using the measured value of the top quark mass and the apparently very good agreement between the NLO (perturbative!) QCD predictions for the top-quark production cross-sections and the experimentally observed value, this implies that $M_A {\stackrel{>}{\!\ _{\sim}}}72\, GeV$, which already improves the older $e^+e^-$ bound, pushing the limit closer to the $Z^0$ mass.
Other more specialized collider data might also be used to bound the axigluon mass. The associated production of an axigluon with a weak boson via the subprocess $q+ \overline{q}' \rightarrow W/Z + A$ is similar to that used for the production of the standard model Higgs boson via $p \overline{p} \rightarrow W/Z + X^0$ with $X^0 = H^0 \rightarrow b
\overline{b}$. Given the expected large branching ratio of the axigluon to $b\overline{b}$ final states ($BR(A \rightarrow b\overline{b})
= 1/5$ for $5$ active quark flavors) and the much larger coupling of the axigluon to the initial quarks, this channel might easily be used to extract limits on $M_A$.
The tree-level partonic cross-section for the largest subprocess, namely $q\overline{q}' \rightarrow W A$, is easily found to be $$\frac{d \hat{\sigma}}{d\hat{t}}
(q\overline{q}' \rightarrow W A)
= \frac{4\alpha_s}{9}
\left[\frac{G_F M_W^2}{\sqrt{2}}\right]
\frac{|V_{qq'}|^2}{\hat{u}\hat{t}\hat{s}^2}
\left[\hat{u}^2 + \hat{t}^2 + 2\hat{s}(M_W^2 + M_A^2)
-
\frac{M_A^2 M_W^2(\hat{u}^2+\hat{t}^2)}{\hat{u}\hat{t}} \right]$$ and at TEVATRON energies we find the following values for total cross-sections times branching ratios (assuming $BR(A \rightarrow
b\overline{b}) = 0.2$) for various axigluon masses in the allowed window:
$$\begin{array}{cccc}
M_A & \sigma \cdot BR & M_A & \sigma \cdot BR \\ \hline
50\,GeV & 51\,pb & 90 \,GeV & 26\,pb \\
60\,GeV & 46\,pb & 100\,GeV & 17\,pb \\
70\,GeV & 40\, pb & 110\,GeV & 16\,pb \\
80\,GeV & 28\, pb & 120\,GeV & 14\,pb \\
\end{array}$$
Bhat [@bhat] has recently surveyed limits for many new physics searches at the TEVATRON and presented preliminary data for the production cross-section times branching ratio for $p\overline{p} \rightarrow W + X^0$ with $X^0 = H^0 \rightarrow b\overline{b}$ from CDF (which uses an $l\nu$ tag for the $W$ bosons). The limits cover the mass range $75\,GeV {\stackrel{<}{\!\ _{\sim}}}M_H {\stackrel{<}{\!\ _{\sim}}}125 \,GeV$ and are of order $\sigma \cdot BR \approx 15-20\,pb$. Given the estimates above, a complete analysis of this process for the axigluon analog could likely exclude axigluons up to $M_A \approx 80-90 \,GeV$, with $M_A > 70\,GeV$ a seemingly safe estimate of the current bound possible.
Given the huge statistical sample of $Z^0$ hadronic decay events at LEP, it is perhaps most natural to extend the analyses of Ref. [@cuypers_2] using LEP data. Following the same strategy as employed previously, we compare the value of $\alpha_s$ extracted from low-energy experiments (which is then run up to $M_Z$) with the value extracted from the hadronic width of the $Z^0$ at the pole. (The improved limits on $M_A$ mentioned above imply that any changes to the running of $\alpha_s$ due to axigluon effects will be small.) The inclusion of real and virtual axigluons increases the hadronic decay rate (or $R$ value in $e^+e^-$ collisions) by a factor of $(1 + \alpha_s(\sqrt{s})f(\sqrt{s}/M_A)/\pi
+ {\cal O}(\alpha_s^2))$ where the function $f(\sqrt{s}/M_A)$ is derived in Ref. [@cuypers_2]. The Particle Data Group perturbative QCD analysis [@pdg_qcd] quotes a value of $\alpha_s$ derived from low-energy data (such as deep-inelastic scattering (excluding HERA), $\tau$ decay, $\Upsilon$ width, and lattice calculations), namely, $\alpha_s^{(LE)}(M_Z)
= 0.118 \pm 0.004$. The value of $\alpha_s$ extracted from the ratio of hadronic to leptonic decay widths of the $Z^0$ ($\Gamma_h/\Gamma_{\mu}
= 20.788 \pm 0.0032$, which probes the same QCD corrections as the $R$ value at lower energies) is $\alpha_s^{(HE)}(M_Z) = 0.123
\pm 0.004 \pm 0.002$ and the evaluation includes the effect of ordinary QCD (gluonic) corrections up to ${\cal O}(\alpha_s^2)$. Using bounds on the possible difference between these two values, after combining errors, we find that the contribution from axigluons, due to the $f(\sqrt{s}/M_A)$ term, is bounded by $0.042 \pm 0.05 \geq f(M_Z/M_A)$ which we take to imply roughly that $f(M_Z/M_A) < 0.092 \,(0.142)$ at the $1\sigma\,(2\sigma)$ or $65\% \,(95\%)$ confidence level. Using the expression for $f(z)$ in Ref. [@cuypers_2], we find that this corresponds to $M_Z/M_A < 0.16 \,(0.25)$ or $M_A > 6.2M_Z \,(4M_Z)$ or $M_A > 570\,GeV \,(365\,GeV)$ at $65\% \,(95\%)$ confidence level. While this is a very simple estimate, given the substantial agreement of the various $\alpha_s$ measurements, axigluons as light as $120\,GeV$ are obviously excluded. Measurements of similar quantities at higher energies at LEP (starting at $\sqrt{s} \approx 130\!-\!40\,GeV$ [@lep_2] and beyond the $W^{+}W^{-}$ threshold), but with much lower statistics, do not improve on these limits. Nonetheless, this analysis easily excludes axigluons in the low-mass window once and for all. Chiral-color models and axigluons, if they have any relevance in nature, can only appear as new physics beyond the $TeV$ scale.
One of us (M.A.D) acknowledges the support of Penn State University through a Research Development Grant (RDG). Both authors thank the organizers of the Pheno - CTEQ 98 meeting, where this work was begun, for their hospitality.
[99]{} P. H. Frampton and S. L. Glashow, Phys. Lett. [**190B**]{}, 157 (1987); Phys. Rev. Lett. [**58**]{}, 2168 (1987). F. Cuypers and P. H. Frampton, Phys. Rev. Lett. [**60**]{}, 1237 (1988) ; M. A. Doncheski, H. Grotch, and R. W. Robinett, Phys. Lett. [**206B**]{}, 137 (1988); M. A. Doncheski, H. Grotch, and R. W. Robinett, Phys. Rev. [**D38**]{}, 412 (1988). F. Cuypers and P. H. Frampton, Phys. Rev. Lett. [**63**]{}, 125 (1989); A. Falk, Phys., Lett. [**B230**]{}, 119 (1989); F. Cuypers, A. F. Falk, and P. H. Frampton, Phys. Lett. [**B259**]{}, 173 (1991). J. Bagger, C. Schmidt, and S. King, Phys. Rev. [**D37**]{}, 1188 (1988). C. Albajar [*et al.*]{} (UA1 Collaboration), Phys. Lett. [**209B**]{}, 127 (1988). F. Abe [*et al.*]{} (CDF Collaboration), Phys. Rev. [**D41**]{}, 1722 (1990); F. Abe [*et al.*]{} (CDF Collaboration), Phys. Rev. Lett. [**74**]{}, 3538 (1995); F. Abe [*et al.*]{} (CDF Collaboration) Phys. Rev. [**D55**]{}, R5263 (1997). Review of Particle Physics, Particle Data Group, Phys. Rev. [**54**]{}, 1 (1996); see page 236 for the detailed review. The most recent web version ([http://pdg.lbl.gov/]{}) has the same limits. M. A. Doncheski and R. W. Robinett, Phys. Lett. [**B412**]{}, 91 (1997). M. S. Chanowitz, M. A. Furman, and I. Hinchliffe, Phys. Lett. [**78B**]{}, 285 (1978); Nucl. Phys. [**B153**]{}, 402 (1979). R. W. Robinett, Phys. Rev. [**D39**]{}, 834 (1989). P. Bhat, invited talk at PHENO-CTEQ-98: Frontiers of Phenomenology from Non-Perturbative QCD to New Physics, Madison, March 1998. See Ref. [@pdg], pp. 77-83. M. Acciarri (L3 Collaboration), Phys. Lett. [**B370**]{}, 195 (1996); [*ibid*]{} [**B371**]{}, 137 (1996).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We survey a variety of possible explications of the term “Individual Risk.” These in turn are based on a variety of interpretations of “Probability,” including Classical, Enumerative, Frequency, Formal, Metaphysical, Personal, Propensity, Chance and Logical conceptions of Probability, which we review and compare. We distinguish between “groupist” and “individualist” understandings of Probability, and explore both “group to individual” (G2i) and “individual to group” (i2G) approaches to characterising Individual Risk. Although in the end that concept remains subtle and elusive, some pragmatic suggestions for progress are made.'
author:
- |
Philip Dawid\
University of Cambridge, UK
bibliography:
- 'strings.bib'
- 'StatViol.bib'
- 'indrisk.bib'
- 'causal.bib'
- 'dna.bib'
- 'allclean.bib'
- 'afms.bib'
- 'apdpubs.bib'
title: On Individual Risk
---
Introduction {#sec:intro}
============
“Probability” and “Risk” are subtle and ambiguous concepts, subject to a wide range of understandings and interpretations. Major differences in the interpretation of Probability underlie the fundamental Frequentist/Bayesian schism in modern Statistical Science. However, these terms, or their synonyms, are also in widespread general use, where lack of appreciation of these subtleties can lead to ambiguity, confusion, and outright nonsense. At the very least, such usages deserve careful attention to determine whether and when they are meaningful, and if so in how many different ways. The focus of this article will be on the concept of “Individual Risk,” which I shall subject to just such deep analysis.
To set the scene, presents some examples displaying a variety of disparate usages of the concept of Individual Risk. These, in turn, are predicated on a variety of different understandings of the concept of Probability. I survey these in , returning to discuss the examples in their light in . Section \[sec:expert\] describes the concept of “expert assignment,” which is is common to a number of understandings of Risk. In and I consider various aspects of “group to individual” (G2i) inference—the attempt to make sense of Individual Risk when Probability is understood as a group phenomenon—and conclude that this can not be done in an entirely satisfactory way. So reverses this process and considers “individual to group” (i2G) inference, where we take Individual Risk as a fundamental concept, and explore what that implies about the behaviour of group frequencies. The i2G approach appears to lead to essentially unique individual risk values; but these can not be considered absolute, but are relative to a specified information base, and moreover are typically not computable. In I take stock of the overall picture. I conclude that the concept of “Individual Risk” remains highly problematic at a deep philosophical level, but that does not preclude our making some pragmatically valuable use of that concept—so long as we are aware of the various pitfalls that may lie in our path.
Examples {#sec:ex}
========
We start with a ménagerie of examples of “individual risk” in public discourse.
\[ex:weather\] A weather forecaster appears on television every night and issues a statement of the form “The probability of precipitation tomorrow is $30\%$” (where the quoted probability will of course vary from day to day). Different forecasters issue different probabilities.
\[ex:arai\] There has been much recent interest concerning the use of Actual Risk Assessment Instruments (ARAIs): statistical procedures for assessing “the risk” of an individual becoming violent [@BSL]. Thus a typical output of the Classification of Violence Risk (COVR) software program, that can be used to inform diagnostic testimony in civil commitment cases, might be: “The likelihood that XXX will commit a violent act toward another person in the next several months is estimated to be between 20 and 32 percent, with a best estimate of 26 percent.”
\[ex:aharoni\] @aharoni:2013 tested a group of released adult offenders on a go/no-go task using fMRI, and examined the relation between task-related activity in the anterior cingulate cortex (ACC) and subsequent rearrest (over four years), allowing for a variety of other risk factors. They found a significant relationship between ACC activation on the go/no-go task and subsequent rearrest; whereas subjects with high ACC activity had a 31% chance of rearrest, subjects with low ACC activity had a 52% chance. They conclude: “These results suggest a potential neurocognitive biomarker for persistent antisocial behavior.”
A newly released offender has low ACC activity: how should we judge his chance of rearrest?
\[ex:jolie\] Writing in the [New York Times (May 14 2013)](http://www.nytimes.com/2013/05/14/opinion/my-medical-choice.html) about her decision to have a preventive double mastectomy, the actress Angelina Jolie said: “I carry a faulty gene, BRCA1, which sharply increases my risk of developing breast cancer and ovarian cancer. My doctors estimated that I had an 87 percent risk of breast cancer and a 50 percent risk of ovarian cancer, although the risk is different in the case of each woman.”
\[ex:climate\] The [Fifth Assessment Report of the Intergovernmental Panel on Climate Change](http://www.ipcc.ch/news_and_events/docs/ar5/ar5_wg1_headlines.pdf), issued in September 2013, contains the statements “It is [*extremely likely*]{} that human influence has been the dominant cause of the observed warming since the mid-20th century,” “It is [*virtually certain*]{} that the upper ocean ($0$–$700$ m) warmed from 1971 to 2010, and it [*likely*]{} warmed between the 1870s and 1971,” and “It is [*very likely*]{} that the Arctic sea ice cover will continue to shrink and thin and that Northern Hemisphere spring snow cover will decrease during the 21st century” (their italics). [It is explained that](http://www.climatechange2013.org/images/uploads/WGIAR5-SPM_Approved27Sep2013.pdf) [*virtually certain*]{} is equivalent to a probability of at least $99\%$, [*extremely likely*]{}, at least $95\%$, [*very likely*]{}, at least $90\%$, and [*likely*]{}, at least $66\%$.
\[ex:obama\] On 4 May 2011, three days after he announced that American troops had killed Osama bin Laden in Pakistan, US President Barack Obama said in an [interview](http://www.cbsnews.com/8301-504803_162-20060530-10391709.html) with “60 Minutes” correspondent Steve Kroft:
> At the end of the day, this was still a 55/45 situation. I mean, we could not say definitively that bin Laden was there.
\[ex:civil\] In a civil court case, the judgment might be expressed as: This is typically interpreted as “with probability exceeding $50\%$.”
In all the above examples we can ask: How were the quoted probabilities interpreted? How might they be interpreted? And how might the quality of such probability forecasts be measured?
Interpretations of Probability {#sec:interp}
==============================
We have already remarked that the concept of “Probability” is a hotly contested philosophical issue.[^1] Even the many who have no patience for such philosophising are usually in thrall to some implicit philosophical conception, which shapes their approach and understanding, and their often fruitless arguments with others who (whether or not so recognised) have a different understanding.
One axis along which the different theories and conceptions of Probability can be laid out—and which is particularly germane to our present purpose—is whether they regard Probability as fundamentally an attribute of groups, or of individuals. I will refer to these as, respectively, “groupist” and “individualist” theories.
Below, in a necessarily abbreviated and admittedly idiosyncratic account, I outline some of the different conceptions of Probability, and hope to bring out their relationships, similarities and differences. For a fuller discussion see @hajek:sep [@galavotti:2014].
Classical Probability {#sec:classprob}
---------------------
If you studied any Probability at school, it will have focused on the behaviour of unbiased coins, well-shuffled packs of cards, perfectly balanced roulette wheels, , In short, an excellent training for a life misspent in the Casino. This is the ambit of [ *Classical Probability*]{}.
The underlying conception is that we have a number of [*elementary outcomes*]{} of an experiment, exactly one of which will actually be realised when the experiment is performed. For example, there are $N=53,644,737,765,488,792,839,237,440,000$ ways in which the cards at Bridge can be distributed among 4 players, and just one of these ways will materialise when the cards are dealt. Any event of interest, for example “North holds 3 aces,” can be represented by the set of all the elementary outcomes for which it is the case; and the number $n$ of these, divided by the total number $N$ of all elementary outcomes, is taken as the measure of the probability of the event in question. The mathematics of Classical Probability is thus really a branch of Combinatorial Analysis, the far from trivial mathematical theory of counting.
Since the focus is on the specific outcome of (say) a particular deal of cards or roll of a die, this classical conception is individualist. But questions as to the interpretation of the “probabilities” computed rarely raise their heads. If they do, it would typically be assumed (or regarded as a precondition for the applicability of the theory) that each of the elementary outcomes is just as likely as any other to become the realised outcome. However, in the absence of any independent understanding of the meaning of “likely,” such an attempt at interpretation courts logical circularity. Furthermore, paradoxes arise when there is more than one natural way to describe what the elementary events should be. If we toss two coins, we could either form 3 elementary events: “0 heads,” “1 head,” “2 heads”; or, taking order into account, 4 elementary events: “tail tail,” “tail head,” “head tail,” “head head.” There is nothing within the theory to say we should prefer one choice over the other. Also, it is problematic to extend the classical conception to cope with an infinite number of events: for example, to describe a “random positive integer,” or a line intersecting a given circle “at random.”
The principal modern application of Classical Probability is to situations—for example, clinical trials—where randomisation is required to ensure fair allocation of treatments to individuals. This can be effected by tossing a “fair coin” (or by simulating such tosses on a computer).
Enumerative Probability {#sec:enumprob}
-----------------------
What I [^2] here term [*Enumerative Probability*]{} can also be regarded as an exercise in counting. Only now, instead of counting elementary outcomes of an experiment, we consider a finite collection of individuals (of any nature), on which we can measure one or more pre-existing attributes.
Thus consider a set of individuals, $I_1, I_2, \ldots, I_N$, and an attribute $E$ that a generic individual $I$ may or may not possess: “smoker.” A specific individual $I_k$ can be classified according to whether or not its “instance,” $E_k$, of that attribute is present—, whether or not $I_k$ is a smoker. If we knew this for every individual, we could compute the [*relative frequency*]{} with which $E$ occurs in the set, which is just the number of individuals having attribute $E$, divided by the total number $N$ of individuals. This relative frequency is the “enumerative probability” of $E$ in the specified set. Clearly this is a “groupist” conception of probability.
Of course, the value of such an enumerative probability will depend on the set of individuals considered, as well as what attributes they possess. We will often care about the constitution of some specific large population (say, the population of the United Kingdom), but have data only for a small sample. Understanding the relationship between the known frequencies in the sample and the unknown frequencies in the whole population is the focus of the theory and methodology of sample surveys. This typically requires the use of randomisation to select the sample—so relying, in part, on the Classical conception.
Although based on similar constructions in terms of counting and combinatorial analysis, Classical and Enumerative Probability are quite different in their scope and application. Thus suppose we ask: What is “the probability” that a new-born child in the United Kingdom will be a boy? Using the Classical interpretation, with just two outcomes, boy or girl, the answer would be $1/2 = 0.5$. But in the UK population about $53\%$ of live births are male, so the associated Enumerative Probability is $0.53$.
Frequency Probability {#sec:freqprob}
---------------------
“Frequency probability” can be thought of as enumerative probability stretched to its limits: instead of a finite set of individuals, we consider an infinite set.
There are two immediate problems with this:
1. \[it:prob1\] Infinite populations do not exist in the real world, so any such set is an idealisation.
2. \[it:prob2\] Notwithstanding that the great statistician Sir Ronald Fisher manipulated “proportions in an infinite population” with great abandon and to generally good effect, this is not a well-defined mathematical concept.
These problems are to some extent resolved in the usual scenario to which this concept of probability is attached: that of [*repeated trials under identical conditions*]{}. The archetypical example is that of tosses of a possibly biased coin, supposed to be repeated indefinitely. The “individuals” are now the individual tosses, and the generic outcome is (say) heads (H), having instances of the form “toss $i$ results in heads.” It is important that the “individuals” are arranged in a definite sequence, for example time order. Then for any finite integer $N$ we can restrict attention to the sequence of all tosses from toss 1 up to toss $N$, and form the associated enumerative probability, $f_N$ say, of heads: the relative frequency of heads in this finite set. We next consider how $f_N$ behaves as we increase the total number $N$ without limit. If $f_N$ approaches closer and closer to some mathematical limit $p$—the “limiting relative frequency” of heads—then that limiting value may be termed the “frequency probability” of heads in the sequence.
Of course we can never observe infinitely many tosses, so even the existence of the limit must remain an assumption: one that is, however, given some empirical support by the observed behaviour of real-world frequencies in such repeated trial situations. But even when we can happily believe that the limit $p$ exists, we will never have the infinite amount of data that would be needed to determine its value precisely. Much of the enterprise of [*statistical inference*]{} addresses the subtle relationship between actual frequencies, observed in finite sequence of trials, and the ideal “frequency probabilities” that inhabit infinity.
Richard @vonmises attempted to build a sophisticated mathematical theory of probability built upon the above frequency conception as its primitive element, but that is now largely of historical interest.
Formal Probability {#sec:statprob}
------------------
A prime task for an applied statistician is to build a “statistical model” of a phenomenon of interest. That phenomenon could be fairly simple, such as recording the outcomes of a sequence of 10 tosses of a coin; or much more sophisticated, such as a description of the earth’s temperature as it varies over space and time. Any such model will have symbols representing particular outcomes and quantities (, the result of the 10th toss of the coin, or the recorded temperature at the Greenwich Royal Observatory at noon GMT on 24 September 2020), and will model these (jointly with all the other outcomes and quantities under consideration) as “random variables,” having a joint probability distribution. Typically the probabilities figuring in this distribution are treated as unknown, and statistical data-analysis is conducted to learn something about them. There does not seem to be a generally accepted terminology for these probabilities figuring in a statistical model: I shall term them “Formal Probabilities.”
Since the focus of statistical attention is on learning these formal probabilities, I find it remarkable how little discussion is to be found as to their meaning and interpretation. In particular, is Formal Probability an individualist or a groupist conception? The elementary building blocks of the statistical model are assignments of probabilities to specific events, so [*prima facie*]{} it looks like an individualist conception. One possible individualist interpretation of a formal probability value is as a propensity—see below—although this raises problems of its own.
In any case, many users of statistical models would not be happy to accept an individualist interpretation of their formal probabilities, and prefer a groupist account of them. Thus consider the simple case of coin-tossing. The usual statistical model for this assigns some common (though typically unknown) formal probability $p$ to each event $E_k$: “the $k$th toss $I_k$ results in heads,” for every $k$; and further models all these tosses, as $k$ varies, as independent. An application of Bernoulli’s Law of Large Numbers shows that this implies that, with probability 1, the limiting relative frequency of heads in the sequence will exist and have value $p$. And many statisticians would claim that—notwithstanding that the formal probabilities are attached to individual tosses—this groupist Frequency Probability interpretation is the [*only*]{} meaningful content of the (superfically individualist) formal “Bernoulli model.”
However, such a ploy is far from straightforward for more complex models, where there may be no natural way of embedding an individual event in a sequence of similar events. For example, if we construct a spatio-temporal statistical model of the weather, the model might well contain a parameter that measures the correlation between the temperatures in Greenwich and New York at noon GMT on 24 September 2020. But just how is such a correlation, between two “one-off” quantities, to be interpreted? Does it have any real-world counterpart?
One possible way of taking the above “Law of Large Numbers” approach for the Bernoulli model and extending it to more general models [@apd:synthesis] is by identifying those events that are assigned probability 1 by the assumed model, and asserting that the [ *only*]{} valid interpretation of the model lies in its claim that these events will occur. But this may be seen as throwing too much away.
Because there is little discussion and no real shared understanding of the interpretation of Formal Probability, pointless disagreements can spring up as to the appropriateness of a statistical model of some phenomenon. An example of practical importance for forensic DNA profiling arises in population genetics, where there has been disagreement as to whether the genes of distinct individuals within the same subpopulation should be modelled as independent, as claimed by @Roeder, or correlated, as claimed by @for/smith. However, without a shared understanding of what (if anything) the correlation parameter in the formal model relates to in the real world, this is a pointless argument. It was pointed out by @apd:isba that (like blind men’s disparate understandings of the same elephant) both positions can be incorporated within the same hierarchical statistical model, where the formal correlation can come or go, according to what it is taken as conditioned on. So while the two approaches superficially appear at odds, at a deeper level they are in agreement.
### Metaphysical Probability {#sec:meta}
There are those who, insistent on having some sort of frequency-based foundation for probability, would attempt to interpret a “one-off” probability as an average over repetitions of the whole underlying phenomenon. For example, the correlation between the temperatures in Greenwich and New York at noon GMT on 24 September 2020 would be taken to refer to an average over—necessarily hypothetical—independent repetitions of the whole past and future development of weather on Earth. This conception appears close to the “many worlds” interpretation of quantum probabilities currently popular with some physicists. I confess I find it closer to science fiction than to real science—“metaphysical,” not “physical.” At any rate, since we can never observe beyond the single universe we in fact inhabit, we can not make any practical use of such hypothetical repetitions.
That said, this way of thinking can be useful, purely as an analogy, in helping people internalise a probability value (whatever its provenance or philosophical back-story). Thus one of the options at\
[ http://understandinguncertainty.org/files/animations/RiskDisplay1/RiskDisplay.html](http://understandinguncertainty.org/files/animations/RiskDisplay1/RiskDisplay.html)\
for visualising a risk value is a graphic containing a number of symbols for your “possible futures,” marked as red to indicate that the event occurs in that future, green that it does not. The relevant probability is then represented by the proportion of these symbols that are red. It is not necessary to take the “possible futures” story seriously in order to find this enumerative representation of a probability value helpful. The same purely psychological conceit explains the misleading name given to the so-called “frequentist” approach to statistical inference. This involves computing the probabilistic properties of a suggested statistical procedure. For example, a test of a null hypothesis at significance level $5\%$ is constructed to have the property that probability of deciding to reject the hypothesis, under the assumption that it is in fact correct, is at most $0.05$. Fisher’s explication of such a test was that, when the it results in the decision to reject, either the null hypothesis was indeed false, or else an event of small probability has happened—and since we can largely discount the latter alternative, we have evidence for the former. Although for this purpose it would be perfectly satisfactory to interpret the probability value $0.05$ as an “individual chance,” relating solely to the current specific application of the procedure, it is almost universally expressed metaphysically, by a phrase such as “over many hypothetical repetitions of the same procedure, when the null hypothesis is true, it will rejected at most $5\%$ of the time.”
Personal Probability {#sec:persprob}
--------------------
[*Personal*]{} (often, though less appropriately, termed [*Subjective*]{}) [*Probability*]{} is very much an individualist approach. Indeed, not only does it associate a probability value, say $p$, with an individual event, say $E$, it also associates it with the individual, say “You,” who is making the assignments, as well as (explicitly or implicitly) with the information, say $H$, available to You when You make the assignment. Thus if You are a weather forecaster, You might assess Your probability of rain tomorrow, given Your knowledge of historical weather to date, at $40\%$. Another forecaster would probably make a different personal assessment, as would You if You had different information (perhaps additional output from a meteorological computer system) or were predicting for a different day.
What is the meaning of Your $40\%$? The standard view is that it simply describes the odds at which You would be willing to bet on “rain tomorrow.” If Classical Probability is modelled on the casino, then Personal Probability is modelled on the race-track.
If You really really had to bet, You would have to come up with a specific numerical value for Your personal probability. Thus to a Personalist there is no such thing as an “unknown probability,” and it appears that these should not feature in any probability model You might build. There is however an interesting and instructive relationship between such Personalist models and the statistical models, as described in , that do feature unknown formal probability values—for more on this see below.
Propensity and chance {#sec:propensity}
---------------------
The interpretation of Probability as a “Propensity” was championed by @popper:propen, and is still much discussed by philosophers—though hardly at all by statisticians. The overall idea is that a particular proposed coin-toss (under specified circumstances) has a certain (typically unknown) “propensity” to yields heads, if it were to be conducted—just as a particular lump of arsenic has a propensity to cause death if it were to be ingested. This is clearly an individualist (though non-Personalist) conception of Probability, but there is little guidance available as to how a propensity probability is to be understood, or how its value might be assessed, except by reference to some “groupist” frequency interpretation. Indeed, some versions of the propensity account [@gillies:bjps2000] take them as referring directly to the behaviour of repetitive sequences, rather than to that of individual events.
Another individualist term that is very close in spirit to “propensity” is “(objective) chance.” The “Principal Principle” of @lewis:80, while not defining chance, relates it to personal probability by requiring that, if You learn (don’t ask how!) that the chance of an event $A$ is (say) $0.6$, and nothing else, then Your personal probability of $A$ should be updated to be $0.6$. And further conditions may be required: for example, that this assessment would be unaffected if You learned of any other “admissible” event, where “admissible” might mean “prior to $A$,” or “independent of $A$ in their joint chance distribution.” Such conditions are difficult to make precise and convincing [@pettigrew:2012]. In any case, the Principal Principle gives no guidance on how to compute or estimate an individual chance.
Logical Probability {#sec:logprob}
-------------------
Yet another non-Personalist individualist conception of Probability is Logical Probability, associated with @jeffreys:39 [@carnap] and others. This is similar to the propensity/chance account in considering that there is an “objective” probability value associated with a given outcome, but differs from those in emphasising the relativity of that value to the information upon which it is premissed. Once again there would seem to be no routine and unambiguous way of computing such logical probability values.
Examples Revisited {#sec:revisit}
==================
We here revisit the examples of in the light of some of the above discussion. My own comments are interpolated in \[…\].
\[ex:weather2\] @gigerenzer:riskanal randomly surveyed pedestrians in five metropolises located in countries that have had different degrees of exposure to probabilistic forecasts: Amsterdam, Athens, Berlin, Milan, and New York. Participants were told to imagine that the weather forecast, based on today’s weather constellation, predicts “There is a $30\%$ chance of rain tomorrow,” and to explain what they understood by that.
> Several people in New York and Berlin thought that the rain probability statement means “3 out of 10 meteorologists believe it will rain” \[a form of enumerative probability?\]. A woman in Berlin said, “Thirty percent means that if you look up to the sky and see 100 clouds, then 30 of them are black” \[a different form of enumerative probability\]. Participants in Amsterdam seemed the most inclined to interpret the probability in terms of the amount of rain. “It’s not about time, it indicates the amount of rain that will fall,” explained a young woman in Amsterdam. Some people seemed to intuitively grasp the essence of the “days” interpretation, albeit in imaginative ways. For instance, a young woman in Athens in hippie attire responded, “If we had 100 lives, it would rain in 30 of these tomorrow” \[a metaphysical interpretation?\]. One of the few participants who pointed out the conflict between various interpretations observed, “A probability is only about whether or not there is rain, but does not say anything about the time and region,” while another said, “It’s only the probability that it rains at all, but not about how much” \[two individualist views\]. Many participants acknowledged that, despite a feeling of knowing, they were incapable of explaining what a probability of rain means.
>
> \[According to the authors of the paper, the standard meteorological interpretation is: when the weather conditions are like today, in 3 out of 10 cases there will be (at least a trace of) rain the next day.\]
\[ex:arai2\] There has been heated recent debate centred on the construction and validity of confidence intervals for the “individual risks” output by an ARAI. This has been initiated and promoted by @HartMichieCooke [@CookeMichie; @HartCooke], whose analysis has had a strong influence (see @BSL), but has been widely criticised for serious technical statistical errors and confusions [@HarrisRiceQuinsey; @HansonHoward; @imreydawid; @mossman:2014]. However, to date that debate has largely sidestepped the crucial question of the meaning of such an individual risk.
@HartMichieCooke make the following argument in an attempt to distinguish between group and individual risks:
> To illustrate our use of Wilson’s method for determining group and individual margins of error, let us take an example. Suppose that Dealer, from an ordinary deck of cards, deals one to Player. If the card is a diamond, Player loses; but if the card is one of the other three suits, Player wins. After each deal, Dealer replaces the card and shuffles the deck. If Dealer and Player play 10 000 times, Player should be expected to win $75\%$ of the time. Because the sample is so large, the margin of error for this group estimate is very small, with a $95\%$ CI of $74$–$76\%$ according to Wilson’s method. Put simply, Player can be $95\%$ certain that he will win between $74$ and $76\%$ of the time. However, as the number of plays decreases, the margin of error gets larger. If Dealer and Player play 1000 times, Player still should expect to win $75\%$ of the time, but the $95\%$ CI increases to $72$–$78\%$; if they play only 100 times, the $95\%$ CI increases to $66$–$82\%$. Finally, suppose we want to estimate the individual margin of error. For a single deal, the estimated probability of a win is still $75\%$ but the $95\%$ CI is $12$–$99\%$. The simplest interpretation of this result is that Player cannot be highly confident that he will win—or lose—on a given deal
\[On 10 000, 1000 or 100 deals, the actual success rate will vary randomly about its target value of $75\%$, and the “confidence intervals” (CIs) described for these cases are intended to give some idea of the possible extent of that variation. But on a single deal the actual success rate can only be $0$ (which will be the case with probability $25\%$) or $100\%$ (with probability $75\%$). This purely binary variation is not usefully described by any “confidence interval,” let alone the above one of $12$–$99\%$, based on a misconceived application of Wilson’s method. See @imreydawid for further deconstruction of the spurious philosophical and mathematical arguments presented by these authors.\]
\[ex:aharoni2\] @neurolaw point out that the results given by @aharoni:2013 provide inflated estimates of the predictive accuracy of the model when generalizing to new individuals: their reanalysis using a cross-validation approach found that addition of brain activation to the predictive model improves predictive accuracy by less than $3\%$ compared to a baseline model without activation. The alternative bootstrap-based reanalysis in @aharoni:2014 likewise finds the original estimates to have been over-inflated, though their correction is smaller. The message is that it can be extremely difficult to obtain reliable estimates of reoffending rates, and any causal attribution, to ACC activity, would be even more precarious.
All these objections aside, suppose we had reliable estimates of the statistical parameters in a well-fitting statistical model, from which we could compute a “reoffending rate” for this new individual. Is it appropriate to regard such a formal probability value, based on a statistical analysis of group data, as the appropriate expression of his individual risk?
\[ex:jolie2\] Angelina Jolie’s statement sounds very much like an “individualist” interpretation of the figure “87 percent risk of breast cancer,” especially in the light of “the risk is different in the case of each woman.” However it would appear that this figure is taken from the website of [Myriad Genetics](https://www.myriad.com/products/bracanalysis/), on which it says “People with a mutation in either the BRCA1 or BRCA2 gene have risks of up to 87% for developing breast cancer by age 70.” This is clearly an enumerative probability. So was Angelina (or her doctor) right to interpret it as her own individual risk?
\[ex:climate2\] The events and scenarios considered by the IPCC are fundamentally “one-off,” so any attempt at interpreting the quoted probabilities must be individualistic. But of what nature? (It was suggested in a radio commentary that a probability of $95\%$ means that $95\%$ of scientists agree with the statement).
\[ex:obama2\] It would be very interesting to know just how the President conceived of this $55\%$ probability of success for a one-off event. We might also ask: Was that probability assessment justified (in any sense) by the turn-out of events?
\[ex:civil2\] While there has been much discussion as to relationship between “legal” and “mathematical” probabilities, this is apparently an assertion about uncertainty in an individual case. Such assessments might best be construed as personal probabilities, though they are notoriously subject to bias and volatility [@fox2002].
Risk and Expert Assignments {#sec:expert}
===========================
Our focus in this article is on the concept of the single-case “individual risk,” and we shall be exploring how this is or could be interpreted from the point of view of the various different conceptions of probability outlined above. One theme common to a number of those conceptions is that of risk as an “expert assignment” [@gaifman:1988]. This means that, if You start from a position of ignorance, and then somehow learn (only) that “the risk” (however understood) of outcome $A$ is $p$, then $p$ should be the measure of Your new uncertainty about $A$.
Personal Probability {#sec:persexp}
--------------------
Consider this first from the Personalist point of view. Suppose You learn the personal probability $p$, for event $A$, of an individual $E$ (the “expert”) who started out with exactly the same overall personal probability distribution as You did, but has observed more things, so altering her uncertainty. By learning $p$ You are effectively learning, indirectly, all the relevant extra data that $E$ has brought to bear on her uncertainty for $A$, so You too should now assign personal probability $p$ to event $A$. Note that this definition of expert is itself a personal one: an expert for You need not be an expert for some one else with different opinions or knowledge.
This example shows that the property of being an expert assignment is quite weak, since there could be a number of different experts who have observed different things and so have different updated personal probabilities. Whichever one of these You learn, You should now use that value as Your own.
Regardless of which expert You are considering consulting (but have not yet consulted), Your current expectation of her expert probability $p$ will be Your own, unupdated, personal probability of $A$. In particular, mere knowledge of the existence of an expert has no effect on the odds You should currently be willing to offer on the outcome of $A$.
What if You learn the personal probabilities of several different experts? It is far from straightforward to combine these to produce Your own revised probability [@ddm:1995], since this must depend on the extent to which the experts share common information, and would typically differ from each individual expert assignment—even if these were all identical.
Chance {#sec:chanceexp}
------
As for “objective chance,” the Principal Principle makes explicit that (whatever it may be) it should act as an expert assignment, and moreover that this should hold for every personalist—it is a “universal expert.” But while this constrains what we can take objective chance to be, it is far from being a characterisation.
Frequency Probability {#sec:freqexp}
---------------------
The relationship between frequencies and expert assignments is considered in detail in below.
Group to Individual Inference for Repeated Trials {#sec:g2i}
=================================================
As we have seen, some conceptions of Probability are fundamentally “groupist,” and others fundamentally “individualist.” That does not mean that a groupist approach has nothing to say about individual probabilities, nor that an individualist approach can not address group issues. But the journey between these two extremes, in either direction, can be a tricky one. Our aim in this article is to explore this journey, with special emphasis on the interpretation of individual probabilities, or “risks.” In this Section, we consider the “group to individual” (G2i) journey; the opposite (i2G) direction will be examined in below.
We start by considering a simple archetypical example. A coin is to be tossed repeatedly. What is “the probability” that it will land heads (H) up (event $E_1$) on the first toss ($I_1$)? We shall consider how various conceptions of Probability might approach this question.
Frequency probability {#sec:freqprob2}
---------------------
The frequency approach apparently has nothing to say about the first toss $I_1$. Suppose however that (very) lengthy experimentation with this coin has shown that the limiting relative frequency of heads, over infinitely many tosses, is $0.3$. Can we treat that value $0.3$ as representing uncertainty about the specific outcome $E_1$? Put otherwise, can we treat the limiting relative frequency of heads as a (universal) expert assignment for the event of heads on the first toss?
While there is no specific warrant for this move within the theory itself, it would generally be agreed that we are justified in doing so if all the tosses of the coin (including toss $I_1$) can be regarded as
> “repeated trials of the same phenomenon under identical conditions.”
We shall not attempt a close explication of the various terms in this description, but note that there is a basic assumption of [ *identity of all relevant characteristics*]{} of the different tosses.
Personal probability {#sec:persprob2}
--------------------
By contrast, the Personalist You would be perfectly willing to bet on whether or not the next toss will land heads up, even without knowing how other tosses of the coin may turn out. But how do Your betting probabilities relate to frequencies?
In order to make this connexion, we have to realise that You are supposed to be able to assess Your betting probability, not merely for the outcome of each single toss, but for an arbitrary specified combination of outcomes of different tosses: for example, for the event that the results of the first 7 tosses will be HHTTHTH in that order. That is, You have a full personal [*probability distribution*]{} over the full sequence of future outcomes. In particular, You could assess (say) the [*conditional*]{} probability that the 101th toss would be H, given that there were (say) 75 Hs and 25 Ts on the first 100 tosses. Now, before being given that information You might well have no reason to favour H over T, and so assign [*unconditional*]{} probability close to $0.5$ to getting a H at the 101th toss. However, after that information becomes available You might well favour a [*conditional*]{} probability closer to $0.75$. All this is by way of saying that—unlike for the Bernoulli model for repeated trials introduced in above—in Your joint betting distribution the tosses would typically [*not*]{} be independent; since, if they were, no information about the first 100 tosses could change Your probability of seeing H on the 101th.
But if You cannot assume independence, what can You assume about Your joint distribution for the tosses? Here is one property that might seem reasonable: that You simply do not care what order You will see the tosses in. This requires, for example, that Your probability of observing the sequence HHTTHTH should be the same as that of the sequence HTHHTHT, and the same again for any other sequence containing 4 Hs and 3 Ts. This would not be so if, for example, You felt the coin was wearing out with use and acquiring an increasing bias towards H as time passes.
This property of the irrelevance of ordering is termed [ *exchangeability*]{}. It is much weaker than independence, and will often be justifiable, at least to an acceptable approximation.
Now it is a remarkable fact [@definetti:37] that, so long only as Your joint distribution is exchangeable, the following must hold:
1. \[it:df1\] You believe, with probability 1, that there will exist a [*limiting relative frequency*]{}, $p =
\lim_{N\rightarrow\infty} f_N$, of H’s, as the number $N$ of tosses observed tends to infinity.
2. \[it:df2\] You typically do not initially know the value of $p$ (though You could place bets on that value—You have a distribution for $p$); but [*if*]{} You were somehow to learn the value of $p$, then conditionally on that information You would regard all the tosses as being independent, with common probability $p$ of landing H. (There are echoes here of the Principal Principle).
According to , under the weak assumption of exchangeability, the “individualist” Personalist can essentially accept the “groupist” Frequency story: more specifically, the Bernoulli model (with unknown probability $p$). But the Personalist can go even further than the Frequentist: according to , exchangeability provides a warrant for equating the “individual risk,” on each single toss, with the overall “group probability” $p$ (if only that were known…). That is, the Bernoulli model will be agreed upon by all personalists who agree on exchangeability [@apd:intersub], and the limiting relative frequency $p$, across the repeated trials, constitutes a universal expert assignment for this class.
However, an important way in which this Personalist interpretation of the Bernoulli model differs from that of the Frequentist is in the conditions for its applicability. Exchangeability is justifiable when You have [*no sufficient reason*]{} to distinguish between the various trials. This is a much weaker requirement than the Frequentist’s “identity of all relevant characteristics.” For example, suppose You are considering the examination outcomes of a large number of students. You know there will be differences in ability between the students, so they can not be regarded as identical in all relevant respects. However, if You have no specific knowledge about the students that would enable You to distinguish the geniuses from the dunces, it could still be reasonable to treat these outcomes as exchangeable.
An important caveat is that Your judgment of exchangeability is, explicitly or implicitly, conditioned on Your current state of information, and can be destroyed if Your information changes. In the above example of the students, since You are starting with an exchangeable distribution, You believe there will be an overall limiting relative frequency $p$ of failure across all the students, and that if You were to learn $p$ that would be Your correct revised probability that a particular student, Karl, will fail the exam. Suppose, however, You were then to learn something about different students’ abilities. This would not affect Your belief (with probability 1) in the existence of the group limiting relative frequency $p$—but You would no longer be able to treat that $p$ as the individual probability of failure for Karl (who, You now know, is particularly bright).
As an illustration of the above, suppose You are considering the performances of the students (whom You initially consider exchangeable) across a large number of examination papers. Also, while You believe that some examinations are easier than others, You have no specific knowledge as to which those might be, so initially regard the examinations as exchangeable. Now consider the full collection of all outcomes, labelled by student and examination. As a Personalist, You will have a joint distribution for all of these. Moreover, because of Your exchangeability judgments, that joint distribution would be unchanged if You were to shuffle the names of the students, or of the examinations. Suppose now You are interested in “the risk” that Karl will fail the Statistics examination. You might confine attention to Karl, and use the limiting relative frequency of his failures, across all his other examinations, as his “risk” of failing Statistics. This appears reasonable since You are regarding Karl’s performances across all examinations (including Statistics) as exchangeable. In particular, this limiting relative frequency of Karl’s failure, across all examinations, constitutes an expert assignment for the event of his failing Statistics. Alternatively, You could concentrate on the Statistics examination, and take the limiting relative frequency of failure on that examination, by all the other students, as measuring Karl’s risk of failing. This too seems justifiable—and supplies an expert assignment—because You consider the performances of all the students (including Karl) on the Statistics examination as exchangeable. However, these two “risks” will typically differ—because they are conditioned on different information. Indeed, neither can take account of the full information You might have, about the performances of all students on all examinations. Given that full information, You should be able to assess both Karl’s ability (by comparing his average performance with those of the other students) and the difficulty of the Statistics examination (on comparing the average performance in Statistics with those for other examination). Somehow or other You need to use all this information (as well as the performances of other students on other examinations) to come up with Your “true” risk that Karl will fail Statistics—but it is far from obvious how You should go about this.[^3]
We can elaborate such examples still further. Thus suppose students are allowed unlimited repeat attempts at each examination, and that (for each student-examination combination) we can regard the results on repeated attempts as exchangeable. Then yet another interpretation of Karl’s risk of failing Statistics on some given attempt—yet another universal expert assignment—would be the limiting relative frequency of failure across all Karl’s repeat attempts at the Statistics examination. This would typically differ from all the values discussed above.
Now it can be shown that, for this last interpretation of risk, its value would be unaffected by further taking into account all the rest of the data on other students’ performances on other examinations: it is truly conditional on all that is (or could be) known about the various students’ performances. Does this mean that we have finally identified the “true” risk of Karl failing Statistics?
Not so fast… Suppose You consider that Karl’s confidence, and hence his performance, on future attempts will be affected by his previous results, his risk of future failure going up whenever he fails, and down whenever he passes. It is not difficult to make this behaviour consistent with the exchangeability properties already assumed [@hill:87]. As a simple model, consider an urn that initially contains 1 red ball (representing success) and 1 green ball (representing failure). Karl’s successive performances are described by the sequence of draws of balls from this urn, made as follows: whichever ball is drawn is immediately replaced, together with an additional ball of the same colour. At each stage it is assumed the draws are made “at random,” with each ball currently in the urn being “equally likely” to be the next to be drawn (this being a reasonably straightforward application of the classical interpretation). Thus Karl’s future performance is influenced by his past successes and failures, with each success \[failure\] increasing \[decreasing\] the (classical) chance of success at the next attempt.
Now it can be shown that the sequence of colours drawn forms an exchangeable process, and it follows that the relative frequency with which a red ball is drawn, over a long sequence of such draws, will converge to some limit $p$. However, at the start of the process $p$ is not known, but is distributed uniformly over the unit interval. How would you assess the probability that the first draw will result in a red ball? Useless to speculate about the currently unknown value of the limiting relative frequency $p$; the sensible answer is surely the “classical” value $1/2$.
In this case, while You still believe that there will exist a limiting relative frequency of failure across all Karl’s attempts at Statistics (and this constitutes a universal expert assignment), not only is this initially unknown to You, but its very value can be regarded as being constructed over time, as Karl experiences successes and failures and his level of performance gets better and worse accordingly. So why should You regard the limiting relative frequency of future failures, dependent as this is on Karl’s randomly varying performance in his future attempts, as an appropriate measure of his risk of failing on this, his first attempt? As commented by @cane:1977 in a parallel context: “…if several clones were grown, each under the same conditions, an observer[…]{}might feel that the various values \[[*, of a limiting proportion—APD*]{}\] they showed needed explanation, although these values could in fact be attributed to chance events.” This point is relevant to the assessment of risk in the context of an individual’s criminal career (), where the very act of committing a new offence might be thought to raise the likelihood of still further offences. “An increase in criminal history increases the likelihood of recidivism, and a lack of increase can reduce that likelihood. Because criminal history can increase (or not) over time and each crime’s predictive shelf life may be limited, it seems important to conceptualize criminal history as a variable marker” [@monahan:14].
It is interesting to view this ambiguity as to what should be taken as the “real” risk of success—the “classical” proportion of red balls currently in the urn, or the “frequentist” proportion of red balls drawn over the whole sequence—from a Personalist standpoint. Both constitute expert assignments for You, but they differ. The former is more concrete, in that You can actually observe (or compute) it at any stage, which You can not do for the latter.
Suppose You are forced to bet on the colour of the first ball to be drawn. From the classical view, the probability value to use is $0.5$ (and that classical value is fully known to You). Alternatively, taking a frequency view, You would consider the unknown limiting proportion $p$, with its uniform distribution over the unit interval. For immediate betting purposes, the relevant aspect of this is its expectation—which is again $0.5$. Thus the personalist does not have to choose between the two different ways of construing “the probability” of success. (This is a special case of the result mentioned in above). And this indifference extends to each stage of the process: if there are currently $r$ red and $b$ black balls in the urn, Your betting probability for next drawing a red, based on Your current expectation (given Your knowledge of $r$ and $b$, which are determined by the results of previous draws) of the unknown limiting “frequency probability” $p$, will be $r/(r+b)$—again agreeing with the known “classical” value.
But what is the relevance of the above “balls in urn” model to the case of Karl’s repeat attempts? Even though the two stories may be mathematically equivalent, there is no real analogue, for Karl, of the “classical” probability based on counting the balls in the urn: Your probability $r/(r+b)$ of Karl’s failing on his first attempt is merely a feature of Your Personalist view of the world, with limited relevance for any one else. So—what is “the risk” in this situation?
Individual Risk {#sec:indrisk}
===============
Stories of coin tosses and such are untypical of real-world applications of risk and probability. So now we turn to a more realistic example:
\[ex:sam\] What is “the risk” that Sam will die in the next 12 months?
You might have good reason to be interested in this risk: perhaps You are Sam’s life assurance company, or You Yourself are Sam. There is plenty of mortality data around; but Sam is an individual, with many characteristics that, in sum (and in Sam), are unique to him. This makes it problematic to apply either of the G2i arguments above. As a Frequentist, You would need to be able to regard Sam and all the other individuals in the mortality data-files as “identical in all relevant characteristics,” which seems a tall order; while the Personalist You would need to be able to regard Sam and all those other individuals as exchangeable—but will typically know too much about Sam for that condition to be appropriate.
How then could You tune Your risk of Sam’s death to the ambient data? A common way of proceeding is to select a limited set of background variables to measure on all individuals, Sam included. For example, we might classify individuals by means of their age, sex, smoking behaviour, fruit and alcohol intake, and physical activity. We could then restrict attention to the subset of individuals, in the data, that match Sam’s values for these variables, and regard the (ideal limiting) relative frequency of death within 12 months in that subset[^4] as a measure of Sam’s own risk.[^5] To the extent that You can regard Sam as exchangeable with all the other individuals sharing his values for the selected characteristics, the limiting relative frequency in that group constitutes an expert assignment for Sam. There can be no dispute that. from a pragmatic standpoint, such information about frequencies in a group of people “like Sam” (or “like Angelina”) can be extremely helpful and informative. In @imreydawid such a group frequency, regarded as relevant to an individual member of the group, is termed an “individual[*[ized]{}*]{} risk.” The foundational philosophical question, however, is: Can we consider this as supplying a measure of “individual risk”? For the Frequentist, that would require a belief that the chosen attributes capture “all relevant characteristics” of the individuals; for the Personalist, it would require that You have no relevant additional information about Sam (or any of the other individuals in the data), and can properly assume exchangeability—conditional on the limited information that is being taken into account. Neither of these requirements is fully realistic.
In any case, irrespective of philosophical considerations, there are two obvious difficulties with the above “individualization” approach:
- The “risk” so computed will depend on the choice of background variables.
- We may be obliged to ignore potentially relevant information that we have about Sam.
These difficulties are often branded “the problem of the reference class,” and would seem to bedevil any attempt to construct an unassailable definition of “individual risk” from a groupist perspective.
“Deep” risk {#sec:deep}
-----------
The above difficulties might disappear if it were the case that, as we added more and more background information, the frequency value in the matching subpopulation settled down to a limit: what we might term the “deep” risk, conditional on [*all*]{} there is to know about Sam. However, it is not at all clear why such limiting stability should be the case, nor is there much empirical evidence in favour of such a hypothesis. Indeed, it would be difficult to gather such evidence, since, as we increase the level of detail in our background information, so the set of individuals who match Sam on that information will dwindle, eventually leaving just Sam himself.
Even for the case of tossing a coin, it could be argued that if we know “too much” about the circumstances of its tossing, that would lead to a very different assessment of the probability of H; perhaps even, with sufficiently microscopic information about the initial positions and momenta of the molecules of the thumb, the coin, the table and the air, the outcome of the toss would become perfectly predictable, and the “deep” probability would reduce to 0 or 1 (or perhaps not, if we take quantum phenomena into account…).
I would personally be sympathetic to the view that this [*reductio ad absurdum*]{} is a “category error,” since no one really intended the “deep” information to be quite [*that*]{} deep. Perhaps it is the case that the addition of more and more “appropriate” background information, at a more superfical level, would indeed lead to a stabilisation of the probability of H at some non-trivial value. But this account is full of vagueness and ambiguities, and begs many questions. And even if we could resolve it for the extremely untypical case of coin-tossing, that would not give us a licence to assume the existence of “deep risk” for more typical practical examples, such as Sam’s dying in the next 12 months.
A Different Approach: Individual to Group Inference {#sec:i2g}
====================================================
All our discussion so far—even for the “individualist” Personalist conception of Probability—has centred on “Group to Individual” (G2i) inference: taking group frequencies as our fundamental starting point, and asking how these might be used to determine individual risks. And it has to be said that we have reached no conclusive answer to this question. So instead we now turn things upside down, and ask: Suppose we take individual risks as fundamental—how then could we relate these to group frequencies, and with what consequences? This is the “Individual to Group” (i2G) approach. As we shall see, although we will not now be defining individual risks in terms of group frequencies, those frequencies will nevertheless severely constrain what we can take the individual risks to be.
Our basic framework is again an ordered population of individuals $(I_1, I_2, \ldots)$. For individual $I_k$, You have an outcome event of interest, $E_k$, and some background information, $H_k$. We denote Your “information base”—Your full set of background information $(H_1, H_2, \ldots)$ on all individuals—by $\cal H$.
Note, importantly, that in this approach we need not assume that You have similar information for different individuals, nor even that the different outcome events are of the same type (though in applications that will typically be the case). In particular, we shall not impose any analogue of either the Frequentist’s condition of “repeated trials under identical conditions,” or the Personalist’s judgment of exchangeability.
We start with an initial bold assumption: that You have been able to assess, for each individual $I_k$, a probability, $p_k = \Pr(E_k \mid
H_k)$, for the associated event $E_k$, in the light of the associated background information $H_k$. This might, but need not be, interpreted as a Personalist betting probability: all that we want is that it be some sort of “individualist” assessment of uncertainty. We put no other constraint on these “individual risks”—most important, they will typically vary from one individual to another. We term $p_k$ a [*probability forecast*]{} for $E_k$. For some purposes we will require that the forecasts are based [*only*]{} on the information in ${\cal H}$, and nothing else. There are some subtleties involved in making this intuitively meaningful condition mathematically precise—one approach is through the theory of computability [@apd:empprob]. Here we will be content to note that this can be done. We shall term such forecasts [*${\cal
H}$-based*]{}.
There are two similar but slightly different scenarios that we can analyse, with essentially identical results. In the first, which we may term the [*independence*]{} scenario, the background information $H_k$ pertains solely to individual $I_k$, and it is supposed that, given $H_k$, Your uncertainty about $E_k$, as expressed by $p_k$, would not change if You were to receive any further information (be it background information or outcome information) on any of the other individuals. In the second, [*sequential*]{} scenario, $H_k$ represents the total background information on all previous and current individuals $I_1, \ldots, I_k$, as well as the outcomes of $E_1, \ldots, E_{k-1}$ for the previous individuals; no other conditions need be imposed.
An important practical application, in the sequential formulation, is to weather forecasting, where $I_k$ denotes day $k$, $E_k$ denotes “rain on day $k$,” $H_k$ denotes the (possibly very detailed) information You have about the weather (including whether or not it rained) up to and including the previous day $k-1$, and You have to go on TV at 6pm each evening and announce Your probability $p_k$ that it will rain the following day $I_k$. We will often use this particular example to clarify general concepts.
Our proposed relationship between individual probabilities and group frequencies will be based on the following idea: Although You are free to announce any probability values You want for rain tomorrow, if You are to be trusted as a reliable weather forecaster, these values should bear some relationship to whether or not it does actually rain on the days for which You have issued forecasts. Note that this approach judges Your probabilities by comparison with the outcomes of the events—not with the “true probabilities” of the events. No commitment as to the existence of “true probabilities” is called for.
Calibration {#sec:simpcal}
-----------
Let $e_k$ denote the actual outcome of event $E_k$, coded $1$ if $E_k$ happens, and $0$ if it does not.
### Overall calibration {#sec:overallcal}
We start by proposing the following [*overall calibration criterion*]{} of agreement between the probability forecasts $(p_k)$ and the outcomes $(e_k)$: over a long initial sequence $I_1, I_2,
\dots, I_N$, the overall proportion of the associated events that occur, $$\label{eq:prop1}
\frac{\sum_{i=1}^Ne_{i}}N,$$ which is a “groupist” property, should be close to the average of the “individualist” forecast probabilities, $$\label{eq:av1}
\frac{\sum_{i=1}^Np_{i}}N.$$
This seems a plausible requirement, but do we have a good warrant for imposing it? Yes. It can be shown [@apd:wellcal] that (for either of the scenarios) the overall calibration property is assigned probability 1 by Your underlying probability distribution. That is to say, You firmly believe that Your probability forecasts will display overall calibration. So, if overall calibration turns out [*not*]{} to be satisfied, an event that You were convinced was going to happen has failed to occur—a serious anomaly, that discredits Your whole distribution, and with it Your probability forecasts [@apd:synthesis]. Overall calibration thus acts as a minimal “sanity check” on your probabiity forecasts: if it fails, You are clearly doing something wrong.
However, that does not mean that, if it holds, You are doing everything right: overall calibration is a weak requirement. For example, in an environment where it rains $50\%$ of the time, a weather forecaster who, ignoring all information about the past weather, always announces a probability of $50\%$, will satisfy overall calibration—but if in fact it rains every alternate day (and much more generally) he will be showing little genuine ability to forecast the changing weather on a day-by-day basis. Another forecaster, who has a crystal ball, always gives probability 1 or 0 to rain, and always gets it right. Her perfect forecasts also satisfy overall calibration.
In order to make finer distinctions between forecasts of such very different quality, we will progressively strengthen the calibration criterion, through a number of stages. Note, importantly, that every such strengthened variant will share the “sanity check” function described above for overall calibration: according to Your probability distribution, it will be satisfied with probability 1. So, if it fails, Your probability forecasts are discredited.
### Probability calibration {#sec:probcal}
For our next step, instead of taking the averages in and over all days until day $N$, we focus on just those days $I_k$ for which the forecast probability $p_k$ was equal to (or very close to) some pre-assigned value. If that value is, say, $30\%$ (and that value is eventually used infinitely often) then [*probability calibration*]{} requires that, in the limit, the proportion of these days on which it in fact rains should be (close to) $30\%$; and similarly for any other pre-assigned value.
However, although probability calibration is again a very natural idea, it is still too weak for our purpose. In particular, it will still be satisfied for both our above examples of uninformative forecasts and of perfect forecasts.
### Subset calibration {#sec:subsetcal}
For our next attempt, we allow the averages in and to be restricted to a subset of the individuals, arbitrarily chosen except for the requirement that it must be selected without taking any account of the values of the $E$ and the $H$’s. For example, we might choose every second day. If in fact the weather alternates wet, dry, wet, dry,…, then the uninformative forecaster, who always says $50\%$, will now fail on this criterion, since if we restrict to the odd days alone his average forecast probability is still $50\%$, but the proportion of rainy days will now be $100\%$. The perfect forecaster will however be announcing a probability forecast of $100\%$ for every odd day, and so will satisfy this criterion.
However, although subset calibration has succeeded in making the desired distinction, even this is not strong enough for our purposes.
### Information-based calibration {#sec:infcal}
In all our attempts so far, we have not made any essential use of the background information base ${\cal H}$, and the requirement that the forecasts be ${\cal H}$-based. But we cannot properly check whether a forecaster is making appropriate use of this background information without ourselves taking account of it.
In the sequential weather forecasting scenario, the forecaster is supposed to be taking account of (at least) whether or not it rained on previous days, and to be responding appropriately to any pattern that may be present in those outcomes. To test this, we could form a test subset in a dynamic way, ourselves taking account of all the forecaster’s background information (but, to be totally fair to the forecaster, nothing else[^6]). Thus we might consider, for example, the subset comprising just those Tuesdays when it had rained on both previous days. If the forecaster is doing a proper job, he should be calibrated (, his average probability forecast should agree with the actual proportion of rainy days), even if we restrict the averages to be over such an “${\cal H}$-based” subset. An essentially identical definition applies in the independence scenario.
When this property is satisfied for all ${\cal H}$-based subsets, we will call the probability forecasts [*${\cal H}$ calibrated*]{}. A set of forecasts that is both ${\cal H}$-based and ${\cal
H}$-calibrated will be called [*${\cal H}$-valid*]{}. Again we stress that You assign probability 1 to Your $\cal H$-based forecasts being $\cal H$-valid, so this is an appropriate condition to impose on them.
Finally, we have a strong criterion relating individual probability forecasts and frequencies. Indeed, it constrains the individuals forecasts so much that, in the limit at least, their values are fully determined. Thus suppose that we have two forecasters, who issue respective forecasts $(p_k)$ and $(q_k)$, and that both sets are ${\cal H}$-valid. It can then be shown [@apd:wellcal] that, as $k$ increases without limit, the difference between $p_k$ and $q_k$ must approach $0$. We may term this result [*asymptotic identification*]{}.
This is a remarkable result. We have supposed that “individual risks” are given, but have constrained these only through the “groupist” ${\cal H}$-calibration criterion. But we see that this results “almost” in full identification of the individual values, in the sense that, if two different sets of probability forecasts both satisfy this criterion, then they must be essentially identical. In particular, if there exists any set of $\cal H$-valid forecasts,[^7] then those values are the “essentially correct” ones: any other set of $\cal H$-based forecasts that does not agree, asymptotically, with those values can not be $\cal
H$-valid—and is thus discredited. So in this sense the i2G approach has succeeded—where the more traditional G2i approaches failed—in determining the values of individual risks on the basis of group frequencies.
As an almost too simple example of this result, in the independence scenario, suppose that $H_k$, the background information for individual $I_k$, is his score on the Violence Risk Appraisal Guide ARAI [@VRAG], with 9 categories. If this limited information ${\cal H}$ is all that is available to the forecaster, his forecasts will be ${\cal H}$-based just when they announce the identical probability value for all individuals in the same VRAG category. And they will be ${\cal H}$-calibrated if and only if, within each of the 9 categories, his announced probability agrees with the actual reoffending rates. That is, ${\cal H}$-validity here reduces to the identification of [*individual*]{} risk with [*individualized*]{} risk; and this identification is thus justified so long as Your complete information is indeed restricted to that comprised by ${\cal
H}$.
What this simple example fails to exhibit is that, in more general cases, the “determination” of the individual forecasts is only asymptotic: given a set of $\cal H$-valid forecasts, we could typically change their values for a finite number of individuals, without affecting ${\cal H}$-validity. So we are not, after all, able to associate a definitive risk value with any single individual.
Another big downside of the above result is that, while assuring us of the essential uniqueness of $\cal H$-valid forecasts, it is entirely non-constructive. It seems reasonable to suppose that, once we know that such uniquely determined forecasts exist, there should be a way to construct them: for example, it would be nice to have an algorithm that would issue probability forecasts for the following day, based on past weather (${\cal H})$, that will be properly calibrated, however the future weather in fact turns out. But alas! in general this is not possible, and there is no ${\cal H}$-based system that can be guaranteed to be ${\cal H}$-calibrated [@oakes:1985; @apd:oakes].
Another potentially serious limitation of the i2G approach is its dependence on the population of individuals considered, and moreover on the order in which they are strung out as a single sequence. It does not seem easy to accommodate a structure, such as considered in , where many students take many resits of many examinations. For one thing, it seems problematic to incorporate the judgements of exchangeability made there into the i2G approach; for another, different ways of forming a sequence of student-examination-resit combinations could well lead to mutually inconsistent calibration requirements.
Varying the information base {#sec:vary}
----------------------------
A fundamental aspect of the i2G approach to individual risk is that this is a relative, not an absolute, concept: its definition and interpretation depend explicitly on the information that is being taken into account,[^8] as embodied in ${\cal H}$. If we change the information base, the associated valid probability values will change.
We can relate the i2G risks based on different information bases, one more complete than the other. Thus suppose that ${\cal K} = (K_k)$ is an information base that is more detailed than ${\cal H}$: perhaps $H_k$ only gives information about whether or not it rained on days prior to day $k$, while $K_k$ also contains information about past maximum temperature, wind speed, Suppose $(p_k)$ is a set of ${\cal H}$-valid forecasts, and $(q_k)$ a set of ${\cal K}$-valid forecasts. These would typically differ, even asymptotically: We would expect the $(q_k)$, being based on more information, to be “better” than the $(p_k)$, and so would [*not*]{} expect $p_k -
q_k \rightarrow 0$. This does not contradict asymptotic identification. The $(q_k)$ are not ${\cal H}$-based, so not ${\cal
H}$-valid. Also, while the $(p_k)$ are ${\cal K}$-based, they have not been required to be ${\cal K}$-calibrated, so they are not ${\cal
K}$-valid. So, whether we take the underlying information base to be ${\cal H}$ or ${\cal K}$, the conditions implying asymptotic identification simply do not apply.
It does however follow from the argument for asymptotic identification that, if we consider the subsequence for which, say, $p_k = 0.4$ (to a good enough approximation), then—while we can expect the $(q_k)$ in that subsequence to vary (because of variations in the additional information in ${\cal K}$ that they take into account)—their limiting average value in the subsequence will be $0.4$ (and similarly for any other target value for the $(p_k)$). In this sense the “deeper” risks $(q_k)$ can be regarded as varying “randomly” about the “shallower” risks $(p_k)$—just as the actual outcomes $(E_k)$ do. Some of the attempts to model individual risks as random, such that of @jr+sg:89, might be understood as contemplating an expanded but unobservable information base ${\cal K}$, and interpreting the “true risks” as the ${\cal K}$-valid ones. However, with access only to the shallower information base ${\cal
H}$, we can only “observe” certain crude averages of these “true risks.” If ${\cal H}$ is indeed all the information at our disposal, there is nothing of value to be gained by extending consideration to the unobservable deeper risks $(q_k)$—which in any case depend on the essentially arbitrary specification of the unobserved deeper information base ${\cal K}$. The flip-side of this is that, when we do have access to the deeper information, we should use that to identify the relevant “individual risk,” rather than be satisfied with a cruder average value based on more superficial information.
An interesting implication of the above argument is that our “individual-to-group” approach allows for the assignment of non-extreme probabilities to events, even when we believe in “deep determinism”—that is, we believe that, given a suitably detailed information base ${\cal K}$, it would be possible to forecast the future perfectly (in which case ${\cal K}$-valid probabilities $(q_k$) would have to be, asymptotically, $0$ or $1$). However, if we only have access to a less detailed information base ${\cal H}$, the associated ${\cal H}$-valid risk values would normally be non-extreme. This approach thus justifies the use of probability as a description of a system we believe to operate deterministically, so long as we are in ignorance of the deep determining circumstances.
Discussion {#sec:disc}
==========
I have surveyed a number of conceptions of the meaning of “individual risk,” and found all of them wanting to some degree. The various G2i approaches all founder on the “problem of the reference class,” which does not have an unambiguous solution—although reasonable pragmatic choices can often be made, and defended as such. The i2G approach avoids ambiguity and delivers individual risk values, nicely calibrated to the information considered available—but only “at infinity,” and even then these asymptotic values are uncomputable. It also does not seem able to take account of sophisticated exchangeability requirements.
What then is one to do? Should the whole idea of individual risk be abandoned? I think this is too extreme, but certainly the concept and nomenclature should not be bandied about carelessly, and it behoves any one using the term to give an account of what they (think they) mean by it.
For what it is worth, my own tentative attitude is as follows, based on an essentially personalistic viewpoint. Suppose I am tasked with assessing “the risk” that Cain, a prisoner up for possible parole, will if released commit a violent act (say, within the next 24 months). Then I should think about Cain, and, taking full account of all I know about him and others like him (and even unlike him), compute my probability forecast by assessing the odds at which I would be willing to bet that he will commit a violent act. This probability can in principle be considered as formed, by conditioning on the information I have on Cain and others, from my full personal joint probability distribution for the properties and outcomes of all these individuals. Under certain strong conditions on that joint distribution, such as exchangeability, my forecast probability would be close to the proportion of such events in people “like Cain”; but more generally, that proportion, and other relevant proportions, would inform, but not directly constrain, my personal, properly conditioned, individual probability forecast for Cain.
But, you object, what of objectivity? You want to know “the risk” that Cain will be violent. Why should you care about my personal probability assessment?
Well, in the large my probability forecasts can be tested against realised outcomes [@apd:enc(probfore)], , using finite-data analogues of calibration [@fs/apd] to see how closely they align with actual frequencies in appropriately selected sets.[^9] If, by such tests (formal or informal), my announced probability forecasts are shown to be out of line with reality, you have every right to discount my risk assessment for Cain. But if they have passed a suitable battery of such tests—are, provisionally, “valid”—then (arguing informally by analogy with our asymptotic identification result) we might expect them to be close to the valid risk assessments of others. So, in the presence of a sufficient quantity of relevant data to allow us to conduct such statistical tests, you should be able to judge whether my risk assessments are reasonably “objective,” and if so have some limited confidence in my—now statistically justified—announced risk for Cain.
But recall that “valid” probability forecasts depend on the information that is being taken into account. Ideally this should be the most detailed information we have about Cain and those like him; but, the more detailed the information is, the harder it will be for me to make my forecasts valid.[^10] So I might choose to artifically restrict my information base, perhaps to just a few simple characteristics. I would thus be sacrificing incisiveness—the possibility (however remote) of making “deep” probability forecasts, valid with respect to the deep information that I could, in principle, take into account—in favour of robustness—the enhanced prospect of achieving validity with respect to deliberately restricted information. In the judicial context, and many others, this sacrifice might well be considered worthwhile: indeed, a suitable specification of what would be “appropriate” characteristics to take into account could be enshrined in statute.[^11] That done, suitable statistical methodology applied to these could produce valid risk assessments—albeit relating to a shallower level of information than would ideally be desirable. The various ARAI systems that have been developed can be thought of as addressing this task.
Acknowledgments {#acknowledgments .unnumbered}
===============
I am grateful to David Faigman, Alan Hájek, Peter Imrey, John Monahan, Russell Poldrack and David Spiegelhalter for helpful comments, and to the MacArthur Foundation for support through its Research Network on Law and Neuroscience.
[^1]: Interestingly, this is not the case for its mathematical properties, where the “Kolmogorov axioms” are largely accepted by all the main factions.
[^2]: Though probably no one else. @hajek:sep uses the term “finite frequentism.”
[^3]: There is in principle a way to do this [@aldous; @hoover:82], supplying a universal expert assignment for the class of all personalists agreeing on exchangeability both across students and across examinations—but it relies on advanced mathematics and is highly non-trivial to implement.
[^4]: Or, we could set up a statistical model for the form of the dependence of this risk on the given attributes, estimate its parameters from the ambient data, and apply it to Sam.
[^5]: An animation for exploring the dependence of this frequency on these characteristics (as well as calendar year) can be found at the website [ understandinguncertainty.org](http://understandinguncertainty.org/files/animations/Survival1/Survival.html).
[^6]: Again, this requirement can be formalised using computability theory.
[^7]: which is however not guaranteed [@schervish:85; @belot]
[^8]: In this it has some of the flavour of Logical Probability
[^9]: One criterion I should definitely [*not*]{} be judged on is how good my probability forecasts are as estimates of the “objective” individual risks. This is for two reasons: first, in the unavoidable absence of knowledge of the values of those risks, this can not be done; and, secondly, there are no such things!
[^10]: Individual risk assessment based on very detailed personal information becomes essentially a matter of “clinical judgment.” Although practitioners in many disciplines—medicine, law, psychiatry, —often have great confidence in their own clinical judgments, these can be very far from being valid [@meehl].
[^11]: This leaves the possibility that an individual might appeal on the grounds that relevant information, specific to his case, was left out of consideration. But it would be for the appellant to make a good case for this.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We derive properties of powers of a function satisfying a second-order linear differential equation. In particular we prove that the $n$-th power of the function satisfies an $(n+1)$-th order differential equation and give a simple method for obtaining the differential equation. Also we determine the exponents of the differential equation and derive a bound for the degree of the polynomials, which are coefficients in the differential equation. The bound corresponds to the order of differential equation satisfied by the $n$-fold convolution of the Fourier transform of the function. These results are applied to some probability density functions used in statistics.'
author:
- 'Naoki Marumo[^1], Toshinori Oaku[^2] and Akimichi Takemura '
bibliography:
- 'second-order-holonomic.bib'
date: 'May, 2014'
title: 'Properties of powers of functions satisfying second-order linear differential equations with applications to statistics'
---
[*Keywords and phrases:*]{} characteristic function, exponents, holonomic function, indicial equation, skewness
Introduction {#sec:intro}
============
In statistics it is important to study the distribution of a sum (i.e. convolution) of $n$ independent random variables. Usually the distribution is studied through the characteristic function, because the convolution of probability density functions corresponds to the product of characteristic functions. If the random variables are identically distributed, then we study the $n$-th power of a characteristic function. The central limit theorem is proved by analyzing the limiting behavior of the $n$-th power of a characteristic function as $n\rightarrow\infty$. Often the technique of asymptotic expansion is employed to improve the approximation for large $n$. However for finite $n$, the exact distribution of the sum of random variables is often difficult to treat. Hence it is important to develop methodology for studying properties of the $n$-th power of a function.
Recently techniques based on holonomic functions ([@kauers_book_11], Chapter 6 of [@hibi_book_13]) have been introduced to statistics and successfully applied to some difficult distributional problems (e.g. [@nakayama], [@hashiguchi]). In this paper we investigate the case that the function satisfies a second-order linear differential equation with rational function coefficients, which we call [*holonomic differential equation*]{}. In Section 2 we prove that the $n$-th power satisfies an $(n+1)$-th order differential equation and give a simple method for obtaining the differential equation. Also we determine the exponents of the differential equation and derive a bound for the degree of the polynomials which appear as coefficients of the differential equation.
As shown in Section 3, there are some important examples in statistics which falls into this case. We discuss sum of beta random variables and sum of cubes of standard normal random variables. The differential equations reveal many interesting properties of the characteristic function and the probability density function of the sum of random variables. These properties are hard to obtain by other methods. We end the paper with some discussions in Section 4.
Main results {#sec:main}
============
In this section we present our main results in Theorems \[thm:ode\], \[thm:deg\] and \[thm:exponents\]. Theorem \[thm:ode\] gives the differential equation satisfied by the $n$-th power. Theorem \[thm:deg\] bounds the degree of coefficient polynomials. Theorem \[thm:exponents\] derives exponents of the differential equation.
Let $\mathbf C(x)$ denote the field of rational functions in $x$ with complex coefficients and let $$R=\mathbf C(x) \langle \p_x \rangle, \quad \partial_x = \frac{d}{dx},$$ denote the ring of differential operators with rational function coefficients. In $R$, the product of $\p_x$ and $a(x)\in \mathbf C(x)$ is defined as $\p_x a(x) = a(x) \p_x + a'(x)$, where $a'(x)$ is the derivative of $a(x)$ with respect to $x$. In order to distinguish the product in $R$ and the action of $\p_x$ to a function, we denote the latter by the symbol $\bullet$.
If we write $\p_x x$, both $\p_x$ and $x$ are the elements of $R$. Hence $\p_x x = x\p_x+1$. On the other hand, if we write $\p_x \bullet x$, this $x$ is a function. Hence $\p_x \bullet x = 1$.
In this paper we study $f(x)$ which is a holonomic function satisfying a second-order differential equation: $$\Big[ \p_x^2 - a_1(x)\p_x - a_0(x) \Big] \bullet f(x)
= 0, \qquad a_0(x), a_1(x)\in \mathbf C(x). \label{ode 1}$$
Order of the differential equation of the $n$-th power and its Fourier transform {#subsec:order-n-th-power}
--------------------------------------------------------------------------------
Let $\bm q_0 = \T{(1,\ 0,\ \dots,\ 0)}$ be an $(n+1)$ dimensional column vector and let $$A(x)=
\bma
0 & a_0(x) \\
n & a_1(x) & 2a_0(x) \\
& n-1 & 2a_1(x) & \ddots \\
&& \ddots & \ddots & na_0(x)\\
&&& 1 & na_1(x)
\ema
\label{eq:qk0}$$ be an $(n+1)\times (n+1)$ tridiagonal matrix with entries from $\mathbf C(x)$. Furthermore define $$\begin{aligned}
\tilde A (x,\p_x) &= A(x) + \p_x I \label{eq:q_k}
\\
&=\bma
\p_x & a_0(x) \\
n & \p_x + a_1(x) & 2a_0(x) \\
& n-1 & \p_x + 2a_1(x) & \ddots \\
&& \ddots & \ddots & (n-1)a_0(x) \\
&&& 2 & \p_x + (n-1)a_1(x) & na_0(x)\\
&&&& 1 & \p_x + na_1(x)
\ema
\nonumber\end{aligned}$$ with entries from $R$. Let $$\begin{aligned}
Q(x)=(q_{ij}(x))_{\substack{0\leq i \leq n\\0 \leq j \leq n+1}} = (\bm q_0 ,\ \tilde A(x,\p_x) \bullet \bm q_0 ,\ \tilde A(x,\p_x)^2 \bullet \bm q_0 ,\ \dots,\ \tilde A(x,\p_x)^{n+1} \bullet \bm q_0)\end{aligned}$$ be an $(n+1)\times (n+2)$ matrix with entries from $\mathbf C(x)$. If we write ${\bm q}_j = \tilde A(x,\p_x)^j \bullet {\bm q}_0$, $j=0,\dots,n+1$, then $${\bm q}_{j+1}=\tilde A(x,\p_x) \bullet {\bm q}_j,$$ or writing down the elements we have $$\label{eq:qij}
q_{i,j+1}(x)
= (n+1-i)q_{i-1,j}(x) + (\p_x + ia_1(x)) \bullet q_{i,j}(x) + (i+1)a_0(x) q_{i+1,j}(x),$$ where $q_{-1,j}(x)=q_{n+1,j}(x)=0$. Hence it is easy to compute the elements of the columns of $Q(x)$ recursively, starting from the first column.
Define $$\label{eq:falling-factorial}
[n]_i= \prod_{k=0}^{i-1} (n-k), \qquad ([n]_0=1).$$ From we can easily prove that $Q(x)$ is an upper-triangular matrix with non-zero diagonal elements, although $Q(x)$ is not a square matrix (cf. Example \[ex:qx\] below).
\[lem:tri\] $q_{ij}(x)=0$ if $i>j$. $q_{ii}(x) = [n]_i \neq 0 \ \ (i=0,1,\dots,n)$.
We use induction on $j$. The result is trivial for $j=0$. Assume $q_{ij}(x) = 0\ (i>j)$ and $q_{j,j}(x) = [n]_j$. Then by we have $q_{i,j+1}(x) = 0\ (i>j+1)$ and $q_{j+1,j+1}(x) = (n-j)[n]_j=[n]_{j+1}$.
This lemma implies $\rank Q(x)=n+1$, or $\dim \Ker Q(x) = 1$. Hence the element of $\Ker Q(x)$ is unique up to the multiplication of a rational function. Here note that we are using the linear algebra over $\mathbf C(x)$.
Let $$\label{eq:kernel-vector}
\bm v(x)=(v_{i}(x))_{0\leq i \leq n+1} \in \Ker Q(x), \quad \bm v(x)\neq \bm 0,$$ where $v_i(x)\in \mathbf C(x)$, $i=0,\dots,n+1$. Once we set $v_{n+1}(x)\neq 0$, then by the triangularity of $Q(x)$, $v_{n}(x), v_{n-1}(x), \dots, v_0(x)$ are successively determined. Moreover, if we set $v_{n+1}(x)=0$, then we obtain $v_n(x)=\dots=v_0(x)=0$. Hence $v_{n+1}(x)\neq 0$ for $\bm v(x)\neq \bm 0$. Often we set $v_{n+1}(x)=1$. For theoretical investigation it is convenient to clear the common denominators of $v_i(x)$’s and take $v_i(x)$’s as polynomials.
\[ex:qx\] Let $n=3$ and let $a_0(x) = 1+x^{-2},\ a_1(x) = -x^{-1}$. Then $$\begin{aligned}
Q(x)=
\bma
1 & 0 & 3 + 3x^{-2} & -3x^{-1} - 9x^{-3} & 21 + 51x^{-2} + 54x^{-4} \\
0 & 3 & -3x^{-1} & 21 + 27x^{-2} & -66x^{-1} - 144x^{-3}\\
0 & 0 & 6 & -18x^{-1} & 60 + 126x^{-2} \\
0 & 0 & 0 & 6 & -36x^{-1} \\
\ema\end{aligned}$$ If we set $v_4(x)=1$, we successively obtain $$\begin{aligned}
v_3(x) = 6x^{-1},\quad
v_2(x) = -10-3x^{-2},\quad
v_1(x) = -30x^{-1}-9x^{-3},\quad
v_0(x) = 9 + 6x^{-2} + 9x^{-4}.\end{aligned}$$ Multiplying by $x^4$ we obtain $\bm v$ with polynomial elements.
We now derive a holonomic differential equation satisfied by the $n$-th power of the holonomic function $f(x)$.
\[thm:ode\] The $n$-th power of $f(x)$ satisfies the following $(n+1)$-th order holonomic differential equation: $$\begin{aligned}
\Big[
v_{n+1}(x)\p_x^{n+1} + v_n(x) \p_x^n + \dots + v_1(x) \p_x + v_0(x)
\Big]\bullet f(x)^n
=0, \label{ode n}\end{aligned}$$ where $v_i(x)$’s are given in .
By induction we prove $$\begin{aligned}
\p_x^k \bullet f^n
= (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)\ \tilde A(x,\p_x)^k \bullet \bm q_0 \label{ind 1}\end{aligned}$$ for any $k\ge 0$. It is obvious for $k=0$. By , $$\begin{aligned}
&\quad \p_x \bullet \Big(f(x)^{n-j}f'(x)^j\Big)\\
& = j f(x)^{n-j} f'(x)^{j-1}f''(x) + (n-j)f(x)^{n-j-1} f'(x)^{j+1}\\
& = j a_0(x)f(x)^{n-j+1}f'(x)^{j-1} + j a_1(x) f(x)^{n-j} f'(x)^j + (n-j)f(x)^{n-j-1} f'(x)^{j+1}\end{aligned}$$ holds for all $j=0,1,\dots,n$, and this leads to $$\p_x \bullet
(f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)
= (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)\ A(x),
\label{eq:ax1}$$ where $A(x)$ is given in .
Hence, assuming for $k$, we obtain $$\begin{aligned}
\p_x^{k+1} \bullet f^n
&= \Big( \p_x \bullet (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n) \Big)\ \tilde A(x,\p_x)^k \bullet \bm q_0\\
&\qquad + (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)\ \p_x \bullet \Big( \tilde A(x,\p_x)^k \bullet \bm q_0 \Big)\\
&= (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)\,A(x) \ \tilde A(x,\p_x)^k \bullet \bm q_0 \qquad\qquad (\text{by \eqref{eq:ax1}})\\
&\qquad + (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)\ \p_x \tilde A(x,\p_x)^k \bullet \bm q_0\\
&= (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)\ (A(x) +\p_x I)\ \tilde A(x,\p_x)^k \bullet \bm q_0 \\
&= (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)\ \tilde A(x,\p_x)^{k+1} \bullet \bm q_0.
\qquad\qquad (\text{by \eqref{eq:q_k}})\end{aligned}$$ Thus is proved. By arranging for $k=0,1,\dots,n+1$, we have $$\begin{aligned}
(f^n,\ \p_x\bullet f^n,\ \dots,\ \p_x^{n+1}\bullet f^n) = (f^n ,\ f^{n-1} f',\ \dots,\ f f'^{n-1},\ f'^n)\ Q(x).\end{aligned}$$ By multiplying it by $\bm v(x) \in \Ker Q(x)$ from the right, we obtain .
If we just want to show the existence of a holonomic differential equation of order $n+1$, we have only to consider $$\begin{aligned}
M = \mathbf C(x) f^n + \mathbf C(x) f^{n-1}f' + \cdots + \mathbf C(x) ff'^{n-1} + \mathbf C(x) f'^n.\end{aligned}$$ Then $M$ is a left $R$-module as well as a vector space over $\mathbf C(x)$ of dimension at most $n+1$. Hence $n+2$ elements, $f^n$, $\p_x\bullet f^n$, …, $\p_x^{n+1}\bullet f^n$, which belong to $M$, are linearly dependent over $\mathbf C(x)$. Similarly, we see that when $f(x)$ satisfies a holonomic differential equation of order $r$ $(\geq 3)$, $f(x)^n$ satisfies a holonomic differential equation of order $\binom{n+r-1}{r-1}$.
There exists a function $f(x)$ satisfying a second-order holonomic differential equation, such that $f(x)^n$ does not satisfy any holonomic differential equation of order less than $n+1$.
Let $f(x)=\sin x$, with $f''(x)+f(x)=0$. We prove by contradiction that $f^n,\ f^{n-1}f', \dots, f'^n$, or $\sin^n x$, $\sin^{n-1}x \cos x$, $\dots$, $\cos^n x$ are linearly independent over $\mathbf C(x)$. It is obvious for $n=0$. Let $m \geq 1$ be the smallest integer such that $\sin^m x, \dots, \cos^m x$ are linearly dependent. Then, there exist rational functions $q_0(x),\dots,q_m(x)$, not all zero, such that $$\begin{aligned}
q_0(x)\sin^m x + q_1(x)\sin^{m-1}x \cos x+\dots+ q_{m-1}(x)\sin x \cos x^{m-1}+q_m(x)\cos^m x=0. \label{l i}\end{aligned}$$ By putting $x=k\pi\ (k=0,1,\dots)$, $q_m(x)$ has infinite number of zeros, and therefore $q_m(x)$ is identically zero. Divide the equation by $\sin x$, and we obtain $q_0(x)\sin^{m-1} x +\dots+q_{m-1}(x)\cos^{m-1} x=0$, which is a contradiction.
Since $f^n,\ f^{n-1}f', \dots, f'^n$ are linearly independent and the matrix $\tilde Q(x) = (\bm q_0 ,\ \tilde A \bullet \bm q_0 ,\ \dots,\ \tilde A^n \bullet \bm q_0)$ is non-singular by Lemma \[lem:tri\] over $\mathbf C(x)$, $(f^n,\ \p_x \bullet f^n,\ \dots,\ \p_x^n \bullet f^n) = (f^n,\ f^{n-1}f', \dots, f'^n)\tilde Q(x)$ are linearly independent over $\mathbf C(x)$. Thus, there does not exist a holonomic differential equation of order less than $n+1$ satisfied by $f(x)^n=\sin^n x$.
We have already remarked that we can take $v_i(x)$, $0\le i\le n+1$, as polynomials in . Also we can cancel common factors in them. Hence we can assume that they are coprime polynomials. We now investigate the highest degree of these polynomials, which is important when the differential equation is Fourier transformed, because it is equal to the order of the transformed equation.
For the rest of this subsection we assume that $a_0(x),a_1(x)$ are Laurent polynomials. Here, we define mindeg and maxdeg of a Laurent polynomial.
For a non-zero Laurent polynomial $f(x)=\sum_{k=m}^M c_k x^k\ (m<M,\ c_m\neq 0,\ c_M\neq 0)$, we define $$\begin{aligned}
\mindeg f(x) = m,\qquad
\maxdeg f(x) = M.\end{aligned}$$ We define $\mindeg 0 = \infty$, $\maxdeg 0 = -\infty$.
Note that for a polynomial $f(x)$, $\maxdeg f(x)=\deg f(x)$.
Now we state the following theorem on the largest degree of the polynomials.
\[thm:deg\] Assume that $a_0(x), a_1(x)$ in are Laurent polynomials and let $m_i = \mindeg a_i(x)$, $M_i = \maxdeg a_i(x)$, $i=0,1$. Let $v_0(x),\ v_1(x),\ \dots,\ v_{n+1}(x)$ be coprime polynomials in . If $m_1 \leq -1,\ M_1\geq -1,\ m_0\geq 2m_1,\ M_0\leq 2M_1$, then $$\begin{aligned}
\max_{0\leq k \leq n+1} \deg v_k(x)
\leq \max\{M_0+(n-1)M_1, nM_1, 0 \} - \min\{ m_0, m_1 \} - (n-1)m_1. \label{maxdeg}\end{aligned}$$
Let $m_{ij}$ denote $\mindeg q_{ij}(x)$. We prove $$\begin{aligned}
m_{0j} \geq m_0 + (j-2)m_1,\qquad
m_{ij} \geq (j-i)m_1,\quad
(0<i<j),\label{degin}\end{aligned}$$ for $j=2,3,\dots,n+1$ by induction. It is easy to check them for $j=2$. Assuming them up to $j$, by , we have $$\begin{aligned}
m_{0,j+1}
&\geq \min\{ m_{0j} + m_1,\ m_{1j} + m_0\}\\
&= m_0 + (j-1)m_1,\\
m_{i,j+1}
&\geq \min\{ m_{i-1,j},\ m_{ij} + m_1,\ m_{i+1,j} + m_0 \}\\
&\geq \min\{ m_{i-1,j},\ m_{ij} + m_1,\ m_{i+1,j} + 2m_1 \}\\
&= (j-i+1)m_1, \qquad(0<i<j+1).\end{aligned}$$ Thus, the results are shown by induction.
Hence choosing an element $\tilde {\bm v}(x) = \T{(\tilde v_0(x),\ \dots,\ \tilde v_n(x),\ 1)}$ $\in \Ker Q(x)$, we successively obtain $$\begin{aligned}
&\mindeg \tilde v_n(x) \geq m_1,\quad
\mindeg \tilde v_{n-1}(x) \geq 2m_1,\\
&\qquad \dots,\quad \mindeg \tilde v_1(x) \geq nm_1,\quad
\mindeg \tilde v_0(x) \geq m_0 + (n-1)m_1.\end{aligned}$$ This implies that $\min_{k} \mindeg \tilde v_k(x) \geq \min\{ m_0+(n-1)m_1,\ nm_1,\ 0 \} = \min\{m_0,\ m_1\} + (n-1)m_1$. By regarding as Laurent polynomials of $x$ as those of $x^{-1}$, we also have $\max_{k} \maxdeg \tilde v_k(x) \leq \max\{ M_0+(n-1)M_1,\ nM_1,\ 0 \}$. Therefore, clearing the denominators of $\tilde {\bm v}$ of , we obtain for the polynomials $v_i(x)$ of $\bm v$.
Let $D=\mathbf C\langle x, \p_x \rangle$ denote the polynomial ring in $x$ and $\p_x$ with complex coefficients. The Fourier transform $\mathcal F$, which is a ring isomorphism of $D$, is defined by (Section 6.10 of [@hibi_book_13]) $$\begin{aligned}
\mathcal F : x\mapsto i\p_x,\qquad
\mathcal F : \p_x\mapsto ix,\qquad
(i=\sqrt{-1}).
\label{eq:fourier-corespondence}\end{aligned}$$ Hence the Fourier transform $\hat L(x,\p_x)$ of $L(x,\p_x)\in D$ is given by $L(i\p_x, ix)$.
This definition is based on the fact that if a function $f(x)$ satisfies the differential equation $L(x,\p_x) \bullet f(x) = 0$, then the Fourier transform $\hat f(\xi) = \int_{-\infty}^\infty e^{-ix\xi} f(x)\, dx$ satisfies the differential equation $\hat L(x,\p_x) \bullet \hat f(x)=0$ under some regularity conditions. If $f$ is a rapidly decreasing holonomic function, then the correspondence is immediate (Section 5.1.4 of [@stein_book]). The correspondence can be justified in the class of slowly increasing functions. See Chapter 5 of [@grubb_book].
We take $v_i(x)$’s as coprime polynomials in and then take the Fourier transform. By the correspondence , the highest degree of the coefficient polynomials of $L$ equals the order of $\hat L$. Hence we have the following corollary.
\[cor:con\] Under the condition of Theorem \[thm:deg\], there exists a holonomic differential equation satisfied by the $n$-th convolution of $\mathcal F[f(x)]$ whose order is less than or equal to the right-hand side of : $$\max\{M_0+(n-1)M_1,\ nM_1,\ 0 \} - \min\{ m_0,\ m_1 \} - (n-1)m_1.$$
Exponents for the differential equation of the $n$-th power and the Fourier transformed equation {#subsec:exponents}
------------------------------------------------------------------------------------------------
Consider an $r$-th order differential equation $$\Big[ (x-a)^r \p_x^r + (x-a)^{r-1} b_{r-1}(x)\p_x^{r-1} + \dots + (x-a) b_1(x)\p_x + b_0(x) \Big] \bullet f(x) = 0. \label{eq:r-th-equation}$$ If $b_0(x),\dots,b_{r-1}(x)$ are all analytic at $x=a$, then $a$ is said to be a regular singular point for the equation. If $a=0$ and $b_0(1/x),\dots,b_{r-1}(1/x)$ are all analytic at $x=0$, then $\infty$ is said to be a regular singular point for the equation.
When the equation is holonomic, $a$ is a regular singular point if the denominators of $b_0(x),\dots,b_{r-1}(x)$ do not have a factor $(x-a)$, and $\infty$ is a regular singular point if $a=0$ and $b_0(x),\dots,b_{r-1}(x)$ are all proper.
When $x_0 \in \mathrm C \cup \{\infty\}$ is a regular singular point for the equation, the $r$-th degree equation $$\begin{aligned}
b(\lambda) = [\lambda]_r + b_{r-1}(x_0) [\lambda]_{r-1} + \dots + b_1(x_0)[\lambda]_1 + b_0(x_0) = 0,\end{aligned}$$ where $[\lambda]_i=\lambda(\lambda-1)\dots(\lambda-i+1)$ (cf. ), is called the [*indicial equation*]{} (Section 9.5 of [@hille_book], Chapter 15 of [@ince_book]) for relative to the regular singular point $x_0$. The roots of the indicial equation are called the [*exponents*]{}.
The case $x_0\neq \infty$ can be reduced to the case $x_0=0$ by the transform $x-x_0 \mapsto x$ and the case $x_0 = \infty$ can be reduced to $x_0=0$ by $x \mapsto 1/x$. Hence in the following we put $x_0=0$.
The equation is equal to $$\Big[ [\theta_x]_r + b_{r-1}(x) [\theta_x]_{r-1} + \dots + b_1(x)[\theta_x]_1 + b_0(x) \Big] \bullet f(x) = 0,$$ where $\theta_x=x\p_x$ is the Euler operator, since $x^k\p_x^k = [\theta_x]_k$. This shows that $b(\lambda)$ is obtained by expressing the differential equation in terms of $x$ and $\theta_x$, and substituting $x=0$ and $\theta_x = \lambda$ formally.
In this subsection we assume that $x_0 \in \mathrm C \cup \{\infty\}$ is a regular singular point for the equation for $f(x)$. Let $\lambda_1,\lambda_2$ be the exponents for relative to the regular singular point $x_0$.
We show the following lemma on the eigenvalues of a matrix before the proof of Theorem \[thm:exponents\] on the exponents for relative to $x_0$.
\[lem:eig\] The eigenvalues of an $(n+1)\times(n+1)$ tridiagonal matrix $$\begin{aligned}
M =
\bma
0 & -\lambda_1\lambda_2 \\
n & \lambda_1+\lambda_2 & -2\lambda_1\lambda_2 \\
& n-1 & 2(\lambda_1+\lambda_2) & \ddots \\
&& \ddots & \ddots & -(n-1)\lambda_1\lambda_2 \\
&&& 2 & (n-1)(\lambda_1+\lambda_2) & -n\lambda_1\lambda_2\\
&&&& 1 & n(\lambda_1+\lambda_2)
\ema\end{aligned}$$ are $$(n-k) \lambda_1 + k \lambda_2, \qquad (k=0,1,\dots,n).$$
The eigenvalues of $M$ are equal to those of the matrix $$\begin{aligned}
&M' =
\bma
0 & \lambda_1 \\
-n\lambda_2 & \lambda_1+\lambda_2 & 2\lambda_1 \\
& -(n-1)\lambda_2 & 2(\lambda_1+\lambda_2) & \ddots \\
&& \ddots & \ddots & (n-1)\lambda_1 \\
&&& -2\lambda_2 & (n-1)(\lambda_1+\lambda_2) & n\lambda_1\\
&&&& -\lambda_2 & n(\lambda_1+\lambda_2)
\ema
.\end{aligned}$$ because the determinant of a tridiagonal matrix $T=(t_{ij})$ depends only on the diagonal elements $t_{ii}$ and the products of off-diagonal elements $t_{i,i+1}t_{i+1,i}$.
If $\lambda_1=0$, it is obvious that the eigenvalues are $0,\ \lambda_2,\ \dots,\ n\lambda_2$. Otherwise, putting $z=\lambda_2/\lambda_1$, we prove that the eigenvalues of the matrix $M'/\lambda_1$ are $\mu_k = kz+(n-k)\ (k=0,1,\dots,n)$.
For $z\neq 1$, all of $\mu_k$’s are different. We show that the eigenvector corresponding to $\mu_k$ is $\bm v^k =(v_l^k)_{0\leq l \leq n}$ where $$\begin{aligned}
v_l^k
= \sum_{j} \binom{n-k}{l-j} \binom{k}{j} z^j. \label{v_l^k}\end{aligned}$$ Here, the summation for $j$ is over the finite interval $\max\{0,\ k+l-n \} \leq j \leq \min\{ k,\ l \}$.
The $l$-th entry $(0\leq l \leq n)$ of $(\mu_k I-M'/\lambda_1)\bm v^k$ equals $$\begin{aligned}
\sum_j \Bigg[
&(n-l+1)\binom{n-k}{l-j}\binom{k}{j-1}
+ (n-k-l)\binom{n-k}{l-j}\binom{k}{j}\\
&\qquad + (k-l) \binom{n-k}{l-j+1}\binom{k}{j-1}
- (l+1) \binom{n-k}{l-j+1}\binom{k}{j}\Bigg] z^j. \label{sumbin}\end{aligned}$$ The first two terms equal $$\begin{aligned}
&\quad \Bigg[ (k+1)\binom{k}{j-1} + (n-k-l)\binom{k+1}{j} \Bigg] \binom{n-k}{l-j} \\
& = (n-k-l+j)\binom{n-k}{l-j} \binom{k+1}{j}
=(n-k)\binom{n-k}{l-j+1}\binom{k+1}{j},\end{aligned}$$ by the relations $\binom{n}{k-1}+\binom{n}{k}=\binom{n+1}{k}$, $k\binom{n}{k}=n\binom{n-1}{k-1}$ and $\binom{n}{k}=\binom{n}{n-k}$. Similarly, the last two terms equal $-(n-k)\binom{n-k}{l-j+1}\binom{k+1}{j}$. Those show that $(\mu_k I-M'/\lambda_1)\bm v^k = \bm 0$.
For $z=1$, all of $\mu_k$’s are identical. Let $\bm v^k =(v_l^k)_{0\leq l \leq n}$ $(k=0,1,\dots,n)$ be $(n+1)$ dimensional vectors where $$\begin{aligned}
v_l^k = \binom{n-k}{l}, \quad
(0\leq l \leq n-k),\qquad
v_l^k = 0,\quad
(n-k < l \leq n).\end{aligned}$$ Then, we can show $(nI-M'/\lambda_1)\bm v^0=\bm 0$ and $(nI-M'/\lambda_1)\bm v^k = k \bm v^{k-1}$ $(k=1,\dots,n)$ as above. Hence $\bm v^k$’s, which are linearly independent, are the generalized eigenvectors of the matrix.
can be formally written as follows: $$\begin{aligned}
v_l^k
= \binom{n-k}{l} \,{}_2F_1(-k,-l,n-k-l+1;z),\end{aligned}$$ where $$\begin{aligned}
{}_2F_1(a,b,c;z)
= \sum_{n=0}^\infty \frac{(a)_n (b)_n}{(c)_n} \frac{z^n}{n!},\qquad
(a)_n = \prod_{k=0}^{n-1} (a+k).\end{aligned}$$ Then, the $l$-th entry $(0\leq l \leq n)$ of $(\mu_k I-M'/\lambda_1)\bm v_k = \bm 0$ is equivalent to $$\begin{aligned}
& c(c-1) \, {}_2F_1(-k,-l-1,c-1;z)
- c\big( c-1+(k-l)z \big) \, {}_2F_1(-k,-l,c;z)\notag\\
&\qquad\qquad\qquad -l(c+k) z\,{}_2F_1(-k,-l+1,c+1;z) = 0,
\label{eq:2f1-recursion}\end{aligned}$$ where $c=n-k-l+1$. The recursion can be confirmed by `HolonomicFunction` ([@hfs]), a package of [Mathematica]{}.
We now show the following theorem on the exponents for .
\[thm:exponents\] If $x_0$ is a regular singular point for , then $x_0$ is a regular singular point for . Moreover, its exponents for are $$\begin{aligned}
(n-k) \lambda_1 + k \lambda_2, \qquad (k=0,1,\dots,n).\end{aligned}$$
We put $x_0=0$ without loss of generality by translation. Then, the equation can be rearranged to $$\Big[ \theta_x^2 - b_1(x)\theta_x - b_0(x) \Big] \bullet f(x)
= 0,
$$ where $b_0(x) = x^2a_0(x)$ and $b_1(x) = xa_1(x)+1$ are analytic at $x=0$. Let $$B(x)=
\bma
0 & b_0(x) \\
n & b_1(x) & 2b_0(x) \\
& n-1 & 2b_1(x) & \ddots \\
&& \ddots & \ddots & nb_0(x)\\
&&& 1 & nb_1(x)
\ema$$ be an $(n+1)\times(n+1)$ tridiagonal matrix and let $\tilde B(x,\theta_x) = B(x)+\theta_x I$, similarly to and . Then, as in we have $$\theta_x^k \bullet f^n
= (f^n ,\ f^{n-1} (\theta_x \bullet f),\ \dots,\ f(\theta_x \bullet f)^{n-1}, \ (\theta_x \bullet f)^n)\ \tilde B(x,\theta_x)^k \bullet \bm q_0.
\label{theta k}$$ Let $$\begin{aligned}
P(x) = (p_{ij}(x))_{\substack{0\leq i \leq n\\0 \leq j \leq n+1}} = (\bm q_0,\ \tilde B(x,\theta_x) \bullet \bm q_0,\ \dots,\ \tilde B(x,\theta_x)^{n+1} \bullet \bm q_0)\end{aligned}$$ be an $(n+1)\times(n+2)$ matrix and let $\bm w(x)=(w_0(x),\ \dots,\ w_n(x),\ 1) \in \Ker P(x)$. Then, the differential equation $$\begin{aligned}
\Big[
\theta_x^{n+1} + w_n(x) \theta_x^n + \dots + w_1(x) \theta_x + w_0(x)
\Big]\bullet f(x)^n
=0\end{aligned}$$ is equal to .
Every entry of $P(x)$ is analytic at $x=0$ since $b_0(x)$ and $b_1(x)$ are analytic. Moreover $w_0(x),\dots,w_n(x)$ are all analytic because each of $p_{ii}(x)$ is a constant and $w_{n+1}(x)=1$. Thus, $x_0=0$ is a regular singular point for since $\theta_x^k$ is a linear combination of $1,x\p_x,\dots,x^k\p_x^k$. Furthermore, the indicial equation for is $$\begin{aligned}
\lambda^{n+1} + w_n(0) \lambda^n + \dots + w_1(0) \lambda + w_0(0) = 0.\end{aligned}$$
On the other hand, since $$\begin{aligned}
\tilde B(x,\theta_x)^k \bullet \bm q_0 \Big|_{x=0}
&= \tilde B(x,\theta_x)^{k-1} \bullet B(0) \bm q_0 \Big|_{x=0}
= \cdots = \tilde B(x,\theta_x) \bullet B(0)^{k-1} \bm q_0 \Big|_{x=0}
= B(0)^k \bm q_0,\end{aligned}$$ we have $$\begin{aligned}
P(0)
= (\bm q_0,\ B(0)\bm q_0,\ \dots,\ B(0)^{n+1}\bm q_0).\end{aligned}$$ Hence, by $P(0)\bm w(0)=\bm 0$, we obtain $$\begin{aligned}
\Big( B(0)^{n+1} + w_n(0) B(0)^n + \dots + w_1(0) B(0) + w_0(0)I \Big) \bm q_0 = \bm 0.\end{aligned}$$ Then by the Cayley–Hamilton theorem and by the uniqueness of $\bm w(0)$, the characteristic equation of the matrix $B(0)$ is equal to the indicial equation for .
On the other hand, by Vieta’s formula, we have $b_0(0) = -\lambda_1\lambda_2$ and $b_1(0) = \lambda_1 + \lambda_2$. Hence the matrix $B(0)$ is equal to the matrix $M$ in Lemma \[lem:eig\]. Thus the exponents for are proved to be $(n-k)\lambda_1+k\lambda_2$ $(k=0,1,\dots,n)$.
We have described the exponents for the differential equation satisfied by $f(x)^n$. From now on, we investigate the exponents for the Fourier transformed equation.
Consider a differential equation $$\begin{aligned}
L\bullet f(x) = 0,\qquad
L = p_r(x) \p_x^r +p_{r-1}(x) \p_x^{r-1} + \dots + p_1(x)\p_x + p_0(x),\end{aligned}$$ where $p_0(x),\dots,p_r(x)$ are coprime polynomials. Let $d$ be the degree of $p_r(x)$. We assume that $\deg p_k(x) \leq d$ $(k=0,1,\dots,r-1)$.
By the definition of a regular singular point, if $x=0$ is a regular singular point, then $d\leq r$. Similarly, if $x=\infty$ is a regular singular point, then $d\geq r$.
If $x=0$ is a regular singular point, the main terms of the differential equation in the neighborhood of $x=0$ are $x^d\p_x^r,\ x^{d-1}\p_x^{r-1},\ \dots,\ \p_x^{r-d}$. Because of the relation $x^a\p_x^b = x^{a-b}\theta_x(\theta_x-1)\cdots(\theta_x-b+1)$, the indicial equation has $0,\ 1,\ \dots,\ r-d-1$ as its roots.
The regular singular point and its exponents are transformed by the Fourier transform as follows.
Suppose that $p_r(x)=x^d$, $\deg p_k(x) \leq d$ $(k=0,1,\dots,r-1)$, $x=0$ is a regular singular point for the equation $L\bullet f(x)=0$ and its exponents are $\mu_1,\dots,\mu_d$ and $0,\ 1,\ \dots,\ r-d-1$. Then $x=\infty$ is a regular singular point for the Fourier transformed equation $\hat L \bullet \hat f(x) = 0$ and its exponents are $-\mu_1-1,\ -\mu_2-1,\ \dots,\ -\mu_d-1$.
Suppose that $\deg p_k(x) \leq \deg p_r(x)$ $(k=0,1,\dots,r-1)$, $x=\infty$ is a regular singular point for the equation $L\bullet f(x)=0$ and its exponents are $\mu_1,\dots,\mu_r$. Then $x=0$ is a regular singular point for the transformed equation and its exponents are $-\mu_1-1,\ -\mu_2-1,\ \dots,\ -\mu_r-1$ and $0,\ 1,\ \dots,\ d-r-1$.
By the assumption there exists a term with the highest degree both in $x$ and $\p_x$. By the Fourier transform $x\mapsto i\p_x,\ \p_x \mapsto ix$, the highest degree is not changed and their weights are reversed. Hence if $x=0\ (\infty)$ is a regular singular point for $L\bullet f=0$, then $x=\infty\ (0)$ is a regular singular point for $\hat L\bullet \hat f=0$, and the main terms in the neighborhood of the singular point are not changed by the Fourier transform.
If $x=0$ is a regular singular point for $L\bullet f(x)=0$, its main term equals $x^{-(r-d)}\theta_x(\theta_x-1)\cdots(\theta_x-(r-d)+1)(\theta_x-\mu_1)\cdots(\theta_x-\mu_d)$. The main term is Fourier transformed to $$\begin{aligned}
&\quad \p_x^{-(r-d)}(\theta_x+1)(\theta_x+2)\cdots(\theta_x+(r-d))(\theta_x+\mu_1+1)\cdots(\theta_x+\mu_d+1)\\
&= x^{r-d}(\theta_x+\mu_1+1)\cdots(\theta_x+\mu_d+1),\end{aligned}$$ by the formula $\p_x^k x^k = (\theta_x+1)(\theta_x+2)\cdots(\theta_x+k)$. This gives the exponents for the Fourier transformed equation at $x=\infty$.
If $x=\infty$ is a regular singular point for $L\bullet f(x)=0$, its main term equals $x^{d-r}(\theta_x-\mu_1)\cdots(\theta_x-\mu_r)$. The main term is Fourier transformed to $$\begin{aligned}
&\quad \p_x^{d-r}(\theta_x+\mu_1+1)\cdots(\theta_x+\mu_r+1)\\
&= x^{-(d-r)}\theta_x(\theta_x-1)\cdots(\theta_x-(d-r)+1)(\theta_x+\mu_1+1)\cdots(\theta_x+\mu_r+1),\end{aligned}$$ by the formula $x^k\p_x^k = \theta_x(\theta_x-1)\cdots(\theta_x-k+1)$. This gives the exponents for the Fourier transformed equation at $x=0$.
Applications to statistics {#sec:applications}
==========================
Sum of beta random variables {#subsec:beta}
----------------------------
Let $f_n(a,b;x)$ be the probability density function of sum of $n$ beta random variables $\mathrm{Beta}(a,b)$. The moment generating function of the beta random variable is $$\begin{aligned}
M(a,b;t)
= \frac{\Gamma(a+b)}{\Gamma(a)\Gamma(b)} \int_0^1 e^{tu} u^{a-1}(1-u)^{b-1}du.\end{aligned}$$ Since $M(a,b;t)$ equals the confluent hypergeometric function ${}_1F_1(a,a+b;t)$ (c.f. [@NIST_book_10]), the characteristic function $\phi(a,b;t) = M(a,b;it)$ satisfies the following second order differential equation: $$\begin{aligned}
\bigg[ \p_t^2 - \Big(i-\frac{a+b}{t} \Big)\p_t - \frac{ia}{t} \bigg]
\bullet \phi(a,b;t) = 0,\qquad
(i=\sqrt{-1}).
\label{ode beta}\end{aligned}$$
An $(n+1)$-th order differential equation satisfied by $M(a,b;t)^n$ is derived by Theorem \[thm:ode\], and by the Fourier transform, we obtain a holonomic differential equation satisfied by $f_n(a,b;x)$. By Corollary \[cor:con\], putting $m_0=M_0=-1,\ m_1=-1,\ M_1=0$, the equation for $f_n(x)$ is at most of the $n$-th order. In fact, the equation derived by the procedure of Section 2 is exactly of the $n$-th order.
We define the initial term of a formal power series, before the proposition on the order.
For a formal power series $f(x)= \sum_{k=0}^\infty c_n x^{\lambda + k}$ $(c_0\neq 0)$, we define $$\begin{aligned}
\ini f(x) = c_0 x^\lambda.\end{aligned}$$ We denote the matrix (vector) whose $(i,j)$ entry is $\ini f_{ij}(x)$ by $\ini F(x)$, where $F(x)=(f_{ij}(x))$.
\[prop:beta\] The differential equation for $f_n(a,b;x)$ derived by the procedure of Section 2 is of the $n$-th order.
We prove $$\begin{aligned}
\ini q_{0,j}(t) = i(-1)^j n\frac{a}{b}(a+b)_{j-1} t^{-(j-1)},\quad
\ini q_{1,j}(t) = (-1)^{j-1} n (a+b)_{j-1}t^{-(j-1)},\end{aligned}$$ for $j=2,3,\dots,n+1$ by induction based on . It is easy to check them for $j=2$. Assuming them up to $j$, we have $$\begin{aligned}
&\ini q_{0,j+1}(t)
= \p_t q_{0,j}(t) + a_0(t) \ini q_{1,j}(t)
= i(-1)^{j+1} n \frac{a}{b} (a+b)_{j}t^{-j},\\
&\ini q_{1,j+1}(t)
= (\p_t - bt^{-1}) \ini q_{1,j}(t)
= (-1)^{j} n (a+b)_j t^{-j},\end{aligned}$$ since $\mindeg q_{2,j} \geq -(j-2)$ by . Thus, the results are shown by induction.
Hence, we obtain $\mindeg q_{0,n+1}(t) = -n$. As in the proof of Theorem \[thm:deg\], by clearing the denominators, we see that the highest degree of $t$ of the equation for $\phi(a,b;t)^n$ is $n$.
\[ex:f\_3\] $f_3(a,b;x)$ satisfies the differential equation $$\begin{aligned}
&\Big[
x(x-1)(x-2)(x-3)\p_x^3 + \big(-6(a+b-2)x^3 + 2(16a + 11b - 27)x^2 -6(8a + 3b - 11)x\\
& + 18(a-1)\big)\p_x^2 + \big((a+b-2)(11(a+b)-18)x^2 - (48a^2 + 66ab + 18b^2 - 145a - 95b\\
& + 108)x + 3(a - 1)(15a + 12b - 22)\big)\p_x - (a+b-2)(2(a+b)-3)(3(a+b)-4)x\\
& + 3(a-1)(2(a+b)-3)(3(a+b)-4)
\Big] \bullet f_3(a,b;x) = 0.\end{aligned}$$
Note that $x=0$ is a regular singular point for , and its exponents are $0,1-(a+b)$ since $$\begin{aligned}
\p_t^2 - \Big(i-\frac{a+b}{t} \Big)\p_t - \frac{ia}{t}
= t^{-2} \theta_t(\theta_t -1+a+b) -it^{-1}(\theta_t + a).\end{aligned}$$ Hence we can obtain the exponents for the equation satisfied by $f_n(a,b;x)$ relative to regular singular point $\infty$. However, it is not informative since $f_n(a,b;x)$ has a compact support.
On the other hand, the equation in Example \[ex:f\_3\] has regular singular points at $x=0,1,2,3$. In general, the degree of the coefficient polynomial of the highest order term $\p_x^n$ is less than or equal to $n+1$. On the other hand, a differential equation satisfied by $f_n$, or the $n$-th convolution of $f_1$, has to have singular points at $x=0,1,\dots,n$, because $f_1$ has singular points at $x=0,1$. Therefore, the highest order term of the differential equation derived as mentioned above is $x(x-1)\cdots(x-n)\p_x^n$, and this implies that $x=0,1,\dots,n$ are all regular singular points.
Especially, in the case of $a=b=1$ (then, the beta distribution becomes the uniform distribution), the differential equation is simply $x(x-1)\cdots(x-n) \p_x^n \bullet f_n(1,1;x)=0$. This is because $\phi(1,1;t) = (e^{it}-1)/(it)$ and thus $$\begin{aligned}
&\quad 2\pi \mathcal F^{-1}[x(x-1)\cdots(x-n) \p_x^n \bullet f_n(1,1;x)]\\
&= -i\p_t(-i\p_t-1)\cdots(-i\p_t-n) \bullet (-1)^n (e^{it}-1)^n\\
&= -i\p_t(-i\p_t-1)\cdots(-i\p_t-(n-1)) \bullet (-1)^n n(e^{it}-1)^{n-1}\\
&=\cdots = -i\p_t \bullet (-1)^n n! = 0.\end{aligned}$$ This shows that $f_n(1,1;x)$ has to be a piece-wise $(n-1)$-th degree polynomial.
The exact form of $f_n(1,1;x)$ is given in Section 1.9 of [@feller_vol2_book]. For $n=2$, $f_2(1,1;x)$ is a continuous piece-wise linear function and $f_2\in C^0$. By induction it follows that $f_n(1,1;x)=\int_0^1 f_{n-1}(1,1;x-y)dy$ belongs to $C^{n-2}$. Hence we can put $f_n(1,1;x) = c_0x^{n-1}\ (0\leq x \leq1)$ by the smoothness at $x=0$. We can also put $f_n(1,1;x) = c_0x^{n-1} + c_1(x-1)^{n-1}\ (1\leq x \leq2)$ by the smoothness at $x=1$. In the same way, we can put $f_n(1,1;x)= \sum_{j=0}^{k} c_j (x-j)^{n-1}\ (k\leq x \leq k+1)$. By the smoothness at $x=n$, the $k$-th $(k=0,1,\dots,n-2)$ derivative of $g_n(x) = \sum_{j=0}^{n-1} c_j (x-j)^{n-1}$ at $x=n$ is zero. Moreover, we have $$\begin{aligned}
\sum_{j=0}^{n-1} \int_j^n c_j (x-j)^{n-1} dx
= \sum_{j=0}^{n-1} c_j\frac{(n-j)^n}{n}
= 1,\end{aligned}$$ because of $\int_0^n f_n(1,1;x) dx = 1$. Hence, $c_0,\dots,c_{n-1}$ satisfy the equation $$\begin{aligned}
\bma
n^{n} & (n-1)^{n} &\cdots & 1^{n}\\
n^{n-1} & (n-1)^{n-1} &\cdots & 1^{n-1}\\
\vdots & \vdots & \vdots & \vdots \\
n^1 & (n-1)^1 &\cdots & 1^1
\ema
\bma
c_0\\
c_1\\
\vdots\\
c_{n-1}
\ema
=
\bma n \\ 0\\ \vdots \\0 \ema.\end{aligned}$$ The matrix on the left is invertible by the Vandermonde determinant. Therefore, we can determine the probability density function $f_n(1,1;x)$.
The $k$-th moment of the beta random variable is $(\alpha)_k/(\alpha+\beta)_k$. The moments of sum of $n$ beta random variables are the coefficients of $$\begin{aligned}
\bigg(
\frac{(\alpha)_0}{(\alpha+\beta)_0} + \frac{(\alpha)_1}{(\alpha+\beta)_1}t^1 + \frac{(\alpha)_2}{(\alpha+\beta)_2}t^2 + \cdots
\bigg)^n.\end{aligned}$$ The probability density function can be approximated in terms of orthogonal polynomials by fitting the moments. However the information provided by the differential equation can not be easily derived from the moments.
Sum of cubes of standard normal random variables {#subsec:cube}
------------------------------------------------
In this section we study characteristic functions and probability density functions of sum of cubes of standard normal variables. Concerning the probability distribution of sample skewness from normal population, Geary ([@geary]) and Mulholland ([@mulholland]) give very detailed results. However the distribution of the sum of cubes of standard normal variables, which is a more basic quantity than the sample skewness, has not been studied in detail.
Let $f_n(x)$ denote the probability density function of sum of cubes of $n$ standard normal variables. The characteristic function of the cube of a standard normal variable is $$\begin{aligned}
\phi(t)
= \frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty e^{-x^2/2} e^{itx^3} \,dx,\qquad
(i=\sqrt{-1}).\end{aligned}$$ Let $$\begin{aligned}
I_j(t)
= \frac{i^j}{\sqrt{2\pi}}\int_{-\infty}^\infty x^j e^{-x^2/2} e^{itx^3}\,dx,\end{aligned}$$ for $j=0,1,2,\dots$. $I_j(t)$ satisfies the recursion $$\begin{aligned}
3t I_j(t)
&= \frac{i^{j-1}}{\sqrt{2\pi}}\int_{-\infty}^\infty x^{j-2} e^{-x^2/2} \, \Big(e^{itx^3}\Big)' \,dx
= -\frac{i^{j-1}}{\sqrt{2\pi}}\int_{-\infty}^\infty e^{itx^3} \,\Big(x^{j-2}e^{-x^2/2}\Big)' \,dx \\
&= I_{j-1}(t) + (j-2)I_{j-3}(t).\end{aligned}$$ Since $\phi(t)=I_0(t),\ \p_t\phi(t) = -I_3(t),\ \p_t^2\phi(t) = I_6(t)$, we can express $\p_t\phi(t), \p_t^2\phi(t)$ in terms of $I_0(t), I_1(t)$. By eliminating $I_0(t), I_1(t)$, we obtain the second-order differential equation for $\phi(t)$ $$\begin{aligned}
\Big[ 27t^3\p_t^2 + (81t^2+1)\p_t + 15t \Big]\bullet \phi(t)=0. \label{phi 1}\end{aligned}$$ This can be also derived by integration algorithm ([@oaku1997]).
A differential equation satisfied by $\phi(t)^n$ is derived by Theorem \[thm:ode\], and by Fourier transform, we obtain a differential equation satisfied by $f_n(x)$. In Corollary \[cor:con\], putting $m_0=M_0=-2,\ m_1=-3,\ M_1=-1$, the equation for $f_n(x)$ is at most of the $3n$-th order. In fact, it is exactly of the $3n$-the order.
The differential equation for $f_n(x)$ derived by the procedure of Section 2 is of the $3n$-th order.
As in the proof of Proposition \[prop:beta\], we can show $$\begin{aligned}
\ini q_{0,j}(t) = (-1)^{j-1} 5n(3t)^{-3(j-1)+1},\quad
\ini q_{1,j}(t) = (-1)^{j-1} n(3t)^{-3(j-1)},\end{aligned}$$ for $j=2,3,\dots,n+1$. Hence, we obtain $\mindeg q_{0,n+1}(t) = -3n$, and thus the highest degree of $t$ of the equation is $3n$.
$f_4(x)$ satisfies the differential equation $$\begin{aligned}
&\Big[
177147 x^5\p_x^{12} +5314410 x^4\p_x^{11} +52455195 x^3\p_x^{10} +(65610 x^4+202242825 x^2)\p_x^9\\
&+ (1180980 x^3+278372295 x)\p_x^8+ (6145470 x^2+89579520)\p_x^7+ (8505 x^3+9950850 x)\p_x^6\\
&+ (76545 x^2+3408480)\p_x^5+155655 x\p_x^4 +(450 x^2+56160)\p_x^3 +1350 x\p_x^2 +480 \p_x+8 x
\Big]\\
& \bullet f_4(x)
=0.\end{aligned}$$
Note that $x=\infty$ is a regular singular point of the differential equation , and its exponents are $-1/3, -5/3$ since $$\begin{aligned}
27x^3\p_x^2 + (81x^2+1)\p_x + 15x
= 3x(3\theta_x+1)(3\theta_x+5) + x^{-1}\theta_x.\end{aligned}$$ This implies that there exists a differential equation satisfied by $\phi(t)^n$ which is regular at $x=\infty$ and its exponents are $-n/3,\ -(n+4)/3,\ \dots,\ -(5n-4)/3,\ -5n/3$. Moreover, there exists a differential equation satisfied by $f_n(x)$ which is regular at $x=0$ and its exponents are $n/3-1,\ (n+4)/3-1,\ \dots,\ (5n-4)/3-1,\ 5n/3-1$ and $0,\ 1,\ \dots,\ 2n-2$.
We now briefly discuss issues in numerical evaluation of $f_n$ based on our differential equation and computation of initial values. For numerically solving the differential equation satisfied by $f_n$, an initial value of $(f_n,\ \p_x \bullet f_n, \dots, \p_x^{3n-1} \bullet f_n)$ at $x = x_0 \neq 0$ is needed. Note that we can not use $x=0$ as the initial point, because it is the singular point of the differential equation.
By dividing the interval of integration of the inversion formula for the characteristic function and integrating by parts repeatedly, for any integer $m$ $(\geq 0)$, we obtain $$f_n(x)
= \frac{1}{\pi} \Re \bigg[ \int_{0}^T \phi_n(t) e^{itx} \,dt
+ e^{iTx} \sum_{j=1}^{m} \phi_n^{(j-1)}(T) \bigg(\frac{i}{x}\bigg)^j
+ \bigg(\frac{i}{x}\bigg)^m \int_T^\infty \phi_n^{(m)}(t) e^{itx}\,dt \bigg],
$$ where $\phi_n(t) = \phi(t)^n$. Integration by parts is needed for numerical evaluation of derivatives of $f_n$.
The formal $k$-th derivative of $f_n(x)$ at $x=x_0$ is $$\begin{aligned}
f_n^{(k)}(x_0)
&= \frac{1}{\pi} \Re \bigg[ \int_{0}^T (it)^k \phi_n(t) e^{ix_0t} \,dt
+ e^{iTx_0} \sum_{j=1}^{m} \phi_n^{(j-1)}(T) \sum_{l=0}^k \binom{k}{l} \frac{(j+l-1)!}{(j-1)!} \frac{i^{k+j+l}}{x_0^{j+l}}T^{k-l}\notag\\
&\qquad + \sum_{l=0}^k \binom{k}{l} \frac{(m+l-1)!}{(m-1)!} \frac{i^{m+k+l}}{x_0^{m+l}} \int_T^\infty t^{k-l} \phi_n^{(m)}(t) e^{ix_0t}\,dt
\bigg]. \label{k-th}\end{aligned}$$
Define $$\begin{aligned}
n!m = \prod_{0\leq k < n/m} (n-km),\ (n>0),\qquad
n!m = 1,\ (n\leq 0).\end{aligned}$$
\[lem:exp\] The following expansion of $\phi(t)$ holds at any $t\ (\neq0)$: $$\begin{aligned}
\phi(t)
&= \frac{\sqrt{2\pi}}{3\Gamma(2/3)}|t|^{-1/3}
\sum_{k=0}^\infty \frac{(6k-5)!6}{(6k)!6\,(6k-4)!6}(3t)^{-2k}
-\frac{\sqrt{2\pi}}{9\Gamma(1/3)}|t|^{-5/3}
\sum_{k=0}^\infty \frac{(6k-1)!6}{(6k)!6\,(6k+4)!6} (3t)^{-2k}.
$$
Airy function $\rm{Ai}(x)$ can be written as follows ([@copson], [@vallee], Chapter 9 of [@NIST_book_10]): $$\mathrm{Ai}(x)
= \frac{1}{\pi} e^{-\frac{2}{3}x^{3/2}} \int_0^\infty e^{-\sqrt{x}u^2} \cos\big( u^3/3\big) \,du \qquad(x>0).
$$ By the transform $u=2^{-1/2}x^{-1/4}s$, we obtain $$\begin{aligned}
\mathrm{Ai}(x)
&= \frac{1}{\sqrt{2}\pi}x^{-1/4} e^{-\frac{2}{3}x^{3/2}} \int_0^\infty e^{-s^2/2} \cos\bigg( \frac{x^{-3/4}}{2^{3/2}} \frac{s^3}{3} \bigg) \,ds.\end{aligned}$$ From this and $$\begin{aligned}
\phi(t) = \sqrt{\frac{2}{\pi}}\int_0^\infty e^{-x^2/2} \cos(tx^3)\, dx,\end{aligned}$$ we obtain the following relation between $\phi(t)$ and $\rm{Ai}(t)$: $$\begin{aligned}
\phi(t)
&= \frac{\sqrt{2\pi}}{3^{1/3}} t^{-1/3} e^{t^{-2}/108} \mathrm{Ai}\bigg( \frac{t^{-4/3}}{4\cdot 3^{4/3}}\bigg)
\qquad(t>0). \label{ai}\end{aligned}$$ Since the Maclaurin expansion of $\rm{Ai}(t)$ is known, the expansion of $\phi(t)$ in $1/t$ can be derived.
By we see that $\phi(t) = O(t^{-1/3})\ (t\to\infty)$. Hence, is justified for $k<m+n/3$ since the integral on $[T,\infty)$ converges for $k<m+n/3$.
In the integral on $[0,T]$ can be numerically computed by using the relation .
The integral on $[T,\infty)$ can be computed without numerical integral. The values of $J_l = \int_{T'}^\infty e^{it}/t^{l/3} \, dt \ (l=1,2,\dots)$, where $T' = Tx_0$, is needed for computation of the integral on $[T,\infty)$. $J_l$ satisfies the following recurrence relation: $$\begin{aligned}
J_{l+3}
= \frac{3}{l} \bigg( \frac{e^{iT'}}{{T'}^{l/3}} + i J_l \bigg),\end{aligned}$$ and thus it is sufficient to compute $J_1$, $J_2$ and $J_3$. $J_1$ and $J_2$ can be reduced to the integral on $[0,T]$ from the formulae (cf. Section 5.9 of [@NIST_book_10]) $$\begin{aligned}
\int_0^\infty \frac{\sin x}{x^p}\,dx = \frac{\pi}{2\Gamma(p)\sin(p\pi/2)},\qquad
\int_0^\infty \frac{\cos x}{x^p}\,dx = \frac{\pi}{2\Gamma(p)\cos(p\pi/2)}\qquad(0<p<1).\end{aligned}$$ $J_3$ can be computed by Maclaurin expansion of trigonometric integrals.
The second term of the right side of can be computed by Lemma \[lem:exp\].
From the above, the value of $f_n^{(k)}(x_0)\ (n=0,1,\dots,3n-1)$ can be computed and the differential equation satisfied by $f_n(x)$ can be numerically solved.
Some discussions {#sec:discussion}
================
In this paper we investigated properties of powers of functions satisfying a second-order holonomic differential equation. Our motivating example was the distribution of convolutions of cubes of standard normal random variables presented in Section \[subsec:cube\], which was in turn motivated by the algorithm given in [@Oaku2013]. In the course of our study of distribution of cubes of standard normal random variables, we noticed some remarkable properties satisfied by the characteristic function of the cube of a standard normal random variable. Based on this example, we developed more general theory presented in Section \[sec:main\], which may be relevant to problems in other areas of applied mathematics.
From a mathematical viewpoint, it is of interest to generalize the results of Section \[sec:main\] to the case of powers of a general holonomic function. From a statistical viewpoint, it is of interest to investigate the distribution of the sum of the $r$-th power ($r\ge 4$) of standard normal random variables.
Acknowledgment {#acknowledgment .unnumbered}
==============
We are grateful to C. Koutschan for computation of .
[^1]: Graduate School of Information Science and Technology, University of Tokyo
[^2]: Department of Mathematics, Tokyo Woman’s Christian University
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For many known non-compact embeddings of two Banach spaces $E\hookrightarrow F$, every bounded sequence in $E$ has a subsequence that takes form of a *profile decomposition* - a sum of clearly structured terms with asymptotically disjoint supports plus a remainder that vanishes in the norm of $F$. In this paper we construct a profile decomposition for arbitrary sequences in the Sobolev space $H^{1,2}(M)$ of a Riemannian manifold with bounded geometry, relative to the embedding of $H^{1,2}(M)$ into $L^{2^*}(M)$, generalizing the well-known profile decomposition of Struwe [@Struwe Proposition 2.1] to the case of any bounded sequence and a non-compact manifold.'
address:
- 'TIFR Centre for Applicable Mathematics, Post Bag no. 6503,Sharada nagar,Yelahanka New Town, Bangalore-560065,India.'
- 'Sankt Olofsgatan 66B, 75 330 Uppsala, Sweden'
author:
- Kunnath Sandeep
- Cyril Tintarev
title: 'Profile decomposition of Struwe-Solimini for manifolds with bounded geometry'
---
[^1]
Introduction
============
*Defect of compactness*, relative to an embedding of two Banach spaces $E\hookrightarrow F$, is a difference $u_{k}-u$ between a weakly convergent sequence $u_{k}\rightharpoonup u$ in $E$ and its limit, taken up to a remainder that vanishes in the norm of $F$. In particular, if the embedding is compact and $E$ is reflexive, the defect of compactness is null. For many embeddings there exist well-structured descriptions of the defect of compactness known as *profile decompositions*. In particular, profile decompositions relative to Sobolev embeddings are sums of terms with asymptotically disjoint supports, which are called *elementary concentrations* or *bubbles.* Profile decompositions were motivated by studies of concentration phenomena in the early 1980’s by Uhlenbeck, Brezis, Coron, Nirenberg, Aubin and Lions, and they play significant role in verification of convergence of functional sequences in applied analysis, particularly when the information available via the classical concentration compactness method is not enough detailed.
Profile decompositions are known when the embedding $E\hookrightarrow F$ is *cocompact* relative to some group $\mathcal{G}$ of bijective isometries on $E$ (embedding $E\hookrightarrow F$ is called $\mathcal{G}$ -cocompact if any sequence $(u_{k})$ in $E$ satisfying $g_{k}u_{k}\rightharpoonup0$ for any sequence of operators $(g_{k})$ in $\mathcal{G}$ vanishes in the norm of $F$). It is easy to show (the example is due to Jaffard, [@Jaffard]) that $\ell^{\infty}({\mathbb{Z}})$ is cocompactly embedded into itself relative to the group of shifts $\mathcal{G}=\lbrace(a_{n})\mapsto(a_{n+m})\rbrace_{m\in{\mathbb{Z}}}$: $g_{k}u_{k}\rightharpoonup0$ means in this case uniform convergence. The earliest cocompactness result for functional spaces known to the authors is the proof of cocompactness of embedding of the inhomogeneous Sobolev space $H^{1,p}({\mathbb{R}}^{N})$, $N>p$, into $L^{q}$, $q\in(p,p^{*})$, where $p^{*}=\frac{pN}{N-p}$, relative to the group of shifts $u\mapsto u(\cdot-y),\;y\in{\mathbb{R}}^{N}$, by E. Lieb [@LIeb], although expressed in very different terms (the term *cocompactness* appeared in literature only a decade ago).
A profile decomposition relative to a group $\mathcal{G}$ of bijective isometries expresses the defect of compactness as a sum of *elementary concentrations*, or *bubbles*, $\sum_{n\in{\mathbb{N}}}g_{k}^{(n)}w^{(n)}$ , with some $g_{k}^{(n)}\in\mathcal{G}$ and $w^{(n)}\in E$, $k\in{\mathbb{N}}$, $n\in{\mathbb{N}}$. The elements $w^{(n)}$, called *concentration profiles*, are then obtained as weak limits of $(g_{k}^{(n)})^{-1}u_{k}$ as $k\to\infty$. Typical examples of groups $\mathcal{G}$, involved in profile decompositions, are the above mentioned group of shifts and the rescaling group, which is a product group of shifts and dilations $u\mapsto t^{r}u(t\cdot)$, $t>0$, where $r=\frac{N-p}{p}$ for $\dot{H}^{1,p}({\mathbb{R}}^{N})$, $N>p$. For a smooth Riemannian manifold $M$, the Sobolev space $H^{1,2}(M)$ is defined as a completion of $C_{0}^{\infty}(M)$ with respect to the norm $$\|u\|^{2}=\int_{M}(|du|^{2}+|u|^{2})\mathrm{d}\mu,$$ where $d$ stays for the covariant derivative and $\mu$ for the Riemannian measure on $M$. We will discuss later also the matters concerning definition of a counterpart of the space $\dot{H}^{1,2}({\mathbb{R}}^{N})$ for manifolds. In what follows the unqualified notation of the norm will always refer to the $H^{1,2}(M)$-norm, and this extends to the notation of the corresponding scalar product. The $L^{p}$ norm for the manifold will be denoted as $\|u\|_{p}$.
This paper is motivated by a profile decomposition for sequences in Sobolev spaces of compact manifolds by Struwe [@Struwe Proposition 2.1], where the defect of compactness for the embedding $H^{1,2}(M)\hookrightarrow L^{2^{*}}(M)$ is a finite sum of bubbles of the form $t_{k}^{\frac{N-2}{2}}w\circ\exp_{y_k}^{-1}(t_{k}\exp_{y_k}(x))$, $t_{k}\to\infty$, $y_k\in M$, multiplied by a suitable cut-off function. Struwe’s result is restricted to Palais-Smale sequences of semilinear elliptic functionals, which assures that the number of bubbles is finite. Struwe’s argument for vanishing of the remainder in $L^{2^{*}}$ is also based on properties of Palais-Smale sequences.
The earliest known result on profile decompositions for general bounded sequences in $\dot{H}^{1,p}({\mathbb{R}}^{N})$ equipped with the rescaling group was proved by Solimini [@Solimini]. It was repeated later, and with a weaker form of asymptotics, in [@Gerard] and [@Jaffard] ([@Jaffard] also extended the result to fractional Sobolev spaces). Subsequently, profile decompositions were found in the general functional-analytic setting relative to a general group of isometries, for Hilbert spaces in [@SchinTin], and for uniformly convex Banach spaces with the Opial condition in [@SoliTi] (without the Opial condition weak convergence has to be replaced in the statement by the less-known Delta convergence). However, despite the general character of this result, some known profile decompositions do not follow directly from [@SoliTi], in particular, when the space $E$ is not reflexive ([@AT_BV]), when one has only a semigroup of isometries (e.g. [@AdiTi-Pisa]), or when the profile decomposition can be expressed without a group ([@DevT; @SkrTi4]). The latter papers are dealing with generalizations of the profile decomposition of Solimini to the case of Riemannian manifolds. Paper [@SkrTi4] proves profile decomposition for the embedding $H^{1,2}(M)\hookrightarrow L^{p}(M)$, $2<p<2^{*}$, when $M$ is a non-compact manifold of bounded geometry, where the loss of compactness is due to bubbles shifting to infinity (concentration compactness of type I in the framework of Lions), and paper [@DevT] deals with the case of compact manifold and $p=2^{*}$, where the loss of compactness is due to Struwe’s bubbles, thus generalizing Struwe’s profile decomposition to the case of general sequences. (This corresponds to concentration compactness of type II in the sense of Lions). However, the results of [@DevT] and [@SkrTi4] do not generalize the original profile decomposition for ${\mathbb{R}}^{N}$ to the case of manifolds in full: the former is restricted to compact manifolds and the latter is restricted to the case of subcritical exponent. The present paper gives an answer to the question if there is a profile decomposition for the embedding $H^{1,2}(M)\hookrightarrow L^{2^{*}}(M)$ with both shifts and Struwe’s bubbles. Our method of proof is different from [@DevT], whose argument does not yield an obvious generalisation to non-compact manifolds, while the result of [@SkrTi4] is employed here as an intermediate step.
Certain limitations have to be imposed on $M$, since the embedding $H^{1,2}(M)\hookrightarrow L^{2^{*}}(M)$ itself does not exist for every manifold. Following the reasoning of [@SkrTi4] we deal here only with manifolds of bounded geometry, which generate concentration profiles at infinty which are still functions in a Sobolev space of a smooth Riemannian manifold, although this manifold is generally not the original $M$.
The other limitation concerns the use of the full Sobolev norm and not the gradient norm of $\dot{H}^{1,2}$ as in [@Solimini]. For sequences bounded in $H^{1,2}({\mathbb{R}}^{N})$, the $L^{2}$-bound eliminates from the profile decomposition of Solimini any deflations (“reverse bubbles”) $t_{k}^{\frac{N-2}{2}}w(t_{k}(x-y_{k}))$ with $t_{k}\to0$, because their $L^{2}$-norms are unbounded. In comparison, on the hyperbolic space (or any space of negative curvature bound away from zero, see [@chavel-eigen]) the gradient norm dominates the $L^{2}$-norm, which suppresses appearance of deflations in the profile decomposition. On the other hand, on the sphere (as on any other compact manifold) quadratic form $\int_{M}|du|^{2}d\mu$ defines only a seminorm whose null space consists of constants. Appending it, for example, by $\left(\int_{M}u\,d\mu\right)^{2}$ yields a norm that also dominates the $L^{2}$-norm which eliminates deflations.
The example of $M={\mathbb{R}}^N$ shows that when $\dot{H}^{1,2}(M)\neq H^{1,2}(M)$, the defect of compactness may have to account also for some analog of “deflation” terms $t_k^\frac{N-2}{2}w(t_k\cdot)$, $t_k\to 0$, but while every manifold admits local “zoom-in” maps, there is no natural global “zoom out” map that could describe concentration of deflative character on every (approximately flat) manifold. The paper is organized as follows. In Section 2 we show that sequences bounded in $H^{1,2}(M)$, that already vanish in $L^{p}(M)$ for some $p\in(2,2^{*})$, vanish also in $L^{2^{*}}(M)$ if they satisfy a “no bubbles” condition. In Section 3 we show that sequences bounded in $H^{1,2}(M)$ that vanish in $L^{p}(M)$ for some $p\in(2,2^{*})$, have a subsequence that satisfies the “no bubbles” condition (and thus vanishes in $L^{2^{*}}(M)$ ) after one subtracts from it all its bubbles, and that the bubbles are asymptotically decoupled. In Section 4 we recall the result of [@SkrTi4] on the defect of compactness for the embedding $H^{1,2}(M)\hookrightarrow L^{p}(M)$, $p\in(2,2^{*})$, which says that a sequence bounded in $H^{1,2}(M)$ has a subsequence that, after subtraction of its weak limit and all suitably defined shift-concentration terms, vanishes in $L^{p}(M)$,$p\in(2,2^{*})$. Combining this result with the result of Section 3 we obtain the main result of this paper, Theorem \[thm:finalPD\], which says that the defect of compactness for the embedding $H^{1,2}(M)\hookrightarrow L^{2^*}(M)$ is a sum of a series of bubbles and a series of shift-concentrations, all of them mutually decoupled. Appendix summarizes some statements concerning manifolds of bounded geometry and manifolds at infinity used in the main body of the paper.
Vanishing lemma
===============
From now on we assume that $M$ is a smooth, complete, connected $N$-dimensional Riemannian manifold of bounded geometry. The latter property is defined as follows.
\[def:bg\](Definition A.1.1 from [@Shubin]) A smooth Riemannian manifold $M$ is of bounded geometry if the following two conditions are satisfied:
\(i) The injectivity radius $r(M)$ of $M$ is positive.
\(ii) Every covariant derivative of the Riemann curvature tensor $R^{M}$of M is bounded, i.e., $\nabla^{k}R^{M}\in L^{\infty}(M)$ for every $k=0,1,\dots$
We refer the reader to the appendix for elementary properties of manifolds of bounded geometry used in this paper, and existence of an appropriate covering of such manifolds.
In what follows $B(x,r)$ will denote a geodesic ball in $M$ and $\Omega_{r}$ will denote the ball in ${\mathbb{R}}^{N}$ of radius $r$ centered at the origin. Let $e_{x}:\Omega_{r}\to B(x,r)$ , $r<r(M)$. be an exponential map of $M$ under identification of the ball of radius $r$ centered at the origin in $T_{x}M$ as the ball $\Omega_{r}\subset{\mathbb{R}}^{N}$ (we reserve the standard notation $\mathrm{exp}_{x}$ for a standard exponential map $T_{x}M\to M$, so that $e_{x}=\exp_{x}\circ i_{x}$ where $i_{x}$ is an arbitrarily fixed linear isometry between ${\mathbb{R}}^{N}$ and $T_{x}M$).
We will deal first with sequences that are bounded in $H^{1,2}(M)$ and vanish in $L^{p}(M)$ for some $p\in(2,2^{*})$ (note that if this is the case for some value of $p$ in the interval than it is true for any other value in the interval). Such sequences may still have $L^{2^{*}}$-norm bounded away from zero. For example, let $v\in C_{0}^{\infty}(\Omega_{r})$ and let $u_{t}(x)=\sum_{n=1}^{m}t^{\frac{N-2}{2}}v(te_{y_{n}}^{-1}(x))$, where $y_{1},\dots,y_{m}$ are distinct points on $M$, $m\in{\mathbb{N}}$, $t>1$, and each term in the sum is understood as extended by zero outside of $B(y_{n},\Omega_{r/t}$). Then $u_{t}\to0$ in $L^{p}(M)$ for any $p\in(2,2^{*}),$ while the sequence is bounded in the $H^{1,2}$-norm and $\|u_{k}\|_{2^{*}}^{2^{*}}\to m\|v\|_{2^{*}}^{2^{*}}$. If we provide a profile decomposition for such sequence with the remainder vanishing in $L^{p}(M)$ for all $p\in(2,2^{*}],$ then we will have a profile decomposition for all sequences relative to the embedding $H^{1,2}(M)\hookrightarrow L^{2^{*}}(M)$, consisting of the sum of the profile decomposition relative to the embedding $H^{1,2}(M)\hookrightarrow L^{p}(M)$, $p\in(2,2^{*})$ (given in [@SkrTi4]) and our new profile decomposition for this remainder (which vanishes in $L^{p}(M)).$ For these ends we identify a sequence that vanishes in $L^{p}(M)$ and passes a “nonconcentration” test (relation (\[eq:noconc\]) below) as vanishing in $L^{2^{*}}(M).$
In what follows we will often consider sequences of the form $t_{k}^{\frac{N-2}{2}} u_{k}\circ e_{y_{k}}(t_{k}\cdot)$, $t_k\to 0$, which are defined on $\Omega_R\subset{\mathbb{R}}^N$ for every $R$, provided that $k$, dependently on $R$, is sufficiently large, which allows a natural definition of their pointwise limit on ${\mathbb{R}}^N$. It is easy to see that if $(u_k)$ is a bounded sequence in $H^{1,2}(M)$, then $(t_{k}^{\frac{N-2}{2}} u_{k}\circ e_{y_{k}}(t_{k}\cdot))$ converges pointwise on ${\mathbb{R}}^N$ to some $w$ if and only if for any $\varphi\in C_c({\mathbb{R}}^N)$, $\int_Md(t_{k}^{\frac{N-2}{2}} u_{k}\circ e_{y_{k}}(t_{k}\cdot))\cdot d\varphi d\mu\to \int_Mdw\cdot d\varphi d\mu$ as $k\to\infty$, which we for short will denote, slightly abusing the notation of weak convergence, as $u_k\rightharpoonup w$ or $w={\operatorname{w-lim}}u_k$. Moreover, for every sequence $(u_k)$ bounded in $H^{1,2}(M)$ and every sequence $(t_k)$ with $t_k\to 0$, the sequence $(t_{k}^{\frac{N-2}{2}} u_{k}\circ e_{y_{k}}(t_{k}\cdot))$ has a subsequence that converges pointwise and weakly in the above sense.
From now on we fix a positive number $r<r(M)$.
\[thm:vanishing\] Let $M$ be a complete smooth $N$-dimensional Riemannian manifold of bounded geometry, and let $(u_{k})$ be a bounded sequence in $H^{1,2}(M)$ such that for any sequence of points $y_{k}\in M$, and any sequence $t_{k}\to0$,
$$t_{k}^{\frac{N-2}{2}}u_{k}\circ e_{y_{k}}(t_{k}\xi)\longrightarrow0\mbox{ a.e. in }{\mathbb{R}}^{N}.\label{eq:noconc}$$
Furthermore, assume that $u_{k}\to0$ in $L^{p}(M)$ for some $p\in(2,2^{*})$. Then $u_{k}\to0$ in $L^{2^{*}}(M)$.
Step1. For any $u\in H^{1,2}(M)$ the following holds: $$\|u\|_{2^{*}}^{2^{*}}\le C\|u\|_{H^{1,2}}^{2}\sup_{j\in2^{\frac{N-2}{2}{\mathbb{Z}}}}\left(\int_{j\le|u(x)|\le2^{\frac{N-2}{2}}j}|u|^{2^{*}}\mathrm{d}\mu\right)^{\frac{2}{N}}.\label{eq:prange}$$ Indeed, let $\chi\in C_{0}^{1}(2^{-\frac{N-2}{2}},2^{N-2})$, extended by zero to $[0,\infty)$ be such that $\chi(s)\in[0,1]$ for all $s$ and $\chi(s)=1$ whenever $s\in[1,2^{\frac{N-2}{2}}]$. Let $\chi_{j}(s)=j\chi(j^{-1}s)$, ${\normalcolor j\in2^{\frac{N-2}{2}{\mathbb{Z}}}}$.
Applying Sobolev inequality to $\chi_{j}(|u|)$ we get
$$\left(\int_{j\le|u(x)|\le2^{\frac{N-2}{2}}j}|u|^{2^{*}}\mathrm{d}\mu\right)^{2/2^{*}}\le C\int_{2^{-\frac{N-2}{2}}j\le|u(x)|\le2^{N-2}j}(|du|^{2}+|u|^{2})\mathrm{d}\mu,$$ from which we have $$\int_{j\le|u(x)|\le2^{\frac{N-2}{2}}j}|u|^{2^{*}}\mathrm{d} \mu \ \ \le$$ $$C\int_{2^{-\frac{N-2}{2}}j\le|u(x)|\le2^{N-2}j}(|du|^{2}+|u|^{2})\mathrm{d}\mu\left(\int_{j\le|u(x)|\le2^{\frac{{N-2}}{2}}j}\mathrm{|u|^{2^{*}}d}\mu\right)^{1-\frac{2}{2^{*}}}.$$ Adding the inequalities over $j\in2^{\frac{N-2}{2}{\mathbb{Z}}}$ , while evaluating the last term by its upper bound, we get (\[eq:prange\]).
Step 2. Apply (\[eq:prange\]) to the sequence $u_{k}$.
Let $j_{k}\in2^{\frac{N-2}{2}{\mathbb{Z}}}$ be such that $$\sup_{j\in2^{\frac{N-2}{2}{\mathbb{Z}}}}\int_{j\le|u_{k}(x)|\le2^{\frac{N-2}{2}}j}|u|_{k}^{2^{*}}\mathrm{d}\mu\le2\int_{j_{k}\le|u_{k}(x)|\le2^{\frac{N-2}{2}}j_{k}}|u|^{2^{*}}\mathrm{d}\mu.\label{eq:nearmax}$$ Then we have from (\[eq:prange\]) $$\|u_{k}\|_{2^{*}}\le C\left(\int_{j_{k}\le|u(_{k}x)|\le2^{\frac{N-2}{2}}j_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu\right)^{\frac{2}{2^{*}N}}\label{eq:redux1}$$
It suffices to consider two cases: $j_{k}\le L$ for all $k$, $L\in{\mathbb{Z}}$, and $j_{k}\to+\infty$. In the first case we have from (\[eq:redux1\]) with any small $\varepsilon>0,$ $$\|u_{k}\|_{2^{*}}\le C\left(\int_{j_{k}\le|u_{k}(x)|\le2^{\frac{N-2}{2}}j_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu\right)^{\frac{2}{N2^{*}}}\le C\left(L^{2^{*}-p}\int_{M}|u_{k}|^{p}\mathrm{d}\mu\right)^{\frac{2}{N2^{*}}}$$ which converges to $0$ by the assumption on the sequence. Step 3. Consider now the second case, $j_{k}\to\infty$. For each $k\in{\mathbb{N}}$, let $Y_{k}\subset M$ be a discretization of $M$ given by Lemma \[lem:covering\], that is, a collection of balls $\lbrace B(y,t_{k}r)\rbrace_{y\in Y_{k}}$ with $t_{k}=j_{k}^{-\frac{2}{N-2}},$ form a covering for $M$ of uniformly finite multiplicity. Let $D_{k}=\lbrace x\in M:|u_{k}(x)|\in[j_{k},2^{\frac{N-2}{2}}j_{k}]\rbrace$ and $D'_{k}=\lbrace x\in M:|u_{k}(x)|\in[2^{-\frac{N-2}{2}}j_{k},2^{N-2}j_{k}]\rbrace$.
By scaling of the Sobolev inequality, applied to $\chi_{j_{k}}(|u_{k}|)$ on the geodesic ball $B(y,t_{k}r)$ there is a constant $C$ independent of $k$ such that $$\left(\int_{B(y,t_{k}r)\cap D_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu\right)^{2/2^{*}}\le C\int_{B(y,t_{k}r)\cap D'_{k}}(|du_{k}|^{2}+t_{k}^{-2}|u_{k}|^{2})\mathrm{d}\mu.$$ Taking into account that the integration domain in the right hand side is a subset of $D'_{k},$ we have $t_{k}^{-2}|u_{k}|^{2}\le C|u_{k}|^{2^{*}}$ uniformly in $k$ , and thus we have $$\begin{array}{c}
\int_{B(y,t_{k}r)\cap D_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu \ \le \\ \int_{B(y,t_{k}r)\cap D'_{k}}(|du_{k}|^{2}+|u_{k}|^{2^{*}})\mathrm{d}\mu\left(\int_{B(y,t_{k}r)\cap D_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu\right)^{\frac{2}{N}}.
\end{array}$$
Adding the inequalities above over $y\in Y_{k}$ and replacing the second term in the right hand side by its upper bound, we have $$\int_{D_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu\le C\sup_{y\in Y_{k}}\left(\int_{B(y,t_{k}r)\cap D_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu\right)^{\frac{2}{N}}.\label{eq:redux2}$$ Choosing points $y_{k}\in Y_{k}$ so that $$\sup_{y\in Y_{k}}\int_{B(y,t_{k}r)\cap D_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu\le2\int_{B(y_{k},t_{k}r)\cap D_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu,$$ we have from (\[eq:redux1\]) and (\[eq:redux2\]), $$\|u_{k}\|_{2^{*}}\le C\left(\int_{B(y_{k},t_{k}r)\cap D_{k}}|u_{k}|^{2^{*}}\mathrm{d}\mu\right)^{\frac{4}{2^{*}N^{2}}}.$$ Changing variables from a small ball on $M$ to small ball on ${\mathbb{R}}^{N}$ by the geodesic map $e_{y_{k}}$, and taking into account that $M$ is a manifold of bounded geometry, so that the estimate of the Jacobian is uniform in $k$, we have, with $\Delta_{k}=\lbrace\xi\in\Omega_{r}:|u_{k}\circ\exp_{y_{k}}(\xi)|\in(j_{k},2^{\frac{N-2}{2}}j_{k})\rbrace$, for all $k$ large enough $$\|u_{k}\|_{2^{*}}\le C\left(\int_{\Omega_{t_{k}r}\cap\Delta_{k}}|u_{k}\circ e_{y_{k}}(\xi)|^{2^{*}}\mathrm{d}\xi\right)^{\frac{4}{2^{*}N^{2}}}.$$ Let us now change variables again by setting $\xi=t_{k}\eta$, $\eta\in\Omega_{r}$: $$\|u_{k}\|_{2^{*}}\le C\left(\int_{\lbrace\eta\in\Omega_{r}:|u_{k}\circ e_{y_{k}}(t_{k}\eta)|\in[j_{k},2^{\frac{N-2}{2}}j_{k}]\rbrace}|t_{k}^{\frac{N-2}{2}}u_{k}\circ e_{y_{k}}(t_{k}\eta)|^{2^{*}}\mathrm{d}\eta\right)^{\frac{4}{2^{*}N^{2}}},$$ and note that the integrand converges to zero almost everywhere by assumption, and is bounded by the constant $2^{N}$, so by Lebesgue dominated convergence theorem the right hand side vanishes, which proves the theorem.
Profile decomposition for sequences vanishing in $L^{p}$, $p<2^{*}$.
====================================================================
We will start with a characterization of decoupling of bubbles involved in our profile decomposition.
We shall say that two sequences, $U_{k}$ and $V_{k}$, in a Hilbert space $H$ are asymptotically orthogonal, if $(U_{k},V_{k})\to0$.
Let us fix $\chi\in C_{0}^{\infty}(\Omega_{r})$ such that $\chi(\xi)=1$ whenever $|\xi|\le\frac{r}{2}$, extended by zero to a function on ${\mathbb{R}}^N$.
\[lem:aorth\]Let (with extension to $M\setminus B_{r}(y_{k})$ by zero): $$V_{k}(x)=2^{\frac{N-2}{2}j_{k}}\,\chi\circ e_{y_{k}}^{-1}(x)\,v(2^{j_{k}}e_{y_{k}}^{-1}(x)),\;x\in M,$$ $$W_{k}(x)=2^{\frac{N-2}{2}\ell_{k}}\,\chi\circ e_{z_{k}}^{-1}(x)\,w(2^{\ell_{k}}e_{z_{k}}^{-1}(x)),\;x\in M$$ where $j_{k},\ell_{k}\in{\mathbb{N}}$, $j_{k},\ell_{k}\to\infty$ and $y_{k},z_{k}\in M$. Then $V_{k}$ and $W_{k}$ are asymptotically orthogonal in $H^{1,2}(M)$ for any choice of $v$ and $w$ in $\dot{H}^{1,2}({\mathbb{R}}^{N})$ if and only if the following condition holds: $$\label{eq:orth3}
|\ell_k-j_k| \ + \ (2^{\ell_k}+2^{j_k}) d(y_k,z_k) \rightarrow \infty$$ as $k\rightarrow \infty.$
Furthermore, given $w\in\dot H^{1,2}({\mathbb{R}}^N)$, $(V_k,W_k)\to 0$ for any $v\in\dot H^{1,2}({\mathbb{R}}^N)$ if and only if $$\label{eq:zlm0}
2^{-j_{k}\frac{N-2}{2}}W_{k}\circ e_{y_{k}}(2^{-j_{k}}\xi)\to 0\mbox{ a.e. in }{\mathbb{R}}^N .$$
1\. Note that $\int_{M}V_{k}^{2}d\mu\to0$ and $\int_{M}W_{k}^{2}d\mu\to0$, so asymptotic orthogonality of $V_{k}$ and $W_{k}$ is equivalent to $\int_{M}dV_{k}\cdot dW_{k}d\mu\to0$. By density of $C_0^\infty({\mathbb{R}}^N)$ in $\dot H^{1,2}({\mathbb{R}}^N)$ we may assume without loss of generality that $v, w \in C_0^\infty({\mathbb{R}}^N)$.
2\. [ *Sufficiency:*]{} First note that supports of $V_k$ and $W_k$ are contained in $B(y_k,r) \cap e_{y_k}(2^{-j_k}\mathrm{supp}\ v)$ and $B(z_k,r) \cap e_{z_k}(2^{-\ell_k}\mathrm{supp}\ w)$ respectively. Thus, if along a subsequence $\inf d(y_k,z_k) >0$, then the supports of $V_{k}$ and $W_{k}$ are disjoint for large $k$ and the conclusion follows. Hence we assume in the rest of the proof that $d(y_k,z_k)\rightarrow 0$ as $k\rightarrow \infty$.\
\
The support of $dV_{k}\cdot dW_{k}$ is contained in $B(z_k,r)$ and hence calculating the integral in the coordinate chart $e_{z_k}$ we get $$\int_{M}dV_{k}\cdot dW_{k}d\mu = \int_{\Omega_r} g^{\alpha \beta}(e_{z_k}(\xi))\partial_\alpha (V_k\circ e_{z_k})\partial_\beta (W_k\circ e_{z_k}) \sqrt{g(e_{z_k}(\xi))}d\xi$$ where $\partial_1,\dots,\partial_N$ is a shorthand for respective partial derivatives $\frac{\partial}{\partial \xi_1}, \dots , \frac{\partial}{\partial \xi_N}$, $\xi_1,\dots,\xi_N$ are the components of $\xi\in \Omega_r$, $g_{\alpha\beta}=\langle\frac{\partial}{\partial \xi_{\alpha}},\frac{\partial}{\partial \xi_{\beta}}\rangle_g,\ (g^{\alpha\beta}) = (g_{\alpha\beta})^{-1},\ g = \det (g_{\alpha\beta})$ and we use the summation convension over $\alpha,\beta \in \{1,...,N\}.$\
\
Denoting $j_{k}-\ell_{k}=m_{k},\ e_{y_k}^{-1}\circ e_{z_k}\ = \ \psi_k $, using the change of variable $\eta=2^{\ell_{k}}\xi $, and taking into account that $g^{\alpha\beta}(z_k) = \delta_{\alpha\beta} $ and $g(z_k)=1$, the above expression simplifies to $\int_{M}dV_{k}\cdot dW_{k}d\mu = \int_{B(z_k,r)\cap B(y_k,r)}dV_{k}\cdot dW_{k}d\mu = 2^{\frac{N-2}{2}m_k} \times$ $$\int_{D_k}g^{\alpha \beta}(e_{z_k}(\xi))\partial_\alpha (\chi \circ \psi_k (\xi)v(2^{j_{k}}\psi_k(\xi)))\partial_\beta (\chi(\xi) w(2^{\ell_k}\xi)) \sqrt{g(e_{z_k}(\xi))}d\xi,$$ $$= 2^{\frac{N}{2}m_k}\int_{2^{\ell_k}D_k \cap \mathrm{supp}w}\left[\nabla w (\eta) \psi'_k(2^{-\ell_{k}}\eta)\cdot \nabla v (2^{j_k}\psi_k(2^{-\ell_{k}}\eta)) + o(1)\right]d\eta$$ where $D_k= e_{z_k}^{-1}\left(B(z_k,r)\cap B(y_k,r)\right)$, $\psi_k'$ is the $N\times N$-matrix derivative of $\psi_k$, and $o(1)$ denotes a sequence converging to zero uniformly in ${\mathbb{R}}^{N}$. Since $M$ is of bounded geometry using standard arguments we may assume $\psi_k$ and its derivatives will converge locally uniformly in $\Omega_r$ to some $\psi \in C^\infty(\Omega_r \rightarrow \overline{\Omega_r})$ and its respective derivatives.\
Since holds, we have up to subsequences two cases :\
[*Case 1 :*]{} $|\ell_k-j_k| \rightarrow \infty $.\
In this case we may assume without loss of generality that $j_{k}-\ell_{k}=m_{k} \rightarrow -\infty $ and hence using the boundednesss of $\psi_k$ and the fact that $v,w \in C_c^\infty({\mathbb{R}}^N)$ we get $$\left|\int_{M}dV_{k}\cdot dW_{k}d\mu\right|\ \le \ C 2^{\frac{N}{2}m_k}\ \rightarrow 0.$$ [*Case 2 :*]{} $|\ell_k-j_k|$ is bounded and $ \ (2^{\ell_k}+2^{j_k}) d(y_k,z_k) \rightarrow \infty $.\
For $\eta \in 2^{\ell_k}D_k \cap \mathrm{supp}\ w$,we have $$2^{j_k}\psi_k(2^{-\ell_{k}}\eta)= 2^{j_k}\psi_k(0) + 2^{j_k}\psi'_k(0)2^{-\ell_{k}}\eta \ + \ O(2^{j_k-2\ell_{k}})$$ Note that since the exponential map is preserving distance from the origin in $\Omega_r$, we have $|\psi_k(0)|= d(y_k,z_k)$. Using this with the fact that $m_k$ is bounded we get $|2^{j_k}\psi_k(2^{-\ell_{k}}\eta)| \rightarrow \infty\ , \ {\rm as}\ \ k\rightarrow \infty$ (with uniform convergence) and hence $\nabla v (2^{j_k}\psi_k(2^{-\ell_{k}}\eta))=0$ for all $\eta \in \mathrm{supp}\ w$ provided that $k$ is sufficiently large, and hence the conclusion follows.
3\. [*Necessity:*]{} Assume that the condition (\[eq:orth3\]) does not hold. Then, on a renamed subsequence, $j_{k}-\ell_{k}=m\in{\mathbb{Z}}$ and $2^{j_k}d(y_k,z_k)= |2^{j_k}\psi_k(0)|$ remains bounded. Hence passing to a further subsequence if necessary we may assume $2^{j_k}\psi_k(0)\to\eta_0\in{\mathbb{R}}^N$. Then, repeating calculations in the proof of sufficiency, we get $$\int_{M}dV_{k}\cdot dW_{k}d\mu\rightarrow 2^{\frac{N}{2}m}\int_{{\mathbb{R}}^N}\nabla w (\eta) \cdot \psi'(0)\nabla v (\eta_0+2^{m}
\psi'(0)\eta)\,d\eta$$
Using elementary properties of the exponential map we can show that $\psi'(0)$ is invertible. Hence the above expression is nonzero, for example, when $v\neq 0$ and $w(\eta)= v(\eta_0+ 2^m\psi'(0)\eta)$.
4\. Finally, writing the scalar product of $H^{1,2}(M)$ in the normal coordinates at $y_k$ we have $$\begin{aligned}
(W_k,V_k)=\int_{{\mathbb{R}}^N}\nabla 2^{-j_{k}\frac{N-2}{2}}W_{k}\circ e_{y_{k}}(2^{-j_{k}}\xi)\cdot\nabla v(\xi) d\xi+o(1),\end{aligned}$$ which proves that is equivalent to asymptotic orthogonality of $V_k,W_k$ for all $v$.
We now provide a profile decomposition for sequences that are bounded in $H^{1,2}(M)$ and vanish in $L^{p}(M)$ for some $p\in(2,2^{*})$. This will allow us to consider a general bounded sequence in $H^{1,2}(M)$, for which a profile decomposition with a remainder vanishing in $L^{p}$ is already known, identify blowups in this remainder, and after subtracting them, obtain a profile decomposition for the remainder up to a term that, by Theorem \[thm:vanishing\], will vanish already in $L^{2^{*}}$.
\[thm:PD-blowupsonly\]Let $M$ be a complete smooth $N$-dimensional Riemannian manifold of bounded geometry, and let $(u_{k})$ be a bounded sequence in $H^{1,2}(M)$ convergent to zero in $L^{p}(M)$ for some $p\in(2,2^{*})$. Then, there exist sequences of points $(y_{k}^{(n)})_{k\in{\mathbb{N}}}$ in $M$ and of numbers $(j_{k}^{(n)})_{k\in {\mathbb{N}}}$ in ${\mathbb{N}}$, $j_{k}^{(n)}\to+\infty$, as well as functions $w^{(n)}\in\dot{H}^{1,2}({\mathbb{R}}^{N})$, $n\in{\mathbb{N}},$ such that $$2^{-j_{k}^{(n)}\frac{N-2}{2}}u_{k}\circ e_{y_{k}^{(n)}}(2^{-j_{k}^{(n)}}\xi)\to w^{(n)}(\xi)\;\mbox{ a.e. in }{\mathbb{R}}^{N},\label{eq:profiles01}$$ $\mathbf{(AO)}$Whenever $m\neq n$, the sequences $j_{k}^{(m)},$ $y_{k}^{(m)}$, $j_{k}^{(n)},$ $y_{k}^{(n)}$ satisfy the condition (\[eq:orth3\]) with $j_{k}=j_{k}^{(m)},$ $y_{k}=y_{k}^{(m)}$, $\ell_{k}=j_{k}^{(n)}$ and $z_k=y_{k}^{(n)}$.
The series $S_{k}=\sum_{n\in{\mathbb{N}}}W_{k}^{(n)},$ where $$W_{k}^{(n)}(x)=2^{j_{k}^{(n)}\frac{N-2}{2}}\,\chi\circ e_{y_{k}^{(n)}}^{-1}(x)\,w^{(n)}(2^{j_{k}^{(n)}}e_{y_{k}^{(n)}}^{-1}(x)),\:x\in M,\label{eq:Sksum}$$
converges in $H^{1,2}(M)$ unconditionally with respect to $n$ and uniformly in $k$,
$$u_{k}-S_{k}\to0\mbox{ in }L^{2^{*}}(M)\, ,\label{eq:remainder}$$
and
$$\sum_{n\in{\mathbb{N}}}\int_{{\mathbb{R}}^{N}}|\nabla w^{(n)}|^{2}dx\le\liminf\int_{M}|du_{k}|^{2}d\mu.\label{eq:Planch01}$$
The proof is analogous to the proofs for profile decompositions in [@SkrTi4; @SchinTin; @Solimini; @SoliTi] and is given in a somewhat abbreviated form.
1\. Consider any sequences $(j_{k}^{(1)})$ in ${\mathbb{N}}$, $j_{k}^{(1)}\to\infty$, and $(y_{k}^{(1)})$ in $M$. The sequence $2^{-j_{k}^{(1)}\frac{N-2}{2}}u_{k}\circ e_{y_{k}^{(1)}}(2^{-j_{k}^{(1)}}\cdot)$ is a bounded sequence in $H^{1,2}(\Omega_{a})$ for any $a\in(0,r(M))$ and thus has a weakly convergent subsequence there. Consequently, by diagonalization, there is a renamed subsequence of $(u_{k})$ such that $2^{-j_{k}^{(1)}\frac{N-2}{2}}u_{k}\circ e_{y_{k}^{(1)}}(2^{-j_{k}^{(1)}}\cdot)$ converges almost everywhere on ${\mathbb{R}}^{N}$to some $w^{(1)}$. Note that, since $M$ is of bounded geometry, $\|\nabla w^{(1)}\|_{2}^{2}\le\limsup C\int_{B_{r}(y_{k}^{(1)})}|du_{k}|^{2}d\mu<\infty$ with the constant $C$ independent of the sequence $(y_k^{(n)})_{k\in{\mathbb{N}}}$. Thus it suffices to prove the theorem for the case $w^{(1)}\neq 0$. Let $\Xi_1$ be a set of all $w\in\dot H^{1,2}(M)$ such that $$2^{-j_{k}^{(1)}\frac{N-2}{2}}u_{k}\circ e_{y_{k}^{(1)}}(2^{-j_{k}^{(1)}}\to w \mbox{ a.e. in } {\mathbb{R}}^N$$ for some choice of $j_{k}^{(1)}\to\infty$ and $(y_{k}^{(1)})$ in $M$, and let $$\beta_1=\sup_{w\in \Xi_1}\|w\|_{\dot{H}^{1,2}}$$ and fix an element $w^{(1)}\in \Xi$ and corresponding sequences $j_{k}^{(1)}$ and $(y_{k}^{(1)})$ so that $\|w^{(1)}\|_{\dot{H}^{1,2}}\ge \frac12 \beta_1$. If $\beta_1=0$, then by Theorem \[thm:vanishing\] $u_k\to 0$ in $L^{2^*}(M)$ and the theorem is proved. We consider therefore the case $\beta_1>0$.
2\. Let us now make a recursive definition of sequences $j_{k}^{(n)}\to\infty$, $(y_{k}^{(n)})$ in $M$, functions $w^{(n)}$ and numbers $\beta_n$, $n\in{\mathbb{N}}$. Let $\nu\in{\mathbb{N}}$ and assume that for every $n=1,\dots,\nu$ we have constructed the elementary concentrations for the sequence $(u_k)$, that is, we make the following assumptions.
Assume, for $n=1,\dots,\nu$, that we have a renamed subsequence of $(u_{k})$, sequences $(j_{k}^{(n)})$ in ${\mathbb{N}}$, $j_{k}^{(n)}\to\infty$, and $(y_{k}^{(n)})$ in $M$. Assume that functions $w^{(n)}\in\dot{H}^{1,2}({\mathbb{R}}^{N})$ satisfy . Assume that (AO) is satisfied for $m,n=1,\dots,\nu$. Furthermore, let $\Xi_n$ be a set of all $w\in\dot H^{1,2}(M)$ such that $$2^{-j_{k}^{(n)}\frac{N-2}{2}}u_{k}\circ e_{y_{k}^{(n)}}(2^{-j_{k}^{(n)}})\to w \mbox{ a.e. in } {\mathbb{R}}^N$$ for some choice of $j_{k}^{(n)}\to\infty$ and $(y_{k}^{(n)})$ in $M$ satisfying (AO) for $m<n=1,\dots,\nu$, and let $$\beta_n=\sup_{w\in \Xi_n}\|w\|_{\dot{H}^{1,2}}$$ and fix an element $w^{(n)}\in \Xi$ and corresponding sequences $j_{k}^{(n)}$ and $(y_{k}^{(n)})$ so that $\|w^{(n)}\|_{\dot{H}^{1,2}}\ge \frac12 \beta_n$.
We now set $S_{k}^{(\nu)}=\sum_{n=1}^{\nu}W_{k}^{(n)}$ and $v_{k}^{(\nu)}=u_{k}-S_{k}^{(\nu)}$. Consider sequences $(j_{k}^{(\nu+1)})$ in ${\mathbb{N}}$, $j_{k}^{(\nu+1)}\to\infty$, and $(y_{k}^{(\nu+1)})$ in $M$. As in the Step 1, we have a renamed subsequence of $(u_{k})$ such that $2^{-j_{k}^{(\nu+1)}\frac{N-2}{2}}u_{k}\circ e_{y_{k}^{(\nu+1)}}(2^{-j_{k}^{(\nu+1)}}\cdot)$ converges almost everywhere on ${\mathbb{R}}^{N}$ to some $w^{(\nu+1)}\in\dot{H}{}^{1,2}({\mathbb{R}}^{N})$. As before, we consider the class $\Xi_{\nu+1}$ of all such limits, and fix $w^{(\nu+1)}$ and corresponding $j_{k}^{(\nu+1)}$ and $(y_{k}^{(\nu+1)})$ so that $\|w^{(\nu+1)}\|_{\dot{H}^{1,2}}\ge \frac12 \beta_{\nu+1}$.
3\. Note that for every $n\le\nu$, $$2^{-j_{k}^{(n)}\frac{N-2}{2}}W_{k}^{(n)}\circ e_{y_{k}^{(n)}}(2^{-j_{k}^{(n)}}\xi)=\chi(2^{-j_{k}^{(n)}}\xi)w^{(n)}(\xi)\to w^{(n)}(\xi)\mbox { a.e. in }{\mathbb{R}}^N,$$ and for each $n'\le\nu$, $n'\neq n$, by (AO) and Lemma \[lem:aorth\], $$2^{-j_{k}^{(n)}\frac{N-2}{2}}W_{k}^{(n')}\circ e_{y_{k}^{(n)}}(2^{-j_{k}^{(n)}}\xi)\to 0\mbox{ a.e. in }{\mathbb{R}}^N,$$ so for every $n\le\nu$, $$2^{-j_{k}^{(n)}\frac{N-2}{2}}S^{(\nu)}_{k}\circ e_{y_{k}^{(n)}}(2^{-j_{k}^{(n)}}\xi)\to w^{(n)}(\xi) \mbox{ a.e. in } {\mathbb{R}}^N,$$ and thus $$\begin{aligned}
\nonumber
2^{-j_{k}^{(n)}\frac{N-2}{2}}v_{k}^{(\nu)}\circ e_{y_{k}^{(n)}}(2^{-j_{k}^{(n)}}\xi)=
\\
\label{eq:proj4}
2^{-j_{k}^{(n)}\frac{N-2}{2}}(u_{k}-S_{k}^{(\nu)})\circ e_{y_{k}^{(n)}}(2^{-j_{k}^{(n)}}\xi)
\to 0 \mbox{ a.e. in }{\mathbb{R}}^N.\end{aligned}$$ 4. Let us show that (AO) is satisfied for $m=1,\dots,\nu$ and $n=\nu+1$ (or vice versa). Once we show this, we will have completed the construction of $(y_{k}^{(n)})_{k\in{\mathbb{N}}}$ in $M$ and of numbers $(j_{k}^{(n)})_{k\in N}$ in ${\mathbb{N}}$, $j_{k}^{(n)}\to+\infty$, as well as functions $w^{(n)}\in\dot{H}^{1,2}({\mathbb{R}}^{N})$, $n\in{\mathbb{N}},$ such that, for a renamed subsequence of $(u_{k}),$ condition (AO) is satisfied for all $n\in{\mathbb{N}}$. If $w^{(\nu+1)}=0$, then, since $\beta_{\nu+1}=0$ we are free replace $(y_{k}^{(\nu+1)})_{k\in{\mathbb{N}}}$ in $M$ and $(j_{k}^{(\nu+1)})_{k\in N}$ with any sequence that saisfies for respective scaling sequences, namely $$|j_k^{(\nu+1)}-j_k^{(n)}| \ + \ (2^{j_k^{\nu+1}}+2^{j_k^{(n)}}) d(y_k^{(\nu+1)},y_k^{(n)}) \rightarrow \infty, n=1,\dots\nu.$$ The renamed $w^{(\nu+1)}$ will be necessarily zero since $\beta_{\nu+1}=0$.
We now may assume that $w^{(\nu+1)}\neq 0$. Assume that (AO) does not hold for indices $\nu+1$ and some $\ell\le\nu$. Then there exist $m\in{\mathbb{Z}}$ and $\lambda\in{\mathbb{R}}$, such that, on a renamed subsequence, $2^{j_{k}^{(\ell)}}d(y_{k}^{(\ell)},y_{k}^{(\nu+1)})$ is bounded and $j_{k}^{(\nu+1)}=j_{k}^{(\ell)}-m$. Let $\psi_k=e_{y_{k}^{(\ell)}}^{-1}\circ e_{y_{k}^{(\nu+1)}}$. Note that $d(y_{k}^{(\ell)},y_{k}^{(\nu+1)})\to 0$, and since $M$ has bounded geometry, on a renamed subsequence we have $\psi_k$ convergent uniformly, together with its derivatives of every order, and its limit is the identity map. Also on a renamed subsequence we have an $\eta_0\in {\mathbb{R}}^N$ such that $2^{j_{k}^{(\nu+1)}}\psi_k(0)\to \eta_0$ (since $|\psi_k(0)|=d(y_{k}^{(\nu+1)},y_{k}^{(\ell)})$). Then we have, uniformly on compact subsets of ${\mathbb{R}}^N$, $$\begin{aligned}
2^{j_{k}^{(\ell)}}\psi_k(2^{-j_{k}^{(\nu+1)}}\xi)=
\\
2^{m}2^{j_{k}^{(\nu+1)}}\psi_k(0)+2^{m}\psi_k'(0)\xi+2^{j_{k}^{(\ell)}}O(2^{-2j_{k}^{(\nu+1)}}\xi)\to 2^{m}\eta_0+2^m\xi.\end{aligned}$$ Note also that for any $n=\ell+1,\dots,\nu$ from (AO) and Lemma \[lem:aorth\] we have $$2^{j_{k}^{(\nu+1)}\frac{N-2}{2}}W_k^{(n)}\circ e_{y_{k}^{(\nu+1)}}(2^{-2j_{k}^{(\nu+1)}}\xi)\to 0.$$
Then, substituting the two last calculations into the expression below, we have $$\begin{aligned}
2^{j_{k}^{(\nu+1)}\frac{N-2}{2}}v_k^{(\nu)}\circ e_{y_{k}^{(\nu+1)}}(2^{-2j_{k}^{(\nu+1)}}\xi)=
\\
2^{j_{k}^{(\nu+1)}\frac{N-2}{2}}v_k^{(\ell)}\circ e_{y_{k}^{(\nu+1)}}(2^{-2j_{k}^{(\nu+1)}}\xi)+o(1)=
\\
2^{-m\frac{N-2}{2}}
2^{j_{k}^{(\ell)}\frac{N-2}{2}}v_k^{(\ell)}\circ e_{y_{k}^{(\ell)}}\circ \psi_k(2^{-2j_{k}^{(\nu+1)}}\xi)=\\
2^{-m\frac{N-2}{2}}
2^{j_{k}^{(\ell)}\frac{N-2}{2}}v_k^{(\ell)}\circ e_{y_{k}^{(\ell)}}(2^{j_{k}^{(-\ell)}}
[2^{j_{k}^{(\ell)}}\psi_k(2^{-2j_{k}^{(\nu+1)}}\xi)])=
\\
2^{-m\frac{N-2}{2}}
2^{j_{k}^{(\ell)}\frac{N-2}{2}}v_k^{(\ell)}\circ e_{y_{k}^{(\ell)}}(2^{j_{k}^{(-\ell)}}
[2^m\eta_0+2^m\xi+o(1)])
\\
\to 0,
\end{aligned}$$ by , which by definition of $w^{(\nu+1)}$ implies $w^{(\nu+1)}=0$, a contradiction.\
6. Let us expand by bilinearity a trivial inequality\
$\int_{M}|d(u_{k}-S_{k}^{(\nu)})|^{2}d\mu\ge0$. We will have $$\int_{M}|du_{k}|^{2}d\mu\ge2\int_{M}du_{k}\cdot dS_{k}^{(\nu)}d\mu-I_{k}-I'_{k},\label{eq:trivpos}$$
where $$I_{k}=\sum_{n=1}^{\nu}\int_{M}\left|d\left(2^{j_{k}^{(n)}\frac{N-2}{2}}\,\chi\circ e_{y_{k}^{(n)}}^{-1}(x)\,w^{(n)}(2^{j_{k}^{(n)}}e_{y_{k}^{(n)}}^{-1}(x))\right)\right|^{2}d\mu,$$ and $$I'_{k}=\sum_{m\neq n,\,m,n=1,\dots\nu}\int_{M}d\left(2^{j_{k}^{(n)}\frac{N-2}{2}}\,\chi\circ e_{y_{k}^{(n)}}^{-1}(x)\,w^{(n)}(2^{j_{k}^{(n)}}e_{y_{k}^{(n)}}^{-1}(x))\right)
\cdot$$ $$d\left(2^{j_{k}^{(m)}\frac{N-2}{2}}\,\chi\circ e_{y_{k}^{(m)}}^{-1}(x)\,w^{(m)}(2^{j_{k}^{(m)}}e_{y_{k}^{(m)}}^{-1}(x))\right)d\mu.$$ The first term in (\[eq:trivpos\]) can be evaluated by writing the integration in rescaled geodesic coordinates $\xi=2^{j_{k}^{(n)}}e_{y_{k}^{(n)}}^{-1}(x)$ and denoting by
$o(1)$ any term vanishing uniformly in ${\mathbb{R}}^{N}$ as $k\to\infty$, we $$\begin{array}{l}
\int_{M}du_{k}\cdot dS_{k}^{(\nu)}d\mu=
\\
\sum_{n=1}^{\nu}2^{j_{k}^{(n)}\frac{N-2}{2}}\,\int_{B_{r}(y_{k}^{(n)})}du_{k}\cdot d\left(\chi\circ e_{y_{k}^{(n)}}^{-1}(x)\,w^{(n)}(2^{j_{k}^{(n)}}e_{y_{k}^{(n)}}^{-1}(x))\right)d\mu
\\
=\sum_{n=1}^{\nu}2^{-j_{k}^{(n)}\frac{N-2}{2}}\int_{\Omega_{2^{j_{k}^{(n)}}r}}\sum_{\alpha,\beta=1}^{N}(\delta_{\alpha\beta}+o(1))\partial_{\alpha}u_{k}\circ e_{y_{k}^{(n)}}(2^{-j_{k}^{(n)}}\xi)\times
\\
\left(\chi(2^{-j_{k}^{(n)}}\xi)\partial_{\beta}w^{(n)}(\xi)+o(1)
\partial_{\beta}w^{(n)}(\xi)\right)d\xi\\
=\sum_{n=1}^{\nu}\int_{{\mathbb{R}}^{N}}|\nabla w^{(n)}(\xi)|^{2}d\xi \ + \ o(1)
\end{array}$$\
An analogous evaluation of $I_{k}$ gives, $I_{k}=$ $$\begin{array}{l}
\sum_{n=1}^{\nu}2^{-j_{k}^{(n)}\frac{N-2}{2}}\int_{\Omega_{2^{j_{k}^{(n)}}r}}\sum_{\alpha,\beta=1}^{N}
\left(\chi(2^{-j_{k}^{(n)}}\xi)\partial_{\alpha}w^{(n)}(\xi)+o(1)\partial_{\alpha}w^{(n)}(\xi)\right)\times
\\
\left(\chi(2^{-j_{k}^{(n)}}\xi)\partial_{\beta}w^{(n)}(\xi)+o(1)\partial_{\beta}w^{(n)}(\xi)\right)(\delta_{\alpha\beta}+o(1))d\xi \\
= \sum_{n=1}^{\nu}\int_{{\mathbb{R}}^{N}}|\nabla w^{(n)}(\xi)|^{2}d\xi \ + \ o(1).
\end{array}$$ By (AO) and Lemma \[lem:aorth\], we have $I'_{k}\to0$. Consequently, (\[eq:trivpos\]) implies $$\begin{aligned}
\int_{M}|du_{k}|^{2}d\mu\ge2\sum_{n=1}^{\nu}\int_{{\mathbb{R}}^{N}}|\nabla w^{(n)}(\xi)|^{2}d\xi-\sum_{n=1}^{\nu}\int_{{\mathbb{R}}^{N}}|\nabla w^{(n)}(\xi)|^{2}d\xi+o(1)
\\
=\sum_{n=1}^{\nu}\int_{{\mathbb{R}}^{N}}|\nabla w^{(n)}(\xi)|^{2}d\xi+o(1).\label{eq:prePl}\end{aligned}$$
Since $\nu$ is arbitrary, we have (\[eq:Planch01\]).
7\. Inequality (\[eq:Planch01\]) implies that $\beta_{\nu}\to0$ as $\nu\to\infty$. Then, following the argument of [@SoliTi] with only trivial modifications, one can show that, for a suitably renamed sequence, the series $S_{k}$ converges in $H^{1,2}(M)$ unconditionally in $n$ and uniformly in $k$.
8\. Let us finally show that the sequence $(S_k)$ gives indeed the defect of compactness. Let $j_{k}\in{\mathbb{N}}$, $j_{k}\to\infty$ and let $y_{k}\in M$. It suffices to consider two cases.
Case A. For every $n\in{\mathbb{N}}$, $j_{k}^{(n)}, j_k$ and $y_{k}^{(n)},y_{k}$ satisfy the condition (\[eq:orth3\]). Then $2^{-j_{k}\frac{N-2}{2}}S_{k}\circ e_{y_{k}}(2^{-j_{k}}\cdot)\to0$ a.e., as this is true for each term in the series of $S_{k}$ by Lemma \[lem:aorth\], and the convergence in the series for $S_{k}$ is uniform. On the other hand, if there is a renamed subsequence such that $2^{-j_{k}\frac{N-2}{2}}u_{k}\circ e_{y_{k}}(2^{-j_{k}}\cdot)\to w$ a.e., we will have necessarily $\|\nabla w\|_{2}\le\beta_{\nu}$ for every $\nu\in{\mathbb{N}}$, i.e. $w=0$. We conclude that $2^{-j_{k}\frac{N-2}{2}}(u_{k}-S_{k})\circ e_{y_{k}}(2^{-j_{k}}\cdot)\to0$ a.e. in this case.
Case B. Without loss of generality we may assume that for some $\ell\in{\mathbb{N}}$ $j_{k}-j_{k}^{(\ell)}=m\in{\mathbb{Z}}$ and that $2^{j_k}d(y_{k}^{(\ell)},y_{k})$ is bounded. Then, repeating the argument of Step 4, we have $$2^{-j_{k}\frac{N-2}{2}}(u_{k}-S_{k}^{(\ell)}\circ e_{y_{k}}(2^{-j_{k}}\cdot))\to0\mbox{ a.e. },$$ while by (AO), Lemma \[lem:aorth\] and the uniform convergence of the series $S_{k}$, $$2^{-j_{k}\frac{N-2}{2}}((S_{k}-S_{k}^{(\ell)})\circ e_{y_{k}}(2^{-j_{k}}\cdot))\to0\mbox{ a.e.},$$ which implies that $2^{-j_{k}\frac{N-2}{2}}(u_{k}-S_{k})\circ e_{y_{k}}(2^{-j_{k}}\cdot)\to0$ a.e. in this case as well.
Consequently, by Theorem \[thm:vanishing\] we have $u_{k}-S_{k}\to0$ in $L^{2^{*}}.$
We supplement Theorem \[thm:PD-blowupsonly\] with the estimate of $L^{2^{*}}$-norms.
\[prop:BL\*\]Let $u_{k}$ and $w^{(n)}$, $n\in{\mathbb{N}}$, be provided by Theorem \[thm:PD-blowupsonly\]. Then $$\int_{M}|u_{k}|^{2^{*}}d\mu\to\sum_{n\in{\mathbb{N}}}\int_{{\mathbb{R}}^{N}}|w^{(n)}|^{2^{*}}dx.$$
By (\[eq:remainder\]) it suffices to show that $$\int_{M}|\sum_{n\in{\mathbb{N}}}W_{k}^{(n)}|^{2^{*}}d\mu\to\sum_{n\in{\mathbb{N}}}\int_{{\mathbb{R}}^{N}}|w^{(n)}|^{2^{*}}dx.$$
Since sequences $(W_{k}^{(n)})_{k\in{\mathbb{N}}}$ have asymptotically disjoint supports in the sense of (\[eq:orth3\]), and since one can by density of $C_c({\mathbb{R}}^N)$ in $\dot H^{1,2}({\mathbb{R}}^N)$ assume that every profile $w^{(n)}$ has compact support, it is easy to show that $$\int_{M}|\sum_{n\in{\mathbb{N}}}W_{k}^{(n)}|^{2^{*}}d\mu-\sum_{n\in{\mathbb{N}}}\int_{M}|W_{k}^{(n)}|^{2^{*}}d\mu\to0,$$ so it suffices to show that for each $n\in{\mathbb{N}}$ $$\int_{M}|W_{k}^{(n)}|^{2^{*}}d\mu\to\int_{{\mathbb{R}}^{N}}|w^{(n)}|^{2^{*}}dx.$$ Indeed, passing to geodesic coordinates at $y_{k}^{(n)}$, and subsequently rescaling them, we have $$\int_{M}|W_{k}^{(n)}|^{2^{*}}d\mu=2^{j_{k}^{(n)}N}\int_{M}\,|\chi\circ e_{y_{k}^{(n)}}^{-1}(x)\,w^{(n)}(2^{j_{k}^{(n)}}e_{y_{k}^{(n)}}^{-1}(x))|^{2^{*}}d\mu=$$ $$2^{j_{k}^{(n)}N}\int_{\Omega_{\rho}}\,|\chi(\xi)\,w^{(n)}(2^{j_{k}^{(n)}}\xi)|^{2^{*}}\sqrt{g(\xi)}d\xi=$$ $$\int_{_{{\mathbb{R}}^{N}}}|\chi(2^{-j_{k}^{(n)}}\eta)\,w^{(n)}(\eta)|^{2^{*}}\sqrt{g(2^{-j_{k}^{(n)}}\eta)}d\eta\to\int_{_{{\mathbb{R}}^{N}}}|w^{(n)}(\eta)|^{2^{*}}d\eta.$$ Passing to the limit at the last step is justified by Lebesgue dominated convergence theorem, and by the fact that in normal coordinates $g(0)=1$.
Profile decomposition for $H^{1,2}(M)\protect\hookrightarrow L^{2^{*}}(M)$
==========================================================================
Profile decomposition for $H^{1,2}(M)\protect\hookrightarrow L^{p}(M)$, subcritical case
----------------------------------------------------------------------------------------
In order to quote the profile decomposition for the subcritical Sobolev embedding from [@SkrTi4], we need to refer to Appendix for definitions of the objects involved there. 1. Discretization $Y_{\rho}$ of a metric space $M$, $\rho>0$, is an at most countable subset of $M$ such that $$d(y,y')\ge\rho/2\mbox{ whenever }y\neq y',y,y'\in Y_{\rho}\label{eq:discrete-1}$$ the collection of balls $\lbrace B(y,\rho)\rbrace_{y\in Y_{\rho}}$ is a covering of $M$, and for any $a\ge\rho$ the multiplicity of the covering $\lbrace B(y,a)\rbrace_{y\in Y}$ is uniformly finite. Manifolds of bounded geometry always admit a discretization for sufficiently small $\rho$, see Appendix.
2\. Manifold at infinity $M_{\infty}^{(y_{k})}$, of a manifold $M$ with finite geometry, associated with a sequence $(y_{k})$ from a discretization $Y_{\rho}$ of $M$ is a Riemannian manifold defined in the Appendix. Like $M$ itself, $M_{\infty}^{(y_{k})}$ has an atlas $\lbrace\varphi_{i}^{(y_{k})},\Omega_{\rho}\rbrace_{i\in{\mathbb{N}}}$ with identical coordinate neighborhoods $\Omega_{\rho}$.
3\. Notation $M_{\infty}^{(y_{k})}$ is somewhat abbreviated as the construction of the manifold at infinity is in fact dependent on the choice of a trailing system $\lbrace(y_{k,i})_{k\in{\mathbb{N}}}\rbrace_{i\in{\mathbb{N}}}$ of the sequence $(y_{k})$. For each $k\in{\mathbb{N}}$, $\lbrace y_{k,i}\rbrace_{i\in{\mathbb{N}}}$ is an ordering (not necessarily unique) of $Y_{\rho}$ by distance from the point $y_{k}$.
\[thm:pd-subcr\]Let $M$ be a manifold of bounded geometry with a discretization $Y_{\rho}$, $\rho<r(M)/8$, and let $(u_{k})$ be a sequence in $H^{1,2}(M)$ weakly convergent to some function $u$ in $H^{1,2}(M)$. Then there exists a renamed subsequence of $(u_{k})$, sequences $(y_{k}^{(n)})_{k\in{\mathbb{N}}}$ on $Y_{\rho}$, and associated with them global profiles $w^{(n)}$ on the respective manifolds at infinity $M_{\infty}^{(n)}{\stackrel{\mathrm{def}}{=}}M_{\infty}^{(y_{k}^{(n)})}$, $n\in{\mathbb{N}}$, such that $d(y_{k}^{(n)},y_{k}^{(m)})\to\infty$ when $n\neq m$, $$u_{k}-u-\sum_{n\in{\mathbb{N}}}W_{k}^{(n)}\to0\mbox{ in }L^{p}(M),\;p\in(2,2^{*}),\label{eq:PD}$$ where $$W_{k}^{(n)}=\sum_{i\in{\mathbb{N}}_{0}}\chi\circ e_{y_{k,i}^{(n)}}^{-1}w_{i}^{(n)}\circ e_{y_{k,i}^{(n)}}^{-1},$$ $$w_{i}^{(n)}=w^{(n)}\circ\varphi_{i}^{(n)}={\operatorname{w-lim}}u_k\circ e_{y_{k,i}}^{(n)},$$ $\lbrace\varphi_{i}^{(n)},\Omega_{\rho}\rbrace_{i\in{\mathbb{N}}}$ is the atlas of $M_{\infty}^{(n)}$, and the series $\sum_{n\in{\mathbb{N}}}W_{k}^{(n)}$ converges in $H^{1,2}(M)$ unconditionally (with respect to $n$) and uniformly in $k\in{\mathbb{N}}$. Moreover, $$\|u\|_{H^{1,2}(M)}^{2}+\sum_{n=1}^{\infty}\|w^{(n)}\|_{H^{1,2}(M_{\infty}^{(n)})}^{2}+\|u_{k}-u-\sum_{n\in{\mathbb{N}}}W_{k}^{(n)}\|_{H^{1,2}(M)}^{2}$$ $$\le\limsup\|u_{k}\|_{H^{1,2}(M)}^{2}\ ,\label{eq:Plancherel}$$ and $$\int_{M}|u_{k}|^{p}d\mu\to\int_{M}|u|^{p}d\mu+\sum_{n=1}^{\infty}\int_{M_{\infty}^{(n)}}|w^{(n)}|^{p}d\mu_{\infty}^{(n)}.\label{eq:newBL}$$
Note that the original version of (\[eq:Plancherel\]) $$\|u\|_{H^{1,2}(M)}^{2}+\sum_{n=1}^{\infty}\|w^{(n)}\|_{H^{1,2}(M_{\infty}^{(n)})}^{2}\le\limsup\|u_{k}\|_{H^{1,2}(M)}^{2}\label{eq:oldPln}$$ in [@SkrTi4] omits the last term in the left hand side of (\[eq:Plancherel\]). The inequality in the present form can be derived, however, from its counterpart in [@SkrTi4] as follows. Let $U_{k}=u+\sum_{n\in{\mathbb{N}}}W_{k}^{(n)}$. Once we show that $(u_{k}-U_{k},U_{k})_{H^{1,2}(M)}\to0$, we have (\[eq:Plancherel\]) by expanding $$0\le\|u_{k}\|{}_{H^{1,2}(M)}^{2}=\|(u_{k}-U_{k})+U_{k}\|{}_{H^{1,2}(M)}^{2}$$ by bilinearity and applying (\[eq:oldPln\]) to the sequence $(U_{k})$. By the uniform convergence of the series in $U_{k}$, it suffices to show that $(u_{k}-U_{k},u)_{H^{1,2}(M)}\to0$ and that $(u_{k}-U_{k},W_{k}^{(n)})_{H^{1,2}(M)}\to0$. Since $u_{k}\rightharpoonup u$ and $W_{k}^{(n)}\rightharpoonup0$, $u_{k}-U_{k}\rightharpoonup0$, the first expression indeed vanishes. For vanishing of the second expression it suffices to show that $(u_k,W_{k}^{(n)})\to\|w^{(n)}\|^2_{H^{1,2}(M_\infty^{(n)})}$ ([@SkrTi4 relation (5.9)]), $(W_{k}^{(m)},W_{k}^{(n)})_{H^{1,2}(M)}\to 0$ whenever $m\neq n$ ([@SkrTi4 proof of Lemma 5.5 below (5.19)]), and\
$(W_{k}^{(n)},W_{k}^{(n)})_{H^{1,2}(M)}\to\|w^{(n)}\|^2_{H^{1,2}(M_\infty^{(n)})}$ ([@SkrTi4 relation (5.14)]).
Main result and corollaries
---------------------------
We will now assume that the parameter $r\in(0,r(M))$ involved in the statements of Sections 2 and 3 satisfies the constraint of Section 4 as well, $r<r(M)/8$.
\[thm:finalPD\]Let $M$ be a Riemannian manifold of bounded geometry and let $u_{k}$ be a bounded sequence in $H^{1,2}(M)$ weakly convergent to some $u$.Then there is a renamed subsequence of $u_{k}$, sequences $(\bar y_k^{(m)})_{k\in{\mathbb{N}}}$, $m\in{\mathbb{N}}$, and $(y_k^{*(n)})_{k\in{\mathbb{N}}}$, $n\in{\mathbb{N}}$, of points in $M$, sequences $(j_k^{(n)})_{k\in{\mathbb{N}}}$, $n\in{\mathbb{N}}$ of integers, $j_{k}^{(n)}\to+\infty$ as $k\to\infty$, such that $$u_{k}-u-\sum_{m\in{\mathbb{N}}}\bar{W}_{k}^{(m)}-\sum_{n\in{\mathbb{N}}}W_{k}^{*(n)}\to0\mbox{ in }L^{2^{*}}(M),\label{eq:finalremainder}$$ where $\bar{W}_{k}^{(m)}$ are given by Theorem \[thm:pd-subcr\] (relative to sequences $(\bar y_k^{(m)})_{k\in{\mathbb{N}}}$), and $$W_{k}^{*(n)}(x)=2^{j_{k}^{(n)}\frac{N-2}{2}}\,\chi\circ e_{y_{k}^{*(n)}}^{-1}(x)\,w^{*(n)}(2^{j_{k}^{(n)}}e_{y_{k}^{*(n)}}^{-1}(x)),\:x\in M,\label{eq:Sksum*}$$ where $$2^{-j_{k}^{(n)}\frac{N-2}{2}}u_{k}\circ e_{y_{k}^{*(n)}}(2^{-j_{k}^{(n)}}\cdot)\to w^{*(n)}\;\mbox{ a.e. in }{\mathbb{R}}^{N},\label{eq:profiles01*}$$ (as in Theorem \[thm:PD-blowupsonly\]), and both series in converge unconditionally and uniformly in $k$.
Furthermore, sequences $j_{k}^{(n)}$, $y_{k}^{*(n)}$, $j_{k}^{(n')},$ $y_{k}^{*(n')}$ satisfy the condition (\[eq:orth3\]).
Moreover, with $M_\infty^{(m)}$, $m\in{\mathbb{N}}$, as in Theorem \[thm:pd-subcr\], we have
$$\sum_{n\in{\mathbb{N}}}\int_{{\mathbb{R}}^{N}}|\nabla w^{*(n)}|^{2}dx+\sum_{m\in{\mathbb{N}}}\int_{M_{\infty}^{(m)}}(|d\bar{w}^{(m)}|^{2}+|\bar{w}^{(m)}|^{2})d\mu_{\infty}^{(m)}$$ $$+ \int_{M}(|du|^{2}+u^{2})d\mu \le\int_{M}(|du_{k}|^{2}+u_{k}^{2})d\mu+o(1),\label{eq:finalPlancherel}$$
and
$$\int_{M}|u_{k}|^{2^{*}}d\mu\to\sum_{n\in{\mathbb{N}}}\int_{{\mathbb{R}}^{N}}|w^{*(n)}|^{2^{*}}dx+\sum_{m\in{\mathbb{N}}}\int_{M_{\infty}^{(m)}}|\bar{w}^{(m)}|^{2^{*}}d\mu_{\infty}^{(m)}+\int_{M}|u|^{2^{*}}d\mu.\label{eq:finalBL}$$
Apply Theorem \[thm:pd-subcr\] to $u_{k}$ and let $v_{k}=u_{k}-u-\sum_{m\in{\mathbb{N}}}\bar{W}_{k}^{(m)}$ be the left hand side of in (\[eq:PD\]) Note that $v_{k}$ is a bounded sequence in $H^{1,2}(M)$ because so are both $u_{k}$ and $\sum_{m\in{\mathbb{N}}}\bar{W}_{k}^{(m)}$ (for the latter it can be inferred from (\[eq:Plancherel\])). Apply Theorem \[thm:PD-blowupsonly\] to $v_{k}.$ Then (\[eq:finalremainder\]) is immediate from combining (\[eq:PD\]) and (\[eq:remainder\]). Relation (\[eq:finalPlancherel\]) follows from substitution of (\[eq:Planch01\]) for $v_{k}$ into (\[eq:Plancherel\]).
By Brezis-Lieb Lemma, we have $$\int_{M}|u_{k}|^{2^{*}}d\mu=\int_{M}|u-u-\sum_{m\in{\mathbb{N}}}\bar{W}_{k}^{(m)}|^{2^{*}}d\mu+\int_{M}|u+\sum_{m\in{\mathbb{N}}}\bar{W}_{k}^{(m)}|^{2^{*}}+o(1).\label{eq:tempBL}$$ Then (\[eq:finalBL\]) follows by evaluating the first term in the right hand side of (\[eq:tempBL\]) by Proposition \[prop:BL\*\] and evaluating the second term as $\int_{M}|u|^{2^{*}}d\mu+\sum_{m=1}^{\infty}\int_{M_{\infty}^{(m)}}|\bar{w}^{(m)}|^{2^{*}}d\mu_{\infty}^{(n)}$. The latter has the same form as (\[eq:newBL\]), but for $p=2^{*}$, and can be obtained by literally repeating the argument in [@SkrTi4] applied to the sequence $(u+\sum_{m\in{\mathbb{N}}}\bar{W}_{k}^{(m)})_{k}$.
Let $M$ be a manifold of negative curvature with bounded geometry (in particular a hyperbolic space). Let $u_{k}\in H^{1,2}(M)$ satisfy $\int_{M}|du_{k}|^{2}d\mu\le C. $Then $u_{k}$ has a renamed subsequence satisfying the assertions of Theorem \[thm:finalPD\].
Let $M$ be a non-compact symmetric space ( in particular, ${\mathbb{R}}^N$ or the hyperbolic space $\mathbb H^N$ with $N>2$). Then Theorem \[thm:finalPD\] holds with $M_\infty^{(m)}=M$ for every $m\in{\mathbb{N}}$, and with $\bar W_k^{(m)}=\bar w^{(m)}\circ\eta_k^{(m)}$, where $w^{(m)}={\operatorname{w-lim}}u_k\circ{\eta_k^{(m)}}^{-1}$, $\eta_k^{(m)}$ are discrete sequences of isometries on $M$, and the sequences ${\eta_k^{(m)}}^{-1}\circ {\eta_k^{(m')}}$ are discrete whenever $m\neq m'$. This is immediate from the corresponding simplification of Theorem \[thm:pd-subcr\] for the case homogeneous spaces, [@SkrTi4 Theorem 1.1]. The same holds also when $M$ coincides with homogeneous space outside of a compact subset.
Appendix
========
Manifolds of bounded geometry and covering lemma
------------------------------------------------
In this subsection we list some elementary properties of manifolds of bounded geometry.
The following lemma is immediate from [@Eich Theorem A ff.] .
\[lem:2.1\]Let $p\in(0,\infty)$ and $a\in(0,r(M))$. There exists a constant $C>1$ such that for any $x\in M$
$$C^{-1}\int_{B_{a}(x)}|u|^{p}d\mu \ \le \ \int_{\Omega_{a}}|u\circ e_{x}\text{\ensuremath{|^{p}}dx \ \ensuremath{\le} \ C\ensuremath{\int}}_{B_{a}(x)}|u|^{p}d\mu,\label{eq:peq}$$
and
$$$$
For any two points $x,y\in M$ define $$\Omega_{a}(x,y)=e_{x}^{-1}(e_{x}\Omega_{a}\cap e_{y}\Omega_{a})\subset\Omega_{a},$$ and a diffeomorphism $$\psi_{yx}=e_{y}^{-1}\circ e_{x}:\Omega_{a}(x,y)\to\Omega_{a}(y,x).$$
\[lem:bdd-derivatives\]If the manifold $M$ has bounded geometry, then for any $\alpha\in{\mathbb{N}}_{0}^{N}$ there exists a constant $C_{\alpha}>0$, such that $$|D^{\alpha}\psi_{yx}(\xi)|\le C_{\alpha}\mbox{ whenever }x,y\in M,\text{\;\ensuremath{\Omega_{a}}(x,y)\ensuremath{\neq\emptyset},}\;\xi\in\Omega_{a}(x,y).$$
whenever $\Omega_{a}(x,y)$ is nonempty.
This lemma is immediate from Definition \[def:bg\].
The following lemma is found, in particular, in [@Hebey Lemma 1.1].
\[lem:covering\]Let $M$ be an $N$-dimensional connected Riemannian manifold with Ricci curvature uniformly bounded from below . Let $\rho>0$. There exists an at most countable set $Y\in M$ such that
$$d(y,y')\ge\rho/2\mbox{ whenever }y\neq y',y,y'\in Y,\label{eq:discrete}$$
the collection of balls $\lbrace B(y,\rho)\rbrace_{y\in Y}$ is a covering of $M$, and for any $a\ge\rho$ the multiplicity of the covering $\lbrace B(y,a)\rbrace_{y\in Y}$ is uniformly finite.
Manifolds at infinity
---------------------
In this subsection we give a cursive summary of the construction of manifolds at infinity for a Riemannian manifold of bounded geometry, following [@SkrTi4]. We consider a radius $a\in(1,r(M))$ and the discretization $Y=Y_{a}$ of $M$ fixed assured by the previous lemma with $\rho=a$. We will use the notation ${\mathbb{N}}_{0}{\stackrel{\mathrm{def}}{=}}{\mathbb{N}}\cup\lbrace0\rbrace$.
\[def:trailing\] Let $(y_{k})_{k\in{\mathbb{N}}}$ be a sequence in $Y$. A countable family $\lbrace(y_{k;i})_{k\in{\mathbb{N}}}\rbrace_{i\in{\mathbb{N}}_{0}}$ of sequences on $Y$ is called a family of trailing sequences for $(y_{k})_{k\in{\mathbb{N}}}$ if for every $k\in{\mathbb{N}}$ $(y_{k;i})_{i\in{\mathbb{N}}_{0}}$ is an ordering of $Y$ by the distance from $y_{k}$, that is, an enumeration of $Y$ such that $d(y_{k,i},y_{k})\le d(y_{k,i+1},y_{k})$ for all $i\in{\mathbb{N}}_{0}$. In particular, $y_{k,0}=y_{k}$.
The trailing family is generally not uniquely defined when for some $k\in{\mathbb{N}}$ there are several points of $Y$ with the same distance from $y_{k}$, so strictly speaking the manifold at infinity is determined in the construction below not by the sequence $(y_{k})$ but by its trailing family.
To each pair $(i,j)\in{\mathbb{N}}_{0}^{2}$ we associate a subset $\Omega_{ij}$ of $\Omega_{\rho}$, where $\rho=\frac{a}{5}$. Note that for each $i,j\in{\mathbb{N}}_{0}$, such that $\liminf_{k\to\infty}d(y_{k,i},y_{k,j})\le\rho$, and with $\xi,\eta\in\Omega_{\rho}$, $$d(e_{y_{k,j}}\xi,e_{y_{k,i}}\eta)\le d(e_{y_{k,j}}\xi,y_{k,j})+d(y_{k,j},y_{k,i})+d(y_{k,i},e_{y_{k,i}}\eta)\le3\rho+o(1)<a$$ for all $k$ large enough, so that we have a sequence of diffeomorphisms
$$\psi_{ij,k}{\stackrel{\mathrm{def}}{=}}e_{y_{k,i}}^{-1}\circ e_{y_{k,j}}:\;\bar{\Omega}_{\rho}\to\Omega_{a},\,k\mbox{ large enough}.$$
By an argument combining uniform bounds on derivatives of exponential maps on a manifold of bounded geometry, Arzela-Ascoli theorem, and diagonalization, one may arrive at a renamed subsequence such that for all $i,j\in{\mathbb{N}}_{0},$ $(\psi_{ij,k})_{k\in{\mathbb{N}}}$ converges in $C^{\infty}(\bar{\Omega}_{\rho})$ to some smooth function $\psi_{ij}:\bar{\Omega}_{\rho}\to\Omega_{a}$. Setting $\Omega_{ij}{\stackrel{\mathrm{def}}{=}}\psi_{ij}\Omega_{\rho}\cap\Omega_{\text{\ensuremath{\rho}}}$, we have $\psi_{ij}\circ\psi_{ji}=\mathrm{id}$ on $\Omega_{ij}$ and $\psi_{ji}\circ\psi_{ij}=\mathrm{id}$ on $\Omega_{ji}$. Therefore $\psi_{ji}=\psi_{ij}^{-1}$ in restriction to $\Omega_{ij}$, and $\psi_{ji}$ is a diffeomorphism between $\Omega_{ij}$ and $\Omega_{ji}$. Note that this construction gives that $\psi_{ii}=\mathrm{id}$ , $\Omega_{ii}=\Omega_{\rho}$ for all $i\in{\mathbb{N}}_{0}$. Furthermore, $$\psi_{\ell i}=\lim_{k\to\infty}e_{y_{k,\ell}}^{-1}\circ e_{y_{k,i}}=\lim_{k\to\infty}e_{y_{k,\ell}}^{-1}\circ e_{y_{k,j}}\circ e_{y_{k,j}}^{-1}\circ e_{y_{k,i}}$$ $$=\lim_{k\to\infty}e_{y_{k,\ell}}^{-1}\circ e_{y_{k,j}}\circ\lim_{k\to\infty}e_{y_{k,j}}^{-1}\circ e_{y_{k,ji}}=\psi_{\ell j}\circ\psi_{ji},$$ and $$\psi_{ij}(\Omega_{ji}\cap\Omega_{jk})=\psi_{ij}(\psi_{ji}(\Omega_{\rho})\cap\Omega_{\text{\ensuremath{\rho}}}\cap\psi_{jk}(\Omega_{\rho})\cap\Omega_{\text{\ensuremath{\rho}}})=\Omega_{ij}\cap\Omega_{ik},$$ This allows to invoke a gluing theorem (e.g. [@Gallier3 Theorem 3.1]) to conclude that exists a smooth differential manifold $M_{\infty}$ with an atlas $\lbrace\Omega_{\rho},\varphi_{i}\rbrace_{i\in{\mathbb{N}}_{0}}$ whose transition maps satisfy $\varphi_{j}^{-1}\varphi_{i}=\psi_{ij}:\Omega_{ji}\to\Omega_{ij}$.
The manifold $M_{\infty}$ can be endowed with the following Riemannian metric. The metric on $M$ is a bilinear form on $TM$, expressed in the atlas $\lbrace\Omega_{\rho},e_{y}\rbrace_{y\in Y}$ as follows. For each $y\in Y$ we have an orthonormal frame $\lbrace\nu_{\alpha}^{y}\rbrace_{\alpha=1,\dots,N}$ on $Te_{y}(\Omega_{\rho})\subset TM$ by using fixed Euclidean coordinates for $\Omega_{\rho}$, and setting a local frame $\nu_{\alpha}^{y}{\stackrel{\mathrm{def}}{=}}e_{y*}\partial_{\alpha}$, $\alpha=1,\dots,N$. Then the metric tensor on the chart $(\Omega_{\rho},e_{y})$ is given by $g[\nu_{\alpha,}^{y}\nu_{\beta}^{y}]\circ e_{y}$. We define a local frame on the chart$(\Omega_{\rho},\varphi_{i})$ of $M_{\infty}$, $i\in{\mathbb{N}}_{0}$, by $\nu_{\alpha}^{(i)}{\stackrel{\mathrm{def}}{=}}\varphi_{i*}\partial_{\alpha}$. Then we set $$g[\nu_{\alpha}^{(i)},\nu_{\beta}^{(i)}]\circ\varphi_{i}{\stackrel{\mathrm{def}}{=}}\lim_{k\to\infty}g[\nu_{\alpha,}^{y_{k,i}}\nu_{\beta}^{y_{k,i}}]\circ e_{y_{k,i}}\mbox{ on }\Omega_{\rho},\,i\in{\mathbb{N}}_{0}.\label{eq:metric}$$ It can be shown that this defines a Riemannian metric on $M_{\infty}$ by verifying a compatibility relation for overlapping charts $$g_{\alpha\beta,i}=\sum_{\gamma,\delta=1}^{N}\partial_{\alpha}\psi_{ji}^{\gamma}\,\partial_{\beta}\psi_{ji}^{\delta}\;g_{\gamma\delta,j}\circ\psi_{ji}\mbox{ on }\Omega_{ij},i,j\in{\mathbb{N}}_{0},\label{eq:gluing-R}$$ where $\psi_{ji}^{\alpha}$ are components of $\psi_{ji}$, i.e. $\psi_{ji}=\sum_{\alpha=1}^{N}\psi_{ji}^{\alpha}\nu_{\alpha}^{(i)}$ and $g_{\alpha\beta,i}{\stackrel{\mathrm{def}}{=}}g[\nu_{\alpha}^{(i)},\nu_{\beta}^{(i)}]\circ\varphi_{i}$.
A manifold at infinity $M_{\infty}^{(y_{k})}$ of a manifold $M$ with bounded geometry, generated by a sequence $(y_{k})$ in $Y$, is a differentiable manifold given by the construction above supplied with a Riemannian atlas $\lbrace\varphi_{i},\Omega_{\rho},[g_{\alpha\beta,i}]\rbrace_{i\in{\mathbb{N}}_{0}}$, whose transition maps are $\varphi_{j}^{-1}\varphi_{i}=\psi_{ji}:\Omega_{ij}\to\Omega_{ji}$, where $\psi_{ji}=\lim_{k\to\infty}e_{y_{k,j}}^{-1}\circ e_{y_{k,i}}$ (as a sequence of maps $\Omega_{\rho}\to\Omega_{a}$) and $\Omega_{ij}=\psi_{ij}\Omega_{\rho}\cap\Omega_{\text{\ensuremath{\rho}}}$, $i,j\in{\mathbb{N}}_{0}$; and the metric $\lbrace[g_{\alpha\beta,i}]\rbrace_{i\in{\mathbb{N}}_{0}}$ is given by (\[eq:metric\]).
In terms of the definition above the argument of this subsection proves the following statement.
Let $M$ be a smooth manifold with bounded geometry. Then every sequence $(y_{k})$ in $Y$ has a renamed subsequence that generates a Riemannian manifold at infinity $M_{\infty}^{(y_{k})}$ of the manifold $M$.
[10]{} Adimurthi, C. Tintarev, On compactness in the trudinger-moser inequality. Ann. Sc. Norm. Sup. Pisa Cl. Sci. 5 (2014), 118.
Adimurthi, C. Tintarev, Defect of compactness in spaces of bounded variation, J. Func. Anal. 271 (2016), 3748.
I. Chavel, Eigenvalues in Riemannian geometry, Pure and Applied Mathematics, 115. Academic Press, Inc., Orlando, FL, 1984.
G. Devillanova, C. Tintarev, A profile decomposition for the limiting Sobolev embedding, ArXiv: 1812.04248
O. Druet, E. Hebey, F. Robert, Blow-up Theory for Elliptic PDEs in Riemannian Geometry, Princeton University Press, 2004.
J. Eichhorn, The boundedness of connection coefficients and their derivatives, Math. Nachr. 152 (1991), 144158 .
J. Gallier, J. Quaintance, Notes on Differential Geometry and Lie Groups, Book in progress (2017), http://www.cis.upenn.edu/\~jean/gbooks/manif.html
J. Gallier, D. Xu, M. Siqueira, Parametric pseudo-manifolds, Differential Geometry and its Applications 3**0** (2012) 702736.
P. Gérard, Description de compacité de l’injection de sobolev. [*ESAIM Control Optim. Calc. Var.*]{}, 3 (1988), 213233.
E. Hebey, Nonlinear Analysis on Manifolds: Sobolev Spaces and Inequalities, Courant lecture Notes **5**, Courant Institute of mathematical Sciences and American Mathematical Society, 1999.
S. Jaffard, Analysis of the lack of compactness in the critical Sobolev embeddings, J. Funct. Analysis. 161 (1999), 384-396.
Lieb, E., On the lowest eigenvalue of the Laplacian for the intersection of two domains. Invent. Math. 74 (1983), 441-448
I. Schindler and K. Tintarev, An abstract version of the concentration compactness principle, Revista Mat. Complutense, 15 (2002), 1-20.
M. A. Shubin, Spectral theory of elliptic operators on noncompact manifolds, Méthodes semi-classiques, Vol. 1 (Nantes, 1991), Astérisque 207, 35108 (1992).
L. Skrzypczak, C. Tintarev, Defect of compactness for Sobolev spaces on manifolds with bounded geometry, Ann. Sc. Norm. Sup. Pisa Cl. Sci. 5 (to appear).
S. Solimini, A note on compactness-type properties with respect to Lorentz norms of bounded subsets of a Sobolev space, Ann. Inst. H. Poincaré Anal. Non Linéaire 12 (1995), 319-337.
S. Solimini, C. Tintarev, Analysis of concentration in the Banach space, Comm. Contemp. Math. 18 (2016), 1550038 (33 pages).
M. Struwe, A global compactness result for elliptic boundary value problems involving limiting nonlinearities, Math. Z. **187**, 511-517 (1984).
K. Tintarev, K.-H. Fieseler, Concentration compactness: functional-analytic grounds and applications, Imperial College Press 2007.
[^1]: One of the authors (C.T.) expresses his gratitude to Professor Sandeep and the faculty of TIFR-CAM for their warm hospitality.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper describes the surface environment of the dense plasma arcs that damage rf accelerators, tokamaks and other high gradient structures. We simulate the dense, non-ideal plasma sheath near a metallic surface using Molecular Dynamics (MD) to evaluate sheaths in the non-Debye region for high density, low temperature plasmas. We use direct two-component MD simulations where the interactions between all electrons and ions are computed explicitly. We find that the non-Debye sheath can be extrapolated from the Debye sheath parameters with small corrections. We find that these parameters are roughly consistent with previous PIC code estimates, pointing to densities in the range $10^{24} - 10^{25}\mathrm{m}^{-3}$. The high surface fields implied by these results could produce field emission that would short the sheath and cause an instability in the time evolution of the arc, and this mechanism could limit the maximum density and surface field in the arc. These results also provide a way of understanding how the properties of the arc depend on the properties (sublimation energy, for example) of the metal. Using these results, and equating surface tension and plasma pressure, it is possible to infer a range of plasma densities and sheath potentials from SEM images of arc damage. We find that the high density plasma these results imply and the level of plasma pressure they would produce is consistent with arc damage on a scale 100 nm or less, in examples where the liquid metal would cool before this structure would be lost. We find that the sub-micron component of arc damage, the burn voltage, and fluctuations in the visible light production of arcs may be the most direct indicators of the parameters of the dense plasma arc, and the most useful diagnostics of the mechanisms limiting gradients in accelerators.'
author:
- 'I. V. Morozov, G. E. Norman'
- 'Z. Insepov, J. Norem$^*$'
title: 'Sheath parameters for non-Debye plasmas: simulations and arc damage'
---
Introduction
============
Comparatively little is known about the vacuum arcs and gradient limits that are important in determining the cost and overall parameters of large linear accelerator facilities. Vacuum arcs are involved in many fields, from particle accelerators, plasma devices, high power switching, surface coating and a variety of laboratory and commercial applications, and these arcs have been under study for many years. Nevertheless, the properties of these dense plasmas are not well understood, although the general behavior of these arcs has been known and under study since 1901 [@earhart; @Anders; @Juttner; @mesyats; @kajita] and these plasmas seem to limit the performance of both major accelerator and tokamak projects and facilities [@CLIC; @MAP; @ITER]. In part, the reason for this situation is that the arcs are small, and many parameters (which are individually hard to measure) evolve very rapidly over a very wide range. Theory and modeling are complicated by the large number of mechanisms that seem to be involved in arc evolution and high density plasmas, which requires a non-Debye analysis of basic properties.
While arc damage has been measured and catalogued for over a hundred years, there has not been any clear correspondence between specific types of arc damage and the past or subsequent behavior of the arc. We argue in this paper that the causes of arc damage are due to the high density, high surface electric field plasma that produces a high plasma pressure in the liquid metal surface underneath the plasma arc. This pressure produces very small scale structures, at or below a few hundred nm. We have developed a self consistent model of arc evolution and show that these structures may or may not survive the subsequent cool-down of the liquid surface, however the cool down itself also seems to produce cracks with small scale structures [@NoremPAC09; @NoremLINAC10; @Noremrf2011; @InsepovNorem11]. This damage can produce future breakdown events.
Recent work has shown that the development of the arc can be explained by two mechanisms: 1) mechanical failure of the solid surface due to Coulomb explosions caused by high surface fields [@InsepovNorem11], and, 2) the development of unipolar arcs [@Schwirzke91], that can act as virtual cathodes and produce currents that can short the driving potential. Once an arc starts, the surface electric field, and field emission increase, increasing ionization of neutrals, causing an increase in the plasma density. The density increase decreases the Debye length and causes an increase in the surface electric field, thus both the electric field and the density increase exponentially with time, roughly described by the arrow in Fig. 1. PIC simulations of the unipolar arc model for vacuum arcs relevant to rf cavity breakdown [@NoremPAC09; @NoremLINAC10; @Noremrf2011; @InsepovNorem11] show that the density of plasma formed above the field emitting asperities can be as high as $10^{26}\ \mathrm{m}^{-3}$. The temperature of such plasma is low, in the range of $1-10$ eV.
These high densities can make the Debye screening length, $$\begin{aligned}
\lambda_D = \sqrt{\epsilon_0 k_B T / n_e e^2,}\end{aligned}$$ become smaller than the mean inter-particle distance, or the number of particles in the Debye sphere, $$\begin{aligned}
N_D = 4\pi n_e \lambda_D^3 / 3,\end{aligned}$$ to become less that unity. This implies the failure of the ideal plasma approximation, as well as most of the assumptions used in simple calculations. Processes in such a dense plasma can be affected by three body particle collisions so that the Particle In Cell (PIC) method which relies on a simple collisional model, with two body collisions, becomes inappropriate, as shown in Fig 1, where the arrow shows the approximate range of parameters for evolving arcs from Ref [@Noremrf2011; @InsepovNorem11], as well as the approximate region of validity for PIC and Molecular Dynamics (MD) codes.
In this paper we calculate the parameters of the surface environment underneath the plasma sheath for the high density plasma conditions using direct two-component MD simulations where the interactions between all electrons and ions are computed explicitly. Although MD simulations have limited space and time scales their results can be considered as the lower level output for the multiscale approach.
Equilibrium and non-equilibrium nonideal plasmas have been studied extensively by MD in the past several decades [@Zwicknagel-JPA03; @JETP05; @Murillo; @Donko; @Rostock-CPP09]. Nevertheless there are few studies of the spatially inhomogeneous systems such as electric double layers or plasma sheath. In this paper we report on the first results for MD simulations of the nonideal plasma sheath near a metallic surface.
We can compare modeling with experimental measurements, but both the modeling and experiments are somewhat indirect. An understanding of the surface fields, combined with plasma density give estimates of the plasma pressure that can be compared with experimental data. It is possible to make indirect experimental measurements of the plasma pressure by comparing the linear dimensions of structures seen in the surface with estimates obtained from comparing the surface tension of the liquid metal with the plasma properties. We describe how the scale of structures frozen in damage as the arc cools, can be used to set limits on the plasma properties.
Simulation technique
====================
The two-component fully ionized electron-ion plasma is considered. Neutral atoms are not taken into account which can affect relaxation times at relatively low plasma densities when the density of neutrals is high enough. It should not, however, affect the stationary distribution of charges. In the present work the simulations are restricted to the singly ionized plasma with $Z=1$.
The electron-electron and ion-ion interactions are given by the Coulomb potential. For electrons and ions it is modified at short distances to account for quantum effects. The equation below assumes a Gaussian wave function for an electron $$\begin{aligned}
V_{ei} (r) = \frac{Z e^2}{4 \pi \epsilon_0 r}
\mathrm{erf} \left( \frac{r}{r_0} \right), \label{erf}\end{aligned}$$ where the $r_0$ parameter that equals to $0.21$ nm in out case to match the ionization energy for copper at $r=0$: $U(0) = - 7.73$ eV (see Fig. \[potential-ei\]). The similar interaction model was used e.g. in [@Rostock-CPP09; @Suraud06; @Zwicknagel-CPP03] for simulations of ionized metallic clusters. More accurate electron-ion and electron-electron interaction models are discussed e.g. in [@CKelbg; @WPMD-JPA09] although they seem to be redundant for this particular case. In fact the results are weakly dependent on the short distance part of the potential as the change of the $U(0)$ value from 7.73 eV to 5.1 eV does not change the results within simulation accuracy.
The Leap-Frog integration scheme is used to solve the classical equations of motion for electrons and ions [@velocityVerlet]. The method takes into account the conservation of the total energy of the finite system, as long as there is no external potential. To follow the electron dynamics, time steps of $0.001-0.01$ fs were taken to calculate the time evolution.
The general simulation scheme follows the method described in [@MolSimul05] and shown in Fig. \[averaging\]. First an equilibrium MD trajectory is calculated for the system at given density and temperature using the nearest image method (periodic boundary conditions) for all dimensions. The simulation box size and other parameters are summarized in Table \[SimParam\]. The Langevin thermostat [@LangevinTherm] is used initially to bring the system to an equilibrium state while it is switched off for a production run. Then the system becomes adiabatic which ensures that all thermodynamic quantities are conserved in average. The ion mass is set to be equal to the electron mass for better mixing of ionic trajectories at this phase. The nonideality parameter, $\Gamma$, is the ratio of the average Coulomb potential energy and the average kinetic energy per electron [@EKKR_book].
$T, \mathrm{eV}$ $n_e, 10^{27}\mathrm{m}^{-3}$ $L_x, \mathrm{nm}$ $L_z, \mathrm{nm}$ $N_i$ $\Gamma$ $\Theta$ $\lambda_D, \mathrm{nm}$
------------------ ------------------------------- -------------------- -------------------- ------- ---------- ---------- --------------------------
1 0.0001 120 360 518 0.11 0.001 23.5
1 0.001 55 165 499 0.23 0.004 7.43
1 0.01 25 75 468 0.50 0.017 2.35
1 0.1 11 33 399 1.08 0.079 0.74
1 1.0 5 15 375 2.32 0.36 0.24
1 5.0 2.8 8.4 329 3.97 1.07 0.11
10 0.01 25 75 468 0.05 0.002 7.43
10 1.0 5 15 375 0.23 0.036 0.74
10 100 1 2.5 300 1.08 0.79 0.07
: \[SimParam\]MD simulation parameters: $T$ is the initial electron temperature, $n_e$ is the initial number density of electrons (or ions), $L_z$ is the transversal simulation box size, $L_x$ is the longitudinal simulation box size, $N_i$ is the number of ions which is equal to the initial number of electrons, $\Gamma = e^2(4\pi
n_e/3)^{1/3}/(4\pi\epsilon_0 k_B T)$ is the nonideality parameter, $\Theta
= \hbar^2 (3\pi^2 n_e)^{2/3} / (2m_e k_B T)$ is the degeneracy parameter, $\lambda_D$ is the Debye length.
At the second phase the particle positions and velocities at particular time moments are taken from the equilibrium trajectory to be used as the initial states for nonequilibrium simulations of the plasma sheath. The interval between those points should be large enough for the initial states to be statistically independent from the microscopical point of view. However, all these states correspond to the same macroscopical conditions as they are taken from a single equilibrium trajectory. Then a bunch of trajectories is computed starting from the given ensemble of initial states and the results are averaged over the ensemble.
In order to study the plasma sheath, the XY plane at $z = 0$ axis is considered as a metallic surface whereas a reflecting wall is introduced on the other side of the box at $z = L_z$. The periodic boundary conditions are still applied for transverse dimensions $x$ and $y$. When an electron crosses the surface it is always meant to be absorbed. Therefore it is removed from the system and the overall surface charge is incremented by its charge $q_{\mathrm{surf}} \leftarrow q_{\mathrm{surf}} - e$.
A non-zero surface charge produces an electrostatic field which influence the particles. where $\sigma = q_{\mathrm{surf}}/(L_x L_y)$ is the surface charge density and $L_x$, $L_y$ are the box sizes in the transverse dimensions. Assuming $L_x = L_y = 2a$ one can obtain (see e.g. [@Jackson_book]) $$\begin{aligned}
E_z(z)
= \frac{\sigma}{\pi\epsilon_0} \arctan\left[ \frac{a^2}{z\sqrt{2a^2 + z^2}} \right],
\label{Ez}\end{aligned}$$ where $E_z$ is the longitudinal component of the electric field. It can be shown that Eq. (\[Ez\]) tends to the field expression of a uniformly charged plane $E_z = \sigma/(2\epsilon_0)$ as $z\to 0$ and to the Coulomb field $E_z =
\sigma a^2 / (\pi\epsilon_0) / z$ as $z\to\infty$. It is important to use Eq. (\[Ez\]) in the simulation with the given boundaries instead of the field of a charged plane $E_z = \sigma/(2\epsilon_0)$ as the later cannot be screened by plasma particles at a large distance. As the surface field grows it starts to repel electrons from the surface until the stationary state is reached.
We do not compute dynamics of ions at this phase as the ions are too heavy to contribute to the simulation results at the electron time scale. At the same time the ions are movable at the equilibrium trajectory that is used for generation of the initial states. Thus the averaging over an ensemble means the averaging over different configurations of ions. In a real system the number of ions will vary with time, due to ions entering and leaving the plasma from their thermal motion and self sputtering. Because the ion velocity is 340 times smaller than the electron velocity, this process is very slow, and we have neglected these effects. This is equivalent to assuming that the self sputtering coefficient for copper ions is near unity.
We have checked that the final results are independent of the simulation box size. If the box is doubled the deviation of the results are in within the statistical errors.
The thermodynamics parameters was maintained in the course of simulation. It was found that the overall electron temperature deviates in the range of $1-10$% due to absorption of the most energetic electrons to the surface.
Simulation Results and Fit Formulas
===================================
Typically the relaxation of the electric field is observed for about 1 ps (see Fig. \[surfield-t\]). The development of the electron profile is shown in Fig. \[dp-t\]. The stationary density profiles obtained after the relaxation are shown in Fig. \[dp-z\]. As the ions does not move, their distribution mimics the uniform distribution obtained from the equilibrium trajectory with full periodic boundary conditions. On the contrary, electrons form the well pronounced layer of plasma near the surface with a positive charge which we consider as the plasma sheath.
The plasma charge density profile is given by the difference between the ion and electron densities as presented in Fig. \[charge-z\] in the semilogarithmic scale. It is seen that starting from the surface the charge density $\sigma(z)$ decays exponentially which agrees with the Debye approximation. At high densities, however, the exponential decay is preceded by a non-exponential area. This regions makes difference between calculation of the sheath length from the slope $\lambda_\mathrm{exp}$ of the exponential decay $\sigma(z)\sim\mathrm{e}^{-z/\lambda_\mathrm{exp}}$ and from the distance $\lambda$ at which the charge density decreases at the value of $\mathrm{e}
= 2.71$ ($\sigma(\lambda) = \sigma(0)/\mathrm{e}$) as illustrated in Fig. \[charge-z\]c.
Both quantities $\lambda_\mathrm{exp}$ and $\lambda$ are presented in Fig. \[lambda-n\] depending on the plasma density and temperature. It is seen that $\lambda_\mathrm{exp}$ follows the Debye-like dependence on density $\lambda_\mathrm{exp} \sim n_e^{-1/2}$ whereas the real sheath length $\lambda$ scales with a slightly different exponent.
The fits for MD data are $$\begin{aligned}
T_e = 1 \mathrm{eV}:\qquad &
\lambda_\mathrm{exp} = 1.7\lambda_D, \quad
\lambda[\mathrm{nm}] = 1.0\cdot10^{11} (n_e[m^{-3}])^{-0.405}, \label{lambdaT1}\\
T_e = 10 \mathrm{eV}:\qquad &
\lambda_\mathrm{exp} = 1.7\lambda_D, \quad
\lambda[\mathrm{nm}] = 3.18\cdot10^{12} (n_e[m^{-3}])^{-0.449}. \label{lambdaT10}\end{aligned}$$
Fig. \[lambdar-n\] shows how the ratios $\lambda_\mathrm{exp}/\lambda_D$ and $\lambda/\lambda_D$, depend on plasma temperature, electron density and the nonideality parameter. While Fig. \[lambdar-n\]a shows the dependence of the screening length on plasma temperature, Fig. \[lambdar-n\]b shows that the values of $\lambda/\lambda_D$ for different temperatures are close to each other when plotted versus the parameter $\Gamma$. This implies that the nonexponential charge density decay and the difference between $\lambda$ and $\lambda_D$ are primarily a function of the plasma nonideality, defined as the ratio of average electrostatic potential energy divided by average kinetic energy.
The values of the electric field at the metal surface are presented in Fig. \[surfiled-n\] depending on both temperature and density. The solid line corresponds to Eq. (2) from [@Schwirzke91] $$\begin{aligned}
\label{EfSchwirzke}
E = \frac{V_f}{\lambda_D} = [n_e k_B T_e / (4\epsilon_0)]^{1/2}
\log[M_i / (2\pi m_e)],\end{aligned}$$ where $M_i$ and $m_e$ are the masses of electron and ion. If the Debye radius in Eq. (\[EfSchwirzke\]) is substituted by the MD result (\[lambdaT1\]) or (\[lambdaT10\]) it results in the values shown by crosses in Fig. \[surfiled-n\] which are in a better agreement with the MD results.
The fits for MD data shown by dashed lines are $$\begin{aligned}
T_e = 1\mathrm{eV}:\qquad &
E[\mathrm{GV/m}] = 2.57\cdot10^{-15} (n_e[m^{-3}])^{-0.577}, \label{EfT1}\\
T_e = 10\mathrm{eV}:\qquad &
E[\mathrm{GV/m}] = 1.21\cdot10^{-13} (n_e[m^{-3}])^{-0.531}. \label{EfT10}\end{aligned}$$
Fig. \[plasmapot-n\] shows the plasma potential calculated using the simple relation of $\phi = E / \lambda$ where both the surface electric field $E$ and the sheath length $\lambda$ are obtained from MD simulations. A more rigorous result can be found by integration of the electric field distribution in plasma but it requires a more accurate evaluation of the space charge away from the sheath area and will be the subject of future work.
Experimental estimates of plasma parameters
===========================================
The internal parameters of the arc, the surface field and the metal under it, are not directly accessible, but experimental measurements can provide some indirect evidence of the internal structure and the active mechanisms. Three phenomena are, in principle, sensitive to these numbers: 1) the mechanism that limits the exponential increase of the density and electric field with time, 2) the properties of the metal surface (melting point, self sputtering yield, etc.) may determine the burn voltage through the mechanism of self sputtering, and, 3) the scale of surface damage frozen into the surface should be sensitive to the plasma pressure that created it. We discuss these briefly.
Field Emission Driven Plasma Instabilities
------------------------------------------
The surface fields and plasma densities described above can be very high and it has been shown that these fields and densities increase exponentially with time during the evolution of the discharge [@NoremLINAC10; @Noremrf2011; @InsepovNorem11]. The magnitude of these fields suggests that field emission over the entire active area could short out the sheath, causing an instability or oscillation in the plasma limiting this exponential increase.
As the surface field increases above 2 GV/m, it becomes possible for field emission currents to short out (or significantly reduce) the sheath field in times on the scale of nanoseconds. Assuming this occurs, we would expect that the sheath would rapidly reestablish itself due to the short collision time, the comparatively large plasma volume, and the high plasma density, and this behavior would produce fluctuations in the visible emission of the arc and fluctuations in the thickness of the sheath. This phenomenon could be described as the plasma “bouncing” on the metal surface. It is known that arcs are unstable, and fluctuations of this sort have been described by Jüttner [@Juttner] and Anders, see Fig 3.22 of Ref. [@Anders]. In rf accelerator cavities, we see oscillations in visible light detected by a phototube with a frequency of approximately 200 MHz that could be due to this effect, see Fig \[PMT\] and Fig \[FFT\].
We can understand the parameter range involved by estimating the current density required to short the sheath in a given time. $$\begin{aligned}
i_s = \Delta \sigma / \Delta t_s = \epsilon_0 E,\end{aligned}$$ where $\Delta t$ = 1 ns implies currents of about 20 MA/m$^2$ and $\sigma$ is the surface charge density. We can approximate Fowler-Nordheim field emission expression for current density at low fields [@dolan] with, $$\begin{aligned}
i_{FN} = 1.8 \times 10^7 (E/3 \times 10^9)^{16},\end{aligned}$$ where the current density $i$ is expressed in A/m$^2$ and the electric field is in V/m. Thus, when electric fields of 3 GV/m appear on the surface they will produce field emission currents capable of shorting the sheath in 1 ns. Assuming the plasma takes a few ns to return to the original density this would imply an instability with a time constant of a few ns.
Instabilities in arc evolution are a well studied phenomenon. These strong oscillations may be related to the “ecton” model Mesyats has described, where he assumes micro-explosions with a timescale of $10^{-8}$ s [@mesyats]. The characteristic, discontinuous “chicken track” traces on the interior of tokamak cavities could also be driven by these instabilities [@Anders; @kajita]. These instabilities are the subject of a further study, and will be be reported elsewhere.
We note that the field emission current densities discussed here, multiplied by the areas of melted copper in 805 MHz cavity damage spots, on the order of 2.5 $\times 10^{-7}$ m$^2$, would produce currents on the order of 4 A, roughly equal to the shorting current expected in 805 MHz breakdown events, see Fig 1 of Ref [@InsepovNorem11].
Material Dependence
-------------------
We expect some dependence of the sheath potential and sheath parameters on the properties of the surface material. It has been shown that the evolution of the plasma is primarily driven by field emitted electron beams at high electric field and self-sputtering of surface material driven by ions falling through the sheath potential [@NoremLINAC10; @Noremrf2011; @InsepovNorem11]. Self-sputtering produces a flux of neutral atoms that can raise the plasma density and also fuel the plasma, permitting long plasma lifetimes [@kajita]. Numerical simulations of sputtering yields show that this mechanism is very sensitive to the sheath potential, ion charge distribution, surface (melting) temperature [@NIMBselfsput], and even grain orientation [@NIMBselfsput1].
Anders, in Section 3.7 of Ref [@Anders] explains the “burn voltage” of arcs, [*i.e.*]{} the voltage drop at the cathode, in terms of the Cohesive Energy Rule, where the cohesive energy of the cathode material is essentially the binding energy of the surface atoms. The larger the cohesive energy, the larger the burn voltage (proportional to the sheath potential) required to maintain a plasma. (The burn voltage is related to the sheath potential, but not equal to it, since electrons emitted from the plasma are not necessarily emitted at the plasma potential.) We believe that the mechanism that correlates the cohesive energy to the burn voltage is likely to be self-sputtering, which is determined by interatomic bonding. This data suggests that the sheath potential should be related to the atomic bonding energy, since we have shown that the sheath potential is primarily a function of the plasma temperature and weakly dependent on the density (see Fig. \[plasmapot-n\]), we assume that this the plasma temperature is primarily involved. Measurements relating the burn voltage to plasma temperature (and perhaps crystal orientation) in different materials might explore this correlation.
Plasma Pressure and Surface Damage
----------------------------------
We believe that the nature of the surface damage can provide information on the parameters of the sheath and the arc. The plasma ion flux hitting the surface should rapidly melt the top few layers of the surface. The plasma pressure pushing on the liquid metal surface can generate a Tonks-Frankel like instability [@He], and uneven surfaces produced by this instability will be opposed by the surface tension force, which will tend to flatten the surface. As the dimensions of this instability become smaller, the surface tension force becomes more dominant, producing a correlation between the plasma pressure and the spatial scale of damage. The experimental problem is that surface tension will tend to smooth over the whole melted area when the liquid surface cools, making the melted areas polished and smooth. Our approach is to look for the smallest scale structure visible in arc damage, and assume that cooling has been rapid enough to preserve some evidence of the plasma pressure. For the scale of damage we observe ( $\sim 0.2 \times 10^{-6}$ m) thermal decay times are on the order of, $$\begin{aligned}
x = \sqrt{Dt},
\end{aligned}$$ where $D$ is the thermal diffusivity constant, approximately $1.1 \times10^{-4}$ m$^2$/sec, implying times on the order a few ns or less, depending very strongly on the geometry of both the material and the details of the heat pulsing. We find experimentally that arcs moving in a transverse magnetic field produce the most fine structure.
The pressure exerted on the surface by the plasma is due to: a) the plasma, and b) the electric field. The thermal plasma pressure is due to ion impacts, $p_i = nkT$, and the electric field pressure is defined by, $p_E = \epsilon_0 E^2/2$ [@Jackson_book], the total is then, $p = p_i - p_E$, since the ions push on the surface and the electric field, generally much smaller, pulls. In the limit of small $E$ and $T$, the ion pressure can be a function primarily of the sheath potential, $p_i=ne\phi$. If, due to a variety of reasons, the pressure is unevenly applied, it will produce a deformation in the liquid surface that is opposed by surface tension, see Fig \[p-gamma\]. The approximate scale of these effects is set by the equilibrium radius, $r$, where the radius where the surface tension is balanced by the plasma pressure can obtained by equating the surface tension force around the circumference with the pressure over the whole area of a hemispherical bubble [@Shortley], $$\begin{aligned}
2 \pi r \gamma = \pi r^2 p,\end{aligned}$$ with $\gamma$ equal to the surface tension constant, approximately 1 N/m, depending on temperature, giving $r \sim 2\gamma/p$ [@copper]. For small structures it has been shown by Tolman that this expression should be corrected by a factor $\delta$, using the expression, $$\begin{aligned}
r = \frac{2\gamma}{p} \left( 1-\frac{\delta}{r} + \cdots \right),\end{aligned}$$ where $\delta$ is the Tolman length [@Tolman49]. The Tolman length is generally evaluated using Molecular Dynamics, and estimates vary from tenths of molecular dimensions to hundredths of atomic dimensions. For radii, $r$, on the order of 100 nm this correction is not significant.
There are many types of arc damage that have been seen SEM images [@Anders; @kajita; @castano]. The damage from a single event is generally circular, in the range of 5 - 200 $\mu$m in diameter, and frequently craterlike with a raised rim. The damage usually shows signs of melting. If the surface has absorbed significant energy, fine structure from the arc can be lost as the metal solidifies, However, if the arc deposits little energy to the surface or cools quickly, for example in SEM images of damage, Fig \[SEM\], a) from 201 MHz rf coupler, and b) from arc damage from Castano [@castano] and images from laser damage [@Schwirzke91], we find complex structures on the scale of 100 - 300 nm, which are not seen in arc damage where large amounts of energy ($\sim$1 J) were present. We assume that if large amounts of energy are transmitted through an arc crater there is less small scale structure, consistent with high stored heat keeping the metal liquid until the surface tension smoothed off the surface. Classic unipolar arc tracks [@kajita] (where magnetic fields move the arc in rapid retrograde motion) are associated with more fine structure, consistent with faster liquid cool down preserving this fine structure.
Simulations of unipolar arcs using PIC codes [@InsepovNorem11] have shown that the plasma potential seems to stay approximately 50 to 75 V during the development of the arc, thus the variation in plasma pressure is primarily due to the plasma ion density. Schwirzke showed that unipolar arcs could produce holes 5 times deeper than their diameter ( 0.7 $\mu$m) [@Schwirzke91]. If we assume that these structures grew from craters with $r \leq$ 0.2 - 0.35 $\mu$m, and the plasma potential, $\phi$ was 50 V, this would imply that the density of the plasma had to be at least $1 - 4 \times 10^{24}$ m$^{-3}$, see Fig \[r-n\]. This is consistent with estimates made from the PIC code, which would not be expected to be reliable at these high densities.
The primary arc damage that results in further high enhancement factors and further breakdown events is likely due to this sub-micron damage, coming either from the plasma pressure itself producing a turbulent surface if it can quickly cool, or cracks produced when the large molten area beneath the arcs cools from the melting point of copper to room temperature leaving a network of surface cracks. The production of high enhancement factors in surface cracks has been demonstrated in Ref [@NoremLINAC10; @Noremrf2011]. The sub-micron component of arc damage thus appears to be both the most direct indicator of the internal parameters of the arc plasma, and (when cracks and crack junctions are considered) the most likely to produce further breakdown events due to high enhancement factors.
Conclusions
===========
We used molecular dynamics simulations to study the non-ideal plasma sheath at a metal surface for conditions we believe are appropriate to those found in accelerator cavities or unipolar arcs. The simulations started from the uniform equilibrium plasma state. Then the relaxation of the electron density profile with formation of the plasma sheath was observed. The relaxation time was found to be of the order of $\sim$ 100 fs. It was shown that the plasma sheath length depends on the electron number density in a slightly different way than the usual expression for the Debye radius due to a non-exponential charge profile at short distances. The values of the sheath length and the surface field were obtained for two values of temperatures and a wide density range with the non-ideality parameter $\Gamma = 0.1 - 4$. We compare the MD results with the contemporary theoretical models and with experimental data from damage. When we compare the plasma conditions that would result from these sheaths with data we find damage consistent with the high plasma pressures implied by the MD and PIC results.
We find that the high density plasma these results imply and the level of plasma pressure they would produce is consistent with the spatial scale of arc damage in rf cavities, in examples where the arc would cool before this structure would be lost. It appears that the sub-micron component of arc damage is both the most direct indicator of the internal parameters of the arc plasma, and, in the case of cracks, the most likely source of further breakdown events due to high enhancement factors. The high surface fields implied by these results could produce field emission that would short the sheath and cause an instability in the time structure of the arc. The relation between self sputtering and the burn voltage is not well understood but the two seem to be closely correlated. We find that the sub-micron component of arc damage, the burn voltage, and fluctuations in the visible light production of arcs may be the most direct indicators of the sheath parameters of the dense plasma.
Acknowledgements
================
We thank the staff of the Accelerator and Technical Divisions at Fermilab and the Muon Accelerator Program (MAP) for supporting and maintaining the MAP experimental program in the MTA experimental area. The work at Argonne is supported by the U.S. Department of Energy Office of High Energy Physics under Contract No. DE-AC02-06CH11357. I. Morozov acknowledges the support by the Programs of Fundamental Research of RAS Nos. 2, 13 and 14. Computations were performed on clusters MIPT-60 and K100 (KIAM RAS).
[99]{}
R. F. Earhart, Phil. Mag. [**1,**]{} 147 (1901).
A. Anders, [*Cathodic Arcs, From Fractal Spots to Energetic Condensation*]{}, Springer, New York (2008), Chapter 3.
B. Jüttner, J. of Phys. D: Appl. Phys., [**34**]{}, (2001) R1-3-R123.
G. Mesyats, IEEE Trans. Plasma Sci., [**23**]{} (1995) 879.
S. Kajita, N. Ohno, S. Kakamura, J. Nuc. Mat. [**415**]{} (2011) 542.
http://clic-study.web.cern.ch/clic-study/.
http://map.fnal.gov/
http://www.iter.org/mach
Z. Insepov, J. Norem, D. Huang, S. Mahalingam, S. Veitzer, Proceedings of PAC09, Vancouver, B. C., Canada, May 4-8, (2009) 800.
Z. Insepov, J. Norem, T. Proslier, S. Mahalingam, S. Veitzer, Proceedings of LINAC10, Tsukuba Japan, Sept 12-17, (2010) 205.
Z. Insepov, J. Norem, S. Veitzer, S. Mahalingam, Proceedings of RF2011, June 1-3, Newport R. I. (to be published), arXiv:1108.0861.
Z. Insepov, J. Norem, Th. Proslier, A. Moretti D. Huang, S. Mahalingam, S. Veitzer; arXiv:1003.1736v3 (2010).
F.R. Schwirzke; IEEE T. Plasma Sci. **19**, No. 5, 690 (1991).
T. Pschiwul, G. Zwicknagel; J. Phys. A, **36**, 6251 (2003).
I.V. Morozov, G.E. Norman; JETP, **100**, 370 (2005).
L.X. Benedict, J.N. Glosli, D.F. Richards, F.H. Streitz, P. Hau-Riege, R.A. London, F.R. Graziani, M.S. Murillo, J.F. Benage; Phys. Rev. Lett. **102**, 205004 (2009).
Z. Donko; J. Phys. A, **42**, 214029 (2009).
T. Raitza, H. Reinholz, G. Röpke, I. Morozov, E. Suraud; Contrib. Plasma Phys. **49**, 498 (2009).
M. Belkacem, F. Megi, P.-G. Reinhard, E. Suraud, and G. Zwicknagel; Eur. Phys. J. D **40**, 247 (2006).
G. Zwicknagel, T. Pschiwul; Contrib. Plasma Phys. **43**, 393 (2003).
A.V. Filinov, M. Bonitz, W. Ebeling; J. Phys. A **36**, 5957 (2003).
I.V. Morozov, I.A. Valuev; J. Phys. A, **42**, 214044 (2009).
M.P Allen, D.J. Tildesley, [ it Computer simulation of liquids]{}, Oxford University Press Inc., NY, 1987.
A.Y. Kuksin, I.V. Morozov, G.E. Norman, V.V. Stegailov, I.A. Valuev; Mol. Simulat. **31**, 1005 (2005).
G. S. Grest and K. Kremer, Phys. Rev. A, [**33**]{}, (1986) 3628.
W. Ebeling, W.D. Kraeft, D. Kremp, G. Ropke, [*Quantum Statistics of Charged Particle Systems*]{}, Springer, New York (2005), Section 2.9.
J. D. Jackson, *Classical Electrodynamics*, Wiley, New York (1999).
W. W. Dolan, Phys. Rev. , [**91**]{}, 510, (1953).
Z. Insepov, J. Norem, S. Veitzer, Nucl. Instrum. and Meth in Phys. Res. B, [**268**]{}, (2010) 642.
Z. Insepov, J. Norem, A. Moretti, Proceedings of PAC 2011, New York, NY, USA, p895.
J. He, N. M. Miskovsky, P. H. Cutler and M. Chung, J. Appl. Phys. [**68(4)**]{}, (1990) 1475.
G. Shortley and D.Williams, [*Elements of Physics: for Students of Science and Engineering*]{}, Third Edition, Prentice-Hall, Inc., Englewood Cliffs, N.J. 1961 p 401.
T. Matsumoto, H. Fujii, T. Ueda, M. Kamai and K. Nogi, Meas. Sci. Technol. [**16**]{}, (2005) 432.
R. C. Tolman, J. Chem. Phys. [**17**]{}, 333 (1949).
C. H. Castano G., presented at “Workshop on Unipolar Arcs”, Argonne, Jan. 29 (2010), https://twindico.hep.anl.gov/indico/conferenceDisplay.py?confId=69
![\[PIC-MD\] The range of PIC and MD codes. The arrow shows the time development on an arc, as described in ref [@NoremPAC09; @NoremLINAC10; @Noremrf2011; @InsepovNorem11] ](PIC-MD){width="\myfigwidth"}
![\[potential-ei\]Electron-ion interaction potential: dashed line – pure Coulomb, solid line – the one used in this work.](potential-ei){width="\myfigwidth"}
![\[averaging\]General simulation scheme: averaging over an ensemble of initial states taken from an auxiliary equilibrium trajectory (solid curve with points).](averaging){width="\myfigwidth"}
![\[surfield-t\]Dependence of the electric field strength at the surface on time for different electron number densities (shown on the plot); $T = 1$eV.](surfield-t){width="\myfigwidth"}
![\[dp-t\]Development of the electron density profile with time. The time moments are shown in the legend. The density is normalized by the mean density in the initial state; $n_e = 10^{23}\mathrm{m}^{-3}$, $T = 1$eV.](dp-t){width="\myfigwidth"}
![\[dp-z\]Density profiles for electrons and ions at the final stationary state depending on the mean electron number density: (a) $n_e = 10^{23}\mathrm{m}^{-3}$ (b) $n_e = 10^{25}\mathrm{m}^{-3}$, (c) $n_e = 10^{27}\mathrm{m}^{-3}$. In all cases $T = 1$eV.](dp-z-t17 "fig:"){width="0.335\linewidth"} ![\[dp-z\]Density profiles for electrons and ions at the final stationary state depending on the mean electron number density: (a) $n_e = 10^{23}\mathrm{m}^{-3}$ (b) $n_e = 10^{25}\mathrm{m}^{-3}$, (c) $n_e = 10^{27}\mathrm{m}^{-3}$. In all cases $T = 1$eV.](dp-z-t19 "fig:"){width="0.303\linewidth"} ![\[dp-z\]Density profiles for electrons and ions at the final stationary state depending on the mean electron number density: (a) $n_e = 10^{23}\mathrm{m}^{-3}$ (b) $n_e = 10^{25}\mathrm{m}^{-3}$, (c) $n_e = 10^{27}\mathrm{m}^{-3}$. In all cases $T = 1$eV.](dp-z-t21 "fig:"){width="0.30\linewidth"}
![\[charge-z\]Distribution of the plasma charge (in arbitrary units) over the longitudinal direction (log-linear plot). The solid line represents the exponential fit, (a) $n_e = 10^{23}\mathrm{m}^{-3}$ (b) $n_e = 10^{25}\mathrm{m}^{-3}$, (c) $n_e = 10^{27}\mathrm{m}^{-3}$. In all cases $T = 1$eV.](dp-z-tln17 "fig:"){width="0.325\linewidth"} ![\[charge-z\]Distribution of the plasma charge (in arbitrary units) over the longitudinal direction (log-linear plot). The solid line represents the exponential fit, (a) $n_e = 10^{23}\mathrm{m}^{-3}$ (b) $n_e = 10^{25}\mathrm{m}^{-3}$, (c) $n_e = 10^{27}\mathrm{m}^{-3}$. In all cases $T = 1$eV.](dp-z-tln19 "fig:"){width="0.322\linewidth"} ![\[charge-z\]Distribution of the plasma charge (in arbitrary units) over the longitudinal direction (log-linear plot). The solid line represents the exponential fit, (a) $n_e = 10^{23}\mathrm{m}^{-3}$ (b) $n_e = 10^{25}\mathrm{m}^{-3}$, (c) $n_e = 10^{27}\mathrm{m}^{-3}$. In all cases $T = 1$eV.](dp-z-tln21 "fig:"){width="0.315\linewidth"}
![\[lambda-n\]The Debye length (lower solid line) and the sheath lengths obtained from MD simulations depending on the electron number density. Dashed line is related to the exponential fit (see Fig. \[charge-z\]), higher solid line represents the width $\lambda$ given by the relation $\sigma(\lambda) = \sigma(0)/e$ where $\sigma(z)$ is the plasma charge profile (Fig. \[charge-z\]). Left figure: $T = 1$eV, right figure: $T = 10$eV.](lambda-n-t1 "fig:"){width="0.48\linewidth"} ![\[lambda-n\]The Debye length (lower solid line) and the sheath lengths obtained from MD simulations depending on the electron number density. Dashed line is related to the exponential fit (see Fig. \[charge-z\]), higher solid line represents the width $\lambda$ given by the relation $\sigma(\lambda) = \sigma(0)/e$ where $\sigma(z)$ is the plasma charge profile (Fig. \[charge-z\]). Left figure: $T = 1$eV, right figure: $T = 10$eV.](lambda-n-t10 "fig:"){width="0.48\linewidth"}
![\[lambdar-n\]Dependence of the ratio between the sheath lengths obtained from MD and the Debye length for different temperatures and plasma densities (see the legends). Abscissa axis in plot (b) is the plasma nonideality parameter.](lambdar-n "fig:"){width="0.50\linewidth"} ![\[lambdar-n\]Dependence of the ratio between the sheath lengths obtained from MD and the Debye length for different temperatures and plasma densities (see the legends). Abscissa axis in plot (b) is the plasma nonideality parameter.](lambdar-g "fig:"){width="0.48\linewidth"}
![\[surfiled-n\]Dependence of the final surface electric field strength on the electron number density for two values of temperature (shown on prot). MD results are compared with the theoretical estimations from [@Schwirzke91]. Crosses correspond to the Eq. (2) from [@Schwirzke91] where the Debye length is replaced by the sheath lengths $\lambda$ obtained from MD (Fig. \[lambda-n\]).](surfiled-n-t1 "fig:"){width="0.49\linewidth"} ![\[surfiled-n\]Dependence of the final surface electric field strength on the electron number density for two values of temperature (shown on prot). MD results are compared with the theoretical estimations from [@Schwirzke91]. Crosses correspond to the Eq. (2) from [@Schwirzke91] where the Debye length is replaced by the sheath lengths $\lambda$ obtained from MD (Fig. \[lambda-n\]).](surfiled-n-t10 "fig:"){width="0.49\linewidth"}
![\[plasmapot-n\] Plasma potential depending on the density for different temperatures. MD results are compared with the theoretical estimations from [@Schwirzke91].](plasmapot-n){width="\myfigwidth"}
![\[PMT\] The time evolution of visible light from an rf arc in an accelerator cavity during breakdown. The red trace shows the conditions with $B=0$ and the blue trace shows the trace with a solenoidal field of $ B=3$ T. Many experiments have seen similar behavior [@Anders]](PMT){width="\myfigwidth"}
.
![\[FFT\] A Fast Fourier Transform of the B = 0 data in the previous Figure. The peak around 200 MHz is similar to FFT plots of instabilities by Anders [@Anders]. ](FFT){width="\myfigwidth"}
![\[p-gamma\] Plasma pressure, $p$ (red arrows), is opposed by the surface tension 2$\pi\gamma$r forces in the liquid metal.](p-gamma){width="\myfigwidth"}
![\[r-n\] Equilibrium radius from surface tension and plasma pressure for two electron temperature plasmas compared with dimensions from ref [@Schwirzke91] and other measurements in the range ($\sim$80 - 350 $\mu$m), see Fig 14.](r-n){width="\myfigwidth"}
![\[SEM\] a) SEM image of unipolar arc tracks from a 201 MHz cavity coupler, showing considerable structure below 1 $\mu$m., b) Image of unipolar arc damage from Castano [@castano].](SEM){width="\myfigwidth"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
For each $d \in \N$ we construct a $3$-generated group $H_d$, which is a subdirect product of free groups, such that the cohomological dimension of $H_d$ is $d$. Given a group $F$ and a normal subgroup $N \lhd F$ we prove that any right angled Artin group containing the special HNN-extension of $F$ with respect to $N$ must also contain $F/N$. We apply this to construct, for every $d \in \N$, a $4$-generated group $G_d$, embeddable into a right angled Artin group, such that the cohomological dimension of $G_d$ is $2$ but the cohomological dimension of any right angled Artin group, containing $G_d$, is at least $d$. These examples are used to show the non-existence of certain “universal” right angled Artin groups.
We also investigate finitely presented subgroups of direct products of limit groups. In particular we show that for every $n\in \N$ there exists $\delta(n) \in \N$ such that any $n$-generated finitely presented subgroup of a direct product of finitely many free groups embeds into the $\delta(n)$-th direct power of the free group of rank $2$. As another corollary we derive that any $n$-generated finitely presented residually free group embeds into the direct product of at most $\delta(n)$ limit groups.
address: 'Mathematical Sciences, University of Southampton, Highfield, Southampton, SO17 1BJ, United Kingdom.'
author:
- Ashot Minasyan
title: On subgroups of right angled Artin groups with few generators
---
Introduction
============
A *right angled Artin group*, also called a *graph group* or a *partially commutative group* in the literature, is a group which has a finite presentation, where the only permitted defining relators are commutators of the generators. To get such a group, one normally starts with any finite simplicial graph ${\Gamma}$, with vertex set $V{\Gamma}$ and edge set $E{\Gamma}$, and defines the associated right angled Artin group $A=A({\Gamma})$ by the presentation $$A=\langle V{\Gamma}\,\|\, [u,v]=1, \, \forall \{u,v\}\in E{\Gamma}\rangle.$$
The structure of these groups and their subgroups has been a subject of intensive study in the recent years. The results of Haglund, Wise and Agol [@H-W; @Wise-QH; @Agol] show that many previously studied groups (e.g., one-relator groups with torsion, limit groups, fundamental groups of closed hyperbolic $3$-manifolds) virtually embed into right angled Artin groups. Thus the class of subgroups of right angled Artin groups is rather rich. However, the first theorem about right angled Artin groups, proved by Baudisch in [@Baud], asserts that a $2$-generated[^1] subgroup of a right angled Artin group is either free or free abelian. In particular, there exists a single right angled Artin group (e.g., $\Z*\Z^2$) which contains all two-generated subgroups of right angled Artin groups. In [@C-K] Casals-Ruiz and Kazachkov ask whether a similar fact is true for $n$-generated subgroups of right angled Artin groups:
\[q:0\] Given $n \in \N$, does there exist a “universal” right angled Artin group which contains all $n$-generated subgroups of arbitrary right angled Artin groups?
Let ${\mathcal{A}}$ denote the class of all groups which are isomorphic to subgroups of (finitely generated) right angled Artin groups. Motivated by Question \[q:0\], in this note we show that there are no “universal” right angled Artin groups for various subclasses of the class ${\mathcal{A}}$. We start with the following statement, where $\mathbb{F}_2$ denotes the free group of rank $2$, $\mathbb{F}_2^d$ denotes its $d$-th direct power and $\operatorname{cd}(H)$ denotes the cohomological dimension of a group $H$ over the integers.
\[prop:H\_d\] For every $d \in \N$ there exists a $3$-generated subgroup $H_d \leqslant \mathbb{F}_2^d$ such that $H_d$ contains the free abelian group of rank $d$ as a subgroup and $\operatorname{cd}(H_d)=d$.
Since $\mathbb{F}_2^d$ is a right angled Artin group, Proposition \[prop:H\_d\] shows that the answer to Question \[q:0\] is negative already for $n=3$. Indeed any group containing all $H_d$, $d \in \N$, must have infinite cohomological dimension, but it is well-known that the cohomological dimension of a right angled Artin group $A$, associated to a finite graph ${\Gamma}$, is finite and equals to the clique number of ${\Gamma}$[^2] (in particular, $\operatorname{cd}(A) \le |V|$). In other words, $$\operatorname{cd}(A)=\max\{|U| \mid U \subseteq V{\Gamma}\mbox{ such that } \{u,v\} \in E{\Gamma}\mbox{ for all distinct } u,v \in U\}.$$
It is now natural to wonder whether imposing a bound on the cohomological dimension, in addition to the bound on the number of generators, would help. Namely, one can ask whether for every $n \in \N$ there is a single right angled Artin group which contains all $n$-generated groups $G \in {\mathcal{A}}$ with $\operatorname{cd}(G) \le n$. To show that this is not the case, we employ special HNN-extensions.
Let $F$ be a group and let $N \leqslant F$ be any subgroup. The *special HNN-extension of $F$ with respect to $N$* is the group $G$ defined by the (relative) presentation $$\label{eq:spec-HNN}
G=\langle F,t \,\|\, tht^{-1}=h,~\forall\,h \in N \rangle.$$
If $G$ and $A$ are groups then we will write $G \hookrightarrow A$ to say that $G$ can be isomorphically embedded into $A$.
\[thm:rope\] Suppose that $F \in {\mathcal{A}}$ and $N \lhd F$ is a normal subgroup such that $F/N \in {\mathcal{A}}$. Let $G$ be the special HNN-extension of $F$ with respect to $N$, defined by presentation . Then
- $G \in {\mathcal{A}}$;
- if $A$ is any right angled Artin group such that $G \hookrightarrow A$ then $F/N \hookrightarrow A$.
The main technical result, used in the proof of Theorem \[thm:rope\], is Proposition \[prop:emb\_quot\] below, which implies that any right angled Artin group containing the special HNN-extension , with $N \lhd F$, must also contain the quotient $F/N$.
Given any $d \in \N$, let us apply Theorem \[thm:rope\] to the case when $F$ is the free group of rank $3$ and $N \lhd F$ is the normal subgroup such that $F/N \cong H_d$, where $H_d$ is the group from Proposition \[prop:H\_d\]. Then the special HNN-extension $G_d$, of $F$ with respect to $N$, will be generated by $4$ elements and will belong to the class ${\mathcal{A}}$ by Theorem \[thm:rope\]. And for any right angled Artin group $A$ with $G_d \hookrightarrow A$, we will have $H_d \hookrightarrow A$. In particular, $\operatorname{cd}(A) \ge \operatorname{cd}(H_d)=d$. On the other hand, we see that $\operatorname{cd}(G_d)=2$ because $G_d$ is not free and acts on a tree with free vertex stabilizers (cf. [@Brown Ex. 4 in Ch. VIII.2]). Thus we obtain the following corollary:
\[cor:G\_d\] For every $d \in \N$ there exists a $4$-generated group $G_d \in {\mathcal{A}}$ such that $\operatorname{cd}(G_d)=2$ but for any right angled Artin group $A$, containing a copy of $G_d$, one has $\operatorname{cd}(A) \ge d$.
The above corollary shows there there exists no “universal” right angled Artin group containing all $4$-generated groups from ${\mathcal{A}}$ of cohomological dimension $2$.
The next naturally arising question could ask whether for a given $n \in \N$ there are right angled Artin groups containing copies of all finitely presented $n$-generated groups $G \in \mathcal A$. This question seems to be more subtle. The groups $H_d$ from Proposition \[prop:H\_d\] are not finitely presented (see Remark \[rem:not\_fp\] below), and a result of Bridson and Miller [@B-M] implies that finitely presented subdirect products of free groups, intersecting each of the factors non-trivially, are virtually surjective on pairs, which shows that they are quite scarce. In [@B-H-M-S] Bridson, Howie, Miller and Short extended this result to finitely presented subgroups in direct products of finitely generated fully residually free groups (a.k.a. limit groups). We use this generalization to prove the following theorem.
\[thm:adeq\_free\] There exists a computable function $\delta: {\mathbb N}\to \N$ such that the following holds. Suppose that $F_1,\dots,F_d$ are limit groups and $H \leqslant F_1 \times \dots \times F_d$ is a finitely presented subgroup such that $H \cap F_i \neq \{1\}$ for each $i=1,\dots,d$. If $H$ is $n$-generated then $d \le \delta(n)$.
In Remark \[rem:delta\_estimate\] we give an estimate $\delta(n) < 2^{18n^3}$ for all $n \in \N$. This estimate is not sharp, and it would be interesting to see whether there exists a polynomial (or even a linear) upper bound for $\delta(n)$ in terms of $n$.
Theorem \[thm:adeq\_free\] implies that no finitely presented analogues of the groups $H_d$ from Proposition \[prop:H\_d\] can be found among subgroups of direct product of free groups:
\[cor:univ\_for\_free\] Given any $n \in \N$, every $n$-generated finitely presented subgroup $H$ of a direct product of finitely many free groups can be embedded into ${\mathbb{F}}_2^{\delta(n)}$. In particular, $\operatorname{cd}(H) \le \delta(n)$.
We would also like to mention another corollary of Theorem \[thm:adeq\_free\] concerning residually free groups, which may be of independent interest. It is known that every finitely generated residually free group $G$ can be embedded in the direct product of finitely many limit groups (see Baumslag, Myasnikov and Remeslennikov [@B-M-R], Kharlampovich and Myasnikov [@Kh-M], or Sela [@Sela]). In the case when $G$ is finitely presented Theorem \[thm:adeq\_free\] can be used to give a bound on the number of factors:
\[cor:res\_free\] Let $G$ be an $n$-generated finitely presented residually free group, for some $n \in \N$. Then $G$ can be embedded into the direct product of at most $\delta(n)$ limit groups.
Using the commutative transitivity of limit groups it is easy to show that for any $d \in \N$ the residually free group $H_d$ from Proposition \[prop:H\_d\] cannot be embedded into the direct product of less than $d$ limit groups. Thus it is not possible to remove the assumption that $G$ is finitely presented in Corollary \[cor:res\_free\].
Unfortunately the characterization of finitely presented subdirect product of free groups given by Bridson and Miller in [@B-M] does not extend to finitely presented subgroups of more general right angled Artin groups. And while Theorem \[thm:rope\] can be used to construct some finitely presented examples (see Corollary \[cor:fp\_ex\] in Section \[sec:5\]), the following question remains open.
\[q:1\] Do there exist finitely presented analogues of groups $H_d$ from Proposition \[prop:H\_d\]? More precisely, does there exist $n \in \N$ such that for any $d \in \N$ there is an $n$-generated finitely presented group $Q_d \in {\mathcal{A}}$ which contains the free abelian subgroup of rank $d$ (and/or which has $\operatorname{cd}(Q_d)=d$)?
A celebrated theorem of Higman, Neumann and Neumann [@H-N-N] states that any countable group $H$ can be embedded into a $2$-generated group $G$; moreover, one can take $G$ to have the same number of defining relations as $H$. It is natural to ask if there is an analogue of this statement within the class ${\mathcal{A}}$:
\[q:2\] Does there exist $n \in \N$ such that every group $H \in {\mathcal{A}}$ can be embedded into some $n$-generated group $G \in {\mathcal{A}}$? Can one choose such $G$ to be finitely presented, in addition to being generated by $n$ elements?
Note that by the theorem of Baudisch [@Baud], mentioned above, the number $n$ in Question \[q:2\] must be at least $3$.
Proof of Proposition \[prop:H\_d\]
==================================
Let $F=F(x,y,z)$ be the free group of rank $3$ with the free generating set $\{x,y,z\}$. Given any $d \in \N$, choose any elements $w_1,\dots,w_d \in \langle x,y \rangle\leqslant F$ which freely generate a free subgroup of rank $d$ in $\langle x,y \rangle \cong\mathbb{F}_2$. For each $i=1,\dots,d$, define the normal subgroup $L_i \lhd F$ as the normal closure of the element $z^{-1}w_i$, and let $$K_i\coloneq \bigcap_{j=1, j \neq i}^d L_j \lhd F.$$ By construction, for every $i$ the quotient $F/L_i$ is canonically isomorphic to the rank $2$ free subgroup $\langle x,y \rangle \leqslant F$. Observe that $K_d \setminus L_d \neq \emptyset $. Indeed, clearly the long commutator $$[z^{-1}w_1,\dots,z^{-1}w_{d-1}] \coloneq [[[z^{-1}w_1,z^{-1}w_2],z^{-1}w_3],\dots,z^{-1}w_{d-1}]$$ belongs to $K_d$, and its image in $F/L_d \cong \langle x, y \rangle \leqslant F$ is equal to the long commutator $[w_d^{-1}w_1,\dots,w_d^{-1}w_{d-1}]$. The latter is non-trivial in $\langle x,y \rangle \leqslant F$ by the choice of $w_1,\dots,w_d$ (because the elements $w_d^{-1}w_1,\dots,w_d^{-1}w_{d-1}$ freely generate a free subgroup of rank $d-1$), hence $[z^{-1}w_1,\dots,z^{-1}w_{d-1}] \in K_d \setminus L_d$ in $F$. In view of the symmetry, we can deduce that $$\label{eq:K_i-L_i}
K_i \setminus L_i \neq \emptyset \mbox{ in $F$, for each } i \in \{1,\dots,d\}.$$
Let $\varphi_i:F \to F/L_i$ denote the canonical epimorphism, and let $\varphi: F \to F/L_1 \times \dots \times F/L_d$ denote the product homomorphism, defined by the formula $\varphi(f) \coloneq \bigl( \varphi_1(f),\dots,\varphi_d(f)\bigr)$ for all $f \in F$. Finally, we define $H_d \coloneq \varphi(F)$ as the image of $F$ in $F/L_1 \times \dots \times F/L_d\cong \mathbb{F}_2^d$.
Notice that $\varphi_j(K_i)=\{1\}$ if $j \neq i$ by the definition of $K_i$, and $\varphi_i(K_i) \neq \{1\}$ by for any $i \in \{1,\dots,d\}$. It follows that the subgroup $\varphi(K_1 \cdots K_d) \leqslant H_d$ is naturally isomorphic to the direct product $\varphi_1(K_1) \times\dots\times \varphi_d(K_d)$ of non-trivial free groups. Consequently, $H_d$ contains a copy of $\Z^d$ as a subgroup, and so $\operatorname{cd}(H_d) \ge d$. On the other hand, $\operatorname{cd}(H_d) \le \operatorname{cd}(\mathbb{F}_2^d)=d$, hence $\operatorname{cd}(H_d)=d$, as claimed.
Finitely presented subdirect products of limit groups
=====================================================
The goal of this section is to prove Theorem \[thm:adeq\_free\] and Corollaries \[cor:univ\_for\_free\],\[cor:res\_free\] formulated in the Introduction.
Let $H$ be a subgroup of the direct product $P\coloneq F_1 \times \dots \times F_m$ of some groups $F_1, \dots,F_m$. Recall that $H$ is called a *subdirect product* if it projects surjectively onto each of the direct factors. Following [@B-H-M-S] we will say that $H$ is *virtually surjective on pairs* (VSP) if for any pair of distinct indices $i,j \in \{1,\dots,m\}$ the image of $H$ under the canonical projection $\rho_{ij}: P \to F_i \times F_j$ has finite index in $F_i \times F_j$. Let $N_i \coloneq H \cap \ker \rho_i \lhd H$, $i=1,\dots,m$, where $\rho_i:P \to F_i$ denotes the canonical projection to the $i$-th coordinate.
\[lem:VSP-reform\] Let $H\leqslant F_1 \times \dots \times F_m$ be a subdirect product. Then $H$ is VSP if and only if $|H:N_iN_j|<\infty$ for all $1 \le i <j \le m$.
If $H$ is VSP then $|F_i \times F_j:\rho_{ij}(H)|<\infty$ whenever $i < j$, hence $F_i \cap \rho_{ij}(H)=\rho_{ij}(N_j)=\rho_i(N_j)$ has finite index in $F_i$. Therefore $\rho^{-1}_i(\rho_i(N_j)) \cap H=N_iN_j$ has finite index in $H$.
On the other hand, if $|H:N_iN_j|<\infty$ for some $0 \le i <j \le m$, then $\rho_{i}(N_iN_j)=\rho_i(N_j)$ has finite index in $\rho_i(H)=F_i$. But $\rho_i(N_j)=\rho_{ij}(N_j) \subseteq \rho_{ij}(H)$ because $\rho_{ij}(N_j)$ is contained in $F_i$ as $N_j \subseteq \ker \rho_j$. Thus $\rho_{ij}(H)$ contains a finite index subgroup of $F_i$. Similarly, it also contains $\rho_{j}(N_i)$, which has finite index in $F_j$. Hence $|F_i \times F_j :\rho_{ij}(H)|<\infty$.
\[rem:not\_fp\] In [@B-M] Bridson and Miller proved that if a finitely presented subdirect product of free groups intersects each of the direct factors non-trivially then it is VSP[^3]. It follows that for any $d \in \N$, the group $H_d$, constructed in the proof of Proposition \[prop:H\_d\], is not finitely presented. Indeed, if $H_d$ was VSP then $|H_d/(N_iN_j)|<\infty$ for any $1 \le i < j \le d$ by Lemma \[lem:VSP-reform\]. But clearly $H/(N_iN_j) \cong F/(L_iL_j)$ is an infinite group because it has a presentation with $3$ generators and two relators (in particular, it surjects onto $\Z$).
The following elementary observation will be useful.
\[lem:image\_of\_prod\] Suppose that $H$ and $F$ are groups and $\psi_i:H \to F$, $i=1,2$, are two epimorphisms. Assume that $H$ is generated by some subset $X \subseteq H$, and let $M \lhd F$ denote the normal closure of the subset $\{ \psi_1^{-1}(x)\psi_2(x) \mid x \in X \}$. Then $\psi_1(\ker \psi_2) \subseteq M$; in particular, if one sets $L \coloneq \ker\psi_1\ker\psi_2 \lhd H$, then $H/L$ maps onto $F/M$.
If $h \in \ker\psi_2$, then $h=x_1\dots x_k$, for some $x_1,\dots,x_k \in X^{\pm 1}$, and $\psi_2(x_1) \dots \psi_2(x_k)=1$ in $F$. Since $\psi_2(x_j) \equiv \psi_1(x_j)$ (mod $M$), we see that $\psi_1(h)=\psi_1(x_1)\dots\psi_1(x_k) \in M$. Thus $\psi_1(\ker\psi_2) \subseteq M$, as claimed.
\[prop:vsp\] There is a computable function $\gamma: {\mathbb N}\to \N$ such that the following holds. Let $H$ be a group with a collection of normal subgroups $N_1,\dots,N_m \lhd H$, such that $H/N_i$ maps onto the free group of rank $2$ for each $i=1,\dots,m$, and $|H:N_iN_j|<\infty$ for all $1\le i<j\le m$. If $H$ is $n$-generated, for some $n \in \N$, then $m \le \gamma(n)$.
Let $F$ denote the free group of rank $2$. By the assumptions, for every $i=1,\dots,m$, there exists an epimorphism $\psi_i: H \to F$ whose kernel contains $N_i$.
Suppose that some elements $x_1,\dots,x_n \in H$ generate $H$, and let $u_{ik} \coloneq \psi_i(x_k) \in F$, $k=1,\dots,n$, $i=1,\dots,m$. For any two distinct indices $i,j \in \{1,\dots,m\}$ let $L_{ij}\coloneq \ker\psi_i \ker\psi_j \lhd H$ and let $M_{ij}$ be the normal closure of the subset $\{u_{i1}^{-1}u_{j1},\dots,u_{in}^{-1}u_{jn}\}$ in $F$. By construction, the group $H/(N_iN_j)$ maps onto the group $H/L_{ij}$, and the latter group maps onto the group $F/M_{ij}$ by Lemma \[lem:image\_of\_prod\]. Since $|H/(N_iN_j)|<\infty$ by the assumptions, the group $F/M_{ij}$ must also be finite whenever $1 \le i<j\le m$.
Let $\{D_lF\}_{l=1}^\infty$ denote the Jennings-Zassenhaus filtration of the free group $F$ for the prime $p=2$. Namely, $D_lF\coloneq \{f \in F \mid f-1 \in I^n\}$, where $I$ is the augmentation ideal of the group ring $\mathbb{K}[F]$ and $\mathbb{K}$ is the field of two elements. Then $D_lF$ is a normal subgroup of finite index in $F$ for each $l \in \N$, and $F=D_1F \geqslant D_2F \geqslant \dots$.
Now, for the given $n \in \N$ choose $\tau \coloneq 2/3$ and $r\coloneq\lceil 2\log_2(3n)\rceil \in \N$ (so that $n \tau^r<2\tau-1$), and set $\gamma(n)\coloneq |F/D_rF|^n \in \N$. Let us show that $m \le \gamma(n)$. Arguing by contradiction, suppose that $m>\gamma(n)$. Then, by the pigeon hole principle, there must exist a pair of indices $i,j \in \{1,\dots,m\}$, $i<j$, such that $u_{ik} \equiv u_{jk}$ (mod $D_rF$) for all $k=1,\dots,n$. It follows that the elements $u_{i1}^{-1}u_{j1},\dots,u_{in}^{-1}u_{jn}$, normally generating $M_{ij}$, are all contained in $D_rF$. Since $F$ is the free group of rank $2$ and $$1-2 \tau +\sum_{i=1}^n \tau^{r}<0,$$ the Golod-Shafarevich theorem (see [@Ershov]) implies that the group $F/M_{ij}$ must be infinite. This contradiction shows that $m \le \gamma(n)$, finishing the proof of the proposition.
\[rem:gamma\_estimate\] We can estimate the value of $\gamma(n)$, for the choice of $\tau=2/3$, $r=\lceil 2\log_2(3n)\rceil$ and the field of two elements $\mathbb{K}$, as follows. Let $Y$ be a generating set for the free group $F$ with $|Y|=2$. Then, for any $l \in \N$, the vector space $I^l/I^{l+1}$, over $\mathbb K$, is spanned by the elements $(y_1-1)(y_2-1) \cdots(y_l-1)+I^{l+1}$, for all $y_1,\dots,y_l \in Y$ (see the proof of [@Grig Lemma 8]). Hence $\dim_\mathbb{K}(I^l/I^{l+1}) \le 2^l$, and so $|I^l/I^{l+1}|\le 2^{2^l}$. Since $|D_lF/D_{l+1}F| \le |I^l/I^{l+1}|$ and $F=D_1F$, we obtain $$|F/D_rF|=|D_1F/D_2F| \cdots |D_{r-1}F/D_rF| \le 2^{2^{r}-1} \le 2^{18 n^2-1}.$$ Thus $\gamma(n) \le 2^{18n^3-n}$ for all $n \in \N$.
For any $\varepsilon>0$ by taking the parameter $\tau$ to be very close to $1/2$ (but still larger than 1/2) and choosing an appropriate $r$ we can get a better estimate $\gamma(n) \le 2^{Cn^{2+\varepsilon}-n}$, where $C$ will be some large constant depending on $\tau$.
Recall that a group $G$ is said to be *residually free* if for every $g \in G\setminus \{1\}$ there exists a free group $F$ and a homomorphism $\phi:G \to F$ such that $\phi(g) \neq \{1\}$ in $F$. The group $G$ is *fully residually free* if for every finite subset $S \subseteq G$ there is a homomorphism from $G$ to a free group whose restriction to $S$ is injective. Finitely generated fully residually free groups are commonly called *limit groups*. These groups play an important role in the study of the first order theory of free groups (see [@Kh-M; @Sela]).
In [@B-H-M-S] Bridson, Howie, Miller and Short proved that a finitely presented subgroup of a direct product of finitely many non-abelian limit groups, which is subdirect and intersects each of the direct factors non-trivially, must be VSP. We will now use this result to prove Theorem \[thm:adeq\_free\].
Let $\rho_i: F_1 \times \dots \times F_d \to F_i$ denote the canonical projection onto $F_i$. Since finitely generated subgroups of limit groups are also limit groups, we can replace each $F_i$ by $\rho_i(H)$ to further assume that $H$ is subdirect.
After renumbering, if necessary, we can suppose that $F_1,\dots, F_s$ are abelian and $F_{s+1},\dots,F_d$ are non-abelian limit groups, for some $s \in \{0,\dots,d\}$. Let $H_1$ and $H_2$ denote the images of $H$ under the canonical projections to $F_1 \times \dots \times F_s$ and $F_{s+1} \times \dots \times F_d$ respectively. Since $\{1\} \neq F_i \cap H \subseteq F_i \cap H_1$ for all $i =1,\dots,s$, the rank $r$, of the free abelian group $H_1$, must be at least $s$. On the other hand, $r \le n$ because $H_1$ is $n$-generated, as a quotient of $H$, and the free abelian group of rank $r$ cannot be generated by fewer than $r$ elements. Hence $s \le n$.
Note that $H_2$ is finitely presented as a quotient of the finitely presented group $H$ by a finitely generated central subgroup $Z \coloneq H \cap (F_1 \times \dots \times F_s)$ ($F_1 \times \dots \times F_s$ is a free abelian group of rank at most $ns$, so $Z$ can be generated by no more than $ns$ elements). By construction, $H_2 \leqslant F_{s+1} \times \dots \times F_d$ is a subdirect product of non-abelian limit groups intersecting each factor non-trivially. Therefore, by [@B-H-M-S Thm. D(5)], $H_2$ must be VSP in $F_{s+1} \times \dots \times F_d$.
From Lemma \[lem:VSP-reform\] it follows that $|H_2:N_iN_j|<\infty$ for all $s+1 \le i<j \le d$, where $N_i\lhd H_2$ denotes the kernel of the projection onto $F_i$, $i=s+1,\dots,d$. Since $H_2/N_i \cong F_i$ is a non-abelian limit group for $i=s+1,\dots,d$, it has an epimorphism onto ${\mathbb{F}}_2$, so all the assumptions of Proposition \[prop:vsp\] are satisfied, and we can use it to conclude that $d-s \le \gamma(n)$. Thus, $d \le \delta(n)$, where $\delta(n) \coloneq n +\gamma(n)$ for all $n \in \N$, and the theorem is proved.
\[rem:delta\_estimate\] In view of Remark \[rem:gamma\_estimate\] we can estimate that $\delta(n)\le n+ 2^{18n^3-n} <2^{18n^3}$ for all $n \in \N$.
If a subgroup $H$, of a direct product $F_1 \times \dots \times F_m$, has trivial intersection with one of the factors, then $H$ is isomorphic to its image under the projection away from this factor. So we can always find an embedding of $H$ into $F_{i_1} \times \dots \times F_{i_k}$, for some $\{i_1,\dots,i_k\} \subseteq \{1,\dots,m\}$, such that $H \cap F_{i_j} \neq \{1\}$ for every $j=1,\dots,k$. In view of Theorem \[thm:adeq\_free\], this observation together with the fact that every finitely generated free group embeds into ${\mathbb{F}}_2$ yields Corollary \[cor:univ\_for\_free\]. Corollary \[cor:res\_free\] is obtained similarly, by first embedding the finitely generated residually free group $G$ into the direct product of finitely many limit groups (which can be done by [@B-M-R Cor. 19]; see also [@Kh-M Cor. 2] or [@Sela Claim 7.5]), and then removing the redundant direct factors that intersect the image of $G$ trivially.
Background on right angled Artin groups {#sec:b-n}
=======================================
Let ${\Gamma}$ be a finite simplicial graph with vertex set $V{\Gamma}=V$ and the edge set $E{\Gamma}=E$. For any vertex $v \in V$, the *link* $\operatorname{link}_{\Gamma}(v) \subseteq V$ is defined as the set of all vertices of ${\Gamma}$ adjacent to $v$ (not including $v$ itself). If $U \subseteq V$ then $\operatorname{link}_{\Gamma}(U) \coloneq \bigcap_{u \in U} \operatorname{link}_{\Gamma}(u)$. A subset $U \subseteq V$ is *reducible* if it can be decomposed in a disjoint union $U=X \sqcup Y$ of non-empty subsets $X,Y \subset V$ such that $Y \subseteq \operatorname{link}_{\Gamma}(X)$. If there is no such decomposition, then $U$ is said to be [*irreducible*]{}.
Let $A=A({\Gamma})$ be the right angled Artin group, associated to ${\Gamma}$. Then the elements of $V$ can be thought of as generators of $A$, and every element $g \in A$ can be represented by a word $w$ over the alphabet $V^{\pm 1}$. The *support* of $w$ is the set of all vertices $v \in V$ such that either $v$ or $v^{-1}$ appears in $w$. The *length* $|g|_{\Gamma}$ and the *support* $\operatorname{supp}_{\Gamma}(g)$, of $g \in A$, are defined as the length and the support of a shortest word over $V^{\pm 1}$ representing $g$ in $A$, respectively. It is a standard fact that $|g|_{\Gamma}$ and $\operatorname{supp}_{\Gamma}(g)$ only depend on $g$ and are independent of the choice of a shortest word representing it.
Right angled Artin groups are special cases of *graph products of groups*, when all the vertex groups are infinite cyclic (see [@A-M Subsec. 2.2] for some background on graph products).
Given any subset $U \subseteq V{\Gamma}$, the subgroup $A_U \leqslant A$, generated by $U$, is called a *full subgroup* (or a *special subgroup*) of $A$. It is not difficult to see that $A_U$ is naturally isomorphic to the right angled Artin group corresponding to the full subgraph ${\Gamma}_U$ of ${\Gamma}$, spanned by the vertices from $U$. Moreover, $A_U$ is a retract of $A$, where the *canonical retraction* $\rho_U:A \to A_U$ is defined on the generating set $V$ of $A$ by the formula $\rho_U(u)=u$ if $u \in U$ and $\rho_U(v)=1$ if $v \in V \setminus U$.
Any conjugate of a full subgroup in $A$ is said to be *parabolic*. Every subset $M \subseteq A$ is contained in a unique minimal (with respect to inclusion) parabolic subgroup ${{\rm Pc}}_{\Gamma}(M)$, which is called the *parabolic closure* of $M$ in $A$ (see [@A-M Prop. 3.10]).
An element $t \in A$ is *cyclically reduced* if it has minimal length in its conjugacy class in $A$. Servatius [@Serv Prop. on p. 38] proved that for every element $g \in A$ there exist $u,t \in A$ such that $g=utu^{-1}$, $t$ is cyclically reduced and $|g|_{\Gamma}=|t|_{\Gamma}+2|u|_{\Gamma}$.
If $t \in A$ and $H \leqslant A$ then we will write ${{\rm C}}_H(t)$ and ${{\rm N}}_A(H)$ to denote the centralizer of $t$ in $H$ and the normalizer of $H$ in $A$ respectively.
\[lem:centr\_of\_powers\] If $A$ is a right angled Artin group then for any $t \in A$ and $m \in \Z\setminus\{0\}$, ${{\rm C}}_A(t^m)={{\rm C}}_A(t)$.
Suppose that $s t^m s^{-1}=t^m$ for some $s \in A$. Then $(sts^{-1})^m=t^m$, and since right angled Artin groups have the Unique Root property (see, for example, [@M-RAAG Lemma 6.3]), we can deduce that $sts^{-1}=t$. Thus ${{\rm C}}_A(t^m) \subseteq {{\rm C}}_A(t)$. The reverse inclusion is obvious, hence the lemma is proved.
Special HNN-extensions in right angled Artin groups {#sec:5}
===================================================
In this section we give a criterion for embedding the special HNN-extension of a group with respect to a normal subgroup into a right angled Artin group. The next lemma uses the obvious fact that the class of right angled Artin groups contains the infinite cyclic group and is closed under forming free and direct products.
\[lem:spec-HNN-emb\] Suppose that $F$ is a subgroup of some right angled Artin group $A$, $N \lhd F$ and $F/N$ is embeddable into some right angled Artin group $B$. Then the special HNN-extension can be embedded into the right angled Artin group $C\coloneq A \times (B*\langle t \rangle)$, where $\langle t \rangle$ denotes the infinite cyclic group generated by $t$.
Abusing the notation, let us assume that $F \leqslant A$ and $F/N \leqslant B$. Let $\phi:F \to B$ denote the natural homomorphism with $\ker(\phi)=N$ and $\mathrm{im}(\phi)=F/N$. Clearly the subgroup ${\overline{F}}\coloneq \{(f,\phi(f)) \mid f \in F\}\leqslant A \times B$ is isomorphic to $F$ and ${\overline{F}}\cap A=N$. Now, let ${\overline{G}}$ be the subgroup of $C= A \times (B*\langle t \rangle)$ generated by ${\overline{F}}$ and $t$.
Evidently, $t$ commutes with $(a,b) \in A \times B$ in $C$ if and only if $b=1$, thus ${{\rm C}}_{{\overline{F}}}(t)={\overline{F}}\cap A=N$. This naturally gives rise to a homomorphism $\psi: G \to {\overline{G}}$, defined according to the formula $\psi(f)\coloneq (f,\phi(f))\in {\overline{F}}$ and $\psi(t)=t$, where $G$ is the special HNN-extension , of $F$ with respect to $N$. Obviously $\psi$ is surjective, so it remains to show that it is injective. This can be easily done by using normal forms: suppose that $g \in G \setminus \{1\}$. Then $g=f_0t^{k_1}f_1 \cdots t^{k_n}f_n$, where $n \ge 0$, $f_0, \dots,f_n \in F$, $k_1,\dots,k_n \in {\mathbb Z}\setminus\{0\}$, $f_i \notin N$ for $1 \le i \le n-1$, and $f_0 \neq 1$ if $n=0$.
If $n=0$ then $g=f_0\neq 1$ in $F$, hence $\psi(g)=(f_0,\phi(f_0)) \neq 1$ in $A \times B \leqslant C$. So, suppose that $n \ge 1 $ and let $\rho: C \to B*\langle t \rangle$ denote the canonical retraction. Then $\rho(\psi(g))=\phi(f_0)t^{k_1} \phi(f_1) \cdots t^{k_n} \phi(f_n)$ is a non-empty reduced word in the free product $B * \langle t \rangle$, hence $\rho(\psi(g)) \neq 1$. Therefore $\psi(g) \neq 1$ in $C$, and $\psi$ is injective.
Thus $G \cong {\overline{G}}\leqslant C$, and the lemma is proved.
Let $A$ be a right angled Artin group associated to a finite graph ${\Gamma}$ with vertex set $V=V{\Gamma}$. Recall that for any subset $M \subseteq A$, ${{\rm Pc}}_{\Gamma}(M)$ denotes the parabolic closure of $X$ in $A$ (see Section \[sec:b-n\]).
\[prop:emb\_quot\] Let $F$ be a subgroup of $A$, and let $t \in A$ be an element such that ${{\rm Pc}}_{\Gamma}(G)=A$, where $G\coloneq \langle F,t \rangle \leqslant A$. Suppose that $N\coloneq {{\rm C}}_F(t)$ is a normal subgroup of $F$. Then
- ${{\rm Pc}}_{\Gamma}(N)$ is a direct factor of $A$. More precisely, ${{\rm Pc}}_{\Gamma}(N)=A_X$ for some $X \subseteq V$, and $V=X \sqcup \operatorname{link}_{\Gamma}(X)$.
- $F \cap A_X=N$, hence $F/N$ embeds into $A_Y \leqslant A$, where $Y \coloneq V \setminus X=\operatorname{link}_{\Gamma}(X)$.
Clearly $N$ is normal in $G$, hence $G \leqslant {{\rm N}}_A(N) \leqslant {{\rm N}}_A({{\rm Pc}}_{\Gamma}(N))$, where the second inclusion was proved in [@A-M Lemma 3.12]. Suppose that ${{\rm Pc}}_{\Gamma}(N)=h A_X h^{-1}$ for some $X \subseteq V$ and some $h \in A$. By [@A-M Prop. 3.13], ${{\rm N}}_A(hA_Xh^{-1})=hA_{X \cup \operatorname{link}_{\Gamma}(X)}h^{-1}$ is a parabolic subgroup of $A$ containing $G$. Since ${{\rm Pc}}_{\Gamma}(G)=A$, it follows that $A=hA_{X \cup \operatorname{link}_{\Gamma}(X)}h^{-1}=A_{X \cup \operatorname{link}_{\Gamma}(X)}$. Therefore $A=A_X \times A_{\operatorname{link}_{\Gamma}(X)}$, ${{\rm Pc}}_{\Gamma}(N)=A_X$ and $V=X \sqcup Y$, where $Y=\operatorname{link}_{\Gamma}(X)$. Thus (i) is proved.
Let us prove (ii) now. Clearly, after replacing $t$, $F$ and $N$ with their conjugates by the same element, we can assume that $t$ is cyclically reduced in $A$ (note that this does not affect the equality ${{\rm Pc}}_{\Gamma}(N)=A_X$, as $A_X \lhd A$). Let $\operatorname{supp}_{\Gamma}(t)=T_1 \sqcup \dots \sqcup T_k$ be a decomposition into the disjoint union of non-empty irreducible subsets $T_i$ of $V$, such that $T_i \subseteq \operatorname{link}_{\Gamma}(T_j)$ in ${\Gamma}$ whenever $i \neq j$. Then one can write $t=t_1^{n_1} t_2^{n_2} \cdots t_k^{n_k}$, where $t_i \in A_{T_i}\setminus\{1\}$ are not proper powers and $n_i \in\N$ for $i=1,\dots, k$.
In [@Serv] Servatius showed that the centralizer, ${{\rm C}}_A(t)$, is generated by $t_1,\dots,t_k$ and $S\coloneq \operatorname{link}_{\Gamma}(\operatorname{supp}_{\Gamma}(t)) \subseteq V$. In other words, $$\label{eq:centralizer}
{{\rm C}}_A(t)=\langle t_1\rangle \langle t_2 \rangle \cdots \langle t_k \rangle A_S \cong \langle t_1\rangle \times \langle t_2 \rangle \times \dots \times \langle t_k \rangle \times A_S.$$
For every $i \in \{1,\dots,k\}$, let $\rho_i:A \to A_{T_i}$ denote the canonical retraction. Suppose that $\rho_j(N) \neq \{1\}$ for some $j \in \{1,\dots,k\}$. Evidently, since $N \leqslant {{\rm C}}_A(t)$, we have $\rho_j(N) \leqslant \rho_j(C_A(t))=\langle t_j \rangle$ by . Consequently, $\rho_j(N)=\langle t_j^{m_j} \rangle$ for some $m_j \in \Z\setminus\{0\}$. Recalling that $N \lhd F$, we see that $\rho_j(N) \lhd \rho_j(F)$. Since $\rho_j(N) \cong \Z$, as right angled Artin groups are torsion-free, we can use Lemma \[lem:centr\_of\_powers\] to conclude that $\rho_j(N)$ is actually central in $\rho_j(F)$ (see [@A-M-S Lemma 9.1] for an alternative proof), thus $\rho_j(F) \leqslant {{\rm C}}_{A_{T_j}}(t_j^{m_j})$. Therefore, applying Lemma \[lem:centr\_of\_powers\] one more time, we see that $\rho_j(F) \leqslant{{\rm C}}_{A_{T_j}}(t_j)=\langle t_j \rangle$ in $A_{T_j}$.
After renumbering, if necessary, we can assume that there exists $l \in \{0,1, \dots,k\}$ such that $\rho_i(N) \neq \{1\}$ if $1 \le i \le l$, and $\rho_i(N)=\{1\}$ if $l+1 \le i \le k$. The inclusion $N \leqslant {{\rm C}}_A(t)$, together with , implies that ${{\rm Pc}}_{\Gamma}(N) \leqslant A_Z$, where $Z \coloneq \bigcup_{j=1}^l T_j \cup S \subseteq V$. In the previous paragraph we have shown that $\rho_j(F) \leqslant \langle t_j \rangle$ whenever $1 \le j \le l$. Let $\rho_S:A \to A_S$ denote the canonical retraction. Since $A_Z$ is canonically isomorphic to the direct product $A_{T_1} \times \dots \times A_{T_l} \times A_S$, for any $g \in F \cap A_Z$ we have $$g=\rho_1(g) \cdots \rho_l(g) \rho_S(g) \in \langle t_1 \rangle \cdots \langle t_l \rangle A_S \subseteq {{\rm C}}_A(t).$$ Therefore, $F \cap A_Z \subseteq F \cap {{\rm C}}_A(t)=N$ by the assumption, yielding that $F \cap A_X=N$, as $A_X={{\rm Pc}}_{\Gamma}(N) \leqslant A_Z$. Thus the proof of claim (ii) is complete.
Special HNN-extensions play an important role in group theory. For example, Higman used them in the proof of his famous embedding theorem – see [@Higman] and [@L-S Ch. IV.7]. In particular he showed that if $F$ is a finitely generated group and $N \lhd F$ is such that the special HNN-extension can be embedded into a finitely presented group then the quotient $F/N$ can also be embedded into a finitely presented group. Proposition \[prop:emb\_quot\] allows us to obtain the following analogue of the latter statement, where the class of finitely presented groups is replaced by the class of right angled Artin groups.
\[cor:rope-RAAG\] Suppose that $F$ is a group, $N \lhd F$ and $G$ is the special HNN-extension , of $F$ with respect to $N$. Then the following are equivalent:
- $F \in {\mathcal{A}}$ and $F/N \in {\mathcal{A}}$;
- $G \in {\mathcal{A}}$.
The implication (a)$\Rightarrow$(b) is given by Lemma \[lem:spec-HNN-emb\]. To prove the reverse implication, assume that $G \hookrightarrow B$ for some right angled Artin group $B$, corresponding to a finite simplicial graph ${\Gamma}'$. Evidently, since $F \leqslant G$, we have $F \in {\mathcal{A}}$.
Note that $A \coloneq {{\rm Pc}}_{{\Gamma}'}(G)$ is a right angled Artin group, corresponding to some full subgraph ${\Gamma}$ of ${\Gamma}'$. Moreover, the parabolic closure ${{\rm Pc}}_{{\Gamma}}(G)$, of $G$ in $A$, is $A$ itself. This is because every parabolic subgroup of $A$ is a parabolic subgroup of $B$ and $A$ is the minimal parabolic subgroup of $B$ containing $G$, by definition. Britton’s lemma for HNN-extensions (see [@L-S Sec. IV.2]) easily yields that $C_F(t)=N$. Therefore all the assumptions of Proposition \[prop:emb\_quot\] are satisfied and we can deduce that $F/N \hookrightarrow A$. Thus (b) implies (a).
We are now ready to prove Theorem \[thm:rope\] from the Introduction.
Claim (i) is an immediate consequence of Lemma \[lem:spec-HNN-emb\]. Suppose that $G \hookrightarrow A$ for some right angled Artin group $A$. Abusing the notation, let us identify $G$ with its isomorphic image in $A$. Arguing as in the proof of Corollary \[cor:rope-RAAG\], we can replace $A$ with a parabolic subgroup to ensure that $A$ is the parabolic closure of $G$. As before, we know that $C_F(t)=N$ by Britton’s lemma, hence $F/N$ embeds in $A$ by Proposition \[prop:emb\_quot\], and claim (ii) holds.
It is not difficult to see that the group $G_d$ from Corollary \[cor:G\_d\], formulated in the Introduction, is not finitely presented for any $d \in \N$. This is because the normal subgroup $N \lhd F$, with $F/N \cong H_d$, cannot be finitely generated (as it is both infinite and has infinite index in the free group $F$ – see [@L-S Prop. I.3.12]), and the special HNN-extension is finitely presented if and only if $F$ is finitely presented and $N$ is finitely generated. One can overcome this obstacle by using the famous construction of Rips [@Rips], where the free group is replaced by a hyperbolic group.
\[cor:fp\_ex\] For every $d \in \N$ there is a finitely presented group $P_d \in {\mathcal{A}}$ such that $\operatorname{cd}(P_d) \le 3$ but any right angled Artin group $A$, with $P_d \hookrightarrow A$, satisfies $\operatorname{cd}(A)\ge d$.
Take any $d \in \N$. Let us use the modification of Rips’s construction suggested by Haglund and Wise in [@H-W] (in view of the results of Wise [@Wise Thm. 1.2] and Agol [@Agol Thm. 1.1] we can also use the original construction of Rips from [@Rips]). Namely, according to [@H-W Thm. 10.1], there is a group $S$ with a normal subgroup $K \lhd S$ such that $S$ is the fundamental group of some compact non-positively curved square complex, $K$ is finitely generated and $S/K \cong \Z^d$. Moreover, Theorems 5.5, 5.7 and 4.2 from [@H-W] imply that $S$ contains a finite index subgroup $F$ such that $F \in {\mathcal{A}}$.
Let $N \lhd F$ be the intersection of $F$ and $K$. Note that $N$ is finitely generated as $|K:N|\le |S:F|<\infty$, and $F/N$ is still isomorphic to the free abelian group of rank $d$ (because it has finite index in $S/K $). Now, let $P_d$ be the special HNN-extension of $F$ with respect to $N$. By Theorem \[thm:rope\], $P_d \hookrightarrow {\mathcal{A}}$ and for any right angled Artin group $A$, containing $P_d$, one has $\Z^d \hookrightarrow A$, hence $\operatorname{cd}(A) \ge d$.
Let $X \subseteq N$ be some finite generating set of $N$. Then $P_d$ can be defined by the (relative) presentation $$\label{eq:pres}
P_d=\langle F,t \,\|\, txt^{-1}=x,~\forall\,x \in X \rangle.$$ Observe that $S$ is finitely presented, as the fundamental group of a finite square complex, hence $F$ is finitely presented as well. Equation shows that a presentation for $P_d$ can be obtained from a finite presentation of $F$ by adding one generator and finitely many defining relations, thus $P_d$ is also finitely presented.
Finally, $\operatorname{cd}(S) \le 2$ as $S$ is the fundamental group of a $2$-complex with contractible universal cover, therefore $\operatorname{cd}(F)\le 2$. And since $P_d$ is an HNN-extension of $F$, we can conclude that $\operatorname{cd}(P_d) \le \operatorname{cd}(F)+1=3$ – see [@Brown Ex. 4 in Ch. VIII.2].
Corollary \[cor:fp\_ex\] demonstrates that there is no “universal” right angled Artin group for the class of all finitely presented groups $P \in {\mathcal{A}}$ with $\operatorname{cd}(P) \le 3$. However, the main result of Kim and Koberda from [@K-K-obstr] implies that an even stronger fact is true: there does not exist a right angled Artin group which contains copies of all $2$-dimensional right angled Artin groups (see also [@K-K Thm. 1.16] for a weaker statement). Indeed, one just needs to recall that every right angled Artin group embeds into the mapping class group Mod($\Sigma$), of some orientable compact surface $\Sigma$ (cf. [@Koberda Prop. 1.3]), and combine this with [@K-K-obstr Thm. 1.2], asserting that for every such surface $\Sigma$ there exists a right angled Artin group $A$, with $\operatorname{cd}(A)=2$, such that $A \not\hookrightarrow \mbox{Mod($\Sigma$)}$.
[**Acknowledgements.**]{} The author would like to thank Montserrat Casals-Ruiz, François Dahmani and Nansen Petrosyan for discussions, and Alexander Olshanskii for suggesting the idea to use the Golod-Shafarevich theorem in the proof of Proposition \[prop:vsp\]. The author is also grateful to the referee for bringing the paper [@K-K-obstr] to his attention.
[99]{} I. Agol, The virtual Haken conjecture. With an appendix by I. Agol, D. Groves and J. Manning. [*Documenta Math*]{}. **18** (2013), 1045–1087.
Y. Antolín, A. Minasyan, Tits alternatives for graph products. *J. Reine Angew. Math.*, to appear.
Y. Antolín, A. Minasyan, A. Sisto, Commensurating endomorphisms of acylindrically hyperbolic groups and applications. Preprint (2013).` arXiv:1310.8605`
A. Baudisch, Subgroups of semifree groups. *Acta Math. Acad. Sci. Hungar.* **38** (1981), no. 1-4, 19–28.
G. Baumslag, A. Myasnikov, V. Remeslennikov, Algebraic geometry over groups. I. Algebraic sets and ideal theory. *J. Algebra* **219** (1999), no. 1, 16–79.
M.R. Bridson, J. Howie, C.F. Miller III, H. Short, On the finite presentation of subdirect products and the nature of residually free groups. *Amer. J. Math.* **135** (2013), no. 4, 891–933.
M.R. Bridson, C.F. Miller III, Structure and finiteness properties of subdirect products of groups. *Proc. Lond. Math. Soc. (3)* **98** (2009), no. 3, 631–651.
K.S. Brown, Cohomology of groups. Graduate Texts in Mathematics, 87. *Springer-Verlag, New York-Berlin*, 1982. x+306 pp.
M. Casals-Ruiz, I. Kazachkov, Limit groups over partially commutative groups and group actions on real cubings. *Geom. Topol.*, to appear.
R. Charney, An introduction to right-angled Artin groups. *Geom. Dedicata* **125** (2007), 141–158.
M. Ershov, Golod-Shafarevich groups: a survey. *Internat. J. Algebra Comput.* **22** (2012), no. 5, 1230001.
R. Grigorchuk, On the Hilbert-Poincaré series of graded algebras that are associated with groups. (Russian) *Mat. Sb.* **180** (1989), no. 2, 207–225, 304. English translation in *Math. USSR-Sb.* **66** (1990), no. 1, 211–229.
F. Haglund, D.T. Wise, Special cube complexes. *Geom. Funct. Anal.* **17** (2008), no. 5, 1551–1620.
G. Higman, Subgroups of finitely presented groups. *Proc. Roy. Soc. Ser. A* **262** (1961), 455–475.
G. Higman, B.H. Neumann, H. Neumann, Embedding theorems for groups. *J. London Math. Soc.* **24**, (1949). 247–254.
O. Kharlampovich, A. Myasnikov, Irreducible affine varieties over a free group. II. Systems in triangular quasi-quadratic form and description of residually free groups. *J. Algebra* **200** (1998), no. 2, 517–570.
S-h. Kim, T. Koberda, An obstruction to embedding right-angled Artin groups in mapping class groups, *Int. Math. Res. Notices* **14** (2014), no 14, 3912–3918.
S-h. Kim, T. Koberda, Embedability between right-angled Artin groups. *Geom. Topol.* **17** (2013), no. 1, 493–530.
T. Koberda, Right-angled Artin groups and a generalized isomorphism problem for finitely generated subgroups of mapping class groups. *Geom. Funct. Anal.* **22** (2012), no. 6, 1541–1590.
R.C. Lyndon, P.E. Schupp, Combinatorial group theory. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 89. *Springer-Verlag, Berlin-New York*, 1977. xiv+339 pp.
A. Minasyan, Hereditary conjugacy separability of right angled Artin groups and its applications. *Groups Geom. Dyn.* **6** (2012), no. 2, 335–388.
E. Rips, Subgroups of small cancellation groups. *Bull. London Math. Soc.* **14** (1982), no. 1, 45–47.
Z. Sela, Diophantine geometry over groups. I. Makanin-Razborov diagrams. *Publ. Math. Inst. Hautes Études Sci.* No. 93 (2001), 31–105.
H. Servatius, Automorphisms of graph groups. *J. Algebra* **126** (1989), no. 1, 34–60.
D.T. Wise, Cubulating small cancellation groups. *Geom. Funct. Anal.* **14** (2004), no. 1, 150–214.
D.T. Wise, The structure of groups with a quasiconvex hierarchy. Preprint (2011). Available from\
<http://www.math.mcgill.ca/wise/papers.html>
[^1]: We say that a group $G$ is *$n$-generated*, for some $n \in \N$, if $G$ has a generating set of cardinality at most $n$.
[^2]: This can be derived from the fact that $A$ is isomorphic to the fundamental group of a non-positively curved Salvetti cube complex whose dimension is precisely the clique number of ${\Gamma}$ – see [@Charney Sec. 3.6].
[^3]: This is implicit in the proof of [@B-M Thm. 4.7] and is explicitly stated in the Introduction of [@B-H-M-S], where it is proved that being VSP is also sufficient for finite presentability of subdirect products of finitely presented groups.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We extend an argument of S.Lichtenbaum involving codimension one cycles to higher codimensions and obtain a generalization of the well-known Picard-Brauer exact sequence for a smooth variety $X$. The resulting exact sequence connects the codimension $n$ Chow group of $X$ with a certain “Brauer-like" group.'
address: 'Departamento de Matemáticas, Universidad de La Serena, La Serena, Chile'
author:
- 'Cristian D. González-Avilés'
title: 'A generalization of the Picard-Brauer exact sequence'
---
Introduction.
=============
Let $k$ be a field and let $X$ be a geometrically integral algebraic $k$-scheme. We write $\kb$ for a fixed separable algebraic closure of $k$ and set $\g=\text{Gal}\big(\e\kb/k\big)$. The $\kb$-scheme $X\times_{\e\spec k}\spec \kb$ will be denoted by $\xb$. Let $\kb[X]^{*}=H_{\text{\'et}}^{\e 0}(\xb,{\Bbb G}_{m})$ and $\br^{\prime}X=H_{\text{\'et}}^{\e 2}(X,{\Bbb G}_{m})$ be, respectively, the group of invertible regular functions on $\xb$ and the cohomological Brauer group of $X$. The exact sequence mentioned in the title is the familiar exact sequence
$$\begin{array}{rcl}
0&\ra & H^{\e 1}(k,\kb[X]^{*})\ra\pic X \ra
\big(\e\pic\xb\e\big)^{\lbe\g}\ra H^{\e 2}(k,\kb[X]^{*})
\ra\br_{\be 1}^{\e\prime}X\\
&\ra & H^{\e 1}\big(k,\pic\xb\e\big)\ra H^{\e 3}(k,\kb[X]^{*})
\end{array}$$
where $H^{\e i}(k,-)=H^{\e i}(\g,-)$ and $\br^{\e\prime}_{1}X=\krn\big(\br^{\e\prime}X\ra
\br^{\e\prime}\e\xb\,\big)$. This sequence may be obtained from the exact sequence of terms of low degree belonging to the Hochschild-Serre spectral sequence $$H^{\e r}\be\big(k,H_{\text{\'et}}^{\e s}\big(\e\xb,{\Bbb
G}_{m}\big)\big)\implies H_{\text{\'et}}^{\e r+s}(X,{\Bbb G}_{m}).$$ When $X$ is [*smooth*]{} (which we assume from now on), there exists an alternative derivation of (1) which makes use of the following (no less familiar) exact sequence: $$0\ra\kb[X]^{*}\ra\kb(X)^{*}\ra\text{Div}\e\xb\ra \pic\xb\ra 0$$ where $\kb(X)^{*}$ (resp. $\text{Div}\e\xb\e$) is the group of invertible rational functions (resp. Cartier divisors) on $\xb$. This approach, seemingly first used by S.Lichtenbaum in \[4\] and then reconsidered by Yu.Manin \[5, p.403\], consists in splitting (2) into two short exact sequences of $\g$-modules and then taking $\g$-cohomology of these sequences. The resulting long $\g$-cohomology sequences are then appropriately combined to produce (1). This paper is a generalization of this idea. The key observation to make is that (2) may be seen as arising from the Gersten-Quillen complex corresponding to the Zariski sheaf ${\mathcal K}_{\e 1,\e\xb}$, which is the sheaf on $\xb$ associated to the presheaf $U\mapsto K_{1}(U)=H^{\e 0}(U,{\mathcal
O}_{U})^{*}$. In Section 2 we work with the Gersten-Quillen complex corresponding to the Zariski sheaf ${\mathcal K}_{\e n,\e\xb}$ associated to the presheaf $U\mapsto\kn(U)$, where $\kn$ is Quillen’s $n$-th $K$-functor ($1\leq n\leq d=\text{dim}(X)$), and obtain the following result. Let $\dn\colon\bigoplus_{y\in \xb^{\e
n-1}}\kb(y)^{*}\ra Z^{\e n}\lbe\big(\e\xb\e\big)$ be the “sum of divisors" map and let $\Bnx$ be the kernel of the induced map $$H^{\e 2}\!\left(k,\textstyle\bigoplus_{y\in \xb^{\e
n-1}}\kb(y)^{*}\right)\ra H^{\e 2}\big(k,Z^{\e
n}\lbe\big(\e\xb\e\big)\big).$$
Let $X$ be a smooth, geometrically integral, algebraic $k$-scheme. Then there exists a natural exact sequence $$\begin{array}{rcl}
0 &\ra& H^{\e 1}(k,\krn\dn)\ra\chn\ra\chnb^{\g}\!\ra H^{\e
2}(k,\krn\dn)\\
&\ra&\Bnx\ra H^{\e 1}\big(k,\chnb\big)\ra H^{\e 3}(k,\krn\dn)\,.
\end{array}$$
The case $n=1$ of the theorem is precisely the exact sequence (1).
In Section 4, which concludes the paper, we show that the group $\Bnx$ in the exact sequence of the theorem is “Brauer-like", in the sense that it contains a copy of $\br_{1} Y=\krn\be\big[\br
Y\ra\br\yb\,\big]$ for every smooth closed integral subscheme $Y\subset X$ of codimension $n-1$.
Preliminaries
=============
We keep the notations of the Introduction. In particular, $X$ is a smooth, geometrically integral algebraic $k$-scheme of dimension $d$ and $n$ denotes a fixed integer such that $1\leq n\leq d$.
There exists a natural bijection between the set of schematic points of $X$ and the set of closed integral subschemes of $X$. This is defined by associating to a point $x\in X$ the schematic closure $V(x)$ of $x$ in $X$. The codimension (resp. dimension) of $x$ is by definition the codimension (resp. dimension) of $V(x)$. The set of points of $X$ of codimension (resp. dimension) $i$ will be denoted by $X^{i}$ (resp. $X_{i}$), and $\eta$ (resp. $\overline{\eta}$) will denote the generic point of $X$ (resp. $\xb$). If $x\neq\eta$, the function field of $V(x)$ will be denoted by $k(x)$. We use the standard notation $k(X)$ for the function field of $X=V\be(\eta)$. For each $x\in X$, $i_{x}$ will denote the canonical map $\text{Spec}\e\e k(x)\ra X$. The function field of $\xb$ will be denoted by $\kb(X)$. For simplicity, we will write $\vbx$ for $V(x)\times_{\spec k}\spec \kb$.
Since $\xb$ is regular \[3, 6.7.4\], the sheaf ${\s K}_{\e n,\e\xb}$ admits the following flasque resolution, known as the Gersten-Quillen resolution (see \[7, p.72\]):
$$\begin{array}{rcl}
0\ra {\s K}_{\e n,\e\xb} &\ra(\e i_{\e\overline{\eta}}\e)_{*}\e
K_{n}\kb(X) \ra\displaystyle\bigoplus_{y\in\xb^{\e 1}}(\e i_{\e
y})_{*}K_{n-1}\e\kb(y)\ra
\dots\\
&\ra\!\!\displaystyle\bigoplus_{y\in\xb^{\e n-1}}(\e i_{\e
y})_{*}\e\kb(y)^{*} \ra \displaystyle\bigoplus_{y\in\xb^{\e n}}(\e
i_{\e y})_{*}\e\bz\e\ra 0
\end{array}$$ where, for $y\in\xb^{\, i}$, $K_{n-i}\,\kb(y)$ is regarded as a constant sheaf on $\kb(y)$. It follows that the groups $H^{\e
i}\lbe\big(\e\xb,{\s K}_{\e n,\xb}\e\big)=H^{\e i}\lbe\big(\e\xb,{\s
K}_{\e n}\e\big)$ are the cohomology groups of the complex $$\kn\kb(X)\overset{\partial^{\e
0}}\longrightarrow\displaystyle\bigoplus_{y\in\xb^{\e 1}}\be
K_{n-1}\e\kb(y)\overset{\partial^{\e 1}}\longrightarrow\dots
\overset{\partial^{\e n\be-\be
2}}\longrightarrow\displaystyle\bigoplus_{y\in\xb^{\e
n-1}}\ng\be\kb(y)^{*}\overset{\partial^{\e n\be-\be
1}}\longrightarrow\displaystyle\bigoplus_{y\in\xb^{\e n}}\be\Bbb
Z\,.$$ Now, if $q\colon\xb\ra X$ is the canonical morphism and $x\in X$, we write $\xb^{\, n-i}_{\be x}$ for the set of points $y\in \xb^{\e
n-i}$ such that $q(y)=x$. For $i=1,2,\dots, n\be -\be 1$ and $\,
x\in X^{n-i}$, set $$\kbix=\bigoplus_{y\e\in\e\xb^{\e n-i}_{\be x}} K_{i}\e\kb(y).$$ Further, write $Z^{\e n}\be\big(\e\xb\e\big)$ for the group of codimension $n$ cycles on $\xb$, i.e., $$Z^{\e n}\be\big(\e\xb\e\big)=\displaystyle\bigoplus_{y\in\xb^{\e
n}}\Bbb Z\,.$$ Then (3) may be written as $$\kn\kb(X)\overset{\partial^{\e 0}}\longrightarrow\bigoplus_{x\in
X^{1}}\be \overline{K}_{n-1}(x)\overset{\partial^{\e
1}}\longrightarrow\dots \overset{\partial^{\e n\be-\be
2}}\longrightarrow\ng\ng\bigoplus_{x\in X^{n-1}}\ng\be
\kblx\overset{\partial^{\e n\be-\be 1}}\longrightarrow Z^{\e
n}\be\big(\e\xb\e\big)\,.$$ The differential $\dn$ equals $\sum_{\e x\in X^{\lbe n-1}}\dnx\,$, where, for each $x\in X^{n-1}$, $$\dnx\colon\kblx=\displaystyle\bigoplus_{y\in\xb^{\e n-1}_{\be x}}
\kb(y)^{*} \ra Z^{\e n}\be\big(\e\xb\e\big)$$ is the sum of the divisor maps $$\text{div}_{\be y}\colon\kb(y)^{*}\ra Z^{\e n}\be\big(\e\xb\e\big).$$ For definition of the latter, see \[7, p.72\]. We note that each of the maps $\text{div}_{ y}$ factors through $Z^{\e 1}(\e V(y))$, whence each $\dnx$ factors through $Z^{\e 1}(\e\vbx)$.
We will write $CH^{\e n}\lbe(X)$ for the Chow group of codimension $n$ cycles on $X$ modulo rational equivalence. Then $H^{\e n}(X,\s
K_{n})=CH^{\e n}\lbe(X)$ (“Bloch’s formula").
Proof of the main theorem
=========================
The complex (4) induces the following short exact sequences of $\g$-modules: $$0\ra \img\dn\ra Z^{\e
n}\be\big(\e\xb\e\big)\ra\chnb\ra 0$$ and $$0\ra\krn\dn\ra\displaystyle\bigoplus_{x\in
X^{n-1}} \kblx\ra\img\dn\ra 0.$$ Observe that the natural morphism $q\colon\xb\ra X$ induces a homomorphism $\chn\ra\chnb^{\g}$.
There exist canonical isomorphisms $$\begin{array}{rcl}
\krn\!\be\left[\chn\ra\chnb^{\g}\e\right]&=& H^{\e
1}\be\big(k,\krn\dn\e\big)\\\\
\cok\!\be\left[\chn\ra\chnb^{\g}\e\right]&=& H^{\e
1}\be\big(k,\img\dn\e\big)
\end{array}$$ and a canonical exact sequence $$0\ra H^{\e 1}\be\big(k,\chnb\big)\ra H^{\e
2}\be\big(k,\img\dn\e\big)\ra H^{\e 2}\big(k,Z^{\e n}
\be\big(\e\xb\e\big)\big).$$
This follows by taking $\g$-cohomology of (5), using the fact that $Z^{\e n}(\xb)$ is a permutation $\g$-module and arguing as in \[1, proof of Proposition 3.6\] to establish the first isomorphism.
The exact sequence (6) induces an exact sequence $$\begin{array}{rcl}
0&\ra & H^{\e 1}\be\big(k,\img\dn\big)
\ra H^{\e 2}\be\big(k,\krn\dn\big)\ra
\displaystyle\bigoplus_{x\in X^{\lbe n-1}} H^{\e 2}\big(k,\kblx\big)\\
&\ra & H^{\e 2}\be\big(k,\img\dn\big)\ra H^{\e
3}\be\big(k,\krn\dn\big)\,.
\end{array}$$
By Shapiro’s Lemma, for each $x\in X^{n-1}$ there exists a (non-canonical) isomorphism $$H^{*}(k,\kblx)\simeq
H^{*}\be\big(\e\text{Gal}\big(\e\kb(y)/k(x)\big),\kb(y)^{*}\big)$$ where, on the right, we have chosen a point $y\in\xb^{\, n-1}$ such that $q(y)=x$. The result now follows by taking $\g$-cohomology of (6), using Hilbert’s Theorem 90.
Combining Lemmas 3.1 and 3.2, we obtain
There exists a canonical exact sequence $$\begin{array}{rcl}
0&\ra & H^{\e 1}\be\big(k,\krn\dn\e\big)\ra\chn\ra\chnb^{\g}\\
&\ra & H^{\e
2}\be\big(k,\krn\dn\e\big)\ra\displaystyle\bigoplus_{x\in X^{n-1}}
H^{\e 2}(k,\kblx).\qed
\end{array}$$
Now define $$\Bnx=\krn\!\be\left[\e H^{\e 2}\!\left(k,\textstyle\bigoplus_{x\in
X^{n-1}}\kblx\right)\ra H^{\e 2}\big(k,Z^{\e
n}\be\big(\e\xb\e\big)\big)\right],$$ where the map involved is induced by $\dn$. Since the composite $$\krn\dn\ra\bigoplus_{x\in X^{n\be -\be 1}}
\kblx\overset{\dn}\longrightarrow Z^{\e n}\be\big(\e\xb\e\big)$$ is zero, the natural map $H^{\e 2}(k,\krn\dn)\ra\be\bigoplus_{x\in
X^{n-1}}\be H^{\e 2}(k,\kblx)$ factors through $\Bnx$. Thus Proposition 3.3 yields a natural exact sequence $$\begin{array}{rcl}
0&\ra & H^{\e 1}\be\big(k,\krn\dn\e\big)\ra\chn\ra\chnb^{\g}\\
&\ra & H^{\e 2}\be\big(k,\krn\dn\e\big)\ra\Bnx.
\end{array}$$ We will now extend the above exact sequence by defining a map $\Bnx\ra H^{\e 1}(k,\chnb)$ whose kernel is exactly the image of the map $H^{\e 2}(k,\krn\dn)\ra\Bnx$ appearing in (8).
It is not difficult to check that the map $$\displaystyle\bigoplus_{x\in X^{n-1}} H^{\e 2}\big(k,\kblx\big)\ra
H^{\e 2}\be\big(k,\img\dn\big)$$ intervening in the exact sequence of Lemma 3.2 maps $\Bnx$ into the kernel of the map $H^{\e 2}(k,\img\dn)\ra H^{\e 2}\be\big(k,Z^{\e n}
\be\big(\e\xb\e\big)\big)$. The latter is naturally isomorphic to $H^{\e 1}\be\big(k,\chnb\big)$ (see Lemma 3.1). Thus there exists a canonical map $\Bnx\ra H^{\e 1}\be\big(k,\chnb\big)$. Again, it is not difficult to check that the kernel of the map just defined is exactly the image of the map $H^{\e 2}(k,\krn\dn)\ra\Bnx$ appearing in (8). Thus we obtain a natural exact sequence $$\begin{array}{rcl}
0 &\ra& H^{\e 1}(k,\krn\dn)\ra\chn\ra\chnb^{\g}\ra H^{\e 2}
(k,\krn\dn)\\
&\ra &\Bnx\ra H^{\e 1}(k,\chnb)\,.
\end{array}$$ Finally, the homomorphisms $H^{\e 1}\be(k,\chnb)\be\ra\be H^{\e
2}\be(k,\img\dn)$ and $H^{\e 2}\be(k,\img\dn)\be\ra\be H^{\e
3}\be(k,\krn\dn)$ from Lemmas 3.1 and 3.2 induce a map $H^{\e
1}(k,\chnb)\ra H^{\e 3}(k,\krn\dn)$ whose kernel is exactly the image of the map $\Bnx\ra H^{\e 1}(k,\chnb)$ defined above. Thus the following holds.
Let $X$ be a smooth $k$-variety. Then there exists a natural exact sequence $$\begin{array}{rcl}
0 &\ra & H^{\e 1}(k,\krn\dn)\ra\chn\ra\chnb^{\g}\ra
H^{\e 2}(k,\krn\dn)\\
&\ra &\Bnx\ra H^{\e 1}\be\big(k,\chnb\big)\ra H^{\e 3}(k,\krn\dn),
\end{array}$$ where $\Bnx$ is the group (7).
When $n=1$, there are natural isomorphisms $\chn=\pic X$ and $\chnb=\pic\xb\,$ \[3, 21.6.10 and 21.11.1\]. Further, $X^{ n-1}=\{\eta\}\e$, $\,\dn=\partial^{\e
n-1}_{\eta}\colon\overline{K}_{\be 1}(\eta)=\kb(X)^{*}\ra\iv\xb$ is the usual divisor map (whose kernel equals $H^{0}(\xb,\Bbb
G_{m})\overset{\text{def.}}{=}\kb[X]^{*}$) and $$\Bnx=B_{\e 1}(X)=\krn\!\be\left[\e H^{\e 2}(k,\kb(X)^{*})\ra H^{\e
2}\be\big(k,\iv\xb\,\big)\right]= \br_{\be 1}X,$$ where $\br_{\be 1}X=\krn\!\left(\e\br X\ra\br\xb\,\right)$ (see the next section). Thus the exact sequence of the theorem is indeed a generalization of (1).
The group $\Bnx$
================
In this Section we show that the group $\Bnx$ appearing in the exact sequence of Theorem 3.4 contains a copy of $\br_{1}
Y=\krn\be\big(\e\br Y\ra\br\yb\,\big)$ for [*every*]{} smooth closed integral subscheme $Y\subset X$ of codimension $n-1$.
Recall that $\dn=\sum_{\e x\in X^{\lbe n-1}}\dnx\,$, where, for each $x\in X^{n-1}$, $$\dnx\colon\kblx=\displaystyle\bigoplus_{y\in\xb^{\e n-1}_{\be x}}
\kb(y)^{*} \ra Z^{\e 1}\be\big(\vbx\big)$$ is the sum of divisors map. For each $x\in X^{n-1}$, set $$\bnx=\krn\!\left[\,H^{\e 2}(k,\kblx)\ra H^{\e
2}(k,Z^{\e1}(\vbx\e))\e\right],$$ where the map involved is induced by $\dnx$, and let $$\Sigma\,\,\colon\bigoplus_{x\in X^{n-1}} H^{\e 2}(k,Z^{\e
1}(\e\vbx))\ra H^{\e 2}(k,Z^{\e n}(\xb))$$ be the natural map $(\xi_{x})\mapsto\sum c_{x}(\xi_{x})$, where $c_{x}\colon H^{\e 2}(k,Z^{\e 1}(\e\vbx))\ra H^{\e 2}(k,Z^{\e
n}(\xb))$ is induced by the inclusion $Z^{\e 1}(\e\vbx)\subset Z^{\e
n}(\xb)$. Then there exists a canonical exact sequence $$0 \ra \displaystyle\bigoplus_{x\in X^{n-1}}\bnx\ra\Bnx\ra
\krn\Sigma\,\,.$$ We will relate the groups $\bnx$ to more familiar objects.
Fix $x\in X^{ n\be -\be 1}$ and set $Y=V(x)$. Then $Y$ is a geometrically reduced algebraic $k$-scheme \[3, 4.6.4\]. Further, the map $\kblx\ra Z^{\e 1}\be\big(\e\yb\e\big)$ factors through $\iv\yb$, the group of Cartier divisors on $\yb$. Consider $$\bnpx=\krn\ng\left[H^{\e2}(k,\kblx)\ra
H^{\e2}\be\big(k,\iv\yb\e\big)\e\right]\subset\bnx.$$ Let ${\s R}_{\,\yb}^{*}$ denote the étale sheaf of invertible rational functions on $\yb$. Note that $\kblx\be=\be H^{\e
0}(\yb,{\s R}_{\,\yb}^{*})$. Now, since $\yb$ is reduced, there exists an exact sequence of étale sheaves $$0\ra\Bbb G_{m,\yb}\ra{\s R}_{\,\yb}^{*}\ra{\s Div}_{\e\yb}\ra 0,$$ where ${\s Div}_{\yb}$ is the sheaf of Cartier divisors on $\yb$ \[3, 20.1.4 and 20.2.13\]. This exact sequence gives rise to an exact sequence of étale cohomology groups $$0\ra H_{\text{\'et}}^{\e 1}\be\big(\e\yb,{\s
Div}_{\e\yb}\e\big)\ra\br'\,\yb\ra H_{\text{\'et}}^{\e
2}\be\big(\e\yb,{\s R}_{\,\yb}^{*}\big) \ra H_{\text{\'et}}^{\e
2}\be\big(\e\yb,{\s Div}_{\e\yb}\big)$$ where $\br'\,\yb=H_{\text{\'et}}^{\e 2}\be\big(\e\yb, \Bbb
G_{m}\big)$ is the cohomological Brauer group of $\yb$ \[2, II, p.73\]. Similarly, there exists an exact sequence $$0\ra H_{\text{\'et}}^{\e 1}(Y,{\s Div}_{Y})\ra\br'\,Y\ra
H_{\text{\'et}}^{\e 2}(Y,{\s R}_{\,Y}^{*}) \ra H_{\text{\'et}}^{\e
2}(Y,{\s Div}_{Y}).$$ We will regard $H_{\text{\'et}}^{\e 1}\be\big(\e\yb,{\s
Div}_{\e\yb}\e\big)$ (resp. $H_{\text{\'et}}^{\e 1}(Y,{\s
Div}_{Y})$) as a subgroup of $\br'\,\yb$ (resp. $\br'\,Y$).
Now the exact sequence of terms of low degree $$0\ra E_{2}^{\e 1,0}\ra E^{\e 1}\ra E_{2}^{\e 0,1}\ra E_{2}^{\e 2,0}
\ra\krn(E^{\e 2}\ra E_{2}^{\e 0,2})\ra E_{2}^{\e 1,1}\ra E_{2}^{\e
3,0}$$ belonging to the Hochschild-Serre spectral sequence $$E_{2}^{\e p,q}=H^{\e p}\be\big(k,H^{\e
q}_{\text{\'et}}\be\big(\e\yb,{\s R}_{\,\yb}^{*}\big)\e\big)\implies
H_{\text{\'et}}^{\e p+q} (Y,{\s R}_{\, Y}^{*})$$ yields, using \[2, II, Lemma 1.6, p.72\], an exact sequence $$0\ra H^{\e 2}(k,\kblx)\ra H_{\text{\'et}}^{\e 2}(Y,{\s R}_{\,Y}^{*})
\ra H_{\text{\'et}}^{\e 2}\be\big(\e\yb,{\s R}_{\,\yb}^{*}\big).$$ Similarly, the spectral sequence $$H^{\e p}\be\big(k,H_{\text{\'et}}^{\e q}\be\big(\e\yb,{\s
Div}_{\e\yb}\big)\big) \implies H_{\text{\'et}}^{\e p+q}(Y,{\s
Div}_{Y})$$ yields a complex $$\begin{array}{rcl}
0 &\ra& H^{\e 1}(k,\iv\yb\e)\ra H_{\text{\'et}}^{\e 1}(Y,{\s
Div}_{Y}) \overset{\psi}\longrightarrow H_{\text{\'et}}^{\e
1}\be\big(\e\yb,{\s Div}_{\e\yb}\big)^{\g}\\
&\overset{\varphi}\longrightarrow & H^{\e
2}\be\big(k,\iv\yb\e\big)\ra H_{\text{\'et}}^{\e 2}(Y, {\s
Div}_{Y})\ra H_{\text{\'et}}^{\e 2}\be\big(\e\yb,{\s
Div}_{\e\yb}\big)
\end{array}$$ which is exact except perhaps at $H_{\text{\'et}}^{\e 2}(Y, {\s
Div}_{Y})$. The map labeled $\psi$ in (13) is induced by the canonical morphism $\yb\ra Y$ , while the map $\varphi$ is the differential $d_{2}^{\, 0,1}$ coming from the spectral sequence (see \[6, II.4, pp.39-52\]). Now we have a commutative diagram
$$\xymatrix{0\ar[r] & H^{\e 2}(k,\kblx)\ar[d]\ar[r]&
H_{\text{\'et}}^{\e 2} (Y,{\s R}_{\,Y}^{*})\ar[d]\ar[r] &
H_{\text{\'et}}^{\e 2}\be\big(\e\yb,{\s
R}_{\,\yb}^{*}\big)\ar[d]\\
0\ar[r] & H^{\e 2}\be\big(\e k,\iv\yb\e\big)/\e\img\varphi \ar[r]&
H_{\text{\'et}}^{\e 2}(Y, {\s Div}_{\e Y})\ar[r] &
H_{\text{\'et}}^{\e 2}\be\big(\e\yb,{\s Div}_{\e\yb}\big).\\
}$$
in which the top row is the exact sequence (12), the bottom row (which is only a complex) is derived from (13), and the middle and right-hand vertical maps are the maps in (11) and (10), respectively. Set $$\widehat{\text{Br}_{1}^{\,\prime}}\e Y=\krn\be\Bigg[\e\br'\e
Y/H_{\text{\'et}}^{\e 1}(Y,{\s Div}_{Y})\ra
\br'\,\yb/H_{\text{\'et}}^{\e 1}\be\big(\e\yb,{\s
Div}_{\e\yb}\big)\e\Bigg].$$ Then the above diagram yields a natural isomorphism $$\widehat{\text{Br}_{1}^{\,\prime}}\e Y= \krn\!\be\left[\e H^{\e
2}(k,\kblx)\ra H^{\e 2}\be\big(\e k,\iv\yb\,\big)/\e\img
\varphi\,\right].$$ (Note: only the exactness of the top row of (14) is needed to obtain the above isomorphism.) On the other hand, there exists an obvious exact sequence $$0\ra\bnpx\ra\krn\!\be\left[\e H^{\e 2}(k,\kblx)\ra H^{\e
2}\be\big(\e k,\iv\yb\,\big)/\e\img\varphi\,\right]\ra\img\varphi,$$ where $\bnpx$ is the group (9). Using (15) and the fact that $\img\varphi$ is naturally isomorphic to $\cok\psi$, where $\psi$ is the map appearing in (13), we conclude that there exists a natural exact sequence $$0\ra\bnpx\ra\widehat{\text{Br}_{1}^{\,\prime}}\e Y
\overset{h}\longrightarrow\cok\psi\,.$$ The map labeled $h$ in the above exact sequence can be briefly described as $``\e\varphi^{-1}\circ h^{\e 2}(\text{div})\e\circ\e
u^{-1}"$, where $u\colon H^{\e 2}(k,\kblx)\ra H^{\e
2}_{\text{\'et}}(Y,{\s R}_{\,Y}^{*})$ is the map intervening in (14) and $h^{\e 2}(\text{div})\colon H^{\e 2}(k,\kblx)\ra H^{\e
2}\be\big(\e k,\iv\yb\,\big)$ is induced by $\text{div}\colon\kblx\ra \iv\yb$. Next, set $$\br_{1}^{\e\prime}\e
Y=\krn\!\left[\e\br^{\e\prime}Y\ra\br^{\e\prime}\e \yb\,\right].$$ There exists a natural exact commutative diagram $$\xymatrix{0\ar[r] & H^{\e 1}_{\text{\'et}}(Y,{\s
Div}_{Y})\ar[d]^{\psi}\ar[r]& \br_{\be 1}^{\e\prime}\e
Y\ar[d]\ar@{->>}[r] & \br_{\be 1}^{\e\prime}\e Y/\e
H^{\e 1}_{\text{\'et}}(Y,{\s Div}_{Y})\ar[d]\\
0\ar[r] & H^{\e 1}_{\text{\'et}}\be\big(\e\yb,{\s
Div}_{\e\yb}\big)^{\g} \ar[r]& \left(\e\br^{\prime}\e\yb\,\right
)^{\g}\ar[r] & \left(\e\br^{\prime}\e\yb/\e H^{\e
1}_{\text{\'et}}\be\big(\e\yb,{\s Div}_{\e\yb}\big)\e\right)^{\g}.}$$ An application of the snake lemma to the above diagram yields a natural exact sequence $$0\ra H^{\e 1}(k,\iv\yb\e)\ra \br_{\be 1}'\e Y\ra
\widehat{\text{Br}_{1}^{\,\prime}}\e Y
\overset{\delta}\longrightarrow\cok\psi\,.$$ Now using the explicit description of the map $\delta$ \[8, Lemma 1.3.2, p.11\] together with the description of the map $\varphi=d_{2}^{\, 0,1}$ from \[6, §II.4\], it can be shown (with some work) that the maps $h$ in (16) and $\delta$ in (17) [*are the same*]{}. Thus we obtain
There exists a canonical isomorphism $$\bnpx=\br_{1}^{\e\prime}\e Y/\e H^{\e 1}\be\big(k,\iv\yb\e\big).$$
Let $x\in X^{ n-1}$ be such that $\yb=\vbx$ is [[locally factorial]{}]{} (this holds, for example, if $Y=V(x)$ is regular). Then there exists a canonical isomorphism $$\bnx=\br_{\be 1}^{\e\prime}\e Y.$$
The hypothesis implies that $\iv\yb=Z^{\e 1}\be\big
(\e\yb\e\big)$ \[3, 21.6.9\], so $\bnx=\bnpx$. On the other hand, since $Z^{\e 1}\be\big (\e\yb\e\big)$ is a permutation $\g$-module, $H^{\e 1}\be\big(k,\iv\yb\e\big)=H^{\e 1}\be\big(k,Z^{\e 1}\be\big
(\e\yb\e\big)\big)=0$. The result is now immediate from the proposition.
Colliot-Thélène, J.-L. and Raskind, W.: *${\s K}_{2}$-Cohomology and the second Chow group* Math. Ann. [**[270]{}**]{}, pp.165-199 (1985).
Grothendieck, A.: *Le Groupe de Brauer I-III.* In: Dix Exposés sur la Cohomologie des Schémas. North-Holland, Amsterdam, pp.46-188 (1968).
Grothendieck, A. and Dieudonné, J.: *Eléments de Géométrie Algébrique IV.* Publ. Math. IHES [**[20,24,28,32]{}**]{}, 1960-1967.
Lichtenbaum, S.: *Duality theorems for curves over $P$-adic fields.* Invent. Math. [**[7]{}**]{}, pp. 120-126 (1969).
Manin, Yu.: *Le groupe de Brauer-Grothendieck en géométrie diophantienne.* In: Actes du Congrès Intern. Math. Nice I, pp. 401-411 (1970).
Shatz, S.: *Profinite groups, Arithmetic, and Geometry.* Ann. of Math. Studies [**[67]{}**]{}, Princeton Univ. Press (1972).
Srinivas, V.: *Algebraic $K$-theory (2nd. Edition)* Progress in Math. [**[90]{}**]{}, Birhäuser, Boston, 1996.
Weibel, C.: *An introduction to homological algebra.* Cambridge Studies in Advanced Math. [**[38]{}**]{}, Cambridge University Press, Cambridge, 1994.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'M. Cocchi , A. Bazzano , L.Natalucci , P. Ubertini , J. Heise , E. Kuulkers , R. Cornelisse and J.J.M. in ’t Zand'
date: 'Received ?; accepted ?'
title: 'Discovery of the X-ray burster SAX J1752.3–3138'
---
Introduction: the [*BeppoSAX*]{}-WFC Galactic bulge monitoring program
======================================================================
One of the main scientific goals of the Wide Field Cameras (WFC, [@Jage97]) on board the [*BeppoSAX*]{} satellite ([@Boel97]) is the study of the timing and spectral behaviour of both transient and persistent sources in the Galactic bulge, X-ray binaries in particular, on time scales ranging from seconds to years. To this end, a monitoring program of systematic wide field observation of the Sgr A sky region is carried out since August 1996 ([@Heis98; @Heis99; @Uber99]). The program consists of a series of observations, each lasting 60 ks net time, almost weekly spaced throughout the two visibility periods (August-October and February-April) of the Galactic Centre region. Up to January 2001, an amount of 4.0 Ms of data has been accumulated this way. The WFC program on the Galactic bulge is significantly contributing in the study of transient X-ray binary systems, mainly X-ray bursters (see e.g. [@Zand00] for a recent overview). SAX J1752.3$-$3138 was discovered by the WFCs during the Summer–Fall 1999 monitoring campaign of the Galactic bulge. The source was observed during an X-ray burst (see Fig.1 and Fig. 2, top panel) and no persistent emission was found ([@Cocc99a]). No other X-ray bursts are detected in the whole WFC data archive. The net exposure time available for burst detection is 6.3 Ms up to January 2001, being the attitude constraints not so stringent as for steady emission detection. This makes the observed event the only source of information about SAX J1752.3$-$3138 available so far. Nevertheless, burst detection itself provides important details on a newly discovered source, being such events uniquely associated to weakly magnetised neutron stars (NS) in Low-Mass X-ray Binary systems (LMXBs). Besides the nature of the source, in several cases X-ray bursts allow to estimate useful parameters such as the source distance and luminosity, and the NS radius (see Lewin, van Paradijs & Taam 1993, hereafter [@Lewi93], for a comprehensive review). A single detected burst in a fairly big amount of observing time is even more interesting, since its uniqueness puts important constraints on the characteristics of the accretion onto the neutron star. In the case of SAX J1752.3$-$3138 in particular, since no steady emission was observed, we explore the bursting behaviour of sources at very low persistent luminosities. The study of the lower luminosity boundary when NS start bursting is an important ingredient for our knowledge of thermonuclear bursts.
In the next section we report on the observation and the spectral/timing data analysis of the (so far) only burst of SAX J1752.3$-$3138. We then briefly discuss the results in Section 3, proposing SAX J1752.3$-$3138 either as a weak transient NS-LMXB or, more intriguingly, as a member of a possibly new class of low-luminosity, slowly accreting X-ray bursters. An estimate of the distance to the source will be also given.
\[t:ae\]
[lc]{} Burst date & 1999, September 2\
Burst UT time (h) & 3.3936\
[*e*]{}-folding time (2–19 keV) & $21.9\pm 1.3$ s\
[*e*]{}-folding time (2–6 keV) & $21.6\pm 2.8$ s\
[*e*]{}-folding time (6–19 keV) & $16.0\pm 2.2$ s\
peak intensity (2–19 keV) & $710\pm 32$ mCrab\
blackbody [*k*]{}T & $1.64^{+0.10}_{-0.09}$ keV\
$R_{\rm km}/d_{10~{\rm kpc}}$ & $9.2^{+1.5}_{-0.9}$\
Reduced $\chi^{2~(a)}$ & 1.00\
steady emission $^{(b)}$ & $< 6$ mCrab $^{(c)}$\
SAX J1752.3$-$3138: Observation and Data Analysis
=================================================
The [*BeppoSAX*]{}-WFCs are two identical coded aperture multi-wire proportional counters, each covering a $40^{\circ} \times 40^{\circ}$ field of view, the largest ever flown for an arcminute ($1^{\prime}-3^{\prime}$, 99% confidence) location accuracy X-ray telescope. The cameras point at opposite directions and operate in the 2–28 keV bandpass. The time resolution of the imaging operative mode is 0.488 ms, and the energy resolution is 18% at 6 keV. The imaging capability and the good instrument sensitivity (5-10 mCrab in $10^{4}$ s) allow an accurate monitoring of complex sky regions, like the Galactic bulge.
Thanks to their scientific capabilities, the [*BeppoSAX*]{}-WFCs are well designed to systematically investigate transient phenomena lasting from seconds to minutes, like the X-ray bursts. The data of the two cameras is searched for bursts and flares by analysing the time profiles of the detectors with a time resolution down to 1 s. Reconstructed sky images are generated for any statistically significant event, in order to identify possible new bursters. The accuracy of the reconstructed position, which of course depends on the burst intensity, is typically better than $5\arcmin$. This analysis procedure demonstrated its effectiveness throughout the Galactic bulge WFC monitoring program, leading to the identification of more than $2.2\times 10^{3}$ X-ray bursts (or $\sim 1.6\times 10^{3}$ when excluding the Rapid Burster and the Bursting Pulsar) from 38 different sources (e.g. [@Zand00]).
The X-ray burst from the new source SAX J1752.3$-$3138 was found during a post-facto analysis of a 50 ks observation of the Galactic centre region performed by the WFC unit 2 on September 2, 1999 ([@Cocc99a]). The bursting source is located $\sim 3\degr$ SW of Sgr A, at $\alpha=17^{\rm h}52^{\rm m}24^{\rm s}$, $\delta=-31^{\circ}37.7^{\prime}$ (J2000, error radius $2.9^{\prime}$, 99% confidence). The burst peak intensity, corrected for dead time effects, was $\sim 0.7$ Crab. No steady emission was detected during the whole observation, with a $3\sigma$ upper limit of 6 mCrab in the 2–28 keV band. No other X-ray bursts were reported from the same sky position during the observation, even if other events could be missed due to the $\sim 53\%$ source covering efficiency associated to the low Earth orbit of [*BeppoSAX*]{}. Search for bursts and/or persistent emission from SAX J1752.3$-$3138 was performed on all the data available from the 1996-1999 [*BeppoSAX*]{}-WFC Galactic Bulge monitoring campaigns but, as mentioned above, we report no other significant detections. The typical WFC sensitivity (on-axis) to burst events is approximately 180 mCrab (2–28 keV), corresponding to a peak of $\sim 7\times 10^{37} {\rm erg~s}^{-1}$ ($\sim 30\%~{\rm L}_{\rm Edd}$) for a 10 kpc distance. This value is given for a $6 \sigma$ detection of a burst with exponential decay and characteristic time of 10 s.
In Fig. 1 the time profiles of the burst, obtained in two energy bands, are shown. The time histories of the bursts are constructed by accumulating only the detector counts associated with the shadowgram of the analysed source, thus improving the signal-to-noise ratio of the profile. The background, which is the sum of (part of) the diffuse X-ray background, the particles background and the contamination of other sources in the field of view, has been subtracted. Source contamination is the dominating background component for crowded sky fields like the Galactic bulge. Nevertheless, the probability of source confusion during a short time-scale event (10–100 s) like an X-ray burst is negligible.
The burst spectrum of SAX J1752.3$-$3138 is consistent with absorbed blackbody radiation with average colour temperature of $\sim 1.6$ keV (see Table 1). Due to the rather poor statistics, the $N_{\rm H}$ parameter could not be satisfactorily constrained, so it was kept fixed according to the interpolated value for the source sky direction, $5.64\times 10^{21} {\rm cm}^{-2}$ ([@Dick90]). An absorbed power-law model does not fit the data adequately.
Time-resolved spectra were accumulated for the observed burst, in order to study the time evolution of the spectral parameters (Table 2). The spectra are all consistent with blackbody radiation, and spectral softening is observed during the burst decay, in agreement with the longer decay time associated to the low-energy time profile (see Fig. 1). Blackbody spectra allow to determine the relationship between the average radius of the emitting sphere $R_{\rm km}$ (in units of km) and the source distance $d_{\rm 10~kpc}$ (in units of 10 kpc). In Fig. 2 the time histories of the measured $R_{\rm km}$/$d_{\rm 10~kpc}$ ratios are shown, assuming isotropic emission and not correcting for gravitational redshift and conversion to true blackbody temperature from colour temperature (see [@Lewi93] for details).
\[t:hc\]
[lccc]{} time range$^{(a)}$ & [*k*]{}T (keV) & $R_{\rm km}/d_{10~{\rm kpc}}$ & $\chi^{2~(b)}_{\rm r}$\
$0 - 6$ s & $1.16\pm 0.10$ & $22.1^{+5.2}_{-3.5}$ & 0.73\
$6 - 12$ s & $2.39^{+0.23}_{-0.21}$ & $5.2^{+1.1}_{-0.8}$ & 0.65\
$12 - 21$ s & $2.26^{+0.31}_{-0.29}$ & $4.5^{+1.6}_{-0.9}$ & 1.10\
$21 - 34$ s & $1.40^{+0.15}_{-0.14}$ & $11.4^{+3.1}_{-1.9}$ & 0.71\
Discussion
==========
The neutron star LMXB SAX J1752.3$-$3138
----------------------------------------
As anticipated, the spectral and timing properties of the burst detected from SAX J1752.3$-$3138 are clearly suggestive of a type I event ([@Hoff78]), typically associated to weakly magnetised neutron stars in low-mass binary systems (e.g. [@Lewi93]). In fact the blackbody emission and the measured colour temperatures of $\sim 1.6$ keV are consistent with this hypothesis. Spectral softening is observed in the time resolved spectra of the bursts (Table 2), and the bursts time profiles can be fitted with exponential decays whose characteristic times are energy dependent, being shorter at higher energies (Fig. 1).
The time history of the blackbody radius as obtained from the time resolved spectral analysis of the burst (Fig. 2, lower panel) shows evidence for radius expansion during the very first seconds of the event. This is generally interpreted as adiabatic expansion of the neutron star photosphere during a high luminosity (Eddington-limited) type I burst. The WFC lightcurve of this event is not sensitive enough to show evident flat-top or double-peaked profile (e.g. [@Lewi93]), due the prompt photospheric expansion (see later). An expansion by a factor $\ga 3$ with respect to the average decay radius is observed, while lower luminosity bursts show almost constant blackbody radius during the whole event. The observed expansion is consistent with the time profiles in Fig. 1, as the soft emission precedes by $\sim 4$ s the harder one. Relatively high luminosity and no hard emission ([*k*]{}T is $\sim 1$ keV) in the very first seconds of the burst imply high values of the blackbody radius. If we narrow the time bin, concentrating on the first 4 s of the burst, we obtain an even larger radius, $33^{+10}_{-6}$ km for a 10 kpc distance. This points to a prompt expansion of the neutron star photospere after the ignition. The fast rise and the prompt photospheric expansion are suggestive of helium burning, while the duration of the burst ([*e*]{}-folding time of $\sim 20$ s), rather long for a pure helium flash, indicates that the burning environment was not hydrogen-free ([@Lewi93; @Bild00]).
Eddington-luminosity X-ray bursts can lead to an estimate of the source distance, if we assume isotropic emission and an Eddington bolometric luminosity of $2.5\times 10^{38} {\rm erg~s}^{-1}$, which is appropriate for a $1.4~{\rm M}_{\odot}$ neutron star and helium-rich fuel. This value also includes a moderate (20%) gravitational redshift correction ([@VanP94]). Taking into account the observed peak flux of the burst (Table 1), which extrapolates to an unabsorbed bolometric intensity of $(2.48\pm 0.21)\times 10^{-8}{\rm erg~cm}^{-2}{\rm s}^{-1}$, we obtain $d=9.2\pm 0.4$ kpc. The reported error is $1\sigma$, but the systematic uncertainties due the assumption of standard burst parameters are of course larger.\
Assuming a Crab-like spectrum, we derive for the source persistent (bolometric) luminosity an upper limit of $\sim4.6\times 10^{36}{\rm erg~s}^{-1}$ when the burst was observed (Sep. 2, 1999). An average upper limit twice as low applies to the whole 1996–2000 data set. Taking into account the inferred distance of 9.2 kpc, the average blackbody radius during the decay of the burst was $\sim 6.7$ km. This value is within the range commonly observed (e.g. [@VanP78]; also [@Cocc00] for recent results). For a canonical $1.4~{\rm M}_{\odot}$ neutron star with $R = 10$ km, we derive an upper limit of $4.0\times 10^{-10}{\rm M}_{\odot}{\rm yr}^{-1}$ ($8.3\times 10^{3}{\rm g~cm}^{-2}{\rm s}^{-1}$) to the accretion rate.
A new class of low-luminosity bursters?
---------------------------------------
SAX J1752.3$-$3138 is not the only example of a WFC-discovered X-ray burster with no persistent emission observed. In ’t Zand et al., 1998, already reported on the discovery of two burst sources (SAX J1753.5$-$2349, SAX J1806.5$-$2215) with no measured steady emission. Also GRS 1741.9$-$2853 and 1RXS J171824.2$-$402934 did not show persistent emission at the time when bursts were discovered ([@Cocc99b] and [@Kapt00]), even though GRS 1741.9$-$2853 was once detected with a luminosity of $\sim 2\times 10^{36} {\rm erg~s}^{-1}$ in 1990 ([@Suny90]). The upper limits on their luminosities not being very constraining (of the order of $10^{36}{\rm erg~s}^{-1}$), it is unclear if these sources are to be regarded as weak persistent emitters instead of transients. But, as X-ray bursts are rarely observed (in 4.5 y WFC monitoring, 3 from SAX J1806.5$-$2215 and GRS 1741.9$-$2853, only 1 from SAX J1752.3$-$3138, SAX J1753.5$-$2349 and 1RXS J171824.2$-$402934), the possibility that these sources are persistent emitters with luminosities around or just below $10^{36}{\rm erg~s}^{-1}$ seems unlikely. In fact more regular bursting behaviour is to be expected in this case, as typically observed in the medium-luminosity (atoll) NS LMXB ([@Lewi93]).\
So these sources could be usually quiescent systems where occasional mass transfer to the neutron star occurs, thus causing type I bursts. If so, they should not be very different from other weak transients observed by the WFCs, where burst activity was observed several days after the outburst (e.g. XTE J1709$-$267, SAX J1712.6$-$3739, SAX J1750.8$-$2900, SAX J1810.6$-$2609, SAX J1808.4$-$3658, see e.g. [@Zand00]). The rather large distance (in excess of 8 kpc for SAX J1752.3$-$3138 and GRS 1741.9$-$2853) and the incomplete time coverage by sensitive enough all-sky monitors could explain the missed detection of the (reasonably weak, a few $10^{36}{\rm erg~s}^{-1}$) outbursts triggering the occasional type I burst activity of these sources.\
There is another intriguing possibility left: this small sample of burst-only objects could be part of a class of intrinsically weak persistent bursters (${\rm L} \la 10^{35}{\rm erg~s}^{-1}$), whose observation would be very useful to investigate the type I bursting behaviour at low accretion rates. The sporadic burst activity of such sources would be then related to their unusual accretion regime instead of their transient behaviour. Actually this could be the case for 1RXS J171824.2$-$402934, whose [*ROSAT*]{} observations suggest weak persitent (even though variable) activity, with bolometric luminosity, taking into account the inferred distance of $\sim 7$ kpc ([@Kapt00]), not exceeding $\sim 2\times 10^{35}{\rm erg~s}^{-1}$. Moreover, an X-ray burst at very low persistent luminosity (${\rm L} \la 10^{33}{\rm erg~s}^{-1}$) was observed by Gotthelf & Kulkarni (1997) in the globular cluster M28. But that burst was notably subluminous ($L_{\rm peak} \sim 4\times 10^{36}{\rm erg~s}^{-1}$), perhaps due to magnetic confinement of the burning fuel ([@Gott97]), while the bursts in the WFC sample reached peak luminosities of $\sim 10^{38}{\rm erg~s}^{-1}$, commonly observed in NS-LMXBs. Further measurements with more sensitive instruments could unveil the nature (triansient or weakly persistent) of these burst-only sources.
We thank the staff of the [*BeppoSAX Science Operation Centre*]{} and [*Science Data Centre*]{} for their help in carrying out and processing the WFC Galactic Centre observations. The [*BeppoSAX*]{} satellite is a joint Italian and Dutch program. M.C., A.B., L.N. and P.U. thank Agenzia Spaziale Nazionale ([*ASI*]{}) for grant support. M.C. also thanks M. Federici (IAS staff) for technical support.
Boella, G., Butler, R.C., Perola, G.C., Piro, L., Scarsi, L., and Bleeker, J.A.M., 1997, A&AS 122, 299 Bildsten, L., 2000, proc. $10^{{\rm th}}$ Annual October Astrophysics Conference, ed. S.S. Holt and W. W. Zhang (astro-ph/0001135) Cocchi, M., Bazzano, A., Natalucci, L., et al., 1999a, IAUC 7307 Cocchi, M., Bazzano, A., Natalucci, L., et al., 1999b, A&A 346, L45 Cocchi, M., Bazzano, A., Natalucci, L., et al., 2000, proc. $4^{{\rm th}}$ [*INTEGRAL*]{} Workshop (astro-ph/0105112) Dickey, J.M., & Lockman, F.J., 1990, A&A Rev. 28, 215 Gotthelf, E.V., & Kulkarni, S.R., 1997, ApJ 490, L161 Heise, J., 1998, [*Nucl. Phys.*]{} B 69/1-3, 186 Heise, J., in ’t Zand, J.J.M., Smith, M.J.S., et al., 1999, [*Astrophys. Letters and Comm.*]{} 38/1-6, 297 Hoffman, J.A., Marshall, H.L., & Lewin, W.H.G., 1978, Nature 271, 630 Jager, R., Mels, W.A., Brinkman, A.C., et al., 1997, A&A 125, 557 in ’t Zand, J.J.M., Heise, J., Muller, J.M., et al., 1998, [*Nucl. Phys.*]{} B 69/1-3, 228 in ’t Zand, J.J.M., 2000, proc. $4^{{\rm th}}$ [*INTEGRAL*]{} Workshop (astro-ph/0104299) Kaptein, R.G., in ’t Zand, J.J.M., Kuulkers, E., Heise, J., and Cornelisse, R., 2000, A&A 358, L71 Lewin, W.H.G., van Paradijs, J., & Taam, R.E., 1993, Space Sci. Rev. 62, 223 Sunyaev, R.A., 1990, IAUC 5104 Ubertini, P., Bazzano, A., Cocchi, M., et al., 1999, [*Astrophys. Letters and Comm.*]{} 38/1-6, 301 van Paradijs, J., 1978, Nature 274, 650 van Paradijs, J., & Mc Clintock, J.E., 1994, A&A 290, 133
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper addresses the problem of tracking mobile intruders in a polygonal environment. We assume that a team of diagonal guards is deployed inside the polygon to provide mobile coverage. First, we formulate the problem of tracking a mobile intruder inside a polygonal environment as a multi-robot task allocation (MRTA) problem. Leveraging on guard deployment strategies in art gallery problems for mobile coverage, we show that the problem of finding the minimum speed of guards to persistently track a single mobile intruder is NP-hard. Next, for a given maximum speed of the intruder and the guards, we propose a technique to partition a polygon, and compute a feasible allocation of guards to the partitions. We prove the correctness of the proposed algorithm, and show its completeness for a specific class of inputs. We classify the guards based on the structural properties of the partitions allocated to them. Based on the classification, we propose motion strategy for the guards to track the mobile intruder when it is located in the partition allocated to the guard. Finally, we extend the proposed technique to address guard deployment and allocation strategies for non-simple polygons and multiple intruders.'
author:
- 'Guillermo J. Laguna'
- Sourabh Bhattacharya
bibliography:
- 'new\_references.bib'
date: 'Received: date / Accepted: date'
title: 'Tracking Mobile Intruders in an Art Gallery: Guard Deployment Strategies, Fundamental Limitations, and Performance Guarantees '
---
[example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore
[Target tracking, Mobile coverage, Art gallery, Multi-robot task allocation, Sliding cameras]{}
Introduction {#sec:intro}
============
Security is an important concern in infrastructure systems. For decades, autonomous mobile robots have been utilized as surveillance [@Mittal:2016][@Witwicki:2017] and crime-fighting agents for barrier assessments [@Theodoridis:2012][@Ciccimaro:1999], intruder detection [@Tuna:2012][@Bazydlo:2017], building virtual terrains or maps [@Oriolo:1998][@Montemerlo:2006], neutralizing explosives [@Majerus:2010][@Nicoud:1995], and recognizing abnormal human behaviors [@Rezazadegan:2017][@Popoola:2012]. Such robots have been designed with the ability to counter threats, limit risks to personnel, and reduce manpower requirements in hazardous environments [@Lee:2003][@Yamamoto:1992][@Theodoridis:2012]. Deployment of these surveillance platforms in teams has given rise to challenging problems in collaborative sensing and decision-making. Motivated from recent surge in interest in autonomous surveillance, we address an asset protection problem in which a team of mobile sensors collaborate to track suspicious mobile entities to secure an environment.
Although advanced electronic and biometric techniques can be used to secure facilities, vision-based monitoring is widely used for persistent surveillance. The idea is to visually cover the environment in order to obtain sufficient information so that appropriate measures can be taken to secure the area in case of any suspicious/malicious activity. The general formulation of a tracking problem consists of a team of autonomous sensing platforms, called [*observers*]{}, that visually track mobile entities, called [*targets*]{}. In this work, we consider the aforementioned formulation in the presence of features in the environment that can occlude the targets from the observers, for example, presence of reflex vertices on the boundary of the environment and obstacles. Precocious planning and coordination between observers can prevent the intruders from escaping the visual footprint of the observers around such occlusions. Deploying a network of autonomous sensing platforms has been an active area of research in multi-robot systems (MRS). Multi-robot systems have emerged as an important area of research in robotics due to their potential applications in several areas, for example, autonomous sensor networks [@Ghosh:2017], building surveillance [@Dipaola:2010], transportation of large objects [@Sakuyama:2012], air and underwater pollution monitoring [@Dunbabin:2012], forest fire detection [@Afzaal:2016], transportation systems [@Matsuhira:2010], or search and rescue after large-scale disasters [@Ko:2009]. Even problems that can be handled by a single multi-skilled robot may benefit from the alternative usage of a robot team, since robustness and reliability can often be increased by combining several robots which are individually less robust and reliable. In case of target tracking, multiple points of view from multiple robots add extra information on the target resulting in a better estimate of its position. However, the uncertainty in the future actions of the target, and the tight coupling between sensing, coordination and control within the team of mobile observers gives rise to challenging problems in multi-robot motion planning.
A simple solution to the tracking problem in bounded environments is to cover the environment with sufficient number of observers. This leads to the [*art gallery problem*]{}, a well-studied topic in computational geometry. In the classical art gallery problem, the goal is to determine the number of guards needed to visually cover a bounded polygon [@Fisk:1978]. A guard can only see the portion of the polygon unobstructed by the boundary of the polygon or by internal obstacles. Over the years several variants of the problem have been studied based on the shape of the polygon (orthogonal [@Kahn:1983; @Durocher:2017; @Durocher:2014; @Durocher:2015], monotone [@Deberg:2017], etc), the type of guards employed (static [@O'Rourke:1987; @Kahn:1983; @Hoffmann:1990; @Czyzowicz:1994; @Urrutia:2000], mobile [@O'Rourke:1983; @Toussaint:1982; @Colley:1995; @Kay:1970]), and the notion of visibility ($k$-transmitters [@Mahdavi:2014], multi-guarding [@Fazli:2010; @Fazli:2010a] etc). Stationary guards can either be point guards (placed anywhere inside the polygon) or vertex guards (restrict them only to the vertices of the polygon). It has been shown that the problem of computing the minimum number of stationary guards required to cover a simply connected polygon is NP-hard [@Lee:1986]. In [@O'Rourke:1987], it is shown that $\lfloor n/3 \rfloor$ static guards with omni-directional field-of-view is sufficient, and sometimes necessary to cover the entire polygon, where $n$ is the number of vertices of the polygon representing the environment. Efforts to obtain a bound on the minimum number of guards required to cover a polygon for special cases include approximation [@Ghosh:2010] and heuristic techniques [@Amit:2010].
The notion of mobile guards was first introduced by Toussaint [@Toussaint:1982] where each guard can travel back and forth along a segment inside the polygon, and every point in the polygon must be seen by at least one guard at some point of time along its path. Since the guards do not visually cover the entire polygon at all times, this notion of coverage is called [*mobile coverage*]{}. In [@O'Rourke:1983] it is proved that $\lfloor n/4 \rfloor$ diagonal guards are sufficient to provide mobile coverage. Therefore, if the guards have sufficient speed, at most $\lfloor n/4 \rfloor$ of them are required to track a mobile intruder in a polygonal environment. This naturally leads to the following question: What is the minimum speed required for the diagonal guards to track a mobile intruder in the environment with known maximum speed? The answer depends on the coordination and control strategy used by the guards.
The need for coordination arises in multi-robot systems (MRS) when individual agents work together to complete a common task [@Barnes:1991]. Depending on the communication scheme adopted by the MRS, coordination can be explicit or implicit. As the name suggests, explicit communication requires the presence of an on-board communication module on each robot of the MRS. On the other hand, implicit communication is generally contextual, based on the sensor data available to a robot regarding its teammates. Due to significant improvements in mobile communication hardware, the coordination scheme prevalent in current MRSs is predominantly explicit in nature. In this work, we propose an explicit coordination scheme for the team of observers. Our scheme explicitly relies on the location of the observers relative to each other thereby building a connection between the geometric aspects of the team formation, and communication topology required for persistent tracking.
In this work, we formulate the tracking problem as a multi-robot task allocation problem (MTAP) [@Gerkey:2004a]. The allocation tries to balance the workload among the observers thereby minimizing the speed required to ensure tracking. The main contributions of this work are as follows:
1. We investigate a variant of the multi-robot target tracking problem in which observers are constrained to move along diagonals of the polygonal environment. We leverage guard deployment strategies proposed for mobile coverage [@O'Rourke:1983] to design deployment and tracking strategies for multiple observers. To the best of our knowledge, our work is the first to build a connection between the well known art gallery problem which deals with coverage and multi-robot tracking.
2. We propose an algorithm that jointly partitions a polygon, and allocates observers to track a target inside the partition. The algorithm is a solution to a resource allocation problem in which the observers are resources for tracking that need to be assigned to partitions of the polygon. The joint partitioning and allocation algorithm is based on the triangulation of the polygon, and therefore, can be extended to polygon with holes. Under the restriction that a single observer can be allocated to each partition, we show that the problem of finding the minimum observer speed required to track an intruder is NP-hard.
3. We present a taxonomy of the observers based on their task allocation, and the level of cooperation needed from other members of the team to perform the tracking task. We introduce the notion of [*critical curves*]{} to construct activation strategies for the mobile observers.
4. We derive the maximum number of targets that can be tracked using the deployment and partitioning algorithm proposed in this work.
The paper is organized as follows. Section \[sec:rel\] presents a brief description of the related work. Section \[sec:prob\] presents the problem formulation and a deployment strategy for the guards. Section \[sec:alloc\] formulates the tracking problem as a multi-robot task allocation problem. Section \[sec:gen\] presents a partitioning technique for the polygon, and an allocation algorithm to assign the partitions to the guards to track a single intruder in the environment. Section \[sec:holes\] addresses the case of non-simple polygons. Section \[sec:multi\] extends the proposed deployment and allocation algorithm to address the case of multiple intruders in the environment. Section \[sec:conclusion\] presents the conclusions and future work.
Related Work {#sec:rel}
============
The multi-robot target-tracking problem was originally introduced by Parker and Emmons in [@Parker:1997]. The authors proposed the framework of Cooperative Multi-robot Observation of Multiple Moving Targets (CMOMMT) for a team of observers that simultaneously maximize the number of targets under observation, and the duration of observation of each target. It is shown that the CMOMMT problem is NP-hard. Consequently, several variants of CMOMMT have been studied to propose implementable solutions, for example, Approximate CMOMMT [@Parker:2002], personality-CMOMMT [@Ding:2006], behavioral-CMOMMT [@Kolling:2006; @Kolling:2007], formation-CMOMMT [@Werger:2000], to name a few. Alternate formulations of the CMOMMT based on particle filtering [@Spletzer:2003] and mixed non-linear integer programming formulation [@Xu:2013a] have also been proposed in the past. These frameworks lead to a numerical approach for generating the trajectory of the observers.
In the past, several variants of the original tracking problem posed in [@Parker:1997] have been addressed. In [@Isler:2005], the authors analyze the [*focus of attention*]{} problem [@Goossens:2016] for specific formations of the observers, and provide approximation algorithms for optimal target allocation. Dames et al. address the problem of detecting, localizing and tracking a team of targets with unknown and time-varying cardinality [@Dames:2018]. In [@Hausman:2015], the authors address the cooperative control of a team of UAVs that try to solve the joint problem of self localization and tracking with on-board sensors. In [@Jung:2002], a region-based coarse approach is proposed to simultaneously observe several mobile targets. The observers use a local method to maximize the number of observed targets. The proposed technique does not require any communication among the robots. In [@Markov:2007], CMOMMT is addressed for cooperative targets. The targets are able to communicate which allows active cooperation through sharing data. Thus the average observation time is minimized through a force field approach. In [@Tang:2005], a team of winged UAVs tries to minimize the average time elapsed between two consecutive observations of each member of a group of targets. An optimal control scheme is used to obtain the motion strategy of the observers.
In this work, we formulate the problem of tracking as a multi-robot task allocation problem. Multi-robot task allocation (MRTA) can be considered as an instance of the well-known optimal assignment problem. In [@Nam:2014], algorithms to solve the matching problem for weighted bipartite multi-graphs are used to solve MTAP. In [@Gerkey:2003] a domain-independent taxonomy of multi-robot task allocation problems is presented. They also analyzed and compared some iterated assignment architectures: ALLIANCE [@Parker:1999], BLE [@Werger:2000], and M+ [@Botelho:1999] and some on-line assignment architectures: MURDOCH (auction-based MRTA) [@Gerkey:2001], first-price auctions (market-based MRTA) [@Stentz:1999] and dynamic role assignment [@Chaimowicz:2002], for MRTA, respectively. In [@Lerman:2006], the authors present a mathematical modeling and analysis of the collective behavior of dynamic task allocation. In [@Yan:2011], a lightweight and robust multi-robot task allocation approach based on trade rules in market economy is presented. Other strategies are based on vacancy chains [@Dahl:2009], auction-based mechanisms [@Hanna:2005] and distributed market-based coordination [@Michael:2008].
In this work, we assume that the observers are omni-directional cameras that slide along the diagonals of the polygon. In [@Durocher:2017], the authors introduce the problem of placing sliding cameras on the boundary of the polygon for mobile coverage. Unlike the standard pin-hole camera model, a sliding camera in [@Durocher:2017] can see a point $p_1$ inside the environment if there exists a point $p_2$ along its trajectory such that the segment $p_1p_2$ is perpendicular to the trajectory of the guard and $p_1p_2$ is fully contained in the environment [@Durocher:2013]. The objective in [@Durocher:2017] is to find the minimum number of sliding cameras that can provide mobile coverage, referred to as the minimum sliding camera (MSC) problem. This problem is a variant of the art gallery problem which is NP-hard even for orthogonal polygons [@Schuchardt:1995]. However, it has been proven to be APX-hard on simple polygons [@Eidenbenz:1998]. Polynomial time approximation algorithms have been developed for orthogonal polygons [@Durocher:2017; @Durocher:2014; @Durocher:2015]. In [@Durocher:2013], it is shown that MSC is NP-hard for non-simple polygons. For monotone orthogonal polygons, a linear time solution to MSC is presented in [@Deberg:2017]. In [@Durocher:2013], the minimum length sliding camera problem is introduced. The objective is to find a set of sliding cameras that minimizes the total length of the trajectory while providing mobile coverage [@Almahmud:2016]. A polynomial time solution for orthogonal polygons with holes is presented in [@Durocher:2013]. Some variants include the concept of cameras with $k$-transmitters, introduced in [@Mahdavi:2014], which allows the cameras to see through $k$ boundary walls of the polygon. In [@Mahdavi:2014], the problem is shown to be NP-complete, and the authors present a $2$-approximation algorithm.
Problem Statement {#sec:prob}
=================
Let $P$ be a simple $n$-vertex polygon representing a closed polygonal environment. First, we consider the case of a single unpredictable target $I$, referred to as *intruder*, inside $P$. In Section \[sec:multi\], we extend our analysis to address multiple intruders inside the polygon. Let $\bar{v}_e$ and $x_I$ denote the maximum speed of the intruder and its location respectively at time $t$. In order to track the mobile intruder, a team of mobile observers, referred to as [*guards*]{}, is deployed inside the polygon. Let $\mathbb{G}$ denote the set of all guards. Each guard is equipped with an omni-directional camera with infinite range. The objective of the guards is to ensure that $I$ is visible to at least one guard at all times. The guards have knowledge of $\bar{v}_e$. Each guard has a maximum speed $\bar{v}_g$ and the *speed ratio* between the guards and the intruders is defined as $r= \bar{v}_g/\bar{v}_e$.
Next, we define some graph-theoretic concepts associated with a polygon and its partitioning. A [*triangulation*]{} of $P$ is defined as a partition of $P$ into a set of disjoint triangles, such that the vertices of the triangles are vertices of the polygon. In general, the triangulation of a polygon is non-unique. The edges of the triangles of the triangulation are called *diagonals* [@O'Rourke:1983], and they can be segments inside $P$ (internal diagonals) or edges of $P$. Thus, the triangulation of $P$ is trivially represented as a planar graph $G=G(P)$ called *triangulation graph*. We assume that a triangulation of $P$ is given. Let $\mathbb{V}(G)$ denote the vertex set of $G$ (corresponds to the vertices of $P$), $\mathbb{E}(G)$ denote the edge set of $G$ (corresponds to the diagonals of the triangulation of $P$), and $\mathbb{T}(G)$ (triangle set of $P$) denote the faces of $G$. Clearly, there is a bijection between the set of vertices of $P$ and $\mathbb{V}(G)$. Also, there is a bijection between the set of diagonals of the triangulation of $P$ and $\mathbb{E}(G)$. Hence, we do not make any distinction between the following pairs: (i) vertices of $P$ and the vertices in $\mathbb{V}(G)$ (ii) diagonals of the triangulation of $P$ and the edges in $\mathbb{E}(G)$ (iii) triangles of the triangulation of $P$ and the faces in $\mathbb{T}(G)$. Let $G_D$ be the dual graph of $G$. Each vertex in $G_D$ corresponds to a face in $\mathbb{T}(G)$. An edge exists between a pair of vertices in $G_D$ if the triangles in $\mathbb{T}(G)$ that correspond to such pair of vertices share an edge. Since $P$ is simply connected, $G_D$ is a tree.
The objective is to keep the intruder in the line-of-sight of at least one guard at all times, which is ensured if the triangle in which the intruder is located is covered[^1] by at least one guard. Each $g_i \in \mathbb{G}$ is a [*diagonal guard*]{} i.e., it is constrained to move along a unique diagonal $h_i \in \mathbb{H}$, where $\mathbb{H} \subset \mathbb{E}(G)$. Let $l_i$ be the length of $h_i \in \mathbb{H}$. The endpoints of $h_i$ are denoted as $v_1(i)$ and $v_2(i)$. Guards $g_i$ and $g_k$ are *neighboring guards* if $g_i$ and $g_k$ are incident[^2] to the same triangle. We define $\mathbb{G}(T_k) \subseteq \mathbb{G}$ as the set of guards incident to $T_k$.
In this paper, the distance between two points $x,y \in P$ is defined as the length of the shortest path between $x$ and $y$ on the visibility graph[^3] constructed using the vertices of $P$ (and the vertices of internal obstacles), $x$ and $y$. It is denoted as $d(x,y)$. The distance between $x\in P$ and a set $R_1 \subset P$ is defined as $d(x,R_1)=\min \{d(x,p) : p \in R_1 \}$. The distance between two sets of points $R_1,R_2\subset P$ is defined as $d(R_1,R_2)= \min \{d(q,R_1) : q \in R_2 \}$.
[**Note: Appendix \[sec:appc\] contains a list of important variables, and their definitions.** ]{}
In the next section, we describe the deployment of the guards inside the polygon.
Selection of Dominating Diagonals {#subsec:dep}
---------------------------------
In [@O'Rourke:1983], it is shown that at most $ \lfloor n/4 \rfloor$ diagonals are sufficient to dominate[^4] $G$. Hence, $\lfloor n/4 \rfloor$ diagonal guards are sufficient to provide mobile coverage. In this section, we describe the strategy proposed in [@O'Rourke:1987] to identify at most $\lfloor n/4 \rfloor$ *dominating diagonals* [^5] of a polygon’s triangulation. In [@O'Rourke:1987], it is shown that there exists a set of dominating diagonals of size $\lfloor n/4 \rfloor$ in every triangulation graph of a polygon with $n \geq 5$. The correctness of the strategy is based on the following results [@O'Rourke:1987]: (i) For any triangulation graph $G$ of a simple polygon $P$ with $n \geq 10$ edges, it is always possible to find a diagonal $d \in \mathbb{E}(G)$ that partitions $G$ into subgraphs $G_1$ and $G_2$ such that $G_1$ is the triangulation graph of a hexagon, heptagon, octagon or nonagon (we call these *basic polygons*). (ii) Any triangulation graph of a heptagon (or any polygon with fewer sides) has one dominating diagonal, and any triangulation graph of an octagon or nonagon has two dominating diagonals.
Algorithm \[alg:guards\] recursively partitions $G$ into triangulation subgraphs of basic polygons. At each iteration, the algorithm searches for a diagonal $d$ that separates a triangulation subgraph of a basic polygon (denoted as $G_p$) such that there is no other diagonal in $\mathbb{E}(G_p)$ (edge set of $G_p$) that can separate a subgraph of a smaller basic polygon. Additionally, for each subgraph $G_p$ obtained, a minimal set of diagonals $\mathbb{E}(G_p)$ is found such that the diagonals in the set can dominate all the triangles[^6] in $\mathbb{T}(G_p)$ (set of faces of $G_p$) that are still not dominated. The process is repeated until the remaining non-partitioned subgraph has $9$ vertices or less. The first **while** cycle (Line $5$) finds a diagonal $d$ that partitions $G$ into a pair of triangulation subgraphs, such that one of those subgraphs ($G_p$) corresponds to a basic polygon and such that there is no other diagonal in $\mathbb{E}(G_p)$ (edge set of $G_p$) that can separate a subgraph of a smaller basic polygon. This can be completed in $O(n)$ time by traversing $G_D$ (dual graph of $G$).
$G_{pol}$ (Line $10$) is a tree such that each $v \in \mathbb{V}(G_{pol})$ is associated with each subgraph $G_p$ found in Line $6$, and one vertex corresponds to the remaining subgraph $G'$ after Line $9$. An edge $e \in \mathbb{E}(G_{pol})$ exists between vertices that correspond to a pair of triangulation subgraphs $G_p$ that share a common diagonal $d \in \mathbb{E}(G)$. In the second **while** loop the minimum set of diagonals that can cover the triangles in $\mathbb{T}(G_i)$ that are not dominated by other diagonals is found (set of *appropriate* diagonals, Line $14$). The second **While** loop takes $O(n)$ time by using $G_{pol}$. Hence, Algorithm \[alg:guards\] takes $O(n)$ time.
**Input**: $G$ **Output**: $\mathbb{H}$ $S_D \leftarrow \emptyset$ is the set of diagonals $d$ $G' \leftarrow G$ find $d$ that separates a triangulation subgraph $G_p$ from $G'$ such that there is no triangulation subgraph of $G_p$ that corresponds to a smaller basic polygon $G'$ becomes the subgraph obtained by removing $G_p$ excepting $d$ and its vertices add $d$ to $S_D$ create $G_{pol}$ from $G$ using the diagonals in $S_D$ (all vertices in $\mathbb{V}(G_{pol})$ are unmarked) $v_i \leftarrow$ unmarked vertex in $G_{pol}$ with at most one unmarked neighbor $G_i \leftarrow$ subgraph of $G$ that corresponds to $v_i$ $H_{i} \leftarrow$ *appropriate* diagonals of $\mathbb{E}(G_i)$ add diagonals of $H_{i}$ to $\mathbb{H}$ mark $v_i$
\[alg:guards\]
We call $\mathbb{H}$ a *deployment* of the guards, since each $g_i \in \mathbb{G}$ is assigned to a diagonal $h_i \in \mathbb{H}$ (it is deployed along the diagonal). In the next section, we address the problem of finding the minimum speed of the guards that can ensure tracking.
A Multi-robot Task Allocation Problem {#sec:alloc}
=====================================
A sufficient condition to track the intruder is to cover at all times the triangle in which it lies. In the previous section, we described an algorithm to select at most $ \lfloor n/4 \rfloor$ diagonals of a polygon that can dominate the triangles of the triangulation of a polygon. Given sufficient speed, the guards assigned to the diagonals in $\mathbb{H}$ can move back and forth on their diagonals to cover the triangle in which the mobile intruder lies. This gives rise to a MRTA problem in which the task for each robot is to cover the triangles allocated to it whenever the intruder lies inside them.
First, we consider the case in which a single guard is allocated to a triangle of the triangulation. An allocation $A:{\mathbb{G}}\rightarrow2^{\mathbb{T}(G)}$ maps each guard to a subset of triangles in the triangulation of the polygon. Let $\mathcal{A}$ denote the set of all possible allocations. We say that an allocation is *complete* if there is no triangle with no guard allocated to it. Next, we classify the triangles of the triangulation of $P$ (the triangles in $\mathbb{T}(G)$) based on the number of incident guards and their position relative to the triangles. Refer to Figure \[fig:triangles\]. The red segments in the figure represent the diagonals $h_i \in \mathbb{H}$ on which the guards move. In the subsequent figures, those diagonals are labeled as $g_i$ instead.
1. Safe Triangle: A triangle $T_k \in \mathbb{T}(G)$ is called *safe* if $g_i$ covers $T_k$ at all times regardless of the location of $x_I$. Clearly, if there is a guard $g_i \in \mathbb{G}$ deployed on one of the edges of $T_k$, the triangle is a safe one. We use $\mathbb{T}^{safe}(G)$ to denote the set of safe triangles. In Figure \[fig:triangles\], safe triangles are shaded in blue.
2. Unsafe Triangle: A triangle $T_k \in \mathbb{T}(G) \backslash \mathbb{T}^{safe}(G)$ is called *unsafe* if $|\mathbb{G}(T_k)|=1$, where $|\cdot|$ is the cardinality operator. In Figure \[fig:triangles\], unsafe triangles are shaded in orange.
3. Regular Triangle: A triangle $T_k$ is *regular* if it is neither safe nor unsafe. In Figure \[fig:triangles\], regular triangles are shaded in white.
![Classification of a set of triangles in $\mathbb{T}(G)$. Notice that the regular triangles $T_1$ and $T_2$ can be covered by more than one guard.[]{data-label="fig:triangles"}](triangles_new){width="0.62\linewidth" height="0.43\linewidth"}
Let $\mathbb{T}_{\alpha}^{safe}(g_i)$ denote the set of triangles that are incident to the endpoint $v_{\alpha}(i)$ of $h_i$ ($\alpha \in \{ 1,2 \}$). Also, let $\overline{\mathbb{T}^{safe}}(G)$ $(=\mathbb{T}(G)\backslash\mathbb{T}^{safe}(G))$ denote the set of non-safe (Regular$+$Unsafe) triangles in $\mathbb{T}(G)$ and $\overline{\mathbb{T}_{\alpha}^{safe}}(g_i)$ denote the set of non-safe triangles that are incident to $v_{\alpha}(i)$. By definition, $\mathbb{T}_{\alpha}^{safe}(g_i)$ is always covered by $g_i$ for every allocation $A(\mathbb{G})\in\mathcal{A}$. We want to assign a $g_i \in \mathbb{G}$ to every $T \in \overline{\mathbb{T}^{safe}}(G)$ so that $g_i$ can cover $T$ when the intruder lies in it. Clearly, only the guards in $\mathbb{G}(T)$ can be assigned to $T$. The deployment strategy proposed in Section \[subsec:dep\] guarantees that $|\mathbb{G}(T)| \geq 1$ for every $T \in \overline{\mathbb{T}^{safe}}(G)$. Therefore, an allocation that assigns a guard to each triangle in $\mathbb{T}(G)$ exists. Moreover, every unsafe triangle can only be assigned to the single guard incident to it.
Given an allocation $A(\mathbb{G})$, there are two possibilities for each $g_i \in \mathbb{G}(T)$:
1. All the non-safe triangles assigned to $g_i$ are in $\overline{\mathbb{T}_{\alpha}^{safe}}(g_i)$. In this case, $g_i$ is static, and can cover all the triangles assigned to it from $v_{\alpha}(i)$ regardless of the speed ratio $r$.
2. There are triangles assigned to $g_i$ in $\overline{\mathbb{T}_1^{safe}}(g_i)$ and $\overline{\mathbb{T}_2^{safe}}(g_i)$. This implies that $g_i$ needs to be at $v_1(i)$ ($v_2(i)$) when the intruder $I$ lies in $\bigcup_{T \in \overline{\mathbb{T}_1^{safe}}(g_i)}{T}$ ($\bigcup_{T \in \overline{\mathbb{T}_2^{safe}}(g_i)}{T}$). Therefore, $g_i$ has to move from one endpoint of $h_i$ to the other depending on $x_I$. Assume that $I$ is initially located in any triangle $T_j\in \overline{\mathbb{T}_1^{safe}}(g_i)$ assigned to $g_i$, and $I$ moves to a triangle $T_k\in \overline{\mathbb{T}_2^{safe}}(g_i)$ also assigned to $g_i$. As a result, the guard $g_i$, initially located at $v_1(i)$, moves to $v_2(i)$. Therefore, $g_{i}$ should reach $v_2(i)$ before the intruder can reach $T_k$. If the aforementioned condition is satisfied for every $T_j \in \overline{\mathbb{T}_1^{safe}}(g_i)$ and every $T_k \in \overline{\mathbb{T}_2^{safe}}(g_i)$ assigned to $g_i$, then $g_i$ can track $I$ when it is inside any triangle allocated to it.
To summarize the above discussion, an allocation $A(\mathbb{G})$ should satisfy the following conditions:
1. $A$ is complete.
2. Unsafe triangles are allocated to the single guard incident to them.
3. $\bar{v}_g\geq\max_i \bar{v}_e l_i/d(T_1^{alloc},T_2^{alloc})$, where $T_1^{alloc} \subset \overline{\mathbb{T}_1^{safe}}(g_i), T_2^{alloc} \subset \overline{\mathbb{T}_2^{safe}}(g_i)$ are the sets of triangles allocated to $g_i$ that correspond to $v_1(i)$ and $v_2(i)$ respectively.
An allocation $A$ is called [*feasible*]{} if it satisfies the above conditions. Given the constraint that a single guard can be allocated to a non-safe triangle, we pose the following problem:
[**Problem 1:**]{} Given $\bar{v}_e$, what is the minimum value of $\bar{v}_g$ and the corresponding allocation $A$ for which the intruder can be persistently tracked?
$G_D$ encodes the adjacency between faces of $G$ based on common diagonals. Analogous to a dual graph, we define a [*Guard Adjacency Graph*]{} (GAG), denoted as $G^{\#}$, which encodes the adjacency between faces of the triangulation graph $G$ based on common guards. For a given triangulation graph $G$ of the polygon and deployment of the guards, $G^{\#}$ is constructed as follows. Each vertex of $G^{\#}$ corresponds to an non-safe triangle in $G$. An edge $e_{j,k}(g_i)$ exists between vertices $v_j$ and $v_k$ in $G^{\#}$ if there exists a guard $g_i$ such that $T_j \in \overline{\mathbb{T}_1^{safe}}(g_i)$ and $T_k \in \overline{\mathbb{T}_2^{safe}}(g_i)$. The weight of the edge $e_{j,k}(g_i)$ is given by $w_{j,k}(g_i)=l_{i}/d(T_j,T_k)$.
Next, we define terms related to allocation of guards to triangles. Given an allocation $A(\mathbb{G}) \in\mathcal{A}$, let $A(g_i)$ be the set of triangles assigned to $g_i$. Then we define the cost of the allocation of $g_i$ as $c(A(g_i))=\displaystyle\max_{T_j,T_k \in A(g_i)} \{w_{j,k}(g_i)\}$. The overall cost of the allocation for $G$ is defined as $c(A(\mathbb{G}))=\displaystyle\max_{g_i\in \mathbb{G}} \{c(A(g_i))\}$. Based on these definitions, Problem $1$ can be formulated as the following problem on $G$:\
\
[**UNIALLOC**]{}: Find a feasible allocation $A(\mathbb{G})\in\mathcal{A}$ for which $c(A(\mathbb{G}))$ is minimized. In order to prove the above theorem, first we define a set of representatives from $G^{\#}$. For each non-safe triangle $T_j \in \overline{\mathbb{T}^{safe}}(G)$, we define a set $S_g(T_j)=\{(j,g_i): g_i \in \mathbb{G}(T_j) \}$. $S_g(T_1),\ldots,S_g(T_{|\mathbb{V}(G_1)|-1}),S_g(T_{|\mathbb{V}(G^{\#})|})$ is a collection of disjoint sets. For any $x=(j,g_a) \in S_g(T_j)$ and $y=(k,g_b) \in S_g(T_k)$ with $j \neq k$ and guards $g_a,g_b$, define $c_2(x,y)$ as follows:
$$\label{eq:cost}
c_2(x,y)=\left \{ \begin{array}{cc}
w_{j,k}(g_a), & a=b\\
0, & \mbox{ otherwise.}
\end{array} \right.$$
$c_2(x,y)=0$ for $a \neq b$ models the fact that $g_a$ and $g_b$ can cover $T_j$ and $T_k$, respectively, by staying static at the corresponding endpoints. If $a=b$ and $e_{j,k}(g_a) \in \mathbb{E}(G^{\#})$, $T_j$ and $T_k$ are are located at opposite ends of $h_a$. In this case, $g_a$ should have a minimum speed of $c_2(x,y) =l_{i}/d(T_j,T_k)$ to cover triangles $T_j$ and $T_k$ if the intruder moves between them along the shortest path between them. Finally, $a=b$ and $e_{j,k}(g_a) \notin \mathbb{E}(G^{\#})$ implies that $T_j$ and $T_k$ are incident to the same endpoint of $h_a$. Since $g_a$ can cover both triangles from that endpoint, it can remain static. Therefore, $c_2(x,y) =0$ in that case.
We say that a set $S_{rep} \subset \displaystyle\bigcup_{j\in\{1,\ldots, |\mathbb{V}(G^{\#})|\}} S_g(T_j)$ is a *set of representatives* if $|S_{rep} \cap S_g(T_j)|=1$ for all $j$. Based on the definition of a set of representatives, we define the following problem which is equivalent to UNIALLOC:
[**MAXREP**]{}: Find the set $S_{rep}$ for which $\max \{ c_2(x,y): x,y \in S_{rep} \mbox{ and } x \neq y \}$ is minimized.
The following problem casts MAXREP as a decision problem:
[**GAMMAREP:**]{} Given $\gamma \in \mathbb{R}_{\geq 0}$, does there exist a set of representatives $S_{rep}$ of size $|\mathbb{V}(G^{\#})|$ such that $c_2(x,y) \leq \gamma$ for all $x,y \in S_{rep}$ with $x \neq y$?
The next lemma proves that GAMMAREP is NP-hard which in turn implies that MAXREP is at least NP-hard.
\[lemma:gamma\] GAMMAREP is NP-hard.
We reduce the problem of finding a $K$-clique **[@Karp:1972]** in a graph (a NP-hard problem [@Chen:2005]) to GAMMAREP. Consider a graph without self-loops $G_2$ such that $|\mathbb{V}(G_2)| > |\mathbb{V}(G^{\#})|$. The problem of finding a $|\mathbb{V}(G^{\#})|$-clique in $G_2$ is NP-hard [@Garey:2002]. For each $v_j \in \mathbb{V}(G_2)$, we define a set $S_j=\{ j \} \times \mathbb{V}(G_2)$. For a pair of vertices $v,w \in \mathbb{V}(G_2)$, we define the following cost function:
$$\label{eq:cost2}
c_3((j,v),(k,w))= \left\{ \begin{array}{cc} 0, & \mbox{ if an edge exists between $v$ and}\\& \mbox{$w$ in $G_2$} \\ 1, & \mbox{ otherwise} \end{array} \right.$$
Given sets $S_j$ and the cost function $c_3$ stated above, we prove that a $|\mathbb{V}(G^{\#})|$-clique in $G_2$ exists if and only if a set of $|\mathbb{V}(G^{\#})|$ representatives $S_{rep} \subset \displaystyle\bigcup_{j\in\{1,\ldots, |\mathbb{V}(G^{\#})|\}} S_j$ such that $c_3((j,v),(k,w)) = 0 \text{ }\forall \text{ }(j,v),(k,w) \in S_{rep}$ can be found.
($\Rightarrow)$ First, assume that there exists such a $S_{rep}$ but there is no $|\mathbb{V}(G^{\#})|$-clique in $G_2$. By definition, for each pair $(j,v),(k,w) \in S_{rep}$, $j \neq k$, and $v \neq w$ ($v=w$ implies that $G_2$ has a self-loop which is not possible). Therefore, each $(j,v) \in S_{rep}$ is associated to a different $v \in \mathbb{V}(G_2)$. Since there is no $|\mathbb{V}(G^{\#})|$-clique in $G_2$, the subgraph induced by the vertices associated with $S_{rep}$ is not complete. Therefore, there is at least one pair $(j,v),(k,w) \in S_{rep}$ such that there is no edge shared between $v$ and $w$ in $G_2$ which implies $c_3((j,v),(k,w)) \neq 0$ (a contradiction). Therefore, if there exists $S_{rep}$ of size $|\mathbb{V}(G^{\#})|$ such that $c_3((j,v),(k,w)) = 0 $ for all $(j,v),(k,w) \in S_{rep}$ then a $|\mathbb{V}(G^{\#})|$-clique in $G_2$ exists.
($\Leftarrow$) Now assume that there is no $S_{rep}$ which meets the aforementioned constraints, but there is a $|\mathbb{V}(G^{\#})|$-clique in $G_2$. Since $G_2$ contains a $|\mathbb{V}(G^{\#})|$-clique, there is a subset of $|\mathbb{V}(G^{\#})|$ vertices such that its induced subgraph $\hat{G}_2$ is complete. Since $\hat{G}_2$ is complete, choosing $(j,v_j) \in S_j$ for each $v_j \in \mathbb{V}(\hat{G}_2)$ as an element of $S_{rep}$ implies that for any $(j,v_j),(k,v_k) \in S_{rep}$, an edge exists between $v_j$ and $v_k$ in $G_2$. Therefore, $c_3(v_j,v_k) =0$ for every $(j,v_j),(k,v_k) \in S_{rep}$ which contradicts the assumption that there is no $S_{rep}$ of size $|\mathbb{V}(G^{\#})|$ such that $c_3((j,v_j),(k,v_k)) = 0 $ for all $(j,v_j),(k,v_k) \in S_{rep}$. Therefore, if there exists a $|\mathbb{V}(G^{\#})|$-clique in $G_2$ then there is $S_{rep}$ of size $|\mathbb{V}(G^{\#})|$ such that $c_3((j,v),(k,w)) = 0 $ for all $(j,v),(k,w) \in S_{rep}$, and the reduction can be done in polynomial time. Thus, we have a polynomial-time reduction from $K$-clique to GAMMAREP.
Based on the above discussion, we can state the following theorem.
\[theorem:unialloc\] UNIALLOC is NP-hard.
The proof follows from the fact UNIALLOC is equivalent to MAXREP which is at least NP-hard.
Suboptimal Algorithm for Computing Approximate minimum speed ratio {#subsec:subop}
------------------------------------------------------------------
MAXCLIQUE refers to the problem of finding the maximum-sized clique in a graph. MAXCLIQUE is NP-hard [@Pardalos:1994]. An approximation algorithm for MAXCLIQUE can be used to obtain a suboptimal solution for Problem 1. The procedure is as follows. Let $G_3$ be a graph such that $\mathbb{V}(G_3)=\{v_j^i: \exists S_g(T_j) \mbox{ such that } (j,g_i) \in S_g(T_j) \}$. Recall that $S_g(T_j)=\{(j,g_i): g_i \in \mathbb{G}(T_j) \}$ for all $1 \leq j \leq |\mathbb{V}(G^{\#})|$. Since $|\mathbb{V}(G^{\#})|=O(n)$ ($|\mathbb{V}(G^{\#})| \leq n-2$), and $ 1 \leq |S_g(T_j)| \leq |\mathbb{G}|=O(n)$, $|S_g(T_j)|=O(n)$. Therefore, $|\mathbb{V}(G_3)|=O(n^2)$. Additionally, let $\mathbb{E}(G_3)=\{e_{j,k}^{a,b}: v_j^a,v_k^b \in \mathbb{V}(G_3) \mbox{ with } j \neq k \}$. Each edge $e_{j,k}^{a,b}$ has an associated weight defined as follows:
$$\label{eq:cost3}
w_{j,k}^{a,b}= \left\{ \begin{array}{cc} 0, & \mbox{ if } a \neq b\mbox{ or } \frac{l_a}{d(T_j,T_k)} \leq r\\ 1, & \mbox{ otherwise} \end{array} \right.$$
We define $G_3'$ as a subgraph of $G_3$ such that $\mathbb{V}(G_3')=\mathbb{V}(G_3)$ and $\mathbb{E}(G_3') = \{ e_{j,k}^{a,b} \in \mathbb{E}(G_3): w_{j,k}^{a,b}=0 \}$. Therefore, $\mathbb{E}(G_3')$ consists of edges in $\mathbb{E}(G_3)$ with weights less than or equal to $r$. We can use any approximation algorithm for MAXCLIQUE to find a $|\mathbb{V}(G_1)|-$clique in $G_3'$ for a given $r$. Since $r_{min}=\min \{ r=\frac{l_i}{d(T_j,T_k)}: e_{j,k}(i) \in G^{\#} \}$, we perform the aforementioned check for values of $r$ that correspond to the weights of the edges in $G^{\#}$. The suboptimal allocation corresponds to the minimum value of $w_{j,k}(g_i)$ for which a $|\mathbb{V}(G_1)|-$clique is found in the graph $G_3'$. If no such $w_{j,k}(g_i)$ is found, $r_{min}$ is equal to the minimum $w_{j,k}(g_i)$. Each vertex $v_j^i \in \mathbb{V}(G_3')$ in the $|\mathbb{V}(G_1)|-$clique corresponds to a distinct triangle in $T_j \in \overline{\mathbb{T}^{safe}}$, and each one of them corresponds to a $g_i \in \mathbb{G}$ which is the guard assigned to $T_j$. This gives an approximate optimal allocation $A(g_i)$ for each $g_i \in \mathbb{G}$.
It has also been shown that approximating the MAXCLIQUE within a factor of $n^\epsilon$ for some $\epsilon>0$ is NP-hard [@Arora:1998; @Hastad:1999]. In the past, probabilistic techniques [@Babel:1991; @Balas:1996; @Wood:1997; @Pardalos:1994] have been proposed in the literature to find largest cliques without any guarantees on the optimality. [@Feige:2004] presents an approximation algorithm that finds a clique with an approximation ratio of $O(n (\log\log n)^2 /(\log n)^3)$ when the size of the maximum clique is between $n/ \log n$ and $n / (\log n)^3$. Since $\mathbb{V}(G_3)=O(n^2)$, where $n$ is the number of vertices of the polygon, the performance ratio in our case is $O(n^2 (\log (\log n))^2 /(\log n)^3)$.
Next, we consider the case in which more than one guard can be assigned to each non-safe triangle. In this case, we partition $T_j$ into disjoint regions, each covered by a unique guard (incident on $T_j$). The problem of computing the minimum guard speed ($\bar{v}_g$) when multiple guards can be assigned to a single triangle is as hard as UNIALLOC which itself is NP-hard.
In the next section, we address the problem of allocating guards to triangles for a given maximum speed of the intruder and the guards.
Computing a Feasible Allocation for Known Guard Speed {#sec:gen}
=====================================================
In the previous section, we proved that the problem of finding the minimum speed of the guards that can ensure tracking is NP-hard. In this section, we propose a technique to search and compute a feasible allocation for given maximum speed of the intruder and guards. Therefore, the speed ratio $r$ is known. We address the general allocation problem in which multiple guards can be assigned to cover an unsafe triangle in the polygon. Specifically, we focus on techniques that partition the polygon, and activate guards to track an intruder located in its allocated partition.
Let $R_j^\alpha(i)$ denote the region inside triangle $T_j$ assigned to the guard $g_i \in \mathbb{G}(T_j)$ incident to vertex $v_{\alpha}(i)$ ($\alpha = \{1,2\}$). We define $S_R^{\alpha}(i)=\{ R_j^\alpha(i) | R_j^\alpha(i) \subseteq T_j\text{ }\in \overline{\mathbb{T}_\alpha^{safe}}(g_i) \}$ as the set of regions assigned to $g_i$ that can be covered from $v_{\alpha}(i)$. We define $\hat{U}_R^{\alpha}(i)= \displaystyle\bigcup_{R_j^{\alpha}(i) \in S_R^{\alpha}(i)} R_j^{\alpha}(i)$ as the union of the regions belonging to $S_R^{\alpha}(i)$. Let $\mathbb{R}^{alloc}(g_i)=\hat{U}_R^{1}(i) \cup \hat{U}_R^{2}(i)$ denote the region assigned to $g_i$. The cost associated with an assignment is defined as $c(\mathbb{R}^{alloc}(g_i))= l_{i}/d(\hat{U}_R^{1}(i),\hat{U}_R^{2}(i))$. In this section, we address the following problem:
[**Problem 2:**]{} For the $\lfloor n/4 \rfloor$ deployment of guards inside a polygon described in Section \[subsec:dep\] and a given $r$, determine $\mathbb{R}^{alloc}(g_i)$ for each $g_i \in \mathbb{G}$ such that $\max \{c(\mathbb{R}^{alloc}(g_i)): g_i \in \mathbb{G} \} \leq r$, and the region in which the intruder lies can be covered by the guard allocated to it.
Since $c(\mathbb{R}^{alloc}(g_i))= l_{i}/d(\hat{U}_R^{1}(i)$,$\hat{U}_R^{2}(i)) \leq r$, $d(\hat{U}_R^{1}(i),\hat{U}_R^{2}(i)) \geq l_{i}/r$ for all $g_i \in \mathbb{G}$. Let $d_I^i=l_{i}/r$ denote the maximum distance traveled by the intruder during the time in which $g_i$ can travel from one endpoint of $h_i$ to the other.
In the next section, we present the procedure to find $\mathbb{R}^{alloc}(g_i)$.
Sequential computation of partitions {#subsec:cont}
------------------------------------
The allocation algorithm proposed in the next section sequentially computes $\hat{U}_R^{\alpha}(i)$ for each guard $g_i$. The end point of $h_i$ at which $\hat{U}_R^{\alpha}(i)$ is computed first gets the label $\alpha=1$. Subsequently, $\hat{U}_R^{\alpha}(i)$ is computed at the other end point of $h_i$ which is assigned the label $\alpha=2$.
From the definition of $\hat{U}_R^{1}(i)= \displaystyle\bigcup_{ R_j^{1}(i) \in S_R^{1}(i)} R_j^{1}(i)$, it is clear that $R_j^{1}(i)$ must be defined for every $T_j \in \overline{\mathbb{T}_{1}^{safe}}(g_i)$ for computing $\hat{U}_R^{1}(i)$. Based on the classification of $T_j \in \overline{\mathbb{T}_{1}^{safe}}(g_i)$, we can have the following scenarios:
1. If every $T_j \in \overline{\mathbb{T}_{1}^{safe}}(g_i)$ is an unsafe triangle, $R_j^{1}(i)=T_j$ by definition since $T_j$ can only be covered by $g_i$. In this case, $\hat{U}_R^{1}(i)= \displaystyle\bigcup_{T_j \in \overline{\mathbb{T}_{1}^{safe}}(g_i)} T_j$.
2. If $T_j$ is a regular triangle, $R_j^{1}(i)$ can be computed only if the region allocated to other guards incident to $T_j$ is known. In this case, $R_j^{1}(i) =\displaystyle\bigcap_{g_k \in \mathbb{G}(T_j) \backslash \{ g_i \} }{\overline{R}_j^2(k)} \cap T_j$, where $\overline{R}_j^2(k)$ is the complement of $R_j^2(k)$.
The following equation allows computing $\hat{U}_R^2(i)$ from $\hat{U}_R^{1}(i)$: $$\label{eq:r_2}
R_j^2(i)= \left\{ \begin{array}{cc}
\displaystyle\bigcap_{p \in \hat{U}_R^1(i)} \beta^{c}_{d_I^i}(p) \cap T_j^{free}, & \hat{U}_R^1(i) \neq \emptyset\\
T_j^{free}, & \mbox{otherwise}
\end{array}
\right.$$ where $\beta_{d_I^i}(p)$ is an open ball (using the metric defined in Section \[sec:intro\]) of radius $d_I^i$ centered at $p$, and $\beta^{c}_{d_I^i}(p)$ is its complement. Also, $T_j^{free} \subseteq T_j$ is the region inside $T_j$ that has not yet been assigned to a guard. In the absence of obstacles between $R_j^2(i)$ and $\hat{U}_R^1(i)$: $$\label{eq:r_minkowski}
R_j^2(i)= (\hat{U}_R^1(i) \bigoplus B_{d_I^i})^c \cap T_k^{free},$$ where $B_{d_I^i}$ is an open ball in $\mathbb{R}^2$, $\bigoplus$ denotes the Minkowski sum, and $(\hat{U}_R^1(i) \bigoplus \beta_{d_I^i})^c$ is the complement of the Minkowski sum. The set $(\hat{U}_R^1(i) \bigoplus B_{d_I^i})$ is called the offset of $\hat{U}_R^1(i)$. The offset of a set bounded by a polyline curve (line-segments/arcs of circle) is a set bounded by polyline curve [@Liu:2007]. Appendix \[sec:appa\] shows that $R_j^2(i)$ and $\hat{U}_R^2(i)$ are sets bounded by polyline curves even in the presence of obstacles. From $\hat{U}_R^{1}(i)$ and $\hat{U}_R^{2}(i)$, we can compute $\mathbb{R}^{alloc}(g_i)$ using the equation $\mathbb{R}^{alloc}(g_i)=\hat{U}_R^{1}(i) \cup \hat{U}_R^{2}(i)$.
In Figure \[fig:cont\], there are three guards, $g_1$, $g_2$, and $g_3$. All the non-safe triangles incident to one endpoint of the diagonals of $g_1$ and $g_3$ are unsafe triangles. Therefore, $\hat{U}_R^{1}(1)$ and $\hat{U}_R^{1}(3)$ are as shown in the figure. However, that is not the case for $g_2$. In order to define $\hat{U}_R^{1}(2)$, either $g_1$ or $g_3$ needs to have a region assigned to it in the triangles shared between it and $g_2$. In Figure \[fig:cont\], $\hat{U}_R^{2}(1)$ (which is obtained from (\[eq:r\_2\])) is used to define $\hat{U}_R^{1}(2)$, which in turn is used to define $\hat{U}_R^{2}(2)$ using (\[eq:r\_2\]). Once that all the guards in $\mathbb{G}(T_j)$ have been assigned to a region $R_j^{\alpha}(i)$, it is possible to determine if no allocation was found for the given $r$. By definition, a region $R_j^{1}(i)$ inside $T_j$ ensures that every location within the triangle is assigned to a guard in $\mathbb{G}(T_j)$. However, that may not hold if all the regions allocated to the guards in $T_j$ are $R_j^{2}(i)$. Every $R_j^{2}(i)$ is constructed using (\[eq:r\_2\]), thereby, guaranteeing that $g_i$ is assigned to the largest possible region within $T_j$ such that $c(\mathbb{R}^{alloc}(g_i)) \geq r$. Otherwise, there is no guarantee to keep track of the intruder if it follows the shortest path between $\hat{U}_R^{1}(i)$ and $\hat{U}_R^{2}(i)$. Clearly, an allocation cannot be found by the above process if there is a region inside the triangle that is not assigned to any guard in $\mathbb{G}(T_j)$ after every $R_j^{2}(i)$ is defined for $T_j$. In Figure \[fig:cont\], there is a region $R_{\emptyset}$ inside the regular triangles shared by $g_2$ and $g_3$ that cannot be assigned to those guards. The existence of a region $R_{\emptyset} \neq \emptyset$ implies that an allocation that guarantees successful tracking was not found. The blue shaded triangles correspond to safe triangles, which by definition are covered all the time.
![Three guards deployed in a polygon.[]{data-label="fig:cont"}](continuous_example){width="0.62\linewidth" height="0.43\linewidth"}
Guard Allocation {#subsec:disc}
----------------
In this section, we present an allocation algorithm that respects the constraints in Problem 2. Algorithm \[alg:allocation\] presents the pseudocode of the allocation technique. Given a triangulation graph $G$ and the speed ratio $r$ as the input, Algorithm \[alg:allocation\] provides a partition of the polygon $P$, and an allocation of the guards to the partitions. The initialization, update and termination step of Algorithm \[alg:allocation\] are as follows:
- Initialization:
- Initialize two queues: $\mathbb{G}_{ready}=\mathbb{G}_{alloc}= \emptyset$.
- If there is a $g_i \in \mathbb{G}$ such that $h_i$ has an endpoint $v_{\alpha}(i)$ at which all the non-safe triangles incident to it are unsafe triangles, then $g_i$ is added to $\mathbb{G}_{ready}$.
- If no guard satisfies the previous criteria, $\mathbb{G}_{ready}=\emptyset$. A guard $g_i$ is selected using Algorithm \[alg:arbitrary\], and all the triangles incident to one of its endpoints are assigned to it.
- Update:
- If there is a guard $g_i \notin \mathbb{G}_{ready} \cup \mathbb{G}_{alloc}$ such that for each $T_j \in \overline{\mathbb{T}_{\alpha}^{safe}}(g_i)$, the regions $\overline{R}_j^2(k)$ are already defined and allocated for all $g_k \in \mathbb{G}(T_j) \backslash \{ g_i \}$. $v_{\alpha}(i)$ is called the [*preferential endpoint*]{} of $g_i$, and is labeled as $v_{1}(i)$.
- A guard $g_i \in \mathbb{G}_{ready}$ is selected. Algorithm \[alg:orient\] is used to compute $\hat{U}_R^{1}(i)$ and $\hat{U}_R^{2}(i)$.
- If $\mathbb{G}_{ready}= \emptyset$, a guard is arbitrarily selected using Algorithm \[alg:arbitrary\]. Thus, Algorithm \[alg:allocation\] can continue as in the previous step.
- Termination: The algorithm can terminate in two ways described as follows.
- Algorithm \[alg:orient\] finds a region $R_{\emptyset}(j) \neq \emptyset$ inside a non-safe triangle $T_j$. Since the region $R_{\emptyset}(j)$ cannot be assigned to any guard, the algorithm terminates. Therefore, no allocation is found.
- Each $T_j \in \overline{\mathbb{T}^{safe}}(G)$ is partitioned, and each partition is assigned to a guard. In this case a feasible allocation is found.
Figure \[fig:ex\_allocation1\] (a) shows a polygonal environment with $21$ ($=n$) edges and $4$ ($\leq\lfloor n/4 \rfloor$) guards. There are $8$ safe and $11$ non-safe triangles. The corresponding graph $G^{\#}$ is shown in Figure \[fig:ex\_allocation1\] (c). The orientation of the edges of $G^{\#}$ illustrates a partial order in which the vertices of $G^{\#}$ are allocated by the algorithm. $\mathbb{V}(G^{\#})$ has $11$ vertices, each one associated with an non-safe triangle. Edges exist between vertices that correspond to triangles incident to opposite endpoints of the diagonal of a guard. For example, $T_9$ and $T_{11}$ are incident to one endpoint of $h_4$ and $T_{10}$ is incident to the other. Hence, edges in $G^{\#}$ that correspond to $g_4$ are $e_{10,9}(g_4)$ and $e_{10,11}(g_4)$. In Algorithm \[alg:allocation\], once a $g_i \in \mathbb{G}_{ready}$ is chosen (and $\mathbb{R}^{alloc}(g_i)$ is obtained by Algorithm \[alg:orient\]), the endpoints of its diagonal can be labeled as $v_1(i)$ and $v_2(i)$.
In the example shown in Figure \[fig:ex\_allocation1\], when the algorithm starts, $\mathbb{G}_{ready}= \{ g_1,g_4 \}$. $g_1$ meets the definition of the $\mathbb{G}_{ready}$ queue since all the non-safe triangles incident to one endpoint of $h_1$ are unsafe triangles ($T_1$ and $T_2$). The same is true for one endpoint of $h_4$ ($T_{10}$ is an unsafe triangle). Next, $g_1$ is selected and its corresponding preferential endpoint is named $v_1(1)$. Algorithm \[alg:orient\] is called. It uses the procedure to compute $\mathbb{R}^{alloc}(g_i)$ from Section \[subsec:cont\] for a given $g_i$. Thus, $R_1^1(1)=T_1$ and $R_2^1(1)=T_2$, $S_R^1(1)= \{ T_1,T_2 \}$ and $\hat{U}_R^1(1)=T_1 \cup T_2$. Edges $e_{1,3}(1)$, $e_{1,4}(1)$, $e_{1,5}(1)$ and $e_{1,8}(1)$ become outgoing edges of $v_1$, and edges $e_{2,3}(1)$, $e_{2,4}(1)$, $e_{2,5}(1)$ and $e_{2,8}(1)$ become outgoing edges of $v_2$. Next, the regions that correspond to the endpoint $v_2(1)$ are computed. They are $R_3^2(1)=T_3$, $R_5^2(1)=T_5$, $R_8^2(1)=T_8$ and $R_4^2(1)$ which is the region inside $T_4$ assigned to $g_1$ shown in Figure \[fig:ex\_allocation1\] (b). Finally, the algorithm searches for the existence of any region $R_{\emptyset}(j) \neq \emptyset$ with $j \in \{3,4,5,8\}$. Since it is not found, the process returns to Algorithm \[alg:allocation\] where $\mathbb{G}_{alloc}$ (which contains all guards that have been allocated) and $\mathbb{G}_{ready}$ are updated to $\mathbb{G}_{alloc}=\{g_1\}$ and $\mathbb{G}_{ready}=\{g_4,g_2\}$. In the second iteration, $g_4$ is selected, and the same process is repeated. $R_{10}^1(4)=\hat{U}_R^1(4)=T_{10}$. Edges $e_{10,9}(4)$ and $e_{10,11}(4)$ become outgoing edges of edges of $v_{10}$. $R_{11}^2(4)=T_{11}$ and $R_9^2(4)$ (the region inside $T_9$ assigned to $g_4$ shown in Figure \[fig:ex\_allocation1\] (b)) are obtained. At the end of the iteration, $\mathbb{G}_{alloc}=\{g_1,g_4\}$ and $\mathbb{G}_{ready}=\{g_2,g_3\}$. The algorithm continues until the end of the fourth iteration at which point $\mathbb{G}_{alloc}=\{g_1,g_4,g_2,g_3\}$ and $\mathbb{G}_{ready}= \mathbb{G}_{\Omega} (=\mathbb{G} \backslash (\mathbb{G}_{ready} \cup \mathbb{G}_{alloc}))= \emptyset$. The allocation of all the regions of the environment is shown in Figure \[fig:ex\_allocation1\] (b), and the resulting directed graph $G^{\#}$ is illustrated in Figure \[fig:ex\_allocation1\] (c). Notice that Line $15$ of Algorithm \[alg:orient\] is only reached when all the edges incident to a vertex $v_j$ in $G^{\#}$ are incoming edges. Thus, the regions $\hat{U}_R^1(g_i)$ and $\hat{U}_R^2(g_i)$ for each guard $g_i \in \mathbb{G}(T_j)$ are already defined. Therefore, the algorithm checks whether the triangle is completely covered by the regions allocated to the guards or there is a region inside the triangle that cannot be assigned to guards. In the example, the aforementioned check is performed for triangles $T_8$, $T_7$, $T_3$, $T_{11}$, $T_6$ and $T_5$. We can see that for all the other vertices, there are outgoing edges incident to them, and those edges are associated with a single guard. According to the definition of $\mathbb{G}_{ready}$, at each iteration Algorithm \[alg:allocation\] selects a guard $g_i$ such that the regions of the other guards that can cover the triangles in $\overline{\mathbb{T}_{1}^{safe}}(g_i)$ are already defined. It follows that Algorithm \[alg:arbitrary\] selects the vertices that correspond to the triangles in $\overline{\mathbb{T}_{1}^{safe}}(g_i)$ and orients all the edges incident to them that correspond to $g_i$ as outgoing edges of those vertices. Since all the other guards that can cover those triangles had their regions defined, it means that Algorithm \[alg:arbitrary\] was called before to orient their edges so all of those edges are outgoing edges of other vertices. Hence, they are incoming edges of the vertices associated with the triangles in $\overline{\mathbb{T}_{1}^{safe}}(g_i)$. This explains why either all the edges incident to a vertex in $G^{\#}$ become incoming edges, or only the edges associated to one guard become outgoing edges.
[0.9]{} {width="0.75\linewidth" height="0.55\linewidth"}
[0.55]{} {width="0.84\linewidth" height="0.55\linewidth"}
[0.3]{} {width="0.8\linewidth" height="1.6\linewidth"}
\[fig:ex\_allocation1\]
**Input**:**$P$,$G$,$r$ and $\mathbb{G}$.** **Output**: Triangles of the triangulation of $P$ partitioned into regions assigned to the guards.
**$\mathbb{V}(G^{\#}) \leftarrow \{ v_j : \exists T_j \in \overline{\mathbb{T}^{safe}}(G) \}$** **$\mathbb{E}(G^{\#}) \leftarrow \{ e_{j,k}(g_i) : \exists g_i \in \mathbb{G} \mbox{ such that } T_j \in \overline{\mathbb{T}_1^{safe}}(g_i) \mbox{ and } T_k \in \overline{\mathbb{T}_2^{safe}}(g_i) \}$** **update $\mathbb{G}_{ready}$** choose $g_i \in \mathbb{G}_{ready}$ $v_1(i) \leftarrow$ preferential endpoint of $h_i$ $\overline{\mathbb{T}_{1}^{safe}}(g_i)$ is obtained $stop \leftarrow$ call Algorithm \[alg:orient\] the allocation is not possible update $\mathbb{G}_{ready}$ and $\mathbb{G}_{alloc}$ $\mathbb{G}_{\Omega} \leftarrow \mathbb{G} \backslash (\mathbb{G}_{ready} \cup \mathbb{G}_{alloc})$ call Algorithm \[alg:arbitrary\] update $\mathbb{G}_{ready}$ and $\mathbb{G}_{alloc}$ go to $3$
\[alg:allocation\]
**Input**: $g_i$,$\overline{\mathbb{T}_{1}^{safe}}(g_i)$,$G^{\#}$ and $\mathbb{G}$. **Output**: Regions $\hat{U}_R^1(i)$ and $\hat{U}_R^2(i)$ are defined and assigned to $g_i$ or no feasible allocation found. $\mathbb{E}(i) \leftarrow \emptyset$ $R_j^1(i) \leftarrow \bigcap_{g_k \in \mathbb{G}(T_j) \backslash \{ g_i \} }{\overline{R}_j^2(k)} \cap T_j$ $\mathbb{E}^j(i) \leftarrow $ edges incident to $v_j$ that correspond to $g_i$ orient edges $e_{j,\cdot}(i) \in \mathbb{E}^j(i)$ as outgoing edges of $v_j$ add edges in $\mathbb{E}^j(i)$ to $\mathbb{E}(i)$ compute $S_R^{1}(i)$ and $\hat{U}_R^1(i)$ $v_j \leftarrow$ head of $e$ compute $R_j^2(i)$ using (\[eq:r\_2\]) $R_{\emptyset}(j) \leftarrow \bigcap_{g_k \in \mathbb{G}(T_j)}{\overline{R}_j^2(k)} \cap T_j$ True compute $S_R^{2}(i)$ and $\hat{U}_R^2(i)$ False
\[alg:orient\]
If $\mathbb{G}_{ready}=\emptyset$ and $\mathbb{G}_{\Omega}\neq \emptyset$ (Line $18$), Algorithm \[alg:allocation\] calls Algorithm \[alg:arbitrary\] to allocate the guards in $\mathbb{G}_{\Omega}$. It happens when the iterative procedure cannot find a unique allocation for the non-safe triangles that can be covered by the guards in the set $\mathbb{G}_{\Omega}$. Let $\mathbb{T}_{\Omega}$ denote the triangles that can be covered by the guards in $\mathbb{G}_{\Omega}$ such that they contain regions that have not yet been assigned. Let $\mathbb{V}_{\Omega}$ be the set of vertices in $\mathbb{V}(G^{\#})$ that correspond to the triangles in $\mathbb{T}_{\Omega}$. Triangles in $\mathbb{T}_{\Omega}$ contain regions that can be assigned to more than one guard, and hence, it is not possible to determine a unique partition. Thus, for each $v_j \in \mathbb{V}_{\Omega}$ there are at least two guards with non-oriented edges incident to $v_j$.
[0.64]{} {width="0.9\linewidth" height="0.65\linewidth"}
[0.59]{} {width="1\linewidth" height="0.8\linewidth"}
[0.35]{} {width="1\linewidth" height="1\linewidth"}
\[fig:ex\_allocation2\]
Figure \[fig:ex\_allocation2\] shows such a scenario involving $2$ guards and $6$ non-safe triangles. If Algorithm \[alg:orient\] was executed, it could find $R_3^1(2)=T_3$ or $R_6^1(2)=T_6$, or also $R_1^1(1)=T_1$ or $R_5^1(1)=T_5$. The presence of the regular triangles $T_2$ and $T_4$ prevents Algorithm \[alg:orient\] to find $\hat{U}_R^1(1)$ and $\hat{U}_R^1(2)$. There is no region that has initially been assigned to any of the guards inside $T_2$ and $T_4$ which can be used to construct the region of the other guard. Therefore, $\mathbb{G}_{ready}= \emptyset$ and $\mathbb{V}_{\Omega}=\{v_2,v_4\}$. Let $G_{\Omega}$ be the subgraph induced by $\mathbb{V}_{\Omega}$. We know that there are at least two guards with non-oriented edges incident to each $v_j \in \mathbb{V}_{\Omega}$. Since the number of vertices of $G_1$ is finite ($G_{\Omega}$ cannot be a tree with an infinite length path), each $v_j \in \mathbb{V}_{\Omega}$ is inside a cycle involving the diagonal of guards in $\mathbb{G}_{\Omega}$. Otherwise, if there is a $v_j \in \mathbb{V}_{\Omega}$ with only one neighbor in $\mathbb{V}_{\Omega}$, there is only one guard with a non-oriented edge incident to it which implies $v_j \notin \mathbb{V}_c$. In Figure \[fig:ex\_allocation2\] (c), $G_{\Omega}$ is the graph induced by $v_2$ and $v_4$. It is a cycle of length $2$ with edges $e_{2,4}(g_1)$ and $e_{2,4}(g_2)$.
Algorithm \[alg:arbitrary\] resolves the aforementioned deadlock in $G_{\Omega}$. It arbitrarily selects a $g_i \in \mathbb{G}_{\Omega}$ and an endpoint of $h_i$. Subsequently, it assigns to $g_i$ all the unassigned regions that it can cover from the selected endpoint. Thus, the edges of the other guards in $\mathbb{G}_{\Omega}$ that can cover those regions are deleted from $\mathbb{E}(G^{\#})$. Consequently, $g_i$ meets the requirement to get added to $\mathbb{G}_{ready}$ which in turn allows Algorithm \[alg:allocation\] to continue. In the example of Figure \[fig:ex\_allocation2\], Algorithm \[alg:arbitrary\] arbitrarily selects $g_1$ and the endpoint that is a vertex of $T_4$ is chosen as $v_1(1)$. Thus, the edges $e_{4,2}(2)$ and $e_{4,3}(2)$ are deleted from $G^{\#}$, and $\mathbb{G}_{ready}=\{g_1\}$, so Algorithm \[alg:allocation\] can continue. Next $g_1$ is chosen, and Algorithm \[alg:orient\] finds $R_5^1(1)=T_5$, $R_4^1(1)=T_4$, $S_R^1(1)= \{ T_5,T_4 \}$ and $\hat{U}_R^1(1)=T_5 \cup T_4$. Thus, edges $e_{5,1}(1)$ and $e_{5,2}(1)$ become outgoing edges of $v_5$, and $e_{4,1}(1)$ and $e_{4,2}(1)$ become outgoing edges of $v_4$. It follows that $R_1^2(1)=T_1$ and $R_2^2(1)$ is the region inside $T_2$ assigned to $g_1$ shown in Figure \[fig:ex\_allocation2\] (b). At the end of this iteration, $\mathbb{G}_{alloc}=\{g_1\}$ and $\mathbb{G}_{ready}=\{g_2\}$. At the end of the second iteration all the regions of the environment have been assigned as shown in Figure \[fig:ex\_allocation2\] (b). The resulting directed graph $G^{\#}$ is illustrated in Figure \[fig:ex\_allocation2\] (c) in which the edges deleted by Algorithm \[alg:arbitrary\] are shown as dotted segments.
Since the number of edges in $G^{\#}$ that can be associated to a given guard is upper bounded by $n(n-1)/2$, the time complexity of Algorithm \[alg:orient\] is $O(n^2)$. Also, since the total number of edges that can be deleted is upper bounded by $O(n^2)$, the time complexity of Algorithm \[alg:arbitrary\] is $O(n^2)$. At each iteration of Algorithm \[alg:allocation\], one guard enters $\mathbb{G}_{alloc}$ and Algorithm \[alg:orient\] is called. It follows that the time complexity of Algorithm \[alg:allocation\] is $O(n^3)$.
choose $g_i \in \mathbb{G}_{\Omega} $ $v_1(i) \leftarrow$ arbitrary endpoint of $h_i$ $\mathbb{G}^{del}(T_k) \leftarrow (\mathbb{G}(T_k) \cap \mathbb{G}_{\Omega}) \backslash \{ g_i \})$ delete from $G^{\#}$ all edges incident to $v_k$ that correspond to $g_j$
\[alg:arbitrary\]
Completeness and Correctness of GENALLOC {#subsec:comp}
----------------------------------------
The following lemma proves that Algorithm \[alg:allocation\] terminates.
\[lemma:5\] Algorithm \[alg:allocation\] terminates in a finite number of steps.
Each iteration of Algorithm \[alg:allocation\] computes the region to be allocated to an unassigned guard. Since the number of guards is finite, Algorithm \[alg:allocation\] terminates in finite number of steps.
The next lemma states a condition under which Algorithm \[alg:allocation\] is complete.
\[lemma:6\] If $G^{\#}$ is a forest, Algorithm \[alg:allocation\] is complete.
The proof is by contradiction. Consider a polygonal environment $P$ for which $G^{\#}$ is a forest, and Algorithm \[alg:allocation\] fails to find a feasible allocation of guards even though one exists. From hereon, any variable with a symbol $\tilde{}$ on top of it is associated with the feasible allocation. For example, $\tilde{R}_j^2(i)$ denotes a region in $T_j$ assigned to $g_i \in \mathbb{G}(T_j)$ based on the feasible allocation. Since Algorithm \[alg:arbitrary\] cannot find a feasible allocation in $P$, it terminates at Line $15$. Hence, there is a vertex $v_j \in \mathbb{V}(G^{\#})$ such that all the edges incident to $v_j$ are incoming edges, and the triangle $T_j \in \overline{\mathbb{T}^{safe}}$ has a region $R_{\emptyset}(j) \neq \emptyset$ which by definition cannot be assigned to any guard by Algorithm \[alg:arbitrary\]. Since a feasible allocation exists, every point inside $R_{\emptyset}(j)$ can be assigned to a guard in $\mathbb{G}(T_j)$. Let $R_{\emptyset}^i(j)\subseteq R_{\emptyset}(j)$ denote the region that can be covered by a guard $g_i\in\mathbb{G}(T_j)$ in the feasible allocation. Therefore, $R_{\emptyset}^i(j)$ exists such that $R_{\emptyset}^i(j) \subseteq \tilde{R}_j^2(i)$ and $R_{\emptyset}^i(j) \not\subseteq {R}_j^2(i)$. For the feasible allocation, $d(\tilde{\hat{U}}_R^1(i),\tilde{R}_j^2(i)) \geq d_I^i$. Since $R_{\emptyset}(j) \cap \tilde{R}_j^2(i) \neq \emptyset$, $d(\hat{U}_R^1(i),\tilde{R}_j^2(i)) < d_I^i$. By definition, $d(\tilde{\hat{U}}_R^1(i),\tilde{R}_j^2(i)) \geq d_I^i$. Therefore, a region $R_{\emptyset}(k)\subset T_k \in \overline{\mathbb{T}_{1}^{safe}}(g_i)$ exists such that $R_{\emptyset}(k)\subset R^1_j(i)$ and $R_{\emptyset}(k)\not\subset \tilde{R}^1_j(i)$. Therefore, $R_{\emptyset}(k)\subset \bigcup_{g_a \in \mathbb{G}(T_k) \backslash \{ g_i \}}\tilde{R}^2_k(a)$. Hence, there exists a guard $g_a \in \mathbb{G}(T_k) \backslash \{ g_i \}$ and a region $R_{\emptyset}^a(k) \subseteq R_{\emptyset}(k)$ such that $R_{\emptyset}^a(k) \subseteq \tilde{R}_k^2(a)$ and $R_{\emptyset}^a(k) \not\subseteq {R}_k^2(a)$. Therefore, we can find a sequence $v_{i_1}\xleftarrow{g_{j_1}}v_{i_2}\xleftarrow{g_{j_2}}\cdots$ of vertices and guards in $G^{\#}$ such that $g_{j_k} \in \mathbb{G}(T_{i_k}) \backslash \{ g_{j_{k-1}} \}$, and corresponding regions $R_{\emptyset}^{j_k}(i_k) \subseteq R_{\emptyset}(i_k)$ such that $R_{\emptyset}^{j_k}(i_k) \subseteq \tilde{R}_{i_k}^2(j_k)$ and $R_{\emptyset}^{j_k}(i_k) \not\subseteq {R}_{i_k}^2(j_k)$. The sequence terminates if $\{g_{j_{k-1}}\} = \mathbb{G}(T_{i_k})$ in which case $T_{i_k}$ is an unsafe triangle. Since $j_k$ does not exist $\tilde{R}_{i_k}^2(j_k)=\emptyset\Rightarrow R_{\emptyset}^{j_k}(i_k)=\emptyset$. Since $R_{\emptyset}(i_k)=R_{\emptyset}^{j_k}(i_k)$ and $R_{\emptyset}(i_k)\neq \emptyset$, we arrive at a contradiction. If the sequence does not terminate, then there exist $i_k$ and $i_j$ such that $i_k=i_j$ since the number of vertices in $G^{\#}$ are finite. This implies the existence of a cycle in $G^{\#}$. Therefore, we arrive at a contradiction since $G^{\#}$ is a forest.
In the appendix (Lemma \[lemma:7\]), we prove a more general result which shows that Algorithm \[alg:allocation\] is complete if Algorithm \[alg:arbitrary\] is never called during execution. Figure \[fig:simple\_env\] shows an example in which $G^{\#}$ does not contain cycles. In Figure \[fig:simple\_env\] (a), a simple polygonal environment is shown along with the deployment of guards. There are four non-safe triangles $T_1$, $T_2$, $T_3$ and $T_4$. $T_1$ and $T_2$ can be covered by $g_1$. $T_3$ and $T_4$ can be covered by $g_2$. In Figure \[fig:simple\_env\] (b), the corresponding graph $G^{\#}$ is shown. $G^{\#}$ is a forest that consists of two paths. Lemma \[lemma:6\] states that Algorithm \[alg:allocation\] will always find the feasible allocation for this environment.
[0.6]{} {width="0.92\linewidth" height="0.8\linewidth"}
[0.34]{} {width="0.9\linewidth" height="0.69\linewidth"}
\[fig:simple\_env\]
Classification of Guards {#subsec:class}
------------------------
In this section, we present a classification of the guards. It is based on the regions $\hat{U}_R^{\alpha}(i)$ $\alpha=\{1,2\}$ constructed from Algorithm \[alg:allocation\].
1. Type $0$ guard: These are guards for which either $\hat{U}_R^1(i) = \emptyset$ or $\hat{U}_R^2(i) = \emptyset$. Since the region allocated to a type $0$ guard can be covered from one endpoint of its diagonal, it is a static guard.
2. Type $1$ guard: There are guards for which all the non-safe triangles allocated to the guard incident to one endpoint are unsafe triangles. Notice that each edge in $\mathbb{E}(G^{\#})$ that corresponds to a type $1$ guard is an outgoing edge of a vertex $v_j$ such that $|\mathbb{G}(T_j)|=1$. Clearly, if $\mathbb{G}(T_j) \backslash \{ g_i \} = \emptyset$ for each $T_j \in \overline{\mathbb{T}_{1}^{safe}}(g_i)$, then $g_i$ is a type $1$ guard.
3. Type $2$ guard: Any guard which is neither Type $0$ nor Type $1$ is a Type 2 guard.
Consider the example shown in Figure \[fig:type\_2\_guard\]. Assume $T_1$, $T_2$ and $T_3$ are the only non-safe triangles incident to $v_1(1)$, $v_1(2)$ and $v_1(3)$ respectively. Algorithm \[alg:allocation\] selects the non-safe triangle $T_2$ to be assigned to the only guard that can cover it ($g_2$). Since $|\mathbb{G}(T_2)|=1$, $\hat{U}_R^{1}(2)=T_2$ $\Rightarrow$ $g_2$ is a type $1$ guard. Once $\hat{U}_R^{1}(1)\in T_1$ is computed after a few steps, Algorithm \[alg:allocation\] selects the non-safe triangle $T_3$ and allocates the unshaded region to $g_3$. Besides $g_3$, $T_3$ can also be covered by $g_1$ and $g_2$. Regions $\hat{U}_R^{1}(1)$ and $\hat{U}_R^{1}(2)$ are known. Consequently, $R_{3}^{2}(1)$ and $R_{3}^{2}(2)$ are also computed. It follows that the unshaded region in $T_3$ is labeled as $R_{3}^{1}(3)$. Moreover, since $\overline{\mathbb{T}_{1}^{safe}}(g_3)=\{T_3\}$, $\hat{U}_R^{1}(3)=R_{3}^{1}(3)$. It follows that $g_3$ is an example of a type $2$ guard.
![Example of a type $1$ guard $g_2$ and a type $2$ guard $g_3$.[]{data-label="fig:type_2_guard"}](fig_1_prop){width="0.85\linewidth" height="0.45\linewidth"}
In Algorithm \[alg:arbitrary\] an unassigned guard $g_i \in \mathbb{G}$ is arbitrarily chosen to cover the unassigned regions inside the triangles incident to one of the endpoints of its diagonal. $g_i$ is arbitrarily chosen since Algorithm \[alg:allocation\] failed to find a unique partition of such regions. It implies that $\hat{U}_R^{1}(i)$ cannot be constructed unlike in the case of a type $0$, a type $1$ nor a type $2$ guard. However, the arbitrary allocation in Algorithm \[alg:arbitrary\] assigns those regions to $g_i$. As a result, $g_i$ is converted to a type $1$ or a type $2$ guard.
Motion strategy for the Guards {#subsec:react}
------------------------------
In this section, we present a motion strategy for the guards to move on their diagonals. We introduce the concept of [*critical curves*]{} to propose activation strategies for type $1$ and type $2$ guards. From the discussion in the previous section, $\hat{U}_R^1(i)$ is the region assigned to a type $1$ or type $2$ guard $g_i$. Therefore, it is the responsibility of $g_i$ to cover the triangles incident to $v_1(i)$ when the intruder lies in $\hat{U}_R^1(i)$.
We define an [*internal critical curve*]{}, denoted by $s_{int}^1(i)$, as the boundary of $\hat{U}_R^1(i)$. Corresponding to an internal critical curve, we define an [*external critical curve*]{} as follows:
$$\label{eq:ext}
s_{ext}^1(i)=\{p \in P \backslash \hat{U}_R^1(i) : d(p,s_{int}^1(i)) = d_I^i \}\\
$$
Comparing the definition of $s_{ext}^1(i)$ to the definition of $R_j^2(i)$ (see Equation (\[eq:r\_2\])), we can conclude that a part of the boundary of $R_j^2$ can belong to $s_{ext}^1(i)$. We define a *critical region* associated with the guard $g_i$ as follows: $$\label{eq:critical}
C_1(i)=\{p \in P \backslash \mathring{\hat{U}}_R^1(i) : d(p,s_{int}^1(i)) \leq d_I^i \},$$ where $\mathring{\hat{U}}_R^1(i)$ is the interior of $\hat{U}_R^1(i)$. Note that, by definition, the boundary of $C_1(i)$ contains both curves $s_{int}^1(i)$ and $s_{ext}^1(i)$, and since $d(\hat{U}_R^1(i),\hat{U}_R^2(i)) \geq d_I^i$, it is clear that $(\hat{U}_R^2(i) \cap C_1(i)) \subset s_{ext}^1(i)$.
Figure \[fig:critical\] shows the region $\hat{U}_R^1(i)=R_1^1(i) \cup R_2^1(i) \cup R_3^1(i)$ for the guard $g_i$. The neighboring guards are not shown for the sake of simplicity. $s_{int}^1(i)$, the boundary of $\hat{U}_R^1(i)$, is represented as blue segments and arcs that form the boundary of the regions $R_1^1(i)$, $R_2^1(i)$ and $R_3^1(i)$. Since triangles $T_4$ and $T_5$ are safe triangles (they have $h_i$ as an edge), there is no internal critical curve inside them. The green segments and curves denote $s_{ext}^1(i)$, and the unshaded region inside $P$ is $C_1(i)$. The boundary of $C_1(i)$ is formed by $s_{int}^1(i)$, $s_{ext}^1(i)$ and edges of the environment. The dark colored regions represent $R_6^2(i)$ and $R_7^2(i)$ which are part of $\hat{U}_R^2(i)$.
![Example of the definition of the critical region of guard.[]{data-label="fig:critical"}](criticalreg){width="0.74\linewidth" height="0.5\linewidth"}
For an intruder located in $C_1(i)$, the following equation maps the position of the intruder ($x_I$) to the position of $g_i$ (denoted as $x_{g_i}$) along its diagonal:
$$\label{eq:1}
x_{g_i}= x_{v_1(i)}+ \frac{d(s_{int}^{1}(i),x_I)}{d_I^i}(x_{v_2(i)}-x_{v_1(i)}),$$
where $x_{v_\alpha(i)}$ is the location of vertex $v_{\alpha}(i)$ ($\alpha =\{1,2\}$). If $x_I \in \hat{U}_R^1(i)$, $g_i$ remains static at $v_1(i)$. Otherwise, if $x_I(t) \notin (\hat{U}_R^1(i) \cup C_j(i))$, $g_i$ remains static at $v_2(i)$.
By definition, $d(\hat{U}_R^1(i),\hat{U}_R^2(i)) \geq d_I^i$ while $d(\hat{U}_R^1(i),s_{ext}^1(i))=d_I^i$. Hence, (\[eq:1\]) guarantees that $g_i$ will always cover the regions assigned to it when the intruder is located in them.
Consider the case of a guard $g_i$ such that $g_i$ is incident to $T_k$ at a vertex $v_2(i)$. The motion strategy proposed in (\[eq:1\]) ensures that $\hat{U}_R^1(i)$ is covered if there is an intruder inside $\hat{U}_R^1(i)$. However, if the intruder is located in $\hat{U}_R^1(i) \cup \mathring{C}_1(i)$, where $\mathring{C}_1(i)$ is the interior of $C_1(i)$, $g_i$ cannot cover $T_k$ because $g_i$ can only be located at $v_2(i)$ when the intruder is outside $\hat{U}_R^1(i) \cup \mathring{C}_1(i)$ according to (\[eq:1\]). Now, consider the case when $g_i$ is incident to $T_k$ at a vertex labeled $v_1(i)$. Since $v_1(i)$ is a vertex of $T_k$, $T_k \cap \hat{U}_R^1(i)$ will be covered by $g_i$ if there is an intruder inside it according to (\[eq:1\]). However, if the intruder lies outside $T_k \cap \hat{U}_R^1(i)$, then $T_k$ is not covered by $g_i$ since $g_i$ is not located at $v_1(i)$ according to (\[eq:1\]). Consequently, for each $g_i\in\mathbb{G}(T_k)$, there exists a region which prevents $g_i$ to cover $T_k$ when the intruder lies inside it. This region, denoted by $\hat{C}_1(i)$, is called the *extended critical region*. It is given by the following expression:
$$\hat{C}_1(i)=
\left\{
\begin{array}{cc}
\hat{U}_R^1(i) \cup \mathring{C}_1(i) & v_2(i) \text{ is a vertex of } T_k\\
P \backslash \hat{U}_R^1(i) & \text{otherwise}
\end{array}
\right.,$$
Based on the concept of extended critical regions, Lemma \[lemma:5.10\] presents a necessary and sufficient condition for the guards to cover a non-safe triangle when an intruder lies in it.
\[lemma:5.10\] For the guards in $\mathbb{G}(T_k)$ ($T_k \in \overline{\mathbb{T}^{safe}}(G)$), (\[eq:1\]) guarantees that the triangle $T_k$ is covered when an intruder is located in it if and only if $\displaystyle\bigcap_{g_i \in \mathbb{G}(T)}( \hat{C}_1(i) \cap T_k) = \emptyset$.
($\Rightarrow$) Assume that $\bigcap_{g_i \in \mathbb{G}(T_k)} (\hat{C}_1(i) \cap T_k)= \emptyset$, and $T_k$ is not covered when the intruder lies in $T_k$. It implies that there is a location inside $T_k$ for the intruder that prevents every $g_i\in \mathbb{G}(T_k)$ to cover $T_k$. According to (\[eq:1\]), such a region must belong to $\bigcap_{g_i \in \mathbb{G}(T_k)} (\hat{C}_1(i) \cap T_k)$ which contradicts our assumption.
($\Leftarrow$) Next, assume that $\bigcap_{g_i \in \mathbb{G}(T_k)} (\hat{C}_1(i) \cap T_k) \neq \emptyset$ and $T_k$ is covered when the intruder is located in it. Since $\bigcap_{g_i \in \mathbb{G}(T_k)} \hat{C}_1(i) \cap T_k \neq \emptyset$ when $x_I \in \bigcap_{g_i \in \mathbb{G}(T_k)} \hat{C}_1(i) \cap T_k$, there is no guard covering $T_k$ according to (\[eq:1\]), which is a contradiction. The lemma follows.
Polygons with Holes {#sec:holes}
===================
In this section, we assume that $P$ has polygonal holes which represent obstacles inside the polygon. Let $\mathbb{Q}=\{Q_1,\ldots,Q_N\}$ represent the set of polygonal holes. Let $\hat{n}=n+n_{Q_1}+\ldots+n_{Q_N}$ denote the total number of vertices of $G$, where $n_i$ is the number of vertices of the hole $Q_i$ and $i \in \{ 1,2,\ldots,N \}$. Figure \[fig:simply\] shows a polygonal environment with an internal polygonal hole $Q_1$. In Theorem $5.1$ of [@O'Rourke:1987], it is shown that one can find an internal diagonal of the triangulation of $P$ between any two holes (or between a hole and the outer boundary) which merges two holes (or the hole with the boundary) if a wall of thickness $0$ is placed on the diagonal. This reduces the value of $N$ by $1$. See Figure \[fig:simply\] where $Q_1$ is merged with the outer boundary through the diagonal shared by triangles $T_1$ and $T_2$. Therefore, for any polygon $P$ with $n$ vertices and $N$ internal polygonal holes, we can construct a simply-connected polygon $P'$ with $n+2N$ vertices. We can apply all techniques proposed in the previous sections for deploying guards, and allocating them to triangles of the triangulation of $P'$ for tracking the intruder.
{width="0.80\linewidth" height="0.54\linewidth"}
\[fig:simply\]
Tracking Multiple intruders {#sec:multi}
===========================
In this section, we analyze the performance of the proposed algorithm for multiple intruders. We assume that all the intruders have the same maximum speed $\bar{v}_e$. We use the symbol $\mathbb{I}$ $(|\mathbb{I}| > 1)$ to denote the set of intruders, and the vector $x_{\mathbb{I}}(t)\in\mathbb{R}^{|\mathbb{I}|}$ to denote their positions inside the polygon. We assume that the deployment of the guards and the allocation of the different regions of the environment are obtained using the techniques presented in sections \[subsec:dep\] and \[sec:alloc\], respectively.
In Section \[subsec:react\], a motion strategy for the guards was proposed for a single intruder. (\[eq:1\]) is a reactive motion strategy that depends on the location of the intruder. It ensures that each guard $g_i$ can cover $\hat{U}_R^1(i)$ when the intruder is inside it. In the presence of multiple intruders, the priority for each $g_i$ is to cover $\hat{U}_R^1(i)$ as long as there is an intruder inside it. Hence, the motion strategy only needs to consider the intruder closer to $\hat{U}_R^1(i)$, i.e. if there is an intruder inside $\hat{U}_R^1(i)$, $g_i$ stays at $v_1(i)$ regardless of the positions of other intruders. Therefore, the motion strategy of the guard in this case is given by (\[eq:3\]), wherein $d(s_{int}^{1}(i),x_I)$ is replaced by $d_{min}(s_{int}^{1}(i),x_{\mathbb{I}}(t))$ defined as follows:
$$\label{eq:3}
d_{min}(i,x_{\mathbb{I}}(t))= \min_{I_k\in\mathbb{I} }{d(s_{int}^{1}(i),x_{I_k}(t))}.$$
In Section \[subsec:react\], we showed that an intruder inside each $\hat{C}_1(i)$ associated to $g_i \in \mathbb{G}(T_k)$ will prevent anon-safe triangle $T_k$ from being covered by any guard incident to it. Therefore, $| \mathbb{G}(T_k)|$ intruders are sufficient to keep $T_k$ uncovered. However, if for example, there are two guards $g_i,g_j \in \mathbb{G}(T_k)$ such that $\hat{C}_1(i) \bigcap \hat{C}_1(j)\neq\emptyset$, a single intruder lying inside the intersection will prevent $g_i$ and $g_j$ from covering $T_k$. Therefore, fewer than $|\mathbb{G}(T_k)|$ intruders can prevent $T_k$ from being covered by any guard incident to it if there are non-empty intersections between the extended critical regions corresponding to distinct guards that can cover $T_k$.
Consider the power set $2^{\mathbb{G}(T_k)}$ of all guards incident to $T_k$. Let $\mathcal{S}\subseteq 2^{\mathbb{G}(T_k)}$ be a collection of all sets $S\in 2^{\mathbb{G}(T_k)}$ for which the extended critical regions of the guards belonging to $S$ have a non-empty intersection. The problem of finding the minimum number of intruders that can be placed at the intersection of extended critical regions to uncover $T_k$ is equivalent to the problem of finding the minimum cover $\mathcal{C}\subseteq \mathcal{S}$ of $\mathbb{G}(T_k)$.
Let $n_I(T_k)$ denote the maximum number of intruders that can be tracked by the guards incident to $T_k$ without uncovering $T_k$ when there is an intruder inside it. The following lemma relates $n_I(T_k)$ to $|\mathcal{C}|$.
\[lemma:6.3\] Let $T_k \in \overline{\mathbb{T}^{safe}}(G)$. $$n_I(T_k)=\left\{\begin{array}{cc}|\mathcal{C}|&T_k \cap \mathcal{I}(C) = \emptyset\text{, }\forall C\in \mathcal{C}\\
|\mathcal{C}|-1& \text{otherwise,}
\end{array}\right.$$ where $\mathcal{I}(C)=\displaystyle\bigcup_{g_k \in C}{\hat{C}_1(k)}$
$|\mathcal{C}|$ intruders, each placed in a distinct $\mathcal{I}(C)$, are sufficient to prevent all guards incident to $T_k$ from covering it when an intruder is located inside $T_k$. If $T_k\cap\mathcal{I}(C)=\emptyset\text{ }\forall C\in \mathcal{C}$, the intruder inside the $T_k$ cannot lie inside any $\mathcal{I}(C), C\in\mathcal{C}$. Therefore, $n_I(T_k)=|\mathcal{C}|$ in this case. Otherwise, the intruder located inside $T_k$ can cover an $\mathcal{I}(C)$. Therefore, $n_I(T_k)<|\mathcal{C}|$ in this case.
Since $\mathcal{C}$ is the minimum set cover of $\mathbb{G}$, $|\mathcal{C}|-1$ intruders cannot prevent $T_k$ from being covered by at least one guard $g_{i}\in\mathbb{G}(T_k)$. Therefore, $n_I(T_k)\geq|\mathcal{C}|-1$. The theorem follows.
In Figure \[fig:examplem\] , two type $1$ guards $g_1$ and $g_2$, and a type $2$ guard $g_3$ are shown with their corresponding external and internal critical curves. The corresponding endpoints $v_1(i)$ are shown with a green disc. The regions $\hat{U}_R^1(i)$ are shaded in orange, and the safe triangles are shaded in blue. $T$ is a regular triangle that can be covered by $g_1$, $g_2$ and $g_3$. Therefore, the set $\mathcal{C}$ consists of the external critical regions and has cardinality $3$. Since there is no intersection between any region in $\mathcal{C}$ and $T$ then $n_I(T)=3$. Figure \[fig:examplem\] , shows the same case but for a smaller value of $r$. $s_{ext}^1(1) \cap s_{ext}^1(2) \neq \emptyset$. This implies that $\mathcal{C}$ consists of the extended critical region of $g_3$ and the intersection of the extended critical regions of $g_1$ and $g_2$. In this case, $\mathcal{C}$ consists of $\hat{C}_1(4)$ and $\hat{C}_1(1) \cap \hat{C}_1(2)$, so it has cardinality $2$. Since none of the regions in $\mathcal{C}$ intersects with $T$, then $n_I(T)=2$.
[0.47]{} {width="1\linewidth" height="1\linewidth"}
[0.45]{} {width="1\linewidth" height="1\linewidth"}
\[fig:examplem\]
\[corollary:1\] The minimum number of intruders that can be tracked based on the strategy proposed in (\[eq:3\]) is $n_I^* = \min\{ n_I(T_k) : T_k \in \overline{\mathbb{T}^{safe}}(G) \}$.
Assume that $|\mathbb{I}| > \min\{ n_I(T_k) : T_k \in \overline{\mathbb{T}^{safe}}(G) \}$. It implies that there is at least one $T_k \in \overline{\mathbb{T}^{safe}}(G)$ for which $n_I(T_k) < |\mathbb{I}|$ and therefore, $T_k$ cannot be covered at all times according to Lemma \[lemma:6.3\]$\implies$ $n_I^* \ngtr \min\{ n_I(T_k) : T_k \in \overline{\mathbb{T}^{safe}}(G) \}$. Now assume that $ |\mathbb{I}| \leq \min\{ n_I(T_k) : T_k \in \overline{\mathbb{T}^{safe}}(G) \}$. According to Lemma \[lemma:6.3\], the guards have a strategy to cover every non-safe triangle if $ |\mathbb{I}| \leq \min\{ n_I(T_k) : T_k \in \overline{\mathbb{T}^{safe}}(G) \}$. Therefore, $n_I^* = \min\{ n_I(T_k) : T_k \in \overline{\mathbb{T}^{safe}}(G) \}$.
The set cover problem is NP-complete [@Karp:1972]. Several polynomial time approximation schemes (PTAS) for the set cover problem have been proposed in the literature [@Chvatal:1979; @Gandhi:2006; @Goldschmidt:1993; @Johnson:1974; @Lovasz:1975; @Slavak:1997]. Better approximation ratios can be obtained at the expense of computational complexity slightly higher than a PTAS [@Bougeois:2009]. For example, it has been shown that any $(1 - \alpha \ln {n})$ - approximation algorithm for the set cover problem must run in time at least $2^{n^{c \alpha}}$ for some small constants $0 < c < 1$ [@Moshkovitz:2012]. [@Cygan:2008; @Cygan:2009] present some efforts to tighten the running time by reducing the value of $c$ in $2^{n^{c \alpha}}$. We can use either of the aforementioned approaches to obtain $\mathcal{C}$. For a problem instance of large size, one might prefer a PTAS, whereas a moderately exponential algorithm is more preferable when the number of guards covering each $T_k$ is small enough.
Conclusion {#sec:conclusion}
==========
In this work, we addressed the problem of tracking mobile intruders in a polygonal environment using a team of diagonal guards. Leveraging on deployment strategies for mobile coverage in art gallery problems, we proposed control and coordination strategies for the guards to track intruders inside a polygonal environment. At first, we formulated the tracking problem as a multi-robot task assignment problem [**on the triangulation graph of a polygon**]{}. We classified the guards based on their position with respect to the triangles of the triangulation. Next, we showed that the problem of finding the minimum speed of the guards to cover the triangles of the triangulation under the constraint that each triangle can only be covered by a single guard is NP-hard. Given the maximum speed of the intruder, we proposed an algorithm to find a feasible allocation of guards to the triangles of the triangulation when multiple guards are allowed to cover the triangle. We proved the correctness of the proposed algorithm, and its completeness for a specific set of inputs. Based on the task allocated to a guard, we proposed control laws for the guards to move along their diagonals. Finally, we extended the algorithm to address deployment and allocation strategies for non-simple polygons and multiple intruders.
We believe that our paper is a first step towards MRS deployment for [**persistent tracking with provable guarantees**]{}. An important direction of future research is to address the tracking problem for guards with sensing and motion constraints, for example, edge guards, which are more constrained in their motion, or line guards, which are less constrained than diagonal guards. Another future research direction is to study the tracking problem for special polygons, for example, orthogonal polygons, monotone polygons etc. For these polygons, it has been shown that fewer guards are required for coverage. Finally, the problem of tracking with mixed team of guards (static and mobile) is an interesting direction of future research.
Construction of $\hat{U}^{2}_R(i)$ and $R^{2}_j(i)$ {#sec:appa}
===================================================
Obtaining all the regions $R_j^{2}(k)$ yields the set $S_R^{2}(i)$ and the region $\hat{U}_R^2(i)$.
We claim that in general, the boundary of any region $\hat{U}_R^{1}(i)$, denoted by $\delta (\hat{U}_R^{1}(i))$, consists of arcs of circle. Let $\mathbb{S}(\delta(\hat{U}_R^1(i)))$ be the set of arcs of circle of $\delta(\hat{U}_R^1(i))$. For each $s \in \mathbb{S}(\delta(\hat{U}_R^1(i)))$, we also define $c(s)$ as the center of the circle generating $s$ (center at infinite in the case of a line segment), $rad(s)$ as its radius. We define the *expanded boundary* of $\hat{U}_R^1(i)$ as $\gamma (\hat{U}_R^1(i))=\{ p \in P \backslash \hat{U}_R^1(i): d(p,\delta (\hat{U}_R^1(i)))=d_I^i \}$. Lemma \[lemma:boundary\] shows that $\gamma(\hat{U}_R^1(i))=s_{int}^1(i)$ consists only of arcs of circle grouped in a set $\mathbb{S} (\gamma(\hat{U}_R^1(i)))$, and Lemma \[lemma:boundary2\] shows that the boundary of $\hat{U}_R^{1}(i)$ and $\hat{U}_R^{2}(i)$ consist of arcs of circle.
It is important to make this remark since most of the computational geometry libraries include segments and circle arcs as basic classes, which are required to build regions $\hat{U}_R^{1}(i)$ and $\hat{U}_R^{2}(i)$. Some computational geometry libraries such as CGAL [@Wein:2007] and [@Mehlhorn:1999], include the implementation of approximation techniques to compute offset curves of polygons. For the case of offsets of polylines there are some approximation algorithms [@Liu:2007; @Choi:1999; @Jian:2001] which may be implemented using the aforementioned libraries with their line segment and circle classes. For this paper we used the LEDA $6.5$ library in the simulations.
\[lemma:boundary\] Given $\mathbb{S}(\delta(\hat{U}_R^1(i)))$, its corresponding $\gamma(\hat{U}_R^1(i))$ consists of arcs of circle.
Trivially, if there are no obstacles between $\delta(\hat{U}_R^1(i))$ and $\gamma(\hat{U}_R^1(i))$, $\gamma (\hat{U}_R^1(i))=\{ p \in P \backslash \hat{U}_R^1(i): d(p,\delta (\hat{U}_R^1(i)))=d_I^i \}$ is the offset of $\delta(\hat{U}_R^1(i))$ (with $d_I^i$ as the offset distance). Therefore, $\gamma (\hat{U}_R^1(i))$ must be a polyline curve [@Liu:2007].
The presence of obstacles between $\delta(\hat{U}_R^1(i))$ and $\gamma(\hat{U}_R^1(i))$ implies that the shortest path from some points in $\gamma(\hat{U}_R^1(i))$ to $\delta(\hat{U}_R^1(i))$ is a chain of connected line segments instead of a line segment, as in the case where there are no obstacles between $\gamma(\hat{U}_R^1(i))$ and $\delta(\hat{U}_R^1(i))$. In Figure \[fig:bound\] a region $\hat{U}_R^1(i)$ is shown as an orange triangle, its corresponding $\delta(\hat{U}_R^1(i))$ is a black dotted segment, $\gamma(\hat{U}_R^1(i))$ is represented as a dotted curve divided into four arcs of circle. $s_1 \in \mathbb{S}(\gamma(\hat{U}_R^1(i)))$ illustrates the case of points in $\gamma(\hat{U}_R^1(i))$ such that the shortest path between them and $\delta(\hat{U}_R^1(i))$ is a line segment. Now consider $s_2 \in \mathbb{S}(\gamma(\hat{U}_R^1(i)))$ the presence of an obstacle implies that for all the points in $s_2$, such as the one illustrated as a black circle, the shortest path between $\gamma(\hat{U}_R^1(i))$ and $\delta(\hat{U}_R^1(i))$ consists of two connected line segments, one with endpoints in $s_2$ and vertex $v_1 \in \mathbb{V}(G)$ and other with $v_1$ as an endpoint and the other at $\delta(\hat{U}_R^1(i))$. Also, for the points in $s_3 \in \mathbb{S}(\gamma(\hat{U}_R^1(i)))$, the shortest path between $\gamma(\hat{U}_R^1(i))$ and $\delta(\hat{U}_R^1(i))$ consists of three connected line segments, one from $s_3$ to $v_2$, another from $v_2$ to $v_1$ and the last one from $v_1$ to $\delta(\hat{U}_R^1(i))$. Consider the points $p \in \gamma(\hat{U}_R^1(i))$ for which the shortest path between them and $\delta(\hat{U}_R^1(i))$ is not a line segment due to the presence of reflex vertices of the environment. As we can see in Figure \[fig:bound\], for any of such $p$ points, the shortest path from $p$ to $\delta(\hat{U}_R^1(i))$ must visit first a reflex vertex $v_{p} \in \mathbb{V}(G)$ ($v_1$ or $v_2$ for instance) of the environment. Clearly, the union of such points $p$ for which the first segment of the shortest path between them and $\delta(\hat{U}_R^1(i))$ has $v_{p}$ as an endpoint, is a subset of the union of all points in the plane that are equidistant to $v_{p}$. Thus, they form an arc of circle, which is centered at $v_{p}$ with radius $d_I^i- d(v_{p}, \delta(\hat{U}_R^1(i)))$, where $d(v_{p}, \delta(\hat{U}_R^1(i)))$ is the length of the shortest path between $v_{p}$ and $\delta(\hat{U}_R^1(i))$. Therefore, every point in $\gamma(\hat{U}_R^1(i))$ belongs to an arc of circle, so $\gamma(\hat{U}_R^1(i))$ is the union of a set $\mathbb{S}(\gamma(\hat{U}_R^1(i)))$ of arcs of circle.
![Boundary $\gamma(\hat{U}_R^1(i))$ decomposed in four different arcs of circle.[]{data-label="fig:bound"}](lemmaboundary){width="0.62\linewidth" height="0.43\linewidth"}
\[lemma:boundary2\] Every $\hat{U}_R^1(i)$ and $\hat{U}_R^2(i)$ is bounded by arcs of circle.
Assume that $\hat{U}_R^1(i)$ is bounded by arcs of circle. According to (\[eq:r\_2\]), $R_j^2(i)$ where $T_j \in \overline{\mathbb{T}_{2}^{safe}}(g_i)$, is the intersection of $T_j$ and the complement of the region enclosed by $\delta (\hat{U}_R^{1}(i))$ and $\gamma (\hat{U}_R^{1}(i))$. Hence the boundary of $R_j^2(i)$ consists of arcs of circle in $\mathbb{S}(\gamma(\hat{U}_R^1(i)))$ and the edges of $T_j$ (arcs of circle with center at infinity). Since the boundary of $R_j^2(i)$ consists of arcs of circle, it follows that $\delta (\hat{U}_R^2(i))$ also consists of arcs of circle from the definition of $\hat{U}_R^2(i)$. Consider the base case where $\hat{U}_R^{1}(i)$ is the union of unsafe triangles. $\delta (\hat{U}_R^{1}(i))$ is then a set of line segments, and the lemma holds. Consider the case of an unassigned guard $g_i$ such that all the guards $g_k \in \mathbb{G}(T_j)\backslash \{ g_i \}$ have their regions $\hat{U}_R^1(k)$ and $\hat{U}_R^2(k)$ defined for each $T_j \in \overline{\mathbb{T}_{\alpha}^{safe}}(g_i)$, where $\alpha = 1 \vee 2 $. We assume that the result holds for the regions of those guards $g_k \in \mathbb{G}(T_j)\backslash \{ g_i \}$. Since the regions of those guards are defined, then $\alpha=1$ and each $R_j^{1}(i)$ can be defined. Recall that $R_j^{1}(i) = \bigcap_{g_k \in \mathbb{G}(T_j) \backslash \{ g_i \} }{\overline{R}_j^2(k)} \cap T_j$, where $T_j \in \overline{\mathbb{T}_{1}^{safe}}(g_i)$. Each $R_j^{1}(i)$ is an intersection of regions $\hat{U}_R^{2}(k)$ which are bounded by arcs of circle. Hence, the intersection region is also bounded by arcs of circle, which trivially implies that $\delta (\hat{U}_R^{1}(i))$ also consists of arcs of circle. And according to the first part of this proof, it follows that $\delta (\hat{U}_R^2(i))$ consists of arcs of circle.
Generalization of Lemma \[lemma:6\] {#sec:appb}
===================================
\[lemma:7\] If Algorithm \[alg:arbitrary\] is never called for a specific input, Algorithm \[alg:allocation\] is complete for such an input.
The proof is by contradiction. We assume that Algorithm \[alg:allocation\] does not find a feasible allocation but there exists one. Additionally, we assume that Algorithm \[alg:arbitrary\] is never called during execution. In the proof of Lemma \[lemma:6\], we show the existence of a sequence $v_{i_1}\xleftarrow{g_{j_1}}v_{i_2}\xleftarrow{g_{j_2}}\cdots$ of vertices and guards in $G_1$ such that $g_{j_k} \in \mathbb{G}(T_{i_k}) \backslash \{ g_{j_{k-1}} \}$ and $R_{\emptyset}^{j_k}(i_k) \subseteq R_{\emptyset}(i_k)$ such that $R_{\emptyset}^{j_k}(i_k) \subseteq \tilde{R}_{i_k}^2(j_k)$ and $R_{\emptyset}^{j_k}(i_k) \not\subseteq {R}_{i_k}^2(j_k)$. It is also stated that the sequence terminates if $\{g_{j_{k-1}}\} = \mathbb{G}(T_{i_k})$ in which case $T_{i_k}$ is an unsafe triangle. The problem of finding the existence of an allocation that works when Algorithm \[alg:allocation\] fails is reduced to the problem of showing that the aforementioned sequence of vertices and guards in $G^{\#}$ does not terminate. Since the number of vertices in $G^{\#}$ is finite, the sequence is stuck in a cycle of vertices of $G^{\#}$. According to the definition of $G^{\#}$ there should be a cycle $\overline{C}$ in $G^{\#}$ involving the vertices and the guards of the cycle in the sequence. Now we prove that such a cycle $\overline{C}$ cannot exist unless Algorithm \[alg:arbitrary\] was called. First, we show that for the first vertex of the cycle $\overline{C}$ that appears in the sequence, $v_{i_k}$, the pair of edges in $\overline{C}$ incident to it are both incoming edges. If $v_{i_k}$ is the vertex that corresponds to the triangle where Algorithm \[alg:arbitrary\] determined that it could not find an allocation, the claim is trivially proved since all edges incident to $v_j$ are incoming edges. Otherwise, if $v_{i_k}$ is no such a vertex, then it corresponds to a vertex where all the incident edges are incoming edges excepting the edges that correspond to guard $g_{j_{k-1}}$. Notice that the sequence $v_{i_1}\xleftarrow{g_{j_1}}v_{i_2}\xleftarrow{g_{j_2}}\cdots$ follows a direction opposite to the orientation of the edges. Therefore, when the sequence is in $v_{i_{k}}$, the next guard cannot be $g_{j_{k-1}}$. It implies that both edges of $\overline{C}$ incident to $v_{i_{k}}$ do not correspond to $g_{j_{k-1}}$, so they are by definition incoming edges. The sequence then continues with a different guard $g_{j_{k}}$ followed by a vertex $v_{i_{k+1}}$. By definition, the edge in $\overline{C}$ incident to $v_{i_{k+1}}$ that corresponds to $g_{j_{k}}$ is an outgoing edge, so the next edge in $\overline{C}$ corresponds to a different guard $g_{j_{k+1}}$ and its corresponding edge in $\overline{C}$ is an incoming edge of $v_{i_{k+1}}$. Clearly, every vertex in $\overline{C}$ has an outgoing and an incoming edge. Since $\overline{C}$ is a cycle, vertex $v_{i_{k}}$ is eventually reached. However, it does not have an incoming and an outgoing edge in $\overline{C}$ (both are incoming edges), which is not possible. Therefore, such a cycle does not exist in $G^{\#}$. This contradicts the initial definition of $G^{\#}$. Thus, at least one edge of $\overline{C}$ was removed during the execution of Algorithm \[alg:allocation\]. This implies that Line $8$ of Algorithm \[alg:arbitrary\] was reached, which is impossible since Algorithm \[alg:arbitrary\] was never called. The result follows.
List of Variables {#sec:appc}
=================
[L[1.8cm]{}|L[6cm]{}]{} **Variable** & **Definition**\
$P$ & Polygon\
$n$ & Number of vertices of $P$\
$\bar{v}_e$ & Maximum speed of intruder\
$x_I$ & Location of intruder\
$t$ & Time\
$\mathbb{G}$ & Set of guards\
$\bar{v}_g$ & Maximum speed of guards\
$r$ & $\bar{v}_g/\bar{v}_e$\
$G$ & Triangulation graph\
$\mathbb{V}(G)$ & Vertex set of graph $G$\
$\mathbb{E}(G)$ & Edge set of graph $G$\
$\mathbb{T}(G)$ & Faces of graph $G$\
$\mathbb{H}$ & Set of diagonals\
$g_i$ & Guard $i$\
$l_i$ & Length of $h_i$\
$v_{\alpha}(i)$ & Endpoint $\alpha$ of $h_i$\
$\mathbb{G}(T)$ & Set of guards incident to $T$\
$d(\cdot,\cdot)$ & Distance function\
$A(g_i)$ & Non-safe triangles allocated to $g_i$\
$A(\mathbb{G})$ & Allocation of all non-safe triangles\
$\mathcal{A}$ & Set of all $A(\mathbb{G})$\
$\overline{\mathbb{T}_{\alpha}^{safe}}(g_i)$ & Set of non-safe triangles incident to $v_{\alpha}(i)$\
$\overline{\mathbb{T}^{safe}}(G)$ & Set of non-safe triangles\
$G^{\#}$ & Guard adjacency graph\
$e_{j,k}(g_i)$ & Edge of $G^{\#}$\
$w_{j,k}(g_i)$ & Weight of $e_{j,k}(g_i)$\
$S_{rep}$ & Set of representatives of $\bigcup_{j\in\{1,\ldots, |\mathbb{V}(G^{\#})|\}} S_g(T_j)$\
$R_j^{\alpha}(i)$ & Region inside $T_j$ assigned to $g_i$ incident to $v_{\alpha}(i)$\
$S_R^{\alpha}$ & Set of regions $R_j^{\alpha}(i)$\
$\mathbb{R}^{alloc}(g_i)$ & Region allocated to $g_i$\
$\hat{U}_R^{\alpha}(i)$ & Union of regions $R_j^{\alpha}(i)$\
$c(\mathbb{R}^{alloc}(g_i))$ & Cost of $\mathbb{R}^{alloc}(g_i)$\
$d_I^i$ & $l_i/r$\
$T_k^{free}$ & Region inside $T_k$ that has not been assigned\
$R_{\emptyset}(j)$ & Region inside $T_j$ that cannot be assigned\
$\mathbb{G}_{ready}$ & Set of guards ready to be allocated\
$\mathbb{G}_{alloc}$ & Set of allocated guards\
$\mathbb{G}_{\Omega}$ & Set of guards that cannot be allocated\
$\mathbb{V}_{\Omega}$ & Set of vertices corresponding to non-safe triangles that can be covered by guards in $\mathbb{G}_{\Omega}$\
$s_{int}^1(i)$ & Internal critical curve\
$s_{ext}^1(i)$ & External critical curve\
$C_1(i)$ & Critical region\
$\hat{C}_1(i)$ & Extended critical region\
$\mathbb{I}$ & Set of intruders
[^1]: In this work, we define a triangle to be *covered* if and only if there is a guard located at the boundary of the triangle.
[^2]: We say that a guard $g_i$ is incident to triangle $T$ if at least one of the endpoints of $h_i$ is a vertex of $T$.
[^3]: A visibility graph of a polygon is a graph whose nodes corresponds to the vertices of the polygon, and there is an edge between any two vertices if the segment joining them is contained inside the polygon.
[^4]: A triangulation graph $G$ is said to be *dominated by a set of diagonals* $\mathbb{H} \subset \mathbb{E}(G)$ if at least one vertex of each triangle in $\mathbb{T}(G)$ is an endpoint of a diagonal in $\mathbb{H}$.
[^5]: A set of dominating diagonals is any set of diagonals that dominate a triangulation graph.
[^6]: A triangle is said to be *dominated* if at least one of its vertices is an endpoint of a diagonal in $\mathbb{H}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The orbital boundary value problem, also known as Lambert Problem, is revisited. Building upon Lancaster and Blanchard approach, new relations are revealed and a new variable representing all problem classes, under L-similarity, is used to express the time of flight equation. In the new variable, the time of flight curves have two oblique asymptotes and they mostly appear to be conveniently approximated by piecewise continuous lines. We use and invert such a simple approximation to provide an efficient initial guess to an Householder iterative method that is then able to converge, for the single revolution case, in only two iterations. The resulting algorithm is compared, for single and multiple revolutions, to Gooding’s procedure revealing to be numerically as accurate, while having a significantly smaller computational complexity.'
author:
- Dario Izzo
bibliography:
- 'main.bib'
date: 'Received: date / Accepted: date'
title: 'Revisiting Lambert’s Problem'
---
Introduction {#intro}
============
Lambert’s problem, sometimes referred to as orbital boundary value problem, is a fascinating problem in astrodynamics that intrigued, over the years, most famous mathematicians. Just like Kepler’s equation, its solution is at the very heart of fundamental astrodynamical and space engineering questions [@celmech1; @celmech2; @izzolambert]. Following the fundamental work laid down, among others, by Euler, Lambert, Lagrange and Gauss, the need of having one robust algoritmic procedure able to function for a wide set of conditions led to revisit the Lambert’s problem during the space era. Among the many contributions made during that period, the work of Lancaster and Blanchard [@lancaster] is to be highlighted as it reduced the solution to Lambert’s problem to performing iterations each one requiring the computation of one only inverse trigonometric or hyperbolic function. Later, Gooding [@gooding] built upon these results and published a procedure achieving high precision in only three iterations for all geometries. Gooding’s algorithm makes use of Halley’s iterations sided to well designed heuristics to set the initial guess of the iterated variable. His methodology to reconstruct the terminal velocity vectors is also remarkable as it is purely algebraic. The resulting procedure is extremely efficient having low computational cost and high accuracy. A number of studies [@peterson], [@klumpp] and [@parrish] have tested Gooding approach extensively, suggesting its superiority with respect to other Lambert solvers. His procedure is most accurate and considered as the fastest existing approach to solve Lambert’s problem [@arora]. Aside from Gooding’s algorithm, many other proposal have been put forward to design Lambert solvers, they all differ in the details of at least one of three fundamental ingredients: a) the iteration variable (directly connected to the time of flight equation), b) the iteration algorithm c) the initial guess and d) the reconstruction of the terminal velocity vectors. More recently iimprovements on the original Gooding algorithm were also claimed [@arora] making use of the universal variable formulation [@bate] and an original cosine transformation. At the same time, a number of works recently addressed the possibility of deploying a large number of Lambert’s algorithms on modern GPU architectures [@parrish], [@arora2] and [@wie]. Interestingly, in the first of these works, a comparison is also made between Gooding procedure, a universal variable Lambert’s solver and an early (slow) version of the algorithm here described (unpublished at that time) showing already its promising nature.
In this paper, we build upon Lancaster and Blanchard work, first deriving some new results, and then proposing and testing a new algorithm. The new algorithm a) iterates on the Lancaster-Blanchard variable $x$ using b) a Householder iteration scheme c) feeded by a simple initial guess found exploiting new analytical results found. The resulting procedure is simple to implement, does not make use of heuristics for the initial guess generation and is able to converge, on average, in only 2 iterations for the single revolution case and 3 in the multirevolution case, introducing a significant reduction in the overall solver complexity.
Background {#sec:1}
==========
From Lambert to Gauss {#sec:2}
---------------------
Lambert’s theorem states that the time of flight $t$ to travel along a keplerian orbit from $\mathbf r_1$ to $\mathbf r_2$ is a function of the orbit semi-major axis $a$, of the sum $r_1+r_2$ and of the chord $c$ of the triangle having $\mathbf r_1$ and $\mathbf r_2$ as sides. The complete formal proof was first delivered by Lagrange and is here sketched briefly in the form reported by Battin [@battin] as some of the quantities and equations involved will prove to be useful in our later developments. We start introducing the eccentric anomaly $E$ and the hyperbolic anomaly $H$ via the corresponding Sundmann transformations $rdE = ndt$ $rdH = Ndt$. The mean motion $n = \sqrt{\mu / a^3}$ and its hyperbolic equivalent $N = \sqrt{-\mu / a^3}$ are also introduced. As we do not make use of universal variables we will be forced to give all our arguments twice: one for the elliptic case $a>0$ and one for the hyperbolic case $a<0$. To this purpose some of the equations will be split in two lines, in which case the line above holds for the elliptic case and the line below holds in the hyperbolic case. We also make use of the reduced eccentric anomaly $E_r \in [0,2\pi]$ so that when $\tilde M$ full revolutions are made $E = E_r + 2\tilde M\pi$. To ease the notation, in the following, we will drop the subscript $r$ so that $E$ will be the reduced eccentric anomaly. The following relations are then valid for an elliptic orbit ($a > 0$): $$\label{eq:ellipse}
\begin{array}{c}
r = a(1-e \cos E) \\
nt = E - e \sin E +2\tilde M\pi\\
r \cos f = a (\cos E - e) \\
r \sin f = a \sqrt{(1-e^2)} \sin E
\end{array}$$ The first one relates the orbital radius $r$ to the eccentric anomaly $E$, the second one is the famous Kepler’s equation relating the eccentric anomaly to the time of flight and the following two relation define the relations between true anomaly $f$ and eccentric anomaly $E$. Similar equations hold in the case of hyperbolic motion: $$\label{eq:hyperbola}
\begin{array}{c}
r = a(1-e \cosh H) \\
Nt = e \sinh H - H\\
r \cos f = a (\cosh H - e) \\
r \sin f = -a \sqrt{(e^2-1)} \sinh H
\end{array}$$ The above equations are valid along a Keplerian orbit, including $\mathbf r_1$ and $\mathbf r_2$. The time of flight can thus be written as: $$\begin{array}{c}
\sqrt{\mu}(t_2-t_1)= \left\{ \begin{array}{l} a^{3/2} \left(E_2-E_1 + e\cos E_1 - e\cos E_2+2M\pi\right) \\ -a^{3/2} \left(e\cosh H_2 - e\cosh H_1 - (H_2 - H_1)\right) \end{array} \right.
\end{array}$$ where $M = \tilde M_2 - \tilde M_1$ is the number of complete revolutions made during the transfer from $r_1$ to $r_2$. We may then define two new quantities such that: $$\begin{array}{c}
\psi = \left\{ \begin{array}{l} \frac{E_2-E_1}2 \\ \frac{H_2-H_1}2 \end{array} \right.,
\cos\varphi = \left\{ \begin{array}{l} e\cos\frac{E_2+E_1}2 \\ e\cosh\frac{H_2+H_1}2 \end{array} \right.
\end{array}$$ so that, by construction, in both the elliptic and hyperbolic motion case $\psi\in[0,\pi]$. We also restrict $\varphi\in[0,\pi]$ (elliptc case) and $\varphi\ge0$ (hyperbolic case) as to avoid ambiguity in the definition of the new angle. The time of flight equation is then written as: $$\label{eq:tof_psi}
\sqrt{\mu}(t_2-t_1)= \left\{ \begin{array}{l} 2 a^{3/2} \left(\psi - \cos\varphi\sin\psi + M\pi)\right. \\ -2a^{3/2} \left(\cosh\varphi\sinh\psi - \psi\right) \end{array} \right.$$ The two new quantities introduced, $\varphi$ and $\psi$ only depend on the problem geometry and the sami-major axis $a$ as can be easily found by computing $c^2 = (r_2\cos f_2 - r_1\cos f_1)^2 - (r_2\sin f_2 - r_1\sin f_1)^2$ and $r_1 + r_2$ from Eq.(\[eq:ellipse\]) and Eq.(\[eq:hyperbola\]), holding: $$\label{eq:g1}
r_1+r_2 = \left\{ \begin{array}{l} 2a(1-\cos\psi\cos\varphi) \\ 2a(1-\cosh\psi\cosh\varphi) \end{array} \right.$$ $$\label{eq:g2}
c = \left\{ \begin{array}{l} 2a\sin\psi\sin\varphi \\ -2a\sinh\psi\sinh\varphi \end{array} \right.$$ Thus, one can conclude that the time of flight, given in Eq.(\[eq:tof\_psi\]), is a function of $a$, $c$ and $r_1+r_2$. To further investigate the functional relation of the time of flight to these quantities it is convenient to introduce two new angles: $$\alpha = \varphi+\psi, \beta = \varphi-\psi$$ Clearly, in the elliptic case $\alpha\in[0.2\pi]$ and $\beta\in[-\pi,\pi]$ while for the hyperbolic case $\alpha\ge0$ and $\beta\ge-\pi$. These bounds are very important, as we shall see, in solving a quadrant ambiguity of the newly defined quantities. The time of flight equation now takes the elegant form: $$\label{eq:tof_alpha}
\sqrt{\mu}(t_2-t_1)= \left\{ \begin{array}{l} a^{3/2} \left(\left(\alpha-\sin\alpha)-(\beta-\sin\beta\right) + 2M\pi\right) \\ -2a^{3/2} \left(\left(\sinh\alpha-\alpha\right)-\left(\sinh\beta-\beta\right)\right) \end{array} \right.$$ and computing $r_1+r_2 \pm c$ from Eq.(\[eq:g1\]) and Eq.(\[eq:g2\]) one easily finds: $$\label{eq:g5}
\frac s{2a} = \left\{ \begin{array}{l} \sin^2\frac\alpha 2 \\ -\sinh^2\frac\alpha 2 \end{array} \right.$$ $$\label{eq:g6}
\frac {s-c}{2a} = \left\{ \begin{array}{l} \sin^2\frac\beta 2 \\ -\sinh^2\frac\beta 2 \end{array} \right.$$ These last three equations were first derived by Lagrange and used in his proof of the Lambert’s theorem. The angles $\alpha$ and $\beta$ cannot be determined univoquely from the equations above as their quadrant is not defined. We thus appear to have two possible solutions for $\alpha$ and $\beta$. The quadrant of $\beta$ can actually be resolved by expanding $\cos \theta /2 = \cos (f_2-f_1)/2$ using trigonometric identities and eventaully showing that the following holds: $$\label{eq:cost2}
\sqrt{r_1r_2}\cos\frac\theta 2 =
\left\{
\begin{array}{l}
2a\sin\frac\alpha 2\sin \frac\beta 2 \\
-2a\sinh\frac\alpha 2\sinh \frac\beta 2
\end{array}
\right.$$ since $\sin\frac\alpha 2,\sinh\frac\alpha 2 \ge 0$ the above equations dictate that $\sin\frac \beta 2, \sinh\frac \beta 2$ have the same sign as $\cos\frac\theta 2$, thus $\beta \in [-\pi,0]$ when $\theta \ge \pi$ and $\beta > 0$ when $\theta \in[0,\pi]$. The ambiguity on the $\alpha$ angle, instead, cannot be resolved as it derives from the fact that exactly two different ellipses, having the same semi-major axis $a$, link $\mathbf r_1$ and $\mathbf r_2$ and thus two different time of flights exist that satisfy Eq.(\[eq:tof\_alpha\]). From Eq.(\[eq:g5\]) and Eq.(\[eq:g6\]) one can also derive the useful relation: $$\label{eq:g7}
\sin\frac\alpha 2 = \lambda\sin\frac\beta 2$$
The Lambert’s problem revival
-----------------------------
During the 18th-19th century, the work on the orbital boundary value problem culminated with Gauss masterpiece Theoria Motus Corporum Coelestium in Sectionibus Conicis Solem Ambientium [@gauss] where the prince of mathematicians conceives was is probably the first procedure able to accurately solve the Lambert’s problem (see Battin [@battin] for an excellent account of Gauss method). In the following years science drifted slowly away this topic, only to revisit it in the second half of the 20th century when the orbital boundary value problem received more attention in the context of Moon exploration. Hence, the work of Lancaster, Blanchard, Battin, Bate and many others introduced several advances on the topic. We here follow the approach from Lancaster and Blanchard that inspred most of our developments and we will thus rederive some of their relations which are needed to explain our new ideas. Consider the parameter $\lambda$ defined as: $$s \lambda = \sqrt{r_1r_2}\cos\frac\theta 2$$ using Eq.(\[eq:cost2\]) and substituting the expressions in Eq.(\[eq:g5\]) and Eq.(\[eq:g6\]) it is simple to show that: $$\lambda^2 = \frac{s-c} s$$ The parameter $\lambda \in [-1,1]$ is positive when $\theta \in[0,\pi]$ and negative when $\theta\in[\pi,2\pi]$. Values of $\lambda^2$ close to unity indicate a chord of zero length, a case which is indeed extremely interesting in interplanetary trajectory design as it is linked to the design of resonant transfers. We also introduce a non dimensional time-of-flight defined as: $$T = \frac 12 \sqrt{\frac \mu{a_m^3}} (t_2-t_1) = \sqrt{2\frac \mu{s^3}} (t_2-t_1)$$ where $a_m = s/2$ is the minimum energy ellipse semi-major axis [@battin]. The advantage of using $\lambda$ and $T$ derives from the fact that $T$ is a function of $a / a_m$ and $\lambda$ alone, which allows to greatly simplify the taxonomy of possible Lambert’s problems. In Gooding’s words [@gooding], all the triangles having equal $c / s$ ratio form a large equivalence class and can be described as L-similar. For them, all Lambert solutions are the same in terms of $a / a_m$ and $T$.
![Non-dimensional time of flight curve for $\lambda=-0.9$ parametrized using $a/a_m$ . \[fig:tof\_a\]](tof_a.png){width="75.00000%"}
If we now plot the time of flight given by Eq.(\[eq:tof\_alpha\]) as a function of the ratio between the semi-major axis and the minimum energy ellipse semi-major axis, for a particular value of $\lambda$ and for single and multiple revolution cases, we get Figure \[fig:tof\_a\]. It is evident how, in order to invert the time of flight relation iterating over $a / a_m$, while possible, is not a good choice. To avoid these problems we follow Lancaster and Blanchard in some further derivations introducing the new quantities: $$\label{eq:def_x}
\begin{array}{cc}
x = \left\{ \begin{array}{l} \cos\frac\alpha 2 \\ \cosh\frac\alpha 2 \end{array}\right., &
y = \left\{ \begin{array}{l} \cos\frac\beta 2 \\ \cosh\frac\beta 2 \end{array}\right.
\end{array}$$ which imply: $$\label{eq:def_x}
\begin{array}{cc}
\left\{ \begin{array}{l} \sqrt{1-x^2} = \sin\frac\alpha 2 \\ \sqrt{x^2-1} = \sinh\frac\alpha 2 \end{array}\right., &
\left\{ \begin{array}{l} \lambda \sqrt{1-x^2} = \sin\frac\beta 2 \\ \lambda \sqrt{x^2-1} = \sinh\frac\beta 2 \end{array}\right.
\end{array}$$ and $y = \sqrt{1-\lambda^2(1-x^2)}$. Using these relations it is possible to relate the auxiliary angles $\varphi$ and $\psi$ directly to $x$: $$\label{eq:g8}
\begin{array}{l}
\cos\varphi = xy - \lambda(1-x^2) \\
\cosh\varphi = xy + \lambda(x^2-1)
\end{array}
, \hskip1cm
\begin{array}{l}
\sin\varphi = (y+x\lambda)\sqrt{1-x^2} \\
\sinh\varphi = (y+x\lambda)\sqrt{x^2-1}
\end{array}$$ and, $$\label{eq:g9}
\begin{array}{l}
\cos\psi = xy + \lambda(1-x^2) \\
\cosh\psi = xy - \lambda(x^2-1)
\end{array}
, \hskip1cm
\begin{array}{l}
\sin\psi = (y - x\lambda)\sqrt{1-x^2} \\
\sinh\psi = (y - x\lambda)\sqrt{x^2-1}
\end{array}$$ which allows to derive the relations $\cos\varphi\sin\psi = (x-\lambda y)\sqrt{1-x^2}$, $\cosh\varphi\sinh\psi = (x-\lambda y)\sqrt{x^2-1}$ and thus have the following time of flight equation valid in all cases: $$\label{eq:tof_x}
T = \frac{1}{1-x^2}\left( \frac{\psi+M\pi}{\sqrt{|1-x^2|}} - x + \lambda y\right)$$ where we must set $M=0$ in the case of hyperbolic motion where unbounded motion prevents complete revolutions to happen. The auxiliary angle $\psi$ is computed using Eq.(\[eq:g9\]) by the appropriate inverse function and thus, the time of flight evaluation is reduced to one only inverse function computation. Given the bounds on $\alpha$, from the definition of $x$, we can see how $x \in [-1,\infty]$. Also, $x>1$ implies hyperbolic motion, while $x<1$ elliptic motion. Since $1-x^2 = \sin^2\frac\alpha 2 = \frac s{2a} = \frac{a_m}a$, we see how $x=0$ corresponds to the minimum energy ellipse. Note that different Lambert’s problems having identical $\lambda$ values (i.e. same $c / s$), result in the same $x$, we then say that $x$ is a Lambert invariant parameter.
Computing Eq.(\[eq:tof\_x\]) in $x=0$ we get: $$\label{eq:t0}
T(x=0)=T_{0M} = \arccos\lambda + \lambda\sqrt{1-\lambda^2} + M\pi = T_{00} + M\pi$$ where we have introduced $T_0$ as the value of $T$ in $x=0$ and $T_{00}$ as the value in the single revolution case $M=0$.
When computing Eq.(\[eq:tof\_x\]) in the single revolution case, a loss of precision is encountered due to numerical cancellation for values of $x\approx 1$ where both $1-x^2$ and $\psi$ tend to zero. In these cases we compute the time of flight equation by series expansion using the elegant result from Battin [@battin] setting: $$\label{eq:tof_x_series}
\begin{array}{l}
\eta = y - \lambda x \\
S_1 = \frac 12 (1-\lambda - x\eta) \\
Q = \frac 43 {}_1F_2(3,1,\frac 52,S_1) \\
2T = \eta^3Q + 4\lambda\eta
\end{array}$$ where ${}_1F_2(a,b,c,d)$ is the Gaussian or ordinary hypergeometric function. This can be evaluated by direct computation of the associated hypergeometric series. Noting that $S_1\to 0$ when $x\to 1$ the number of terms to retain in the series is small whenever the series is used in the neighbourhood of $x = 1$. Departing from Battin, we study the parabolic case substituting $x=1, y=1$ into Eq.(\[eq:tof\_x\_series\]) and thus obtaining the following remarkable expression: $$\label{eq:t1}
T(x=1) = T_1 = \frac{2}{3}(1-\lambda^3)$$ relating the geometry of the triangle created by two different observations of an object on a parabolic keplerian orbit to the non-dimensional time elapsed between them. It is also possible to derive the following formulas for the time of flight derivatives: $$\label{eq:derivatives}
\begin{array}{l}
(1-x^2)\frac {dT}{dx} = 3Tx-2+2\lambda^3\frac xy\\
(1-x^2)\frac {d^2T}{dx^2} = 3T + 5x\frac{dT}{dx}+2(1-\lambda^2)\frac{\lambda^3}{y^3}\\
(1-x^2)\frac{d^3T}{dx^3} = 7x\frac{d^2T}{dx^2} + 8 \frac{dT}{dx} - 6(1-\lambda^2)\lambda^5\frac x{y^5}
\end{array}$$ which are valid in all cases (single and multiple revolutions, elliptic and hyperbolic) except in $\lambda^2=1, x=0$ and $x=1,\forall \lambda$. We then apply de l’Hôpital rule to the first of the above equations, and using the expression derived for $T_1$ we are also able to find the value of the derivative of the time of flight curves in the case of a parabola: $$\label{eq:dt1dx}
\left. \frac{dT}{dx}\right|_{x=1} = \frac 25 (\lambda^5-1)$$ which is valid for $M=0$. By direct substitution, one can also easily show: $$\left. \frac{dT}{dx}\right|_{x=0} = -2$$ A great advantage of the time of flight equation in the form of Eq.(\[eq:tof\_x\]) as derived by Lancaster and Blanchard [@lancaster] is in the low computational cost of computing $T$ and its derivatives, up to the third order. Only one trigonometric (or hyperbolic) function inversion, two square roots and a few multiplications, divisions and sums are indeed necessary to compute these numerical values. Other approaches based on geomerical considerations or on a universal variables formulation are, at best, only able to match such a simple representation. We now summarize all the information relative to all possible Lambert problems in one single graph as done in Figure \[fig:tof\_x\].
![Time of flight curves parametrized using $x$ for different $\lambda$ and $M$ values. \[fig:tof\_x\]](tof_x.png){width="75.00000%"}
A new Lambert invariant variable
================================
Let us consider the following new variables: $$\label{eq:new_variables}
\xi = \left\{
\begin{array}{ll}
\log(1+x), &M=0 \\
\log(\frac{1+x}{1-x}), &M>0
\end{array} \right., \hskip1cm
\tau = \log(T)$$ The domain of the time-of flight curve is now extended to $[-\infty,\infty]$. In the case of $M=0$ the co-domain is also extended similarly. Let us study the resulting time of flight equation $\tau(\xi,\lambda,M)$. In Figure \[fig:tof\_xi\] we plot $\tau$ against $\xi$ for $M=0$ and $M=1$ and thirty equally spaced values of $\lambda\in[-0.9,0.9]$. In the case $M=0$ the curves appear to have two asymptotes having negative inclination coefficient. For the multiple revolution case ($M>0$) the curves have two symmetric asymptotes. The new introduced parameter $\xi$ is Lambert invariant according to Gooding’s definition [@gooding] as it essentially is a transformation of the Lambert invariant variable $x$. We study the differential properties of the new curves, we have:
$$\label{eq:new_variables_differentials}
d\xi = \left\{
\begin{array}{ll}
\frac{1}{1+x}dx, &M=0 \\
\frac{2}{1-x^2}dx, &M>0
\end{array} \right., \hskip1cm
d\tau = \frac 1T dT$$
Substituting these relations into Eq.(\[eq:derivatives\]), after some manipulations we may derive the following hybrid expressions for the derivatives in the case $M=0$: $$\label{eq:derivatives_tau}
\begin{array}{l}
\frac {d\tau}{d\xi} = \frac{1+x}T\frac{dT}{dx}\\
\frac {d^2\tau}{d\xi^2} = \frac{(x+1)^2}{T}\frac{d^2T}{dx^2}+\frac{d\tau}{d\xi}-\left(\frac{d\tau}{d\xi}\right)^2\\
\frac{d^3\tau}{d\xi^3} = \frac{(1+x)^3}{T}\frac{d^3T}{dx^3} + \left(\frac{d^2\tau}{d\xi^2}-\frac{d\tau}{d\xi}+\left(\frac{d\tau}{d\xi}\right)^2\right)\left(2-\frac{d\tau}{d\xi}\right)+\frac{d^2\tau}{d\xi^2} - 2\frac{d\tau}{d\xi}\frac{d^2\tau}{d\xi^2}
\end{array}$$
![Time of flight curves ($\tau$) parametrized using $\xi$ for 30 $\lambda$ values equally spaced in $[-0.9,0.9]$. \[fig:tof\_xi\]](tof_xi_M0.png "fig:"){width="49.00000%"} ![Time of flight curves ($\tau$) parametrized using $\xi$ for 30 $\lambda$ values equally spaced in $[-0.9,0.9]$. \[fig:tof\_xi\]](tof_xi_M1.png "fig:"){width="49.00000%"}
Note that these expressions can be computed, sequentially, after Eq.(\[eq:derivatives\]). The following holds for the $M=0$ case: $$\label{eq:asympt1}
\begin{array}{l}
\lim_{\xi\to\infty} \tau = -\xi+\log(1-\lambda|\lambda|)\\
\lim_{\xi\to-\infty} \tau = -\frac 32 \xi + \log\left(\frac{\pi}{4}\sqrt{2}\right)
\end{array}$$ which describes the asymptotic behaviour of the time of flight as visualized in Figure \[fig:tof\_xi\]. The two asymptotes are thus revealed to have negative coefficients $-1$ and $-3 / 2$. For the multirevolution cases the derivatives are found to be: $$\label{eq:derivatives_tau_mr}
\begin{array}{l}
\frac {d\tau}{d\xi} = \frac{1-x^2}{2T}\frac{dT}{dx}\\
\frac{d^2\tau}{d\xi^2}=\frac{(1-x^2)^2}{4T}\frac{d^2T}{dx^2} - x\frac{d\tau}{d\xi} - \left(\frac{d\tau}{d\xi}\right)^2\\
\begin{split}
\frac{d^3\tau}{d\xi^3}=\frac{(1-x^2)^3}{8T}\frac{d^3T}{dx^3}-\left(\frac{d^2\tau}{d\xi^2}+x\frac{d\tau}{d\xi}+\left(\frac{d\tau}{d\xi}\right)^2\right)\left(2x+\frac{d\tau}{d\xi}\right) - \\
-\frac{1-x^2}{2}\frac{d\tau}{d\xi}-x\frac{d^2\tau}{\xi^2}-2\frac{d\tau}{d\xi}\frac{d^2\tau}{d\xi^2}
\end{split}
\end{array}$$ again computable in cascade and the following asymptotic behavior can be derived: $$\label{eq:asympt2}
\begin{array}{l}
\lim_{\xi\to-\infty} \tau = \log\left(\frac{\pi+M\pi}8\right) - \frac 32 \xi\\
\lim_{\xi\to\infty} \tau = \log\left(\frac{M\pi}8\right) + \frac 32 \xi
\end{array}$$ revealing two symmetric asymptotes having inclination $\pm 3 / 2$.
![Time of flight curves parametrized using $\xi$ for different $\lambda$ and $M$ values. \[fig:tof\_xi\_tot\]](tof_xi.png){width="95.00000%"}
In Figure \[fig:tof\_xi\_tot\] we report the time of flight curves in the new variables for different values of $\lambda$ and $M$. The reader can then compare this $\xi$-$\tau$ plane to the $x$-$T$ plane visualised in Figure \[fig:tof\_x\].
Lambert Solver
==============
A Lambert solver can be defined as a procedure that returns, for a gravitational field of strength $\mu$, all the possible velocity vectors $\mathbf v_1$ and $\mathbf v_2$ along keplerian orbits linking $\mathbf r_1$, $\mathbf r_2$ in a transfer time $T^*$. The ingredient of such an algorithm are, essentially a) the choice of a variable to iterate upon and thus invert the time of flight curve, b) the iteration method, c) the starting guess to use with the iteration method and d) the reconstruction methodology to compute $\mathbf v_1$ and $\mathbf v_2$ from the value returned by the iterations. As we will detail, our Lambert solver a) iterates on the Lancaster-Blanchard variable $x$ using b) a Householder iteration scheme c) feeded by initial guesses found exploiting the curve shape in the $\tau$-$\xi$ plane and the new analytical results found above. Eventually the velocity vectors are reconstructed following, again, the methodology proposed by Gooding [@gooding]. The final pseudo-code of the proposed Lambert solver is reported in Algorithm \[alg:main\], \[alg:findxy\]. Note how we detect the maximum number of revolutions $M_{max}$ at the beginning by computing $T_{min}$ in one case. All other cases (i.e. $M<=M_{max}$) will not require the evaluation of a $T_{min}$ via an iterative procedure. By doing so we do not bound the roots (short and long period) and thus risk cases where the solution jumps between the long and short period branches. While this did not happen in our tests of the new routine, it is a possibility we are not safeguarding against.
The code, written in C++ and exposed to python, is made available as part of the open source project PyKEP from the European Space Agency github repository <https://github.com/esa/pykep/>. The final algorithm is the final result of many different trials to exploit the newly found results detailed above, and in particular the $T_0,T_1$ expressions (and their derivatives) and the $\xi$-$\tau$ plane. It is worth reporting how one very robust set-up, not selected as our final proposed algorithm, was that of iterating directly with a simple derivative free method (regula-falsi) on the $\xi$-$\tau$ plane using constant initial guesses (i.e $x_l = -0.7, x_r=0.7$). We ended up choosing a different set-up (Algorithm \[alg:main\], \[alg:findxy\]) which turned up to be faster in our computational tests.
The Householder iterations
--------------------------
One of the main differences of the proposed Lamber solver with respect to previous work is the use of the Householder iterative scheme as a root finder for the time-of-flight curves $T(x)-T^* = 0$. Householder iterations are not used widely as the added computational effort of computing higher order derivatives is not worth the gain whenever these request further function evaluations. In our case, as the derivatives computation is done using equations \[eq:derivatives\], Householder iterations are able to provide a significant benefit. We report the exact form used to implement the iterations as it is known how different numerical form can produce different behaviours. After experimenting with different implementations the following was used:
$$x_{n+1} = x_n - f(x_n) \frac{f'^2(x_n) - f(x_n)f''(x_n) / 2}{f'(x_n) (f'^2(x_n) - f(x_n)f''(x_n)) + f'''(x_n)f^2(x_n) / 6 }$$ where $f$ is, in our case, $T(x)-T^*$ and the derivatives are indicated as $f',f'',f'''$.
Initial Guess
-------------
To generate an initial guess for $x$ we use the newly introduced $\xi$-$\tau$ variables and the values $T_0$ and $T_1$ as computed from Eq.(\[eq:t0\]) and Eq,(\[eq:t1\]). Our initial guess is obtained inverting the following linear approxmation to the time of flight curves: $$\tau = c\xi + d$$ where we vary the $c$ and $d$ values according to the value of $\tau$ and $M$.
### Single revolution
Let us start form the single revolution case. Clearly, for high values of $\tau$, we must set $c = -1.0$, while for low values $c=-3/2$ so that the asymptotic behavior derived in Eq.(\[eq:asympt1\]) is reproduced. We then consider the following piece-wise linear approximation: $$\begin{array}{ll}
\tau = -\frac 32 \xi + \tau_0, & T \ge T_0 \\
\tau = - \xi + \log 2 +\tau_1, & T < T_1 \\
\tau = \tau_0 + \frac{\xi}{\log 2}(\tau_1 - \tau_0) & T_1 < T < T_0
\end{array}$$ where $\tau_1=\log T_1$ and $\tau_0$ = $\log T_0$. We have basically enforced the lines to pass through the points $(x_0,T_0)$ and $(x_1,T_1)$ and have the desired asymptotic behaviour.
The above relation is easily inverted and thus the following simple starter $\xi_0$ is derived: $$\begin{array}{ll}
\xi_0 = \frac 23(\tau_0-\tau), & T \ge T_0 \\
\xi_0 = \log 2+\tau_1-\tau, & T < T_1 \\
\xi_0 = \frac{\tau-\tau_0}{\tau_1-\tau_0} \log2 & T_1 < T < T_0
\end{array}$$ Transforming these relations back to the $x$-$T$ plane we find the following expression for the starter $x_0$: $$\label{eq:x0}
\begin{array}{ll}
x_0 = \left(\frac{T_0}{T}\right)^{\frac 23} -1, & T \ge T_0 \\ \\
x_0 = 2\frac{T_1}{T} -1, & T < T_1 \\
x_0 = \left(\frac{T_0}{T}\right)^{\log_2\left(\frac {T_1}{T_0}\right)} -1& T_1 < T < T_0
\end{array}$$ having an extremely low computational cost also in view of the fact $\log_2$ admits efficient implementaions. We can improve the expression for the $T<T_1$ case making use of the newly found result expressed in Eq.(\[eq:dt1dx\]). We thus set $$\begin{array}{ll}
x_0 = \left(\frac{T_0}{T}\right)^{\frac 23} -1, & T \ge T_0 \\ \\
x_0 = \frac 52 \frac{T_1(T_1-T)}{T (1-\lambda^5) } + 1, & T < T_1 \\
x_0 = \left(\frac{T_0}{T}\right)^{\log_2\left(\frac {T_1}{T_0}\right)} -1& T_1 < T < T_0
\end{array}$$ where we enforced the derivatives and values at $x=1$ and $x=\infty$. We report the error introduced by using these expressions in Figure \[fig:errors\] where we also show, for comparison, the same plot relative to the Gooding initial guess. The error is defined by the difference between the initial guess computed for a given $T^*$ and the actual value of $x$ resulting in a time of flight $T^*$.
![Absolute errors introduced by the Gooding initial guess (left) and the proposed initial guess (right) for the single revolution $M=0$ case. Each line correspond to a different $\lambda$ value ranging from -0.99 to 0.99 \[fig:errors\]](ic_err_gooding.png "fig:"){width="45.00000%"} ![Absolute errors introduced by the Gooding initial guess (left) and the proposed initial guess (right) for the single revolution $M=0$ case. Each line correspond to a different $\lambda$ value ranging from -0.99 to 0.99 \[fig:errors\]](ic_err_mine.png "fig:"){width="45.00000%"}
### Multiple revolutions
For the multiple revolution case, assuming a solution exists, there are two possible values of $x$ and thus we will need two starters. We obtain these by direct inversion of Eq.(\[eq:asympt2\]) that define the two asymptotes. $$\begin{array}{l}
\xi_{0l} = \frac 23(\log{\frac{\pi+M\pi}8}-\tau) \\
\xi_{0r} = \frac 23(\tau-\log{\frac{M\pi}8})
\end{array}$$ The above equations may then be transformed back to the $x$-$T$ plane so that the following simple starters are derived: $$\label{eq:x0lr}
\begin{array}{l}
x_{0l} = \frac{\left(\frac{M\pi + \pi}{8T}\right)^{\frac 23}-1}{\left(\frac{M\pi + \pi}{8T}\right)^{\frac 23}+1} \\
x_{0r} = \frac{\left(\frac{8T}{M\pi}\right)^{\frac 23}-1}{\left(\frac{8T}{M\pi}\right)^{\frac 23}+1}
\end{array}$$ The above expressions approximate well the time of flight curves as $|x| \to 1$. A great advantage of these expressions is that they do not make use of $T_{min}$, $x_{min}$ (i.e. the minimum of the time of flight curve and its extremal value) which would require a distinct set of iterations to be found. We then avoid to pass to the root solver solution bounds at each $M$, at the risk of allowing, during the iterations, switches between short and long period solutions, a theoretical occurence, though, that was never encountered in our experimental tests.
$t > 0$, $\mu > 0$ $\mathbf c = \mathbf r_2 - \mathbf r_1$ $c=|\mathbf c|$, $r_1=|\mathbf r_1|$, $r_2=|\mathbf r_2|$ $s=\frac 12 (r_1+r_2+c)$ $\hat {\mathbf i}_{r,1} = \mathbf r_1 / r_1$, $\hat {\mathbf i}_{r,2} = \mathbf r_2/ r_2$ $\hat {\mathbf i}_{h} = \hat {\mathbf i}_{r,1} \times \hat {\mathbf i}_{r,2}$ $\lambda^2 = 1-c/s$, $\lambda = \sqrt{\lambda^2}$ $\lambda = -\lambda$ $\hat {\mathbf i}_{t,1} = \hat {\mathbf i}_{r,1} \times \hat{\mathbf i}_h $, $\hat {\mathbf i}_{t,2} = \hat {\mathbf i}_{r,2} \times \hat
{\mathbf i}_{r,2}$ $\hat {\mathbf i}_{t,1} = \hat{\mathbf i}_h \times \hat {\mathbf i}_{r,1}$, $\hat {\mathbf i}_{t,2} = \hat{\mathbf i}_h \times \hat
{\mathbf i}_h$ $T = \sqrt{\frac{2\mu}{s^3}} t$ $x_{list},y_{list} = $findxy($\lambda$, $T$) $\gamma = \sqrt{\frac{\mu s}{2}}$, $\rho = \frac{r_1-r_2}{c}$, $\sigma = \sqrt{(1-\rho^2)}$ $V_{r,1} = \gamma [(\lambda y - x) - \rho(\lambda y + x)] / r_1$ $V_{r,2} = -\gamma [(\lambda y - x) + \rho(\lambda y + x)] / r_2$ $V_{t,1} = \gamma\sigma(y+\lambda x) / r_1$ $V_{t,2} = \gamma\sigma(y+\lambda x) / r_2$ $\mathbf v_1 = V_{r,1} \hat {\mathbf i}_{r,1} + V_{t,1} \hat {\mathbf i}_{t,1}$ $\mathbf v_2 = V_{r,2} \hat {\mathbf i}_{r,2} + V_{t,2} \hat {\mathbf i}_{t,2}$
$|\lambda| < 1$, $T < 0$ $M_{max} = \mbox{floor} (T / \pi)$ $T_{00} = \arccos \lambda + \lambda\sqrt{1-\lambda^2}$ start Halley iterations from $x=0,T=T_0$ and find $T_{min}(M_{max})$ $M_{max} = M_{max} - 1$ $T_1 = \frac 23 (1-\lambda^3)$ compute $x_0$ from Eq.(\[eq:x0\]) start Householder iterations from $x_0$ and find $x,y$ compute $x_{0l}$ and $x_{0r}$ from Eq.(\[eq:x0lr\]) with $M=M_{max}$ start Householder iterations from $x_{0l}$ and find $x_r,y_r$ start Householder iterations from $x_{0r}$ and find $x_l,y_l$ $M_{max}=M_{max}-1$
New solver performances
=======================
To test the performances of our new algorithm we start by assessing its accuracy. We consider, for an assigned $M$, a random $\lambda \in [-0.999,0.999]$ and a random $x_{true} \in [-0.99,3]$ (or $x_{true} \in [-0.999,0.999]$ whenever $M>0$) and we compute the resulting time of flight $T$. We then use Housholder iterations starting from the appropriate initial guess to find back the $x$ value. We find that stopping the iterations whenever the difference between the $x$ value computed at two successive iterations is less than $10^{-5}$ ($10^{-8}$ whenever $M>0$) is a good setting. We record, for each of such trials, the number of iterations made by the Householder method $it$, and the error defined as $\epsilon = |x_{true}-x|$. This is repeated 1,000,000 times for $M=0$ and then 100,000 times for each $M=1,2,....,50$. The result is shown in Figure \[fig:accuracy\]. In the vast majority of cases we obtain an error $<10^{-13}$. Few cases have a slightly larger error (up to $<10^{-11}$) and these are mainly corresponding to multirevolution cases where $T\approx T_{min}$. We also note, on the proximity of $\lambda=1$ values, a distinct rise in the absolute error. This is due to the $M=0$ case and the loss of precision in the computation of $y$ from $\lambda$, a problem that can be avoided computing $y$ directly from the problem geometry, but it is here deemed as not necessary. Looking then at the number of iterations, we compute the mean over all instances having the same $M$ value. We obtain, for the single revolution case, an average of 2.1 iterations while, in the multiple revolution case, we get an average of 3.3 iterations to convergence. Note how in these tests we do not find a case where a switch occurs between the short period and long period solution during the root solving procedure. Such a switch would infact immediately appear in Figure \[fig:accuracy\] as a point with a large absolute error $\epsilon$.
![Absolute error $\epsilon = |x-x_{true}|$ of our Lambert solver as a function of $\lambda$. This is achieved, on average over the $M=0$ cases, in 2 iterations \[fig:accuracy\]](accuracy.png){width="95.00000%"}
We then turn to the anaysis of our algorithm complexity with respect to the known Gooding algorithm, considered by many as the most accurate and efficient Lambert solver up to date. First we note that in terms of accuracy, Gooding algorithm is comparable to ours. We then run a speed test. For the purose of this test we reimplement both our and Gooding algorithm in pure Python language (i.e. no C++ bindings) and we record the execution time to solve the same 100,000 randomly generated problems (using the same bounds as above). In the case $M=0$, the proposed algorithm resulted to be faster by a factor $1.25$, while in the multi revolution cases by a factor $1.5$. This type of test is very sensitive to implementation details and to the underlying computing architecture and even if we did our best to pay as much attention to them in both cases, we support our result with more general considerations. The main difference between our algorithm and Gooding’s is in the initial guess generation and in the iteration method. Gooding algorithm employs Halley iterations, while we make use of Householder iterative scheme. While Halley’s method has a slighlty lower complexity and does not need to compute also the third derivative from Eq.(\[eq:derivatives\]), our iterative scheme reaches, in the case $M=0$, a comparable accuracy in only 2 iterations on average compared to the 3 iterations needed for the Gooding case. For $M>0$ the number of required iterations is comparable in both cases but the initial Guess used in Gooding algorithm has, in general, a higher complexity as it makes use of a higher number of square roots and exponentiations. In the $M>0$ case Gooding initial guess also requires the determination of $x_{min},T_{min}$ via a further Halley iterative scheme, while the initial guess we use does not make use of any particular value, while still allowing the Householder method to converge within a few iterations and in all cases. We must, though, note once more that as we do not compute $x_{min},T_{min}$ we also cannot bound the solution during the root solver iterations and thus allow for the theoretical possibilty of a switch between short and long period solutions. Such a rare event never appeared in our extensive testing of the new routine
Finally, we measure the error also in terms of the computed terminal velocities by comparing all $\mathbf v_2$, returned by our Lambert solver, to the same values as computed via numerical propagation (using Lagrange coefficients) from $\mathbf r_1,\mathbf v_1$. We do this by instantiating at random $\mathbf r_1, \mathbf r_2$ with each component in the range $[-4,4]$ and $t\in[0.1,100]$. For the purpose of this test we consider $\mu=1$ and we measure the norm of the resulting vector of the velocity difference. Repeating this experiment for a total cumulative 10,000,000 Lambert’s Problems, an average error of $10^{-13}$ is obtained, with a maximum error measured to be $10^{-8}$.
Conclusion
==========
We revisit Lambert’s problem building upon the results of Lancaster and Blanchard and finding some new properties of the time of flight curves. We propose a new transformation of such curves able to further simplify the problem suggesting efficient approximations to the final solution. Using our results to design a new procedure to solve the Lambert problem we are able to build a low complexity algorithm that we find able to provide accurate solutions in a shorter time when compared to the state of the art Gooding’s algorithm.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We illustrate the basic notions of [*additional non-isospectral symmetries*]{} and their interplay with the discrete [*[[Darboux-Bäcklund]{} ]{}transformations*]{} of integrable systems at the instance of [*constrained Kadomtsev-Petviashvili*]{} () integrable hierarchies. As a main application we present the solution of discrete multi-matrix string models in terms of Wronskian $\t$-functions of graded $SL(m,1)$ hierarchies.'
---
by -by -
9.5in 0.5in 6.7in -.6in
\#1[(\[eq:\#1\])]{} \#1[\[eq:\#1\]]{} \#1 ${\left(}
\def$[)]{} v \#1 \#1
c ł Ł Ø ¶ §
=cmss12 =cmu10 scaled1
\#1["7017[\#1]{} 23\#1]{}
\[section\] \[section\] \[section\] \[section\] \[section\]
1[[**W\_[1+]{}**]{}]{}
\#1\#2\#3[[*Trans. Am. Math. Soc.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Invent. Math.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Advances in Math.*]{} [**\#1**]{} (\#2) \#3]{} \#1\#2\#3[[*Proc. Natl. Acad. Sci. USA*]{} [**\#1**]{} (\#2) \#3]{}
December, 1995\
${}$\
${}$
.3in
[ H. Aratyn]{}
.1in Department of Physics\
University of Illinois at Chicago\
845 W. Taylor St.\
Chicago, IL 60607-7059, [*e-mail*]{}: [email protected]\
.3in E. Nissimov${}^{a),\, 3}$ and S. Pacheva${}^{a),\, b),}$
.1in ${}^{a)}$ Institute of Nuclear Research and Nuclear Energy\
Boul. Tsarigradsko Chausee 72, BG-1784 Sofia, Bulgaria\
[*e-mail*]{}: [email protected], [email protected]\
and\
${}^{b)}$ Department of Physics, Ben-Gurion University of the Negev\
Box 653, IL-84105 $\;$Beer Sheva, Israel\
[*e-mail*]{}: [email protected], [email protected]
.3in
Integrable systems (for the basics, see refs. [[@Faddeev; @QISM; @Olsha]]{}) constitute an outstanding branch of theoretical physics since they describe a vast variety of fundamental non-perturbative phenomena ranging from $D=2$ (space-time dimensional) nonlinear soliton physics and planar statistical mechanics to string and membrane theories in high-energy elementary particle physics. It turns out that, under plausible assumptions, a variety of physically interesting theories in higher space-time dimensions can be reformulated as lower-dimensional ($D=2$) integrable models which in the same time possess [*infinite-dimensional symmetries*]{} and thus, as a rule, being integrable (see, especially, the recent developments [[@Seiberg-Witten]]{} related with integrability of Seiberg-Witten effective low-energy theory of (extended) supersymmetric gauge theories).
Among the various infinite-dimensional symmetry groups and algebras playing rôle in integrable field theories, a particularly distinguished place belongs to the Lie algebra $W_{1+\infty}$ [[@W-inf]]{} (specific “large $N$ limit” of Zamolodchikov’s $W_N$ conformal algebras [[@Zam]]{}, isomorphic to the Lie algebra of all purely differential operators on the circle). It contains (together with its supersymmetric extension) all previously known infinite-dimensional symmetry algebras – Virasoro and Kac-Moody. Also, it is precisely a Lie-algebraic deformation of the infinite-dimensional generalization of the Virasoro algebra – the algebra of area-preserving diffeomorphisms.
Recently $\, W_{1+\infty}$ symmetries attracted broad interest as they appeared naturally as inherent structures of models in different areas of theoretical physics : theory of black holes and space-time singularities, two-dimensional quantum gravity, nonlinear evolution equations in higher dimensions, self-dual gravity, $N=2$ superstring theory (refs. (a)–(d) in [[@W-appl]]{}), quantum Hall effect [[@Cappelli]]{}. All listed models possess, in one form or in another, [*exactly soluble*]{} features, which naturally suggest an intimate connection of [*integrability*]{} with $W_{1+\infty}$ algebra. This claim may, furthermore, be substantiated by the realization that $W_{1+\infty}$ algebra is a subalgebra of the algebra $\PsDA$ of arbitrary pseudo-differential operators on the circle. Already since the pioneering papers of Adler-Kostant-Symes and of the Faddeev’s school [[@AKS]]{} it was realized that $\PsDA$ forms the foundation of completely integrable systems. In fact, it turns out that most known integrable models (i.e. those admitting Lax or “zero curvature” representation) can be associated with specific coadjoint orbits of various subalgebras of $\PsDA$ or with different Hamiltonian reductions thereof [[@Faddeev]]{}.
The generic integrable system based on $\PsDA$ symmetry algebra is the Kadomtsev-Petviashvili (${\bf KP}$) integrable hierarchy [[@Zakh; @Dickey]]{} of soliton nonlinear evolution equations. Its name derives from the fact that ${\bf KP}$ hierarchy contains the $D=2+1$ dimensional nonlinear soliton ${\bf KP}$ equation which appeared originally in plasma physics. In the last few years the main interest towards ${\bf KP}$ hierarchy originates from its deep connection with the statistical-mechanical models of random matrices ((multi-)matrix models) providing non-perturbative discretized formulation of string theory [[@integr-matrix]]{}.
The purpose of the present talk is to provide a brief discussion of the basic notions of [*additional non-isospectral symmetries*]{} [[@add-symm; @Dickey]]{} and their interplay with the discrete [*[[Darboux-Bäcklund]{} ]{}transformations*]{} [[@DB]]{} of integrable systems within the ${\bf KP}$ integrable hierarchy (being and arch-type integrable system, as pointed out above) and its various constrained versions (hierarchies) relevant in discrete multi-matrix string models. Furthermore, we show how to obtain the solution of the latter in terms of Wronskian $\t$-functions of the hierarchies.
We describe the (generalized) ${\bf KP}$ integrable hierarchy in the language of pseudo-differential operators (for a background, see [[@Dickey]]{}). The main object is the pseudo-differential Lax operator $L$ subject to an infinite set of evolution equations: L = D\^m + \_[j=0]{}\^[m-2]{} v\_j D\^j + \_[i 1]{} u\_i D\^[-i]{} , L = [[L\^[lm]{}\_[(+)]{}]{} ,[L]{}]{} Here, the coefficients of $L$ are (smooth) functions of $x \equiv t_1$ and the higher time-evolution parameters $t_2 ,t_3 ,\ldots$ ; $D \equiv \pa_x$, whereas the subscripts $(\pm)$ denote purely differential (purely pseudo-differential) part of the corresponding pseudo-differential operators. The flows ${{{\partial {}}\over{\partial {t_l}}}}$ in commute (as vector fields on the space of Lax operators ) among themselves which expresses the integrability of the ${\bf KP}$ system.
Within the Sato-Wilson dressing operator formalism, with the following dressing expression for the generalized ${\bf KP}$ Lax operator : L= W D\^m W\^[-1]{} ,W 1 + \_[i 1]{} w\_i D\^[-i]{} the evolution equations for $L$ are equivalent to: W = - $ W D^l W^{-1} $\_[-]{} W
In what follows we shall also need the important notions of [*(adjoint) eigenfunctions*]{} and [*(adjoint) Baker-Akhiezer*]{} functions. The function $\Phi$ ($ \Psi$) is called for the Lax operator $L$ satisfying Sato’s flow equations if its flows are given by expression: = L\^[lm]{}\_[(+)]{} ; = - $ L^{*} $\^[lm]{}\_[(+)]{} for the infinite many times $t_l$ . If, in addition, an (adjoint) eigenfunction satisfies the spectral equation: L(ł) = ł(ł) , (ł) = L\^[lm]{}\_[(+)]{} (ł) ; L\^[\*]{}\^[\*]{} (ł) = ł\^[\*]{}(ł) , \^[\*]{} (ł) = - $ L^{*} $\^[lm]{}\_[(+)]{} \^[\*]{}(ł) $\psi^{(\ast )} (\l )$ is called (BA) function.
The BA function of the generic Lax operator $L$ is obtained from the BA function of the “free” Lax operator $L^{(0)} = D^m$ : \^[(0)]{} (ł) = { \_[l 1]{} t\_l ł\^[l m]{} } e\^[({ t}, ł)]{} by applying the dressing operator $W$ to : (ł) = W e\^[({ t}, ł)]{} = e\^[({ t},ł)]{} The function $\t (\{ t_l \} )$ of all evolution parameters is called of the (generalized) ${\bf KP}$ hierarchy and by itself constitutes an alternative natural way to describe the pertinent integrable system.
The ${\bf KP}$ system is endowed with bi-Hamiltonian Poisson bracket structures (another expression of its integrability) which results from the two compatible Hamiltonian structures on the algebra of pseudo-differential operators $\PsDA$ [[@STS83]]{}. The latter are given by: \_1 - L X, Y\
[[[[|. [X]{} ]{}]{} ,[[|. [Y]{} ]{}]{}]{}]{}\_2 \_A $ \( LX$\_[(+)]{} LY - $ XL$\_[(+)]{} YL\
&+& dx [Res]{}( [[L]{} ,[X]{}]{}) \^[-1]{} [Res]{}( [[L]{} ,[Y]{}]{}) where the following notations are used. $<\cdot \v \cdot >$ denotes the standard bilinear pairing in $\PsDA$ via the Adler trace ${\left\langle {L}\right|\left. {X} \right\rangle} = {\Tr}_A \( LX\)$ with ${\Tr}_A X = \int {\rm Res} X $. Here $X,Y$ are arbitrary elements of the algebra of pseudo-differential operators of the form $X = \sum_{k \geq - \infty} D^k X_k $ and similarly for $Y$. The second term on the r.h.s. of is a Dirac bracket term originating from the second-class Hamiltonian constraint $v_{m-1}=0$ on $L$ .
In terms of the Lax coefficient functions $v_{m-2},\ldots ,v_0 ,u_1 ,u_2 ,\ldots ,$ the first Poisson bracket structure takes the form of an infinite-dimensional Lie algebra which is a direct sum of two subalgebras spanned by $\{ v_j\}$ and $\{ u_i\}$, respectively. The latter is called $\Win1$-algebra [[@W-inf]]{}. Its Cartan subalgebra contains the infinite set of (Poisson-)commuting ${\bf KP}$ integrals of motion $H_{l-1} = {1\over l} {\Tr}_A L^{l\over m}$ whose densities are expressed in terms of the $\t$-function as: \_x = [Res]{} L\^[lm]{}
In turn, the second Poisson bracket structure spans a nonlinear (quadratic) algebra called $\hWinf (m)$ [[@W-h-inf]]{}, which is an infinite-dimensional generalization of Zamolodchikov’s $W_N$ conformal algebras [[@Zam]]{}.
Let us now turn our attention to a specific class of Hamiltonian reductions of the full (generalized) ${\bf KP}$ system (hierarchies, for short), where the purely pseudo-differential part of the ${\bf KP}$ Lax operator is parametrized through a [*finite*]{} number of functions (fields). To this end let us recall the notion of (adjoint) eigenfunction of a Lax operator . As in [[@chengs]]{} let us consider the flow of a vector field $\pa_\a$ given by: \_L = [[L]{} ,[\_[i=1]{}\^M \_i D\^[-1]{} \_i]{}]{} where $\Phi_i , \Psi_i$ are a set of $M$ independent (adjoint) eigenfunctions of $L$ . Using the simple identity valid for any differential operator $B_{(+)}$: \_[-]{} = $ B_{(+)}\Phi_i $ D\^[-1]{} \_i - \_i D\^[-1]{} $ B^{*}_{(+)}\Psi_i $ one can easily show that: \_, L = 0 ,l=1,2,… Now, the constrained KP hierarchy (denoted as ${\sf cKP}_{m,M}$) is obtained by identifying the “ghost” flow $\pa_{\a}$ with the isospectral flow ${{{\partial {}}\over{\partial {t_m}}}}$ which, upon comparison of with , implies the following constrained form of $L$: L L\_[m,M]{} = L\_[(+)]{} + \_[i=1]{}\^M \_i D\^[-1]{} \_i =D\^m+ \_[j=0]{}\^[m-2]{} v\_j D\^j + \_[i=1]{}\^M \_i D\^[-1]{} \_i subject to the same Lax evolution equations as in . Moreover, using again identity one finds that the functions $\Phi_i , \Psi_i$ remain (adjoint) eigenfunctions of the constrained Lax operator $L_{m,M}$ .
As shown in ref.[[@avoda]]{}, the ${\sf cKP}_{m,M}$ hierarchies given by are equivalent to the so called “multi-boson” hierarchies [[@multi-b]]{}: L\_[m,M]{} = L\_[(+)]{} + \_[i=1]{}\^M A\^[(M)]{}\_i $ D - B^{(M)}_i $\^[-1]{} $ D - B^{(M)}_{i+1}$\^[-1]{} $ D - B^{(M)}_M $\^[-1]{}\
A\^[(M)]{}\_k = (-1)\^[M-k]{} \_[s=1]{}\^k \_s\
B\^[(M)]{}\_k = - \_x where Wf\_1 ,…,f\_k \_x\^[i-1]{} f\_j denotes the standard Wronskian.
There is still another useful representation of the ${\sf cKP}_{m,M}$ Lax operator as a ratio of two purely differential Lax operators [[@no2rabn1; @office; @rio]]{}: L\_[m,M]{} = L\_[m+M]{} $ L_M$\^[-1]{} ; m , M 1\
L\_[m+M]{} $ D - b_{m+M}$ $ D - b_{m+M-1} $ $ D - b_1 $ , L\_M $ D - {\ti b}_{M}$ $ D - {\ti b}_{M-1} $ $ D - {\ti b}_1 $ where the coefficients $b_j , {\ti b}_j $ are subject to the constraint: \_[j=1]{}\^[m+M]{} b\_j - \_[l=1]{}\^[M]{} [b]{}\_l = 0
As already proved in detail in ref.[[@office]]{}, the ${\sf cKP}_{m,M}$ Lax operator $L=L_{m,M}$ obeys the same two compatible Poisson bracket structures and , [*i.e.*]{}, ${\sf cKP}_{m,M}$ hierarchies are legitimate Hamiltonian reductions of the full (generalized) ${\bf KP}$ hierarchy. Moreover, the second Poisson bracket structure in terms of the coefficients $ $ takes the form of free-field Poisson bracket algebra: [[b\_i (x)]{} ,[b\_j (y)]{}]{} $ \d_{ij} - {1\over m}$ \^ (x-y) , i,j =1,…, m+M\
[[[b]{}\_k (x)]{} ,[[b]{}\_l(y)]{}]{} - $ \d_{kl} + {1\over m}$ \^ (x-y) ,k,l =1,…,M\
[[b\_i (x)]{} ,[[b]{}\_l (y)]{}]{} \^ (x-y) which, as demonstrated in refs.[[@Yu; @office]]{}, is precisely the Cartan subalgebra of the graded $SL(m+M,M)$ Kac-Moody algebra. This latter property justifies the alternative name of the constrained ${\sf cKP}_{m,M}$ hierarchies – $SL(m+M,M)$ KP-KdV hierarchies.
In other words, – provide via eq. explicit free-field realizations of the nonlinear $\hWinf (m)$ algebra. Similar free-field realizations exist also for $\Win1$ – the first ${\bf KP}$ Poisson bracket structure (see refs.[[@multikp; @office]]{}). Let $L$ be again a pseudo-differential Lax operator of the full generalized KP hierarchy (recall $x \equiv t_1$) and let $M$ be a pseudo-differential operator “canonically conjugated” to $L$ such that: [[L]{} ,[M]{}]{} = , M = [[L\^[lm]{}\_[(+)]{}]{} ,[M]{}]{} Within the Sato-Wilson dressing operator formalism – the $M$-operator can be expressed in terms of dressing of the “bare” $M^{(0)}$ operator: M\^[(0)]{} = \_[l 1]{} t\_l D\^[l-m]{} = X\_[(m)]{} + \_[l 1]{} t\_[m+l]{} D\^l\
X\_[(m)]{} \_[l=1]{}\^[m]{} t\_l D\^[l-m]{} conjugated to the “bare” Lax operator $L^{(0)} = D^m$, [*i.e.*]{}: M W M\^[(0)]{} W\^[-1]{} = W X\_[(m)]{} W\^[-1]{} + \_[l 1]{} t\_[m+l]{} L\^[lm]{} = \_[l 0]{} t\_[m+l]{} L\^[lm]{}\_[(+)]{} + M\_[-]{}\
M\_[-]{} W X\_[(m)]{} W\^[-1]{} - t\_m - \_[l 1]{} t\_[m+l]{} W . W\^[-1]{} where in we used eqs.. Note that $X_{(m)}$ is a pseudo-differential operator satisfying ${\Bigl\lbrack\,{D^m}\, ,\,{X_{(m)}}\,\Bigr\rbrack} = \one$ .
On BA functions the action of $M$ is as follows: M (ł) = $ {{{\partial {}}\over{\partial {\l}}}} + \a_m (\l )$ (ł) where $\a_m (\l )$ is a function of $\l$ only.
Since any eigenfunction $\Phi$ can be represented as a linear “superposition” of BA functions : ({ t}) = \_ dł (ł) (ł,{ t}) (with an appropriate contour $\Gamma$ in the complex $\l$-plane, such that the integral in exists), eq. implies that: M ({ t}) = \_ dł$ -{{{\partial {}}\over{\partial {\l}}}} + \a_m (\l )$(ł)(ł,{ t})
The so called [*additional (non-isospectral) symmetries*]{} [[@add-symm; @Dickey]]{} are defined as vector fields on the space of ${\bf KP}$ Lax operators or, alternatively, on the dressing operator , through their flows as follows: \_[k,n]{} L = - [[$ L^k M^n$\_[-]{}]{} ,[L]{}]{} = [[$ L^k M^n$\_[(+)]{}]{} ,[L]{}]{} + n L\^k M\^[n-1]{} , \_[k,n]{} W = - $ L^k M^n$\_[-]{} W which [*commute*]{} with the usual ${\bf KP}$ flows ${{{\partial {}}\over{\partial {t_l}}}}$ .
.1in
Let us now turn our attention to the notion of [[Darboux-Bäcklund]{} ]{}(DB) transformations of (generalized) ${\bf KP}$ hierarchy and its reductions – ${\sf cKP}_{m,M}$ hierarchies , defined as follows [[@DB; @avoda]]{} : &=& T L T\^[-1]{} \_[(+)]{} + [L]{}\_[-]{} ,T D \^[-1]{}\
[L]{}\_[(+)]{} &=& [L]{}\_[(+)]{} + $\pa_x \(\chi^{-1} L_{(+)}\chi$\_[1]{} D\^[-1]{} \^[-1]{} \_[-]{} &=& \_0 D\^[-1]{} \_0 + D \^[-1]{} L\_[-]{} D\^[-1]{} \^[-1]{} $ \; =
{\wti \Phi}_0 D^{-1} {\wti \Psi}_0 +
\sum_{i=1}^{M} {\wti \Phi}_i D^{-1} {\wti \Psi}_i
\quad {\rm for} \;\; L = L_{m,M} \; $\
\_0 &=& $ \chi D \chi^{-1} L $ TL,\_0 = \^[-1]{}\
\_i &=& \_x $ \chi^{-1} \Phi_i $ ,\_i = - \^[-1]{} \_x\^[-1]{} $ \Psi_i \chi $ where $\chi$ is an (non-BA) eigenfunction of $L$. The DB-transformed Lax operator satisfies the same flow equations w.r.t. $t_l$ as in : ${{{\partial {}}\over{\partial {t_l}}}} {\wti L} = {\Bigl\lbrack\,{{\wti L}^{l\over m}_{(+)}}\, ,\,{{\wti L}}\,\Bigr\rbrack}$ due to the simple identity valid for any pseudo-differential operator $B$ $ \chi D \chi^{-1} B \chi D^{-1} \chi^{-1} $\_[(+)]{} = D \^[-1]{} B\_[(+)]{} D\^[-1]{} \^[-1]{} - \_x $\chi^{-1} (B_{(+)} \chi) $ D\^[-1]{} \^[-1]{} (and using the fact that $\chi$ is an eigenfunction of $L$). Moreover, eq. shows that, in order to preserve the ${\sf cKP}_{m,M}$ form of the DB-transformed Lax operator ${\wti L}={\wti L}_{m,M}$, we have to choose $\chi = \Phi_{i_0}$ where $\Phi_{i_0}$ is any one of the eigenfunctions of the initial $L=L_{m,M}$ .
One can generalize – for successive [[Darboux-Bäcklund]{} ]{}transformations on the initial $L=L_{m,M}\equiv L^{(0)}$ as follows. Within each subset of $m$ successive steps we can perform the DB transformations w.r.t. the $m$ different eigenfunctions of . Repeated use of the following important composition formula for Wronskians [[@Wronski]]{} : T\_k T\_[k-1]{} T\_1 (f ) = where T\_j = [ W\_[j]{} W\_[j-1]{} ]{} D [ W\_[j-1]{} W\_[j]{} ]{} = $ D + \( \ln { W_{j-1} \over W_{j} } $\^ ;W\_[0]{}=1\
W\_k W\_k \_1, …,\_k ,W\_[k-1]{} $f $W\_[k]{} \_1, …,\_[k-1]{}, f and employing short-hand notations: T\^[(k)]{}\_i \^[(k)]{}\_i D $\Phi^{(k)}_i$\^[-1]{} ;\^[(s)]{}\_i $ L^{(0)}$\^s \^[(0)]{}\_i , i=1,…,m where the upper indices in brackets indicate the order of the corresponding DB step, yields the following generalization of – (below $1 \leq l \leq m$) : \_i\^[(km+l)]{} = T\^[(km+l-1)]{}\_l …T\^[(km)]{}\_1 T\^[(km-1)]{}\_m …T\^[((k-1)m)]{}\_1 …T\^[(m-1)]{}\_m …T\^[(0)]{}\_1 \^[(k\_)]{}\_i\
=\
\^[(k\_[+]{})]{}\_i \^[(k+1)]{}\_i 1 i l ; \^[(k\_[-]{})]{}\_i \^[(k)]{}\_i l+1 i m Correspondingly, for the $\t$ function after $km+l$ steps of successive DB transformations we get: = \^[(km+l-1)]{}\_l …\^[(km)]{}\_1 \^[(km-1)]{}\_m …\^[((k-1)m)]{}\_1 …\^[(m-1)]{}\_m …\^[(0)]{}\_1\
= W\^[(0)]{}\_1 ,…,\^[(0)]{}\_m , \^[(1)]{}\_1,…,\^[(1)]{}\_m,…,\^[(k-1)]{}\_1,…,\^[(k-1)]{}\_m , \^[(k)]{}\_1,…,\^[(k)]{}\_l
.1in
We now formulate the main result of this section – the condition for [*compatibility*]{} between additional-symmetry flows and [[Darboux-Bäcklund]{} ]{}transformations .
Let $\Phi$ be an eigenfunction of $L$ defining a [[Darboux-Bäcklund]{} ]{}transformation, [*i.e.*]{} : = L\^[lm]{}\_[(+)]{} , = $\Phi D \Phi^{-1}$ L $\Phi D^{-1} \Phi^{-1}$ or, in terms of dressing operator: = $\Phi D \Phi^{-1}$ W D\^[-1]{} Then the DB-transformed $M$ operator (cf. ) acquires the form: = $\Phi D \Phi^{-1}$ M $\Phi D^{-1} \Phi^{-1}$ = \_[l 0]{} t\_[m+l]{} [L]{}\^[lm]{}\_[(+)]{} + [M]{}\_[-]{}\
[M]{}\_[-]{} = [W]{} [X]{}\_[(m)]{} [W]{}\^[-1]{} - t\_m - \_[l 1]{} t\_[m+l]{}[W]{}.[W]{}\^[-1]{} where ${\wti X}_{(m)} = D X_{(m)} D^{-1}$ with $X_{(m)}$ as in . Clearly ${\wti X}_{(m)}$, like $X_{(m)}$, is also admissible as canonically conjugated to $D^m$.
The DB-transformed BA function reads in accordance with : (ł) = ł\^[-[1m]{}]{} \_x $ \Phi^{-1} \psi (\l )$ and the DB-transformed $M$-operator acts on it as: (ł) = $ {{{\partial {}}\over{\partial {\l}}}} + {\wti \a_m} (\l )$ (ł) , (ł) = \_m (ł) + [1m]{} ł\^[-1]{}
Taking into account , we arrive at the following important
Motivated by applications to (multi-)matrix models (see next sections and ref.[[@oakpark]]{}), one can require invariance under some of the additional-symmetry flows, [*e.g.*]{}, under the lowest one ${\bar \pa}_{0,1}$ known as “string-equation” constraint in the context of (multi-)matrix models: \_[0,1]{} L = 0 [[M\_[(+)]{}]{} ,[L]{}]{} = - ; \_[0,1]{} = 0 M\_[(+)]{} = 0 Eqs., using second eq.,– and , lead to the following constraints for $L$, the BA function $\psi (\l )$ and the DB-generating eigenfunction $\Phi$ of $L$, respectively: \_[l 1]{} t\_[m+l]{} L + [[t\_1]{} ,[L]{}]{} \_[m,1]{} = -\
$ \sum_{l \geq 1} \frac{l+m}{m} t_{m+l} {{{\partial {}}\over{\partial {t_l}}}} +
t_m - \a_m (\l )$ (ł) = (ł)\
$ \sum_{l \geq 1} \frac{l+m}{m} t_{m+l} {{{\partial {}}\over{\partial {t_l}}}} + t_m $ = 0 ({ t}) = \_ dł e\^[(ł)]{} (ł;{ t})
Now let us recall the formula for the $\t$-function ratio for ${\sf cKP}_{m,M}$ hierarchies subject to successive DB transformations. Noticing that the eigenfunctions $\Phi^{(k)}$ of the DB-transformed Lax operators $L^{(k)}$ satisfy the [*same*]{} constraint eq. irrespective of the DB-step $k$, we arrive at the following result (“string-equation” constraint on the $\t$-functions) :
It is well-known that the discrete one-matrix model can be associated to the following chain of the Lax operators connected via DB transformations: L\^[(k+1)]{} $\Phi^{(k)} D {\Phi^{(k)} }^{-1}$ L\^[(k)]{} $\Phi^{(k)} D^{-1} {\Phi^{(k)} }^{-1}$ = D + \^[(k+1)]{} D\^[-1]{} \^[(k+1)]{}\
\^[(k+1)]{} \^[(k)]{} $ \ln \Phi^{(k)}$\^ + $\Phi^{(k)}$\^2 \^[(k)]{} ,\^[(k+1)]{} = $\Phi^{(k)}$\^[-1]{} where \^[(n)]{} = [ W\_[n+1]{} , , …, \^n W\_n , , …, \^[n-1]{} ]{} , = W\_n , , …, \^[n-1]{} with $\p = \int d \l \exp \({\sum_{k=1}^{\infty}} t_k \l^k \)$. The above proposition (with $m=1$) coincides perfectly with the “string-equation” ${\cal L}_{-1}^{(N)} W_N \lb \p, \pa \p, \ldots , \pa^{N-1} \p \rb =0$, with ${\cal L}_{-1}^{(N)} = {\sum_{k=2}^{\infty}} k t_k {{{\partial {}}\over{\partial {t_{k-1}}}}} + N t_1$.
The partition function of the multi-matrix ($q$-matrix) string model reads: Z\_N { t\^[(1)]{}},…,{ t\^[(q)]{}},{ g} = dM\_1 …dM\_q -\_[=1]{}\^q \_[r\_=1]{}\^[p\_]{} t\^[()]{}\_[r\_]{} M\_\^[r\_]{} + \_[=1]{}\^[q-1]{} g\_[,+1]{} M\_M\_[+1]{} where $M_\a$ are Hermitian $N \times N$ matrices, and the orders of the matrix “potentials” $p_{\a}$ may be finite or infinite. In refs.[[@BX]]{} it was shown that, via the method of generalized orthogonal polynomials [[@ortho-poly]]{}, one associates to generalized Toda-like lattice systems subject to specific constraints, so that $Z_N$ and its derivatives w.r.t. the coupling parameters can be expressed in terms of solutions of the underlying Toda-like discrete integrable hierarchy where $\{ t^{(1)}\},\ldots ,\{ t^{(q)}\}$ play the role of “evolution” parameters. This Toda-like discrete integrable hierarchy differs from the full generalized Toda lattice hierarchy [[@U-T]]{} in that the associated Toda matrices in the first hierarchy are semiinfinite and contain in general [*finite*]{} number of non-zero diagonals.
It turns out that, in order to identify the continuum integrable hierarchy which provides the exact solution for , we need the following subset of the associated linear system and the corresponding Lax (“zero-curvature”) representation from the Toda-like lattice system [[@BX]]{} : \_[nm]{} \_m = ł\_n , \_n = - $ {Q(1)}^r_{-}$\_[nm]{} \_m , \_n = - $ {Q(q)}^s_{-}$\_[nm]{} \_m\
Q(1) = \^r\_[(+)]{} , Q(1) , Q(1) = Q(1) , [Q(q)]{}\^s\_[-]{}\
Q(q) = Q\^r\_[(+)]{} , Q(q) , Q(q) = Q(q) , [Q(q)]{}\^s\_[-]{} In what follows it is convenient to introduce the short-hand notations: t\_r t\^[(1)]{}\_r , r=1,…,p\_1 ; \_s t\^[(q)]{}\_s , s=1,…,p\_q ; Q Q(1) ,Q(q) Further, there is a series of additional constraints (“coupling conditions”) relating $Q \equiv Q(1)$ and $\tQ \equiv Q(q)$. In the two-matrix model case ($q=2$) their explicit form is: -g Q , =\
Q\_[(-)]{} = - \_[s=1]{}\^[p\_2 -1]{} \_[s+1]{} \^s\_[(-)]{} - [1g]{} [t]{}\_1\
\_[(+)]{} = - \_[r=1]{}\^[p\_1 -1]{} \_[r+1]{} Q\^r\_[(+)]{} - [1g]{} [t]{}\_1 Here the subscripts $-/+$ denote lower/upper triangular parts, whereas $(+)/(-)$ denote upper/lower triangular plus diagonal parts. In the higher ($q \geq 3$) multi-matrix case the “coupling conditions” have much more intricate form (involving also the “intermediate” $Q(2),\ldots ,Q(q-1)$ matrices). However, their explicit form will not be needed to find the solution for $Z_N$ since we will be able to extract the relevant information solely from the discrete Lax system – and the relations expressing $Q \equiv Q(1),\,\tQ \equiv Q(q)$ in terms of orthogonal polynomial factors (see eqs. below).
The parametrization for the matrix elements of the Jacobi matrices $Q \equiv Q(1)$ and $\tQ \equiv Q(q)$ is as follows: Q\_[nn]{} \_[nn]{} = a\_0 (n) , \_[n,n+1]{} Q\_[n,n+1]{} =1\
[Q(1)]{}\_[n,n-k]{} Q\_[n,n-k]{} = a\_k (n) k=1,…, m(1) , m(1) = (p\_q -1) …(p\_2 -1)\
[Q(1)]{}\_[nm]{} Q\_[nm]{} = 0 m-n 2 , n-m m(1) +1\
[Q(q)]{}\_[nn]{} \_[nn]{} = b\_0 (n) , \_[n,n-1]{} \_[n,n-1]{} = R\_n\
[Q(q)]{}\_[n,n+k]{} \_[n,n+k]{} = b\_k (n) R\_[n+1]{}\^[-1]{} …R\_[n+k]{}\^[-1]{} k=1,…, m(q) , m(q) = (p\_[q-1]{} -1) …(p\_1 -1)\
[Q(q)]{}\_[nm]{} \_[nm]{} = 0 n-m 2 , m-n m(q) +1
In terms of the $Q \equiv Q(1),\,\tQ \equiv Q(q)$ matrix elements the partition function is expressed in the following way [[@BX]]{}: Z\_N = const \_[n=0]{}\^[N-1]{} h\_n\
a\_0 (n) = h\_n ,b\_0 (n) = h\_n ,R\_n =\
h\_n = Q\^r\_[nn]{} , h\_n = \^s\_[nn]{}\
Z\_N = \_[n=0]{}\^[N-1]{} a\_0 (n) , Z\_N = \_[n=0]{}\^[N-1]{} b\_0 (n)\
Z\_N = \_[n=0]{}\^[N-1]{} Q\^r\_[nn]{} , Z\_N = \_[n=0]{}\^[N-1]{} \^s\_[nn]{} where $h_n$ are the normalization factors in the nonlocally generalized orthogonal polynomial formalism [[@ortho-poly]]{} (using notations ) : h\_n \_[nm]{} = \_ \_ dłd P\_n (ł) { \_[r=1]{}\^[p\_1]{} ł\^r t\_r } (ł,; { t\^}, { g}) { \_[s=1]{}\^[p\_q]{} \^s \_s } [P]{}\_m ()\
(ł,; { t\^}, { g}) = \_ \_[=2]{}\^[q-1]{} d\_ { \_[=2]{}\^[q-1]{} \_[r\_=1]{}\^[p\_]{} t\^[(r\_)]{}\_[r\_]{} \_\^[r\_]{} + \_[=2]{}\^[q-2]{} g\_[,+1]{} \_\_[+1]{} + g\_[12]{} ł\_2 + g\_[q-1,q]{} \_[q-1]{} }\
{ t\^} $ t^{(2)},\ldots ,t^{(q-1)}$
As in the case of two-matrix model [[@office; @avoda]]{}, using the lattice equations of motion (eqs.– for $r=1,s=1$) we obtain the following important:
In the two-matrix model the explicit form of the coefficients $\a_s$ reads: $\a_s = - \frac{(s+1)}{g}\tit_{s+1}$.
Similarly, we have the dual statement with the rôles of $Q \equiv Q(1)$ and $\tQ \equiv Q(q)$ interchanged.
As an important consequence of , let us take its diagonal $00$-part and use the last eq. which yields: h\_0 = $ \sum_{s=1}^{m(1)} \a_s {{{\partial^{s} {}}\over{{\partial {\tit_1}}^{s} }}} + \a_0 $ h\_0 This equation is the only remnant of the constraints (“coupling conditions”) on the multi-matrix model $Q$-matrices which will be used in the sequel.
Based on our experience with the two-matrix model [[@office; @avoda]]{}, it turns out natural to introduce the fractional power of $Q \equiv Q(1)$ : = Q\^[1]{} \^[1]{} whose parametrization closely resembles that of $\tQ \equiv Q(q)$ : \_[nn]{} = \_0 (n) , \_[n,n-1]{} = \_n , \_[n,n+k]{} = \_k (n) \_[n+1]{}\^[-1]{} …\_[n+k]{}\^[-1]{} k 1\
\_[nm]{} = 0 n-m 2 From eqs. we find the following relation between the matrix elements of $\hQ \equiv Q(1)^{1\over {m(1)}}$ and $\tQ$ : \_n = \_[11]{} R\_n , \_0 (n) = \_[11]{} b\_0 (n) + \_[10]{} ,\_1 (n) = \_[11]{}\^2 b\_1 (n) + - \_[10]{}\^2 etc., with $\g$-coefficients as in –.
In order to identify the continuum hierarchy associated with the general $q$-matrix model, as a first step we reexpress, using , the Toda-like lattice hierarchy – as a single set of flow equations for $\hQ \equiv {Q(1)}^{1\over {m(1)}}$ : \^[m(1)]{}\_[nm]{} \_m = ł\_n , \_n = - ${\hQ}^s_{(-)}$\_[nm]{} \_m\
= , \^[s]{}\_[-]{} , s=1,…,p\_q , 2m(1), 3m(1), …, p\_1 m(1)\
t\_r \_[rm(1)]{} r=1,…, p\_1 Here, as in the two-matrix case [[@office; @avoda]]{}, we have introduced a new subset of evolution parameters $\lcurl {\hat t}_s \rcurl$ instead of $\lcurl {\ti t}_s \equiv t^{(q)}_s \rcurl$ defined as: = \_[=1]{}\^s \_[s]{} , s=1,…, m(q) with the same $\g_{s\s}$ as in . As a second step, one employs the Bonora-Xiong procedure [[@BX]]{} to get from the discrete Lax system – an equivalent continuum Lax system associated with a fixed lattice site $n$, where the continuum space coordinate is $x \equiv \htt_1$. Namely, the latter continuum integrable system is obtained by writing eqs. in more detail using the parametrization , : ł\_n &=& \_[n+1]{} + a\_0 (n) \_n + \_[k=1]{}\^[p\_2 -1]{}a\_k (n)\_[n-k]{}\
\_n &=& - \_n \_[n-1]{} and further using to express $\psi_{n \pm\ell}$ in terms of $\psi_n$ at a fixed lattice site $n$ in eq. and the higher evolution eqs. (for $s \geq 2$). Upon operator conjugation and an appropriate similarity transformation, it acquires the form (as before $x \equiv \htt_1$) : L(n) = [[$ L^{s\over {m(1)}}(n)$\_[(+)]{}]{} ,[L (n)]{}]{} , s=1,…,p\_q , 2m(1), 3m(1), …,p\_1 m(1)\
L(n) = D\_x\^[m(1)]{} + m(1) \_1 (n) D\_x\^[m(1)-2]{} + + \_[n+1]{} $ D_x - {\hb}_0 (n) $\^[-1]{} where $\hb_{0,1}(n), \hR_{n+1}$ are the matrix elements of $\hQ$ ,. Rewriting in the equivalent “eigenfunction” form: L(n) = D\_x\^[m(1)]{} + m(1) \_1 (n) D\_x\^[m(1)-2]{} + + (n+1) D\_x\^[-1]{} (n+1)\
(n+1) \_[n+1]{} { \_0 (n) } ,(n+1) { -\_0 (n) } and comparing with , we identify the continuum integrable hierarchy , describing equivalently the discrete multi-matrix model, as a constrained $\cKP_{m(1),1}$ hierarchy.
Exactly as in the two-matrix case [[@office; @avoda]]{}, lattice shifts $n \to n+1$ in the underlying discrete Toda lattice system, described by –, generate [[Darboux-Bäcklund]{} ]{}transformations in the continuum $\cKP_{m(1),1}$ hierarchy –. This is due to the fact that the latter continuum hierarchy preserves its form for any value of the discrete label $n$. The solutions for the eigenfunctions and $\t$-functions at each successive step of [[Darboux-Bäcklund]{} ]{}transformation is given explicitly, as particular cases of eqs.–, by: (n) =\
= \_[j=0]{}\^n (j) = W\_[n+1]{} (0), L(-1)(0), …, L(-1)\^n (0)\
(0) = $ L(-1)$\^[s]{}\_[(+)]{} (0) , s=1,…,p\_2 , 2(p\_2 -1), 3(p\_2 -1), …, p\_1 (p\_2 -1) where everything is expressed in terms of the eigenfunction $\Phi (0)$ of the “initial” Lax operator $L(-1)$. The difference with the two-matrix case is only the explicit form of the latter (recall $x \equiv \htt_1$) : L(-1) = e\^[\_[10]{} \_1]{} $ \sum_{s=0}^{m(1)} \a_s \g_{11}^{-s} D^s $ e\^[-\_[10]{} \_1]{} where the coefficients $\a_s ,\g_{10},\g_{11}$ have more complicated dependence on $\{ \htt_s\}$ than in the two-matrix case.
Exactly as in the two-matrix case, we obtain the relation between the $n$-th step DB eigenfunction $\Phi (n)$ and the orthogonal polynomial normalization factor $h_n$ which generalizes : (n) e\^[\_0 (n)]{} = h\_n \_[11]{}\^n \_1 \_[10]{} + (\^) ,\^ $ \htt_2 ,\ldots \htt_{m(q)}$ Substituting into and using the Wronskian formula we get: Z\_N = \_[n=0]{}\^[N-1]{} = [[[\^[i-1]{} ]{}]{}]{} $ L(-1)$\^[j-1]{} (0) e\^[-N$ \htt_1 \g_{10} + \vareps (\htt^{\pr})$]{} \_[11]{}\^[-]{}\
= [[[\^[i-1]{} ]{}]{}]{} $ e^{- \htt_1 \g_{10}} L(-1) e^{\htt_1 \g_{10}}$\^[j-1]{} h\_0 where we absorbed the $\g_{11}$-factors via changing ${{{\partial {}}\over{\partial {\htt_1}}}} \to {{{\partial {}}\over{\partial {\tit_1}}}}$ by the definition , [*i.e.*]{}, $\g_{11}^{-1} {{{\partial {}}\over{\partial {\htt_1}}}} = {{{\partial {}}\over{\partial {\tit_1}}}}$. Now, we find using and : $ e^{- \htt_1 \g_{10}} L(-1) e^{\htt_1 \g_{10}}$\^[j\_1]{} h\_0 = $ \sum_{s=0}^{m(1)} \a_s {{{\partial^{s} {}}\over{{\partial {\tit_1}}^{s} }}} $\^[j-1]{} h\_0 = [[[\^[j-1]{} ]{}]{}]{} h\_0 Substituting into yields the final result for the multi-matrix model partition function: Z\_N = [[[\^[i+j -2]{} [h\_0]{}]{}]{}]{} which is functionally the same as for the two-matrix model, however, with a more complicated expression for $h_0$ : h\_0 = \_ \_ dłd { \_[r=1]{}\^[p\_1]{} ł\^r t\_r } (ł,; { t\^}, { g}) { \_[s=1]{}\^[p\_q]{} \^s \_s } Eq. was previously obtained (see refs.[[@Moroz]]{}) from a different approach.\
E.N. gratefully acknowledges support and hospitality of the [*Deutscher Akademischer Austauschdienst*]{} and Prof. K.Pohlmeyer at the University of Freiburg. S.P. acknowledges support from the Ben-Gurion University, Beer-Sheva.
[99]{} L.D. Faddeev and L.A. Takhtajan, [*“Hamiltonian Methods in the Theory of Solitons”*]{}, Springer (1987) L. Faddeev, [*Integrable Models In 1+1 Dimensional Quantum Field Theory*]{}, Les Houches Lectures, session XXXIX (1982), J.-B. Zuber and R. Stora eds., Elsevier Sci. Publ. (1984);\
L. Faddeev, in [*Fields and Particles*]{}, H. Mitter and W. Schweiger eds., Springer (1990);\
P. Kulish and E. Sklyanin, in [*Lect. Notes in Phys.*]{} [**151**]{}, p.67, Springer (1982);\
L. Takhtajan, in [*Introduction to Quantum Group and Integrable Massive Models of Quantum Field Theory*]{}, Mo-Lin Ge and Bao-Heng Zhao eds., World Sci. (1990) M. Olshanetsky and A. Perelomov, ; [[*Theor. Mat. Phys.*]{} [**45**]{} (1980) 843]{}; [[*Phys. Reports*]{} [**71**]{} (1981) 313]{}; [*ibid*]{} [**94**]{} (1983);\
[*“Bäcklund transformations”*]{}, Lecture notes in Mathematics, Vol. 515, Eds. A. Dold and B. Eckmann, Springer (1976);\
[*“Solitons”*]{}, eds. R. Bullough and P. Caudrey, [*Topics in Current Physics*]{}, Springer (1980) A. Gorsky, I. Krichever, A. Marshakov, A. Mironov and A. Morozov, in [*hep-th/9505035*]{};\
H. Itoyama and A. Morozov, in [*hep-th/9511126; hep-th/9512161*]{};\
E. Martinec, in [*hep-th/9510204*]{};\
E. Martinec and N. Warner, in [*hep-th/9511052*]{} V. Kac and D. Peterson, [*Proc. Natl. Acad. Sci. USA*]{} [**78**]{} (1981) 3308;\
I. Bakas, [[*Commun. Math. Phys.*]{} [**134**]{} (1990) 487]{};\
C. Pope, L Romans and X. Shen, [[*Nucl. Phys.*]{} [**B339**]{} (1990) 191]{};\
A. Radul, [[*Functional Analysis and Its Application*]{} [**25**]{} (1991) 33]{};\
A. Radul and I. Vaysburd, [[*Phys. Lett.*]{} [**274B**]{} (1992) 317]{} A. Zamolodchikov, [[*Theor. Mat. Phys.*]{} [**65**]{} (1985) 1205]{};\
V. Fateev and S. Lukianov, [[*Int. J. Mod. Phys.*]{} [**A7**]{} (1992) 853]{};\
P. Bouwknegt and K. Schoutens, [[*Phys. Reports*]{} [**223**]{} (1993) 183]{} (a) J. Ellis, N. Mavromatos and D. Nanopoulos, [*Phys. Lett.*]{} [**2647B**]{} (1991) 465; [*ibid*]{} [**272B**]{} (1991) 261; [*hep-th/9403133*]{}\
(b) I. Klebanov and A. Polyakov, [*Mod. Phys. Lett.*]{} [**A6**]{} (1991) 3273; E. Witten, [*Nucl. Phys.*]{} [**B373**]{} (1992) 187\
(c) Y. Watanabe, Annali Matem. Pura Appl. [**136**]{} (1984) 77;\
A. Reiman and M. Semenov-Tyan-Shansky, J. Sov. Math. [**31**]{} (1985) 3399;\
K. Yamagishi, Phys. Lett. [**259B**]{} (1991) 436 ; F. Yu and Y.-S. Wu, [*ibid*]{} [**263B**]{} (1991) 220\
(d) Q. Han-Park, Phys. Lett. [**236B**]{} (1990) 429; [*ibid*]{} [**238B**]{} 208 ; [[*Int. J. Mod. Phys.*]{} [**A7**]{} (1992) 1415]{};\
K. Yamagishi and G. Chapline, Class. Quant. Grav. [**8**]{} (1991) 1\
(e) H. Ooguri and C. Vafa, [[*Nucl. Phys.*]{} [**B361**]{} (1991) 469]{}; [*ibid.*]{} [**B367**]{} (1991) 83 A. Cappelli, C, Trugenberger and G. Zemba, [[*Nucl. Phys.*]{} [**B396**]{} (1993) 465]{}; [[*Phys. Rev. Lett.*]{} [**72**]{} (1994) 1902]{}; [[*Nucl. Phys.*]{} [**B448 \[FS\]**]{} (1995) 470]{}; [*hep-th/9502021; hep-th/9502050*]{};\
D. Karabali, [[*Nucl. Phys.*]{} [**B428**]{} (1994) 1994]{};\
B. Sakita, 124;\
R. Ray and B. Sakita, [*Ann. Phys.*]{} [**230**]{} (1994) 131 B. Kostant, [*London Math.Soc.Lect.Notes*]{}, Ser. [**34**]{} (1979) 287;\
M. Adler, [*Inv. Math.*]{} [**50**]{} (1979) 219;\
W. Symes, [*Inv. Math.*]{} [**59**]{} (1980) 13;\
A.G. Reyman and M.A. Semenov-Tian-Shansky, [*Inv. Math.*]{} [**54**]{} (1979) 81; [*ibid*]{} [**63**]{} (1981) 423 S. Manakov, S. Novikov, L. Pitaevski and V. Zakharov, [*“Soliton Theory: The Inverse Problem”*]{}, Nauka, Moscow (1980) L.A. Dickey, [*“Soliton Equations and Hamiltonian Systems”*]{}, World Scientific, Singapore (1991) M. Douglas, [[*Phys. Lett.*]{} [**238B**]{} (1991) 176]{};\
E. Martinec, [[*Commun. Math. Phys.*]{} [**138**]{} (1991) 437]{};\
A. Gerasimov, Yu. Makeenko, A. Marshakov, A. Mironov, A. Morozov and A. Orlov, [[*Mod. Phys. Lett.*]{} [**A6**]{} (1991) 3079]{};\
R. Dijkgraaf, E. Verlinde and H. Verlinde, [[*Nucl. Phys.*]{} [**B348**]{} (1991) 435]{};\
L. Alvarez-Gaumé, C. Gomez and J. Lacki, [*Phys. Lett.*]{} [**B253**]{} (1991) 56 A. Orlov and E. Schulman, [[*Letters in Math. Phys.*]{} [**12**]{} (1986) 171]{};\
A. Orlov, in [*Plasma Theory and Nonlinear and Turbulent Processes in Physics*]{} (World Scientific, Singapore, 1988);\
L. Dickey, [[*Commun. Math. Phys.*]{} [**167**]{} (1995) 227]{} (also in [hep-th/9312015]{}); [[*Mod. Phys. Lett.*]{} [**A8**]{} (1993) 1357]{} (also in [*hep-th/9210155*]{})\
M. Adler, T. Shiota and P. van Moerbecke, [[*Phys. Lett.*]{} [**194A**]{} (1994) 33]{};\
L. Haine and E. Horozov, [*Bull. Sc. Math.*]{}, 2-e serie, [**117**]{} (1993) 485 V. Matveev and M. Salle, [*“Darboux Transformations and Solitons”*]{}, Springer-Verlag (1991);\
A. Leznov, A. Shabat and R. Yamilov, [[*Phys. Lett.*]{} [**174A**]{} (1993) 397]{};\
W. Oevel, [[*Physica*]{} [**A195**]{} (1993) 533]{};\
L.-L. Chau, J.C. Shaw and H.C. Yen, [[*Commun. Math. Phys.*]{} [**149**]{} (1992) 263]{} M. Semenov-Tyan-Shansky, [[*Functional Analysis and Its Application*]{} [**17**]{} (1983) 259]{} I. Bakas and E. Kiritsis, [*Int. J. Mod. Phys.*]{} [**A7**]{} \[Suppl. 1\] (1992) 55;\
F. Yu and Y.-S. Wu, [[*Nucl. Phys.*]{} [**B373**]{} (1992) 713]{} Y. Cheng, W. Strampp and B. Zhang, [[*Commun. Math. Phys.*]{} [**168**]{} (1995) 117]{} H. Aratyn, E. Nissimov and S. Pacheva, [[*Phys. Lett.*]{} [**201A**]{} (1995) 293]{} (also in [*hep-th/9501018*]{}) L. Bonora and C.S. Xiong, [[*Phys. Lett.*]{} [**317B**]{} (1993) 329]{};\
H. Aratyn, E. Nissimov, S. Pacheva and I. Vaysburd, [*Phys. Lett.*]{} [**294B**]{} (1992) 167 (also in [*hep-th/9209006*]{}) H. Aratyn, E. Nissimov and S. Pacheva, [*Phys. Lett.*]{} [**331B**]{} (1994) 82 (also in [*hep-th/9401058*]{}) H. Aratyn, E. Nissimov, S. Pacheva and A.H. Zimerman, [[*Int. J. Mod. Phys.*]{} [**A10**]{} (1995) 2537]{} (also in [*hep-th/9407117*]{}); H. Aratyn, [*Lectures at the VIII J.A. Swieca Summer School, 1995*]{} (also in [*hep-th/9503211*]{}) F. Yu, [[*Letters in Math. Phys.*]{} [**29**]{} (1993) 175]{} (also in [*hep-th/9301053*]{}) H. Aratyn, E. Nissimov and S. Pacheva, [*Phys. Lett.*]{} [**314B**]{} (1993) 41 (also in [*hep-th/9306035*]{}) E.L. Ince, [*Ordinary Differential Equations*]{}, chap. V, London, 1926;\
M.M. Crum, [*Quart. J. Math. Oxford*]{} [**6**]{} (1955) 121;\
M. Adler and J.Moser, [[*Commun. Math. Phys.*]{} [**61**]{} (1978) 1]{} L. Bonora and C.S. Xiong, [[*Int. J. Mod. Phys.*]{} [**A8**]{} (1993) 2973]{}; [[*Nucl. Phys.*]{} [**B405**]{} (1993) 191]{} (also in [*hep-th/9212070*]{}); [[*J. Math. Phys.*]{} [**35**]{} (1994) 5781]{} (also in [*hep-th/9311070*]{}); “[*Correlation functions of two-matrix models*]{}", in [*hep-th/9311089*]{} L. Bonora and C.S. Xiong, [*Phys. Lett.*]{} [**B347**]{} (1995) 41 (also in [*hep-th/9405004*]{}); [*Nucl. Phys.*]{} [**B434**]{} (1995) 408 (also in [*hep-th/9407141*]{});\
L. Bonora, C. Constantinidis and E. Vinteler, in [*hep-th/9511172*]{} D.Bessis, C. Itzykson and J.-B. Zuber, [*Adv. Appl. Math.*]{}[**1**]{} (1980) 109;\
C. Itzykson and J.-B. Zuber, [[*J. Math. Phys.*]{} [**21**]{} (1980) 411]{};\
M. Mehta, [[*Commun. Math. Phys.*]{} [**79**]{} (1981) 327]{};\
S. Chadha, G. Mahoux and M. Mehta, [[*J. Physics*]{} [**A14**]{} (1981) 579]{} K. Ueno and K. Takasaki, [*Adv. Stud. Pure Math.*]{} [**4**]{} (1984) 1 H. Aratyn, E. Nissimov and S. Pacheva, preprint BGU-95/19/Dec-PH (also in [*solv-int/9512nnn*]{}) A. Morozov, [*Physics Uspekhi*]{} [**37**]{} (1994) 1 (also in [*hep-th/9303139*]{}); in [*hep-th/9502091*]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'I apply commonly used regularization schemes to a multiloop calculation to examine the properties of the schemes at higher orders. I find complete consistency between the conventional dimensional regularization scheme and dimensional reduction, but I find that the four-dimensional helicity scheme produces incorrect results at next-to-next-to-leading order and singular results at next-to-next-to-next-to-leading order. It is not, therefore, a unitary regularization scheme.'
author:
- 'William B. Kilgore'
title: Regularization Schemes and Higher Order Corrections
---
Introduction
============
Dimensional regularization [@'tHooft:1972fi] is an elegant and efficient means of handling the divergences that arise in perturbation theory beyond the tree level. Among its many favorable qualities it respects gauge and Lorentz invariance and allows one to handle both ultraviolet and infrared divergences in the same manner. The application of dimensional regularization to different kinds of problems has led to the development of a variety of regularization schemes, which share the dimensional regularization of momentum integrals, but differ in their handling of external (or observed) states and of spin degrees of freedom.
The original formulation of dimensional regularization [@'tHooft:1972fi], known as the ’t Hooft-Veltman ([[HV]{}]{}) scheme, specifies that observed states are to be treated as four-dimensional, while internal states are to be treated as $D_m = 4
- 2\,{{\varepsilon}}$ dimensional. That is, both their momenta and spin degrees of freedom were to be continued from four to $D_m$ dimensions. It turns out that one has the freedom to choose the value of the trace of the Dirac unit matrix to take its canonical value of four, so fermions continue to have two spin degrees of freedom, even though their momenta are continued to $D_m$ dimensions. Internal gauge bosons, however, have $D_m-2$ spin degrees of freedom (internal massive gauge bosons have $D_m-1$ degrees of freedom).
A slight variation on the [[HV]{}]{} scheme has come to be called conventional dimensional regularization ([[CDR]{}]{}) [@Collins:Renorm]. In this variation, all particles and momenta are taken to be $D_m$ dimensional. This often turns out to be computationally more convenient, since one set of rules governs all interactions. This is particularly so when computing higher order corrections to theories subject to infrared sensitivities, like [[QCD]{}]{}. In the [[HV]{}]{} scheme, if two external states have infrared sensitive overlaps, they must be treated as internal, or $D_m$ dimensional states. In the [[CDR]{}]{}scheme, all states are already treated as $D_m$ dimensional, so there is no possibility of failing to properly account for infrared overlaps.
A third variation, called dimensional reduction ([[DRED]{}]{}) [@Siegel:1979wq], was devised for application to supersymmetric theories. In supersymmetry, it is essential that the number of bosonic degrees of freedom is exactly equal to the number of fermionic degrees of freedom. This requirement is violated in the [[HV]{}]{} and [[CDR]{}]{} schemes. In the [[DRED]{}]{} scheme, the continuation to $D_m$ dimensions is taken as a compactification from four dimensions. Thus, while space-time is taken to be four-dimensional and particles have the standard number of degrees of freedom, momenta span a $D_m$ dimensional vector space and momentum integrals are regularized dimensionally.
A fourth variation, the four-dimensional helicity ([[FDH]{}]{}) scheme [@Bern:1992aq; @Bern:2002zk], was developed primarily for use in constructing one-loop amplitudes from unitarity cuts. The most efficient building blocks for such calculations are tree-level helicity amplitudes, which necessarily have two spin degrees of freedom for both fermions and gauge bosons. The [[FDH]{}]{} scheme resembles the [[DRED]{}]{} scheme in that it regularizes momentum integrals dimensionally while maintaining the spin degrees of freedom of a four-dimensional theory (and therefore appears to be a valid supersymmetric regularization scheme [@Bern:2002zk]), but there are crucial differences, which I will discuss in detail.
The fact that the [[HV]{}]{} scheme respects the unitarity of the $S$-matrix was proven at its introduction [@'tHooft:1972fi]. The arguments which establish the validity of the [[HV]{}]{} scheme carry over to the [[CDR]{}]{} scheme and establish that it too is a valid regularization scheme. After some initial confusion over the proper renormalization procedure [@vanDamme:1984ig; @Capper:1980ns; @Jack:1993ws] for the [[DRED]{}]{} scheme, it was established that it too is a proper, unitary regularization scheme [@Jack:1993ws] and that it is indeed equivalent to the [[CDR]{}]{} scheme [@Jack:1994bn]. The [[FDH]{}]{} scheme has never been subjected to such stringent examination. It has been used successfully in a number of landmark next-to-leading order ([[NLO]{}]{}) calculations, but it has never been established whether it is a proper, unitary regularization scheme, or merely a set of shortcuts that allow expert users to obtain correct results.
In this paper, I will perform a well-known multiloop calculation in the various regularization schemes. I will show that while the [[HV]{}]{}and [[CDR]{}]{} scheme calculations yield the correct result and the [[DRED]{}]{}scheme calculation, while far more complicated is completely equivalent, the [[FDH]{}]{} scheme calculation yields incorrect results which inevitably violate unitarity at sufficiently high order. A detailed comparison of the various calculations identifies the source of the unitarity violations in the [[FDH]{}]{} scheme.
The plan of this paper is as follows: in section two, I will describe the test calculation to be performed and present the result to be obtained. In sections three, four and five, I will describe in detail the calculation to next-to-next-to-leading order ([[NNLO]{}]{}) as it is performed in the [[CDR]{}]{}, [[DRED]{}]{} and [[FDH]{}]{} schemes, respectively. In section six, I present partial results at [[N${}^3$LO]{}]{} which solidify the conclusion that the [[CDR]{}]{} and [[DRED]{}]{} schemes are equivalent and correct, but that the [[FDH]{}]{} scheme violates unitarity. In section seven, I will discuss my results and draw my conclusions.
The Test Environment
====================
To test the regularization schemes, I will calculate two quantities: the massless nonsinglet contributions to
1. the hadronic decay width of a fictitious neutral vector boson $V$, of mass $M_V$;
2. the single photon approximation to the total hadronic annihilation cross section for an electron – positron pair.
I will perform these calculations by means of the optical theorem, taking the imaginary part of the forward scattering amplitudes. In both cases, this means taking the imaginary part of the vacuum polarization tensor sandwiched between external states. Since the optical theorem is a direct consequence of the unitarity of the $S$-matrix, any unitary regularization scheme must give the same result, once one expands in terms of a standard coupling. To avoid complications involving prescriptions for handling $\gamma_5$ and the Levi-Civita tensor, I will take $V$ to have only vectorlike couplings. In this way, the vacuum polarization tensor for the $V$ boson will be identical to that of the off shell photon, up to coupling constants and so the [[QCD]{}]{} expansion of the two results will differ only by constant numerical factors.
Each regularization scheme will start from the same four-dimensional Lagrangian, $$\begin{split}
{\cal L} =& - \frac{1}{2}A^{a}_{\mu}{\left(}\partial^{\mu}\partial^{\nu}
(1-\xi^{-1}) - g^{\mu\nu}\Box{\right)}A^{a}_{\nu} -
g\,f^{abc}(\partial^\mu\,A^{a\,\nu})A^b_\mu\,A^c_\nu
- \frac{g^2}{4}f^{abc}\,f^{ade}\,A^{b\,\mu}\,A^{c\,\nu}\,A^d_\mu\,A^e_\nu\\
& + i\sum_f\,
\overline{\psi}_f^{i}{\left(}\delta_{ij}\slashed{\partial}
-i\,g\,t^{a}_{ij}\slashed{A}^a
-i\,g_V\,Q_f\slashed{V}{\right)}\,\psi_f^{j}
- \overline{c}^{a}\Box\,c^a
+ g\,f^{abc}{\left(}\partial_\mu\,\overline{c}^a{\right)}\,A^{b\,\mu}\,c^c\,,
\label{eqn::4dlagrange}
\end{split}$$ where $A^{a\,\mu}$ is the [[QCD]{}]{} gauge field, $V^\mu$ is the massive vector boson, $\psi_f$ is the quark field of flavor $f$, $\overline{c}^a$ and $c^a$ are the Faddeev-Popov ghost fields, $g$ is the [[QCD]{}]{} coupling, $g_V$ is the $V$ gauge coupling and $Q_f$ represents the charge of the quark flavor $f$ under the $V$ symmetry. I will not be computing nontrivial corrections in $g_V$, so there is no need to specify the $V$-self interaction parts of the Lagrangian.
![Sample diagrams of one-, two- and three-loop contributions to the vacuum polarization of $V$. \[fig::samples\]](vpol1.png "fig:"){width="4.cm"} ![Sample diagrams of one-, two- and three-loop contributions to the vacuum polarization of $V$. \[fig::samples\]](vpol2.png "fig:"){width="4.cm"} ![Sample diagrams of one-, two- and three-loop contributions to the vacuum polarization of $V$. \[fig::samples\]](vpol3a.png "fig:"){width="4.cm"} ![Sample diagrams of one-, two- and three-loop contributions to the vacuum polarization of $V$. \[fig::samples\]](vpol3b.png "fig:"){width="4.cm"}
The result to [[N${}^3$LO]{}]{} is well known [@Chetyrkin:1979bj; @Dine:1979qh; @Celmaster:1980ji; @Gorishnii:1988bc; @Gorishnii:1990vf], $$\begin{split}
\Gamma^V_{had} =& \Gamma^V_{0,\,{\rm had}}\,{\cal F}({{\alpha_{s}^{\overline{\rm{{MS}}}}}},Q^2=M_V^2)
\hskip 70pt \Gamma^V_{0,\,{\rm had}} = \frac{\alpha_V\,M_V}{3}\,N_c\sum_f\,Q_f^2\\
\sigma^{e^+\,e^-\to\ {\rm had}}(Q^2) =&
\sigma_0^{e^+\,e^-\to\ {\rm had}}(Q^2)\,{\cal F}({{\alpha_{s}^{\overline{\rm{{MS}}}}}},Q^2)
\hskip 40pt \sigma_0^{e^+\,e^-\to\ {\rm had}}(Q^2) =
\frac{4\,\pi\,\alpha^2}{3\,Q^2}\,N_c\sum_f\,Q_f^2\\
\label{eqn:knownresult}
\end{split}$$ and $$\begin{split}
{\cal F}({{\alpha_{s}^{\overline{\rm{{MS}}}}}},Q^2) =\hskip-12pt&\hskip 12pt\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,C_F\,\frac{3}{4}
{\left[}1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2} +
{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2{\left(}{{\beta_{1}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}
+ {{\beta_{0}^{\overline{\rm{{MS}}}}}}^{\,2}\,\ln^2\frac{\mu^2}{Q^2}{\right)}{\right]}\right.\\
& + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2{\left[}{\left(}-C_F^2\,\frac{3}{32}
+ C_F\,C_A\,{\left(}\frac{123}{32}
- \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{11}{16}
+ \frac{1}{2}\,\zeta_3{\right)}{\right)}\right.\\
&\qquad\qquad\times\left.{\left(}1 + 2{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}
\,\ln\frac{\mu^2}{Q^2}{\right)}{\right]}\\
& + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\left[}-C_F^3\frac{69}{128}
+ C_F^2\,C_A{\left(}-\frac{127}{64} - \frac{143}{16}\zeta_3
+ \frac{55}{4}\,\zeta_5{\right)}\right.\\
&\qquad\qquad+ C_F\,C_A^2{\left(}\frac{90445}{3456} - \frac{2737}{144}\,\zeta_3
- \frac{55}{24}\,\zeta_5{\right)}\\
&\qquad\qquad
+ C_F^2\,N_f{\left(}-\frac{29}{128} + \frac{19}{8}\,\zeta_3
- \frac{5}{2}\,\zeta_5{\right)}+ C_F\,C_A\,N_f{\left(}-\frac{485}{54} + \frac{56}{9}\,\zeta_3
+ \frac{5}{12}\,\zeta_5{\right)}\\
&\left.\left.\qquad\qquad
+ C_F\,N_f^2{\left(}\frac{151}{216} - \frac{19}{36}\,\zeta_3{\right)}- \frac{1}{4}\,\pi^2\,C_F\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}^2{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^4{\right)}\right\}\,.
\label{eqn:knownF}
\end{split}$$ To obtain the hadronic decay width at [[LO]{}]{}, [[NLO]{}]{} and [[NNLO]{}]{}, I need to compute the [[QCD]{}]{} corrections to the vacuum polarization of the $V$ (photon) at $1$, $2$ and $3$ loops, respectively. Sample diagrams are shown in [Fig.(\[fig::samples\])]{}.
Methods
-------
In each scheme, I will need to compute the vacuum polarization of $V$ and the necessary coupling renormalization constants. As a cross-check on the reliability of my calculational framework, I reproduce known results on the [[QCD]{}]{} $\beta$-functions and mass anomalous dimensions to three-loop order, as well as the three-loop [[QCD]{}]{} contributions to the $\beta$-function of $V$ (where needed).
In all calculations, I generate the contributing diagrams using QGRAF [@Nogueira:1993ex]. The symbolic algebra program FORM [@Vermaseren:2000nd] is used to implement the Feynman rules and perform algebraic manipulations to reduce the result to a set of Feynman integrals to be performed and their coefficients. The set of Feynman integrals are then reduced to master integrals using the program REDUZE [@Studerus:2009ye]. Using the method of Ref. [@Davydychev:1997vh], the vertex corrections can be expressed in terms of the same propagator integrals used to compute the vacuum polarization and wave function renormalizations. The complete set of master integrals at one, two and three loops are shown in [Fig.(\[fig::masters\])]{}.
a\) b)\
c)\
Most of the master integrals are trivial iterated-bubble diagrams and the others were evaluated long ago [@Chetyrkin:1980pr; @Kazakov:1983ns]. As an additional cross-check, the integral reduction and evaluation is also performed using the program MINCER[@Gorishnii:1989gt; @Larin:1991fz].
Notation
--------
The various schemes that I will consider span a variety of vector spaces, each with their own metric tensor. To establish some level of consistency, I will denote the metric tensor of classical four-dimensional space-time as $\eta^{\mu\nu}$; the metric tensor of the $D_m$ dimensional vector space in which momentum integrals are regularized will be denoted as $\hat{g}^{\mu\nu}$; and the metric tensor of the largest vector space will be denoted $g^{\mu\nu}$. Where it does not vanish, the complement of $\hat{g}^{\mu\nu}$ will be denoted as $\delta^{\mu\nu} = g^{\mu\nu} - \hat{g}^{\mu\nu}$. Similarly, the Dirac matrices $\gamma^\mu$, will be denoted $\gamma_{(4)}^\mu$ when they are strictly four-dimensional, $\hat{\gamma}^\mu$ when they span the $D_m$ dimensional space and $\bar{\gamma}^{\mu}$ in the space spanned by $\delta^{\mu\nu}$.
I will now present the details of the calculation in the [[CDR]{}]{}, [[DRED]{}]{}and [[FDH]{}]{} schemes.
Conventional Dimensional Regularization
=======================================
In the [[CDR]{}]{} scheme, the calculation is quite straightforward. The Lagrangian and Feynman rules are just the same as for a four-dimensional calculation, except that the Dirac matrices $\gamma^\mu$ and the metric tensor $g^{\mu\nu}$ have been extended to span a $D_m$ dimensional vector space. That is, $$\{\gamma^\mu,\gamma^\nu\} = 2\,g^{\mu\nu}\,,\qquad
g^{\mu\nu}\,g_{\mu\nu} = D_m\,,\qquad
\gamma^\mu\,\gamma_\mu = D_m\,,\qquad
g^{\mu\nu}\equiv\hat{g}^{\mu\nu}\,.
\label{eqn:cdrdirac}$$ The Dirac trace, ${\mathop{\rm Tr{\left[}{1}{\right]}}\nolimits} = 4$, retains its standard normalization.
Although $D_m$ is given the representation $D_m = 4-2\,{{\varepsilon}}$, the sign of ${{\varepsilon}}$ is not determined. If it is taken to be positive, so that $D_m<4$, then the Feynman integrals that one encounters are convergent under the rules of ultraviolet power counting. On the other hand, infrared power counting would prefer ${{\varepsilon}}<0 \Rightarrow D_m > 4$. In practice, the sign of ${{\varepsilon}}$ does not matter and it can be used to regularize both infrared and ultraviolet divergences. Regardless of the sign of ${{\varepsilon}}$, it is important that the vector space in which momenta take values is larger than the standard $3+1$ dimensional space-time. This means that the standard four-dimensional metric tensor $\eta^{\mu\nu}$ spans a smaller space than the $D_m$ dimensional metric tensor, and the four-dimensional Dirac matrices $\gamma^{0,1,2,3}$ form a subset of the full $\gamma^\mu$, $$g^{\mu\nu}\,g_{\mu}^{\rho} = g^{\nu\rho}\,,\qquad\qquad
g^{\mu\nu}\,\eta_{\mu}^{\rho} = \eta^{\nu\rho}\,,\qquad\qquad
\eta^{\mu\nu}\,\eta_{\mu}^{\rho} = \eta^{\nu\rho}\,.
\label{eqn:cdr4dim}$$ These considerations are of particular importance when considering chiral objects involving $\gamma_5$ and the Levi-Civita tensor, but will play a role in our discussion below.
Because the Dirac trace is unchanged, fermions still have exactly two degrees of freedom in the [[CDR]{}]{} scheme. Gauge bosons, however, acquire extra spin degrees of freedom in the $D_m$ dimensional vector space. The spin sum over polarization vectors in a physical (axial) gauge takes the form $$-g_{\mu\nu}\,\sum_{\lambda} \epsilon^{*\,\mu}(k,\lambda)\,\epsilon^{\nu}(k,\lambda)
= g_{\mu\nu}\,{\left(}g^{\mu\nu} - \frac{k^\mu\,n^\nu +
n^\mu\,k^\nu}{k\cdot n}{\right)}= D_m-2 = 2-2\,{{\varepsilon}}\,,
\label{eq::cdrspinsum}$$ where $n$ is the axial gauge reference vector. For massive vector bosons, the spin sum becomes $$-g_{\mu\nu}\,\sum_{\lambda} \epsilon^{*\,\mu}(k,\lambda)\,\epsilon^{\nu}(k,\lambda)
= g_{\mu\nu}\,{\left(}g^{\mu\nu} - \frac{k^\mu\,k^\nu}{M^2}{\right)}= D_m-1 = 3-2\,{{\varepsilon}}\,,
\label{eq::cdrmassivespinsum}$$
Renormalization
---------------
The renormalization constants in the [[CDR]{}]{} scheme are defined as $$\begin{split}
\Gamma^{(B)}_{AAA}&= Z_{1}\Gamma_{AAA}\,,\qquad
\psi^{(B)\,i}_f = Z^{\frac{1}{2}}_2\,\psi^{i}_f\,,\qquad
A^{(B)\, a}_\mu = Z^{\frac{1}{2}}_3\,A^{a}_\mu\\
\Gamma^{(B)}_{c\overline{c}A}&= \widetilde{Z}_1
\Gamma_{q\overline{q}A}\,,\qquad\
c^{(B)\,a} = \widetilde{Z}^{\frac{1}{2}}_3\,c^a\,,\qquad\ \
\overline{c}^{(B)\,a} = \widetilde{Z}^{\frac{1}{2}}_3\,\overline{c}^a\,,\\
\Gamma^{(B)}_{q\overline{q}A}&= Z_{1\,F}\Gamma_{q\overline{q}A}\,,\qquad
\xi^{(B)} = \xi\,Z_3\,,
\end{split}$$ where $\Gamma_{abc}$ represents the vertex function involving fields $a$, $b$ and $c$.
Although we treat the quark fields as massless, we can compute the mass anomalous dimension by introducing a fictitious scalar particle $\phi$ and computing the $\beta$-function of its Yukawa coupling to the quarks. The equivalence is clear from the standard model, where the Higgs Yukawa coupling and the fermion mass are proportional at leading electroweak order and must behave the same under [[QCD]{}]{}renormalization. For this purpose, I introduce one more renormalization constant, $\Gamma^{(B)}_{q\overline{q}\phi}=
Z_{1\,\phi}\Gamma_{q\overline{q}\phi}$. One can introduce a wave function renormalization for $\phi$, $Z_{3\,\phi}$, but it will not contribute because $Z_{3\,\phi} = 1 + {\cal O}(\alpha_\phi)$. Note also that I do not need to compute the [[QCD]{}]{} corrections to the $\beta$-function for $\alpha_V$, which will start at order $\alpha_V^2$ because of the Ward Identity.
In the $\overline{\rm MS}$ scheme, the couplings renormalize as $$\begin{split}
{{\alpha_s^{B}}}&={\left(}\frac{\mu^2\,e^{\gamma_E}}{4\,\pi}{\right)}^{{\varepsilon}}\,
Z_{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}\,{{\alpha_{s}^{\overline{\rm{{MS}}}}}}\,,\qquad Z_{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}} = \frac{Z_1^2}{Z_3^3}
= \frac{Z_{1\,F}^2}{Z_2^2\,Z_3}
= \frac{\widetilde{Z}_1^2}{\widetilde{Z}_3^2\,Z_3}\\
{{\alpha_\phi^{B}}}&={\left(}\frac{\mu^2\,e^{\gamma_E}}{4\,\pi}{\right)}^{{\varepsilon}}\,
Z_{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}\,{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}\,,\qquad Z_{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}
= \frac{Z_{1\,\phi}^2}{Z_2^2\,Z_{3\,\phi}}
\label{eqn::cdrrenorm}
\end{split}$$
The structure of the renormalization constants $Z_{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}$ and $Z_{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}$ is determined entirely by their lowest order ($1/{{\varepsilon}}$) poles, which in turn define the $\beta$-functions.
$$\begin{split}
{{\beta_{}^{\overline{\rm{{MS}}}}}}({{\alpha_{s}^{\overline{\rm{{MS}}}}}}) = \mu^2\frac{d}{d\,\mu^2}\frac{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}{\pi}
&= -{{\varepsilon}}\frac{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}{\pi}{\left(}1 + \frac{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}{Z_{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}}
\frac{\partial Z_{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}}{\partial{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}{\right)}^{-1}\\
&= -{{\varepsilon}}\frac{{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}{\pi} - \sum_{n=0}^\infty\,{{\beta_{n}^{\overline{\rm{{MS}}}}}}\,{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^{n+2}\\
{{\beta_{\phi}^{\overline{\rm{{MS}}}}}}({{\alpha_{s}^{\overline{\rm{{MS}}}}}}) = \mu^2\frac{d}{d\,\mu^2}\frac{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}{\pi}
&= -{\left(}{{\varepsilon}}\frac{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}{\pi} + \frac{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}{Z_{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}}
\frac{\partial Z_{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}}{\partial{{\alpha_{s}^{\overline{\rm{{MS}}}}}}}
\,{{\beta_{}^{\overline{\rm{{MS}}}}}}({{\alpha_{s}^{\overline{\rm{{MS}}}}}}){\right)}{\left(}1 + \frac{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}{Z_{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}}
\frac{\partial Z_{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}}{\partial{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}{\right)}^{-1}\\
&= -\frac{{{\alpha_{\phi}^{\overline{\rm{{MS}}}}}}}{\pi}{\left(}{{\varepsilon}}+ \sum_{n=0}^\infty
\,{{\beta_{\phi\,,n}^{\overline{\rm{{MS}}}}}}\,{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^{n+1}{\right)}\\
\label{eqn:cdrbetadef}
\end{split}$$
The mass anomalous dimension, $${{\gamma_{}^{\overline{\rm{{MS}}}}}}({{\alpha_{s}^{\overline{\rm{{MS}}}}}}) =
\frac{\mu^2}{m^{{{\overline{\rm MS}}}}}\frac{d}{d\mu^2}m^{{{\overline{\rm MS}}}}
= \sum_{n=0}^\infty -{{\gamma_{n}^{\overline{\rm{{MS}}}}}}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^{n+1}$$ is defined in terms of $m$, rather than $m^2$, with the result that ${{\gamma_{n}^{\overline{\rm{{MS}}}}}} = \frac{1}{2}{{\beta_{\phi\,,n}^{\overline{\rm{{MS}}}}}}$. The results for ${{\beta_{n}^{\overline{\rm{{MS}}}}}}$ and ${{\gamma_{n}^{\overline{\rm{{MS}}}}}}$ through three loops are given in [Appendix \[sec:cdrrenorm\]]{}.
Vacuum polarization in the [[CDR]{}]{} scheme
---------------------------------------------
The imaginary part of the unrenormalized vacuum polarization tensor in the [[CDR]{}]{} scheme is $$\begin{split}
\Im{\left[}\left.\Pi^{(B)}_{\mu\nu}(Q)\right|_{{{CDR}}}{\right]}&=
\frac{ -Q^2\,g_{\mu\nu} + Q_{\mu}Q_{\nu}}{3}
{{\alpha_{V}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{\varepsilon}}\left\{ \vphantom{{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}}\right.\\
& \hskip-50pt
1 + {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{{\varepsilon}}}
C_F\,{\left[}\frac{3}{4}
+ {{\varepsilon}}{\left(}\frac{55}{8} - 6\,\zeta_3{\right)}+ {{\varepsilon}}^2\,{\left(}\frac{1711}{48} - \frac{15}{4}\,\zeta_2 - 19\,\zeta_3
- 9\,\zeta_4{\right)}+ {\cal O}({{\varepsilon}}^3){\right]}\\
& \hskip-50pt
+{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}{\left[}\frac{1}{{{\varepsilon}}}{\left(}\,\frac{11}{16}C_F\,C_A
- \frac{1}{8}C_F\,N_f{\right)}\right.\\
&\hskip -30pt - \frac{3}{32}\,C_F^2 + C_F\,C_A{\left(}\frac{487}{48}
- \frac{33}{4}\zeta_3{\right)}+ C_F\,N_f{\left(}- \frac{11}{6} + \frac{3}{2}\,\zeta_3{\right)}\\
&\hskip -30pt +{{\varepsilon}}{\left(}C_F^2{\left(}- \frac{143}{32} - \frac{111}{8}\,\zeta_3
+ \frac{45}{2}\,\zeta_5{\right)}+ C_F\,C_A{\left(}\frac{50339}{576} - \frac{231}{32}\,\zeta_2 - \frac{109}{2}\,\zeta_3
- \frac{99}{8}\,\zeta_4 -
\frac{15}{4}\,\zeta_5{\right)}\right.\\
&\hskip -30pt\left.\left.\left.
+ C_F\,N_f{\left(}- \frac{4417}{288} + \frac{21}{16}\,\zeta_2
+ \frac{19}{2}\,\zeta_3 + \frac{9}{4}\,\zeta_4{\right)}{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^3{\right)}\right\}\,.
\label{eqn:impi0cdr}
\end{split}$$ Upon renormalizing the [[QCD]{}]{} coupling according to [Eq.(\[eqn::cdrrenorm\])]{}, setting ${{\alpha_{V}^{B}}}\to{{\alpha_{V}}}{\left(}\frac{\mu^2\,e^{\gamma_E}}{4\,\pi}{\right)}^{{\varepsilon}}$, and dropping terms of order $({{\varepsilon}})$, I obtain $$\begin{split}
\Im{\left[}\left.\Pi_{\mu\nu}(Q)\right|_{{{CDR}}}{\right]}&=
\frac{ -Q^2\,g_{\mu\nu} + Q_{\mu}Q_{\nu}}{3}
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2\,\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{3}{32}
+ C_F\,C_A\,{\left(}\frac{123}{32}
- \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{11}{16}
+ \frac{1}{2}\,\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,.
\label{eqn:impicdr}
\end{split}$$ In this way of performing the calculation, all of the [[QCD]{}]{} states that appear are internal states, so the [[HV]{}]{} scheme gives exactly the same result.
Total Decay rate and annihilation cross section in the [[CDR]{}]{} scheme {#sec:cdrdecann}
-------------------------------------------------------------------------
The decay rate and the annihilation cross section are determined by computing the imaginary part of the forward scattering amplitude. For the decay rate, this means attaching the polarization vector ${{\varepsilon}}^{\mu}(Q,\lambda)$ and its conjugate ${{\varepsilon}}^{\nu}(Q,\lambda)^{*}$ ($Q^2 = M_V^2$) and averaging over the spins, $$\Gamma^{{{CDR}}}_{V\to\ {\rm hadrons}} = \frac{1}{M_V}\frac{1}{N_{\rm spins}}
\sum_\lambda {{\varepsilon}}^{\mu}(Q,\lambda)\,\Im{\left[}\left.\Pi_{\mu\nu}(Q)
\right|_{{{CDR}}}{\right]}\,{{\varepsilon}}^{\nu}(Q,\lambda)^{*}\,,
\label{eqn:cdrdecay}$$ where $$\frac{1}{N_{\rm spins}}\sum_\lambda {{\varepsilon}}^{\mu}(Q,\lambda)\,
{{\varepsilon}}^{\nu}(Q,\lambda)^{*}
= \frac{1}{N_{\rm spins}}{\left(}-g^{\mu\nu} + \frac{Q^\mu\,Q^\nu}{M_V^2}{\right)}\,.
\label{eqn:cdrspinav}$$
Notice that because the imaginary part of the vacuum polarization tensor is finite, it does not matter whether the spin sum is taken in $D_m = 4 - 2\,{{\varepsilon}}$ dimensions as in the [[CDR]{}]{} scheme or in four dimensions as in the [[HV]{}]{} scheme as the difference is of order ${{\varepsilon}}$. The result is $$\begin{split}
\Gamma^{{{CDR}}}_{V\to\ {\rm hadrons}} =&
\frac{{{\alpha_{V}}}\,M_V}{3}\,N_c\,\sum_f\,Q_f^2\,
\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{3}{32}
+ C_F\,C_A\,{\left(}\frac{123}{32}
- \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{11}{16}
+ \frac{1}{2}\,\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,,
\label{eqn:gamcdrverify}
\end{split}$$ in agreement with [Eqs.(\[eqn:knownresult\]-\[eqn:knownF\])]{}.
For the annihilation cross section $\sigma_{e^+\,e^-\to\ {\rm
hadrons}}$, one attaches fermion bilinears to each end of the vacuum polarization tensor and averages over the spins. $$\sigma^{{{CDR}}}_{e^+\,e^-\to\ {\rm hadrons}} = \frac{2}{Q^2}\frac{e^2}{4}
\sum_{\lambda\,\lambda^{'}} \frac{{{{\left\langle\overline{v}(p_{e^+},\lambda)\left|\gamma^{\mu}\right|u(p_{e^-},\lambda^{'})\right\rangle}}}}{Q^2}\Im{\left[}\left.\Pi_{\mu\nu}(Q)
\right|_{{{CDR}},\,{{\alpha_{V}}}\to\alpha}{\right]}\frac{{{{\left\langle\overline{u}(p_{e^-},\lambda^{'})\left|\gamma^{\nu}\right|v(p_{e^+},\lambda)\right\rangle}}}}{Q^2}\,.
\label{eqn:cdrannav}$$ Because this is a forward scattering amplitude, the spinor bilinears can be combined into a trace, $$\frac{1}{2} \sum_{\lambda\,\lambda^{'}} {{{\left\langle
\overline{v}(p_{e^+},\lambda)\left|\gamma^{\mu}\right|u(p_{e^-},\lambda^{'})\right\rangle}}}
{{{\left\langle\overline{u}(p_{e^-},\lambda^{'})\left|\gamma^{\nu}\right|v(p_{e^+},\lambda)\right\rangle}}} = \frac{1}{2}{\mathop{\rm Tr{\left[}{\slashed{p}_{e^+}\,
\gamma^{\mu}\slashed{p}_{e^-}\,\gamma^{\nu}}{\right]}}\nolimits}
= {\left(}-Q^2\,g^{\mu\,\nu} + Q^\mu\,Q^\nu{\right)}\,,$$ where the last identification results from the fact that $Q^\mu =
p_{e^-}^{\mu} + p_{e^+}^{\mu}$,$p_{e^-}\cdot\,Q =
p_{e^+}\cdot\,Q = Q^2/2$. The result is $$\begin{split}
\sigma^{{{CDR}}}_{e^+\,e^-\to\ {\rm hadrons}} =&
\frac{4\pi\,\alpha^2}{3\,Q^2}\,N_c\,\sum_f\,Q_f^2\,
\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{3}{32}
+ C_F\,C_A\,{\left(}\frac{123}{32}
- \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{11}{16}
+ \frac{1}{2}\,\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,,
\label{eqn:sigcdrverify}
\end{split}$$ again in agreement with [Eqs.(\[eqn:knownresult\]-\[eqn:knownF\])]{}.
Thus, I have established that I can reproduce the known results in the [[CDR]{}]{} scheme through three-loop order, which is a strong check on my computational framework.
Dimensional Reduction
=====================
In dimensional reduction, one starts from standard four-dimensional space-time and compactifies to a [*smaller*]{} vector space of dimension $D_m = 4 - 2\,{{\varepsilon}}< 4$ in which momenta take values. The particles in the spectrum, however, retain the spin degrees of freedom of four dimensions. That is, both fermions and gauge bosons still have two degrees of freedom. This is by design, of course, since it is required by supersymmetry. All Dirac algebra can be treated as four-dimensional. However, now the four-dimensional metric tensor $\eta^{\mu\nu}$ spans a larger space than the $D_m$ dimensional metric $\hat{g}^{\mu\nu}$ that might arise from tensor momentum integrals, $$\hat{g}^{\mu\nu}\,\eta_{\mu}^{\rho} = \hat{g}^{\nu\rho}\,.
\label{eqn::metricdred}$$
There is also a very serious consequence of the fact that the $D_m$ dimensional vector space is smaller than four-dimensional space-time. The Ward Identity only applies to the $D_m$ dimensional vector space. This means that the $2\,{{\varepsilon}}$ spin degrees of freedom that are not protected by the Ward Identity must renormalize differently than the $2-2\,{{\varepsilon}}$ degrees of freedom that are protected. In supersymmetric theories, the supersymmetry provides the missing Ward Identity which demands that the $2\,{{\varepsilon}}$ spin degrees of freedom be treated as gauge bosons. In nonsupersymmetric theories, however, they must be considered to be distinct particles, with distinct couplings and renormalization properties. It is common to refer to these extra degrees of freedom as “${{\varepsilon}}$-scalars” or as “evanescent” degrees of freedom.
Once the evanescent degrees of freedom (which I will label $A_e^{a\,\tilde\mu}$, to distinguish them from the gluons, $A^{a\,\mu}$) are recognized as independent particles, it is apparent that their couplings are also independent, not only of the [[QCD]{}]{}coupling, but of one another. That is, the coupling $g_e$ of the evanescent gluons to the quarks is not only distinct from $g$, the coupling of [[QCD]{}]{}, but is also distinct from $\lambda_i$, the quartic couplings of the evanescent gluons to themselves. (The quartic gauge coupling of [[QCD]{}]{} splits into three independent quartic couplings of the evanescent gluons.) Note that the massive vector boson $V^\mu$ also has evanescent degrees of freedom, $V_e^{\tilde\mu}$, which couple to quarks with strength $g_{Ve}$.
Thus, the Lagrangian in the [[DRED]{}]{} scheme becomes: $$\begin{split}
{\cal L} =& - \frac{1}{2}A^{a}_{\mu}{\left(}\partial^{\mu}\partial^{\nu}
(1-\xi^{-1}) - \hat{g}^{\mu\nu}\Box{\right)}A^{a}_{\nu}
- g\,f^{abc}(\partial^\mu\,A^{a\,\nu})A^b_\mu\,A^c_\nu
- \frac{g^2}{4}f^{abc}\,f^{ade}\,A^{b\,\mu}\,A^{c\,\nu}\,A^d_\mu\,A^e_\nu\\
& + i\sum_f\,
\overline{\psi}_f^{i}{\left(}\delta_{ij}\slashed{\partial}
-i\,g\,t^{a}_{ij}\slashed{A}^a
-i\,g_V\,Q_f\slashed{V}{\right)}\,\psi_f^{j}
- \overline{c}^{a}\Box\,c^a
+ g\,f^{abc}{\left(}\partial_\mu\,\overline{c}^a{\right)}\,A^{b\,\mu}\,c^c\\
& + \frac{1}{2}A_{e\,\tilde\mu}^{a}\Box\ A_e^{a\,\tilde\mu}
- g\,f^{abc}(\partial^\mu\,A_e^{a\,\tilde\nu})A^b_\mu\,A^c_{e\,\tilde\nu}
+ \frac{g^2}{2}f^{abc}\,f^{adf}\,A^{b\,\mu}\,A_e^{c\,\tilde\nu}\,
A^d_\mu\,A_{e\,\tilde\nu}^f
- \frac{1}{4}\sum_i \lambda_i\,H_i^{bcdf}\,A_e^{b\,\tilde\mu}\,A_e^{c\,\tilde\nu}\,
A_{e\,\tilde\mu}^d\,A_{e\,\tilde\nu}^f\\
& + \sum_f\,\overline{\psi}_f^{i}{\left(}g_e\,t^{a}_{ij}\slashed{A}_e^a
+ g_{Ve}\,Q_f\slashed{V}_e{\right)}\,\psi_f^{j}\,.
\label{eqn::dredlagrange}
\end{split}$$ As mentioned above, the quartic coupling of the evanescent gluons splits into three terms, which mix under renormalization. One can choose the tensors $H_i^{bcde}$ to be [@Harlander:2006xq] $$\begin{split}
H_1^{bcde} =& \frac{1}{2}{\left(}f^{abc}\,f^{ade} + f^{abe}\,f^{adc}{\right)}\\
H_2^{bcde} =& \delta^{bc}\delta^{de} + \delta^{bd}\delta^{ce}
+ \delta^{be}\delta^{cd}\\
H_3^{bcde} =& \frac{1}{2}{\left(}\delta^{bc}\delta^{de} + \delta^{be}\delta^{cd}{\right)}- \delta^{bd}\delta^{ce}\,,
\label{eqn::dredquartic}
\end{split}$$ Although the quartic couplings enter the $\beta$-functions and anomalous dimension at three loops and are essential to the renormalization program, they do not explicitly contribute to the calculation at hand.
Now that the correct spectrum has been identified, one must carefully consider the renormalization program. The naïve application of the principle of minimal subtraction leads to the violation of unitarity [@vanDamme:1984ig]. Because the contributions of evanescent states and couplings to scattering amplitudes are weighted by a factor ${{\varepsilon}}$, the leading one-loop contribution is finite and therefore not subtracted. As one proceeds to higher orders, there is a mismatch among the counterterms such that the renormalization program fails to remove all of the ultraviolet singularities.
A successful renormalization program for the [[DRED]{}]{}scheme [@Jack:1993ws; @Jack:1994bn] applies the principle of minimal subtraction to the evanescent Green functions (that is, Green functions with external evanescent states) themselves. At each order, the renormalization scheme renders the evanescent Green functions finite. Since evanescent Green functions enter into the scattering amplitudes of physical particles at order ${{\varepsilon}}$ and they are rendered finite by renormalization, they never contribute to physical scattering amplitudes.
The evanescent coupling still contributes to Green functions with only physical external states, but the contribution is rendered finite by the prescribed renormalization program [@Jack:1993ws; @Jack:1994bn; @Harlander:2006rj; @Harlander:2006xq]. Because the evanescent coupling, $\alpha_e$ renormalizes differently than the gauge coupling $\alpha_s$, the two cannot be identified, even at the end of the calculation. One can choose a renormalization point where the two coincide, but they evolve differently under renormalization group transformations and their values will diverge as one moves away from the renormalization point.
Still, the evanescent coupling is essentially a fictitious quantity and one finds that if one computes a physical quantity in the [[DRED]{}]{}scheme and then converts the running couplings of the [[DRED]{}]{} scheme to those of a scheme such as [[CDR]{}]{} that has no evanescent couplings, the factors of $\alpha_e$ drop out [@Harlander:2006rj; @Harlander:2006xq].
Renormalization
---------------
The renormalization constants in the [[DRED]{}]{} scheme are defined as $$\begin{split}
\Gamma^{(B)}_{AAA}&= Z_{1}\Gamma_{AAA}\,,\qquad
\psi^{(B)\,i}_f = Z^{\frac{1}{2}}_2\,\psi^{i}_f\,,\qquad
A^{(B)\, a}_\mu = Z^{\frac{1}{2}}_3\,A^{a}_\mu\\
\Gamma^{(B)}_{c\overline{c}A}&= \widetilde{Z}_1
\Gamma_{q\overline{q}A}\,,\qquad\
c^{(B)\,a} = \widetilde{Z}^{\frac{1}{2}}_3\,c^a\,,\qquad\ \
\overline{c}^{(B)\,a} = \widetilde{Z}^{\frac{1}{2}}_3\,\overline{c}^a\,,\\
\Gamma^{(B)}_{q\overline{q}A}&= Z_{1\,F}\Gamma_{q\overline{q}A}\,,\qquad
\xi^{(B)} = \xi\,Z_3\,,\\
\Gamma^{(B)}_{q\overline{q}e}&=Z_{1\,e}\Gamma_{q\overline{q}e}\,,\qquad
A^{(B)\,a}_{e\, \mu} = Z^{\frac{1}{2}}_{3\,e}\,A^{a}_{e\,\mu}\,,\qquad
\Gamma^{(B)\,i}_{eeee} = Z^{i}_{1\,eeee}\,\Gamma^{i}_{eeee}\,,\\
\Gamma^{(B)}_{q\overline{q}V_e}&=Z_{1\,Ve}\Gamma_{q\overline{q}V_e}\,,\qquad
V^{(B)}_{e\, \mu} = Z^{\frac{1}{2}}_{3\,Ve}\,V_{e\,\mu}\,.
\label{eqn:dredrenormconst}
\end{split}$$ In addition, I again introduce the fictitious scalar that allows me to compute the mass anomalous dimension for massless quarks. Note that while the Ward Identity protects $\alpha_V$ from leading [[QCD]{}]{}corrections, it does not protect $\alpha_{Ve}$. That is why I need to introduce renormalization constants for the vertex and wave-function and why I need to compute the $\beta$-function of $\alpha_{Ve}$.
In the ${{\overline{\rm DR}}}$ scheme (modified minimal subtraction in the [[DRED]{}]{}scheme), the couplings renormalize as $$\begin{split}
{{\alpha_s^{B}}}&={\left(}\frac{\mu^2\,e^{\gamma_E}}{4\,\pi}{\right)}^{{\varepsilon}}\,
Z_{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}\,{{\alpha_{s}^{\overline{\rm{{DR}}}}}}\,,\qquad Z_{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}} = \frac{Z_1^2}{Z_3^3}
= \frac{Z_{1\,F}^2}{Z_2^2\,Z_3}
= \frac{\widetilde{Z}_1^2}{\widetilde{Z}_3^2\,Z_3}\,,\\
{{\alpha_e^{B}}}&={\left(}\frac{\mu^2\,e^{\gamma_E}}{4\,\pi}{\right)}^{{\varepsilon}}\,
Z_{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}\,{{\alpha_{e}^{\overline{\rm{{DR}}}}}}\,,\qquad Z_{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}
= \frac{Z_{1\,e}^2}{Z_2^2\,Z_{3\,e}}\,,\\
{{\alpha_{Ve}^{B}}}&={\left(}\frac{\mu^2\,e^{\gamma_E}}{4\,\pi}{\right)}^{{\varepsilon}}\,
Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}\,{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}\,,\qquad Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}
= \frac{Z_{1\,Ve}^2}{Z_2^2\,Z_{3\,Ve}}\,,\\\
{{\alpha_\phi^{B}}}&={\left(}\frac{\mu^2\,e^{\gamma_E}}{4\,\pi}{\right)}^{{\varepsilon}}\,
Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}\,{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}\,,\qquad Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}
= \frac{Z_{1\,\phi}^2}{Z_2^2\,Z_{3\,\phi}}\,.
\label{eqn:dredrenorm}
\end{split}$$
and the $\beta$-functions are given by $$\begin{split}
{{\beta_{}^{\overline{\rm{{DR}}}}}} = \mu^2\frac{d}{d\,\mu^2}\frac{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}{\pi}
&= -{\left(}{{\varepsilon}}\frac{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}{\pi}
+\frac{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}}\frac{\partial
Z_{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}\,{{\beta_{e}^{\overline{\rm{{DR}}}}}}
+\frac{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}}\frac{\partial
Z_{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}}{\partial{{\eta_{i}^{{{\overline{\rm DR}}}}}}}\,{{\beta_{\eta_i}^{\overline{\rm{{DR}}}}}}{\right)}{\left(}1 + \frac{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}}
\frac{\partial Z_{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}{\right)}^{-1}\\
&= -{{\varepsilon}}\frac{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}{\pi} - \sum_{i,j,k,l,m}\,{{\beta_{ijklm}^{\overline{\rm{{DR}}}}}}
\,{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{i}\,{{{\left(}\frac{\alpha_{e}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{j}\,{{{\left(}\frac{\eta_{1}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{k}
\,{{{\left(}\frac{\eta_{2}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{l}\,{{{\left(}\frac{\eta_{3}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{m}\\
{{\beta_{e}^{\overline{\rm{{DR}}}}}} = \mu^2\frac{d}{d\,\mu^2}\frac{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}{\pi}
&= -{\left(}{{\varepsilon}}\frac{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}{\pi}
+\frac{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}}\frac{\partial
Z_{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}\,{{\beta_{}^{\overline{\rm{{DR}}}}}}
+\frac{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}}\frac{\partial
Z_{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}}{\partial{{\eta_{i}^{{{\overline{\rm DR}}}}}}}\,{{\beta_{\eta_i}^{\overline{\rm{{DR}}}}}}{\right)}{\left(}1 + \frac{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}}
\frac{\partial Z_{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}{\right)}^{-1}\\
&= -{{\varepsilon}}\frac{{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}{\pi} - \sum_{i,j,k,l,m}\,{{\beta_{e,\,ijklm}^{\overline{\rm{{DR}}}}}}
\,{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{i}\,{{{\left(}\frac{\alpha_{e}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{j}\,{{{\left(}\frac{\eta_{1}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{k}
\,{{{\left(}\frac{\eta_{2}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{l}\,{{{\left(}\frac{\eta_{3}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{m}\\
{{\beta_{Ve}^{\overline{\rm{{DR}}}}}} = \mu^2\frac{d}{d\,\mu^2}\frac{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}{\pi}
&= -{\left(}{{\varepsilon}}\frac{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}{\pi}
+\frac{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}}\frac{\partial
Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}\,{{\beta_{}^{\overline{\rm{{DR}}}}}}
+\frac{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}}\frac{\partial
Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}\,{{\beta_{e}^{\overline{\rm{{DR}}}}}}
+\frac{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}}\frac{\partial
Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}}{\partial{{\eta_{i}^{{{\overline{\rm DR}}}}}}}\,{{\beta_{\eta_i}^{\overline{\rm{{DR}}}}}}{\right)}\\
&\hskip 100pt \times
{\left(}1 + \frac{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}}
\frac{\partial Z_{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}{\right)}^{-1}\\
&= -\frac{{{\alpha_{Ve}^{\overline{\rm{{DR}}}}}}}{\pi}{\left(}{{\varepsilon}}+ \sum_{i,j,k,l,m}\,{{\beta_{Ve,\,ijklm}^{\overline{\rm{{DR}}}}}}
\,{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{i}\,{{{\left(}\frac{\alpha_{e}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{j}\,{{{\left(}\frac{\eta_{1}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{k}
\,{{{\left(}\frac{\eta_{2}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{l}\,{{{\left(}\frac{\eta_{3}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{m}{\right)}\\
{{\beta_{\phi}^{\overline{\rm{{DR}}}}}} = \mu^2\frac{d}{d\,\mu^2}\frac{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}{\pi}
&= -{\left(}{{\varepsilon}}\frac{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}{\pi} + \frac{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}}
\frac{\partial Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}
\,{{\beta_{}^{\overline{\rm{{DR}}}}}} + \frac{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}}
\frac{\partial Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{e}^{\overline{\rm{{DR}}}}}}}
\,{{\beta_{e}^{\overline{\rm{{DR}}}}}} + \frac{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}}
\frac{\partial Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}}{\partial{{\eta_{i}^{{{\overline{\rm DR}}}}}}}
\,{{\beta_{\eta_i}^{\overline{\rm{{DR}}}}}}{\right)}\\
&\hskip 100pt \times{\left(}1 + \frac{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}{Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}}
\frac{\partial Z_{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}}{\partial{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}{\right)}^{-1}\\
&= -\frac{{{\alpha_{\phi}^{\overline{\rm{{DR}}}}}}}{\pi}{\left(}{{\varepsilon}}+ \sum_{i,j,k,l,m}\,{{\beta_{\phi,\,ijklm}^{\overline{\rm{{DR}}}}}}
\,{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{i}\,{{{\left(}\frac{\alpha_{e}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{j}\,{{{\left(}\frac{\eta_{1}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{k}
\,{{{\left(}\frac{\eta_{2}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{l}\,{{{\left(}\frac{\eta_{3}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{m}{\right)}\\
\label{eqn:dredbetadef}
\end{split}$$
Through three-loop order, the $\eta_i$ do not contribute to the [[QCD]{}]{}$\beta$-function, ${{\beta_{}^{\overline{\rm{{DR}}}}}}$, nor to the vacuum polarization of $V$ (or $V_e$). To three-loop order, I find agreement with known results [@Harlander:2006rj; @Harlander:2006xq] and derive new results for the $\beta$-function of $\alpha_{Ve}$. The coefficients of the $\beta$-functions and anomalous dimensions are given in [Appendix \[sec:dredrenorm\]]{}.
By comparing ${{\beta_{Ve,\,\,}^{\overline{\rm{{DR}}}}}}$ and ${{\gamma_{}^{\overline{\rm{{DR}}}}}}$ in [Eqs.(\[eqn:dredgamma\]-\[eqn:dredbetave\])]{}, we see that the term “${{\varepsilon}}$-scalar” is a misnomer. If the evanescent part of $V$ were a true scalar, its $\beta$-function would coincide (but for a factor of $2$) with the mass anomalous dimension. The pure ${{\alpha_{s}^{\overline{\rm{{DR}}}}}}$ terms do coincide, because there is no nonvanishing contraction of the Lorentz indices of the evanescent $V$ and those of the gluons. Because there are contractions between the Lorentz indices of the evanescent $V$ and those of the evanescent gluons, however, terms involving ${{\alpha_{e}^{\overline{\rm{{DR}}}}}}$ do not agree.
Calculations in the [[DRED]{}]{} scheme naturally produce results in terms of ${{\alpha_{s}^{\overline{\rm{{DR}}}}}}$ while the standard result has been expressed in terms of ${{\alpha_{s}^{\overline{\rm{{MS}}}}}}$. One can always convert one renormalized coupling to another. The rule for converting ${{\alpha_{s}^{\overline{\rm{{DR}}}}}}\to{{\alpha_{s}^{\overline{\rm{{MS}}}}}}$, derived in Refs. [@Kunszt:1993sd; @Harlander:2006rj], is $${{\alpha_{s}^{\overline{\rm{{DR}}}}}}= {{\alpha_{s}^{\overline{\rm{{MS}}}}}}{\left[}1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\frac{C_A}{12}
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\frac{11}{72}C_A^2
- {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}{{{\left(}\frac{\alpha_{e}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}\frac{C_F\,N_f}{16}
+ \ldots{\right]}\label{eqn:drbartomsbar}$$ When the result is expressed in terms of ${{\alpha_{s}^{\overline{\rm{{MS}}}}}}$, all ${{\alpha_{e}^{\overline{\rm{{DR}}}}}}$ terms drop out.
Vacuum polarization in the [[DRED]{}]{} scheme
----------------------------------------------
In the [[DRED]{}]{} scheme, there are two independent transverse vacuum polarization tensors, $$\Im{\left[}\left.\Pi^{(B)}_{\mu\nu}(Q)\right|_{{{DRED}}}{\right]}=
\frac{-Q^2\,\hat{g}_{\mu\nu} + Q_{\mu}Q_{\nu}}{3}\,
\Im{\left[}\left.\Pi^{(B)}_{A}(Q)\right|_{{{DRED}}}{\right]}- Q^2\,\frac{\delta_{\mu\nu}}{2\,{{\varepsilon}}}\,\Im{\left[}\left.\Pi^{(B)}_{B}(Q)\right|_{{{DRED}}}{\right]}\,,$$ where $$\begin{split}
\Im{\left[}\left.\Pi^{(B)}_{A}(Q)\right|_{{{DRED}}}{\right]}&=
{{\alpha_{V}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{\varepsilon}}\left\{ \vphantom{{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}}\right.\\
& \hskip-50pt
1 + {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{{\varepsilon}}}
C_F\,{\left[}\frac{3}{4} + {{\varepsilon}}{\left(}\frac{51}{8}
- 6\,\zeta_3{\right)}+ {{\varepsilon}}^2\,{\left(}\frac{497}{16} - \frac{15}{4}\zeta_2
- 15\,\zeta_3 - 9\,\zeta_4{\right)}+ {\cal O}({{\varepsilon}}^3){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_e^{B}}{\pi}{\right)}}}{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{{\varepsilon}}}
C_F\,{\left[}-{{\varepsilon}}\,\frac{3}{4} - {{\varepsilon}}^2\,\frac{29}{8}
+ {\cal O}({{\varepsilon}}^3){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^2\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}
{\left[}\frac{1}{{{\varepsilon}}}{\left(}\frac{11}{16}C_F\,C_A
- \frac{1}{8}C_F\,N_f{\right)}- \frac{3}{32}C_F^2
+ {\left(}\frac{77}{8} - \frac{33}{4}\zeta_3{\right)}\,C_F\,C_A
- {\left(}\frac{7}{4} - \frac{3}{2}\zeta_3{\right)}\,C_F\,N_f \right.\\
&\hskip -30pt +{{\varepsilon}}{\left(}C_F^2{\left(}- \frac{141}{32} - \frac{111}{8}\,\zeta_3
+ \frac{45}{2}\,\zeta_5{\right)}+ C_F\,C_A{\left(}\frac{15301}{192} - \frac{231}{32}\,\zeta_2 - \frac{193}{4}\,\zeta_3
- \frac{99}{8}\,\zeta_4 -
\frac{15}{4}\,\zeta_5{\right)}\right.\\
&\hskip -20pt\left.\left.
+ C_F\,N_f{\left(}- \frac{1355}{96} + \frac{21}{16}\,\zeta_2
+ \frac{17}{2}\,\zeta_3 +
\frac{9}{4}\,\zeta_4{\right)}{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_e^{B}}{\pi}{\right)}}}^2\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}
{\left[}\frac{3}{4}C_F^2 - \frac{3}{8}C_F\,C_A
+ \frac{3}{16}C_F\,N_f
- {{\varepsilon}}{\left(}\frac{47}{8}C_F^2 - \frac{11}{4}C_F\,C_A
+ \frac{7}{4}C_F\,N_f{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}\\
& \hskip-50pt\phantom{1}\left.
+ {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}{{{\left(}\frac{\alpha_e^{B}}{\pi}{\right)}}}\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}
{\left[}- \frac{9}{8}C_F^2 - {{\varepsilon}}{\left(}\frac{141}{16}C_F^2
+ \frac{21}{16}C_F\,C_A{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}+ {\cal O}{\left(}{\left(}\frac{{{\alpha_s^{B}}}}{\pi},\frac{{{\alpha_e^{B}}}}{\pi}{\right)}^3{\right)}\right\}\,,
\label{eqn:impi0Adred}
\end{split}$$ and $$\begin{split}
\Im{\left[}\left.\Pi^{(B)}_{B}(Q)\right|_{{{DRED}}}{\right]}&=
{{\alpha_{Ve}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{\varepsilon}}\left\{{{\varepsilon}}+ 2\,{{\varepsilon}}^2 + {\left(}4-\frac{3}{2}\zeta_2{\right)}{{\varepsilon}}^3
+ {\cal O}({{\varepsilon}}^4)\right.\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{{\varepsilon}}}
C_F\,{\left[}\frac{3}{2} + {{\varepsilon}}\frac{29}{4} + {{\varepsilon}}^2\,{\left(}\frac{227}{8} - \frac{15}{2}\zeta_2
- 6\,\zeta_3{\right)}+ {\cal O}({{\varepsilon}}^3){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_e^{B}}{\pi}{\right)}}}{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{{\varepsilon}}}
C_F\,{\left[}- 1 - 4\,{{\varepsilon}}- {{\varepsilon}}^2{\left(}\frac{27}{2} - 5\,\zeta_2{\right)}+ {\cal O}({{\varepsilon}}^3){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^2\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}
{\left[}\frac{1}{{{\varepsilon}}}{\left(}\frac{9}{8}C_F^2 + \frac{11}{16}C_F\,C_A
- \frac{1}{8}C_F\,N_f{\right)}+ \frac{279}{32}C_F^2
+ \frac{199}{32}C_F\,C_A - \frac{17}{16}C_F\,N_f\right. \\
&\hskip -30pt +{{\varepsilon}}{\left(}C_F^2{\left(}\frac{3139}{64} - \frac{189}{16}\,\zeta_2
- \frac{45}{4}\,\zeta_3{\right)}+ C_F\,C_A{\left(}\frac{2473}{64} - \frac{231}{32}\,\zeta_2
- \frac{75}{8}\,\zeta_3{\right)}\right.\\
&\hskip -20pt\left.\left.
+ C_F\,N_f{\left(}- \frac{207}{32} + \frac{21}{16}\,\zeta_2
+ \frac{3}{2}\,\zeta_3{\right)}{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}{{{\left(}\frac{\alpha_e^{B}}{\pi}{\right)}}}\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}
{\left[}- \frac{1}{{{\varepsilon}}}\frac{9}{4}C_F^2 - \frac{129}{8}C_F^2
- \frac{3}{8}C_F\,C_A\right.\\
& \hskip-30pt\phantom{1}\left.
- {{\varepsilon}}{\left(}{\left(}\frac{671}{8}
- \frac{189}{8}\zeta_2 - 9\,\zeta_3{\right)}\,C_F^2
+ \frac{53}{16}C_F\,C_A{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_e^{B}}{\pi}{\right)}}}^2\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}
{\left[}\frac{1}{{{\varepsilon}}}{\left(}C_F^2 - \frac{1}{4}C_F\,C_A + \frac{1}{8}C_F\,N_f{\right)}+ \frac{13}{2}C_F^2 - \frac{3}{2}C_F\,C_A
+ \frac{15}{16}C_F\,N_f\right.\\
& \hskip-50pt\phantom{1}\left.
+ {{\varepsilon}}{\left(}{\left(}31 - \frac{21}{2}\zeta_2
- \frac{3}{4}\zeta_3{\right)}\,C_F^2 - {\left(}\frac{53}{8}
- \frac{21}{8}\zeta_2 - \frac{3}{8}\zeta_3{\right)}\,C_F\,C_A
+ {\left(}\frac{157}{32} - \frac{21}{16}\zeta_2{\right)}\,C_F\,N_f{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}\\
& \hskip-50pt\phantom{1}\left.
+ {\cal O}{\left(}{\left(}\frac{{{\alpha_s^{B}}}}{\pi},\frac{{{\alpha_e^{B}}}}{\pi}{\right)}^3{\right)}\right\}\,,\\
\label{eqn:impi0Bdred}
\end{split}$$ where $\displaystyle
{\cal O}{\left(}{\left(}\frac{{{\alpha_s^{B}}}}{\pi},\frac{{{\alpha_e^{B}}}}{\pi}{\right)}^3{\right)}$ denotes terms for which the sum of the powers of $\displaystyle{\left(}\frac{{{\alpha_s^{B}}}}{\pi}{\right)}$ and $\displaystyle{\left(}\frac{{{\alpha_e^{B}}}}{\pi}{\right)}$ is at least three.
Upon renormalization according to [Eq.(\[eqn:dredrenorm\])]{} and expanding in terms of ${{\alpha_{s}^{\overline{\rm{{MS}}}}}}$ according to [Eq.(\[eqn:drbartomsbar\])]{}, I find that $$\begin{split}
\Im{\left[}\left.\Pi_{A}(Q)\right|_{{{DRED}}}{\right]}&=
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2 \left\{
1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}\frac{3}{4}C_F
{\left[}1+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}{{\beta_{20}^{\overline{\rm{{DR}}}}}}\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-50pt\phantom{1}\left.
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^2{\left[}-C_F^2\frac{3}{32}
+ C_F\,C_A{\left(}\frac{121}{32} - \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f{\left(}-\frac{11}{16} + \frac{1}{2}\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{\left(}\frac{{{\alpha_s^{B}}}}{\pi},\frac{{{\alpha_e^{B}}}}{\pi}{\right)}^3{\right)}\right\}\\
&\hskip-45pt={{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2\,
\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{3}{32}
+ C_F\,C_A\,{\left(}\frac{123}{32} - \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{11}{16} + \frac{1}{2}\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,,\\
\Im{\left[}\left.\Pi_{B}(Q)\right|_{{{DRED}}}{\right]}&={\cal O}({{\varepsilon}})\,.
\label{eqn:impidred}
\end{split}$$
Total Decay rate and annihilation cross section in the [[DRED]{}]{} scheme {#sec:dreddecann}
--------------------------------------------------------------------------
As in the [[CDR]{}]{} scheme, the decay rate and annihilation cross section are determined from the imaginary part of the forward scattering amplitude. $$\Gamma^{{{DRED}}}_{V\to\ {\rm hadrons}} = \frac{1}{M_V}\frac{1}{N_{\rm spins}}
\sum_\lambda {{\varepsilon}}^{\mu}(Q,\lambda)\,\Im{\left[}\left.\Pi_{\mu\nu}(Q)
\right|_{{{DRED}}}{\right]}\,{{\varepsilon}}^{\nu}(Q,\lambda)^{*}\,,
\label{eqn:dreddecay}$$ where $$\frac{1}{N_{\rm spins}}\sum_\lambda {{\varepsilon}}^{\mu}(Q,\lambda)\,
{{\varepsilon}}^{\nu}(Q,\lambda)^{*}
= \frac{1}{3}{\left(}-\hat{g}^{\mu\nu} + \frac{Q^\mu\,Q^\nu}{M_V^2}
- \delta^{\mu\nu}{\right)}\,.
\label{eqn:dredspinav}$$ The evanescent part of the spin average contracts only with the $\Pi_{B}(Q)$ term, which has been renormalized to be of order $({{\varepsilon}})$, so that the result is $$\begin{split}
\Gamma^{{{DRED}}}_{V\to\ {\rm hadrons}} =&
\frac{{{\alpha_{V}}}\,M_V}{3}\,N_c\,\sum_f\,Q_f^2\,
\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{3}{32}
+ C_F\,C_A\,{\left(}\frac{123}{32}
- \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{11}{16}
+ \frac{1}{2}\,\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,,
\label{eqn:gamdredverify}
\end{split}$$ just like in the [[CDR]{}]{} calculation.
For the annihilation cross section $\sigma_{e^+\,e^-\to\ {\rm
hadrons}}$, one attaches fermion bilinears to each end of the vacuum polarization tensor and averages over the spins. $$\begin{split}
\sigma^{{{DRED}}}_{e^+\,e^-\to\ {\rm hadrons}} &= \frac{2}{Q^2}\frac{e^2}{4}
\sum_{\lambda\,\lambda^{'}} \frac{{{{\left\langle\overline{v}(p_{e^+},\lambda)\left|\hat{\gamma}^{\mu}\right|u(p_{e^-},\lambda^{'})\right\rangle}}}}{Q^2}\Im{\left[}\left.\Pi_{\mu\nu}(Q)
\right|_{{{DRED}},\,{{\alpha_{V}}}\to\alpha}{\right]}\frac{{{{\left\langle\overline{u}(p_{e^-},\lambda^{'})\left|\hat{\gamma}^{\nu}\right|v(p_{e^+},\lambda)\right\rangle}}}}{Q^2}\\
&+ \frac{2}{Q^2}\frac{e_{\ell\,e}^2}{4}
\sum_{\lambda\,\lambda^{'}} \frac{{{{\left\langle\overline{v}(p_{e^+},\lambda)\left|\bar{\gamma}^{\mu}\right|u(p_{e^-},\lambda^{'})\right\rangle}}}}{Q^2}\Im{\left[}\left.\Pi_{\mu\nu}(Q)
\right|_{{{DRED}},\,{{\alpha_{V}}}\to\alpha}{\right]}\frac{{{{\left\langle\overline{u}(p_{e^-},\lambda^{'})\left|\bar{\gamma}^{\nu}\right|v(p_{e^+},\lambda)\right\rangle}}}}{Q^2}\,,
\label{eqn:dredannav}
\end{split}$$ where $e_{\ell\,e}$ represents the coupling of the evanescent photon to the electron. Combining the spinor bilinears into traces, $$\begin{split}
\frac{1}{2} &\sum_{\lambda\,\lambda^{'}} {{{\left\langle
\overline{v}(p_{e^+},\lambda)\left|\hat{\gamma}^{\mu}\right|u(p_{e^-},\lambda^{'})\right\rangle}}}
{{{\left\langle\overline{u}(p_{e^-},\lambda^{'})\left|\hat{\gamma}^{\nu}\right|v(p_{e^+},\lambda)\right\rangle}}} = \frac{1}{2}{\mathop{\rm Tr{\left[}{\slashed{p}_{e^+}\,
\gamma^{\mu}\slashed{p}_{e^-}\,\gamma^{\nu}}{\right]}}\nolimits}
= {\left(}-Q^2\,\hat{g}^{\mu\,\nu} + Q^\mu\,Q^\nu{\right)}\\
\frac{1}{2} &\sum_{\lambda\,\lambda^{'}} {{{\left\langle
\overline{v}(p_{e^+},\lambda)\left|\bar{\gamma}^{\mu}\right|u(p_{e^-},\lambda^{'})\right\rangle}}}
{{{\left\langle\overline{u}(p_{e^-},\lambda^{'})\left|\bar{\gamma}^{\nu}\right|v(p_{e^+},\lambda)\right\rangle}}} = \frac{1}{2}{\mathop{\rm Tr{\left[}{\slashed{p}_{e^+}\,
\bar{\gamma}^{\mu}\slashed{p}_{e^-}\,\bar{\gamma}^{\nu}}{\right]}}\nolimits}
= {\left(}-Q^2\,\delta^{\mu\,\nu}{\right)}\,
\end{split}$$ The final result is $$\begin{split}
\sigma^{{{DRED}}}_{e^+\,e^-\to\ {\rm hadrons}} =&
\frac{4\pi\,\alpha^2}{3\,Q^2}\,N_c\,\sum_f\,Q_f^2\,
\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{3}{32}
+ C_F\,C_A\,{\left(}\frac{123}{32}
- \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{11}{16}
+ \frac{1}{2}\,\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,,
\label{eqn:sigdredverify}
\end{split}$$ again in agreement with [Eqs.(\[eqn:knownresult\]-\[eqn:knownF\])]{}. As promised, under the [[DRED]{}]{} scheme renormalization program, evanescent Green functions are rendered finite by renormalization and contribute to scattering amplitudes at order $({{\varepsilon}})$. Also as promised, the results are completely equivalent to those of the [[CDR]{}]{} scheme.
The Four-Dimensional Helicity Scheme
====================================
In the four-dimensional helicity scheme, one defines an enlarged vector space of dimensionality $D_m = 4-2\,{{\varepsilon}}$, in which loop momenta take values, as in the [[CDR]{}]{} scheme. In addition, one defines a still larger vector space, of dimensionality $D_s=4$, in which internal spin degrees of freedom take values. The precise rules for the [[FDH]{}]{}scheme are given in Ref. [@Bern:2002zk]. They are:
1. As in ordinary dimensional regularization, all momentum integrals are integrated over $D_m$ dimensional momenta. Metric tensors resulting from tensor integrals are $D_m$ dimensional.
2. All “observed” external states are taken to be four-dimensional, as are their momenta and polarization vectors. This facilitates the use of helicity states for observed particles.
3. All “unobserved” or internal states are treated as $D_s$ dimensional, and the $D_s$ dimensional vector space is taken to be larger than the $D_m$ dimensional vector space. Unobserved states include virtual states inside of loops, virtual states inside of trees as well as external states that have infrared sensitive overlaps with other external states.
4. Both the $D_s$ and $D_m$ dimensional vector spaces are larger than the standard four-dimensional space-time, so that contraction of four-dimensional objects with $D_m$ or $D_s$ dimensional objects yields only four-dimensional components.
To keep track of the many vector spaces and their overlapping domains, I give the result of the contractions of the various metric tensors with one another, $$\begin{split}
g^{\mu\nu}\,g_{\mu\nu} &= D_s\,,\qquad\
\hat{g}^{\mu\nu}\,\hat{g}_{\mu\nu} = D_m\,,\qquad
\eta^{\mu\nu}\,\eta_{\mu\nu} = 4\,,\qquad
\delta^{\mu\nu}\,\delta_{\mu\nu} = D_x = D_s - D_m\\
g^{\mu\nu}\hat{g}^{\rho}_{\nu} &= \hat{g}^{\mu\rho}\,,\qquad
g^{\mu\nu}\eta^{\rho}_{\nu} = \eta^{\mu\rho}\,,\qquad
\hat{g}^{\mu\nu}\eta^{\rho}_{\nu} = \eta^{\mu\rho}\,,\\
g^{\mu\nu}\delta^{\rho}_{\nu} &= \delta^{\mu\rho}\,,\qquad
\hat{g}^{\mu\nu}\delta^{\rho}_{\nu} = 0\,,\qquad\quad\
\eta^{\mu\nu}\delta^{\rho}_{\nu} = 0\,.\\
\label{eqn::fdhmetrics}
\end{split}$$
Like the [[HV]{}]{} scheme, the [[FDH]{}]{} scheme treats observed states as four-dimensional. In inclusive calculations, however, where there are infrared overlaps among external states, the external states are taken to be $D_s$ dimensional in the infrared regions.
As in the [[DRED]{}]{} scheme, spin degrees of freedom take values in a vector space that is larger than that in which momenta take values. It would seem, therefore, that the same remarks regarding the Ward Identity and the conclusion that the $D_x = D_s - D_m$ dimensional components of the gauge fields and their couplings must be considered as distinct from the $D_m$ dimensional gauge fields and couplings would apply. That is not, however, how the [[FDH]{}]{} scheme is used. All field components in the $D_s$ dimensional space are treated as gauge fields and no distinction is made between the couplings. It is common, however, to define an interpolating scheme, the “$\delta_R$” scheme, in which $D_s = 4 - 2\,{{\varepsilon}}\,\delta_R$. The parameter $\delta_R$ interpolates between the [[HV]{}]{} scheme ($\delta_R=1$) and the [[FDH]{}]{}scheme ($\delta_R=0$). Using this scheme gives one a handle on the impact of the evanescent degrees of freedom on the result, but not on the impact of a distinct evanescent coupling.
It is claimed [@Bern:2002zk] that the essential difference between the [[FDH]{}]{} and [[DRED]{}]{} schemes is that in the former $D_m > 4$, while in the latter $D_m < 4$. It must be this difference, then, that allows for the very different handling of the evanescent couplings and degrees of freedom. We shall see what impact this choice has in the calculation and discussion below.
Renormalization
---------------
I will not give detailed results for the renormalization parameters of the [[FDH]{}]{} scheme. There is no point in doing so because, as I will show, the rules of the [[FDH]{}]{} scheme enumerated in the previous section are not consistent with a successful renormalization program. The first sign that there is a problem with the renormalization program comes in the computation of the one-loop renormalization constants. In particular, the gluon vacuum polarization tensor splits into two independent components, $\Pi_A^{\mu\nu} =
\Pi_A(Q^2)\,{\left(}(-Q^2\hat{g}^{\mu\nu} + Q^\mu\,Q^\nu{\right)}$ and $\Pi_B^{\mu\nu} = \Pi_B(Q^2)\,\delta^{\mu\nu}$, both of which are singular. This is a clear warning that what the [[FDH]{}]{} scheme calls the gluon is in fact two distinct sets of degrees of freedom. If I ignore $\Pi_B$ and just renormalize $\Pi_A$, I find the usual result that $${{\beta_{0}^{\overline{\rm{{FDH}}}}}} = \frac{11}{12}C_A - \frac{1}{6}N_f\,.$$ Note that I also get this result if I take the spin average (trace) of the full vacuum polarization tensor. Because $\Pi_B$ is weighted by a factor of $2\,{{\varepsilon}}$, its contribution to the spin average is not singular. Because the leading order term in the quantities being calculated is of order one, and the [[NLO]{}]{} term of order $\alpha_s$, this result for the one-loop $\beta$-function is all that is needed to compute the renormalized cross section at [[NNLO]{}]{}. Furthermore, the many [[NLO]{}]{} results that have been obtained using the [[FDH]{}]{} scheme have all renormalized using the above result for ${{\beta_{0}^{\overline{\rm{{FDH}}}}}}$.
When I try to proceed to the two-loop beta function, I find that both $\Pi_A$ and $\Pi_B$ contribute singular terms to the spin-averaged vacuum polarization, while if I again ignore $\Pi_B$ and renormalize $\Pi_A$, I obtain the usual value for $\beta_1$, $${{\beta_{1}^{\overline{\rm{{FDH}}}}}} = \frac{17}{24}C_A^2 - \frac{5}{24}C_A\,N_f
- \frac{1}{8}C_F\,N_f\,.$$ This seems to be the choice made in Ref. [@Bern:2002zk] as they quote only the result for terms proportional to $Q^{\mu}Q^{\nu}$, which would be part of my $\Pi_A$. Since the standard lore has been that ${{\alpha_{s}^{\overline{\rm{{FDH}}}}}}$ and ${{\alpha_{s}^{\overline{\rm{{DR}}}}}}$ coincide, at least through second order corrections, this seems to be the most reasonable choice. Furthermore, it means that the conversion to ${{\alpha_{s}^{\overline{\rm{{MS}}}}}}$ will be [@Kunszt:1993sd; @Bern:2002zk] $${{\alpha_{s}^{\overline{\rm{{FDH}}}}}}= {{\alpha_{s}^{\overline{\rm{{MS}}}}}}{\left[}1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\frac{C_A}{12}
+ \ldots{\right]}\label{eqn:fdhtomsbar}$$ As it turns out, it does not matter what choice one makes as even the one-loop result for ${{\beta_{0}^{\overline{\rm{{FDH}}}}}}$, which seems safe if only because it is familiar, leads to the violation of unitarity.
Vacuum polarization in the [[FDH]{}]{} scheme
---------------------------------------------
Leaving aside the question of renormalization beyond one-loop, I will proceed with the calculation of the $V$-boson vacuum polarization. In performing calculations in the [[FDH]{}]{} scheme, it becomes apparent that the results are identical, term-by-term. to the calculation in the [[DRED]{}]{} scheme, except that the evanescent gluons are identified as gluons and the coupling $\alpha_e$ is set to $\alpha_s$. Therefore I find that $$\Im{\left[}\left.\Pi^{(B)}_{\mu\nu}(Q)\right|_{{{FDH}}}{\right]}=
\frac{-Q^2\,\hat{g}_{\mu\nu} + Q_{\mu}Q_{\nu}}{3}\,
\Im{\left[}\left.\Pi^{(B)}_{A}(Q)\right|_{{{FDH}}}{\right]}- Q^2\,\frac{\delta_{\mu\nu}}{2\,{{\varepsilon}}}\,\Im{\left[}\left.\Pi^{(B)}_{B}(Q)\right|_{{{FDH}}}{\right]}\,,$$ where $$\begin{split}
\Im{\left[}\left.\Pi^{(B)}_{A}(Q)\right|_{{{FDH}}}{\right]}&=
{{\alpha_{V}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{\varepsilon}}\left\{ \vphantom{{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}}\right.\\
& \hskip-50pt
1 + {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{{\varepsilon}}}
C_F\,{\left[}\frac{3}{4} + {{\varepsilon}}{\left(}\frac{45}{8}
- 6\,\zeta_3{\right)}+ {{\varepsilon}}^2\,{\left(}\frac{439}{16} - \frac{15}{4}\zeta_2
- 15\,\zeta_3 - 9\,\zeta_4{\right)}+ {\cal O}({{\varepsilon}}^3){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^2\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}
{\left[}\frac{1}{{{\varepsilon}}}{\left(}\frac{11}{16}C_F\,C_A
- \frac{1}{8}C_F\,N_f{\right)}- \frac{15}{32}C_F^2
+ {\left(}\frac{37}{4} - \frac{33}{4}\zeta_3{\right)}\,C_F\,C_A
- {\left(}\frac{25}{16} - \frac{3}{2}\zeta_3{\right)}\,C_F\,N_f \right.\\
&\hskip -30pt +{{\varepsilon}}{\left(}C_F^2{\left(}- \frac{235}{32} - \frac{111}{8}\,\zeta_3
+ \frac{45}{2}\,\zeta_5{\right)}+ C_F\,C_A{\left(}\frac{14521}{192} - \frac{231}{32}\,\zeta_2 - \frac{193}{4}\,\zeta_3
- \frac{99}{8}\,\zeta_4 -
\frac{15}{4}\,\zeta_5{\right)}\right.\\
&\hskip -20pt\left.\left.\left.
+ C_F\,N_f{\left(}- \frac{1187}{96} + \frac{21}{16}\,\zeta_2
+ \frac{17}{2}\,\zeta_3 +
\frac{9}{4}\,\zeta_4{\right)}{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^3{\right)}\right\}\,,
\label{eqn:impi0Afdh}
\end{split}$$ and $$\begin{split}
\Im{\left[}\left.\Pi^{(B)}_{B}(Q)\right|_{{{FDH}}}{\right]}&=
{{\alpha_{V}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{\varepsilon}}\left\{ \vphantom{{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}}\right.\\
& \hskip-50pt {{\varepsilon}}+
{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{{{\varepsilon}}}
C_F\,{\left[}\frac{1}{2} + {{\varepsilon}}\frac{13}{4} + {{\varepsilon}}^2\,{\left(}\frac{119}{8} - \frac{5}{2}\zeta_2
- 6\,\zeta_3{\right)}+ {\cal O}({{\varepsilon}}^3){\right]}\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^2\,{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{2\,{{\varepsilon}}}
{\left[}\frac{1}{{{\varepsilon}}}{\left(}-\frac{1}{8}C_F^2 + \frac{7}{16}C_F\,C_A{\right)}- \frac{29}{32}C_F^2
+ \frac{139}{32}C_F\,C_A - \frac{1}{8}C_F\,N_f\right. \\
&\hskip -30pt +{{\varepsilon}}{\left(}C_F^2{\left(}-\frac{245}{64} + \frac{21}{16}\,\zeta_2
- 3\,\zeta_3{\right)}+ C_F\,C_A{\left(}\frac{1837}{64} - \frac{147}{32}\,\zeta_2
- 9\,\zeta_3{\right)}\right.\\
&\hskip -20pt\left.\left.\left.
+ C_F\,N_f{\left(}- \frac{25}{16} + \frac{3}{2}\,\zeta_3{\right)}{\right)}+ {\cal O}({{\varepsilon}}^2){\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^3{\right)}\right\}\,.\\
\label{eqn:impi0Bfdh}
\end{split}$$ Upon renormalizing such that $${{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}\to{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{-{{\varepsilon}}}
{\left(}1 - \frac{{{\beta_{0}^{\overline{\rm{{FDH}}}}}}}{{{\varepsilon}}}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}{\right)}\,,\qquad\qquad
{{\alpha_{V}^{B}}}\to{{\alpha_{V}}}{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{-{{\varepsilon}}}\,,$$ I find that $$\begin{split}
\Im{\left[}\left.\Pi_{A}(Q)\right|_{{{FDH}}}{\right]}&=
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2 \left\{
1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}\frac{3}{4}C_F
{\left[}1+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}{{\beta_{0}^{\overline{\rm{{FDH}}}}}}\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-50pt\phantom{1}\left.
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}^2{\left[}-C_F^2\frac{15}{32}
+ C_F\,C_A{\left(}\frac{131}{32} - \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f{\left(}-\frac{5}{8} + \frac{1}{2}\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}^3{\right)}\right\}\\
&\hskip-45pt={{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2\,
\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{15}{32}
+ C_F\,C_A\,{\left(}\frac{133}{32} - \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{5}{8} + \frac{1}{2}\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,,\\
\Im{\left[}\left.\Pi_{B}(Q)\right|_{{{FDH}}}{\right]}&=
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2 \left\{
{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}\frac{1}{2}C_F
{\left[}1+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}{{\beta_{0}^{\overline{\rm{{FDH}}}}}}\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}^2{\left[}\frac{1}{{{\varepsilon}}}{\left(}- C_F^2\frac{1}{8} - C_F\,C_A\frac{1}{48}
+ C_F\,N_f\frac{1}{12}{\right)}{\left(}1 + 3{{\varepsilon}}\ln\frac{\mu^2}{Q^2}{\right)}\right.\\
&\left.\left.
- C_F^2\frac{29}{32}
+ C_F\,C_A\frac{131}{96}
+ C_F\,N_f\frac{5}{12}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}^3{\right)}\right\}\\
&\hskip-45pt=
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2 \left\{
{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\frac{1}{2}C_F
{\left[}1+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}{{\beta_{0}^{\overline{\rm{{FDH}}}}}}\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-50pt\phantom{1}
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2{\left[}\frac{1}{{{\varepsilon}}}{\left(}- C_F^2\frac{1}{8} - C_F\,C_A\frac{1}{48}
+ C_F\,N_f\frac{1}{12}{\right)}{\left(}1 + 3{{\varepsilon}}\ln\frac{\mu^2}{Q^2}{\right)}\right.\\
&\left.\left.
- C_F^2\frac{29}{32}
+ C_F\,C_A\frac{45}{32}
+ C_F\,N_f\frac{5}{12}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,.
\label{eqn:impifdh}
\end{split}$$
Total Decay rate and annihilation cross section in the [[FDH]{}]{} scheme {#sec:fdhdecann}
-------------------------------------------------------------------------
The results of the vacuum polarization calculation look to be disastrous as $\Pi_B$ is singular at order $\alpha_s^2$. However, the rules of the [[FDH]{}]{} scheme, enumerated above, specify that external states are taken to be four-dimensional. This means that the spin average of the vector polarizations is $$\frac{1}{N_{\rm spins}}\sum_\lambda {{\varepsilon}}^{\mu}(Q,\lambda)\,
{{\varepsilon}}^{\nu}(Q,\lambda)^{*}
= \frac{1}{3}{\left(}-\eta^{\mu\nu} + \frac{Q^\mu\,Q^\nu}{M_V^2}{\right)}\,,
\label{eqn:fdhspinav}$$ which annihilates $\left.\Pi_B^{\mu\nu}\right|_{{FDH}}$. For the annihilation rate, the rules are a bit ambiguous, as they could be read to mean that the lepton spinors are four-dimensional but the vertex ($\gamma^\mu$) connecting them to the loop part of the amplitude is $D_s$ dimensional. This would bring $\left.\Pi_B^{\mu\nu}\right|_{{FDH}}$ into the calculation and lead to a singular result at order $\alpha_s^2$. However, Rule $4$ could also be read to mean that the vertex sandwiched between four-dimensional states is also reduced to being four-dimensional.
Assuming this interpretation, I find that $$\begin{split}
\Gamma^{{{FDH}}}_{V\to\ {\rm hadrons}} =&
\frac{{{\alpha_{V}}}\,M_V}{3}\,N_c\,\sum_f\,Q_f^2\,
\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{15}{32}
+ C_F\,C_A\,{\left(}\frac{133}{32}
- \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{5}{8}
+ \frac{1}{2}\,\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,,
\label{eqn:gamfdhverify}
\end{split}$$ and $$\begin{split}
\sigma^{{{FDH}}}_{e^+\,e^-\to\ {\rm hadrons}} =&
\frac{4\pi\,\alpha^2}{3\,Q^2}\,N_c\,\sum_f\,Q_f^2\,
\left\{ 1 + {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\, C_F\,\frac{3}{4}{\left[}1
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}\,{{\beta_{0}^{\overline{\rm{{MS}}}}}}\,\ln\frac{\mu^2}{Q^2}{\right]}\right.\\
& \hskip-45pt\left.
+{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^2\,{\left[}-C_F^2\,\frac{15}{32}
+ C_F\,C_A\,{\left(}\frac{133}{32}
- \frac{11}{4}\zeta_3{\right)}+ C_F\,N_f\,{\left(}-\frac{5}{8}
+ \frac{1}{2}\,\zeta_3{\right)}{\right]}+ {\cal O}{\left(}{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3{\right)}\right\}\,.
\label{eqn:sigfdhverify}
\end{split}$$ The results agree with one another, are correct through [[NLO]{}]{} and are finite through [[NNLO]{}]{}. Unfortunately, the [[NNLO]{}]{} terms are not correct! Because the discrepancy is finite, there remains the possibility that the conversion from ${{\alpha_{s}^{\overline{\rm{{FDH}}}}}}$ to ${{\alpha_{s}^{\overline{\rm{{MS}}}}}}$ given in [Eq.(\[eqn:fdhtomsbar\])]{} is incorrect, although this would contradict previous results [@Kunszt:1993sd; @Bern:2002zk]. If this were the case, then one would expect that the [[N${}^3$LO]{}]{} result would also be finite but incorrect. If, instead, the finite discrepancy at [[NNLO]{}]{}is the result of a failure of the renormalization program, the [[N${}^3$LO]{}]{}result should be singular.
Partial results at [[N${}^3$LO]{}]{}
====================================
Although first computed some time ago, the vacuum polarization at four loops [@Gorishnii:1988bc; @Gorishnii:1990vf] remains a formidable calculation. It is only necessary, however, to look at a small part of the calculation: the terms proportional to the square of the number of fermion flavors, $N_f^2$. This is fortunate for a couple of reasons: 1) there are only three four-loop diagrams to be computed, see [Fig.(\[fig:fourloop\])]{}, (plus three more in the [[DRED]{}]{} scheme, where the gluons are replaced by evanescent gluons); and 2) the contributions from renormalization in the [[CDR]{}]{} and [[FDH]{}]{} schemes come only from the leading term in the [[QCD]{}]{} $\beta$-function ($\beta_{0}$ and $\beta_{0}^2$). Thus, my result will not depend on how the higher order terms of the $\beta$-function are chosen in the [[FDH]{}]{} scheme.
![Four-loop diagrams that contribute to the $N_f^2$ term at [[N${}^3$LO]{}]{}. \[fig:fourloop\]](TetBubA.png "fig:"){width="4.cm"} ![Four-loop diagrams that contribute to the $N_f^2$ term at [[N${}^3$LO]{}]{}. \[fig:fourloop\]](TetBubB.png "fig:"){width="4.cm"} ![Four-loop diagrams that contribute to the $N_f^2$ term at [[N${}^3$LO]{}]{}. \[fig:fourloop\]](TetBubC.png "fig:"){width="4.cm"}
The [[CDR]{}]{} scheme
----------------------
In the [[CDR]{}]{} scheme, there are only three four-loop diagrams that need to be calculated. The first two are simply iterated-bubble diagrams and are essentially trivial. The third is slightly nontrivial, so I again use my QGRAF-FORM-REDUZE suite of programs to address the problem. All of the four-loop master integrals can be found in Ref. [@Baikov:2010hf]. I find the result of the four-loop calculation to be $$\begin{split}
\Im{\left[}\left.\Pi^{(B)}_{\mu\nu}(Q)\right|_{{{CDR}}}{\right]}_{\alpha_s^3\,N_f^2}
&=\frac{ -Q^2\,g_{\mu\nu} + Q_{\mu}Q_{\nu}}{3}
{{\alpha_{V}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}
{\right)}^{4\,{{\varepsilon}}}\\
&\qquad\times {{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^3\,C_F\,N_f^2{\left[}\frac{1}{48\,{{\varepsilon}}^2}
+ \frac{1}{{{\varepsilon}}}{\left(}\frac{121}{288}
- \frac{1}{3}\zeta_3{\right)}+ \frac{2777}{576} - \frac{3}{8}\zeta_2
- \frac{19}{6}\zeta_3 - \frac{1}{2}\zeta_4{\right]}\label{eqn:fourloopcdrbare}
\end{split}$$ Renormalizing, I find $$\begin{split}
\Im{\left[}\left.\Pi_{\mu\nu}(Q)\right|_{{{CDR}}}{\right]}_{\alpha_s^3\,N_f^2}
&=\frac{ -Q^2\,g_{\mu\nu} + Q_{\mu}Q_{\nu}}{3}
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3\,C_F\,N_f^2\\
&\times{\left[}\frac{151}{216} - \frac{1}{24}\zeta_2
- \frac{19}{36}\zeta_3 + {\left(}\frac{11}{48} - \frac{1}{6}\zeta_3{\right)}\ln{\left(}\frac{\mu^2}{Q^2}{\right)}+ \frac{1}{48}\ln^2{\left(}\frac{\mu^2}{Q^2}{\right)}{\right]}\label{eqn:fourloopcdr}
\end{split}$$ Using this term to compute the $\alpha_s^3\,N_f^2$ contribution to the decay rate and annihilation cross section as in Eqs. (\[eqn:cdrdecay\],\[eqn:cdrannav\]), I find the result expected from [Eqs.(\[eqn:knownresult\]-\[eqn:knownF\])]{}.
The [[DRED]{}]{} scheme
-----------------------
In the [[DRED]{}]{} scheme, there are three extra four-loop diagrams to compute, obtained by replacing gluon propagators with evanescent gluon propagators. I find $$\begin{split}
\Im{\left[}\left.\Pi^{(B)}_{A}(Q)\right|_{{{DRED}}}{\right]}_{\alpha_s^3\,N_f^2} &=
{{\alpha_{V}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{4\,{{\varepsilon}}}
C_F\,N_f^2\,\left\{ \vphantom{{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}}\right.\\
&\phantom{+}
{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^3\,{\left[}\frac{1}{48\,{{\varepsilon}}^2} + \frac{1}{{{\varepsilon}}}{\left(}\frac{13}{32} - \frac{1}{3}\zeta_3{\right)}+ \frac{7847}{1728} - \frac{3}{8}\zeta_2
- \frac{53}{18}\zeta_3 - \frac{1}{2}\zeta_4{\right]}\\
&\left.+ {{{\left(}\frac{\alpha_e^{B}}{\pi}{\right)}}}^3\,{\left[}-\frac{1}{{{\varepsilon}}}\frac{3}{64} - \frac{83}{128}{\right]}\right\}\\
\Im{\left[}\left.\Pi^{(B)}_{B}(Q)\right|_{{{DRED}}}{\right]}_{\alpha_s^3\,N_f^2} &=
{{\alpha_{Ve}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{4\,{{\varepsilon}}}
C_F\,N_f^2\,\left\{ \vphantom{{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}}\right.\\
& \phantom{+}{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^3\,{\left[}\frac{1}{72\,{{\varepsilon}}^2}
+ \frac{1}{{{\varepsilon}}}\frac{73}{432} + \frac{3595}{2592}
- \frac{1}{4}\zeta_2 - \frac{1}{3}\zeta_3{\right]}\\
&\left.+ {{{\left(}\frac{\alpha_e^{B}}{\pi}{\right)}}}^3\,{\left[}-\frac{1}{48\,{{\varepsilon}}^2}
- \frac{1}{{{\varepsilon}}}\frac{11}{48} - \frac{155}{96}
+ \frac{3}{8}\zeta_2{\right]}\right\}
\label{eqn:impi40dred}
\end{split}$$ Upon renormalizing according to [Eq.(\[eqn:dredrenorm\])]{} and converting the coupling to ${{\alpha_{s}^{\overline{\rm{{MS}}}}}}$, I obtain $$\begin{split}
\Im{\left[}\left.\Pi_{A}(Q)\right|_{{{DRED}}}{\right]}_{\alpha_s^3\,N_f^2} &\\
&\hskip -55pt =
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2\,C_F\,N_f^2{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{MS}}}}}{\pi}{\right)}}}^3\,{\left[}\frac{151}{216} - \frac{1}{24}\zeta_2
- \frac{19}{36}\zeta_3 + {\left(}\frac{11}{48} - \frac{1}{6}\zeta_3{\right)}\ln{\left(}\frac{\mu^2}{Q^2}{\right)}+ \frac{1}{48}\ln^2{\left(}\frac{\mu^2}{Q^2}{\right)}{\right]}\,,\\
\Im{\left[}\left.\Pi_{B}(Q)\right|_{{{DRED}}}{\right]}_{\alpha_s^3\,N_f^2} &={\cal O}({{\varepsilon}})\,.
\label{eqn:impi4dred}
\end{split}$$ As for the [[CDR]{}]{} scheme, this leads to the expected result for the decay rate and annihilation cross section.
The [[FDH]{}]{} scheme
----------------------
In the [[FDH]{}]{} scheme, however, I find that $$\begin{split}
\Im{\left[}\left.\Pi^{(B)}_{A}(Q)\right|_{{{FDH}}}{\right]}_{\alpha_s^3\,N_f^2} &=
{{\alpha_{V}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{4\,{{\varepsilon}}}
C_F\,N_f^2\\
&\times
{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^3\,{\left[}\frac{1}{48\,{{\varepsilon}}^2} + \frac{1}{{{\varepsilon}}}{\left(}\frac{23}{64} - \frac{1}{3}\zeta_3{\right)}+ \frac{13453}{3456} - \frac{3}{8}\zeta_2
- \frac{53}{18}\zeta_3 - \frac{1}{2}\zeta_4{\right]}\,,\\
\Im{\left[}\left.\Pi^{(B)}_{B}(Q)\right|_{{{FDH}}}{\right]}_{\alpha_s^3\,N_f^2} &=
{{\alpha_{V}^{B}}}\,N_c\,\sum_f\,Q_f^2{\left(}\frac{4\,\pi}{Q^2\,e^{\gamma_E}}{\right)}^{4\,{{\varepsilon}}}
C_F\,N_f^2\\
& \times{{{\left(}\frac{\alpha_s^{B}}{\pi}{\right)}}}^3\,{\left[}-\frac{1}{144\,{{\varepsilon}}^2}
- \frac{1}{{{\varepsilon}}}\frac{13}{216} - \frac{295}{1296}
+ \frac{1}{8}\zeta_2 - \frac{1}{3}\zeta_3{\right]}\,.
\label{eqn:impi40fdh}
\end{split}$$ I renormalize according to $${{\alpha_s^{B}}}={\left(}\frac{\mu^2\,e^{\gamma_E}}{4\,\pi}{\right)}^{{\varepsilon}}\,
{{\alpha_{s}^{\overline{\rm{{FDH}}}}}}\,{\left[}1 - {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}\frac{{{\beta_{0}^{\overline{\rm{{FDH}}}}}}}{{{\varepsilon}}}
+ {{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}^2{\left(}\frac{{{\beta_{0}^{\overline{\rm{{FDH}}}}}}^2}{{{\varepsilon}}^2}
- \frac{1}{2}\frac{{{\beta_{1}^{\overline{\rm{{FDH}}}}}}}{{{\varepsilon}}}{\right)}{\right]}\,,$$ keeping only terms proportional to ${{\alpha_{s}^{\overline{\rm{{FDH}}}}}}^3\,N_f^2$. Such terms can only come from the ${{\beta_{0}^{\overline{\rm{{FDH}}}}}}$ and ${{\beta_{0}^{\overline{\rm{{FDH}}}}}}^2$ terms, so any uncertainty about ${{\beta_{1}^{\overline{\rm{{FDH}}}}}}$ has no effect here. The renormalized result is $$\begin{split}
\Im{\left[}\left.\Pi_{A}(Q)\right|_{{{FDH}}}{\right]}_{\alpha_s^3\,N_f^2} &\\
&\hskip -75pt =
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2\,C_F\,N_f^2{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}^3\,{\left[}-\frac{1}{192\,{{\varepsilon}}} + \frac{1843}{3456} - \frac{1}{24}\zeta_2
- \frac{19}{36}\zeta_3 + {\left(}\frac{3}{16} - \frac{1}{6}\zeta_3{\right)}\ln{\left(}\frac{\mu^2}{Q^2}{\right)}+ \frac{1}{48}\ln^2{\left(}\frac{\mu^2}{Q^2}{\right)}{\right]}\,,\\
\Im{\left[}\left.\Pi_{B}(Q)\right|_{{{FDH}}}{\right]}_{\alpha_s^3\,N_f^2} &\\
&\hskip -75pt =
{{\alpha_{V}}}\,N_c\,\sum_f\,Q_f^2\,C_F\,N_f^2{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{FDH}}}}}{\pi}{\right)}}}^3\,{\left[}\frac{1}{144\,{{\varepsilon}}^2} -\frac{5}{432\,{{\varepsilon}}} - \frac{869}{2592}
+ \frac{1}{18}\zeta_2 - \frac{5}{27}\ln{\left(}\frac{\mu^2}{Q^2}{\right)}- \frac{1}{36}\ln^2{\left(}\frac{\mu^2}{Q^2}{\right)}{\right]}\,.
\label{eqn:impi4fdh}
\end{split}$$ The demand that external states be four-dimensional removes the $\Pi_B$ term, but there is also a pole in $\Pi_A$ and no finite renormalization to put the result in terms of ${{\alpha_{s}^{\overline{\rm{{MS}}}}}}$ can remove it. I must therefore conclude that the [[FDH]{}]{} scheme is not consistent with unitarity.
Discussion
==========
In this paper, I have performed a high-order calculation in each of three regularization schemes: the conventional dimensional regularization ([[CDR]{}]{}) scheme; the dimensional reduction ([[DRED]{}]{}) scheme; and the four-dimensional helicity ([[FDH]{}]{}) scheme. Of these, the [[CDR]{}]{} scheme is by far the most widely used, and was, in fact, used to compute the original results that I use as my test basis. The [[FDH]{}]{} scheme has primarily been used to produce one-loop helicity amplitudes, although it has been used in a few cases in two-loop calculations and also as a supersymmetric regulator. The primary purpose of this paper was to put the [[FDH]{}]{} scheme to a stringent test and determine its reliability in a high-order calculation. The [[DRED]{}]{}scheme is primarily used as a supersymmetric regulator and is quite cumbersome for nonsupersymmetric calculations. It is, however, closely related to the [[FDH]{}]{} scheme and has been demonstrated [@Jack:1993ws; @Jack:1994bn; @Harlander:2006rj; @Harlander:2006xq] to be equivalent to the [[CDR]{}]{} scheme through four loops. A close comparison of the details of the calculations in the [[FDH]{}]{} and [[DRED]{}]{} schemes helps to identify where and when things go wrong with the former.
In the cases of the [[CDR]{}]{} and [[DRED]{}]{} schemes, I have reproduced the known result for the hadronic decay width of a massive vector boson (or equivalently, the $e^+e^-$ annihilation rate to hadrons) through [[NNLO]{}]{}, and a few terms at [[N${}^3$LO]{}]{}. This represents computing the [[QCD]{}]{} corrections to the vacuum polarization of the photon ($V$ boson) through three loops, with partial results at four loops. In addition, I have reproduced the renormalization parameters of [[QCD]{}]{}($\beta$-function(s), mass anomalous dimension) through three-loop order. This establishes that I have theoretical control over all of the needed calculations through three-loop order. In order to obtain the partial [[N${}^3$LO]{}]{} result in the [[DRED]{}]{} scheme, I also needed the three-loop [[QCD]{}]{} corrections to the $\beta$-function of the evanescent photon ($V$ boson).
The calculation of the $V$ boson decay rate provides another instance of the equivalence the [[CDR]{}]{} and [[DRED]{}]{} schemes at the four-loop level [@Harlander:2006xq]. The ability to obtain the correct result using the [[DRED]{}]{} scheme required a delicate balance of the many extra couplings and their renormalization effects upon one another. Indeed, given the complexity needed to make the [[DRED]{}]{} scheme work, it seems that there should be little surprise that the [[FDH]{}]{} scheme, with its greater simplicity, should fail.
Perhaps, it is worth considering how it is that the [[FDH]{}]{} scheme has been used successfully in so many calculations. Its most common use has been in the construction of one-loop scattering amplitudes via unitarity cuts, using four-dimensional helicity amplitudes as the primary building blocks. Thus, it is natural that it restricts observed (external) states to be four-dimensional. Because the [[FDH]{}]{}scheme defines that $D_s > D_m > 4$, this restriction excludes evanescent fields from appearing as external states. This is very important because, as one can see from comparing Eqs. (\[eqn:impidred\]) and (\[eqn:impi4fdh\]), terms involving external evanescent states are the most dangerous. Even though it does not renormalize evanescent states and couplings properly the [[FDH]{}]{} is able to get the nonevanescent part of the vacuum polarization tensor correct at [[NLO]{}]{}, while the evanescent part is ready to contribute a finite error at [[NLO]{}]{}. Because the [[DRED]{}]{} scheme defines $4 > D_m$, the evanescent states are [*parts*]{} of the classical four-dimensional states. It would not seem natural to exclude them from appearing as external states. Instead, they are handled through the renormalization program so that their effects are removed from physical scattering amplitudes. In the [[FDH]{}]{} scheme, the evanescent states are instead [*additions*]{} to the four-dimensional states (as are the extra degrees of freedom that come from regularizing momentum integrals) and there is no barrier to excluding them as observed states.
In an [[FDH]{}]{} scheme calculation, a tree-level term is strictly four-dimensional and is free from evanescent contributions. (Depending on interpretation, this may be a stronger condition than is given in the rules of Ref [@Bern:2002zk], but it is the actual condition imposed if one defines the tree-level amplitude as being a four-dimensional helicity amplitude.) Because evanescent terms are absent at tree-level, they cannot generate ultraviolet poles at one loop. Even if one were to renormalize them properly, as in the [[DRED]{}]{}scheme, there would be nowhere to make the counter-term insertion! In fact, the one-loop contributions are not even finite, as the counting over the number of states ($2{{\varepsilon}}$) makes the result of order ${{\varepsilon}}$. This is clearly illustrated in [Eq.(\[eqn:impi0Adred\])]{}. Neither $\alpha_s$, nor $\alpha_e$ appear at [[LO]{}]{}. Therefore, the contributions at [[NLO]{}]{} are finite for $\alpha_s$ and of order ${{\varepsilon}}$ (because of the counting over the number of states) for $\alpha_e$. In more complicated [[QCD]{}]{} calculations, $\alpha_s$ will appear at [[LO]{}]{}and will therefore contribute an ultraviolet pole at one-loop, which will be removed by renormalization. $\alpha_e$, however, will still make its first appearance at [[NLO]{}]{} and that contribution will be of order ${{\varepsilon}}$. Thus, one can expect that the [[FDH]{}]{} scheme, used as above, should be reliable for computing [[NLO]{}]{} corrections through finite order (${{\varepsilon}}^0$). The error from improperly identifying evanescent quantities should be of order ${{\varepsilon}}$. At [[NNLO]{}]{} and beyond however, the failure to properly identify and renormalize the evanescent parameters leads to incorrect results and the violation of unitarity.
So, as suggested [@Bern:2002zk], one of the [[FDH]{}]{} scheme’s most important assets is that it defines $D_s > D_m > 4$. This feature is also the scheme’s undoing, though not of necessity. Because the effects of external evanescent states can be removed (or indeed never seen) by imposing a four-dimensionality restriction, and because the effects of internal evanescent states therefore contribute at order ${{\varepsilon}}$ at one loop, it appears that one can simply ignore the distinction between gauge and evanescent terms. In contrast, because the [[DRED]{}]{} scheme must deal with external evanescent terms from the beginning, its advocates were forced to develop a successful renormalization program [@Jack:1993ws; @Jack:1994bn]. Extensive testing [@Jack:1993ws; @Jack:1994bn; @Harlander:2006rj; @Harlander:2006xq] has shown that this program works to at least the fourth order and that it handles the effects of both internal and external evanescent contributions. As I remarked earlier, calculations in the [[DRED]{}]{} and [[FDH]{}]{} schemes are term-by-term identical, except for the identification of the couplings and propagating states. Thus, one could make the [[FDH]{}]{} scheme a unitary regularization scheme for nonsupersymmetric calculations by recognizing the distinction between gauge and evanescent terms and adopting the [[DRED]{}]{} scheme’s renormalization program. This would, of course, do away with any notion of the [[FDH]{}]{} scheme being simple, but it would at least be correct. The [[FDH]{}]{} scheme would still be distinguished from the [[DRED]{}]{} scheme by the fact that $D_s > D_m > 4$, which facilitates helicity amplitude calculations and, in chiral theories, improves its situation with regard to $\gamma_5$ and the Levi-Civita tensor [@Siegel:1980qs; @Stockinger:2005gx]. Furthermore, with a valid renormalization program, the requirement of four-dimensional observed states could be made optional. This would lead to two linked, slightly different, schemes, just like the [[HV]{}]{}and [[CDR]{}]{} schemes. This suggestion has already been made by Signer and Stöckinger [@Signer:2008va] who in fact define their version of the [[DRED]{}]{} scheme to have precisely the $D_s > D_m > 4$ hierarchy of the [[FDH]{}]{} scheme.
Thus, in conclusion, the [[CDR]{}]{} and [[DRED]{}]{} schemes are correct and equivalent ways of performing [[QCD]{}]{} calculations through [[N${}^3$LO]{}]{}. The [[FDH]{}]{} scheme, however, has been shown to be incorrect and to violate unitarity beyond [[NLO]{}]{} when applied to nonsupersymmetric theories. It must therefore be viewed as a shortcut for performing [[NLO]{}]{}calculations and should only be used for such calculations with great caution.
#### Acknowledgments: {#acknowledgments .unnumbered}
This research was supported by the U.S. Department of Energy under Contract No. DE-AC02-98CH10886.
Renormalization parameters for the [[CDR]{}]{} scheme {#sec:cdrrenorm}
=====================================================
To three-loop order, I find the coefficients of the $\beta$-function to be $$\begin{split}
{{\beta_{0}^{\overline{\rm{{MS}}}}}} &= \frac{11}{12}C_A - \frac{1}{6}N_f\,,\qquad\qquad
{{\beta_{1}^{\overline{\rm{{MS}}}}}} = \frac{17}{24}C_A^2 - \frac{5}{24}C_A\,N_f
- \frac{1}{8}C_F\,N_f\,,\\
{{\beta_{2}^{\overline{\rm{{MS}}}}}} &= \frac{2857}{3456}C_A^3
- \frac{1415}{3456}C_A^2\,N_f - \frac{205}{1152}C_A\,C_F\,N_f
+ \frac{1}{64}C_F^2\,N_f + \frac{79}{3456}C_A\,N_f^2
+ \frac{11}{576}C_F\,N_f^2\,,
\label{eqn:cdrbeta}
\end{split}$$ while the coefficients of the mass anomalous dimension are $$\begin{split}
{{\gamma_{0}^{\overline{\rm{{MS}}}}}} &= \frac{3}{4}C_F\,,\hskip80pt
{{\gamma_{1}^{\overline{\rm{{MS}}}}}} = \frac{3}{32}C_F^2 + \frac{97}{96}C_F\,C_A
- \frac{5}{48}C_F\,N_f\,,\\
{{\gamma_{2}^{\overline{\rm{{MS}}}}}} &= \frac{129}{128}C_F^3 - \frac{129}{256}C_F^2\,C_A
+ \frac{11413}{6912}C_F\,C_A^2
- {\left(}\frac{23}{64} - \frac{3}{8}\zeta_3{\right)}\,C_F^2\,N_f
- {\left(}\frac{139}{864} + \frac{3}{8}\zeta_3{\right)}\,C_F\,C_A\,N_f
- \frac{35}{1728}C_F\,N_f^2\,,
\label{eqn:cdrgamma}
\end{split}$$ in agreement with known results [@Tarasov:1980au; @Larin:1993tp; @Chetyrkin:1997dh; @Vermaseren:1997fq].
Renormalization parameters for the [[DRED]{}]{} scheme {#sec:dredrenorm}
======================================================
The coefficients of the [[QCD]{}]{} $\beta$-function, ${{\beta_{}^{\overline{\rm{{DR}}}}}}({{\alpha_{s}^{\overline{\rm{{DR}}}}}})$ through three loops are: $$\begin{split}
{{\beta_{20}^{\overline{\rm{{DR}}}}}} &= \frac{11}{12}C_A - \frac{1}{6}N_f\,,\hskip60pt
{{\beta_{30}^{\overline{\rm{{DR}}}}}} = \frac{17}{24}C_A^2 - \frac{5}{24}C_A\,N_f
- \frac{1}{8}C_F\,N_f\,,\\
{{\beta_{40}^{\overline{\rm{{DR}}}}}} &= \frac{3115}{3456}C_A^3
- \frac{1439}{3456}C_A^2\,N_f - \frac{193}{1152}C_A\,C_F\,N_f
+ \frac{1}{64}C_F^2\,N_f + \frac{79}{3456}C_A\,N_f^2
+ \frac{11}{576}C_F\,N_f^2\,,\\
{{\beta_{31}^{\overline{\rm{{DR}}}}}} & = - \frac{1}{16}C_F\,N_f{\left(}\frac{3}{2}C_F{\right)}\,,\qquad\quad
{{\beta_{22}^{\overline{\rm{{DR}}}}}} = - \frac{1}{16}C_F\,N_f{\left(}\frac{1}{2}C_A - C_F
- \frac{1}{4}N_f{\right)}\,,
\label{eqn:dredbeta}
\end{split}$$ where the notation is that $${{\beta_{}^{\overline{\rm{{DR}}}}}}({{\alpha_{s}^{\overline{\rm{{DR}}}}}}) = -{{\varepsilon}}\frac{{{\alpha_{s}^{\overline{\rm{{DR}}}}}}}{\pi} - \sum_{i,j,k,l,m}\,{{\beta_{ijklm}^{\overline{\rm{{DR}}}}}}
\,{{{\left(}\frac{\alpha_{s}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{i}\,{{{\left(}\frac{\alpha_{e}^{\overline{\rm{{DR}}}}}{\pi}{\right)}}}^{j}\,{{{\left(}\frac{\eta_{1}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{k}
\,{{{\left(}\frac{\eta_{2}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{l}\,{{{\left(}\frac{\eta_{3}^{{{\overline{\rm DR}}}}}{\pi}{\right)}}}^{m}\,.$$ The last three indices of ${{\beta_{ijklm}^{\overline{\rm{{DR}}}}}}$ are omitted when they are all equal to $0$.
The $\beta$-function of evanescent [[QCD]{}]{} coupling, ${{\beta_{e,\,\,}^{\overline{\rm{{DR}}}}}}({{\alpha_{e}^{\overline{\rm{{DR}}}}}})$ is $$\begin{split}
{{\beta_{e,\,0\,2}^{\overline{\rm{{DR}}}}}} &= \frac{1}{2}C_A - C_F - \frac{1}{4}N_f\,,\qquad
{{\beta_{e,\,1\,1}^{\overline{\rm{{DR}}}}}} = \frac{3}{2}C_F\,,\\[5pt]
{{\beta_{e,\,0\,3}^{\overline{\rm{{DR}}}}}} &= \frac{3}{8}C_A^2 - \frac{5}{4}C_A\,C_F
+ C_F^2 - \frac{3}{16}C_A\,N_f
+ \frac{3}{8}C_F\,N_f\,,\qquad
{{\beta_{e,\,1\,2}^{\overline{\rm{{DR}}}}}} = - \frac{3}{8}C_A^2 + \frac{5}{2}C_A\,C_F
- \frac{11}{4}C_F^2 - \frac{5}{16}C_F\,N_f\,,\\
{{\beta_{e,\,2\,1}^{\overline{\rm{{DR}}}}}} &= - \frac{7}{64}C_A^2
+\frac{55}{48}C_A\,C_F + \frac{3}{16}C_F^2
+ \frac{1}{16}C_A\,N_f - \frac{5}{24}C_F\,N_f\\[5pt]
{{\beta_{e,\,0\,2100}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt= -\frac{9}{8}\qquad
{{\beta_{e,\,0\,2010}^{\overline{\rm{{DR}}}}}} = \frac{5}{4}\qquad
{{\beta_{e,\,0\,2001}^{\overline{\rm{{DR}}}}}} = \frac{3}{4}\\
{{\beta_{e,\,0\,1200}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt= \frac{27}{64}\qquad
{{\beta_{e,\,0\,1020}^{\overline{\rm{{DR}}}}}} = -\frac{15}{4}\qquad
{{\beta_{e,\,0\,1002}^{\overline{\rm{{DR}}}}}} = \frac{21}{32}\qquad
{{\beta_{e,\,0\,1101}^{\overline{\rm{{DR}}}}}} = -\frac{9}{16}\\
{{\beta_{e,\,0\,4}^{\overline{\rm{{DR}}}}}} &=
- {\left(}\frac{7}{4} + \frac{9}{4}\zeta_3{\right)}\,C_F^3
+ {\left(}\frac{17}{8} + \frac{15}{2}\zeta_3{\right)}\,C_F^2\,C_A
- {\left(}\frac{3}{4} + \frac{69}{16}\zeta_3{\right)}\,C_F\,C_A^2
+ {\left(}\frac{1}{16} + \frac{9}{16}\zeta_3{\right)}\,C_A^3\\
&
+ {\left(}\frac{13}{32} - \frac{33}{16}\zeta_3{\right)}\,C_F^2\,N_f
+ {\left(}\frac{1}{32} + \frac{51}{32}\zeta_3{\right)}\,C_F\,C_A\,N_f
- {\left(}\frac{21}{128} + \frac{9}{32}\zeta_3{\right)}\,C_A^2\,N_f
- {\left(}\frac{1}{128}C_F - \frac{7}{256}C_A{\right)}\,N_f^2\\
{{\beta_{e,\,1\,3}^{\overline{\rm{{DR}}}}}} &=
{\left(}\frac{13}{2} - 3\,\zeta_3{\right)}\,C_F^3
- {\left(}10 - 6\,\zeta_3{\right)}\,C_F^2\,C_A
+ {\left(}\frac{133}{32} - \frac{15}{4}\zeta_3{\right)}\,C_F\,C_A^2
- {\left(}\frac{25}{64} - \frac{3}{4}\zeta_3{\right)}\,C_A^3\\
&
+ {\left(}\frac{13}{16} - \frac{3}{4}\zeta_3{\right)}\,C_F^2\,N_f
- \frac{9}{8}{\left(}1-\zeta_3{\right)}\,C_F\,C_A\,N_f
+ {\left(}\frac{7}{32} - \frac{3}{8}\zeta_3{\right)}\,C_A^2\,N_f
+ \frac{3}{64}\,C_A\,N_f^2\\
{{\beta_{e,\,2\,2}^{\overline{\rm{{DR}}}}}} &=
- {\left(}\frac{139}{64} - \frac{27}{4}\zeta_3{\right)}\,C_F^3
- {\left(}\frac{793}{128} + 18\,\zeta_3{\right)}\,C_F^2\,C_A
+ {\left(}\frac{1587}{256} + \frac{207}{16}\zeta_3{\right)}\,C_F\,C_A^2
- {\left(}\frac{427}{512} + \frac{45}{16}\zeta_3{\right)}\,C_A^3\\
&
- {\left(}\frac{569}{256} - \frac{99}{16}\zeta_3{\right)}\,C_F^2\,N_f
+ {\left(}\frac{31}{16} - \frac{171}{32}\zeta_3{\right)}\,C_F\,C_A\,N_f
- {\left(}\frac{871}{1024} - \frac{45}{32}\zeta_3{\right)}\,C_A^2\,N_f
+ {\left(}\frac{1}{16}C_F - \frac{1}{256}C_A{\right)}\,N_f^2\\
{{\beta_{e,\,3\,1}^{\overline{\rm{{DR}}}}}} &=
\frac{129}{64}C_F^3 - \frac{457}{128}C_F^2\,C_A
+ \frac{11875}{3456}C_F\,C_A^2
- \frac{3073}{4608}C_A^3\\
&
- {\left(}\frac{23}{32} - \frac{3}{4}\zeta_3{\right)}\,C_F^2\,N_f
- {\left(}\frac{157}{1728} + \frac{3}{4}\zeta_3{\right)}\,C_F\,C_A\,N_f
+ \frac{463}{2304}C_A^2\,N_f
- {\left(}\frac{35}{864}C_F + \frac{5}{576}C_A{\right)}\,N_f^2\\
{{\beta_{e,\,0\,3100}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt =
- \frac{9}{64} + \frac{243}{128}N_f\qquad
{{\beta_{e,\,0\,3010}^{\overline{\rm{{DR}}}}}} = \frac{5}{8} - \frac{45}{64}N_f\qquad
{{\beta_{e,\,0\,3001}^{\overline{\rm{{DR}}}}}} = \frac{3}{32} - \frac{81}{64}N_f\\
{{\beta_{e,\,1\,2100}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt = -\frac{219}{16}\qquad
{{\beta_{e,\,1\,2010}^{\overline{\rm{{DR}}}}}} = \frac{145}{48}\qquad
{{\beta_{e,\,1\,2001}^{\overline{\rm{{DR}}}}}} = \frac{73}{8}\\
{{\beta_{e,\,2\,1100}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt = -\frac{1125}{1024}\qquad
{{\beta_{e,\,2\,1010}^{\overline{\rm{{DR}}}}}} = \frac{105}{128}\qquad
{{\beta_{e,\,2\,1001}^{\overline{\rm{{DR}}}}}} = \frac{615}{512}\\
{{\beta_{e,\,0\,2200}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt =
\frac{1413}{512} - \frac{729}{1024}N_f\qquad
{{\beta_{e,\,0\,2020}^{\overline{\rm{{DR}}}}}} = - \frac{115}{32} + \frac{135}{64}N_f\qquad
{{\beta_{e,\,0\,2002}^{\overline{\rm{{DR}}}}}} = - \frac{161}{256} - \frac{567}{512}N_f\\
{{\beta_{e,\,0\,2110}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt =
\frac{75}{8}\qquad
{{\beta_{e,\,0\,2101}^{\overline{\rm{{DR}}}}}} = -\frac{471}{128} + \frac{243}{256}N_f\qquad
{{\beta_{e,\,0\,2011}^{\overline{\rm{{DR}}}}}} = -\frac{85}{8}\\
{{\beta_{e,\,0\,1300}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt = -\frac{1701}{1024}\qquad
{{\beta_{e,\,0\,1210}^{\overline{\rm{{DR}}}}}} = -\frac{405}{128}\qquad
{{\beta_{e,\,0\,1201}^{\overline{\rm{{DR}}}}}} = \frac{1701}{512}\\
{{\beta_{e,\,0\,1120}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt = \frac{135}{32}\qquad
{{\beta_{e,\,0\,1111}^{\overline{\rm{{DR}}}}}} = \frac{135}{16}\qquad
{{\beta_{e,\,0\,1102}^{\overline{\rm{{DR}}}}}} = -\frac{81}{128}\\
{{\beta_{e,\,0\,1021}^{\overline{\rm{{DR}}}}}} \hskip-13pt&\hskip13pt =
- \frac{315}{32}\qquad
{{\beta_{e,\,0\,1012}^{\overline{\rm{{DR}}}}}} = - \frac{315}{32}\qquad
{{\beta_{e,\,0\,1003}^{\overline{\rm{{DR}}}}}} = \frac{63}{128}\qquad
\label{eqn:dredbetae}
\end{split}$$
The mass anomalous dimension in the [[DRED]{}]{} scheme is $$\begin{split}
{{\gamma_{10}^{\overline{\rm{{DR}}}}}} &= \frac{3}{4}C_F \\[5pt]
{{\gamma_{20}^{\overline{\rm{{DR}}}}}} &= \frac{3}{32}C_F^2 + \frac{91}{96}C_A\,C_F
- \frac{5}{48}C_F\,N_f\qquad
{{\gamma_{11}^{\overline{\rm{{DR}}}}}} = - \frac{3}{8}C_F^2 \qquad
{{\gamma_{02}^{\overline{\rm{{DR}}}}}} = \frac{1}{4}C_F^2 - \frac{1}{8}C_A\,C_F
+ \frac{1}{16}C_F\,N_f\\[5pt]
{{\gamma_{30}^{\overline{\rm{{DR}}}}}} &= \frac{129}{128}C_F^3 - \frac{133}{256}C_F^2\,C_A
+ \frac{10255}{6912}C_F\,C_A^2
- {\left(}\frac{23}{64} - \frac{3}{8}\zeta_3{\right)}\,C_F^2\,N_f
- {\left(}\frac{281}{1728} + \frac{3}{8}\zeta_3{\right)}\,C_A\,C_F\,N_f
- \frac{35}{1728}C_F\,N_f^2\\
{{\gamma_{21}^{\overline{\rm{{DR}}}}}} &= - \frac{27}{64}C_F^3 - \frac{21}{32}C_F^2\,C_A
- \frac{15}{256}C_F\,C_A^2 + \frac{9}{64}C_F^2\,N_f\\
{{\gamma_{12}^{\overline{\rm{{DR}}}}}} &= \frac{9}{8}C_F^3 - \frac{21}{32}C_F^2\,C_A
+ \frac{3}{64}C_F\,C_A^2 + \frac{3}{128}C_F\,C_A\,N_f
+ \frac{3}{16}C_F^2\,N_f\\
{{\gamma_{03}^{\overline{\rm{{DR}}}}}} &= - \frac{3}{8}C_F^3 + \frac{3}{8}C_F^2\,C_A
- \frac{3}{32}C_F\,C_A^2 + \frac{1}{16}C_F\,C_A\,N_f
- \frac{5}{32}C_F^2\,N_f - \frac{1}{128}C_F\,N_f^2\\[5pt]
{{\gamma_{02100}^{\overline{\rm{{DR}}}}}} \hskip-10pt&\hskip10pt= \frac{3}{8}\qquad
{{\gamma_{02010}^{\overline{\rm{{DR}}}}}} = -\frac{5}{12}\qquad
{{\gamma_{02001}^{\overline{\rm{{DR}}}}}} = -\frac{1}{4}\qquad\\
{{\gamma_{01200}^{\overline{\rm{{DR}}}}}} \hskip-10pt&\hskip10pt= -\frac{9}{64}\qquad
{{\gamma_{01101}^{\overline{\rm{{DR}}}}}} = \frac{3}{16}\qquad
{{\gamma_{01020}^{\overline{\rm{{DR}}}}}} = \frac{5}{4}\qquad
{{\gamma_{01002}^{\overline{\rm{{DR}}}}}} = - \frac{7}{32}
\label{eqn:dredgamma}
\end{split}$$ The above results for ${{\beta_{}^{\overline{\rm{{DR}}}}}}$, ${{\beta_{e,\,\,}^{\overline{\rm{{DR}}}}}}$ and ${{\gamma_{}^{\overline{\rm{{DR}}}}}}$ all agree with the results of Refs. [@Harlander:2006rj; @Harlander:2006xq]
The [[QCD]{}]{} contributions to the $\beta$-function of the evanescent part of a non-[[QCD]{}]{} gauge coupling is a new result. I find $$\begin{split}
{{\beta_{Ve,\,1\,0}^{\overline{\rm{{DR}}}}}}& = \frac{3}{2}C_F\qquad {{\beta_{Ve,\,0\,1}^{\overline{\rm{{DR}}}}}} = - C_F\\[5pt]
{{\beta_{Ve,\,2\,0}^{\overline{\rm{{DR}}}}}}& = \frac{3}{16}C_F^2 + \frac{91}{48}C_F\,C_A
- \frac{5}{24}C_F\,N_f \qquad
{{\beta_{Ve,\,1\,1}^{\overline{\rm{{DR}}}}}} = - \frac{11}{4}C_F^2 - \frac{3}{4}C_F\,C_A\qquad
{{\beta_{Ve,\,0\,2}^{\overline{\rm{{DR}}}}}} = C_F^2 + \frac{3}{8}C_F\,N_f\\[5pt]
{{\beta_{Ve,\,3\,0}^{\overline{\rm{{DR}}}}}} & = \frac{129}{64}C_F^3
- \frac{133}{128}C_F^2\,C_A
- {\left(}\frac{23}{32} - \frac{3}{4}\zeta_3{\right)}\,C_F^2\,N_f
+ \frac{10255}{3456}C_F\,C_A^2
- {\left(}\frac{281}{864} + \frac{3}{4}\zeta_3{\right)}\,C_F\,C_A\,N_f
- \frac{35}{864}C_F\,N_f^2\\
{{\beta_{Ve,\,2\,1}^{\overline{\rm{{DR}}}}}} & = - {\left(}\frac{139}{64} - \frac{27}{4}\zeta_3{\right)}\,C_F^3
- {\left(}\frac{331}{64} + \frac{81}{8}\zeta_3{\right)}\,C_F^2\,C_A
+ \frac{11}{16}C_F^2\,N_f
- {\left(}\frac{195}{256} - \frac{27}{8}\zeta_3{\right)}\,C_F\,C_A^2
+ \frac{5}{64}C_F\,C_A\,N_f\\
{{\beta_{Ve,\,1\,2}^{\overline{\rm{{DR}}}}}} & = {\left(}\frac{13}{2} - 3\,\zeta_3{\right)}\,C_F^3
- {\left(}\frac{7}{8} - \frac{9}{2}\zeta_3{\right)}\,C_F^2\,C_A
+ {\left(}\frac{63}{64} - \frac{3}{4}\zeta_3{\right)}\,C_F^2\,N_f
+ {\left(}\frac{7}{16} - \frac{3}{2}\zeta_3{\right)}\,C_F\,C_A^2\\
&
- {\left(}\frac{3}{64} - \frac{3}{4}\zeta_3{\right)}\,C_F\,C_A\,N_f\\[5pt]
{{\beta_{Ve,\,0\,3}^{\overline{\rm{{DR}}}}}} & = - {\left(}\frac{7}{4} + \frac{9}{4}\zeta_3{\right)}\,C_F^3
+ {\left(}\frac{1}{8} + \frac{27}{8}\zeta_3{\right)}\,C_F^2\,C_A
- \frac{27}{32}C_F^2\,N_f
+ {\left(}\frac{1}{16} - \frac{9}{8}\zeta_3{\right)}\,C_F\,C_A^2
+ \frac{3}{64}C_F\,C_A\,N_f
+ \frac{3}{64}C_F\,N_f^2\\
{{\beta_{Ve,\,0\,2100}^{\overline{\rm{{DR}}}}}} \hskip -13pt&\hskip13pt= \frac{3}{8}\qquad
{{\beta_{Ve,\,0\,2010}^{\overline{\rm{{DR}}}}}} = -\frac{25}{6}\qquad
{{\beta_{Ve,\,0\,2001}^{\overline{\rm{{DR}}}}}} = -\frac{1}{4}\\
{{\beta_{Ve,\,0\,1200}^{\overline{\rm{{DR}}}}}} \hskip -13pt&\hskip13pt=-\frac{63}{64}\qquad
{{\beta_{Ve,\,0\,1101}^{\overline{\rm{{DR}}}}}} = \frac{21}{16}\qquad
{{\beta_{Ve,\,0\,1020}^{\overline{\rm{{DR}}}}}} = \frac{65}{4}\qquad
{{\beta_{Ve,\,0\,1002}^{\overline{\rm{{DR}}}}}} = -\frac{49}{32}
\label{eqn:dredbetave}
\end{split}$$
[33]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ** (, , ).
, ****, ().
, ****, ().
, , , , ****, (), .
, ****, ().
, , , ****, ().
, , , ****, (), .
, , , ****, (), .
, , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
(), , .
, ****, (), .
, , , ****, (), .
, , , ****, ().
, ****, ().
, , , , ****, ().
, , (), .
, , , , , ****, (), .
, , , , ****, (), .
, , , ****, (), .
, ****, (), , .
, ****, ().
, ****, (), .
, ****, (), .
, , , ****, ().
, ****, (), .
, ****, (), .
, , , ****, (), .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this work we derive a lower bound for the minimum time required to implement a target unitary transformation through a classical time-dependent field in a closed quantum system. The bound depends on the target gate, the strength of the internal Hamiltonian and the highest permitted control field amplitude. These findings reveal some properties of the reachable set of operations, explicitly analyzed for a single qubit. Moreover, for fully controllable systems, we identify a lower bound for the time at which all unitary gates become reachable. We use numerical gate optimization in order to study the tightness of the obtained bounds. It is shown that in the single qubit case our analytical findings describe the relationship between the highest control field amplitude and the minimum evolution time remarkably well. Finally, we discuss both challenges and ways forward for obtaining tighter bounds for higher dimensional systems, offering a discussion about the mathematical form and the physical meaning of the bound.'
address:
- 'Frick Laboratory, Princeton University, Princeton NJ 08544, US'
- 'Frick Laboratory, Princeton University, Princeton NJ 08544, US'
- 'Institute of Mathematics, Physics, and Computer Science, Aberystwyth University, Aberystwyth SY23 2BZ, UK'
- 'Frick Laboratory, Princeton University, Princeton NJ 08544, US'
author:
- Christian Arenz
- Benjamin Russell
- Daniel Burgarth
- Herschel Rabitz
title: The roles of drift and control field constraints upon quantum control speed limits
---
Introduction
============
Future and present quantum technologies, as well as experiments in highly sensitive quantum systems, require a fine degree of control over the considered system. In particular, the preparation of states and the implementation of quantum gates for quantum information processing tasks both critically rely on high fidelity quantum operations. It is vital to be able to implement such operations as accurately as possible with the available control resources, while also operating on a time scale significantly below the typical decoherence time scale of the system employed. Quantum control, which is primarily focused on the task of ‘steering’ a quantum system towards a desired target by using suitably tailored classical fields [@ControlRev1; @ControlRev2], has successfully been applied to a broad class of quantum systems for disparate purposes. Diverse applications include: driving chemical reactions [@HerschMol], entangling spin qubits in nitrogen vacancy centers [@SpinQubits] and noise filtering [@NoiseF]. In general, much attention has been drawn to two aspects of quantum control theory, (i) the identification of the operations/states that can be implemented/prepared, and (ii) the calculation of corresponding pulses. Regarding (i), the Lie theoretic approach, sometimes referred to as geometric quantum control theory, expresses the questions of reachability within the framework of Lie groups and Lie algebras [@BookDalessandro; @BookJurdjevic; @LieGeoTH]. The frequently employed Lie algebra rank criterion [@LieRankC1] is a powerful tool which facilitates the determination of the reachable operations or states for a given quantum system steered by classical control fields. When it comes to the determination of the control fields (ii), both numerical and analytical tools are used. The deployment of optimal control theory [@BookDalessandro; @OptimalC], which is based on the Pontryagin maximum principle, can efficiently maximize the fidelity for reaching a desired target. Typically this is done by numerically optimizing a given cost functional, sometimes subject to additional constraints, with a gradient based search [@Grape; @DYNAMOpaper; @GateFid2; @exactgradient].
While both the aforementioned aspects of quantum control have been extensively studied, much less attention has been devoted to understanding either the relevant time scales or the properties of the control fields necessary to implement a given target. In various studies [@SpeedLimit1; @SpeedLimit2; @SpeedLimit3; @SpeedLimit4; @SpeedLimit5; @SpeedLimit6; @SpeedLimit8] lower bounds (known as *quantum speed limits*, or *QSL*), which characterize how fast a quantum system can evolve from an initial state to some final target state have been established for closed, finite dimensional systems. Additionally the QSL has been studied for implementing a two qubit gate using ultracold atoms in an optical potential [@SpeedLimit7]. In later works, these bounds have been extended to open systems [@SpeedLimitOpenS1; @SpeedLimitOpenS2; @SpeedLimitOpenS3; @SpeedLimitOpenS5; @SpeedLimitOpenS4]. Moreover, based on Lieb-Robinson bounds, speed limits for quantum information tasks such as the creation of entanglement were recently established [@SpeedLimitQuantumInfor]. In the closed system case, it has been shown that, for specific examples such limits can be reached by searching for control pulses using optimal control theory [@OptimalControlSpeedLimit1; @OptimalControlSpeedLimit2]. We remark here that this is a rather special case, while typically the standard QSL bounds are not tight when used for time dependent control systems, which will be discussed in section \[sec:tightness\].
In cases in which the control fields are unconstrained, the minimum time (minimized over all pulses which implement a desired gate) to implement a target unitary transformation can be calculated analytically for simple models [@OptimalControlSpeedLimit2; @ExactCalc1; @ExactCalc2; @ExactCalc3]. Moreover, extensive numerical studies have been carried out to find the minimum gate time for more complex systems [@Carenzspinstar; @Nori]. However, finding the shortest possible control pulses in general remains challenging. It is broadly equivalent to finding geodesics of Randers type Finsler metrics on either (special) unitary groups or complex projective spaces for the tasks of implementing gates or preparing states, respectively [@Geo1; @Geo2; @Geo3; @Geo4; @Geo5]. Other approaches also exist including brachistochrone equations [@Geo6], however as of yet these can only be addressed by numerical approaches and they are not geometrically intrinsic rendering analytical solutions harder to obtain.
In light of the difficulty of finding general solutions to time optimal quantum control problems, it is desirable to establish lower bounds drawing on as much information as possible in order to obtain first estimates. In particular, the effects of a constrained control field, the strength of the internal Hamiltonian and the choice of the target gate upon the minimum time are physically important and relatively unexplored. Although detailed studies for qubit systems exist [@ReachQubit1; @ReachQubit2; @ReachMultiQubit; @Ellie], the characterization of the reachable set of operations as a function of the evolution time and constraints on either or both of the control field and the ‘internal’ Hamiltonian also warrants further investigation.
The purpose of this article is twofold. The first objective is to address the aforementioned questions, while in a second step we will discuss the obstacles to obtaining more accurate bounds on minimum gate times. We emphasize that the standard QSL’s, which are typically formulated for time independent systems, only depend on the geometry of the systems Hilbert space, whereas here we seek a QSL that is system and control dependent. The results serve as first estimates towards controlling complex quantum systems with feasible and robust pulses on an appropriate time scale. Throughout this work we consider systems described by a Hamiltonian of the form $$\begin{aligned}
\label{eq:ControleSG}
H(t)=H_{0}+f(t)H_{c},\end{aligned}$$ where $f(t)$ is the control field and $H_{0}$ and $H_{c}$ are the drift and control Hamiltonians respectively. Based on simple arguments, particularly an inequality from [@Geo1], we derive for a lower bound, $$\begin{aligned}
\label{eq:lowerboundtime}
\frac{C(U_{g},H_{c})}{\Vert H_{0}\Vert}+\frac{C(U_{g},H_{0})}{|f_{\text{max}}|\Vert H_{c}\Vert }\leq T,\end{aligned}$$ for the time $T$ required to implement a target unitary operation $U_{g}$. The quantities $C(U_{g},H_{c})$ and $C(U_{g},H_{0})$ depend on the target gate, the eigenbasis of the control and the drift Hamiltonian respectively, and the dimension of the quantum system being considered (see Eq. and ); $f_{\text{max}}$ is the maximum permitted control field amplitude. Since the bound depends on the target gate, for a given system we can further characterize the set of unitary operations which are not reachable for a fixed $f_{\text{max}}, ~T$ and $\Vert H_{0}\Vert, ~\Vert H_{c}\Vert$. We subsequently establish a lower bound for the time $T_c$ in which all gates become reachable as a corollary of this observation. Numerical gate optimization using gradient ascent pulse engineering (GRAPE) is used in order to study the tightness of the obtained bounds. For a single qubit, we show that optimal control theory allows us to operate at the boundary (similar to a Pareto front) of the viable region in the $T,~f_{\text{max}}$ plane defined by . Finally, we discuss challenges to obtaining bounds which are tighter for higher dimensional systems, and further offer a discussion about the nature of the obtained bound .
Bound on the minimum gate time
==============================
We start by considering the following control system $$\begin{aligned}
\label{eq:controlsystem}
\dot{U}(t)=-iH(t)U(t),~~~~U(0)=\mathds{1}, \end{aligned}$$ on the unitary group $\text{U}(d)$ consisting of unitary $d\times d$ matrices. Throughout this article we set $\hbar=1$. We study Hamiltonians $H(t)$ of form , in which the control field enters in a bilinear way, typically known as the *dipole approximation* in chemical physics and as affine bi-linear control on Lie groups in the mathematics community [@Elliot]. We further denote the set of gates, which can be reached at any time by some specific control field by $\mathcal R$. It is well known that the closure of the reachable set $\overline{\mathcal R}$ is equal to the Lie group $e^{\mathfrak{L}}$ with $\mathfrak{L}=\text{Lie}(iH_{0},iH_{c})$ being the dynamical Lie algebra that is generated by iterated commutators and (real) linear combinations of the drift and the control Hamiltonian [@BookDalessandro]. The system is said to be fully controllable iff $\mathfrak{L}=\mathfrak{u}(d)$ (or $\mathfrak{su}(d)$ for traceless Hamiltonians), where $\mathfrak{u}(d)$ is the Lie algebra of skew-hermitian matrices. Equivalently, for a fully controllable system every unitary goal gate $U_{g}\in \text{U}(d)$ can be implemented arbitrarily well [@BookDalessandro]. Remarkably, this is true for *almost all* (all but a set of measure zero) control system of the form [@BookJurdjevic; @Altafini]. For a more detailed introduction into quantum control theory and its terminology we refer to [@BookDalessandro; @BookJurdjevic; @LieGeoTH]. The dynamical Lie algebra is a powerful tool allowing one to identify the operations that can be implemented within a given control system. However, it does not reveal anything about how much time is needed in order to implement a specific target, neither does it say anything about the strength of the corresponding control field(s). Intuitively one would expect that if the strength of the control or the drift Hamiltonian decreases, then the time or the control field amplitude must correspondingly increase depending on the gate we want to implement. In what follows we verify this intuition by establishing the lower bound .
Using the triangle inequality, one can show that for two unitary operators $U_{1}(T)$ and $U_{2}(T)$, which are solutions to the Schödinger equation at time $T$, the inequality [@Geo1] (see appendix \[app:derivationNielsen\] for a derivation), $$\begin{aligned}
\label{eq:NielsenBound}
\Vert U_{1}(T)-U_{2}(T)\Vert \leq \int_{0}^{T}\Vert H_{1}(t)-H_{2}(t)\Vert\, dt\end{aligned}$$ holds for any unitarily invariant norm with $H_{1}(t)$ and $H_{2}(t)$ being the Hamiltonians corresponding to the two trajectories $U_{1}(t)$ and $U_{2}(t)$ respectively. Now, let $H_{1}(t)=H_{0}+f(t)H_{c}$ and $H_{2}(t)=f(t)H_{c}$ such that $U_{2}(T)=\exp(-i\alpha(T)H_{c})$ and $U_{1}(T)=U_{g}$ is the solution to which implements the desired target and $\alpha(T)=\int_{0}^{T}f(t)dt$ is the integrated control field. We assume here that the target gate can be implemented by the given control system, i.e., $U_{g}\in \overline{\mathcal R}$, and we further note that any corresponding control field is not necessarily unique. There can exist multiple different pulse shapes driving the system to the same target evolution for a given final time $T$. Roughly speaking, Eq. , instantiated with the above choice for $U_{1}$ and $U_{2}$, yields a description of how much the drift Hamiltonian is “needed” in order to reach a gate. We note that a similar separation has been suggested in [@SpeedLimit8] by constructing an observable that commutes with $H_{0}$. However, the speed limit in [@SpeedLimit8] applies only to state-to-state transfer and it captures only the effect of constrained control fields and fails to characterize the role of the strength of the drift Hamiltonian. In the following we derive a speed limit for implementing a unitary transformation that explicitly incorporates the strength of the drift Hamiltonian, as well the maximum control field amplitude. We begin with the above choice for $U_{1}$ and $U_{2}$ in order to obtain a speed limit that depends on the strength of the drift Hamiltonian. Afterwards, we chose $U_{1}$ and $U_{2}$ differently in order to obtain another speed limit that depends on the maximum strength of the control field. A linear combination of both bounds yields the desired result from the introduction. Evaluating for the Frobenius norm $\Vert A\Vert=\sqrt{\text{tr}\{A^{\dagger}A\}}$, which is used throughout this work, we find $$\begin{aligned}
\label{fundermentalultimatebound}
\frac{\sqrt{2(d-\Re[\text{tr}\{U_{2}^{\dagger}(T)U_{g}\}])}}{\Vert H_{0}\Vert }\leq T,\end{aligned}$$ and since $\Re[\text{tr}\{U_{2}^{\dagger}(T)U_{G}\}]\leq \sum_{j}^{d}|{\langle{\phi_{j}^{(c)}}|}U_{g}{|{\phi_{j}^{(c)}}\rangle}|$ with $\{{|{\phi_{j}^{(c)}}\rangle}\}_{j=1}^{d}$ being the eigenbasis of $H_{c}$, we further conclude $$\begin{aligned}
\label{eq:speedlimit}
\frac{\sqrt{2(d-\sum_{j}^{d}|{\langle{\phi_{j}^{(c)}}|}U_{g}{|{\phi_{j}^{(c)}}\rangle}|)}}{\Vert H_{0}\Vert} \leq T.\end{aligned}$$ Similar to the lower bounds that were obtained in [@SpeedLimit1; @SpeedLimit2; @SpeedLimit3], the above inequality is a lower bound for the least time needed to implement a given target unitary gate. Henceforth, we refer to this time as the minimum gate time. The speed with which a desired given unitary can be implemented is inherently limited by the speed with which the propagator $U(t)$ evolves under the free evolution alone. Unless one wants to implement a gate that can be reached by the control and $H_{c}$ alone, which can be done instantaneously if we assume that the amplitude of the control field is unconstrained, the strength of the drift Hamiltonian sets an “intrinsic” limit on how fast we can reach the desired target. However, typically the amplitude of the control field is limited in any experimental situation, and therefore a practical lower bound for $T$ must also depend on the highest control field amplitude $f_{\text{max}}$. Analogous to the derivation of , but now with $H_{2}=H_{0}$, this can be established by using $\int_{0}^{T}\Vert f(t)H_{c}\Vert\,dt\leq T |f_{\text{max}}|\Vert H_{c}\Vert$. We find $$\begin{aligned}
\label{eq:speedlimitfield}
\frac{\sqrt{2(d-\sum_{j}^{d}|{\langle{\phi_{j}^{(0)}}|}U_{g}{|{\phi_{j}^{(0)}}\rangle}|)}}{|f_{\text{max}|}\Vert H_{c}\Vert} \leq T,\end{aligned}$$ where $\{{|{\phi_{j}^{(0)}}\rangle}\}_{j=1}^{d}$ is the eigenbasis of $H_{0}$. Conversely to , the bound represents a speed limit that is enforced by limitations of the control field (extrinsic), rather than intrinsic limitations given by the strength of the drift Hamiltonian. We postpone the discussion about the distinction between extrinsic and intrinsic speed limits to section \[ref:discussionextrinsicintrinsic\] and proceed by defining $$\begin{aligned}
\label{eq:defC1}
C(U_{g},H_{c})&\equiv \frac{\sqrt{2(d-\sum_{j}^{d}|{\langle{\phi_{j}^{(c)}}|}U_{g}{|{\phi_{j}^{(c)}}\rangle}|)}}{2},\\
\label{eq:defC2}
C(U_{g},H_{0})&\equiv \frac{\sqrt{2(d-\sum_{j}^{d}|{\langle{\phi_{j}^{(0)}}|}U_{g}{|{\phi_{j}^{(0)}}\rangle}|)}}{2}.\end{aligned}$$ From and we then find $2C(U_{g},H_{c})/\Vert H_{0}\Vert+2C(U_{g},H_{0})/(|f_{\text{max}}|\Vert H_{c}\Vert)\leq 2T$. As such, the lower bound from the introduction is obtained by linearly combining and . We note that each term of the left-hand side of is weighted in a different manner by the target operation. As described in [@meijfcs], there are many ways to combine two or more speed limit formulas to create novel ones. As in [@MaxSpeedlim], simply taking the maximum of and yields $T~\geq~\max\{2C(U_{g},H_{c})/\Vert H_{0}\Vert,~2C(U_{g},H_{0})/(|f_{\text{max}}|\Vert H_{c}\Vert)\}$. Another method is to take convex combinations, which is done in this work using an equal weighting. The authors have not as of yet determined the combination that produces the tightest bound. However, the numerical simulations in section \[ref:pareto\] suggest that this choice is worth investigation and moreover, it is conjectured in section \[sec:tightness\] that speed limits for the control system generally should be of this form. To summarize, the inequality can be considered as a necessary condition which must be satisfied by $\Vert H_{0}\Vert,~\Vert H_{c}\Vert,~f_{\text{max}}$ and $T$ in order to implement some $U_{g}\in\overline{\mathcal R}$.
Characterization of the reachable set
-------------------------------------
Since the lower bound depends on the target unitary transformation $U_{g}$, it reveals some information about the set of gates $\mathcal G_{T}$, which *provably* cannot be reached for a given evolution time $T$. For instance, consider the simplified case of implementing a gate that can be reached by the control Hamiltonian alone, i.e. $U_{g}=\exp(-i\alpha(T)H_{c})$. This yields $C(U_{g},H_{c})=0$. Here the drift Hamiltonian is not required to reach the target evolution. However, if the control field is not sufficiently large the gate cannot be implemented.
### Single qubit case:
in order to study more complex cases requiring an interplay between the drift and the control Hamiltonian, we consider a single qubit described by the Hamiltonian $$\begin{aligned}
\label{eq:singlequbit}
H(t)=\Omega \sigma_{x}+f(t)\sigma_{z}, \end{aligned}$$ where $\sigma_{j}$, with $j=x,y,z$, are the Pauli matrices. We remark that the reachable set of a single qubit subject to two independent control fields was recently analyzed in great detail [@ReachQubit1; @ReachQubit2]. The system is fully controllable hence every $U_{g}\in \text{SU}(2)$ can be implemented. We can parameterize a $U_{g}=R_{z}(\alpha)R_{y}(\gamma)R_{z}(\beta)$ with the three angles $0\leq \alpha < 2\pi,~0\leq \beta < 4\pi,~0\leq \gamma \leq \pi$ where $R_{z}(\alpha)=\exp(-i\frac{\alpha}{2} \sigma_{z})$ and $R_{y}(\gamma)=\exp(-i\frac{\gamma}{2} \sigma_{y})$ are rotations around $\sigma_{z}$ and $\sigma_{y}$ respectively such that $C(U_{g})=C(\alpha,\beta,\gamma)$. For an unconstrained control field, the time required to implement some $U_{g}\in \text{SU}(2)$ can be calculated exactly [@OptimalControlSpeedLimit2] using the Euler angle decomposition above. This calculation shows that the minimum time is determined by $\Omega$. With the established bound , we can now proceed with analyzing the effect of a constrained control field amplitude. For $\beta=0$, i.e when every state can be reached from an initial eigenstate of $\sigma_{z}$, the lower bound takes the form $$\begin{aligned}
\label{eq:reachablesstatesqubit}
\frac{\sqrt{(2-2\cos(\gamma/2))}}{4|\Omega|}+\frac{\sqrt{3-\cos(\alpha)\cos(\gamma)}}{4|f_{\text{max}}|}\leq T.\end{aligned}$$ Fig. \[fig:Qubit\] a) shows the set of states, parametrized through $\gamma$ and $\alpha$, that do not satisfy (grey area) for a fixed evolution time $T=0.53$ and $\Omega=f_{\text{max}}=1$, hence these states cannot be reached. The white area contains all states that do satisfy , nonetheless, this does not reveal whether they are reachable or not. This question is related to the tightness of the bound , which will be analyzed in the next section \[ref:pareto\].
A way to study which gates are provably not reachable as a function of $T$ for a given $\Omega,~f_{\text{max}}$ is to consider the volume of the set $\mathcal G_{T}$. For $\text{SU}(2)$, the measure such that the volume of the entire group is one, is given by $dV=\frac{1}{16\pi^{2}}\sin(\gamma)\,d\gamma\,d\alpha\,d\beta$. As such, for a single qubit the volume $V(\mathcal G_{T})$ can be calculated as $$\begin{aligned}
\label{eq:volumedef}
V(\mathcal G_{T})=\frac{1}{16\pi^{2}}\int_{\Sigma(\mathcal G_{T})}\sin(\gamma)\,d\gamma\,d\alpha\,d\beta, \end{aligned}$$ where $$\begin{aligned}
\label{eq:intregion}
\Sigma(\mathcal G_{T})=\left\{(\alpha,\beta,\gamma)\,\Big|\, \frac{C(\alpha,\beta,\gamma,\sigma_{z})}{\sqrt{2}|\Omega|}+\frac{C(\alpha,\beta,\gamma,\sigma_{x})}{\sqrt{2}|f_{\text{max}}|}\leq T \right\}, \end{aligned}$$ is the integration region. In Fig. \[fig:Qubit\] b) we numerically integrated for different values $T$. The solid black line shows the case where $\Omega=f_{\text{max}}=1$ and the dashed lines (dashed-dotted lines) show the cases where $|\Omega|<|f_{\text{max}}|$ ($|\Omega|>|f_{\text{max}}|$) with $f_{\text{max}}=1$.
a)\
![\[fig:Qubit\] Characterization of the gates that proveably cannot be reached for a single qubit described by the control system . Based on , a) shows the set of states that cannot be reached (grey area) from an initial eigenstate of $\sigma_{z}$ for an evolution time $T=0.52$ and $\Omega=f_{\text{max}}=1$. Fig. b) shows the volume of gates ( and ) that cannot be reached as a function of the evolution time $T$. The ratio $\frac{|\Omega|}{|f_{\text{max}}|}\in\{10,~2,~1,~0.5,~0.25\}$ was chosen from left to right in decreasing order, whereas the solid black line represents $\Omega=f_{\text{max}}=1$.](Reachablesetqubit.pdf "fig:"){width="1\columnwidth"}
b)\
![\[fig:Qubit\] Characterization of the gates that proveably cannot be reached for a single qubit described by the control system . Based on , a) shows the set of states that cannot be reached (grey area) from an initial eigenstate of $\sigma_{z}$ for an evolution time $T=0.52$ and $\Omega=f_{\text{max}}=1$. Fig. b) shows the volume of gates ( and ) that cannot be reached as a function of the evolution time $T$. The ratio $\frac{|\Omega|}{|f_{\text{max}}|}\in\{10,~2,~1,~0.5,~0.25\}$ was chosen from left to right in decreasing order, whereas the solid black line represents $\Omega=f_{\text{max}}=1$.](Volume1.pdf "fig:"){width="1\columnwidth"}
### The time required to implement all gates:
from Fig. \[fig:Qubit\] we observe that when the evolution time becomes larger the number of gates that proveably *cannot* be implemented becomes monotonically *smaller*. It is known (Theorems 1 and 3 [@BookJurdjevic], which lead to the result in [@qclcl]) that for a fully controllable system there exist a time $T_{c}$ for which all gates can be implemented. In order to establish a lower bound for $T_{c}$ we seek the gate for which $C(U_{g},H_{0})$ and $C(U_{g},H_{c})$ become maximal. In the single qubit case this can be achieved simply through $U_{g}=\sigma_{y}$. Unfortunately, unless $H_{0}$ and $H_{c}$ have eigenbasis which are mutually unbiased [@MUTU], finding maximizing $U_{g}$ in general remains an open problem. However, for a fully controllable qubit system there always exists a $U_{g}$ (see \[app:derivationLB\] for further details) for which $\Re[\text{tr}\{U_{g}^{\dagger}U^{(i)}\}]\leq d/2$ with $i=1,2$ where $U^{(1)}=\exp(-i\alpha(T_{c})H_{c})$ and $U^{(2)}=\exp(-iT_{c}H_{0})$. With $\Vert H_{0}\Vert\leq \sqrt{d}|E_{0}|$ and $\Vert H_{c}\Vert\leq \sqrt{d}|E_{c}|$, where $E_{0}$ and $E_{c}$ are the highest eigenvalue of $H_{0}$ and $H_{c}$ respectively, we thus have $$\begin{aligned}
\label{eq:criticaltimebound}
\frac{1}{2|E_0|}+\frac{1}{2|f_{\text{max}}E_c|} \leq T_{c}.\end{aligned}$$ For $E_0,~E_c,$ and $f_{\text{max}}$ determined by the quantum system and experimental control limitations, respectively, the bound determines how at least much time is in order to be able to implement all gates. In particular, an obstacle to fully controlling the system on a implementable time scale arises when the norm of $H_{0}$ decreases with an increase of the dimension of the system, or, when $f_{\text{max}}$ is not sufficiently large.
In order to be able to implement all gates, $E_0$, $E_{c}$ and $f_{\text{max}}$ must be given in such a way that $T_{c}$ does not reach an order within which other effects, such as decoherence, cannot be neglected. Denoting by $T_{\text{Dec}}$ the typical decoherence time scale, $|2E_0|^{-1}+|2f_{\text{max}}E_c|^{-1}\leq T_{\text{Dec}} $ needs to be satisfied in order to be able to implement all gates. We note here that this is a heuristic argument rather than a rigorous conclusion as the application of control fields can substantially change the effect of the environment [@Kurizki]. For example, in the extreme case of an infinitely fast decoupling sequence the effect of the environment can be completely suppressed for a large class of system-environment interactions [@LVioal1; @Dec1me]. In such a way coherence times can be significantly prolonged [@QubitinSolid]. Opposingly, as shown in [@SchmidtCalarco; @KochReich2; @Me], sometimes the environment and noise that is caused by it are beneficial, even turning the system into a fully controllable one [@Me]. A detailed and rigorous analysis, including the interaction with an environment, is beyond the scope of this work and will be the subject of future studies.
Tightness of the bound and Pareto optimal control {#ref:pareto}
-------------------------------------------------
As mentioned in the introduction, the aim of optimal control theory is to find a pulse that maximizes or minimizes a given cost functional. For the implementation of a target unitary gate $U_{g}$ this is typically done by minimizing the infidelity $$\begin{aligned}
\label{eq:infidelity}
\epsilon=1-\left|\frac{1}{d}\text{tr}\{U_{g}^{\dagger}U(T)\}\right|^{2},\end{aligned}$$ using a gradient based search [@Grape; @DYNAMOpaper; @GateFid2; @exactgradient], such as the GRAPE algorithm [@Grape]. In order to study the tightness of the bounds and we employ in this section numerical minimization of the infidelity $\epsilon$ using the GRAPE algorithm in the QuTip control package [@QuTip1; @QuTip2]. We begin by analyzing the bound , i.e., the case in which the control field amplitude is not constrained such that the speed limits only arise from the limited strength of the drift Hamiltonian. We study the single qubit example from the previous section for a target evolution $U_{g}=\sigma_{y}$ and an N-level system which is known to be fully controllable [@FullControlSingleAcc]. The drift Hamiltonian reads $H_{0}=J\sum_{j=1}^{N-1}({|{j}\rangle}{\langle{j+1}|}+\text{h.c.})$ where control is exerted trough $H_{c}={|{1}\rangle}{\langle{1}|}$ and as a target evolution we consider the SWAP gate $U_{g}=\exp(-i\pi/2({|{1}\rangle}{\langle{N}|}+{|{N}\rangle}{\langle{1}|}))$.
In Fig. \[fig:ParetoNum\] a) we show the minimum gate time as a function of the norm of the drift Hamiltonian, where the inset plot shows the N-level system with $N=4$ levels and $H_{0}$ was normalized in such a way that $\Vert H_{0}\Vert=J$. In both cases the grey curves represent the lower bound and the numerically estimated values for the minimum gates times (black diamonds) were obtained by minimizing $\epsilon$ for different values of the total evolution time $T$ until a threshold of $\epsilon< 10^{-7}$ is reached. As mentioned in the previous section, for the single qubit control system with an unbounded control field amplitude the minimum gate time $T^{*}$ can be calculated exactly [@OptimalControlSpeedLimit2], yielding for $U_{g}$ from above $T^{*}=\frac{\pi}{2|\Omega|}$ (blue curve in Fig. \[fig:ParetoNum\] a)). We emphasize that the numerically obtained values are themselves only an upper bound since the convergence of the optimization algorithm depends on the initial trial pulse, which was chosen randomly in all cases. Nevertheless, from Fig. \[fig:ParetoNum\] a) we conclude that, remarkably, the lower bound is tight for the single qubit control system. Unfortunately, as indicated by the inset in Fig. \[fig:ParetoNum\] a), this is less satisfactory when the dimension of the quantum system increases.
a)\
![\[fig:ParetoNum\] Numerical gate optimization of the infidelity $\epsilon$ given by to study the tightness of the bounds (grey curves) and (black curve) shown in a) and b) respectively. a) Minimum gate time as a function of the strength $\Omega$ and $J$ of the drift Hamiltonian on a logarithmic scale for the single qubit control system with a target evolution $U_{g}=\sigma_{y}$ and (inset plot) a N-level system (details can be found in the main text) with $N=4$ levels and a SWAP gate as a target evolution. The diamonds show the numerically obtained values and in the single qubit case the exact value of the minimum gate time (blue curve) is given by $T^{*}=\frac{\pi}{2|\Omega|}$. b) Numerical gate optimization of the infidelity for different values of the evolution time $T$ and the highest permitted control field amplitude $f_{\text{max}}$ evaluated for a single qubit with $U_{g}=\sigma_{y}$ and $\Omega=1$. The colormap shows the infidelity, where in the orange region, $\epsilon<10^{-10}$ is achieved.](SpeedLimit.pdf "fig:"){width="1.0\columnwidth"}
b)\
![\[fig:ParetoNum\] Numerical gate optimization of the infidelity $\epsilon$ given by to study the tightness of the bounds (grey curves) and (black curve) shown in a) and b) respectively. a) Minimum gate time as a function of the strength $\Omega$ and $J$ of the drift Hamiltonian on a logarithmic scale for the single qubit control system with a target evolution $U_{g}=\sigma_{y}$ and (inset plot) a N-level system (details can be found in the main text) with $N=4$ levels and a SWAP gate as a target evolution. The diamonds show the numerically obtained values and in the single qubit case the exact value of the minimum gate time (blue curve) is given by $T^{*}=\frac{\pi}{2|\Omega|}$. b) Numerical gate optimization of the infidelity for different values of the evolution time $T$ and the highest permitted control field amplitude $f_{\text{max}}$ evaluated for a single qubit with $U_{g}=\sigma_{y}$ and $\Omega=1$. The colormap shows the infidelity, where in the orange region, $\epsilon<10^{-10}$ is achieved.](ParetoFrontPlotComp.pdf "fig:"){width="1.0\columnwidth"}
Besides minimizing $\epsilon$, sometimes there are additional constraints one must take into account. For instance, one wants to find the optimal control pulses that minimize $\epsilon$, while keeping the length of the pulses as short as possible and additionally using the least amount of energy. Such a multi-objective optimization is known as Pareto optimal control [@BookHillermeier; @BookStatMat; @Hersch]. Typically one seeks to identify non-dominated solutions, and it is generally not possible to achieve fully optimal solutions that maximize all objectives [@Hersch]. We consider here the situation where we (i.e., $\epsilon=0$) want to perfectly implement some $U_{g}$ in the shortest possible time $T$ while simultaneously constraining the highest control field amplitude $f_{\text{max}}$ as much as possible. Clearly, there is a trade off between the three objectives, meaning that we lose fidelity when we constrain either the control field or the evolution time too much. A general question is how much we can reduce the length and the highest amplitude of the pulse, while still being able to implement $U_{g}$. Rearrangement of yields $$\begin{aligned}
\label{eq:boundTF}
T-\frac{1}{|f_{\text{max}}|}\frac{C(U_{g},H_{0})}{\Vert H_{c}\Vert}\geq \frac{C(U_{g},H_{c})}{\Vert H_{0}\Vert}, \end{aligned}$$ which defines a region that characterizes how much $T$ and $f_{\text{max}}$ can be minimized while still begin able to implement some $U_{g}$. Clearly, the inequality is a lower bound for the minimization over $T$ and $f_{\text{max}}$, whereas the actual smallest values for which the infidelity $\epsilon$ is still zero might be larger. In the following we therefore want to analyze the tightness of the lower bound . Again, we resort to numerical gate optimization of the infidelity $\epsilon$, but now for different values of $T$ and $f_{\text{max}}$. We focus on a single qubit described though the control system with $\Omega=1$. As a target evolution we again take $U_{g}=\sigma_{y}$, which yields $T-1/(\sqrt{2}|f_{\text{max}}|)\geq 1/(\sqrt{2})$. The results are shown in Fig. \[fig:ParetoNum\] b) wherein the black curve represents the lower bound and the orange area represents the achievement of infidelities of $\epsilon<10^{-10}$. The points that seem to break the continuity in the orange region are numerical artifacts, which can be removed by additionally minimizing over different initial pulses. From Fig. \[fig:ParetoNum\] b) we observe that the bound is tight too for the single qubit control system, but, as before, similar simulations for higher dimensional systems indicate that tightness is lost.
Discussion
==========
The bound has been shown to be an excellent approximation to the numerically obtained values in the case of a single qubit system. However, for higher dimensional systems, the bound becomes diminishingly tight. In the following we want to discuss the reasons that this appears to be the case and a way forward to obtain tighter bounds for higher dimensional systems. Moreover, we make the distinction between two types of bounds for the minimum gate time, to which we refer as *intrinsic* and *extrinsic*. They arise from limitations of the drift or the control Hamiltonian (intrinsic) or some limited control resources (extrinsic), such as constrained control fields.
Tightness {#sec:tightness}
---------
The bound can be considered as a *first order* approximation to the true minimum gate times. The approximating step can be traced back to the use of the triangle inequality in the derivation of (see \[app:derivationNielsen\], Eq. ). In this step higher order commutator expressions contributing to the trajectory of the unitary propagator $U(t)$ are disregarded. By observing the nature of the terms in , we see that while $\Vert H_0 \Vert$ (or $\Vert H_{c}\Vert$) and the eigenbasis of $H_c$ ($H_{0}$) both appear, the norm of the commutator $[iH_0, iH_c]$ does not appear. Furthermore, all additional commutator terms of $iH_0$ and $iH_c$ are also absent. For controllable systems, the set of all such nested commutator expressions [@Altafini] must generate the whole algebra $\mathfrak{u}(d)$. For a visual example of the way in which such bracket expressions appear, see the ‘Lie tree’ diagrams in [@Carenzspinstar]. As such, a critical part of the dynamics of a system evolution is disregarded by any bound on minimum gate times (or any other QSL formula used in quantum control) which does not take these additional commutator terms, that is the structure of the underlying dynamical Lie algebra, into account. In the same work numerical gate optimization suggests that the minimum gate time for a specific model scales exponentially with the number of qubits. As the dimension of a system rises, the nested commutator depth required to span the full algebra $\mathfrak{u}(d)$ grows [@Carenzspinstar; @Elliot]. The authors conjecture that tighter bounds on minimum gate times in terms of maximum control field amplitudes can be obtained by incorporating higher order terms of nested commutator expressions. As such, the authors anticipate the possibility of establishing tighter bounds of the form $$\begin{aligned}
\label{eq:formtigherbound}
\frac{L^{\bot}}{\Vert H_{0}\Vert}+\frac{L}{|f_{\text{max}}|\Vert H_{c}\Vert}\leq T, \end{aligned}$$ where $L$ and $L^{\bot}$ are two contributions to the length of the time optimal trajectory connecting $U(0)=\mathds{1}$ and $U(T)=U_{g}$. In [@ExactCalc1; @ExactCalc2] it was shown that the group is covered (formally, foliated) by a family of subsets, known as *cosets*, which play a crucial role in characterizing the time optimal trajectories in systems with unbounded controls. These subsets are related to the unitary operations corresponding to the control Hamiltonians alone; evolutions of arbitrarily high speed are possible within these sets using the controls alone, provided that the controls are unconstrained. Moreover, the total length of any time optimal trajectory splits into two contributions. Firstly, $L^{\bot}$, the length of the trajectory between the cosets (i.e., orthogonal to each coset) and secondly, $L$, the length of the trajectory within cosets. The speed at which a time optimal trajectory can be traversed also splits into two parts, namely $\Vert H_{0}\Vert$, the speed between cosets and, the speed of the evolution within cosets. For a constrained control field this speed is bounded by $|f_{\text{max}}|\Vert H_{c}\Vert$. Hence tighter bounds than are expected to be still of the form .
Intrinsic and extrinsic speed limits {#ref:discussionextrinsicintrinsic}
------------------------------------
Many works have recently focused on determining minimum gate times, or quantum speed limits. We emphasize here two clear types of bounds which are in regular use, but which have not yet been clearly delineated or contrasted. We first want to distinguish quantum control systems which are fully controllable only in the presence of a drift term (i.e., mathematically removing the drift would cause the system to no longer be fully controllable) from those systems for which this is not the case. Systems of the latter class are known as *strongly controllable* [@Elliot], i.e. they are fully controllable with controls alone regardless of the presence or absence of any drift term. In the case of controllable, but not strongly controllable systems, there is an *intrinsic* quantum speed limit. This is to say, the minimum gate time (over all control fields without any constraints) ultimately has its physical origin in the fact that the implementation of the gate requires exploiting the drift term which is not directly under control, and is bounded in strength. These speed limits are of the form $T > F(H_0, H_c, U_{g})$. This situation is to be contrasted with systems having a constrained control field $f(t)$, for which the bound on the minimum time arises as a consequence of limitations of the control field, such as bounded amplitude, limited bandwidths, power and energy constraints. These bounds are of the form $T>F(H_0, H_c, U_{g},\mathcal F)$ where $\mathcal F$ is the set of admissible controls, and are *extrinsic*, in the sense that they arise not only from limitations on the system Hamiltonian itself, but also from constraints on the control fields. Typically quantum speed limits in the literature are of the former type since they are not formulated within context of quantum control theory [@SpeedLimit1; @SpeedLimit2; @SpeedLimit4; @SpeedLimit5; @SpeedLimit6]. We remark that, contrastingly, every constraint on the overall Hamiltonian $H(t)$ of a controlled quantum system potentially yields a speed limit [@SpeedLimit3]. However, the latter type of limit remains insufficiently investigated, despite being of critical importance for practical applications of quantum control. The distinction between intrinsic and extrinsic limits identifies two significantly different types of actionable information in a quantum control scenario. The extrinsic case indicates when constrained control fields are the limiting factor, whereas the intrinsic limit indicates a physical boundary which cannot be crossed for a given quantum system no matter what type of control is employed.
The bound derived in this work is of both the intrinsic and extrinsic type. In the limit $|f_{\text{max}}|\to\infty$, the bound furnishes information purely about the intrinsic speed limit of a given system as the term containing $f_{\text{max}}$ vanishes. Additionally, in the limit that the term corresponding to the control Hamiltonian goes to zero only the term depending on the drift Hamiltonian persists, and thus the remaining term represents the intrinsic limit.
Conclusions and Outlook
=======================
We have derived a lower bound for the time required to implement a unitary gate through a classical control field. The bound depends on the strength of the drift and the control Hamiltonian, the highest permitted control field amplitude and the target gate one wants to implement. The derived bound can be considered as an extrinsic quantum speed limit since the minimum time to implement a target unitary gate is limited by the maximum control field amplitude and the strength of the internal Hamiltonian. However, if we allow the control field to be unconstrained, the bound yields an intrinsic quantum speed limit that cannot be crossed, since the speed of the evolution is limited by the norm of the drift Hamiltonian.
The results in this work are a step towards characterizing the reachable set of gates given a certain evolution time, and thus further establishing a bound on the minimum time $T_{c}$ needed to implement *all* gates. We have provided a criterion for assessing the time $T_c$ at which all gates are reachable in a given system. This observation has implications for the control landscape [@Land1] of the same quantum systems, since the the non-existence of traps (local minima/maxima of the objective functional considered as a function of control(s)) crucially depends on the assumption that the evolution time is sufficiently long to be able to implement all gates (see [@Land2] and references therein). Moreover, using numerical gate optimization, we found that the derived bound is remarkably tight for a single qubit control system; unfortunately this is no longer the case for higher dimensional systems. We argued that this behavior originates from the underlying structure of the dynamical Lie algebra of the control system, particularly the norms of nested commutators and their relation to the desired gate. Furthermore, the interplay between small matrix elements in the drift Hamiltonian and minimum gate times also warrants further investigation.
In this work, and many other works on the quantum speed limit, it is assumed that the time being sought is the minimum time to *perfectly* implement a specific desired unitary operation $U_{g}$, i.e., that the task of interest is to find a pulse which solves $U(T)=U_{g}$. However, if some error is allowed in the implementation of a gate, does the corresponding minimum gate time defer radically? The authors conjecture that this is the case for gates within a small neighborhood of any given fast gates, i.e., gates which can be implemented by the controls alone. Furthermore, recently it has been shown that simple analytically obtained pulses can lead to high fidelity gates $\epsilon\approx 0.01$ [@Vir1; @Vir2]. Further numerical control optimization yields yet higher fidelities at the cost of requiring significantly higher frequency components within the numerically optimized pulse [@Vir2]. It would be favorable to obtain criteria characterizing the set of gates for which “simple” pulses exist and further to understand the highest frequency required in a pulse to implement a gate perfectly.
C. A. acknowledges the NSF (grant CHE-1464569) and fruitful discussions with Thomas-Schulte Herbrüggen and Robert Zeier. D.B. acknowledges support from the EPSRC Grant No. EP/M01634X/1. B.R. acknowledges the DOE (grant DE-FG-02ER15344) and H. R. acknowledges the ARO (grant W911NF-16-1-0014).
Derivation of the inequality {#app:derivationNielsen}
=============================
Here we verify the inequality from the main text based on the work in [@Geo1], which is valid for any unitarily invariant matrix norm, i.e. $\Vert VAU\Vert=\Vert A\Vert $ with $V,U$ being unitary. Consider $$\begin{aligned}
\frac{d}{dt}(U_{1}^{\dagger}(t)U_{2}(t))&=U_{1}^{\dagger}(t)(iH_{1}(t))U_{2}(t)+U_{1}^{\dagger}(t)(-iH_{2}(t))U_{2}(t) \nonumber \\
&=iU_{1}^{\dagger}(t)(H_{1}(t)-H_{2}(t))U_{2}(t),\end{aligned}$$ such that with $U_{1}(0)=U_{2}(0)=\mathds{1}$ integrating yields $$\begin{aligned}
U_{1}^{\dagger}(t)U_{2}(t)-\mathds{1} =-i\int_{0}^{t}(U_{1}^{\dagger}(t^{\prime})(H_{1}(t^{\prime})-H_{2}(t^{\prime}))U_{2}(t^{\prime}))dt^{\prime}. \end{aligned}$$ We note that for any unitarily invariant norm we have $\Vert U_{2}^{\dagger}(t)U_{1}(t)-\mathds{1} \Vert=\Vert U_{1}(t)-U_{2}(t)\Vert $. Using the triangle inequality and unitary invariance again, $$\begin{aligned}
\Vert U_{1}(t)-U_{2}(t)\Vert &=\left\Vert -i\int_{0}^{t}(U_{1}^{\dagger}(t^{\prime})(H_{1}(t^{\prime})-H_{2}(t^{\prime}))U_{2}(t^{\prime}))dt^{\prime} \right\Vert \nonumber\\
\label{eq:usetriangle}
&\leq \int_{0}^{t} \Vert U_{1}^{\dagger}(t^{\prime})(H_{1}(t^{\prime})-H_{2}(t^{\prime}))U_{2}(t^{\prime}) \Vert dt^{\prime} \\ \nonumber
&=\int_{0}^{t} \Vert H_{1}(t^{\prime})-H_{2}(t^{\prime})\Vert dt^{\prime},\end{aligned}$$ we hence arrive at the desired result .
Derivation of the lower bound {#app:derivationLB}
==============================
In order to derive the lower bound for the time $T_{c}$ at which all gates become reachable we first show that there always exist a $U_{g}\in \text{U}(d)$ with $d$ being even for which $$\begin{aligned}
\label{eq:one}
\Re[\text{tr}\{U_{g}^{\dagger}U^{(i)}\}]\leq \frac{d}{2},~~~~i=1,2,\end{aligned}$$ holds where $U^{(i)}$ is the unitary evolution generated by the drift and the control Hamiltonian respectively at the time $T_{c}$. Using the eigenbasis $\{{|{\varphi_{j}}\rangle}\}_{j=1}^{d}$ of $U_{g}$ with eigenvalues $\exp(-i\lambda_{j})$ respectively the left hand side can be rewritten as $$\begin{aligned}
&\Re[\text{tr}\{U_{g}^{\dagger}U^{(i)}\}]=\sum_{j=1}^{d}\cos(\lambda_{j}+\phi_{j}^{(i)})|{\langle{\varphi_{j}}|}U^{(i)}{|{\varphi_{j}}\rangle}| \nonumber \\
&\leq \frac{d}{2}+\sum_{j~\text{even}~\vee~\text{odd}}\cos(\lambda_{j}+\phi_{j}^{(i)})|{\langle{\varphi_{j}}|}U^{(i)}{|{\varphi_{j}}\rangle}|, \end{aligned}$$ where ${\langle{\varphi_{j}}|}U^{(i)}{|{\varphi_{j}}\rangle}=e^{\phi_{j}^{(i)}}|{\langle{\varphi_{j}}|}U^{(i)}{|{\varphi_{j}}\rangle}|$ was used. We observe that $\lambda_{j}$ can always be chosen in such a way that the sum of the right hand side becomes zero. For example take $\lambda_{j}=-\phi_{j}^{(1)}+\frac{\pi}{2}$ for $j$ even and $\lambda_{j}=-\phi_{j}^{(2)}+\frac{\pi}{2}$ for $j$ odd. Thus, for any target evolution $U_{g}$ constructed in this way, inequality holds. Now, for a fully controllable system there always exists a control pulse such that for the time $T_{c}$ the gate $U_{g}$ from above is implemented. Applying , we thus have $$\begin{aligned}
&\sqrt{d}\leq \Vert U_{g}-U^{(2)}\Vert \leq T_{c}\Vert H_{0}\Vert \leq T_{c} \sqrt{d}|E_0|, \\
&\sqrt{d}\leq \Vert U_{g}-U^{(1)}\Vert \leq \int_{0}^{T_{c}}\Vert f(t)H_{c}\Vert\,dt \leq T_{c} \sqrt{d}|f_{\text{max}}E_c|, \nonumber \end{aligned}$$ where $E_0$ and $E_c$ are the highest eigenvalues of $H_{0}$ and $H_{c}$, respectively. Combining the bounds from above we hence find $$\begin{aligned}
\frac{1}{2|E_0|}+\frac{1}{2|f_{\text{max}}E_c|} \leq T_{c}.\end{aligned}$$
References {#references .unnumbered}
==========
[99]{}
S. J. Glaser, U. Boscain, T. Calarco, C. P. Koch, W. Köckenberger, R. Kosloss, I. Kuprov, B. Luy, S. Schirmer. T. H. Herbrüggen, D. Sugny, and F. K. Wilhelm, Eur. Phys. J. D **69**, 1-24 (2015).
C. Brif, R. Chakrabarti and H. Rabitz, New. J. Phys. **12**, 075008 (2010).
R. J. Levis, G. M. Menkir and H. Rabitz, Science **292**, 5517 (2001).
F. Dolde, V. Bergholm, Y. Wang, I. Jakobi, B. Naydenov, S. Pezzagna, J. Meijer, F. Jelezko, P. Neumann, T. Schulte-Herbrüggen, and J. Biamonte, Nat. Com. **5**, (2014).
A. Soare, H. Ball, D. Hayes, J. Sastrawan, M. C. Jarratt, J. J. McLoughlin, X. Zhen, T. J. Green, and M. J. Biercuk, Nat. Phys. **11**, 825-829 (2014).
D. D’Allesandro, *Introduction to Quantum Control and Dynamics*, Chapman & Hall, New York (2008).
V. Jurdjevic, *Geometric Control Theory*, Cambridge University Press, Cambridge (1997).
G. Dirr, U. Helmke, I. Kurniawan and T. Schulte-Herbrüggen, Rep. Math. Phys. **64**, 93-121 (2009).
H. Sussmann and V. Jurdjevic, J. Diff. Equat. **12**, 313 (1972).
W. Zhu, J. Botina, and H. Rabitz, J. Chem. Phys. **108**, 1953 (1998).
N. Khaneja, T. Reiss, C. Kehlet, T. Schulte-Herbrüggen and S. J. Glaser, J. Magn. Reson. **172**, 296-305 (2005).
S. Manches, U. Sander, S. J. Glaser, P. de Fouquieres, A. Gruslys, S. G. Schirmer and T. Schulte-Herbrüggen, Phys. Rev. A **84**, 022305 (2011).
F. F. Floethe, P. de Fouquieres and S. G. Schirmer, New. J. Phys. **14**, 073023 (2012).
R. Nigmatullin and S. G. Schirmer, New. J. Phys **11**, 105032 (2009).
N. Margolus and L. B. Levitin, Physica D **120**, 188 (1998).
V. Giovannetti, S. Lloyd, and L. Maccone, Phys. Rev. A **67**, 052109 (2003).
B. Russell and S. Stepney, arXiv:1410.3209 \[quant-ph\].
J. P. Jones and P. Kok, Phys. Rev. A **82**, 022107 (2010).
P. Kosiński and M. Zych, Phys. Rev. A **73**, 024303 (2006).
M. H. Yung, Phys. Rev. A **74**, 030303 (2006).
S. Kallush, M. Khasin, and R. Kosloff, New J. Phys. **16**, 015008 (2014).
M. H. Goerz, T. Calarco and C. P. Koch, J. Phys. B: At. Mol. Opt. Phys. **44**, 154011 (2011).
A. del Campo, I. L. Egusquize, M. B. Plenio and S. F. Huelga, Phys. Rev. Lett. **110**, 050403 (2013).
S. Deffner and E. Lutz, Phys. Rev. Lett. **111**, 010402 (2013).
M. M. Taddei, B. M. Escher, L. Davidovich, Phys. Rev. Lett. **110**, 050402 (2013).
C. Liu, Z. Y. Xu, and S. Zhu, Phys. Rev. A **91**, 022102 (2015).
D. P. Pires, M. Cianciaruso, L. C. Céleri, G. Adesso, Phys. Rev. X **6**, 021031 (2016).
J. F. Epstein and B. K. Whaley, arXiv: 1612.04767 \[quant-ph\].
T. Caneva, M. Murphy, T. Calarco, R. Fazio, S. Monangero, V. Giovanetti and G. E. Santoro, Phys. Rev. Lett. **103**, 240501 (240501).
G. C. Hegerfeldt, Phys. Rev. Lett. **111**, 260501 (2013).
N. Khaneja, R. Brockett, and S. J. Glaser, Phys. Rev. A **63**, 032308 (2001).
N. Khaneja, S. J. Glaser and R. Brockett, Phys. Rev. A **65**, 032301 (2002).
A. Carlini, A. Hosoya, T. Koike and Y. Okudaira, Phys. Rev. Lett. **96**, 060503 (2006).
C. Arenz, G. Gualdi and D. Burgarth, New. J. Phys. **16**, 065023 (2014).
S. Ashhab, P. C. De Groot, and F. Nori, Phys. Rev. A **85**, 052327 (2012).
M. A. Nielsen, M. R. Downling, M. Gu and A. C. Doherty, Science **311**, 5764 (2006).
B. Russell and S. Stepney, J. Phys. A **48**, 115303 (2015).
B. Russell and S. Stepney, Phys. Rev. A **90**, 012303 (2014).
D. C. Brody, G. W. Gibbons and D. M. Meier, New. J. Phys. **17**, 033048 (2015).
D. C. Brody and D. M. Meier, Phys. Rev. Lett. **114**, 100502 (2015).
X. Wang, M. Allegra, K. Jacobs, S. Lloyd, C. Lupo and M. Mohseni, Phys. Rev. Lett. **114**, 170501 (2015).
F. Albertini and D. D’Allesandro, J. Math. Phys. **56**, 012106 (2015).
R. Romano, Phys. Rev. A **90**, 062302 (2014).
F. Albertini and D. D’Allesandro, arXiv:1510.07091 \[quant-ph\].
E. Assémat, M. Lapert, Y. Zhang, M. Braun, S. J. Glaser, and D. Sugny, Phys. Rev. A **82**, 013415 (2010).
D. L. Elliot, *Bilinear Control Systems*, Applied Mathematical Science, Vol. **169**, Springer (2000).
C. Altafini, J. Math. Phys. **43**, 2051 (2002).
B. Russel and S. Stepney, arXiv: 1661.09220 \[quant-ph\].
L. B. Levitin and T. Toffoli, Phys. Rev. Lett. **103**, 160502 (2009).
P. De Fouquieres and Sophie G. Schirmer, Infinite Dimensional Analysis, Quantum Probability and Related Topics 16.03 (2013).
W. K. Wootters and B. D. Fields, Ann. Phys. (N.Y.) **191**, 363 (1989).
A. G. Kofman and G. Kurizki, Phys. Rev. Lett. **87**, 270405 (2001).
L. Viola, E. Knill and S. Lloyd, Phys. Rev. Lett. **82**, 2417 (1999).
C. Arenz, R. Hillier, M. Fraas and D. Burgarth, Phys. Rev. A **92**, 022102 (2015).
J. Du, X. Rong, N. Zhao, Y. Wang, J. Yang and R. B. Liu, Nature **461**, 1265-1268 (2009).
R. Schmidt, A. Negretti, J. Ankerhold, T. Calarco, and J. T. Stockburger, Phys. Rev. Lett. **107**, 130404 (2011).
D. M. Reich, N. Katz, and C. P. Koch, Sci. Rep. **5**, 12430 (2015).
C. Arenz, D. Burgarth, P. Facchi, V. Giovannetti, H. Nakazato, S. Pacazio and K. Yuasa, Phys. Rev. A **93**, 062308 (2016).
J. Johansson, P. Nation, and F. Nori, Computer Physics Communications **183**, 1760 (2012).
R. Johansson, P. Nation, A. Pitchford, C. Granade, A. L. Grimsmo, markusbaden, A. Vardhan, P. Migda, kafis- cher, D. Vasilyev, et al., qutip/qutip: Qutip-4.0.0 (2016), URL https://doi.org/10.5281/zenodo.220867.
S. G. Schirmer, I. C. Pullen and P. J. Pemberton-Ross, Phys. Rev. A **78**, 062339 (2008).
C. Hillermeier, *Nonlinear multiobjective optimization: a generalize homotopy approach*, Birkhauser, Basel (2001).
R. B. Statnikov and J. B. Matusov, *Multicriteria Optimization and Engineering*, Springer-Verlag, New York (1990).
R. Chakrabarti, R. Wu and H. Rabitz, Phys. Rev. A , 033414 (2008).
R. Chakrabarti and H. Rabitz, International Reviews in Physical Chemistry **26**, 671-735 (2007).
B. Russell, R. Wu, and H. Rabitz, arXiv:1608.06198 \[quant-ph\].
S. E. Economou and E. Barnes, Phys. Rev. A **91**, 161405 (2015).
E. Barnes, C. Arenz, A. Pitchford, and S. E. Economou, arXiv: 1612.09384 \[quant-ph\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We derive an exact analytic solution to a Klein-Gordon equation for a step potential barrier with cutoff plane wave initial conditions, in order to explore wave evolution in a classical forbidden region. We find that the relativistic solution rapidly evanesces within a depth $2x_p$ inside the potential, where $x_p$ is the penetration length of the stationary solution. Beyond the characteristic distance $2x_p$, a Sommerfeld-type precursor travels along the potential at the speed of light, $c$. However, no spatial propagation of a main wavefront along the structure is observed. We also find a non-causal time evolution of the wavefront peak. The effect is only an apparent violation of Einstein causality.'
address: |
Facultad de Ciencias, Universidad Autónoma de Baja California\
Apartado Postal 1880, Ensenada, Baja California, México.
author:
- Jorge Villavicencio
title: Exact relativistic time evolution for a step potential barrier
---
\#1\#2
Introduction
============
Since the early beginnings of quantum mechanics, the problem of particle propagation in classical forbidden regions has been the subject of both theoretical and experimental investigations. Over the years, several non-relativistic approaches based in cutoff wave initial conditions[* *]{}have been introduced in the literature in order to investigate the time-dependent features of wave evolution in evanescent media. Some of these theoretical models were inspired in the pioneering work of Sommerfeld and Brillouin [@Somm; @Leoa], while others [@muga96; @PRA97; @PRA99; @PRBRC] were based in the seminal work of Moshinsky [@mm; @mm1], who a few decades ago started a fundamental discussion on the non-relativistic and relativistic transient effects. These models represent important steps towards the clarification of the dynamics in classical forbidden regions, and a renewed motivation to explore this problem has been recently stimulated by the issue of superluminal velocities in photon [@chiao; @chiaob] and microwave [@enders; @nimtz] tunneling. Hence, it is clear that a full relativistic approach to describe the wave evolution in evanescent media is needed. Nevertheless, this has become a complex problem due to the lack of exact analytical solutions to relativistic wave equations with appropriate initial conditions. Among the few works in the field , Deutch and Low [@Deutch] have provided a lucid description of barrier penetration of an initial state given by a cutoff Gaussian wavepacket, based on a one-dimensional Klein-Gordon equation. Although no exact relativistic solutions were obtained, the issues of Einstein causality and superluminal phenomena were rigorously discussed using approximate solutions.
In this paper we consider a model based on the Klein-Gordon equation, as in the work of Deutch and Low [@Deutch], for a potential step barrier and cutoff plane wave initial conditions. We obtain an exact analytic solution to the problem along the potential region and study the main features of wave evolution, in particular the regime of transient effects at early times.
The paper is organized as follows. In section 2 we discuss the shutter problem, and present the analytical derivation for the solution to a relativistic wave equation for a step potential barrier. Section 3 deals with a numerical example for the solution along the internal region of the potential, and the results are discussed in section 4. Finally, in section 5 we present the summary and conclusions.
The relativistic shutter problem
================================
To investigate the time evolution of cutoff plane wave in a classical forbidden region, let us consider a classical field $\psi _r^s$ satisfying a one-dimensional Klein-Gordon equation with a variable potential $V(x)$, as in the model of Deutch and Low [@Deutch],
$$\frac{\partial ^2}{\partial x^2}\psi _r^s(x,k_r,t)=\frac 1{c^2}\frac{%
\partial ^2}{\partial t^2}\psi _r^s(x,k_r,t)+V_0(x)\psi _r^s(x,k_r,t).
\label{step}$$
In our case, $V_0(x)$ is given by a step potential barrier,
$$V_0(x)=\left\{
\begin{array}{cc}
\mu _0^2, & x\geq 0, \\
0, & x<0,
\end{array}
\right. \label{pot}$$
where $\mu _0=(m_0c/\hbar )$, and the initial condition at $t=0$ corresponds to a plane wave shutter [@PRA99] (see figure \[fshutter\]), given by,
$$\psi _r(x,t=0)=\left\{
\begin{array}{cc}
e^{ik_rx}-e^{-ik_rx}, & x\leq 0, \\
0, & x>0.
\end{array}
\right. \label{conref}$$
The simplicity of our quasi-monochromatic initial state (\[conref\]), allows a closed analytical solution of the problem. It differs from that of reference [@Deutch], where a cutoff Gaussian wavepacket initial condition was considered. Note that condition (\[conref\]) comes from the fact that for $t<0$, the solution for the left side of the shutter [@note1] is given by $\psi _r(x,k_r,t)={\rm exp}[ik_r(x-ct)]-{\rm exp}%
[-ik_r(x+ct)]$, for $x<0$, and zero for $x>0$.
To obtain the solution for $x>0$ and $t>0$, we shall proceed along the same lines as in our recent work [@PRA99]. We begin by Laplace transforming the equation (\[step\]) using the standard definition
$$\overline{\psi }(x,k_r,s)=\int\nolimits_0^\infty \psi (x,k_r,t)e^{-st}dt,
\label{laplace}$$
with the initial condition given by equation (\[conref\]). As a consequence, one gets a pair of differential equations corresponding to the regions $x>0$ and $x<0$. In order to obtain the transmitted wave function, one must consider the matching conditions for the wave function and its derivative at $x=0$. The Laplace transformed solution reads,
$$\overline{\psi }_r^s(x,s)=\frac{2E}{i\left( s+iE\right) \left( s+p\right) }%
e^{-px/c}, \label{steplap}$$
where $p=(s^2+\mu _0^2c^2)^{1/2},$ and $E=k_r=(E_r/\hbar c)$ corresponds to the relativistic energy $E_r$ given in reciprocal units of length.
The time dependent solution for $x>0$ is readily obtained by performing the inverse Laplace transform of equation (\[steplap\]) using the Bromwich integral formula,
$$\psi _r^s(x,t)=\frac 1{2\pi i}\int\limits_{\gamma ^{\prime }-i\infty
}^{\gamma ^{\prime }+i\infty }\overline{\psi }_r^s(x,s)e^{st}ds,
\label{bromwich1}$$
where the integration path is taken along a straight line $s=\gamma ^{\prime
}$ parallel to the imaginary axis in the complex $s-$plane. The real parameter $\gamma ^{\prime }$ can be chosen arbitrarily as long as all singularities remain to the left-hand side of $s=\gamma ^{\prime }$.
The integral (\[bromwich1\]) expressed in this form, is difficult to manipulate since the integrand (\[steplap\]) has branch points at $s=\pm
i\mu _0c$. To surmount this difficulty, let us introduce the change of variable, $-iu=(s+p)/\mu _0c$, which allows to eliminate the branch points. Thus, $p=i\mu _0c(u^{-1}-u)/2$, and as a consequence, the Bromwich integral may be written as,
$$\psi (x,t)=\frac 1{2\pi i}\int\limits_{i\gamma -\infty }^{i\gamma +\infty
}F(u)du, \label{brom2}$$
where the new integrand $F(u)$ is given by,
$$\begin{aligned}
F(u) &=&\frac{2E}{\mu _0}\frac{(1-u^2)}{u^2(u^2-2Eu/\mu _0+1)} \nonumber \\
&&\times \text{exp}\{i\mu _0[u(x-ct)-u^{-1}(x+ct)]/2\}. \label{integrand}\end{aligned}$$
Note that the branch points go into an essential singularity at $u=0$ and two simple poles $u_{\pm }=(E\pm iq)/\mu _0$, where we defined $q=(\mu
_0{}^2-E^2)^{1/2}$. The integration in equation (\[brom2\]) is performed along a straight line $L$ parallel to the real axis cutting the positive imaginary axis at $i\gamma $. We proceed to evaluate the above integral by considering a closed Bromwich integration contour (see figure \[fig2\]), and Cauchy’s residue theorem. For $x>ct$ we close the integration path $L$ from above, by a large semicircle $\Gamma _1$ of radius $R$, forming a closed contour $C_1$. The contribution along $\Gamma _2$ vanishes as $%
R\rightarrow \infty $, and since there are no poles enclosed inside $C_1,$ $%
\psi (x,t)=0$ for $x>ct$. For the case $x<ct,$ we close the integration path from below with a large semicircle $\Gamma _2$. The closed contour $C_2$ contains three small circles $C_{0,}$ $C_{+}$ and $C_{-}$ enclosing the essential singularity at $u=0$ and the simple poles at $u_{+}$ and $u_{-},$ respectively. Hence by using Cauchy’s theorem, it follows that,
$$\frac 1{2\pi i}\left[ \int\limits_{i\gamma -\infty }^{i\gamma +\infty
}-\int\limits_{\Gamma
_2}+\int\limits_{C_0}+\int\limits_{C_{+}}+\int\limits_{C_{-}}\right]
F(u)du=0. \label{Cauchy}$$
The integrals corresponding to the contours $C_{+}$ $C_{-}$ can be easily evaluated, and yield the exponential contributions to (\[Cauchy\]), namely,
$$\frac 1{2\pi i}\int\limits_{C_{\pm }}F(u)du=k_{\pm }e^{(\mp qx-iEct)},
\label{expon}$$
where we defined $k_{\pm }=2E/(E\pm iq)$.
The contour integration for $C_0$ requires a more elaborate calculation, since involves an essential singularity at $u=0$. For this case, we introduce the change of variable given by $\omega =-iu\xi ^{-1}$, thus the integral now becomes
$$\int\limits_{C_0}F(u)du=\int\limits_{C_0^{^{\prime }}}\frac{2E}{i\mu _0\xi ^3%
}\frac{(1+\omega ^2\xi ^2)exp[\eta (\omega -\omega ^{-1})/2]}{\omega
^2(\omega -\omega _{+})(\omega -\omega _{-})}d\omega , \label{C0}$$
where $\omega _{\pm }=(E\pm iq)/i\mu _0\xi $. To carry out the integration, first let us separate the integrand into partial fractions, and substitute the well known formula for the Bessel generating function,
$$e^{\eta (\omega -\omega ^{-1})/2}=\sum_{n=0}^\infty \omega ^nJ_n(\eta
)+\sum_{n=1}^\infty (-1)^n\omega ^{-n}J_n(\eta ) \label{relation}$$
and the series expansion,
$$(\omega _{\pm }-\omega )^{-1}=(\omega _{\pm })^{-1}\sum_{n=0}^\infty (\omega
/\omega _{\pm })^nJ_n(\eta ). \label{relation2}$$
The resulting integrals can be evaluated by means of the residue theorem. For the case of an essential singularity, the residue may be determined by computing explicitly the coefficient corresponding to $\omega ^{-1}$ from the series expansion and their products. In that case, equation (\[C0\]) becomes,
$$\begin{aligned}
\frac 1{2\pi i}\int\limits_{C_0}F(u)du &=&\left[ \frac{2iE}{\mu _0\xi }%
J_1(\eta )-k_{+}\sum_{n=0}^\infty (-1)^n\frac{J_n(\eta )}{(\omega _{+})^n}%
\right. \nonumber \\
&&\left. -k_{-}\sum_{n=0}^\infty (-1)^n\frac{J_n(\eta )}{(\omega _{-})^n}%
\right] \text{.} \label{C0_f}\end{aligned}$$
Finally, substituting the results given by equations (\[C0\_f\]) and (\[expon\]) into equation (\[Cauchy\]), the solution for the internal region is,
$$\psi _r^s(x,t)=\left\{
\begin{array}{ll}
\psi _{+}(q)+\psi _{-}(q), & t>x/c \\
0, & t<x/c,
\end{array}
\right. \label{stepfinal}$$
with $\psi _{\pm }(q)$ defined as,
$$\begin{aligned}
\psi _{\pm }(q) &=&k_{\pm }\left[ e^{(\mp qx-iEct)}+\frac{iz_{\pm }}{2\xi }%
J_1(\eta )\right. \nonumber \\
&&\left. -\sum\limits_{n=0}^\infty (\xi /iz_{\pm })^nJ_n(\eta )\right] .
\label{simplif}\end{aligned}$$
In the above expression, $J_n(\eta )$ stands for the Bessel function of order $n$. The other parameters are defined as: $\xi =\left[
(ct+x)/(ct-x)\right] ^{1/2},$ $\eta =\mu _0(c^2t^2-x^2)^{1/2}$ and $z_{\pm
}=(E\pm iq)/\mu _0$. From equation (\[stepfinal\]) we see that the solution obeys Einstein causality, i.e. no propagation faster than the speed of light, $c$, is detected along the barrier region. In other words, an observer located at an arbitrary position $x_0$ inside the barrier must wait a time $t=(x_0/c)$ before detecting the arrival of the signal.
For the sake of completeness, let us now consider the asymptotic behavior of $\psi _r^s(x,t)$ for the cases $\mu _0\rightarrow 0$, $t\rightarrow \infty $ and $x\rightarrow ct$. From the solution we have just discussed, one may recover the free propagation solution in the limit $\mu _0\rightarrow 0$. This corresponds to let the variables $\eta \rightarrow 0,$ $q\rightarrow iE$. To illustrate the limit process in equation ( \[stepfinal\]), let us rewrite the solution by using equation (\[relation\]), namely, $$\begin{aligned}
\psi _r^s(x,t) &=&k_{-}\left[ e^{(qx-iEct)}-J_0(\eta )\right. \nonumber \\
&&\left. -\sum\limits_{n=1}^\infty (\xi /iz_{-})^nJ_n(\eta )\right]
\nonumber \\
&&+k_{+}\left[ \sum\limits_{n=2}^\infty (z_{+}/i\xi )^nJ_n(\eta )\right] .
\label{as1}\end{aligned}$$
As $\mu _0$ $\rightarrow 0$, the variable $J_0(\eta )\rightarrow 1$, and since $(z_{-})^{-1}\rightarrow 0$ the first series on the right hand-side clearly vanishes. It can be shown that the second series also vanishes, by replacing the Bessel functions by their asymptotic values for small values of the argument $\eta $,
$$J_n(\eta )\simeq 2^{-n}\eta ^n/n!. \label{bessmall}$$
Therefore, one obtains the solution for the free propagation case,
$$\psi _r^s(x,t)\rightarrow \left\{
\begin{array}{cc}
e^{ik_r(x-ct)}-1, & t>x/c, \\
0, & t<x/c.
\end{array}
\right. \label{solibre}$$
Note that the free case solution rises from zero only after a time $t=(x/c)$ fulfilling relativistic causality, and then oscillates periodically thereafter.
In the case of the long-time limit $(t\rightarrow \infty )$, we have $\xi
\rightarrow 1$ and $\eta \rightarrow \infty $. From the asymptotic expansion of $J_n(\eta )$ for large values of the argument $\eta $,
$$J_n(\eta )\simeq \frac 1{(\pi \eta /2)^{1/2}}cos[\eta -\frac 14(2n+1)\pi ],
\label{aslong}$$
and therefore $J_n(\eta )\rightarrow 0$. One can see from equation (\[simplif\]) that the series in $\psi _{+}(q)$ vanishes. On the other hand, if we rewrite $\psi _{-}(q)$ by means of equation (\[relation\]), the exponential term is canceled and the remaining series vanishes. Consequently, $\psi _r^s(x,t)\ $goes into the stationary solution $\phi
_r^s(x,t)$ given by,
$$\phi _r^s(x,t)=k_{+}e^{-qx}e^{-iEct}. \label{stationary}$$
The asymptotic behavior near the relativistic cutoff, is obtained when $%
x\rightarrow ct$ in equation (\[stepfinal\]). In this case we have $\eta
\rightarrow 0$, which allows us to substitute the asymptotic expansion (\[bessmall\]) in $\psi _{+}(q)$ (equation (\[simplif\])). Thus, the series of equation (\[simplif\]) goes into an exponential function, which cancels exactly with the exponential term, and as a result the solution $\psi _{+}(q)
$ goes like $iEJ_1(\eta )/\mu _0\xi $. From similar considerations on $\psi
_{-}(q)$, an identical expression is obtained and the approximate behavior of $\psi _r^s(x,t)$ near the cutoff is given by,
$$\psi _r^s(x,t)\approx \frac{2iE}{\mu _0\xi }J_1(\eta ), \label{bessel2}$$
where for exactly the value $x=ct$, the above expression goes to zero fulfilling relativistic causality.
Examples
========
In order to exemplify the evolution of the solution given by equation (\[stepfinal\]) along the evanescent region, we decided to study the properties of $|\psi _r^s(x,t)|^2$ as a function of time $t$ and the position $x$. The parameters for the system considered in all the cases for the present study are: barrier height $\mu _0=1.542$ $nm^{-1}$, incidence energy $E_r=10.0$ $%
eV $ ( $E=5.064\times 10^{-2}$ $nm^{-1}$).
The first case corresponds to the spatial evolution of $|\psi _r^s(x,t)|^2$ along the dispersive region. In figure \[birth\] we show at early times the birth of the main wavefront as a function of the position, for increasing values of time: $t_1=.001$ $fs$, $t_2=.0035$ $fs$ and $t_3=.0075$ $fs$. The solution rises as time goes on, and at $t_3$, $|\psi _r^s(x,t)|^2$ has already crossed over the stationary solution $|\phi _r^s(x,t)|^2$ (dashed line). The inset of figure \[birth\] shows the crossover of $|\psi
_r^s(x,t)|^2$ at later time $t_4=.012$ $fs$. This behavior is relevant since it indicates that the relativistic solution fluctuates around $|\phi
_r^s(x,t)|^2$, before reaching its asymptotic regime. At the inset, we can also observe how the solution evanesces within a finite depth given approximately by $2x_p=1.317$ $nm$, where $x_p=(1/q)$ is the [*penetration length*]{} of the stationary solution $|\phi
_r^s(x,t)|=|k_{+}|e^{-qx}$ (equation (\[stationary\])). We find that beyond $2x_p$ the solution exhibits a small maxima, corresponding to the birth of a forerunner. In figure \[earlyt1\] we depict the spatial evolution of $|\psi _r^s(x,t)|^2$ (solid line) for a fixed value of $t=0.05$ $fs$. As can be seen in this example, the main part of the wave rapidly evanesces in the potential region for small values of the position. However, from approximately $2x_p$ onwards, the solution exhibits an oscillatory behavior before reaching the relativistic cutoff at $x=15.0$ $nm$, corresponding to the earliest arrival of the signal at a point located within the potential. The stationary solution $|\phi _r^s(x,t)|^2$ (dashed line) is also included for comparison. It is interesting to note the similarity of the oscillatory structure in figure \[earlyt1\], to the well known Sommerfeld precursor [@Leob; @precexpe], which is one of the essential features of wave propagation in dispersive media. Despite the fact that Sommerfeld’s approach is quite different from ours, the similarities go beyond the numerical results. For instance, their asymptotic analysis showed that the wave function is governed by a first order Bessel function near the relativistic cutoff. Our analysis reproduces such behavior, which is given by equation (\[bessel2\]). For comparison, the value of the Bessel function $J_1(\eta )$ modulated by the prefactor $2iE/\mu _0\xi $, is also included in figure \[earlyt1\] (dotted line). Note that if we define the frequency of the oscillations of the precursor in terms of the distance between successive zeros of $J_1(\eta )$, one sees from the definition of the argument $\eta $ that the value of the frequency depends only on the position $x$ and the potential $\mu _0$ that characterizes the medium i.e. the precursor frequency is independent of the incidence energy.
In figure \[earlyt2\] we show $|\psi _r^s(x,t)|^2$ (solid line) as a function of the position $x$, at a subsequent time $t=0.3$ $fs$. We can see that the solution reaches its stationary value $|\phi _r^s(x,t)|^2$ (dashed line) for small values of $x$; nevertheless, near the relativistic cutoff at $x=90.0$ $nm$, the precursor exhibits a rich oscillatory structure. The inset of figure \[earlyt2\] illustrates the forerunner near the cutoff, and shows that the asymptotic behavior is dictated by the Bessel function of equation (\[bessel2\]) (dotted line).
Up to here we have illustrated the spatial behavior of $|\psi _r^s(x,t)|^2$, and some interesting features of the time evolution. In order to fully explore the relevant features of the time evolution, in figure \[diffraction\] we plot $|\psi _r^s(x,t)|^2$ as a function of time at different positions: $x_1=0.4$ $nm$ , $x_2=0.6$ $nm$ and $x_3=0.8$ $nm$. For all the curves depicted, as soon as $t>(x/c)$ the solution is different from zero along the internal region, fulfilling relativistic causality. As can be seen, the solution rises from zero at $t=(x/c)$ and grows monotonically towards a maximum value, from which it starts to oscillate thereafter, forming a pattern very similar to the diffraction in time phenomenon [@mm].[** **]{}The concept of diffraction in time was originally introduced by Moshinsky [@mm] while discussing the shutter problem for the free particle Schrödinger equation. He observed a time-dependent oscillatory regime of the probability density near the semiclassical wavefront that he named diffraction in time, in analogy to the well known Fresnel optical diffraction. It is interesting to note the resemblance of the oscillatory pattern in figure \[diffraction\], to the diffraction in time phenomenon observed in the free propagation case [@PRA99]. Moreover, in the low-energy regime ($\mu _0/E\gg 1$) the solution (\[stepfinal\]) can be rewritten in a more concise form by using equation (\[relation\]), namely, $$\Psi _r^s(x,t)\approx 2(E/V)\left[ U_3(i\eta /\xi ,\eta )-U_1(i\eta
/\xi,\eta )\right], \label{flommel}$$ where $U_1$ and $U_3$ are the Lommel functions of two variables [@lommel2v], widely used in connection with optical diffraction [@bornwolf]. The resemblance to diffraction phenomena suggests that there exists a more profound link; however, the physical implications of the striking mathematical similarities found above deserves further study.
It is important to mention that the transient effect depicted in figure \[diffraction\] is observed in the low-energy regime i.e. $(\mu _0/E)\gg 1$; this condition is satisfied in the present example, where the effect was observed for values of the ratio[** **]{}$(\mu _0/E)\simeq 30$[**.**]{} Moreover, we only observed the phenomenon in the regime of small values of the position $x$ where the solution decays in the potential region i.e. $%
x<2x_p$. From values greater than $x\simeq 2x_p$ the solution enters into a different oscillatory regime, and the diffraction-type pattern begins to disappear. In figure \[nodiff\] we illustrate the inhibition of the diffraction-type pattern for a fixed value of the position $x=3.0$ $nm$. Clearly, it fades out and is replaced by a series of oscillations, which register the fast crossing of the precursor at $x=3.0$ $nm$, and the remaining forerunners.
There is another interesting feature in the time evolution of $|\psi
_r^s(x,t)|^2$ that can be appreciated in figure \[peakshift\], in which we plot $|\psi _r^s(x,t)|^2$ as a function of time in the main peak region. Surprisingly, the maximum peak of the wave appears on $x=0.5$ nm (dotted line) earlier than the peak at $x=0.3$ nm (dashed line) and $x=0.1$ nm(solid line). This relative [*time shift*]{} of the wave peak is an apparent violation of relativistic causality, and can be interpreted as a non-causal behavior. This comes from the fact that we are comparing the maximum wave peak at different positions. However, we observe that the wavefront always fulfills Einstein causality, and no signal travels faster than $c$ in the dispersive region. Therefore, the observed shift of the main peak may be interpreted as a reshaping of the wave and not as a genuine violation of relativity.
It is interesting to mention that we have observed a similar non-causal behavior in the probability density along a classical forbidden region of a rectangular potential barrier, within a non-relativistic framework. Moreover, some authors have also reported non-causal phenomena in electromagnetic evanescent modes [@nimtz].
Discussion
==========
The possibility of describing the wave evolution from the transient to the stationary regime, offers a clear advantage over the asymptotic methods of solution available in the literature, for which the short and intermediate transient regimes are inaccessible. In what follows, we shall discuss the new features in the dynamical process of evanescent waves observed in the previous section.
The buildup of $|\psi _r^s|^2$ exhibits a very interesting behavior; the solution instead of just grow monotonically towards the stationary solution $%
|\phi _r^s|^2$, fluctuates around such value before reaching the asymptotic regime, as it is shown by the series of curves of $|\psi _r^s|^2$ versus $x$ at different times (see figure \[birth\]). The effect of these fluctuations in a plot of $|\psi _r^s|^2$ versus $t$ ($x$ fixed) is manifested as a series of oscillations similar to a diffraction in time pattern, see figure \[diffraction\]. The inset of figure \[birth\] shows that beyond a certain distance, an interesting structure of the wave appears; this is the birth of the Sommerfeld-type forerunner which travels along the potential region, as illustrated in figures \[earlyt1\] and \[earlyt2\]; the head of this signal propagates at the speed of light and can reproduced by the first order Bessel function $J_1(\eta )$ (dotted line). The birth of the forerunner is an important event since its propagation at longer times becomes the dominant process; this is also the case in the context of different relativistic and non-relativistic approaches , where the characterization of the forerunners has been recently emphasized.
At early times and for small values of the position, the main front of the wave decays exponentially along the potential. As time goes on, the dynamics is dominated by the propagation of the forerunners since the main front rapidly reaches its asymptotic value without propagation. Thus, one may speak of two regimes, which as discussed in the previous section, are characterized by $2x_p$ where $x_p$ is the penetration length. If we choose a position $x>2x_p$ and wait for the main wavefront, instead of detecting its arrival we would only register the fast crossing of the precursor (see figure \[nodiff\]). The absence of main wavefront propagation in the evanescent region is in agreement with a series of works [@muga96; @Ranfagni; @Jauhjo; @Teranishi], which have questioned the existence of semiclassical wavefront propagation proposed by Stevens [@stevens83] and supported by Moretti [@morettipra92; @morettipra92a].
Another important result of this work is the non-causal peakshift exhibited in figure \[peakshift\]. The non-causal behavior observed here is a consequence of the reshaping of the wave; reshaping effects have also been observed in the context of wavepacket evolution within both relativistic [@Deutch] and non-relativistic [@Krenslin] approaches. The role of this effects and the issue of non-causal behavior has been recently discussed by Deutch and Low [@Deutch] for the case of Gaussian wavepacket evolution in the transmitted region of a potential barrier, based on approximate solutions to the Klein-Gordon equation. Although the barrier and the step potential are different systems, both exhibit an evanescent region; hence, we believe that the non-causal behavior observed in reference [@Deutch] could be related to a reshaping process occurring inside the barrier, similar to the reshaping observed in the step discussed in our model. However, in order to investigate such a reshaping inside the barrier, the solution of the Klein-Gordon equation is required for the internal region. In this respect, the analytical techniques used in this work may provide a suitable method of solution to tackle this fundamental problem; nevertheless, this not an easy task since the extension of our model to the case of a barrier of finite width involves more complicated analytical properties of the solution due to the presence of resonances.
Summary and conclusions
=======================
We have derived an exact analytical solution to a Klein-Gordon equation for a step potential barrier, using a cutoff plane wave initial condition. To our knowledge, this is the first model which allows a closed solution for the description of relativistic transient effects in a classical forbidden region.
The main features of the spatial and time evolution along the evanescent region can be summarized in the following points: (i) We found a regime where the solution is exponentially suppressed and thus, decays as a function of $x$ along the potential. This main part of the wave does not propagate along the structure. The regime is characterized by a region extending from $x=0$ to $x\approx 2x_p$, where $x_p=(1/q)$ is the penetration length of the stationary solution (\[stationary\]). However, from $x\approx 2x_p$ onwards, the solution exhibits an oscillating pattern near the relativistic cutoff, traveling at the speed of light, $c$ which can be identified as a Sommerfeld-type precursor. Also, within the finite depth $%
2x_p$, we found in the low-energy situation that the time evolution of $%
|\psi _r^s(x,t)|^2$, exhibits a transient effect similar to the diffraction in time phenomenon [@mm]. (ii) We showed that along the internal region, there exists a time shift associated to the main peak of the wave function, that can be interpreted as a non-causal behavior along the classical forbidden region. This of course is only an apparent violation of relativistic causality, since in our model the wavefront satisfy always Einstein causality i. e. no signal travels faster than the speed of light.
The relevance of these results, comes from the fact that our findings may be of interest to elucidate on the problem of wave propagation in finite width potentials.
I would like to thank Gastón Garcia-Calderón and Alberto Rubio, who suggested the shutter model approach to solve this problem. I am indebted to them for encouragement and valuable discussions. I would like to thank also Marcos Moshinsky for fruitful discussions on the shutter problem. Useful discussions with Roberto Romo are gratefully acknowledged. This work was supported financially by Conacyt-México under Contract No. 431100-5-32082E.
References {#references .unnumbered}
==========
[99]{} Stevens K W H [*1983*]{} [*J. Phys.* ]{}C[*: Solid State Phys.*]{} [**16**]{} 3649.
Moretti P 1992 [*Phys. Rev.*]{} A [**46**]{} 1233.
1992 [*Physica Scripta*]{} [**45**]{} 18.
Büttiker M and Thomas H 1998 [*Superl. Microstruc.*]{} [**23**]{} 781.
Muga J G and Büttiker M 2000 [*Phys. Rev.* ]{}A[* [**62**]{} 023808*]{}.
Sommerfeld A 1914 [*Ann. Phys.*]{} [**44**]{} 177.
Brillouin L 1914 [*Ann. Phys.*]{} [**44**]{} 203.
1960 [*Wave Propagation and Group Velocity*]{} (Academic Press, New York) p. 74.
Brouard S and Muga J G 1996 [*Phys. Rev.*]{} A [**54**]{} 3055.
Garcia-Calderón G and Rubio A 1997 [*Phys. Rev.*]{} A [**55**]{} 3361.
Garcia-Calderón G, Rubio A and Villavicencio J 1999 [*Phys. Rev.*]{} A [**59**]{} 1758.
Romo R and Villavicencio J 1999 [*Phys. Rev.* ]{}B [**60**]{} R2142.
Moshinsky M 1952 [*Phys. Rev.*]{} [**88**]{}, 625.
1976 [*Am. J. Phys.*]{} [**44**]{} (11) 1037.
Steinberg A M, Kwait P G and Chiao R Y 1993 708.
Chiao R Y 1993 [*Phys. Rev.*]{} A [**48**]{} R34.
Enders A and Nimtz G 1993 [*J. Phys. I France*]{} [**3**]{} 1089.
Nimtz G 1998 [*Ann. Phys. (Leipzig)*]{} [**7**]{} 618.
Deutch J M and Low F E 1993 [*Ann. Phys.*]{}, NY. [**228**]{} 184.
Note that the shutter is a device that aids to visualize the initial condition and hence it is not part of the system.
The experimental observation of Sommerfeld and Brillouin precursors has already been reported in the microwave domain \[Pleshko P and Palócz I 1969 [*Phys. Rev. Lett.*]{} [**22**]{} 1201\].
Krenzlin H M, Budczies J and Kehr K W 1998 [*Ann. Phys. (Leipzig)*]{} [**7**]{} 732.
E. N. Dekanosidze,[** **]{}[*Tables of Lommel’s Functions of Two Variables*]{} (Pergamon Press, New York, 1960) p. 1.
M. Born and E. Wolf, [*Principles of Optics*]{} (Pergamon Press, Oxford, 1965) p. 438.
Ranfagni A, Mugnai[** **]{}D and Agresti A 1991 [*Phys. Lett. A*]{} [**158** ]{}161.
Jauho A P and Jonson[** **]{}M [*Superlattices and Microstructures*]{} 1989 [**6** ]{}303.
Teranishi N, Kriman A M and Ferry D K [*Superlattices and Microstructures*]{} 1987 [**3** ]{}509.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We review our current knowledge about a particular case of decoupled gas kinematics – inner ionized-gas polar disks. Though more difficult to be noticed, they seem to be more numerous than their large-scale counterparts; our recent estimates imply about 10% of early-type disk galaxies to be hosts of inner polar disks. Since in the most cases the kinematics of the inner polar gaseous disks is decoupled from the kinematics of the outer large-scale gaseous disks and since they nested around very old stellar nuclei, we speculate that the inner polar disks may be relic of very early events of external gas accretion several Gyr ago. Such view is in agreement with our new paradigm of disk galaxies evolution.'
author:
- 'O. K. Sil’chenko$^1$ and A. V. Moiseev$^2$'
bibliography:
- '001osilchenko.bib'
title: 'Inner polar gaseous disks: incidence, ages, possible origin'
---
Introduction
============
Among gas subsystems with decoupled kinematics, a particular interest is inspired by polar rings/disks. Firstly, they are beautiful, secondly, they seem to be stable over many dynamic times, and thirdly, they imply certainly accretion of external gas from highly inclined orbits. **Inner** polar gaseous disks are less spectacular than large-scale polar rings; however they may be even more numerous though difficult to be detected against the bright bulge background in early-type disk galaxies.
We note that the first evidence of existence of circumnuclear gas on polar orbits in the literature was presented by @Rubin1977 in their interpretation of the large line-of-sight velocity gradient along minor axis in NGC 3672. Further @bettoni90 have claimed inner polar gaseous disk in the southern ringed lenticular galaxy NGC 2217. By studying it through long-slit spectroscopy, @bettoni90 found visible gas counter-rotation in some slit orientations (not all). Their geometrical scheme for the center of NGC 2217 demonstrated clearly that the ionized-gas disk had to be warped in such a way that in the very center it occupied the polar plane orthogonal to the bar major axis. Later we found inner polar disks in unbarred early-type spiral galaxies NGC 2841 [@n2841pr] and NGC 7217 [@n7217pr] by obtaining two-dimensional velocity fields for the ionized gas and for the stellar component with the integral-field unit of the 6-m telescope, Multi-Pupil Fiber Spectrograph (MPFS). The outer neutral hydrogen in both spirals is confined to their main symmetry planes and rotates normally. It was a puzzle how a small amount of polar-orbiting gas could reach the circumnuclear regions without colliding with the main gaseous disks.
Now a few dozens of inner polar gaseous disks/rings are known. Their samples were presented earlier by @corsini03 and @moiseev_malta; the latest statistics based on the data for 47 inner polar disks collected over literature is published by @mois_rev, and here we review briefly some incidence properties.
Incidence
=========
When we analyze all the cases with the inner gaseous disks inclined to the galactic symmetry planes by more than 45, we find that the inclinations of such disks tend strongly to the strictly polar orientation: about two thirds of all such disks are inclined by $\ga 80$. This is consistent with theoretical claims about stability of the strictly polar orientations and instability of the disks inclined by intermediate angles; the latters would precess until they occupy the polar or co-planar orientation.
The inner polar disks – as well as the large-scale ones – prefer to inhabit early-type galaxies. However while large-scale polar rings are seen mostly around gas-poor E/S0 galaxies – about a factor of 3 more often than around spirals, – and it can be explained by them devoiding the hosts with large-scale coplanar gaseous disks [@resh11], the inner polar disks are found in Sa–Sc spiral galaxies in one third of all cases, and large-scale coplanar gaseous disks do not prevent their appearance (see the above mentioned examples of NGC 2841 and NGC 7217); even a few cases are known to be found in very late-type dwarfs. The typical size (radius) of an inner polar disk is 0.2–2 kpc; the lower limit is perhaps defined by our restricted spatial resolution. If to consider inner polar disks together with the large-scale relatives, a continuous sequence in their sizes normalized by a galaxy diameter is observed with a gap at the size $\sim0.5\,D_{25}$. This bimodal distribution can be explained by different agents of stability for polar structures: while the external structures are stabilized by the spheroidal (or even triaxial) potential of halo, the inner disks are usually settled well within the bulge-dominated area [@SmirnovaMoiseev2013]. In any case, the presence of embedding stabilizing potential is important. Is it crucial that this three-dimensional potential has to be also triaxial as in NGC 2217 [@bettoni90]? @mois_rev presents the following statistics: among 40 galaxies with the inner polar disks which have the morphological type S0 and later there are 17 galaxies with bars or triaxial bulges. This gives us the fraction of barred galaxies among galaxies with the inner polar disks, only 43%$\pm$8%, completely consistent with the fraction of barred and/or triaxial-bulge galaxies among all disk galaxies, 45% [@aguerri09].
The list of all known till 2012 inner polar disks by @mois_rev cannot be used to estimate how often the phenomenon can be met: the sample of the hosts of the inner polar disks listed there is quite inhomogeneous. To estimate the inner polar disk incidence, we have used the data of the recent integral-field spectroscopic survey ATLAS-3D [@atlas3d_1]. The ATLAS-3D sample is volume-limited one and includes 60 elliptical galaxies and 200 lenticular galaxies (if we classify NGC 2768 as S0). We have taken the raw science and calibration frames from the open Isaac Newton Group Archive of the Cambridge Astronomical Data Center and have calculated the stellar and ionized-gas line-of-sight velocity fields. Then the orientation of the rotation planes for both components in every galaxy was determined by fitting a circular-rotation model, and the angles between the rotation planes of the stellar and gaseous components were calculated by using the formula (1) from @mois_rev. Among 200 S0 galaxies of the ATLAS-3D volume-limited sample, we have found 8 new inner polar gaseous disks with the inclination to the stellar rotation planes by more than 50(taken into account both solutions of the equation (1) of @mois_rev, because we don’t know which side of the is nearest to the observer); 12 inner polar disks in the S0 galaxies of the ATLAS-3D sample have been already listed in @mois_rev. Having in total 20 inner polar disks in S0 galaxies of the ATLAS-3D volume-limited sample, we conclude that nearby lenticular galaxies have inner polar disks in 10% of all cases. Our estimate refers to the totality of S0 galaxies over [*all*]{} types of environments. This incidence of the inner polar disks in the early-type disk galaxies, 10%, exceeds greatly the frequence of the large-scale polar rings, 0.1–0.4% [@resh11].
Figure \[n2962\] shows a nice example of the newly discovered inner polar disk in the lenticular galaxy NGC 2962 – a member of the ATLAS-3D volume-limited sample. We have observed this galaxy earlier at the Russian 6-m telescope with the integral-field spectrograph MPFS which field of view was $16\arcsec \times 16\arcsec$, and in the very center, inside $R=5\arcsec$, we saw a compact, fastly rotating, nearly edge-on polar gaseous disk. But with the larger field of view of the SAURON, $41\arcsec \times 33\arcsec$, we are now seeing a switch of the gas rotation sense at $R\approx 7\arcsec -10\arcsec$: the galaxy possesses two nested polar gaseous disks counterrotating each other (Fig. \[n2962\]).
{width="80.00000%"}
Origin
======
Is the polar momentum inner or external?
----------------------------------------
This question may seem to sound strange: if a main baryonic component, stars which are formed from the own gas of the galaxy, rotates in the galactic disk symmetry plane, how may the polar gas be of local origin? Meanwhile there are intrinsic secular evolution mechanisms that produce strongly inclined gaseous disks in the very center of a galaxy, and one of them had been revealed by simulations of @friedli_benz. By tracing dynamical evolution of [*initially retrograde*]{} gas in the disk of an isolated barred galaxy, @friedli_benz have found that after about 2 Gyr of angular momentum exchange with the stellar bar the gas inside a few hundred parsec comes to a strongly inclined plane due to vertical instabilities. Since retrograde motions of stars are always present in the barred potential [@pfenniger], and since stars drop gas during their evolution, in principle the inner polar gaseous disks may form in barred galaxies without outer donor contribution. Indeed, we have found several cases when the presence of the inner polar disk in the very center is accompanied by the presence of counterrotating gas in the more outer disk – e.g. in NGC 7280 [@n7280; @sil05]. But the presence of a bar is necessary. However, the statistics in the previous Section does not show prevalence of barred galaxies among the hosts of inner polar disks: less than a half of the hosts of inner polar disks reveal triaxiality of their inner stellar structures. So we are now inclined to the hypothesis of the external gas accretion as the dominant mechanism of inner polar disk formation.
{width="80.00000%"}
How much gas can be in a polar orbit?
-------------------------------------
To identify a source of gas accretion, we must estimate first of all typical amounts of gas populating polar orbits. Here a lot of diversity is observed. In some cases the inner polar ionized-gas disks have their extension into the very outer parts of galaxies when they are observed at the 21cm line of the neutral hydrogen – these are the cases, e.g., of NGC 3414 (with the inner polar disk found by @polardust) or of NGC 7280 or of UGC 9519 mapped in the neutral hydrogen line by @serra_h1. In the prototype of large-scale polar ring galaxies, NGC 2685, the inner ionized gas is also polar [@sil_polar98]. In these cases the total mass of the polar gas can be as large as $10^8 -10^9$ solar masses, and the $M(\mbox{HI})/L_K$ ratios resemble those of spiral galaxies [@serra_h1]. In the volume-limited S0-galaxy sample from ATLAS-3D [@atlas3d_1] about one third of all galaxies with the inner polar ionized-gas disks have polar neutral-hydrogen outer extension. However many galaxies have inner [*polar*]{} ionized-gas component and outer [*coplanar*]{} neutral-hydrogen disk; and they are sometimes also rather gas-rich but their main gaseous components are confined to the galaxy symmetry planes. Among lenticular galaxies, we can mention NGC 2962 where @grossi09 have found $1.1\times 10^9$ M of neutral hydrogen in a disk coplanar to the stellar one but extending much farther from the center. And certainly even more such cases can be found among spiral galaxies with the inner ionized-gas polar disks. An inner ionized-gas polar disk was found in a barred spiral, SB(r)b, galaxy NGC 5850 by @moiseev04; the stellar and gaseous rotations were compared over the $16\arcsec \times 16\arcsec$ field of view of the 6-m telescope IFU MPFS. Now we have calculated larger stellar and gaseous velocity fields by using the archival SAURON data (Fig. \[n5850\]). One can immediately see from Fig. \[n5850\] that the sense of the gas rotation changes at the radius of 7–10 (1.3–1.8 kpc); the more outer ionized gas rotates together with the stars. And the same orientation of the rotation plane is demonstrated by all the $2\times 10^9$ solar masses of neutral hydrogen measured in NGC 5850 by @higdon. The same patterns of stellar and ionized gas circumnuclear kinematics were also presented recently in the paper by @Bremer2013, which is based on VLT observations with the VIMOS IFU. The better spatial resolution [comparing with the early observations by @moiseev04] has allowed to calculate precisely the kinematic orientation parameters in the inner disk velocity field. @Bremer2013 claimed that the angle between the inner and outer disks planes is only $24\deg$, however the equation (1) from @mois_rev gives also the second solution – $54\deg$, that corresponds to the case of strongly inclined inner gaseous disk.
{width="50.00000%"}
NGC 7217
--------
An interesting case of a spiral galaxy with the inner polar ionized-gas disk having the radius of only 350 pc [@n7217pr] is represented by an isolated Sab galaxy NGC 7217; here we show the recent HST image of the central part of the galaxy (Fig. \[n7217\]) where the inner ionized-gas polar disk can be seen ‘by eye’ in the narrow photometric band centered onto the emission lines H$\alpha +$\[NII\]. Its neutral hydrogen disk, $0.7\times 10^9$ M, extending to $R\approx 8$ kpc, is coplanar to the stellar disk and rotates just as the stars; at the outer edge of this disk intense star formation in a ring is observed though the visible gas density is below the gravitational stability threshold [@n7217_h1]. Recently we have studied the origin of the complex structure of NGC 7217 in detail [@last7217], and here we discuss this galaxy as a pure key point revealing possible formation mechanisms of the inner polar disks.
Photometric structure of NGC 7217 can be described as three-tiered: we [@n7217pr] have separated three exponential segments in its surface-brightness radial profile. The innermost segment seen only at $R<10\arcsec$ (0.8 kpc) may be a pseudobulge; then other two segments represent an antitruncated disk. Our deep long-slit spectroscopic observations [@last7217] having allowed to measure stellar rotation and line-of-sight velocity dispersion (close to a vertical velocity dispersion because the galaxy is seen almost face-on) as well as the properties of the stellar populations, have revealed prominent differences in all respects between two exponential parts of the stellar disk. Firstly, the inner part of the disk is substantially thinner than the outer part, and secondly, the mean age of the stellar population in the inner disk is 5 Gyr while the stellar ages in the outer disk, even beyond the starforming ring, is very young, less than 2 Gyr. The galaxy being an early-type spiral without a bar, possesses meantime three rings of current star formation [@verdes95]. Interestingly, the age of the nuclear stellar population, inside the circumnuclear starforming ring, is [*very*]{} old – older than 10 Gyr. Obviously, despite violent processes of gas radial re-distribution and external gas accretion betrayed by the inner polar disk presence, the gas has never reached the very center of NGC 7217 for the last 10 Gyr.
Having in hands the detailed structure of NGC 7217 and evolutionary sequence of building elements of this structure, we have tried to fit observational properties of NGC 7217 with the models provided by on-line service GalMer [@galmer]. We have found that only at least [*two*]{} independent gas-rich minor-merger events can provide a full list of properties: the inner polar disk is formed by an accretion of a gas-rich dwarf from an inclined retrograde orbit, and the outer flaring ringed starforming disk is shaped by merging a prograde-orbiting satellite. The necessity of two minor mergers is due to the fact that minor merging from a retrograde orbit gives an inclined inner gaseous disk but does not thicken the large-scale stellar disk. The latter feature requires minor merging from a prograde orbit. Since the star formation burst in the outer disk of NGC 7217 is very young, we conclude that the minor merging from a retrograde orbit was the first event, and minor merging from a prograde orbit was the last, quite recent one.
Ages
====
The large-scale outer polar rings may be stable in the polar state over a few Gyrs according to theoretical estimations [e.g. @steiman-cameron] as well as to numerical simulations [@Snaith2012]. Stability of their circumnuclear counterparts is still an open question. However some observational evidences in favour of their very long living times also exist: just among lenticular galaxies with the inner polar disks we found [*very*]{} old stellar nuclei, $T>10$ Gyr [@polardust], while over the full sample of nearby lenticular galaxies the typical ages of the stellar nuclei are 2–5 Gyr [@sil06; @sil08].
The whole evolution of disk galaxies is governed by the regime of external gas accretion. Recently we [@sil12] have proposed a scenario according to which all disk galaxies were formed around $z\approx 2$ as lenticular galaxies, and only much later, at $z<1$, most of them started smooth gas accretion and, after having formed thin dynamically cold stellar disks, transformed into spirals. In the frame of this scenario, a natural epoch of forming inner polar gaseous disk is very early stages of the accretion era. If the first accretion event was from a highly inclined orbit, an inner polar long-living gaseous disk would form [*before*]{} the main gas accretion in the galactic symmetry plane proceeded. It is the way to obtain a stable system with mutually orthogonal nested gaseous disks; and then inner polar disks would be relics of very early events of external gas accretion.
We thank Enrica Iodice and the organizers for the interesting and inspiring conference and for the invitation to present this review. A.M. is grateful to the non-profit ‘Dynasty’ Foundation and to the RFBR grant 13-02-00416. This contribution makes use of data obtained from the Isaac Newton Group Archive which is maintained as part of the CASU Astronomical Data Centre at the Institute of Astronomy, Cambridge. The ACS images of NGC 7217 were obtained from the Hubble Legacy Archive, which is a collaboration between the Space Telescope Science Institute (STScI/NASA), the Space Telescope European Coordinating Facility (ST-ECF/ESA) and the Canadian Astronomy Data Centre (CADC/NRC/CSA).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An analysis of free-recall datasets from two independent experiments allows to identify two anomalous instances of non-monotonicity in free recall: a maximum in the dependence of the inter-response intervals on the serial-position lags, and a minimum in the rate of contiguous recall near the beginning of the recall process. Both effects, it is argued, may stem from a hierarchical search protocol in the space of memories. An elementary random-walk model on binary strings is used to test this hypothesis.'
author:
- Francesco Fumarola
title: Hierarchical searching in episodic memory
---
Introduction
============
Free-recall experiments are a key tool for the controlled investigation of episodic memory. A typical free-recall experiment takes place in two stages: during the presentation stage, subjects are shown a list of words; during the memory test, they are requested to recall them in any order (Binet and Henri, 1894; Murray, 1960).
It will be useful to introduce the following terminology: [“serial position”]{}, the numbered position of a word within the list; [“inter-response interval”]{} (IRI), the time interval that elapses between two consecutive recalls; [“transition”]{}, any sequence of two consecutively recalled word; [“lag”]{}, the difference between the serial positions of two words in a given transition. For example, if the $5$th word in the list is recalled right after the $8$th, the corresponding lag is $L= -3$. Transitions with lag $L=\pm 1$ will be called contiguous.
Some of the best-known empirical properties of free recall are:
- power-law scaling: the number of retrieved items scales like a power law of the number of items in the list (Standing, 1973; Murray et al., 1976);
- primacy and recency effects: the first and last words in the list are recalled better than the rest (Murdock, 1962);
- the lag-recency effect: the probability of transitions from one recall to the next is a decreasing function of the lag (Kahana, 1996).
- forward asymmetry: the tendency to recall items in forward order (Ebbinghaus, 1913).
Many other effects have been reported, especially in connection with the semantic and phonological properties of individual words (for a review, see Kahana, 2012).
In this paper, two additional effects are pointed out, concerning the dependence of the IRIs on the lag and the time-dependence of the contiguous recall rate.
I begin by considering, in the next section, two datasets coming from independent experiments and by extracting the variables I have mentioned. The effects emerge clearly from both datasets, despite rather different experimental conditions. I then sketch qualitatively a possible explanation – the presence of a hierarchical component in the search mechanism. I design a simple model that may help test the hypothesis; the model is a variant of the random-walk approaches to free recall recently adopted by various authors. Finally, I compare the findings with the result of simulations on the model and consider possible ramifications of the hypothesis.
Non-monotonicity in Free-Recall Observables
===========================================
The Archival Data
-----------------
The PEERS study (Penn Electrophysiology of Encoding and Retrieval Study) is a study recently conducted at the University of Pennsylania and devoted to assembling a large database on the electrophysiological correlates of memory (Lohnas, 2013).
The sample I have considered corresponds to Experiment I of PEERS. It includes data from trials on $156$ college students (age range: 18$-$30) and on $38$ older adults (age range: 61$-$85 years). In each trial, $16$ words were presented one at a time on a computer screen. Each word was drawn from a pool of $1638$ words with heterogeneous semantic and lexical properties.
Each item was kept on the screen for $3000$ ms, followed by an interstimulus interval of 800$-$1200 ms. After the last item in the list, there was a delay of 1200$-$ 1400 ms, after which the participant was given $75$ s to attempt to recall aloud any of the just-presented items. Multiple trials were performed on each subject, summing up to $3744$ trials for the students’ sample and to $912$ for older adults. For more details on the experimental procedure, see Healey and Kahana (2016).
In order to check my findings against multiple experiments, I have also made use of an older set of data, coming from the free-recall experiments described in Polyn et al., 2009.
These experiments involved 45 participants; the lists contained 24 words each, again presented on a computer screen. Each word was shown for 3 s; in the retrieval stage, participants were given 90 s to recall the words. A total of 1394 trials were performed.
Three differences may be noticed between the experimental conditions of PEERS and those employed by Polyn et al.: (1) the length of the lists, which is $16$ in PEERS and $24$ in Polyn et al.; (2) the longer time allowed by Polyn et al. for both memorization and recall; (3) the different size of the word-pools, as PEERS experiments used a pool of 1638 words, while the pool of Polyn et al. contained only 1297.
Intrusions (i.e. words wrongly recalled from outside the list) have been discarded from all these data; repetitions (less than $0.3\%$ of events) are counted with the lag $L=0$.
Distribution of the IRIs
------------------------
The free-recall literature has shown, as mentioned, that two events are more likely to be recalled at a short distance from each other if the serial-position interval between them is smaller (lag-recency effect). There is therefore a degree of chronological continuity in the way events are stored and recalled, and this fact has been successfully formalised in retrieved-context theories such as the Temporal Context Model of Howard and Kahana, 2002.
Accordingly, we expect the $\textit{time}$ $\textit{interval}$ between two $\textit{consecutive}$ recalls to be also an increasing function of the time interval between the events recalled. Otherwise said, we expect that if two events have occurred at a longer time distance from each other, it will take a longer time to recall them consecutively. In the specific case of free-recall experiments, the inter-response interval (IRI) between two consecutive recalls should then be an increasing function of the absolute value $|L|$ of the lag between the words recalled.
This simple expectation turns out to be contradicted by the data. To see this, let us begin the analysis with a rough approximation, i.e. by regarding all recorded transitions as statistically independent. The resulting histograms of the IRIs for the experiments of Healey and Kahana (2016) are displayed in the upper-left panel of Fig. 1. A different histogram is obtained for each absolute value of the lag; to avoid cluttering, some intermediate lag values are not shown; since the lists in these experiments were $16$ words long, the maximum lag is $15$.
The presence of a peak at small $\tau$ for transitions with $|L|=+1$ is no surprise, and agrees with expectations from the lag-recency effect. For longer lags, the peak is suppressed, as seen in the curves corresponding $|L|=2,3,8$. However, this monotonous suppression has ceased by $|L|=12$, and the peak has begun to grow again. With the maximal lag $|L|=15$, we have recovered a peak as tall as the peak for $|L|=3$. This means that the recall process, slow for transitions to intermediate lags, is faster again for transitions to very long lags.
The upper-right panel shows the corresponding cumulative distribution, visualised for all values of the lag. The growth of the cumulative toward saturation becomes gradually slower as the lag increases from $|L|=+1$. It reaches its slowest point for $|L|^*=10$ (yellow curve); then it gradually grows faster again. The inversion point is the lag $|L|^* \sim 0.6 S$, where $S$ is the size of the lists. For the maximal lag $|L|=15$, the cumulative curve virtually overlaps with the curve for $|L|=3$.
Data from the experiments of Polyn et al. (2009) are similarly displayed in the two lower panels. The lower-left panel shows the distribution of IRIs for a handful of lag values. The behavior of the distribution confirms what we observed in the data of Healey and Kanaha: a suppression of the peak for intermediate lags values, and a new enhancement for near-maximal values. In fact, we find here that the peak for the maximal lag (that is, in this case, $|L|=23$) is taller than the peak for contiguous transitions ($|L|=1$).
These findings are better gauged by plotting again the cumulative distribution, shown in the lower-right panel of Fig. 1 for all lags. It can be seen that the highest suppression of the peak occurs for $|L|^*=14$, so we have once again $|L|^* \sim 0.6 S$, where $S$ is the size of the lists. For the maximal lag $|L|=23$, the cumulative curve reaches saturation faster than the curve for contiguous transitions.
While this is an intriguing result, it relies on the assumption that all transition events could be treated independently. On the other hand, transition events within the same trial are statistically correlated, and the same may be true for transition events within different trials performed on the same subject.
To better compare these sets of data, I have averaged the IRIs corresponding to all transitions performed both on the same subject and with the same value of the lag. In Fig. 2, results are plotted as functions of the ratio between the lag and the size of the lists. Blue dots correspond to individual subjects, black curves are the corresponding histograms, while the red and green curves display the mean and median over the histograms.
Participants in the experiment of Polyn et al. were accorded a longer time for the memory test. Perhaps, this is why the curve of the mean IRI for Polyn et al. lies slightly above the curve for the data of Healey and Kahana. In spite of this, the two curves follow the same overall pattern.
For short lags, a growth in the mean IRI is observed. This requires no explanation, as we expect the thought process to move with greater speed between memories created within a shorter time from one another. As the lags grow, so does the time it takes to go from one memory to the next. This increasing trend, however, slows down in mid-range, and ceases altogether at values of $|L|/S$ in the range between $0.5$ and $0.7$. Then the average IRI begins to decrease as the lag grows.
I have computed the correlation coefficients for the ascending and descending arcs of the two histograms, by regarding the maximum as their divide point. The ascent is characterised by $r =0.31$ in Healey and Kahana, by $r= 0.32$ in Polyn et al.; the descent, by $r= - 0.18$ in Healey and Kahana and $r = - 0.23$ in Polyn et al. All these four coefficients correspond to a p-value $p<10^{-5}$.
The decreasing component of these curves is a highly counterintuitive feature. It suggests that, if two events take place within a shorter time interval, they will be recalled further apart from each other (at least for a certain range of time intervals). The Polyn et al. data are particularly surprising, because the curve reaches further down for maximal lags than for minimal lags. A memory situated $23$ memories away from the most recently visited memory is recalled faster, on average, than a memory contiguous to the last one we recalled.
It is well known that transitions with a positive lag have a higher probability of occurring than transitions with a negative lag. Since the free-recall statistics is not symmetric as a function of the lag, one may wonder if the phenomenon we have just discovered concerns only negative transitions, or only positive ones.
In Figure 3, the answer is sought by plotting the mean IRIs for both sets of experiments. The means were computed by averaging values concerning individual subjects, and the error bars are their standard deviations. The two curves are again nearly overlapping, in spite of differences in the two experimental settings.
More importantly, the non-monotonous pattern is identically displayed by both curves on both sides of the origin, with the increasing trend followed by a decreasing section. The asymmetry can be measured by computing the index $\eta= \frac{1}{S -1} \sum_{n=1}^{S -1 } \frac{\left|\bar{\tau}(n) - \bar{\tau}(-n)\right|}{\tau(n) + \tau(-n)}$ for the two sets of data, which gives $\eta_{2009}= 0.08 $ and $\eta_{2016} = 0.12$. The psychological mechanism at the root of this behavior, we may conclude, is likely to be independent of forward asymmetry.
Contiguous Recall Rate
----------------------
Let $t$ be the descrete time variable labeling steps in the recall process. The contiguous recall rate $p_{cont}(t)$ at the $t$-th step is the probability that the $t$-th recall will be effected contiguously. Otherwise said, $p_{cont}(t)$ is the probability that the serial position of $t$-th word recalled will be contiguous to the serial position of the $t$-th word recalled.
How do we expect $p_{cont}(t)$ to evolve as a function of time? We will first seek a tentative answer under some simple assumptions, then look at the actual answers.
In Fig. 4, I have plotted the distribution of the initial recall probability, as a function of serial position, for the two datasets under consideration. This is the probability distribution $p_1$ of the serial position of the first word to be recalled, and we will refer to this distribution as the [“initial conditions”]{} of the process. It has been studied extensively (Murdock, 1960; Murdock, 1962; Bjork and Whitten, 1974) and, as per the primacy and recency effects, it is concentrated near the beginning and the end of the list.
The time-averaged probability $P^0_{j i}$ of transitions from word $i$ to word $j$ is shown in figure 5A, computed for the case of the Polyn et al. data. These probabilities have been obtained by considering transitions recorded at any point during the recall process and performing an unweighed average over events. Therefore, they do not contain $\textit{all}$ the information about the recall process unless the latter is a homogenous Markov chain.
Immediate repetitions, nearly absent, have not been considered, hence the diagonal is zero. The super- and sub- diagonal elements are the most prominent terms. The transition probability decays quickly as we move toward the off-diagonal corners. Notice that there is always a finite probability for recall termination, shown as an additional column in the plot. This is equivalent to the presence of a sink state in the graph associated to the process.
We can now apply the matrix $\hat{P}^{0}$ to the distribution $p_1$ in order to obtain stochastic realizations of the supposed Markov process, over which we may compute the statistics of arbitrary variables. In particular, we are interested in the variable $P_{cont}(t)$, the contiguous recall probability, defined above.
Results are shown in panel B of figure 5. A large variability in the contiguous-recall probability emerges at the very beginning of the recall process. This is due, of course, to the fact that the initial distribution is distant from the equilibrium state. The two curves approach stationary values rather quickly. The stationary value of the contiguous recall probability has no dependence on the distribution $p_1$; it is higher for the Healey and Kahana data because, in those experiments, shorter lists have been used, hence contiguous transitions are statistically favored.
The homogenous Markovian hypothesis tells us that a quick relaxation to a steady value is the main feature of the contiguous recall rate $p_{cont}(t)$. A direct analysis of data, however, yields a completely different picture, in which the most conspicuous feature is a [“minimum”]{} in contiguous recall near the beginning of the recall process.
This is shown in Fig. 6, where the contiguous recall rate has been extracted by analyzing transition events in the two sets of experimental data. The blue dots are probability values obtained by averaging over all trials corresponding to the same subject. Black curves are histograms corresponding to each given lag, while the red curve is the average of the probability values corresponding to each subject.
As can be seen, both simulations and experiments show a large variability in the contiguous-recall probability at the very beginning of the recall process. This is due, at least in the simulations, to the fact that the initial distribution is distant from the equilibrium state. However, unlike the simulated curve, the experimental curve experiences a minimum at a recall position $t^*$, such that $t^* \sim 5$ for 16-word lists, and $t^* \sim 9$ for 24-word lists.
The contiguous recall rate then enters an increases phase, while the simulated curve stays close to a stationary value. For very large times, the experimental curves are no longer reliable as too few experimental points are available.
For comparison, the average probabilities obtained by regarding all trials as independent have been included as the green curve in Fig. 6. The behavior of the trial-averaged and subject-averaged curves is analogous, even if a different number of transitions is recorded for each subject. Just as we found for the IRI distribution, the subject-by-subject and trial-by-trial statistics are in qualitative agreement.
I have finally computed correlation coefficients for the descending and ascending branches of the contiguous recall curve. Between the beginning of the recall process and the minimum of the curve, the correlation coefficient is found to be $r=- 0.42$ for the Healey and Kahana data and $r=-0.47$ for the Polyn et al. data. Both these correlation coefficients correspond to $p<10^{-5}$. From the minimum to the last experimental point in the plot, correlation coefficients are found to be $r=0.10$ for Healey and Kahana, $r=0.22$ for Polyn et al., with $p < 10^{-4}$ in both cases.
Once again, we were expecting a monotonous trend and discovered a non-monotonous one. While in the IRI distribution we found the presence of a maximum, in the contiguous recall rate we detected a minimum. These two counterintuitive phenomena may, of course, result from independent mechanisms. In the second half of this paper, I will develop a particular hypothesis that may be a viable candidate to explaining both effects.
Hierarchic Search Hypothesis
============================
Hypothesis and toy model
------------------------
A natural approach to modeling free-recall relies on processes exploring stochastically a psychological space populated by available memories. This corresponds for instance to the model of memory retrieval used in Kenett (2014) or Abbott (2015), and it could be argued that conventional retrieved-context models are based on a similar principle, because temporal contexts are encoded by a matrix representation that includes a stochastic element.
Romani et al. (2013) represented the thought process as a random-walker moving in a particular geometry and were able to approximate, under appropriate conditions, more complex neural-network models of associative retrieval, predicting correctly the power-law scaling of free recall that we mentioned above.
In the random-walker scenario, however, it would be difficult to understand either of the two effects described in the previous section. In particular, the average IRI would have to be a monotonously increasing function of the lag. This follows from the fact that the distance between two memories is larger if the events memorised happened further apart in time. But the lag-recency effect entails that the random-walker moves continuously in space, hence a longer distance will be covered on average in a longer time. Therefore, models based on a random walker are not likely to explain why the average inter-response interval $\textit{decreases}$ as the lag grows beyond a certain point.
On the other hand, there has been a lot of work in recent times regarding the brain’s ability to carry out concurrent thought processes simultaneously (Sigman and Dehaene, 2008; Cowan, 2010); in computer science, studies have emerged on the advantages of carrying out certain tasks through a swarm-like organization of computational components (Bonabeau et al, 1999; Eberhart, 2001). This motivates the question of how the random-walker envisaged by Romani et al. (2013), Kenett (2014), Abbott (2015) would behave if replaced by several random-walkers simultaneously searching the psychological space.
The possibility we will focus upon consists in a hierarchical search, where multiple random-walkers move simultaneously in search of memories and the right to effect retrieval resides with one random-walker at a time. The space in which the random-walkers move will be modeled, for the sake of simplicity, as the space of binary arrays where each point corresponds to a string of $0$’s and $1$’s; the length of the strings is the dimensionality $d$ of the space.
The presentation stage of free recall will be modeled as the process that, starting with a given binary array, creates new memories by progressively flipping a random digit. The constraint that no state should correspond to two memories will be enforced, although in sufficiently high dimensions it is mostly unnecessary.
At the beginning of the retrieval stage, each random-walker is made to lie at the location of one of the memories. We pick these initial locations randomly and independently of each other, according to the distribution of initial words found in experimental data (Fig. 4). The retrieval process for each individual walker evolves by flipping random digits in the binary array that describes the walker’s location. Whenever the location coincide with a memory, that memory is a candidate for retrieval.
The right to effect retrieval reside on one random-walker at a time, and only that walker is allowed to retrieve memories. This scenario is hierarchical (fig. 7) because it combines two stochastic motions: the [“microscopic”]{} motion of each individual random-walker, representing parallel thought-tracks, and the [“macroscopic”]{} motion of the retrieval rights, switching stochastically among them. The latter will be assumed to be self-avoiding, no random-walker being visited twice during the recall process.
The number $N$ of walkers deployed for recall may be expected to be proportional to the magnitude of the recall task, that is, to the length $S$ of the list. Random-walkers that the macroscopic walk has not yet visited will be called [“unexplored”]{}. If $n_t$ is their number at a generic time step $t$, the system has a finite probability $q(n_t)$ of hopping onto any of the unexplored random-walkers, and a probability $1 - n_t q(n_t)$ of staying with the current one. Once a random-walker has been abandoned, it is no longer be taken under consideration.
Numerical results
-----------------
Results from simulations of this model are shown in Fig. 8, for the simple choice $ q(n)=\frac{1}{n +1}$. At the beginning of each simulation, each random-walker has been placed at the location of a word-memory, chosen in accordance with the empirical distributions in Fig. 4. The model has two parameters: the dimensionality $d$ of psychological space and the walker-per-word ratio $\kappa$, defined by $N=\kappa S$; the results in Fig. 8 were obtained in $7$ dimensions and for $\kappa=9$.
The left-hand panel of the figure shows the resulting values of the transition probability as a function of the lag. Experimental values have also been included. It can be seen that, in spite of its simplicity, the toy model reproduces qualitatively the lag recency effect for both sets of initial conditions.
In the right-hand panel, the probability of contiguous recall has been plotted as a function of time for the two initial conditions; the variable $t$ in the x-axes represents the order of recall. The same behaviour emerges for both sets of experimental conditions: the contiguous recall probability decreases sharply at small times, experiences a minimum shortly after the beginning of the recall process, then increases steadily, all the way to the end of the process.
This is precisely the behaviour we noticed in experimental data (see Fig. 6). By reasoning in terms of the hierarchical model, we may now try to explain the possible mechanism behind such behavior.
In this model, when a memory has been retrieved, there are two ways it can locate the next memory: by hopping onto another random-walker that is going to locate it, or by continuing to follow the same random-walker that located the first memory. I will refer to the former mechanism as [“zapping”]{}, to the latter as [“free association”]{}.
At the beginning of the recall process, zapping dominates: this happens because many unexplored walkers are still present, and the system is testing the possibilities offered by the many available trajectories. Once most random-walkers have been explored, the system focuses on the few left, and is forced to make a greater use of associative recall.
By construction, zapping is indifferent to serial position; associative recall, on the contrary, favours contiguous memories because each random-walker moves continuously. Therefore, the probability of contiguous recall will begin to grow.
This interpretation is confirmed by Fig. 9, where the associative component and the zapping component of the contiguous recall probability are shown separately for one of the simulations (with the Heleay and Kahana initial conditions). The observable quantity predicted by the figure is the average probability of contiguous recall, shown as a solid curve, which lies between the two dashed curves corresponding to the zapping and associative components. The smaller panel to the right shows the zapping frequency as a function of time.
At small times, zapping predominates. The probability of contiguous transitions coincides therefore with the probability of contiguity among zapping transitions. Since the initial condition is concentrated near few definite locations (the initial and final memories), zapping transitions possess initially a large degree of contiguity. This ceases to be true as the packet of random-walkers spreads away from its initial position. Zapping transitions cease to have a considerable chance of being contiguous, and since they are still the dominant type of transitions, the overall probability of contiguous recall goes down.
Gradually, the zapping frequency decreases and the frequency of associative transitions increases. As a consequence, the average probability of contiguous recall coincides less and less with the probability of contiguous transitions in the zapping process, and migrates toward the associative component.
The lag-recency effect enters now into play. Through its action, as seen in Fig. 9, the curve of contiguous recall for associative transitions lies always above the curve of contiguous recall for zapping transitions. As associative recall comes to dominate the dynamics, the probability of contiguous transitions therefore increases. By the end of the process it virtually coincides with the probability of contiguous transitions in associative recall.
The hierarchical structure of the search for memories leaves its telltale mark in the presence of two different time scales. The initial drop in contiguous recall occurs over a time scale that depends on the dimensionality of psychological space and on the distribution of initial conditions. This is a microscopic time scale, independent on the size of the random-walker population.
The subsequent increase in contiguous recall, on the other hand, happens on a macroscopic time scale, that could not be obtained from studying individual random-walker dynamics. This is the time scale over which the reservoir of unexplored random-walkers is gradually depleted, and its value is controlled by the parameter $\kappa$. We learn from experiments that $\kappa$ is large enough to allow for the depletion to stretch throughout the recall process, but low enough to let the reservoir be substantially depleted.
The rate of contiguous recall has proven to be a highly informative variable. Thinking along similar lines, we may find a qualitative explanation for what we observed in the distribution of the IRIs.
The argument is simple: in the hierarchical scenario, there are two different ways to perform fast transitions between words during the recall process, by exploiting lag-recency or by exploiting zapping. The same mechanism that creates the lag-recency effect can make associative recall faster, but only if the words involved lie nearby within the list. Hence, fast transitions due to association may only occur for small serial-position lags.
Zapping, on the other hand, can provide fast transition with any lags. Going from a random-walker that has found a memory to another random-walker that has just found another memory is equally likely regardless of the locations of the memories, and may produce fast transitions between memories located at any lag from each other. Only for very long lags, however, zapping encounters no rival mechanism, because associative recall performs those transitions too rarely to provide any competition.
Therefore, two different sets of fast transitions should exist, one from association, at short lags, and one from zapping, at long lags.
To verify this conjecture, I employ simulations of the model with the same parameters as above ($7$ dimensions, $\kappa=9$) so as to extract: (1) the mean IRI; (2) the mean IRI among associative transitions; (3) the mean IRI for zapping transitions; (4) the zapping probability. All this is plotted in Fig. 10 as a function of the lag, for both sets of initial conditions. The resulting scenario appears identical for the two initial conditions, and sheds some light on the possible meaning of Fig. 2.
The time analog of the lag-recency effect speeds up associative transitions only for very short lags; in this range, hence, the mean associative IRI (shown in blue) grows quickly as a function of the lag.
For such short lags, the zapping probability (shown in green) is considerably smaller than unity, and associative transitions play a crucial role. Therefore the average IRI (the only observable quantity, shown in black) reflects strongly the increase of the associative IRI – the more so as the zapping IRI (shown in red) is nearly constant. This results in a growth of the mean IRI, as seen in the experiments.
As the lags increases further, the zapping probability experiences a steep growth and associative transitions become rarer. The mean IRI shifts therefore toward the value of the IRI for zapping transitions. But the zapping IRI lies always beneath the associative IRI (in the whole range of lags); therefore for long lags the mean overall IRI decreases.
As an outcome, the behaviour of the average IRI as a function of the lag corresponds to the black curves of Fig. 10 – non-monotonous, increasing for short lags and decreasing for long lags – thus reproducing the experimental results of Fig. 2.
The initial conditions (see Fig.4) produce a further effect clearly discernible in Fig. 10, which adds slightly to the speeding up of transitions with near-maximal lags. Since the random-walker packet is distributed mostly on extremal word memories (near the end and the beginning of the list), zapping transitions with near-maximal lags are [“prepared”]{} by the initial condition, and happen without much searching at the very beginning of the recall process. That makes them faster, so the zapping IRI also decreases for very long lags.
As we have seen, the concentration of random-walkers on early word-memories is greater for shorter lists, i.e. in the experiments of Healey and Kahana. Accordingly, the decrease of the zapping IRI at long lags is more sustained for those initial conditions.
A final note: the indentation in the curves of the associative IRI (and, to a lesser extent, in the other curves) does not come from numerical inaccuracy; it is a corollary of the not-so-high dimensionality used for simulations. In the $d$-dimensional space of binary arrays, any two memories can be at most $d$ flips apart. If the lag between them is odd, and the dimension not very large, there is a non-negligeable chance that their distance will be unity, which makes them easier to recall associatively. This indentation, however, gradually disappears in higher dimensions and has no psychological significance.
Mathematical considerations
---------------------------
The motion of a single random walker in the space of binary arrays of length $d$ is markovian, with transition matrix $M_0 (x, y)= \frac{1}{d} \delta_{{\bf |x - y|}, 1}$. The associated distribution function $\Psi_x(t) $ (i.e. the probability of finding the walker at position $x$ at time $t$) can be found through the techniques of Kac (1947) and is equal to $\Psi_x(t) = \sum_y {{d} \choose {\|x - y\|}}^{-1} P_{\|x-y\|}(t) \Psi_y(0)$, where
$$P_n(t) = \frac{1}{2^{d}} \sum_{i=0}^{d} \ \sum_{k=\max\{ 0 , n-d+i\}}^{\min\{ n, i\}} \frac{(-1)^k d! ( 1 - 2 i/ d)^t
}{k! (i-k)! (n-k)! (d-i-n+k)!}$$
In the full stochastic process involving $\kappa S$ random walkers, the position of the random-walker currently entitled to retrieval evolves according to the transition matrix
$$\label{reformulation}
M_t(x,y) = ( 1 - \alpha_t) M_0(x,y) + \alpha_t \Psi _x(t)$$
where $\{\alpha_t\}_{t\geq 0}$ is random binary sequence constrained by $\sum_t \alpha_t = \kappa N -1$. For each given choice of the sequence $\alpha_t$, the resulting process is a nonhomogeneous Markov chain, and the stochastic object it describes is a single collective random-walker moving in the random field created by the upper level of the hierarchical search.
The model discussed in this paper should not be taken as a realistic portrait of the actual search mechanism. It is mainly designed to grasp one nontrivial aspect of the psychology involved – the coexistence of zapping and free association, with the gradual discarding of random-walkers that causes a dwindling of the zapping phase.
To infer in what ways the real search differs from the basic model I have proposed, a principle of functionality may be invoked. If we accept the above demonstration that multiple thought-particles are being deployed, the actual search mechanism must be designed to utilise them efficiently.
For instance, random-walkers may only be discarded if they have proven inefficient throughout a certain number of steps. Or, when selecting which random-walkers to visit next, the system may use some approximate knowledge on the current findings of the random-walkers: e.g., only walkers that have just located a memory may compete for retrieval rights. These more realistic models turn out to require far fewer random-walkers and far less time to perform retrieval (compare Alon et al., 2008).
For these models, the reformulation of the theory as the motion of a single collective random-walker, as per eq. (\[reformulation\]), will no longer be exact. If the macroscopic random-walk uses feedback on the random-walker trajectories, the macroscopic and microscopic motions do not decouple. However, the reformulation will still be valid at a mean-field level, with the hopping probability $\alpha_t$ now solved for self-consistently.
This approach will be further developed elsewhere. Let us only notice that, for the class of models we are considering, the mean-field theory holds a valuable psychological significance. It describes the viewpoint of an observer who interacts with the gas of random-walkers from the outside, unawares of its multiplicity.
Conclusions
===========
Two instances of non-monotonicity in free recall have been identified through a fresh analysis of data from two independent experiments.
The distribution of the inter-response intervals (IRIs) exhibits a peak at low durations for very short lags, and a similar peak for very long lags, while for intermediate lags the peaks tend to be suppressed. As a consequence, the mean and median values of the IRI experience a minimum at intermediate lag values. The rate of contiguous recall, on the other hand, exhibits a minimum near the beginning of the recall process. This minimum appears to be robust, and its position an increasing function of the list’s size.
While these findings are counterintuitive, they may be understood by assuming that multiple retrieval processes are being carried out simultaneously, and by allowing the retrieval process to switch stochastically among them. This leads to a dynamics where memories can be reached either through the switching mechanism ([“zapping”]{}) or continuously, through free association.
To test the hypothesis, I have considered a simple two-parameter model, where multiple random-walkers are allowed to explore psychological space simultaneously. The right to effect retrieval is passed around among random-walkers and can be retained for any length of time, but no random-walker is granted retrieval rights twice.
This affects drastically the curve of contiguous recall, and produces a behaviour compatible with experiments. In the early stages of the process, transitions between different random-walkers dominate. Later, associative retrieval through single-walker trajectories is prevalent and contiguous recall intensifies accordingly.
The distribution of the IRIs is also given an explanation: the first peak comes from the enhancement of associative retrieval at low lags; the second peak comes from the shorter time interval required by transitions between memories retrieved by two different random-walkers.
I would like to thank Michael J. Kahana, of the University of Pennsylvania, for generously sharing the data obtained in his laboratory.
Bibliography
============
Abbott, J. T. (2015). Random walks on semantic networks can resemble optimal foraging. *Psychol. Rev.*, *122*(3): 558–69.
Alon N., Avin C., Kouchy M., Kozma G., Lotker Z. & Tuttle M. (2008). Many random walks are faster than one. In: *Proc. of SPAA 2008*, pp. 119–128.
Binet A. & Henri. V. (1894). La memoire des mots. *L’année psychologique*, Bd. I, 1:1–23.
Bjork R. A. & Whitten W. B. (1974). Recency-sensitive retrieval processes in long-term free recall. *Cognitive Psychology*, *6*(2): 173–189.
Bonabeau, E., Theraulaz, G., & Dorigo M. (1999). *Swarm intelligence: From Natural to Artificial Systems*. Oxford (UK): Oxford University press.
Cowan, N. (2010). Multiple Concurrent Thought: the Meaning and Developmental Neuropsychology of Working Memory. *Dev. Neuropsychl.*, *35*(5): 447–474.
Ebbinghaus, H. (1913). *Memory: A contribution to experimental psychology*. New York, NY: Teachers College, Columbia University.
Eberhart, R. C., Shi Y., & Kennedy J. (2001). *Swarm Intelligence*. New York, NY: Morgann Kaufmann.
Healey & Kahana(2016). A four-component model of age-related memory change. *Psychological Review*, *123*(1):23–69.
Howard M. W. & Kahana M. J. (2002). A distributed representation of temporal context. *Journal of Mathematical Psychology*, *46*:269–299.
Kac M. (1947). Random Walk and the Theory of Brownian Motion. *The American Mathematical Monthly*, *54* (7), 1: 369–391.
Kahana, M. J. (1996). Associative retrieval processes in free recall. *Memory and Cognition*, *24*: 103–109.
Kahana M. J. (2012). *Foundations of Human Memory*. Oxford, UK: University Press.
Kenett, Y. N. (2014). Examining Search Processes in Low and High Creative Individuals with Random Walks. *Front. Hum. Neurosc.*, *8*:407–413.
Lohnas L. J. & Kahana M. J. (2013). Parametric effects of word frequency effect in memory for mixed frequency lists. *Journal of Experimental Psychology: Learning, Memory, and Cognition*, *39*:1943–1946.
Murdock B. B. (1960). The immediate retention of unrelated words. *Journal of Experimental Psychology*, *60*:222–234.
Murdock B. B. (1962). The serial position effect of free recall. *Journal of Experimental Psychology*, *64*(5):482–488.
Murray D. J., Pye C., & Hockley W. E. (1976). Standing’s power function in long-term memory. *Psychological Research*, *38*(4):319–331.
Polyn S. M., Norman K. A., & Kahana M. J. (2009). A context maintenance and retrieval model of organizational processes in free recall. *Psychological Review*, *116*:129–156.
Romani S., Pinkoviezky I., Rubin A., Tsodyks M. (2013). Scaling laws of associative memory retrieval. *Neural Computation*, *25*:2523–2544.
Sigman and Dehaene (2008). Brain Mechanisms of Serial and Parallel Processing. *Journal of Neuroscience*, *23*(30): 7585–7598.
Standing L. (1973). Learning 10.000 pictures. *Quarterly Journal of Experimental Psychology*, *25*:207-222.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This work deals with the modeling of nonsmooth vibro-impact motion of a continuous structure against a rigid distributed obstacle. Galerkin’s approach is used to approximate the solutions of the governing partial differential equations of the structure, which results in a system of ordinary differential equations (ODEs). When these ODEs are subjected to unilateral constraints and velocity jump conditions, one must use an event detection algorithm to calculate the time of impact accurately. Event detection in the presence of multiple simultaneous impacts is a computationally demanding task. Ivanov (Ivanov, A., $1993$. “Analytical methods in the theory of vibro-impact systems", Journal of Applied Mathematics and Mechanics, $57(2)$, pp. $221$–$236$.) proposed a nonsmooth transformation for a vibro-impacting multi-degree-of-freedom system subjected to a single unilateral constraint. This transformation eliminates the unilateral constraints from the problem and, therefore, no event detection is required during numerical integration. This nonsmooth transformation leads to sign function nonlinearities in the equations of motion. However, they can be easily accounted for during numerical integration. Ivanov used his transformation to make analytical calculations for the stability and bifurcations of vibro-impacting motions; however, he did not explore its application for simulating distributed collisions in spatially continuous structures. We adopt Ivanov’s transformation to deal with multiple unilateral constraints in spatially continuous structures. Also, imposing the velocity jump conditions exactly in the modal coordinates is nontrivial and challenging. Therefore, in this work we use a modal-physical transformation to convert the system from modal to physical coordinates on a spatially discretized grid. We then apply Ivanov’s transformation on the physical system to simulate the vibro-impact motion of the structure. The developed method is demonstrated by modeling the distributed collision of a nonlinear string against a rigid distributed surface. For validation, we compare our results with the well-known penalty approach.'
author:
- Surya Samukham
- 'S. N. Khaderi'
- 'C. P. Vyasarayani [^1]'
title: 'Galerkin–Ivanov transformation for nonsmooth modeling of vibro-impacts in continuous structures'
---
Introduction
============
In many engineering applications, structures are subjected to vibro-impacting motions [@babitsky2013theory; @ibrahim2009vibro; @ibrahim2014recent]. Clearances in mechanical joints due to wear can lead to vibro-impacting motions in machine components. In some vibration-based energy harvesting applications, vibro-impacting motions are deliberately introduced to increase the bandwidth of the frequency response [@vijayan2015non]. In addition, the study of vibro-impact motion of a structure has various applications in the synthesis of sound in stringed musical instruments of Indian origin, like the sitar and the tanpura [@chatziioannou2015energy; @issanchou2017modal; @issanchou2018nonsmooth]. Applications are also found in the elevator-rope collision problem [@shin2018vibration], multi-body dynamics [@shafei2018oblique], and impact-damper systems [@li2006experiments; @yang2019investigation]. However, in this work, we are interested in the problem of modeling the vibro-impact motion of a continuous structure constrained by a rigid distributed obstacle. When a continuous structure is spatially discretized using numerical approximation methods, there are three ways the impacting motion can be simulated. The first method approximates the rigid obstacle as a foundation of springs with high stiffness and is known as the penalty approach [@gilardi2002literature; @bilbao2015numerical; @chatziioannou2015energy; @bilbao2015numerical2; @issanchou2018string; @kang2018calculation]. Alternatively, one can use Lagrange multipliers to impose contact constraints once the structure comes into contact with the obstacle. Complementarity conditions between Lagrange multipliers and gap functions are used to impose contact loss [@wagg2005periodic]. The Lagrange multiplier approach simulates the perfect sticking motion of vibro-impacting systems [@wagg2005periodic]. Impacts can also be simulated using a coefficient of restitution (CoR or $R$) approach [@gilardi2002literature; @li2015modeling; @uchida2015making; @liakou2016fast; @huang2016multi; @rebouccas2019unilateral]. Once contact is detected between the structure and the obstacle, the appropriate velocity jump conditions are imposed at the point of contact. In the Lagrange multiplier and CoR approaches, one must solve an event detection problem to isolate the time of impact accurately. Multiple simultaneous impacts are present when a continuous structure impacts a distributed obstacle, which makes the event detection problem computationally expensive. A few other recently developed numerical time-integration techniques to simulate the nonsmooth dynamical systems are given in [@acary2008numerical], [@studer2009numerics], and [@acary2016energy].
Within the framework of the CoR-based approach, Ivanov [@ivanov1993analytical; @ivanov1994impact] proposed a nonsmooth spatial transformation that automatically satisfies both the unilateral constraints and the velocity jump conditions at the point of contact. This approach eliminates the need for event detection. Ivanov proposed this method to study the vibro-impacting motion of a multi-degree-of-freedom (MDOF) system, where the displacement of a single mass is constrained. Ivanov’s transformation was successfully used to study the nonlinear dynamics of a single-degree-of-freedom (SDOF) vibro-impacting ship motion [@grace2011inelastic; @grace2011inelastic2]. In this work, we adopt Ivanov’s transformation to account for the multiple unilateral constraints in a MDOF system.
In this paper, we have implemented Ivanov’s approach to simulate the vibro-impacting motion of a continuous structure against a rigid distributed obstacle. First, we discuss the implementation of Ivanov’s transformation to a SDOF system, followed by its application to the MDOF system (continuous structure). To demonstrate the application of Ivanov’s approach to a continuous structure subjected to distributed contact, we consider a nonlinear string and present the vibro-impacting motion of the string against the rigid flat and sinusoidal obstacles for different CoR. Galerkin approximation has been used to solve the governing partial differential equations (PDEs) of the string. The set of ordinary differential equations (ODEs) obtained from the Galerkin approach are in modal coordinates and the imposition of impact constraints upon the modal system is a challenging task [@vyasarayani2010modeling; @wagg2002application; @wei2019effect]. To avoid such difficulties, we use a transformation to convert the modal system into its physical coordinates by discretizing the string in space [@van2016simulation; @issanchou2017modal]. We then apply Ivanov’s transformation to the spatially discretized system to incorporate the impact constraints. After Ivanov’s transformation, the resulting system of differential equations are in Ivanov’s coordinates and non-stiff in nature. Therefore, an explicit integrator with a large time-step size can be used to solve the system of differential equations, which is not the case with the penalty method because of the presence of stiff differential equations. This advantage of implementing an explicit integrator with a larger time-step size in a MDOF system significantly reduces the computational cost and time compared to an implicit integrator. The results obtained from Ivanov’s method in this work have been verified with the penalty method for the case of $R=1$. The main advantages of the proposed method are as follows:
1. The nonsmooth nature of the contact problem is preserved, which is not the case with the penalty approach.
2. The equations in Ivanov’s coordinates are non-stiff, in contrast to those obtained using the penalty approach.
3. Explicit integration schemes can be used with larger time-step sizes for integration purposes instead of a computationally expensive implicit integrator.
4. No event detection is necessary for simulating impacts, which is the most important benefit of Ivanov’s approach, since the problem of event detection is challenging in the case of large MDOF systems (like Galerkin approximations and finite element analysis of continuous structures) subjected to distributed collisions.
This paper is organized as follows: In Sec. [\[Model\]]{}, we describe in detail how Ivanov’s transformation works by modeling a point mass falling on a rigid obstacle. Using the example of a vibro-impacting SDOF system, we compare the advantages of Ivanov’s method over the penalty method in Sec. \[vibroimp\]. The equations of motion for a nonlinear MDOF system in Ivanov’s coordinates are derived in Sec. \[mainder\]. In Sec. \[galerkin\], we discuss the Galerkin approximation of the governing differential equation of a nonlinear string and the modal-to-physical coordinate transformation. In Sec. \[results\], we apply Ivanov’s transformation to simulate the vibro-impact motion of the string against a distributed obstacle. The results are validated and compared with the penalty approach. The contributions of the paper have been summarized in Sec. \[concl\].
Mathematical modeling
=====================
In this section, using the example of a point mass bouncing on a rigid surface, we illustrate the idea behind Ivanov’s transformation. Later, by using the example of a vibro-impacting oscillator, we demonstrate the advantages of modeling impact using Ivanov’s method over the penalty approach. When simulating rigid collisions using a penalty approach, the penalty stiffness term is usually selected to be large, which results in stiff differential equations. Due to the stiff nature of the differential equations, one must use a very small time-step size during numerical integration. However, in Ivanov’s method, the obtained differential equations in the transformed coordinates are non-stiff. Unlike the penalty approach, Ivanov’s method accurately captures the nonsmooth behavior in velocity during impact.
Point mass bouncing on a rigid surface {#Model}
--------------------------------------
To illustrate Ivanov’s transformation, we consider the problem of a point mass, under the influence of gravity, impacting a rigid surface (see Fig. \[Figure1\]). The equation of motion for this problem can be written as follows: $$\ddot{p}=-g, \thinspace\thinspace\thinspace p\ge0,
\label{Sec.2.1.eq1}$$ with initial conditions $p(0)=\alpha_0$ and $\dot{p}(0)=\beta_0$. When the mass makes contact with the obstacle ($p(t_c)=0$) at time $t_c$, the following jump condition in velocity must be respected: $$\dot{p}(t_{c}^{+})=-R\dot{p}(t_{c}^{-}),
\label{Sec.2.1.eq2}$$ where $R$ is the coefficient of restitution. Equations (\[Sec.2.1.eq1\]) and (\[Sec.2.1.eq2\]) completely describe the motion of the mass. To apply Ivanov’s transformation, we recast Eqs. (\[Sec.2.1.eq1\]) and (\[Sec.2.1.eq2\]) in state-space form. Introducing $u=p$ and $v=\dot{p}$, we get $$\left\{ \begin{array}{c}
\dot{u}\\
\dot{v}
\end{array} \right\} =\left[\begin{array}{cc}
0 & 1\\
0 & 0
\end{array}\right]\left\{ \begin{array}{c}
u\\
v
\end{array}\right\} +\left\{ \begin{array}{c}
0\\
-g
\end{array}\right\}, \thinspace\thinspace\thinspace u\ge0
\label{Sec.2.1.eq3}$$ and when $u(t_c)=0$, we have $$v(t_{c}^{+})=-Rv(t_{c}^{-})
\label{Sec.2.1.eq4}$$
![Schematic of a point mass falling on a rigid surface.[]{data-label="Figure1"}](point_mass_schematic.eps){width="40.00000%"}
Ivanov introduced the following nonsmooth transformation: $$\begin{aligned}
\label{Sec.2.1.eq5}
&u =\eta ~\text{sgn}(\eta)\\
\label{Sec.2.1.eq6}
&v =\zeta\left(1-k~ \text{sgn}(\eta\zeta)\right)\text{sgn}(\eta)\end{aligned}$$ In Eq. (\[Sec.2.1.eq6\]), $k=\frac{1-R}{1+R}$. Equation (\[Sec.2.1.eq5\]) ensures that $u\ge0$ for all time. To understand the transformation given by Eq. (\[Sec.2.1.eq6\]), we need to look at the phase-space of $\eta-\zeta$, as shown in Fig. \[Figure2\](b); the actual phase-space (in $u-v$) is shown in Fig. \[Figure2\](a) for comparison. In the first and third quadrants of the $\eta-\zeta$ plane, we have $\text{sgn}(\eta\zeta)>0$. In the second and fourth quadrants of the $\eta-\zeta$ plane, we have $\text{sgn}(\eta\zeta)<0$. When a trajectory $AB^{-}$ starting in phase-space $\eta-\zeta$ from point $A$ is about to cross $\eta=0$ at time $t_{B}^{-}$ from the fourth to the third quadrant, from Eq. (\[Sec.2.1.eq6\]), we have $$\label{Sec.2.1.eq7}
v(t_{B}^{-})=\left(1+k\right)\zeta(t_{B}^{-})$$ Immediately after the contact, at time $t_B^{+}$, we have $$\label{Sec.2.1.eq8}
v(t_{B}^{+})=-\left(1-k\right)\zeta(t_{B}^{+})$$ Similarly, when the trajectory $B^{+}C^{-}$ is about to cross $\eta=0$ at time $t_{C}^{-}$ from the second to the first quadrant, we have $$\label{Sec.2.1.eq9}
v(t_{C}^{-})=-\left(1+k\right)\zeta(t_{C}^{-})\\$$ After crossing $\eta=0$, we have $$\label{Sec.2.1.eq10}
v(t_{C}^{+})=\left(1-k\right)\zeta(t_{C}^{+})$$ It should be noted that both $\eta$ and $\zeta$ are continuous, so we have $\zeta(t_{B}^{-})=\zeta(t_{B}^{+})$ and $\zeta(t_{C}^{-})=\zeta(t_{C}^{+})$, which results in the following expressions: $$\begin{aligned}
\label{Sec.2.1.eq111}
v(t_{B}^{+})=-\left(\frac{1-k}{1+k}\right)v(t_{B}^{-})=-Rv(t_{B}^{-})\\
\label{Sec.2.1.eq112}
v(t_{C}^{+})=-\left(\frac{1-k}{1+k}\right)v(t_{C}^{-})=-Rv(t_{C}^{-})\end{aligned}$$ Therefore, it is clear that the nonsmooth transformation given by Eq. (\[Sec.2.1.eq6\]) automatically satisfies Eq. (\[Sec.2.1.eq4\]) at the event of an impact. Due to the jump conditions (Eq. (\[Sec.2.1.eq111\]) and Eq. (\[Sec.2.1.eq112\])) imposed on the velocity $v$ at the time of impact, the trajectories in $\eta-\zeta$ space that enter the third quadrant from the fourth quadrant can only go to the first quadrant through the second quadrant.
Equations (\[Sec.2.1.eq5\]) and (\[Sec.2.1.eq6\]) can be written as follows: $$\left\{ \begin{array}{c}
u\\
v
\end{array}\right\} =\boldsymbol{S}\left\{ \begin{array}{c}
\eta\\
\zeta
\end{array}\right\}
\label{Sec.2.1.eq12}$$ The transformation matrix, $\boldsymbol{S}$ in Eq. (\[Sec.2.1.eq12\]), is given by $$\label{Sec.2.1.eq13}
\boldsymbol{S}=\left[\begin{array}{cc}
\text{sgn}(\eta) & 0\\
0 & \left(1-k~\text{sgn}(\eta \zeta)\right)\text{sgn}(\eta)
\end{array}\right]$$ The matrix $\boldsymbol{S}$ is constant in each quadrant of phase-space (see Fig. \[Figure2\]). Substituting Eq. (\[Sec.2.1.eq12\]) into Eq. (\[Sec.2.1.eq3\]), we get $$\left\{ \begin{array}{c}
\dot{\eta}\\
\dot{\zeta}
\end{array}\right\} =\boldsymbol{S}^{-1}\left[\begin{array}{cc}
0 & 1\\
0 & 0
\end{array}\right]\boldsymbol{S}\left\{ \begin{array}{c}
\eta\\
\zeta
\end{array}\right\} +\boldsymbol{S}^{-1}\left\{ \begin{array}{c}
0\\
-g
\end{array}\right\}
\label{Sec.2.1.eq14}$$ The initial conditions for $\eta$ and $\zeta$ can be calculated as follows: $$\left\{ \begin{array}{c}
\eta(0)\\
\zeta(0)
\end{array}\right\} =\boldsymbol{S}^{-1}\left\{ \begin{array}{c}
u(0)\\
v(0)
\end{array}\right\}
\label{Sec.2.1.eq15}$$ Equation (\[Sec.2.1.eq14\]) can be integrated numerically to obtain $\eta$ and $\zeta$. Later, Eq. (\[Sec.2.1.eq12\]) can be used to calculate $u$ and $v$. It should be noted that there are no constraints on Eq. (\[Sec.2.1.eq14\]) and its solutions are continuous functions of time. Figure \[Figure3a\] shows the displacement of the point mass in transformed coordinates ($\eta$, solid blue line) and physical coordinates ($u$, dashed red line). Similarly, Fig. \[Figure3b\] shows the velocity in transformed coordinates ($\zeta$, solid blue line) and physical ($v$, dashed red line) coordinates. Figure \[Figure3c\] shows the total energy of the point mass, and it can be seen that energy is lost at every collision. If we integrate Eq. (\[Sec.2.1.eq14\]) for a sufficiently long time, the energy will approach zero. Also, the frequency of the transformed solution ($\eta$) continuously increases due to the chattering nature of the physical solution ($u$). If we use an adaptive time-step integrator, the integrator automatically reduces the time-step sizes to account for increasing frequency in the solution. For practical reasons, if the energy falls below a small threshold, we can assume that the point mass has reached equilibrium.
Single-degree-of-freedom vibro-impacting system {#vibroimp}
-----------------------------------------------
To demonstrate the advantage of modeling impact using Ivanov’s method over a penalty approach, we consider the problem of a SDOF vibro-impacting system, as shown in Fig. \[Figure4\](a). In the penalty approach, the rigid obstacle as shown in Fig. \[Figure4\](a) is replaced with a high-stiffness ($k_p$) linear spring (see Fig. \[Figure4\](b)), usually referred to as the penalty stiffness. For a large value of $k_p$, the spring behaves approximately like a rigid obstacle. In a penalty approach, we solve the following equation: $$\label{Sec.2.2.eq1}
m\ddot{p}+c\dot{p}+(k_{s}+\mu k_{p})p=0$$ In Eq. (\[Sec.2.2.eq1\]), $\mu$ is defined as follows: $$\mu=\begin{cases}
0 & p\ge 0\\
1 & p<0
\end{cases}
\label{Sec.2.2.eq2}$$ From Eq. (\[Sec.2.2.eq1\]), we can see that the system has a time-scale of $\tau_1=2\pi \sqrt{\frac{m}{k_s}}$ when the mass is not in contact with the obstacle ($\mu=0$); when the mass is in contact with the obstacle ($\mu=1$), it has a time-scale of $\tau_2=2\pi \sqrt{\frac{m}{k_s+k_{p}}}$. Having two separate time-scales of very different magnitudes makes Eq. (\[Sec.2.2.eq1\]) a stiff differential equation. Therefore, when an explicit integrator is used to solve Eq. (\[Sec.2.2.eq1\]), one must use a time-step size much smaller than $\tau_2$ (the smallest time-scale present in the solution) to obtain accurate results. The equation of motion of the system, considering a perfectly rigid obstacle ($R=1$) (see Fig. \[Figure4a\]), can be written as follows: $$\label{Sec.2.2.eq3}
m\ddot{p}+c\dot{p}+k_{s}p=0, ~~p\ge d$$ To apply Ivanov’s method for solving the vibro-impact problem (Eq. (\[Sec.2.2.eq3\])), we substitute $u=p-d$ and $v=\dot{p}$ in Eq. (\[Sec.2.2.eq3\]) to obtain $$\begin{aligned}
\label{Sec.2.2.eq4}
\left\{ \begin{array}{c}
\dot{u}\\
\dot{v}
\end{array}\right\} =\left[\begin{array}{cc}
0 & 1\\
-k_s/m & -c/m
\end{array}\right]\left\{ \begin{array}{c}
u\\
v
\end{array}\right\} \nonumber \\ + \left\{ \begin{array}{c}
0\\
-k_s d/m
\end{array}\right\},\,~u\ge0 \end{aligned}$$ By following the same transformation as discussed in the point mass bouncing problem (see Eqs. (\[Sec.2.1.eq5\]) and (\[Sec.2.1.eq6\])), we can write the equations of motion for the vibro-impacting system with perfectly rigid impacts in Ivanov’s coordinates as follows: $$\begin{aligned}
\left\{ \begin{array}{c}
\dot{\eta}\\
\dot{\zeta}
\end{array}\right\} =\boldsymbol{S}^{-1}\left[\begin{array}{cc}
0 & 1\\
-k_s/m & -c/m
\end{array}\right]\boldsymbol{S}\left\{ \begin{array}{c}
\eta\\
\zeta
\end{array}\right\} \nonumber \\ + \boldsymbol{S}^{-1} \left\{ \begin{array}{c}
0\\
-k_s d/m
\end{array}\right\}
\label{Sec.2.2.eq5}\end{aligned}$$ (Refer to Eq. (\[Sec.2.1.eq13\]) for the definition of the matrix $\boldsymbol{S}$.) As pointed out earlier (see Fig. \[Figure2\]), the matrix $\boldsymbol{S}$ is diagonal and constant in each quadrant of the phase-space. Further, the magnitude of the elements of $\boldsymbol{S}$ are of the same order in all quadrants (see Fig. \[Figure2\]). Therefore, when Eq. (\[Sec.2.2.eq5\]) is solved using an explicit integrator, it does not appear as a stiff equation to the integrator.
First, we determine the solution of the system shown in Fig. \[Figure4a\] subject to the constraint $p \ge 0.5$ using the event detection technique. The solution of Eq. (\[Sec.2.2.eq3\]) is obtained using the “ode45" (adaptive time-step) integrator in MATLAB with relative and absolute tolerances of $1\times10^{-12}$. The parameters considered for the analysis are $m=1$, $k=1$, $c=0$, and the integration has been carried out over a time of $t=0$ to $t=10$. The initial conditions are $p(0)=1$ and $\dot{p}(0)=0$. This reference solution $p_r(t)$ is used to compare the results from Ivanov’s method and the penalty method with an explicit fixed-time-step-size integrator. Now, we solve the same problem (Fig. \[Figure4\]) using the penalty approach (Eq. (\[Sec.2.2.eq1\])) and Ivanov’s method (Eq. (\[Sec.2.2.eq5\])) using a fixed-time-step-size, fourth-order Runge–Kutta integrator. Different time-step sizes ($\Delta t$) in the range of $10^{-4}$ to $10^{-1}$ are considered for the analysis. Further, the mean squared error (MSE) given by $e=\frac{1}{n}\sum_{i=1}^{n}(p_r(t_i)-p(t_i))^2$ is evaluated between the reference solution $p_r(t)$ and the solutions obtained from Ivanov’s method and the penalty method ($p(t)$) for different time-step sizes of integration. Here, $n$ is the number of sample points considered for the error estimation and $t_i$ is the time instant. The variation of MSE for the two methods with respect to the change in integration time-step size is shown in Fig. \[Figure5\]. In Fig. \[Figure5\], $e$ is calculated by evaluating the solutions ($p(t)$ and $p_r(t)$) at $n=101$ equally spaced time intervals between $t=0$ and $t=10$. The MSE for the penalty approach is calculated for $3$ different penalty stiffness values ($k_p= 10^5, 10^6,$ and $10^7$) and the corresponding variation in $e$ is shown in Fig. \[Figure5\]. From the results shown in Fig. \[Figure5\], it is clearly observed that the MSE in the penalty method increases with an increase in $\Delta t$. This is because of the penalty stiffness $k_p$ in Eq. (\[Sec.2.2.eq1\]), which decreases the fastest time-scale $\tau_2$ and makes the equation stiff. Therefore, as discussed earlier, when an explicit integrator is used for the integration, $\Delta t$ must be chosen much smaller than $\tau_2$. As a result, in the penalty method, the suitable time-step size for integration decreases with an increase in the penalty stiffness $k_p$, which can be clearly observed from Fig. \[Figure5\]. Also, with the increase in $k_p$, the MSE in the solution decreases for the smaller integration time-step sizes (Fig. \[Figure5\]). Therefore, to obtain an accurate solution in the penalty method, a very large value of $k_p$ must be chosen. However, this decreases $\tau_2$ and a correspondingly smaller time-step size must be used for integration, which increases the computation cost significantly. In contrast, the final equations obtained in Ivanov’s method are not stiff (see Eq. (\[Sec.2.2.eq5\])) and hence a larger time-step size can be used for integration without reducing the accuracy of the solution. We can clearly see this in Fig. \[Figure5\]: the error in the solution obtained from Ivanov’s method is of an order smaller than $10^{-4}$ even when $\Delta t=10^{-1}$, which is not the case with the penalty approach.
In summary, with the penalty approach, to obtain an accurate solution of a vibro-impact system, the time-step size must chosen to be very small depending on the penalty stiffness $k_p$. However, by using Ivanov’s method, an accurate solution can be obtained with a much larger time-step size. Moreover, for simulating a rigid collision, one must use a large value for the penalty stiffness, which puts a hard restriction on the numerical time-step size of the integrator. However, in Ivanov’s approach, the nonsmooth transformation simulates the case of infinite penalty stiffness without making the equations stiff.
![Variation of mean squared error ($e=\frac{1}{n}\sum_{i=1}^{n}(p_r(t_i)-p(t_i))^2$) with respect to the time-step size of integration ($\Delta t$) in Ivanov’s method and the penalty method.[]{data-label="Figure5"}](error_iv_pen_ref_ed.eps){width="50.00000%"}
Ivanov’s transformation for distributed impacts {#mainder}
-----------------------------------------------
Having demonstrated the applicability and advantage of Ivanov’s transformation for an SDOF system, we now discuss its extension to MDOF systems where only some degrees of freedom are subjected to impact constraints. To develop Ivanov’s transformation for the most general MDOF system, we consider the following coupled nonlinear model where some degrees of freedom ($\boldsymbol{p}$) are subjected to impact while the others ($\boldsymbol{q}$) are not subjected to any impact. The equations of motion for such a mechanical system can be written as follows: $$\begin{aligned}
\label{Sec.2.3.eq1}
\ddot{\boldsymbol{p}}= & \boldsymbol{f}(\boldsymbol{p},\dot{\boldsymbol{p}},\boldsymbol{q},\dot{\boldsymbol{q}},t)\\
\label{Sec.2.3.eq2}
\ddot{\boldsymbol{q}}= & \boldsymbol{g}(\boldsymbol{p},\dot{\boldsymbol{p}},\boldsymbol{q},\dot{\boldsymbol{q}},t),\end{aligned}$$ with the initial conditions $\boldsymbol{p}(\boldsymbol{0})=\boldsymbol{\alpha}_0$, $\dot{\boldsymbol{p}}(\boldsymbol{0})=\boldsymbol{\beta}_0$, $\boldsymbol{q}(\boldsymbol{0})=\boldsymbol{\gamma}_0$, and $\dot{\boldsymbol{q}}(\boldsymbol{0})=\boldsymbol{\nu}_0$. In Eq. (\[Sec.2.3.eq1\]), the displacement coordinates $\boldsymbol{p}=[p_{1}(t),\thinspace p_{2}(t),...,p_{m}(t)]^\text{T}$ are subjected to impact constraints of the form $$p_{i}\ge d_{i},\thinspace\thinspace i=1,2,...,m
\label{Sec.2.3.eq3}$$ Once any of the above constraints becomes an equality constraint ($p_{i}(t_c)= d_{i}$) at time $t_c$, a velocity jump condition is imposed as follows: $$\dot{p}_{i}(t_c^{+})=-R\dot{p}_{i}(t_c^{-})
\label{Sec.2.3.eq4}$$ By defining $\boldsymbol{d}=[d_1,d_2,...,d_m]^\text{T}$, Eq. (\[Sec.2.3.eq3\]) can be written compactly as $$\boldsymbol{p}-\boldsymbol{d} \ge \boldsymbol{0}
\label{Sec.2.3.eq5}$$ It should be noted that the displacement coordinates $\boldsymbol{q}=[q_{1}(t),\thinspace q_{2}(t),...,q_{n}(t)]^\text{T}$ are not subjected to any impact constraints. By introducing the state variables $\boldsymbol{u}=\boldsymbol{p}-\boldsymbol{d}$, $\boldsymbol{v}=\dot{\boldsymbol{p}}$, $\boldsymbol{r}=\boldsymbol{q}$, and $\boldsymbol{s}=\dot{\boldsymbol{q}}$, Eqs. (\[Sec.2.3.eq1\]) and (\[Sec.2.3.eq2\]) can be written as follows: $$\begin{aligned}
\label{Sec.2.3.eq6}
\dot{\boldsymbol{u}}&=&\boldsymbol{v}\\ \label{eq2_vcp}
\dot{\boldsymbol{v}}&=&\boldsymbol{f}(\boldsymbol{u}+\boldsymbol{d},\boldsymbol{v},\boldsymbol{r},\boldsymbol{s},t)\\ \label{Sec.2.3.eq7}
\dot{\boldsymbol{r}}&=&\boldsymbol{s}\\ \label{eq4_vcp}
\dot{\boldsymbol{s}}&=&\boldsymbol{g}(\boldsymbol{u}+\boldsymbol{d},\boldsymbol{v},\boldsymbol{r},\boldsymbol{s},t)\end{aligned}$$ The impact constraint becomes $\boldsymbol{u} \ge 0$, where $\boldsymbol{u}=[u_{1}(t),\thinspace u_{2}(t),...,u_{m}(t)]^\text{T}$. Once any of the above constraints becomes an equality constraint ($u_{i}(t_c)=0$) at time $t_c$, a velocity jump condition is imposed as follows: $$\label{Sec.2.3.eq8}
v_{i}(t_c^{+})=-Rv_{i}(t_c^{-})$$ To apply Ivanov’s method, the following transformation is introduced: $$\begin{aligned}
\label{Sec.2.3.eq9}
\boldsymbol{u}&=&\boldsymbol{T}\boldsymbol{\eta}\\\label{Sec.2.3.eq10}
\boldsymbol{v}&=&\boldsymbol{W}\boldsymbol{\zeta}\end{aligned}$$ The matrices $\boldsymbol{T}$ and $\boldsymbol{W}$ are defined as follows: $$\begin{aligned}
\label{Sec.2.3.eq11}
\boldsymbol{T}&=&\text{diag}\left(\text{sgn}(\boldsymbol{\eta})\right)\\
\label{Sec.2.3.eq12}
\boldsymbol{W}&=&\text{diag}\left(\left(1-k~\text{sgn}(\boldsymbol{\eta}\circ\boldsymbol{\zeta})\right)\circ\text{sgn}(\boldsymbol{\eta})\right)\end{aligned}$$ In Eq. (\[Sec.2.3.eq12\]), the symbol “$\circ$” represents the element-by-element multiplication (Hadamard product) of two vectors. Substituting Eqs. (\[Sec.2.3.eq9\]) and (\[Sec.2.3.eq10\]) into Eqs. (\[Sec.2.3.eq6\])-(\[eq4\_vcp\]), we get $$\begin{aligned}
\label{Sec.2.3.eq13}
\dot{\boldsymbol{\eta}}&=&\boldsymbol{T}^{-1}\boldsymbol{W}\boldsymbol{\zeta}\\ \label{Sec.2.3.eq14}
\dot{\boldsymbol{\zeta}}&=&\boldsymbol{W}^{-1}\boldsymbol{f}(\boldsymbol{T}\boldsymbol{\eta}+\boldsymbol{d},\boldsymbol{W}\boldsymbol{\zeta},\boldsymbol{r},\boldsymbol{s},t)\\ \label{Sec.2.3.eq15}
\dot{\boldsymbol{r}}&=&\boldsymbol{s}\\ \label{Sec.2.3.eq16}
\dot{\boldsymbol{s}}&=&\boldsymbol{g}(\boldsymbol{T}\boldsymbol{\eta}+\boldsymbol{d},\boldsymbol{W}\boldsymbol{\zeta},\boldsymbol{r},\boldsymbol{s},t) \end{aligned}$$ The initial conditions for Eqs. (\[Sec.2.3.eq13\])–(\[Sec.2.3.eq16\]) are $$\begin{aligned}
\label{Sec.2.3.eq17}
\boldsymbol{\eta}(0)&=&\boldsymbol{T}^{-1}\boldsymbol{u}(0)=\boldsymbol{T}^{-1}\left(\boldsymbol{p}(0)-\boldsymbol{d}\right) \\ \nonumber
&=&\boldsymbol{T}^{-1}\left(\boldsymbol{\alpha}_{0}-\boldsymbol{d}\right)\\ \label{Sec.2.3.eq18}
\boldsymbol{\zeta}(0)&=&\boldsymbol{W}^{-1}\boldsymbol{v}(0)=\boldsymbol{W}^{-1}\dot{\boldsymbol{p}}(0)=\boldsymbol{W}^{-1}\boldsymbol{\beta}_{0}\\ \label{Sec.2.3.eq19}
\boldsymbol{r}(0)&=&\boldsymbol{q}(0)=\boldsymbol{\gamma}_{0}\\ \label{Sec.2.3.eq20}
\boldsymbol{s}(0)&=&\boldsymbol{\dot{q}}(0)=\boldsymbol{\nu}_{0} \end{aligned}$$ The solution of Eqs. (\[Sec.2.3.eq13\])–(\[Sec.2.3.eq16\]) automatically satisfies all the constraints imposed on $\boldsymbol{p}$ (Eq. (\[Sec.2.3.eq3\])) and $\boldsymbol{\dot{p}}$ (Eq. (\[Sec.2.3.eq4\])). Equations (\[Sec.2.3.eq13\])–(\[Sec.2.3.eq16\]) are the closed-form equations for a nonlinear MDOF system (where only some degrees of freedom are subjected to constraints) in Ivanov’s coordinates. Upon solving the system in Ivanov’s coordinates, the physical solution (displacements and velocities) can be reconstructed by using the inverse of the transformations given by Eqs. (\[Sec.2.3.eq9\]) and (\[Sec.2.3.eq10\]).
Vibro-impact motion of a string {#galerkin}
===============================
To demonstrate the application of the proposed methodology, we simulate the vibro-impact motion of a nonlinear string against a rigid distributed obstacle (see Fig. \[Figure7\]). The governing equation of motion for the string is given by [@oplinger1960frequency] $$\rho \frac{\partial^2 Y}{\partial T^2} - T_0 \left(1+ \frac{EA}{2T_0 L} \int_{0}^{L} \left(\frac{\partial Y}{\partial X}\right)^2 dX\right) \frac{\partial^2 Y}{\partial X^2}+C\frac{\partial Y}{\partial T}=0,
\label{galerkin_eq1}$$ with the following boundary conditions: $$Y(0,T)=Y(L,T)=0
\label{galerkin_eq2}$$ Here, $Y$ is the transverse displacement of the string, $L$ is the length, $A$ is its cross-sectional area, $T_0$ is the string tension at the equilibrium position, $\rho$ is the mass density, $C$ is the damping coefficient, $E$ is the Young’s modulus of the material of the string, $X$ is the coordinate along the length of the string, and $T$ is time.
![Schematic of a string with a distributed impact constraint.[]{data-label="Figure7"}](string_impact.eps){width="50.00000%"}
We now substitute the following non-dimensional parameters into the governing equation (Eq. (\[galerkin\_eq1\])) to facilitate the analysis: $$\begin{aligned}
& &y=\frac{Y}{H},~ x=\frac{X}{L},~ t=\frac{T}{\alpha},~ \gamma={\frac{EAH^2}{2T_0L^2}}, ~\alpha=\sqrt{\frac{\rho L^2}{T_0}}, \nonumber\\ & & c=\frac{C \alpha}{\rho},~\text{and}~ d(x)=\frac{H-D(x)}{H}.
\label{galerkin_eq3}\end{aligned}$$ Upon substituting the non-dimensional parameters (Eq. (\[galerkin\_eq3\])), the governing equation of motion (Eq. (\[galerkin\_eq1\])) can be written in non-dimensional form as follows: $$\frac{\partial^2 y}{\partial t^2} - \left(1+\gamma\int_{0}^{L} \left(\frac{\partial y}{\partial x}\right)^2 dx\right) \frac{\partial^2 y}{\partial x^2}+c\frac{\partial y}{\partial t}=0,
\label{galerkin_eq4}$$ with the boundary conditions: $$y(0,t)=y(1,t)=0
\label{galerkin_eq5}$$ Now, the solution of Eq. (\[galerkin\_eq4\]) is assumed as follows: $$y(x,t)=\sum_{j=1}^{N}\phi_j(x)\eta_j(t) = \boldsymbol{\phi}^\text{T}(x)\boldsymbol{\eta}(t)
\label{galerkin_eq6}$$ In Eq. (\[galerkin\_eq6\]), $\boldsymbol{\phi}(x)=\left[\phi_1(x), \,\ \phi_2(x),...,\phi_N(x)\right]^\text{T}$ and $\boldsymbol{\eta}(t)=\left[\eta_1(t), \,\ \eta_2(t),...,\eta_N(t)\right]^\text{T}$. Here, $\phi_j(x)$ are the mass-normalized mode-shapes of the string and $\eta_j(t)$ are the modal coordinates. In this work, $\phi_j(x)$ are chosen to be $\sqrt{2} \sin(j\pi x)$. On substituting Eq. (\[galerkin\_eq6\]) in Eq. (\[galerkin\_eq4\]), pre-multiplying by $\boldsymbol{\phi}(x)$, integrating over the domain $\left[0 ~~1\right]$, and simplifying using orthogonality conditions, the following set of coupled ODEs are obtained: $$\ddot{\eta_j}(t)+ \left(1 + \gamma \sum_{k=1}^{N} \pi^2 k^2 \eta_k(t)^2\right) \omega_j^2 \eta_j(t)+c\dot{\eta}_j(t)=0
\label{galerkin_eq7}$$ Here, $\omega_j=j\pi$ are the natural frequencies of the string. Equation (\[galerkin\_eq7\]) can now be written in matrix form as $$\bar{\boldsymbol{M}}\ddot{\boldsymbol{\eta}}(t)+\bar{\boldsymbol{K}}\left(\eta\left(t\right)\right)\boldsymbol{\eta}(t)+\bar{\boldsymbol{C}}\dot{\boldsymbol{\eta}}(t)=0,
\label{galerkin_eq8}$$ where $\bar{\boldsymbol{M}}$ is an identity matrix and $\bar{\boldsymbol{K}}(\eta(t))$ is a diagonal matrix with the diagonal elements $\bar{K}_{jj}=\left(1 + \gamma \sum_{k=1}^{N} \pi^2 k^2 \eta_k(t)^2\right) \omega_j^2$ and $\bar{\boldsymbol{C}}=c\bar{\boldsymbol{M}}$. Here, Eq. (\[galerkin\_eq8\]) is in modal coordinates and it is difficult to implement impact conditions in the modal system. Therefore, we discretize the string into a set of physical coordinates (see Fig. \[1d\_grid\]) and introduce the following notation: $$y(x_i,t)=p_i(t)
\label{galerkin_eq9}$$
![Schematic of a spatially discretized string.[]{data-label="1d_grid"}](string_1D_grid.eps){width="45.00000%"}
Using Eq. (\[galerkin\_eq9\]), Eq. (\[galerkin\_eq6\]) can be written as $$p_i(t)=\boldsymbol{\phi}^\text{T}(x_i)\boldsymbol{\eta}(t)
\label{galerkin_eq10}$$ Equation (\[galerkin\_eq10\]) can be written in matrix form as follows: $$\boldsymbol{p}(t)=\boldsymbol{\Phi}^\text{T}\boldsymbol{\eta}(t),
\label{galerkin_eq11}$$ where $\boldsymbol{p}(t)=\left[p_1(t),p_2(t),...,p_N(t)\right]^\text{T}$ and $\boldsymbol{\Phi}=\left[\boldsymbol{\phi}(x_1),\boldsymbol{\phi}(x_2),...,\boldsymbol{\phi}(x_N)\right]^\text{T}$. Now, Eq. (\[galerkin\_eq11\]) can be rewritten as $$\boldsymbol{\eta}(t)=\left(\boldsymbol{\Phi}^\text{T}\right)^{-1}\boldsymbol{p}(t)
\label{galerkin_eq12}$$ Upon substituting Eq. (\[galerkin\_eq12\]) into Eq. (\[galerkin\_eq8\]) and pre-multiplying by $\boldsymbol{\Phi}^\text{T}$, Eq. (\[galerkin\_eq8\]) can be written in terms of the physical coordinates as follows: $${\boldsymbol{M}}\ddot{\boldsymbol{p}}(t)+{\boldsymbol{K}}\left(\boldsymbol{p}\left(t\right)\right)\boldsymbol{p}(t){+\boldsymbol{C}}\dot{\boldsymbol{p}}(t)=0,
\label{galerkin_eq14}$$ where ${\boldsymbol{M}}=\boldsymbol{\Phi}^\text{T}\bar{\boldsymbol{M}}\left(\boldsymbol{\Phi}^\text{T}\right)^{-1}$, ${\boldsymbol{K}}\left(\boldsymbol{p}\left(t\right)\right)=\boldsymbol{\Phi}^\text{T}\bar{\boldsymbol{K}}\left(\left(\boldsymbol{\Phi}^\text{T}\right)^{-1}\boldsymbol{p}\left(t\right)\right)\left(\boldsymbol{\Phi}^\text{T}\right)^{-1}$ and ${\boldsymbol{C}}=\boldsymbol{\Phi}^\text{T}\bar{\boldsymbol{C}}\left(\boldsymbol{\Phi}^\text{T}\right)^{-1}$. Now, by applying Ivanov’s transformation as discussed in Sec. \[mainder\], to Eq. (\[galerkin\_eq14\]), the motion of the string impacting a distributed rigid obstacle can be obtained.
Results and discussion {#results}
======================
We now validate the proposed approach by simulating the vibro-impact motion of a string (see Fig. \[Figure7\]) against a distributed obstacle using the Galerkin–Ivanov transformation. The analysis is performed by solving Eq. (\[galerkin\_eq14\]) subjected to the impact constraint using Ivanov’s method and penalty method. For the analysis, the non-dimensional parameter $\gamma$ is chosen to be $1$ and $c$ is chosen to be $0$ unless otherwise specified. The validation is performed by comparing the solutions obtained from Ivanov’s method with the results from the penalty approach. The obstacle in the penalty method is modeled as a foundation of unilateral linear springs, each having a spring constant of $k_p=1\times10^{8}$. All numerical simulations have been carried out in MATLAB using a fixed-time-step-size, fourth-order Runge–Kutta integrator with $\Delta t = 1 \times 10^{-4}$. The penalty stiffness for the string and the integration time-step size have been selected to get a good match between the responses from Ivanov’s method and penalty method. It should be noted that Ivanov’s method is a limiting case of the penalty approach (with infinite penalty stiffness) and it is to be expected that the results from these methods will match only for large values of penalty stiffness. To demonstrate the reliability of the proposed modal-to-physical transformation, we plotted the first $200$ natural frequencies of the linearized system before and after the transformation in Fig. \[omegaVsN\]. The red circles in Fig. \[omegaVsN\] represent the natural frequencies of the linearized system in modal coordinates (${\omega}_m$) (Eq. (\[galerkin\_eq8\])) and the blue line represents the frequencies in physical coordinates (${\omega}_s$) (Eq. (\[galerkin\_eq14\])). This clearly demonstrates the accuracy of the transformation in conserving all the natural frequencies of the system exactly.
![Natural frequencies of a string in modal coordinates ${\omega}_m$ (Eq. (\[galerkin\_eq8\])) and in physical coordinates ${\omega}_s$ (Eq. (\[galerkin\_eq14\])).[]{data-label="omegaVsN"}](omegaVsN.eps){width="48.00000%"}
We now discuss the results for the string impact problem. In the present analysis, we considered $N=201$ modes/grid-points in the Galerkin approximation,for which an accurate match is obtained between Ivanov’s method and the penalty approach. As a first example, we simulate the vibro-impacting motion of a string against a flat obstacle. The considered obstacle is defined as $d(x)=0.025, ~\frac{1}{3}\le x\le \frac{2}{3}$. The initial displacement of the string is considered to be $0.05\times\sin(\pi x)$, and the initial velocity is considered to be zero. On following Ivanov’s approach (discussed in Sec. \[mainder\]), the solution for the problem is obtained.
In Fig. \[Figure8a\], we have shown the shape of the string obtained from numerical simulations at different time instants for $R=1$ (CoR) obtained using both the Ivanov and penalty approaches. A clear match between the results obtained from both the methods has been observed in Fig. \[Figure8a\]. Furthermore, we have also implemented the Ivanov approach using an explicit numerical integrator for different time-step sizes ($\Delta t=~0.0025,~ 0.0013,~ 0.0008$, and $0.0001$). The MSE ($e(t)$) between the solutions obtained from Ivanov and penalty approaches with respect to time has been evaluated and shown in Fig. \[Figure8b\]. The solution from the penalty approach has been determined for the penalty spring stiffness of $k_p=10^8$ and is used to compare the Ivanov solutions. From Fig. \[Figure8b\], it is clear that the error in the Ivanov solution reduces with decrease in the integration time-step size and converges to the penalty solution. It is also observed that the initial MSE between the solutions is of an order less than $10^{-5}$ and increases gradually with respect to time. This happens because of the finite penalty stiffness, due to which the duration of contact is also finite. When the simulation is run for a long time, these accumulated contact times will introduce a phase shift in the solution when compared to Ivanov’s solution. Moreover, for simulating a rigid collision in penalty approach, one has to use a large value for the penalty stiffness, which puts a hard restriction on the numerical integrator time-step size. However, in Ivanov’s approach, the nonsmooth transformation simulates the case of infinite penalty stiffness, but without making the equations stiff. Therefore, the comparison between the Ivanov and penalty methods will be valid only for small simulation times (few impacts). In addition, the unilateral constraints are exactly satisfied in Ivanov’s approach, unlike the penalty method. We can only compare results obtained from Ivanov’s method for the case of $R=1$ with those obtained from the penalty approach. This is because there are no models that relate $R$ with contact dissipation in the penalty approach for discretized continuous systems.
In Fig. \[Figure9\], we show the midpoint deflection of the string for three values of CoR ($R= 1.0$, $0.9$, and $0.8$) for $c=0$. Again, we can see a good agreement between the solutions for the penalty method (dashed blue line) and Ivanov’s method for $R=1$ (solid red line). It is also observed from Fig. \[Figure9\] that the rebounding displacement decreases for a smaller $R$ because of the loss of energy with every impact for $R<1$. We also present the midpoint displacement of the string impacting a flat obstacle for different damping values ($c=0,~0.1$, and $0.2$) in Fig. \[Figure12a\]. The corresponding penalty solutions have also been shown in Fig. \[Figure12b\] with a dashed blue line and an accurate match between both the solutions has been observed. In addition, the MSE between the Ivanov and penalty solutions for different integration time-step sizes ($\Delta t=0.0025,~0.0013,~0.0008$ and $0.0001$) for $c=0.2$ and $R=1$ has been presented in Fig. \[Figure12b\]. It is clearly observed from Fig. \[Figure12b\] that the MSE between both solutions is of the order less than $1.5\times 10^{-5}$ for all $\Delta t$. Also the Ivanov solution seems to converge to the penalty solution for decreasing $\Delta t$.
For the second example, we consider the motion of a string impacting a sinusoidal obstacle of the form $d(x)=0.05-0.025\times\sin(\pi (x-\frac{1}{3})), ~\frac{1}{3}\le x\le \frac{4}{3}$. The initial displacement of the string is considered to be $0.05 \times \sin(\pi x)$, and the initial velocity is taken to be zero. Figures \[Figure10\], \[Figure11\], and \[Figure13\] are similar to Figs. \[Figure8\], \[Figure9\], and \[Figure12\], respectively, except that the results are presented for the case of a sinusoidal obstacle. It is clearly observed from Figs. \[Figure10\], \[Figure11\] and \[Figure13\] that the solutions from Ivanov’s method and the penalty approach match closely with an MSE of order less than $10^{-5}$.
From the results shown in Figs. \[Figure8\]–\[Figure11\], it is clearly demonstrated that Ivanov’s transformation can be successfully used to simulate the vibro-impacting motions of a continuous structure with distributed contacts using a Galerkin approximation. Even though we have reported results for a string, the approach can also be used to simulate impacts in membranes, beams, and plate structures.
Conclusions {#concl}
===========
In this paper, the application of Ivanov’s method to simulate the nonsmooth motion of a continuous structure against a rigid distributed obstacle using Galerkin approximation is demonstrated. A closed-form expression for the nonlinear MDOF system in Ivanov’s transformed coordinates is reported. The transformed differential equations are without any constraints and automatically satisfy the unilateral and velocity jump conditions at impact. The technique is validated by solving three problems: (i) a spring-mass-damper system, (ii) the motion of a string against a flat obstacle, and (iii) the motion of a string against a sinusoidal obstacle. The governing partial differential equations of motion of the string are approximated using the Galerkin approach. We first convert the ODEs obtained from the Galerkin approximation to the physical coordinates using a modal-physical transformation and then apply Ivanov’s approach. The solutions that are obtained from Ivanov’s method are then compared with the results from the penalty method. The equations of motion obtained from Ivanov’s method are not stiff, in contrast to those obtained from the penalty method. Therefore, a larger time-step size can be used for the numerical integration of the equations of motion obtained from Ivanov’s method. The event detection problem for large MDOF systems in the presence of multiple impact constraints is a difficult and challenging task and is completely eliminated by the application of Ivanov’s method.
Acknowledgments
===============
CPV gratefully acknowledges the Department of Science and Technology for funding this research through Inspire fellowship (DST/INSPIRE/04/2014/000972).
[10]{}
V. I. Babitsky, [*Theory of vibro-impact systems and applications*]{}. Berlin, Germany: Springer-Verlag, 1998.
R. A. Ibrahim, [*Vibro-impact dynamics: modeling, mapping and applications*]{}. Berlin, Heidelberg: Springer-Verlag, 2009.
R. A. Ibrahim, “Recent advances in vibro-impact dynamics and collision of ocean vessels,” [*Journal of Sound and Vibration*]{}, vol. 333, no. 23, pp. 5900–5916, 2014.
K. Vijayan, M. Friswell, H. H. Khodaparast, and S. Adhikari, “Non-linear energy harvesting from coupled impacting beams,” [*International Journal of Mechanical Sciences*]{}, vol. 96, pp. 101–109, 2015.
V. Chatziioannou and M. Van Walstijn, “Energy conserving schemes for the simulation of musical instrument contact dynamics,” [*Journal of Sound and Vibration*]{}, vol. 339, pp. 262–279, 2015.
C. Issanchou, S. Bilbao, J.-L. Le Carrou, C. Touz[é]{}, and O. Doar[é]{}, “A modal-based approach to the nonlinear vibration of strings against a unilateral obstacle: Simulations and experiments in the pointwise case,” [*Journal of Sound and Vibration*]{}, vol. 393, pp. 229–251, 2017.
C. Issanchou, V. Acary, F. P[é]{}rignon, C. Touz[é]{}, and J.-L. Le Carrou, “Nonsmooth contact dynamics for the numerical simulation of collisions in musical string instruments,” [*The Journal of the Acoustical Society of America*]{}, vol. 143, no. 5, pp. 3195–3205, 2018.
J.-H. Shin, E. Talib, and M. K. Kwak, “Vibration of a string against multiple spring-mass-damper stoppers,” [*Journal of Sound and Vibration*]{}, vol. 414, pp. 157–173, 2018.
A. M. Shafei and H. R. Shafei, “Oblique impact of multi-flexible-link systems,” [*Journal of Vibration and Control*]{}, vol. 24, no. 5, pp. 904–923, 2018.
K. Li and A. Darby, “Experiments on the effect of an impact damper on a multiple-degree-of-freedom system,” [*Journal of Vibration and Control*]{}, vol. 12, no. 5, pp. 445–464, 2006.
Y. Yang and X. Wang, “Investigation into the linear velocity response of cantilever beam embedded with impact damper,” [*Journal of Vibration and Control*]{}, vol. 25, no. 7, pp. 1365–1378, 2019.
G. Gilardi and I. Sharf, “Literature survey of contact dynamics modelling,” [*Mechanism and Machine Theory*]{}, vol. 37, no. 10, pp. 1213–1239, 2002.
S. Bilbao and A. Torin, “Numerical modeling and sound synthesis for articulated string/fretboard interactions,” [*Journal of the Audio Engineering Society*]{}, vol. 63, no. 5, pp. 336–347, 2015.
S. Bilbao, A. Torin, and V. Chatziioannou, “Numerical modeling of collisions in musical instruments,” [*Acta Acustica United with Acustica*]{}, vol. 101, no. 1, pp. 155–173, 2015.
C. Issanchou, J.-L. Le Carrou, C. Touz[é]{}, B. Fabre, and O. Doar[é]{}, “String/frets contacts in the electric bass sound: Simulations and experiments,” [*Applied Acoustics*]{}, vol. 129, pp. 217–228, 2018.
J. Kang, “Calculation of [L]{}yapunov exponents in impacted beam on distributed contact,” [*Journal of Sound and Vibration*]{}, vol. 431, pp. 295–303, 2018.
D. Wagg, “Periodic sticking motion in a two-degree-of-freedom impact oscillator,” [*International Journal of Non-Linear Mechanics*]{}, vol. 40, no. 8, pp. 1076–1087, 2005.
X. Li, Z. Yao, and R. Wu, “Modeling and sticking motion analysis of a vibro-impact system in linear ultrasonic motors,” [*International Journal of Mechanical Sciences*]{}, vol. 100, pp. 23–31, 2015.
T. K. Uchida, M. A. Sherman, and S. L. Delp, “Making a meaningful impact: modelling simultaneous frictional collisions in spatial multibody systems,” [*Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences*]{}, vol. 471, no. 2177, p. 20140859, 2015.
A. Liakou, V. Deno[ë]{}l, and E. Detournay, “Fast in-plane dynamics of a beam with unilateral constraints,” [*Journal of Engineering Mechanics*]{}, vol. 143, no. 2, p. 04016116, 2016.
D. Huang, W. Xu, D. Liu, and Q. Han, “Multi-valued responses of a nonlinear vibro-impact system excited by random narrow-band noise,” [*Journal of Vibration and Control*]{}, vol. 22, no. 12, pp. 2907–2920, 2016.
G. F. S. Rebou[ç]{}as, I. F. Santos, and J. J. Thomsen, “Unilateral vibro-impact systems-experimental observations against theoretical predictions based on the coefficient of restitution,” [*Journal of Sound and Vibration*]{}, vol. 440, pp. 346–371, 2019.
V. Acary and B. Brogliato, [*Numerical methods for nonsmooth dynamical systems: applications in mechanics and electronics*]{}. Springer Science & Business Media, 2008.
C. Studer, “Numerics of unilateral contacts and friction,” [*Vol. 47 of Lecture Notes in Applied and Computational Mechanics, Springer Verlag, Berlin Heidelberg, 2009.*]{}, 2009.
V. Acary, “Energy conservation and dissipation properties of time-integration methods for nonsmooth elastodynamics with contact,” [*ZAMM-Journal of Applied Mathematics and Mechanics*]{}, vol. 96, no. 5, pp. 585–603, 2016.
A. Ivanov, “Analytical methods in the theory of vibro-impact systems,” [ *Journal of Applied Mathematics and Mechanics*]{}, vol. 57, no. 2, pp. 221–236, 1993.
A. Ivanov, “Impact oscillations: linear theory of stability and bifurcations,” [*Journal of Sound and Vibration*]{}, vol. 178, no. 3, pp. 361–378, 1994.
I. M. Grace, R. A. Ibrahim, and V. N. Pilipchuk, “Inelastic impact dynamics of ships with one-sided barriers. [Part I]{}: analytical and numerical investigations,” [*Nonlinear Dynamics*]{}, vol. 66, no. 4, pp. 589–607, 2011.
I. M. Grace, R. A. Ibrahim, and V. N. Pilipchuk, “Inelastic impact dynamics of ships with one-sided barriers. [Part II]{}: experimental validation,” [ *Nonlinear Dynamics*]{}, vol. 66, no. 4, pp. 609–623, 2011.
C. P. Vyasarayani, J. McPhee, and S. Birkett, “Modeling impacts between a continuous system and a rigid obstacle using coefficient of restitution,” [*Journal of Applied Mechanics*]{}, vol. 77, no. 2, p. 021008, 2010.
D. Wagg and S. Bishop, “Application of non-smooth modelling techniques to the dynamics of a flexible impacting beam,” [*Journal of Sound and Vibration*]{}, vol. 256, no. 5, pp. 803–820, 2002.
H. Wei, G. Li, P. Guo, and J. Zhao, “Effect of method type on the response of continuum vibro-impact,” [*Shock and Vibration*]{}, vol. 2019, 2019.
M. Van Walstijn and J. Bridges, “Simulation of distributed contact in string instruments: a modal expansion approach,” in [*2016 24th European Signal Processing Conference (EUSIPCO)*]{}, pp. 1023–1027, IEEE, 2016.
D. W. Oplinger, “Frequency response of a nonlinear stretched string,” [ *The Journal of the Acoustical Society of America*]{}, vol. 32, no. 12, pp. 1529–1538, 1960.
[^1]: Address all correspondence to this author. E-mail address: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper presents the analysis of [[*Chandra*]{}]{} X-ray snapshot observations of a subsample of the extragalactic sources listed in the revised Third Cambridge radio catalog (3CR), previously lacking X-ray observations and thus observed during [[*Chandra*]{}]{} Cycle 15. This data set extends the current [[*Chandra*]{}]{} coverage of the 3CR extragalactic catalog up to redshift $z$=1.0. Our sample includes 22 sources consisting of one compact steep spectrum (CSS) source, three quasars (QSOs), and 18 FRII radio galaxies. [As in our previous analyses, here we report the X-ray detections of radio cores and extended structures (i.e., knots, hotspots and lobes) for all sources in the selected sample]{}. We measured their X-ray intensities in three energy ranges: soft (0.5–1 keV), medium (1–2 keV) and hard (2-7 keV) and we also performed standard X-ray spectral analysis for brighter nuclei. [All radio nuclei in our sample have an X-ray counterpart. We also discovered X-ray emission associated with the eastern knot of 3CR154, with radio hotspots in 3CR41, 3CR54 and 3CR225B and with the southern lobe of 3CR107. Extended X-ray radiation around the nuclei 3CR293.1 and 3CR323 on a scale of few tens kpc was also found.]{} X-ray extended emission, potentially arising from the hot gas in the intergalactic medium [and/or due to the high energy counterpart of lobes, is ]{} detected for 3CR93, 3CR154, 3CR292 and 3CR 323 over a few hundreds kpc-scale. [Finally, this work also presents an update on the state-of-the-art of [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{} observations for the entire 3CR sample.]{}'
author:
- |
F. Massaro, V. Missaglia, C. Stuardi, D. E. Harris, R. P. Kraft, A. Paggi, E. Liuzzo,\
G. R. Tremblay, S. A. Baum, C. P. O’Dea, B. J. Wilkes, J. Kuraszkiewicz & W. R. Forman
title: 'The 3CR Chandra snapshot survey: extragalactic radio sources with 0.5$<z<$1.0'
---
Introduction {#sec:intro}
============
Since the early 60’s, the ensemble of extragalactic sources listed in the Third Cambridge radio catalog (3C) represents one of the most attractive samples to study the physics of radio-loud active galactic nuclei (AGNs). Originally created using radio observations [performed]{} at 159 MHz [@edge59], and subsequently at 178 MHz [@bennett62], the 3C catalog went through two main revisions later in the 80’s [see e.g. @laing83; @spinrad85].
[Since then, a vast suite of observations became available from the radio to optical wavelengths thus enriching the multifrequency database necessary to carry out broad band analyses. Radio images with arcsecond resolution for almost all 3CR extragalactic sources are already present in the NRAO Very Large Array (VLA) Archive Survey (NVAS)[^1] and in the MERLIN[^2] archives [see e.g., @giovannini05].]{} At higher frequencies, Spitzer [see e.g., @werner12; @dicken14], in particular for high redshift sources [see also @haas08; @leipski10], and Hubble Space Telescope observations cover more than 90% of the 3CR extragalactic catalog [see e.g. @madrid06; @privon08; @tremblay09; @hilbert16]. [In addition near infrared observations are also available for a significant fraction of the 3CR objects]{} [see e.g. @baldi10]. Recently, the Herschel Space Observatory also observed several 3CR sources, mostly focusing on the higher redshift ones, [see e.g. @podigachoski15; @westhues16]. [Moreover, a dedicated spectroscopic campaign was carried out with the Telescopio Nazionale Galileo to provide a detailed optical classification and to study their nuclear emission [@buttiglione09; @buttiglione11]]{}. All these observations make the extragalactic 3CR catalog an ideal sample to investigate AGN [nuclear properties, extended radio structures, such as jet knots, hotspots and lobes, and/or their large-scale environments]{} [see e.g. @ineson13; @chiaberge15].
However, although most of the 3CR extragalactic radio sources were observed thanks to extensive X-ray campaigns carried out with [[*Chandra*]{}]{}, [[*XMM-Newton*]{}]{} and [[*Swift*]{}]{} [see e.g., @hardcastle00; @harvanek01; @hardcastle06; @evans06; @balmaverde12; @wilkes13; @maselli16 and references therein], until [Cycle 9 the [[*Chandra*]{}]{} archive covered only up to $\sim$60% of the 3CR extragalactic sample [see e.g. @massaro15 for a recent review], while the others, such as [[*XMM-Newton*]{}]{}, covered less than 1/3 of the entire catalog. Thus we started our [[*Chandra*]{}]{} snapshot survey to ensure that all 3CR extragalactic sources have at least an exploratory X-ray observation, with an angular resolution similar to those at lower energies, available to the astronomical community.]{} Adopting a step-wise strategy, we [requested]{} observations in narrow redshift, $z$, ranges [resulting in]{} modest proposals each cycle to minimize the impact on the [[*Chandra*]{}]{} schedule. To date all the 3CR sources with $z<$1 have at least a snapshot observation (i.e., less than 20ksec total exposure time) available in the [[*Chandra*]{}]{} archive [@massaro10; @massaro12; @massaro13] [and several of them inspired follow up X-ray observations on interesting objects [see e.g., @3c171; @3c305fol; @3c89 to name a few examples].]{}
Here, we present the analysis of the [[*Chandra*]{}]{} snapshot observations [approved]{} during Cycle 15 including all 3CR radio sources lying between $z$=0.5 and $z$=1 that were previously unobserved by [[*Chandra*]{}]{}.
The paper is organized as follows. The update of the ongoing [[*Chandra*]{}]{} campaign of the 3CR sources is described in § \[sec:history\] together with some details on the current sample. A brief overview of the data reduction procedures are given in § \[sec:obs\] while results are described in § \[sec:results\]. Then, in § \[sec:summary\] we present our summary and conclusions. Finally, in the Appendix, we show X-ray images with radio contours overlaid for all the sources in the current sample (§ \[sec:images\]) together with the updated summary of the [[*Chandra*]{}]{} observations for the entire of 3CR extragalactic catalog (§ \[sec:state\]).
Unless [otherwise stated we adopt cgs units for numerical results and we also assume]{} a flat cosmology with $H_0=69.6$ km s$^{-1}$ Mpc$^{-1}$, $\Omega_{M}=0.286$ and $\Omega_{\Lambda}=0.714$ [@bennett14]. Spectral indices, $\alpha$, are defined by flux density, S$_{\nu}\propto\nu^{-\alpha}$.
State-of-the-art of the 3CR extragalactic Chandra snapshot survey {#sec:history}
=================================================================
The revised 3C extragalactic catalog includes 298 sources [@spinrad85]. We have already [analyzed]{} and published all the data collected to date for the [[*Chandra*]{}]{} observations carried out in Cycles 9, 12 and 13 for a total of 75 sources [@massaro10; @massaro12; @massaro13] and an additional 140 objects, [listed]{} in the [[*Chandra*]{}]{} archive, were also presented adopting the same data reduction procedures [@massaro11; @massaro15]. Several subsets of the 3CR sample have been also observed by other groups [e.g., @wilkes13; @kuraszkiewicz17]. Table \[tab:summary\] summarizes the references for the [[[*Chandra*]{}]{} observations of the 3CR extragalactic sources]{} already analyzed and published as part of this compilation.
In our previous archival analyses we excluded 7 sources, namely 3CR 66A [e.g., @abdo11], 3CR 71 [alias NGC 1068; e.g., @brinkman02], 3CR 84 [alias NGC1275 or Perseus A; e.g., @fabian03], 3CR 186 [@siemiginowska10], 3CR 231 [alias M82; e.g., @griffiths00], 3CR 317 [alias Abell 2052; e.g., @blanton09] and 3CR 348 [alias Hercules A; e.g., @nulsen05], since [each of these source has accumulated an exposure time of greater than 80 ks, and has been discussed extensively in the literature.]{} In addition, we also did not re-analyze: 3CR 236, 3CR 326, 3CR 386 since the PI of these observations is currently carrying out the analysis (M. Birkinshaw, priv. comm.).
The [[*Chandra*]{}]{} archive now includes all the 3CR [extragalactic]{} sources up to $z$=0.5 (i.e., 150 sources), with the only exceptions [being]{} those objects for which spectroscopic observations, performed after the last revision [@spinrad85], reported different redshift estimates, namely: 3CR27, at $z$=0.184, 3CR69 at $z$=0.458 [@hiltner91] and 3CR93 at $z$=0.357, as confirmed by Ho & MinJin (2009). The present analysis extends [the [[*Chandra*]{}]{} ]{} database up to $z$=1.0 including 22 more targets. During these Cycle 15 3CR snapshot observations, we also observed 3CR142.1 and 3CR277 for which the redshift reported in the literature [@hewitt91] updates the earlier [estimates]{} from Spinrad et al. (1985). These two sources, together with 3CR93, belong to the sample of 22 targets analyzed in the present work. [[*Chandra*]{}]{} snapshot observations of 3CR27 and 3CR69 [were proposed]{} and obtained in subsequent observing cycles (Stuardi et al. 2017 in prep.).
Twenty-five of the 298 3CR extragalactic sources are still unidentified, lacking an optical counterpart and/or an optical spectroscopic observation necessary to unveil their nature. We recently observed 21 of these 25 targets with [[*Swift*]{}]{} snapshot observations [discovering]{} X-ray counterparts for eleven of them, but even using optical data [available thanks to the instruments on board of the [[*Swift*]{}]{} satellite]{}, we could not discern and/or confirm their extragalactic nature [see @maselli16 for all details].
A summary table on the state-of-the-art of the [[*Chandra*]{}]{} observations for all the 3CR [extragalactic]{} sources, including detections of extended components: jet knots, hotspots, lobes and X-ray emission from the hot intergalactic medium [present]{} in galaxy groups/clusters, is reported in Appendix § \[sec:state\].
[|lcccll|]{} Program & Cycle & Proposal Number & Number of sources & Redshift range & Reference\
3CR snapshot survey & 9 & 09700745 & 30$^*$ & $z<$0.3 & Massaro et al. (2010)\
XJET$^+$ & — & — & 47 & — & Massaro et al. (2011)\
3CR snapshot survey & 12 & 12700211 & 26 & $z<$0.3 & Massaro et al. (2012)\
3CR snapshot survey & 13 & 13700190 & 19 & 0.3$<z<$0.5 & Massaro et al. (2013)\
Archival project$^+$ & — & — & 93 & — & Massaro et al. (2015)\
3CR snapshot survey & 15 & 15700111 & 22$^\circ$ & 0.5$<z<$1.0 & This work\
\
$^*$ The AO9 sample includes 3CR 346 that was re-observed in Cycle 12 because during Cycle 9 its [[*Chandra*]{}]{} observation was affected by high background\
[see @massaro10 for details].\
$^+$ The redshift ranges for both the archival and the XJET samples are unbounded w.r.t. selection\
(see also [ for more details on the database]{}).\
$^\circ$ The field of view of the 3C255B [[*Chandra*]{}]{} observation also covers the region where 3C255A lies but no X-ray counterpart for its nucleus is detected.
Data reduction and data analysis {#sec:obs}
================================
Data reduction and analysis procedures adopted for all the [[*Chandra*]{}]{} observations presented here were extensively described in our previous papers, thus only a brief overview is reported in the following. We adopted the same procedures since our final aim is to create a uniform database for the entire 3CR extragalactic survey once all the sources listed therein will be observed by [[*Chandra*]{}]{}.
We followed the standard procedure described in the [[*Chandra*]{}]{} Interactive Analysis of Observations (CIAO) threads[^3], to perform our data reduction and we used CIAO v4.7 with the [[*Chandra*]{}]{} Calibration Database (CALDB) version 4.6.2.
X-ray photometry
----------------
We initially generated level 2 event files using the $acis\_process\_events$ task and filtering for grades 0,2,3,4,6. We checked the absence of high background intervals inspecting lightcurves extracted for each data set, but this never occurred. We aligned the X-ray position of each core with that of the radio to perform the astrometric registration [see e.g., @massaro11 for details]. [The source 3CR292 has been observed twice with [[*Chandra*]{}]{}, obsID 16065 and obsID 17488, with exposure times of $\sim$4ksec and $\sim$8ksec, respectively. In this case a merged event file was created using the CIAO routine <span style="font-variant:small-caps;">reproject\_obs</span>, thus reprojecting event files to the reference coordinates of the deeper observation (i.e., obsID 17488).]{}
Then we created flux maps in the X-ray energy ranges: 0.5 – 1 keV (soft), 1 – 2 keV (medium), 2 – 7 keV (hard), taking into account exposure time and effective area. In our procedure, as previously done, we used monochromatic exposure maps set to the nominal energies of 0.75, 1.4, and 4 keV for the soft, medium and hard band, respectively. All flux maps were converted from units of counts/sec/cm$^2$ to cgs units by multiplying each event by the nominal energy of each band. However, we made the necessary correction to recover the observed erg/cm$^2$/s, when performing X-ray photometry [see e.g., @3c17; @3c305 for details].
We measured observed fluxes for all the X-ray detected nuclei and extended components. This was done choosing a region of size and shape appropriate to the observed X-ray emission and matching the radio structure. We also chose two background regions, having the same shape and size so as to avoid X-ray emission from other parts of the source. The flux for each energy band and region was measured using funtools[^4] as in our previous analyses. Uncertainties are computed assuming Poisson statistics (i.e., square root of the number-of-counts) in the source and background regions. X-ray fluxes, not corrected for the Galactic absorption, measured for the cores are reported in Table \[tab:cores\] while those for detected jet knots, hotspots and lobes are given in Table \[tab:features\]. The name of each component (e.g., knot or hotspot) is a combination of one letter indicating the orientation of the radio structure and one number indicating distance from the core in arcseconds.
Since the [[*Chandra*]{}]{} native pixel size for the ACIS instrument is 0.492, the data are undersampled, thus to recover the resolution we regridded our images to 1/2, 1/4, or 1/8 of the native ACIS pixel size. This was dictated by the angular size of each radio source and by the number of counts in each source components. For sources of large angular extent 1/2 or no regridding was adopted [see also @massaro12; @massaro13 for more details].
Finally, we performed a comparison between radio and X-ray images at similar angular resolution to verify if extended structures in radio sources, such as jet knots, hotspots and lobes, have an X-ray counterpart. To achieve this goal, we used radio images retrieved from publicly available websites as that of the National Radio Astronomy Observatory VLA Archive Survey (NVAS) [^5], NASA Extragalactic Database (NED) [^6], the (DRAGN)[^7] website as well as personal websites of our colleagues[^8][^9]. Image parameters for each radio observation used are given in the [figure captions]{} of § \[sec:images\].
X-ray spectral analysis
-----------------------
We performed spectral analysis for the X-ray counterparts of radio cores of those sources having more than 400 counts, to determine their X-ray spectral indices $\alpha_X$, the presence or absence of significant intrinsic absorption, and the role played by mild pileup in artificially hardening the spectrum.
The spectral data were extracted from a 2 aperture, as for photometric measurements, using the CIAO routine <span style="font-variant:small-caps;">specextract</span>, thereby automating the creation of count-weighted response matrices. Background spectra were extracted in nearby circular regions of radius 10 not containing obvious sources. The source spectra were then filtered in energy between 0.5-7 keV and binned to allow a minimum number of 30 counts per bin to ensure the use of the Gaussian statistic. We used the <span style="font-variant:small-caps;">Sherpa</span>[^10] modeling and fitting package to fit our spectra. For each source we adopted two models: (1) a redshifted power-law with Galactic and intrinsic photoelectric absorption components, and (2) the same model with an additional pileup component using the <span style="font-variant:small-caps;">jdpileup</span> *Chandra* CCD pileup model developed by Davis (2001).
Prior to fitting, the Galactic hydrogen column density and the source redshift were fixed at the measured values for each source. When considering the first (1) fitting model, the two main variable parameters - namely the intrinsic absorption (N$_H$(z)) and X-ray photon index $\Gamma$ - were allowed to vary in a first pass fit, but subsequently stepped through a range of possible physical values to explore the parameter space, determine 68% confidence intervals, and quantify the degree to which $N_H(z)$ and $\Gamma$ are degenerate.
We have also explored the possible effect of pileup on our sources adding a <span style="font-variant:small-caps;">jdpileup</span>[^11] component to our baseline model (2). We left the parameters of the <span style="font-variant:small-caps;">jdpileup</span> model fixed to their default values, with the exception of <span style="font-variant:small-caps;">alpha</span> (the probability of a good grade when two photons pile together) and <span style="font-variant:small-caps;">f</span> (the fraction of flux falling into the pileup region). In no case were we able to constrain the value of <span style="font-variant:small-caps;">alpha</span> since it was usually degenerate with <span style="font-variant:small-caps;">f</span> and/or the intrinsic absorption, so we decided to freeze it to its default value of 1. The value of <span style="font-variant:small-caps;">f</span> was left free to vary between 0.85 and 1, and was constrained in two cases. The results of spectral fitting are given in Table \[tab:spectra\].
Results {#sec:results}
=======
General
-------
We detected [the X-ray counterpart of all radio cores]{} in our sample and we measured [their fluxes]{} adopting a circular region of 2$\arcsec$ centered on the radio position used for the astrometric registration. All the results of our X-ray photometry, i.e., nuclear X-ray fluxes in the three different bands (see § \[sec:obs\]) together with their X-ray luminosities, are reported in Table \[tab:cores\] while X-ray images for the whole selected sample are presented in § \[sec:images\]. In Table \[tab:cores\], we [also report the values of the ’extended emission’ parameter,]{} computed as the ratio of the total number of counts in a circular region of radius r=2 circle to that of a circle of radius r=10 both centered on the radio position of each 3CR source (i.e., Ext. Ratio “Extent Ratio”). This ratio is close to unity for unresolved (i.e., point-like) sources since the on-axis encircled energy for 2 is $\simeq$0.97, and only a small increase is expected up to 10. Thus [parameter values]{} significantly less than 0.9 indicate the presence of extended emission around the nuclear component [e.g., @massaro10; @massaro13]. [In our sample this situation clearly occurs]{} for 3CR107, 3CR293.1 and 3CR323.
Four of the 22 sources[./, namely: 3CR93, 3CR114, 3CR154 and 3CR288.1,]{} show X-ray nuclei with more than 400 counts within a circular region of 2 radius. Thus, according to our previous works, we performed a basic X-ray spectral analysis for them. Their spectra are consistent with a mildly absorbed power-law with the absorption consistent with the Galactic value [@kalberla05]. 3CR93 and 3CR154 show a fraction of pileup of 10%, however their fits do not improve significantly when [including]{} a jdpileup [spectral component]{} [^12]. Results of the spectral analysis are reported in Table \[tab:spectra\], the statistical uncertainties quoted refer to the 68% confidence level.
We also discovered X-ray emission associated with three hotspots in three different sources; their [X-ray fluxes are reported]{} in Table \[tab:features\] together with detection significances, all above 3$\sigma$, computed assuming a Poisson distribution for the background. [The X-ray counterpart of a radio knot in the 3CR154 eastern lobe as well as that of the whole southern radio lobe of 3CR107 was also found (see Table \[tab:features\] for details).]{}
Three of 22 3CR sources in our sample [lie within optically known galaxy cluster, namely]{}: 3CR34, 3CR44 and 3CR247 [@spinrad85] and to search for possible X-ray emission due to the [hot gas permeating the intergalactic medium]{} around the radio [structure]{}, we adopted the following procedure for all our targets. We measured the total number of counts in a circular region with diameter equal to the total extent of the radio source and we subtracted those counts within circular regions of 2 corresponding to the radio nucleus and/or to background sources lying within the larger circle. Then assuming a Poisson distribution for the background events, we computed the probability of obtaining the measured value given the expected number of counts in the background.
[We detected an excess of X-ray photons, above 3$\sigma$ significance, around the radio structures of 3CR93, 3CR154 and 3CR323, on a scale of few hundreds kpc as shown in Fig. \[fig:clusters\]. Such X-ray extended emission could be due to the presence of hot gas filling their large-scale X-ray environment (see Fig. \[fig:clusters\]) but it could also be contaminated by the radiation arising from the lobes of these FRII radio galaxies which would be aligned with their large-scale radio structure. These scenarios are indistinguishable due to the low number of counts available, however we tend to favor the former (i.e. presence of hot gas in the intergalactic medium) since the peaks of the X-ray surface brightness do not appear coincident with those of the radio intensity (i.e., hotspots in the lobes). We also found a marginal detection (i.e. 2$\sigma$ significance) of X-ray emission from the group/cluster around 3CR34 (see also Fig. \[fig:clusters\]). For this source we do not claim that the extended X-ray emission could be associated with the radio lobes because it lies in a galaxy-rich large scale environment [@mccarthy95]. Then we discovered extended X-ray emission also for 3CR292 on a hundreds of kpc scale (see Fig. \[fig:3c292cluster\]) above 3$\sigma$ significance. Here an excess of X-ray photons is also found associated with the northern hotspot/lobe but it is consistent with a 3$\sigma$ fluctuation of the X-ray diffuse background (more details are given in the following section). A specific analysis to search for galaxy groups or cluster signatures surrounding the 3CR sources observed during the [[*Chandra*]{}]{} snapshot survey is out of the scope of the paper and it will be presented in a forthcoming work.]{}
{height="6.5cm" width="9cm"} {height="6.5cm" width="9cm"} {height="6.5cm" width="9cm"} {height="6.5cm" width="9cm"}
![[The [[*Chandra*]{}]{} X-ray image for 3CR292 in the energy range 0.5–7 keV. The merged event file has been created combining both observations available in the archive (obsID 16065 and obsID 17488, $\sim$4ksec and $\sim$8ksec exposure time, respectively). X-ray image has been smoothed with a Gaussian function of 8 pixels kernel radius. The five radio contour levels, black at 1.4 GHz and cyan at 8.5 GHz, overlaid on the [[*Chandra*]{}]{} image were computed starting at 8 mJy/beam and 0.4, increasing by a factor of 2 and 4, respectively. The radio map at 1.4 GHz is not registered because the radio core is undetected.]{}[]{data-label="fig:3c292cluster"}](./3c292cluster.pdf){height="5.5cm" width="8.5cm"}
Finally, we remark that X-ray images with radio contours overlaid for all the sources in the current sample are shown in the Appendix (§ \[sec:images\]).
Source details
--------------
*3CR34*. This is an FRII radio galaxy at $z$=0.69 and optically classified as a high-excitation line radio galaxy [HERG; @mullin06]. It lies near the center of a compact cluster of galaxies [@mccarthy95] appearing as one of the reddest members and being surrounded by fainter companions. There is a double hotspot in the western lobe. Best et al. (1997) reported the detection of a strong jet-cloud interaction at 120 kpc distance from the core and HST images also show a narrow region of blue emission orientated along the radio axis and directed towards a radio hotspot. In the [[*Chandra*]{}]{} observation analyzed here we only detected the X-ray core. We tested the presence of diffuse X-ray emission due to the intergalactic medium of the known galaxy cluster where 3C 34 lies [according to the procedure previously described]{}. We found an X-ray excess of about 3$\sigma$ significance (see Fig. \[fig:clusters\]) where the number of counts in an annular region of radii 2and 30, centered on the radio core position, is 90, [compared with 45.7 expected background counts in that region]{}. [We also estimated the luminosity of the X-ray extended emission, $L_X$, adopting an annular region of inner radius 2and outer radius 30, centered on the position of the 3CR34 radio nucleus, to exclude the nuclear contribution of the central radio galaxy. Our estimate is $L_X=$(5.2$\pm$2.2)$\cdot$10$^{43}$erg/s. We note that this $L_X$ estimate, as in the following cases of 3CR93, 3CR154, 3CR292 and 3CR323, is an upper limit on the X-ray luminosity of the hot gas in the intergalactic medium since there could be a possible contamination due to the lobe X-ray emission.]{}
*3CR41*. This is an FRII radio galaxy with $z=$0.79, classified as a narrow emission line radio galaxy. In the radio image at 8.45 GHz there is no detection of a jet in the northern lobe while extended radio emission associated with the jet pointing towards the south-eastern hotspot is detected [@mullin08 see e.g.]. Both radio lobes [are also visible even]{} at 8.45 GHz. At higher energies, in our [[*Chandra*]{}]{} snapshot observations, the southern hotspot is detected [as well as the core of the radio galaxy]{}.
*3CR44*. This quasar, with a radio morphology [similar to that of an FRII radio galaxy]{}, is at $z$=0.66 [being]{} associated with a galaxy cluster visible in the optical image[@odea09]. Kharb et al. (2008) reported a hint of a jet-like structure extending towards the southern hotspot in the 5GHz [radio map]{} while the HST image shows this radio galaxy to be either composed of two [structures]{} oriented north-south or, more likely, bisected by a dust lane running east-west [@mccarthy97]. We clearly detected the nuclear emission in the X-rays and we also found an excess of X-ray counts associated with the northern lobe but at $<$3$\sigma$ significance.
*3CR54*. This is an FRII radio galaxy at redshift $z$=0.8274. The lobe morphology on the southern side of the source is more extended than the northern one [@kharb08]. In the HST image, the galaxy has close companions and the source appears extended towards the southwest side but it is unclear if it is a bridge, a tidal tail, or a jet feature [@mccarthy97]. In the X-ray image the radio core is clearly detected together with the southern hotspot.
*3CR55*. This is an FRII radio galaxy at $z$=0.735 and [optically classified]{} as a narrow emission line radio galaxy. [The radio core is detected in the X-rays while the two hotspots are not.]{}
*3CR93*. This [is a $z$=0.358 quasar with a lobe-dominated radio]{} morphology [see e.g. @bogers94]. In the optical image, [available in the HST archive,]{} 3CR93 has a host galaxy with $\sim$3 diameter [@lehnert99]. In the X-ray image the core is clearly detected with more than 1000 counts but there is no detection of hotspots. X-ray spectral analysis of the nuclear emission shows a power law spectrum with absorption consistent with the Galactic column density. Significant extended X-ray emission on hundreds of kpc scale was found around 3CR93 with a significance above 3$\sigma$ (see Fig. \[fig:clusters\]). The number of counts in an annular region of radii 2and 30, centered on the radio core position, is 114 while those expected for the same region in the background is 45.4. [The X-ray luminosity of the extended emission, $L_X$, estimated adopting an annular region of inner radius 2and outer radius 30, centered on the location of 3CR93, as previously done for 3CR34, is $L_X=$(2.3$\pm$0.5)$\cdot$10$^{43}$erg/s.]{}
*3CR107*. [This FRII - HERG radio galaxy lies at $z$=0.785 [@mccarthy97]. In our [[*Chandra*]{}]{} snapshot observation we clearly detected the nuclear emission as well as X-ray extended emission associated with the southern lobe (see Fig. \[fig:3c107lobe\] and Table \[tab:features\]), as also indicated by the low value of the “ext ratio” in Table \[tab:cores\].]{}
![[The [[*Chandra*]{}]{} X-ray image for 3CR107, centered on its southern lobe, in the energy range 0.5–7 keV. The image has been smoothed with a Gaussian function of 5 pixels kernel radius. Level radio contours (white) overlaid on the [[*Chandra*]{}]{} image were computed starting at 0.2 mJy/beam, increasing by a factor of 2 and drawn using the same 4.9 GHz radio map adopted for the registration. The X ray flux of the southern lobe was measured using the yellow elliptical region drawn in the figure. Here there is spatial coincidence between the excess of X-ray photons and the lobe radio structure.]{}[]{data-label="fig:3c107lobe"}](./3c107lobe.pdf){height="6.5cm" width="9.cm"}
*3CR114*. This FRII radio galaxy is at $z$=0.815 showing fairly weak emission lines in its optical spectrum that led to a LERG classification [@strom90]. A faint compact nucleus with several clumps within the few arcseconds is observed at radio frequencies and a jet-like feature appears on the northern side in the 1.4GHz image [@kharb08] while single hotspots are detected both in the southern and in the northern lobes. Storm et al. (1990) claimed that this radio galaxy could lie in the core of a rich, distant cluster with some signatures of a merger on the basis of their optical observations. In the X-ray image the nucleus is detected but there is no evidence of hotspots and no signatures of X-ray emission from intracluster medium was found.
*3CR142.1*. This double radio source (i.e., FRII morphology) at $z=$0.4061. In the radio there are two clear extended radio lobes with a radio bridge showing a constant spectral index [@kharb08; @odea09]. In the X-ray image only the radio nucleus is detected.
*3CR154*. This is a nearby [lobe-dominated quasar]{} [@bogers94] at $z$=0.58 [@sokolovsky11] [appearing as a point-like]{} source in the HST optical image. In [the [[*Chandra*]{}]{} snapshot observation only]{} the relatively [bright radio core]{} is detected, for which [X-ray spectral analysis was performed]{}. As in 3CR34 and 3CR93, 3CR154 shows extended X-ray emission on kpc scale detected above 5$\sigma$ level of confidence, [measured adopting the same method previously described (see Fig. \[fig:clusters\])]{}. We measured 243 X-ray photons in a annular region of radii 2and 40, centered on the location of the radio core, while 70.1 are expected in the background region. [This X-ray extended emission has a luminosity $L_X=$(2.2$\pm$0.3)$\cdot$10$^{44}$erg/s, measured using an annular region of inner radius 2and outer radius 40, centered on the location of radio nucleus of 3CR154. In the eastern lobe we found an association between the peak of the radio and the X-ray intensities for a radio knot lying at 33 from the nucleus (see marked region in Fig. \[fig:3c154knot\]). Given this spatial coincidence, this excess of X-ray photons is probably due to the high energy counterpart of the radio extended structure. On the other hand, the southern knot in the same eastern lobe does not show a correspondence between the radio and the X-ray emissions ($\sim$1.5 offset), indicating that this high energy emission could be linked to the hot gas in the intergalactic medium.]{}
![[The [[*Chandra*]{}]{} X-ray image for 3CR154, centered on its eastern lobe, in the energy range 0.5–7 keV. The image has been smoothed with a Gaussian function of 8 pixels kernel radius. Radio contour levels, overlaid on the [[*Chandra*]{}]{} image, were computed using radio maps at 1.4 and at 8.46 GHz, starting at 1 and 2 mJy/beam and increasing by a factor of 2 and 4, respectively. A circular region of 4 radius, marked with a red dashed line, was used to measure the X-ray flux of the eastern knot. The southern radio knot, in the eastern lobe, has an offset of 1.7 between the peak of the radio surface brightness and the X-ray one. This second excess of X-ray photons could be due to diffuse hot gas in the large scale environment.]{}[]{data-label="fig:3c154knot"}](./3c154knot.pdf){height="5.5cm" width="9cm"}
*3CR169.1*. This classical FRII radio galaxy is at $z$=0.633. There is no detection of the jet [and/or of the hotspots in the radio map at]{} 8 GHz [@kharb08]. Harvanek et al. (1998) classified this source as a narrow emission line radio galaxy according to its optical spectrum. In the X-ray snapshot image the nucleus is [detected but there is no detection of the extended radio structure]{}.
*3CR217*. This is a FRII radio galaxy with narrow emission lines at $z$=0.898. [In the radio image at]{} 8 GHz there are no jet signatures, in either the eastern or the western lobe [@mullin08] and only the western lobe is detected. In the X-ray band the radio core is detected but the [[*Chandra*]{}]{} image is not registered because [we were not able to locate precisely its position]{}. Only a single photon is associated with the western hotspot [being undetected in the X-rays.]{}
*3CR225B*. This is a FRII radio galaxy at $z$=0.582, optically classified as a narrow emission line galaxy. It does show only hotspots within its radio lobes [@mullin08]. In the X-ray there is a hint of diffuse emission near the nucleus but no detection for the two hotspots. We note that the nearby radio source 3CR225A is also in the field of view of the [[*Chandra*]{}]{} snapshot observation on a nearby CCD, but we did not detect any signature of X-ray emission arising from this radio object.
*3CR237*.This is the only compact steep spectrum (CSS) radio source in our sample. It shows an FRII radio morphology and is at $z$=0.877. The radio source size is less than 2 [(i.e., less than $\sim$15 kpc at $z=$0.877)]{} and it has a clear X-ray counterpart.
*3CR247*. This is an FRII radio galaxy at $z$=0.75, [showing two hotspots on both the eastern and the western side]{}, optically classified as a narrow emission line radio galaxy. The host galaxy associated with the radio source 3C247 lies in a very crowded field. [In the HST optical image it appears as a symmetrical central galaxy]{} with a close companion lying about 0.8 arcsec to the south. Two nearby galaxies also lie within the envelope of the infrared emission [see e.g. @best98; @best00]. Radio spectral ageing analysis of this source has yielded an age of 3-5 million years, corresponding to a hotspot advance velocity of about 0.1c [@liu92]. In the [[*Chandra*]{}]{} X-ray image the radio core and the northern hotspot are detected.
*3CR272*. A classical double FRII radio source at $z$=0.944, showing a relatively faint nuclear component with respect to the two hotspots at radio frequencies. [Its optical spectrum shows a typical HERG spectrum]{} with high ionization emission lines [@strom90]. There is nothing to report in the [[*Chandra*]{}]{} snapshot image other than the core detection.
*3CR277*. This is a giant FRII radio galaxy at $z$=0.414 [@strom90], optically classified as a low excitation type radio galaxy (i.e., LERG). At 1.4 GHz the core and a jet-like structure are detected in the eastern lobe. Both radio lobes show double hotspots in high resolution radio images [see e.g., @harvanek98]. According to McCarthy et al. (1997) the source reveals a basic double morphology with the nuclear component slightly offset from the geometrical center. In the X-ray image the [radio core]{} is detected while [while neither double hotspot has an X-ray counterpart]{}.
*3C277.2*. This FRII radio galaxy at $z$=0.767 is optically classified as a narrow line emission galaxy. It has a clear jet-like component in the southern lobe that is closer to the radio core but no signatures in the X-rays.
*3CR288.1*. This is a lobe-dominated QSO lying at $z$=0.961 having a clear point-like optical counterpart [@gendre08]. In the X-ray image the nucleus is clearly detected and given the high number of photons with respect to the majority of the other targets, we also performed its spectral analysis finding its X-ray spectrum consistent with an absorbed power-law (see Table \[tab:spectra\]). Extended emission on kpc scale seems to be present around the core (see Table \[tab:cores\]) but no evidence of a galaxy cluster or a group of galaxies is present in the optical images.
[*3CR292*. A narrow line FRII radio galaxy at redshift 0.713. Mullin et al. (2006) reported the detection of the two hotspots and the core in the radio map at 8.45 GHz but a part from the core none of these features is detected in the X-rays. Adopting the same procedure as in previous cases we found that 3CR292 shows extended X-ray emission (more than 3$\sigma$ detection significance) detected on kpc scale and potentially due to the presence of hot gas in the intergalactic medium (see Fig. \[fig:clusters\]). The number of X-ray photons measured in a annular region of radii 2and 75, centered on the radio core position, is 525 while those expected in the background, for the same area, is 364.3, where both X-ray counts have been measured using the merged event file. We estimated an X-ray luminosity for the extended emission of $L_X=$(2.5$\pm$0.7)$\cdot$10$^{44}$erg/s, using the same annular region. Belsole et al. (2004) reported the X-ray detection of radio lobes in an [[*XMM-Newton*]{}]{} observation and as previously stated in the merged event file we clearly detected an excess of X-ray photons correspondent to the location of the northern hotspot. However, as shown in Fig. \[fig:3c292lobe\], in a circular region of 3.5 radius, centered on the northern hotspot, we measured 5 X-ray counts in the $\sim$4ksec [[*Chandra*]{}]{} observation (i.e. obsID 16065) but only 1 in the deeper one (i.e. obsID 17488, $\sim$8ksec exposure time). Then, in the latter observation, the peak of the radio surface brightness is not coincident with the X-ray excess. Since this radio structure lies in a large scale environment permeated by X-ray emission from the hot intergalactic medium, we measured the level fo the X-ray background of the $\sim$4ksec observation within 75 from the radio core and excluding the nuclear X-ray emission. We found that, over the same area (i.e., circle of 3.5 radius), the expected number of counts is 1.5$\pm$1.2 (obsID 16065), thus this X-ray excess is consistent with a 3$\sigma$ fluctuation of the diffuse X-ray background. A deeper investigation will be necessary to distinguish between X-ray emission associated with the hotspot or that of the intergalactic medium. Variability on monthly time scale has been excluded because the hotspot size, even if being relatively compact with respect to the lobe structure, covers a region of $\sim$25 kpc.]{}
![[The [[*Chandra*]{}]{} X-ray images for 3CR292, centered on its northern lobe, in the energy range 0.5–7 keV. The left image corresponds to the [[*Chandra*]{}]{} observation with shorter exposure time (obsID 16065, $\sim$4ksec) while the right panel to the deeper one $(obsID 17488 \sim$8ksec of exposure time). Both images have been smoothed with a Gaussian function of 8 pixels kernel radius. The five radio contours levels, yellow at 1.4 GHz and white at 8.5 GHz, overlaid on the [[*Chandra*]{}]{} image were computed starting at 8 and 0.4mJy/beam, increasing by a factor of 2 and 4, respectively. An elliptical region, marked with a red dashed line, was used to compute the X-ray flux of the northern lobe and it is also shown.]{}[]{data-label="fig:3c292lobe"}](./3c292lobe.pdf){height="6.5cm" width="9.cm"}
*3CR293.1*. This faint FRII radio galaxy is at $z$=0.709. [Its radio structures]{}, i.e, the nucleus and both hotspots, are barely detected in the radio image available to us at 4.9 GHz and the [comparison with the X-ray image indicates]{} that the radio core is detected with [possible extended X-ray emission around it (see Table \[tab:cores\])]{}.
*3CR 323*. This is a FRII radio galaxy at $z$=0.679. McCarthy et al. (1997) pointed out that the outer lobes are irregular and rather different in structure, the northern one undergoing a sharp bend to the east and ending in what may be a hotspot. In the X-ray we only detected the nucleus and there is also significant X-ray extended emission surrounding it both on scales of few tens and $\sim$300 kpc (above 5$\sigma$ significance), potentially due to its intergalactic medium (see Fig. \[fig:clusters\]). There are 329 X-ray photons in a annular region of radii 2and 75, centered on the radio core position, while 269.3 background photons are expected in this region. In this case the number of X-ray counts within a circular region of 2 positioned on two point-like sources lying on the western side of 3CR323 were also subtracted from the X-ray photons measured within the annular region. This does not affect the significance of the detection of the X-ray extended emission since only 8 and 20 photons were measured for the two nearby objects. [We estimated an X-ray luminosity of $L_X=$(4.7$\pm$1.6)$\cdot$10$^{43}$erg/s for the extended emission surrounding 3CR323, this was measured over an annular region of inner radius 2and outer radius 21, centered on the location of its radio core. A lower value of the outer radius for the annular region has been chosen to avoid contamination due to the presence of nearby foreground/background X-ray point-like sources.]{}
Summary and Conclusions {#sec:summary}
=======================
This paper presents the analysis of [[*Chandra*]{}]{} snapshot observations of a subsample of the extragalactic sources listed in the revised Third Cambridge radio catalog (3CR), previously lacking [[*Chandra*]{}]{} observations and observed during Cycle 15. This data set extends the current [[*Chandra*]{}]{} coverage of the 3CR extragalactic catalog up to redshift $z$=1.0. [The 3CR extragalactic sample]{} includes 22 sources listing one compact steep spectrum (CSS) source and three quasars (QSOs) while all the remaining sources are FRII radio galaxies. Nineteen targets lie at $z$ in the range 0.5–1.0 plus 3C93, 3CR142.1 and 3CR277. One additional target, 3CR255A, lies in the [[*Chandra*]{}]{} field of view of a nearby source (i.e., 3CR255B) observed during these Cycle 15 observations, but its radio core is not detected in the X-rays.
The main aims of the 3CR [[*Chandra*]{}]{} snapshot survey are: (i) to search for X-ray emission from jet knots, hotspots and lobes, (ii) to study the nuclear emission of their host galaxies and (iii) to [investigate]{} their environments at all scales, aiming to discover new galaxy [groups/clusters via the X-ray emission of the intergalatic medium]{}.
In the present work, the basic source parameters for the newly acquired [[[*Chandra*]{}]{} data are presented. We created fluxmaps for all the X-ray snapshot]{} observations and compared them with radio images to search for the high energy counterparts of extended radio structures (i.e., jet knots, hotspots, lobes). We measured their X-ray intensities in three energy ranges, [namely soft, medium and hard band, for all radio cores and hotspots detected in the X-rays]{}. Then, for the nuclei brighter than 400 X-ray photons, measured in a circular region of 2 radius in the 0.5–7 keV energy range, we also performed X-ray spectral analysis [showing nuclear spectra all consistent with simple power-law model, with an eventual mild intrinsic absorption in a single case.]{}
We found X-ray emission arising from three hotspots in 3CR41, 3CR54 and 3CR225B. We also report the discovery of extended X-ray emission, on tens of kpc scale, around the radio nuclei of 3CR107 and 3CR293.1, [for the former source due to the X-ray counterpart of the southern radio lobe]{}. Three sources in our sample are members of, optically-known galaxy groups/clusters: 3CR41, 3CR44 and 3CR247. In the last two cases we did not detect X-ray emission arising from the intergalactic medium while a marginal detection (i.e., 2$\sigma$) was found for 3CR34. Moreover, we discovered extended X-ray emission on a scale of few hundreds of kpc around the radio structures of 3CR93, 3CR154, 3CR292 and 3CR 323, all above 3$\sigma$ significance. [Then, for 3CR154, we also detected the X-ray counterpart of a knot in the eastern radio lobe at 33 distance from the nucleus.]{}
[Finally, we highlight that a table summarizing the state-of-the-art of the X-ray (i.e., [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{}) observations carried out to date is reported at the end of the present manuscript (see § \[sec:state\]). [[*Chandra*]{}]{} detections are all based on both our current and previous analysis and represent an update with respect to previous works while those regarding [[*XMM-Newton*]{}]{}, shown here for the first time, are only based on literature search.]{}
We thank the anonymous referee for useful comments that led to improvements in the paper. We are grateful to M. Hardcastle and C. C. Cheung for providing several radio images of the 3CR sources while the remaining ones were downloaded from the NVAS[^13] (NRAO VLA Archive Survey), NED[^14] (Nasa Extragalactic Database) and from the DRAGN webpage[^15]. This investigation is supported by the NASA grants GO1-12125A, GO2-13115X, and GO4-15097X. G.R.T. acknowledges support from the National Aeronautics and Space Administration (NASA) through Einstein Postdoctoral Fellowship Award Number PF-150128, issued by the Chandra X-ray Observatory Center, which is operated by the Smithsonian Astrophysical Observatory for and on behalf of NASA under contract NAS8-03060. This work was also supported by contributions of European Union, Valle D’Aosta Region and the Italian Minister for Work and Welfare. The National Radio Astronomy Observatory is operated by Associated Universities, Inc., under contract with the National Science Foundation. This research has made use of data obtained from the High-Energy Astrophysics Science Archive Research Center (HEASARC) provided by NASA’s Goddard Space Flight Center; the SIMBAD database operated at CDS, Strasbourg, France; the NASA/IPAC Extragalactic Database (NED) operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. TOPCAT[^16] [@taylor05] for the preparation and manipulation of the tabular data and the images. SAOImage DS9 were used extensively in this work for the preparation and manipulation of the images. SAOImage DS9 was developed by the Smithsonian Astrophysical Observatory.
[Facilities:]{} , ,
Abdo, A. A., Ackermann, M., Ajello, M. et al. 2011ApJ...726...43 Arnaud, K.A., 1996, “Astronomical Data Analysis Software and Systems V”, eds. Jacoby G. and Barnes J., p17, ASP Conf. Series volume 101 Baldi, R. D.; Chiaberge, M.; Capetti, A.; Sparks, W.; Macchetto, F. D.; O’Dea, C. P.; Axon, D. J.; Baum, S. A.; Quillen, A. C. 2010 ApJ, 725, 2426 Belsole, E.; Worrall, D. M.; Hardcastle, M. J.; Birkinshaw, M.; Lawrence, C. R. 2004 MNRAS, 352, 924 Belsole, E.; Worrall, D. M.; Hardcastle, M. J.; Croston, J. H. 2007 MNRAS, 381, 1109 Balmaverde, B., Capetti, A., Grandi, P., Torresi, E., Chiaberge, M. et al. 2012 A&A, 545A, 143 Best, P. N., Longair, M. S., R$\dot{o}$ttering, H. J. A. 1997 MNRAS, 286, 785 Best, P. N., Longair, M. S., R$\dot{o}$ttering, H. J. A. 1998 MNRAS, 295, 549 Best, P. N. 2000 MNRAS, 317, 720 Bennett, A. S. 1962 MmRAS, 68, 163 Bennett, C. L., et al. 2014 ApJ, 794, 135 Blanton, E. L., Randall, S. W., Douglass, E. M. et al. 2009 ApJ, 697L, 95 Bogers, W. J., Hes, R., Barthel, P. D., Zensus, J. A., 1995 A&A 105, 91 Brinkman, A. C.; Kaastra, J. S.; van der Meer et al. 2002 A&A, 396, 761 Buttiglione, S., Capetti, A., Celotti, A., Axon, D.J., Chiaberge, M., Macchetto, F.D., Sparks, W.B., 2009 A&A 495, 1033 Buttiglione, S.; Capetti, A.; Celotti, A.; Axon, D. J.; Chiaberge, M.; Macchetto, F. D.; Sparks, W. B. 2011 A&A, 525A, 28 Chiaberge, M.; Gilli, R.; Lotz, J. M.; Norman, C. 2015 ApJ, 806, 147 Croston, J. H.; Hardcastle, M. J.; Harris, D. E.; Belsole, E.; Birkinshaw, M.; Worrall, D. M. 2005 ApJ, 626, 733 Croston, J. H.; Hardcastle, M. J.; Birkinshaw, M.; Worrall, D. M.; Laing, R. A. 2008 MNRAS, 386, 1709 Dasadia, S. Sun, M. Morandi, A. et al. 2016 MNRAS, 458, 681 Davis, J. E. 2001 ApJ, 562, 575 2014 ApJ, 788, 98 Dicken, D.; Tadhunter, C.; Morganti, R.; Axon, D.; Robinson, A.; Edge, D. O., Shakeshaft, J. R., McAdam, W. B., Baldwin, J. E.; Archer, S. 1959 MmRAS, 68, 37 Evans, D. A.; Worrall, D. M.; Hardcastle, M. J.; Kraft, R. P.; Birkinshaw, M. 2006 ApJ, 642, 96 Fabian, A. C., Sanders, J. S., Allen, S. W. et al. 2003 MNRAS, 344L, 43 Fanaroff, B. L. & Riley J. M. 1974, MNRAS, 167, P31 Gendre M. A. & Wall, J. V. 2008 MNRAS, 390, 819 Giovannini, G.; Taylor, G. B.; Feretti, L.; Cotton, W. D.; Lara, L.; Venturi, T. 2005, ApJ 618, 635 Griffiths, R. E., Ptak, A., Feigelson, E. D. et al. 2000 Sci, 290, 1325 Grimes, Rawlings & Willott, 2004, MNRAS, 349, 503 Haas, M; Willner, S. P.; Heymann, F.; Ashby, M. L. N.; Fazio, G. G.; Wilkes, B. J.; Chini, R.; Siebenmorgen, R. 2008 ApJ, 688, 122 Hardcastle, M. J.; Worrall, D. M. 2000 MNRAS, 314, 359 Hardcastle, M. J.; Evans, D. A.; Croston, J. H. 2006 MNRAS, 370, 1893 Hardcastle, M. J., Massaro, F., Harris, D. E. 2010 MNRAS, 401, 2697 Hardcastle, M. J., Massaro, F., Harris, D. E. et al. 2012 MNRAS, 424, 1774 Harvanek, M. & Hardcastle, M. J. 1998 ApJS, 119, 1 Harvanek, M.; Ellingson, E.; Stocke, J. T.; Rhee, G. 2001 AJ, 122, 2874 Hes, R.; Barthel, P. D.; Fosbury, R. A. E. 1996 A&A, 313, 423 Hewitt, A.& Burbidge, G. 1991 ApJS, 75, 297 Hilbert, B.; Chiaberge, M.; Kotyla, J. P.; Tremblay, G. R.; Stanghellini, C. et al. 2016 ApJS, 225, 12 Hiltner, P. R. & Roeser, H.-J. 2009 ApJS, 184, 398 Ho, L. C. & Minjin, K. 2009 ApJS, 184, 398 Ineson, J.; Croston, J. H.; Hardcastle, M. J.; Kraft, R. P.; Evans, D. A.; Jarvis, M. 2013 ApJ, 770, 136 Ineson, J.; Croston, J. H.; Hardcastle, M. J.; Mingo, B. 2017 MNRAS, 467, 1586 Kalberla, P.M.W., Burton, W. B., Hartmann, D., et al. 2005, A&A, 440, 775 Kharb, P. , O’Dea, C. P., Baum, S. A., Daly, R. et al. 2008 ApJS, 174, 74 Kuraszkiewicz, J. et al. 2017 ApJ in prep. Laing, R. A., Riley, J. M., Longair, M. S. 1983 MNRAS 204, 151 Laskar, T.; Fabian, A. C.; Blundell, K. M.; Erlund, M. C. 2010 MNRAS, 401, 1500 Lehnert, M. D.; Miley, G. K.; Sparks, W. B.; Baum, S. A.; Biretta, J. et al. 1999 ApJS, 123, 351L Leipski, C.; Haas, M.; Willner, S. P.; Ashby, M. L. N.; Wilkes, B. J.; Fazio, G. G.; Antonucci, R.; Barthel, P.; Chini, R.; Siebenmorgen, R.; Ogle, P.; Heymann, F. 2010 ApJ, 717, 766 Liu, R., Pooley, G. & Riley, J. M. 1992 MNRAS, 257, 545 Madrid, J. P.; Chiaberge, M.; Floyd, D.; Sparks, W. B.; Macchetto, D.; 2006 ApJS, 164, 307 Mannering, E.; Worrall, D. M.; Birkinshaw, M. 2013 MNRAS, 431, 858 Maselli, A.; Massaro, F.; Cusumano, G.; La Parola, V.; Harris, D. E. et al. 2016 MNRAS, 460, 3829 Massaro, F., Chiaberge, M., Grandi, P. et al. 2009 ApJ, 692, L123 Massaro, F., Harris, D. E., Chiaberge M. et al. 2009 ApJ, 696, 980 Massaro, F. et al. 2010 ApJ, 714, 589 Massaro, F., Harris, D. E., Cheung, C. C. 2011 ApJS, 197, 24 Massaro, F. et al. 2012 ApJS, 203, 31 Massaro, F. et al. 2013 ApJS, 203, 31 Massaro, F., D. E., Harris, E., Liuzzo, M., Orienti, R., Paladino et al. 2015 ApJS 220, 5 McCarthy, P. J.; Spinrad, H.; van Breugel, W. 1995 ApJS, 99, 27 McCarthy, P. J., Miley, G. K., de Koff, S., Baum, S. A., Sparks, W. B. et al. 1997 ApJS, 112, 415 Mullin, L. M.; Hardcastle, M. J.; Riley, J. M. 2006 MNRAS, 372, 113 Mullin, L.M., Riley, J. M., Hardcastle, M. J. MNRAS 2008, 390, 595 Nulsen, P. E. J., Hambrick, D. C., McNamara, B. R. 2005 ApJ, 625L, 9 O’Dea, C. P., Daly, R. A., Kharb P., Freeman, K. A., Baum, S. A., 2009 A&A 494, 471 Perryman, M. A. C.; Lilly, S. J.; Longair, M. S.; Downes, A. J. B. 1984 MNRAS, 209, 159 Podigachoski, P.; Barthel, P. D.; Haas, M.; Leipski, C.; Wilkes, B. et al. 2015 A&A, 575A, 80 Privon, G. C.; O’Dea, C. P.; Baum, S. A.; Axon, D. J.; Kharb, P.; 2008 ApJS, 175, 423 Shelton, D. L.; Hardcastle, M. J.; Croston, J. H. 2011 MNRAS, 418, 811 Siemiginowska, A.; Burke, D. J.; Aldcroft, Thomas L.; Worrall, D. M.; Allen, S.; Bechtold, Jill; Clarke, Tracy; Cheung, C. C. 2010 ApJ, 722, 102 Sokolovsky, K. V. , Kovalev, Y. Y., Pushkarev, A. B., Lobanov, A. P. 2011 A&A, 532, 38 Spinrad, H., Marr, J., Aguilar, L., Djorgovski, S. 1985 PASP, 97, 932 Strom, R. G.; Riley, J. M.; Spinrad, H.; van Breugel, W. J. M.; Djorgovski, S.; Liebert, J.; McCarthy, P. J. 1990 A&A, 227, 1 Staurdi, C. et al. 2017 ApJS in prep. Taylor, M. B. 2005, ASP Conf. Ser., 347, 29 Tremblay, G. R.; Chiaberge, M.; Sparks, W. B.; Baum, S. A.; Allen, M. G.; et al. 2009 ApJS, 183, 278 Werner, M. W., Murphy, D. W., Livingston, J. H., Gorjian, V., Jones, D. L., 2012 ApJ, 759, 86 Westhues, C.; Haas, M.; Barthel, P.; Wilkes, B. J.; Willner, S. P. et al. 2016 AJ, 151, 120 Wilkes, B. J., Kuraszkiewicz, J., Haas, M. et al. 2013 ApJ, 773, 15
[|llllllllllll|]{} 3CR & Class & R.A. (J2000) & Dec. (J2000) & z & kpc scale & D$_L$ & N$_H$ & m$_v$ & S$_{178}$ &[[*Chandra*]{}]{} & Obs. Date\
name & & (hh mm ss) & (dd mm ss) & & (kpc/arcsec) & (Mpc) & cm -2 & & Jy & Obs. ID & yyyy-mm-dd\
34 & FRII - HERG & 01 10 18.542 & +31 47 19.51 & 0.69 & 7.191 & 4236.3 & 5.50e20 & 21 & 11.9 & 16046 & 2014-09-25\
41 & FRII - HERG & 01 26 44.325 & +33 13 10.96 & 0.794 & 7.586 & 5035.7 & 5.09e20 & 21 & 10.6 & 16047 & 2014-09-03\
44 & QSO & 01 31 21.647 & +06 23 43.14 & 0.66 & 7.058 & 4011.6 & 3.18e20 & 22 & 7.9 & 16048 & 2014-06-14\
54 & FRII - HERG & 01 55 30.258 & +43 45 59.06 & 0.8274 & 7.693 & 5298.7 & 7.80e20 & 22 & 8.8 & 16049 & 2014-06-15\
55 & FRII - HERG & 01 57 10.539 & +28 51 39.70 & 0.735 & 7.374 & 4578.4 & 5.38e20 & 22 & 21.5 & 16050 & 2014-06-15\
93 & QSO & 03 43 29.996 & +04 57 48.60 & 0.358 & 5.059 & 1924.5 & 1.15e21 & 18.1 & 9.9 & 16051 & 2014-10-10\
107 & FRII - HERG & 04 12 22.620 & -00 59 32.69 & 0.785 & 7.555 & 4965.3 & 8.42e20 & 22 & 10.8 & 16052 & 2014-09-02\
114 & FRII - LERG & 04 20 22.243 & +17 53 56.97 & 0.815 & 7.654 & 5200.7 & 1.61e21 & 22 & 6.5 & 16053 & 2014-09-02\
142.1 & FRII & 05 31 29.334 & +06 30 26.90 & 0.4061 & 5.476 & 2233 & 1.79e21 & 21 & 19.4 & 16054 & 2014-08-16\
154 & QSO & 06 13 50.139 & +26 04 36.64 & 0.580 & 6.654 & 3426.3 & 3.47e21 & 18.0 & 23.1 & 16055 & 2014-08-13\
169.1 & FRII - HERG & 06 51 14.816 & +45 09 28.56 & 0.633 & 6.930 & 3811.7 & 9.30e20 & 20.5 & 7.3 & 16056 & 2014-08-16\
217 & FRII - HERG & 09 08 50.6 & +37 48 19 & 0.898 & 7.892 & 5863.9 & 1.68e20 & 22 & 11.3 & 16057 & 2014-06-13\
225B & FRII -HERG & 09 42 15.396 & +13 45 50.49 & 0.582 & 6.665 & 3440.7 & 3.45e20 & 19 & 21.3 & 16058 & 2014-10-18\
237 & FRII-CSS & 10 08 00.0 & +07 30 16 & 0.877 & 7.836 & 5694.5 & 1.89e20 & 21 & 20.9 & 16059 & 2014-10-31\
247 & FRII - HERG & 10 58 58.973 & +43 01 24.66 & 0.750 & 7.430 & 4693.7 & 8.79e19 & 21.5 & 10.6 & 16060 & 2014-09-26\
272 & FRII & 12 24 28.5 & +42 06 36 & 0.944 & 8.003 & 6238.6 & 2.23e20 & 22 & 8 & 16061 & 2015-03-01\
277 & FRII & 12 51 43.6 & +50 34 25 & 0.414 & 5.540 & 2284.7 & 1.04e20 & 20 & 7.5 & 16062 & 2015-03-03\
277.2 & FRII - HERG & 12 53 33.330 & +15 42 31.18 & 0.767 & 7.492 & 4825.2 & 1.70e20 & 21.5 & 12 & 16063 & 2015-05-07\
288.1 & QSO & 13 42 13.267 & +60 21 42.79 & 0.9610 & 8.041 & 6378.3 & 1.75e20 & 18.1 & 9 & 16064 & 2014-06-08\
292 & FRII - HERG & 13 50 41.852 & +64 29 35.86 & 0.713 & 7.287 & 4410.4 & 1.66e20 & 20.7 & 10.1 & 16065$^+$ & 2014-09-12\
293.1 & FRII & 13 54 40.519 & +16 14 43.14 & 0.709 & 7.271 & 4380.0 & 1.82e20 & 21 & 9.2 & 16066 & 2014-06-05\
323 & FRII & 15 41 45.594 & +60 15 34.03 & 0.6790 & 7.143 & 4153.6 & 1.49e20 & 21 & 8.4 & 16067 & 2014-04-30\
\
Col. (1): The 3CR name. Col. (2): The ‘class’ column contains both a radio descriptor (Fanaroff-Riley class I or II), Compact Steep Spectrum (CSS) and the optical spectroscopic designation, LERG, “Low Excitation Radio Galaxy”, HERG, “High Excitation Radio Galaxy” and QSO for quasars [see also @perryman84; @hes96; @grimes04 for more details]. Col. (3-4): The celestial positions listed are those of the radio nuclei which we used to register the X-ray images: Right ascension and Declination (equinox J2000, see § \[sec:obs\] for details). We reported here the original 3CR position [@spinrad85] of the sources for which the radio core was not clearly detected. Col. (5): Redshift $z$. We also verified in the literature (e.g., NED and/or SIMBAD databases) if new $z$ values were reported after the release of the 3CR catalog. Col. (6): The angular to linear scale factor in arcseconds. Cosmological parameters used to compute it are reported in § \[sec:intro\]. Col. (7): Luminosity Distance in Mpc. Cosmological parameters used to compute it are reported in § \[sec:intro\]. Col. (8): Galactic Neutral hydrogen column densities N$_{H,Gal}$ along the line of sight [@kalberla05]. Col. (8): The optical magnitude in the V band taken from the 3CR catalog [@spinrad85]. Col. (9): S$_{178}$ is the flux density at 178 MHz, taken from Spinrad et al. (1985). Col. (10): The [[*Chandra*]{}]{} observation ID. Col. (11): The date when the [[*Chandra*]{}]{} observation was performed.\
[$^+$ For 3CR292 two Chandra observations are available with obsID: 16065 and 17488, the latter performed in 2014-11-21, (see § \[sec:results\] for details).]{}
[|rrrrrrr|]{} 3CR & Ext. Ratio & F$_{0.5-1~keV}^*$ & F$_{1-2~keV}^*$ & F$_{2-7~keV}^*$ & F$_{0.5-7~keV}^*$ & L$_X$\
name & & (cgs) & (cgs) & (cgs) & (cgs) & (10$^{44}$erg s$^{-1}$)\
34 & 0.72 (0.04) & 0.25 (0.57) & 4.69 (1.43) & 77.53 (9.95) & 82.47 (10.06) & 1.76 (0.22)\
41 & 0.59 (0.06) & 0.32 (0.32) & 0.94 (0.68) & 41.33 (7.42) & 42.6 (7.46) & 1.29 (0.23)\
44 & 0.45 (0.09) & 0.82 (0.59) & 0.68 (0.64) & 9.71 (3.67) & 11.22 (3.77) & 0.22 (0.07)\
54 & 0.83 (0.05) & 1.06 (0.75) & 4.76 (1.33) & 41.04 (7.25) & 46.85 (7.41) & 1.57 (0.25)\
55 & 0.48 (0.09) & 0.82 (0.58) & 2.41 (0.91) & 8.39 (3.47) & 11.61 (3.64) & 0.29 (0.09)\
93 & 0.948 (0.006) & 77.48 (6.26) & 157.41 (7.41) & 429.53 (22.58) & 664.42 (24.58) & 2.94 (0.11)\
107 & 0.15 (0.06) & 0.38 (0.38) & 0.49 (0.49) & 2.95 (2.12) & 3.82 (2.21) & 0.11 (0.07)\
114 & 0.91 (0.02) & 1.16 (0.82) & 43.23 (4.05) & 255.54 (17.3) & 299.94 (17.78) & 9.71 (0.58)\
142.1 & 0.61 (0.09) & 0.57 (0.57) & 2.24 (0.91) & 11.19 (3.77) & 13.99 (3.92) & 0.08 (0.02)\
154 & 0.947 (0.006) & 41.35 (4.59) & 241.5 (9.31) & 923.47 (33.73) & 1206.32 (35.29) & 16.95 (0.5)\
169.1 & 0.64 (0.08) & 2.1 (0.94) & 0.23 (0.51) & 18.65 (4.98) & 20.98 (5.1) & 0.36 (0.09)\
217 & 0.91 (0.20) & 1.42 (1.0) & 15.01 (2.29) & 100.01 (10.98) & 116.44 (11.26) & 4.78 (0.46)\
225B & 0.48 (0.11) & 0.68 (0.48) & 0.12 (0.26) & 5.18 (2.59) & 5.98 (2.65) & 0.08 (0.04)\
237 & 0.84 (0.05) & 4.11 (1.45) & 7.0 (1.53) & 13.76 (4.01) & 24.87 (4.53) & 0.97 (0.18)\
247 & 0.7 (0.06) & 0.99 (0.7) & 3.71 (1.25) & 36.22 (6.61) & 40.92 (6.77) & 1.07 (0.18)\
272 & 0.41 (0.12) & 0.0 (0.0) & 0.46 (0.46) & 4.85 (2.48) & 5.31 (2.52) & 0.25 (0.12)\
277 & 0.84 (0.06) & 0.33 (0.33) & 2.2 (1.0) & 26.81 (6.0) & 29.34 (6.09) & 0.18 (0.04)\
277.2 & 0.43 (0.10) & 3.15 (1.29) & 0.21 (0.47) & 2.26 (2.26) & 5.63 (2.65) & 0.16 (0.07)\
288.1 & 0.946 (0.08) & 76.35 (5.65) & 114.05 (6.1) & 236.49 (16.45) & 426.88 (18.43) & 20.78 (0.9)\
292 & 0.86 (0.04) & 0.9 (0.9) & 1.8 (1.04) & 103.58 (14.13) & 106.29 (14.19) & 2.47 (0.33)\
293.1 & 0.26 (0.10) & 0.0 (0.0) & 0.84 (0.6) & 1.76 (1.24) & 2.59 (1.38) & 0.06 (0.03)\
323 & 0.18 (0.06) & 0.0 (0.0) & 0.9 (0.65) & 5.03 (2.25) & 5.93 (2.34) & 0.12 (0.05)\
\
Col. (1): The 3CR name. Col. (2): The Ext. Ratio defined as the ratio of the net counts in the r=2 circle to the net counts in the r=10 circular region surrounding the core of each 3CR source. The 1$\sigma$ uncertainties are given in parenthesis. Col. (3): Measured X-ray flux between 0.5 and 1 keV. Col. (4): Measured X-ray flux between 1 and 2 keV. Col. (5): Measured X-ray flux between 2 and 7 keV. Col. (6): Measured X-ray flux between 0.5 and 7 keV. Col. (7): X-ray luminosity in the range 0.5 to 7 keV with the 1$\sigma$ uncertainties given in parenthesis.\
Note:\
($^*$) Fluxes are given in units of 10$^{-15}$erg cm$^{-2}$s$^{-1}$ and 1$\sigma$ uncertainties are given in parenthesis. The uncertainties on the flux measurements are computed as described in § \[sec:obs\]\
($^+$) Sources having count rates above the threshold of 0.1 counts per frame for which the X-ray flux measurement is affected by pileup [see @massaro13 and references therein for additional details].
[|lllll|]{} 3CR & $\Gamma_X$ & $N_{H,int}$ & <span style="font-variant:small-caps;">f</span> & $\chi^2/dof$\
93 & 1.78(-0.08,0.15) & $<$0.06 & $<$0.91 & 34.91/25\
114 & 2.06(-0.34,0.36) & 7.21(-2.4,2.54) & -& 3.91/7\
154 & 1.85(-0.14,0.18) & 0.61(-0.26,0.28) & $<$0.92 & 41.52/39\
288.1 & 1.77(-0.08,0.09) & $<$0.1 & - &16.99/19\
\
Col. (1): The 3CR name. Col. (2): The X-ray photon index $\Gamma_X$. Col. (3): The intrinsic absorption at the source redshift. Col. (4): The fraction of flux falling into the pileup region. Col. (5): The $\chi^2$ value divided by the degrees of freedom. Statistical uncertainties quoted refer to the 68% confidence level.
[|rrrrrrrrr|]{} 3CR & Component & class & Counts & F$_{0.5-1~keV}^*$ & F$_{1-2~keV}^*$ & F$_{2-7~keV}^*$ & F$_{0.5-7~keV}^*$ & L$_X$\
name & & & & (cgs) & (cgs) & (cgs) & (cgs) & (10$^{42}$erg s$^{-1}$)\
41 & s11.3 & h & 6 (0.1) & 0.56 (0.56) & 0.55 (0.55) & 4.7 (2.7) & 5.8 (2.8) & 17.5 (8.5)\
54 & s9.3 & h & 3 (0.4) & 0.0 (0.0) & 0.91 (0.52) & 0.0 (0.0) & 0.91 (0.52) & 3.05 (1.74)\
107 & s5.0 & l & 11 (1.5) & 0.62 (0.62) & 1.3 (0.8) & 3.9 (2.3) & 5.9 (2.5) & 17.3 (7.2)\
154 & e33.0 & k & 8 (0.9) & 0.0 (0.0) & <0.24 & 6.3 (2.6) & 6.6 (2.6) & 9.2 (3.7)\
225B & w2.0 & h & 4 (0.1) & 0.0 (0.0) & 0.79 (0.58) & 0.69 (0.69) & 1.5 (0.9) & 2.1 (1.3)\
\
Col. (1): The 3CR name. Col. (2): The component name (e.g., knot or hotspot) is a combination of one letter indicating the orientation of the radio structure and one number indicating distance from the core in arcseconds. Col. (3): The component class: “h” = hotspot - “k” = knot - “l” = lobe. Col. (4): The total counts in the photometric circle together with the average of the 8 background regions, in parentheses; both for the 0.5 to 7 keV band. Col. (5): Measured X-ray flux between 0.5 and 1 keV. Col. (6): Measured X-ray flux between 1 and 2 keV. Col. (7): Measured X-ray flux between 2 and 7 keV. Col. (8): Measured X-ray flux between 0.5 and 7 keV. Col. (9): X-ray luminosity in the range 0.5 to 7 keV with the 1$\sigma$ uncertainties given in parenthesis.\
Note:\
($^*$) Fluxes are given in units of 10$^{-15}$erg cm$^{-2}$s$^{-1}$ and 1$\sigma$ uncertainties are given in parenthesis. The uncertainties on the flux measurements were computed as described in § \[sec:obs\]\
A: Images of the sources {#sec:images}
========================
For all the 3CR sources in our sample, radio morphologies are shown here as contours superposed on the re-gridded/smoothed X-ray event files. The full width half maximum (FWHM) of the Gaussian smoothing function and the binning factor are reported in the figure captions. X-ray event files were limited to the 0.5 to 7 keV band and rebinned to change the pixel size with a binning factor ’f’ (e.g. f=1/4 produces pixels 4 times smaller than the native ACIS pixel of 0.492). The labels on the color bar for each X-ray map are in units of counts/pixel. We included in each caption also the radio brightness of the lowest contour, the factor (usually 2 or 4) by which each subsequent contour exceeds the previous one, the frequency of the radio map, and the FWHM of the clean beam. Figures appear so different from each other mainly because of the wide range in angular size of the radio sources.
![The X-ray image of 3CR44 for the energy band 0.5-7 keV. The event file has been regridded to 1/8 of the native pixel size (i.e., 0.492). The image has been smoothed with a Gaussian of FWHM=5. The radio contours (black) were computed using a 8.4 GHz radio map and start at 0.25 mJy/beam, increasing by factors of four. Radio core is weak but detected in our 8.4 GHz radio image; thus the [[*Chandra*]{}]{} image is registered.[]{data-label="fig:3c44app"}](./3c44radcon.pdf)
![The X-ray image of 3CR54 for the energy band 0.5-7 keV. The event file has been regridded to 1/8 of the native pixel size (i.e., 0.492). The image has been smoothed with a Gaussian of FWHM=7. The radio contours (black) were computed using a 8.4 GHz radio map and start at 0.25 mJy/beam, increasing by factors of four.[]{data-label="fig:3c54app"}](./3c54radcon.pdf)
![The X-ray image of 3CR107 for the energy band 0.5-7 keV. The event file has been regridded to 1/4 of the native pixel size (i.e., 0.492). The image has been smoothed with a Gaussian of FWHM=7. The radio contours (black) were computed using a 4.9 GHz radio map and start at 0.125 mJy/beam, increasing by factors of four.[]{data-label="fig:3c107app"}](./3c107radcon.pdf)
![The X-ray image of 3CR114 for the energy band 0.5-7 keV. The event file has been regridded to 1/8 of the native pixel size (i.e., 0.492). The image has been smoothed with a Gaussian of FWHM=7. The radio contours (black) were computed using a 4.9 GHz radio map and start at 1.0 mJy/beam, increasing by factors of four.[]{data-label="fig:3c114app"}](./3c114radcon.pdf)
![The X-ray image of 3CR169.1 for the energy band 0.5-7 keV. The event file has been regridded to 1/4 of the native pixel size (i.e., 0.492). The image has been smoothed with a Gaussian of FWHM=7. The radio contours (black) were computed using a 8.4 GHz radio map and start at 0.25 mJy/beam, increasing by factors of four.[]{data-label="fig:3c169.1app"}](./3c169p1radcon.pdf)
![The X-ray image of 3CR292 for the energy band 0.5-7 keV (obsID 17488). The event file has been regridded to 1/4 of the native pixel size (i.e., 0.492). The image has been smoothed with a Gaussian of FWHM=5. The radio contours (black) were computed using a 8.5 GHz radio map and start at 0.3 mJy/beam, increasing by factors of four.[]{data-label="fig:3c292app"}](./3c292radcon.pdf)
![The X-ray image of 3CR292 for the energy band 0.5-7 keV. The event file has been regridded to 1/4 of the native pixel size (i.e., 0.492). The image has been smoothed with a Gaussian of FWHM=7. The radio contours (black) were computed using a 8.4 GHz radio map and start at 0.25 mJy/beam, increasing by factors of four.[]{data-label="fig:3c323app"}](./3c323radcon.pdf)
B: The Status of the [[*Chandra*]{}]{} X-ray 3CR observations {#sec:state}
=============================================================
Here we present the current status of the [[*Chandra*]{}]{} and [[*XMM-Newton*]{}]{} observations for the entire 3CR catalog. While [[*Chandra*]{}]{} X-ray observations have been uniformly re-analyzed, as reported in our previous investigations, all the [[*XMM-Newton*]{}]{} information provided here is based on a literature search [see e.g. @croston05; @belsole07; @croston08; @laskar10; @shelton11; @ineson13; @mannering13; @ineson17 and references therein for more details].
For all 3CR sources, we report their classification, labeling: radio galaxies (RG), according to the Fanaroff & Riley criterion [@fanaroff74]; quasars (i.e., QSRs); Seyfert galaxies (Sy) and BL Lac objects (BL). We also indicate as “UNID” those sources which, lacking optical spectroscopy, remain unidentified. We include a column reporting the radio morphology for the radio galaxies (FRI $vs$ FRII types) and indicating those objects that also show the radio structure of: (i) Compact Steep Spectrum (CSS) or X-shaped (XS) radio sources or (ii) have been classified in the literature as wide-angle tailed or narrow-angle tailed radio galaxies (WAT and NAT, resepctively). We also devoted a column to the optical classification of radio galaxies distinguishing them as HERG or LERG. The most updated value of the redshift $z$ is also reported and we used a “cluster flag” to label sources that belong to a known galaxy group/cluster. We considered sources belonging to a galaxy-rich large scale environment those for which there is a known optical group/cluster reported in the literature and/or those for which there is an archival X-ray observation confirming the presence of hot gas in the intergalactic medium.
Regarding the X-ray analysis, we report X-ray detections of radio components adopting the following symbols: $k=$ jet knot; $h=$ hotspot; $l=$ lobe and $igm$ for sources that belong to a galaxy-rich large scale environment. No distinction has been made between sources lying in group or clusters of galaxies. We also adopt the symbol $e$ for those radio objects that show extended X-ray emission of kpc scale as highlighted in our analyses using the “extent ratio” measurements. For [[*XMM-Newton*]{}]{} observations we only adopted $l$ and $igm$ symbols due to the lower angular resolution with respect to [[*Chandra*]{}]{} that does not allow to see counterparts of jet knots and hotspots in the large fraction fo the 3CR sources.
Finally, the “X-ray flag” indicates if the source was already observed by [[*Chandra*]{}]{} (c) and/or [[*XMM-Newton*]{}]{} (x). Sources marked with a $^*$ close to their 3CR name are those not re-analyzed in our previous studies (see § \[sec:history\] for more details). The table present in this work updates and thus supersedes those included in previous publications.
[|rrrrrrrrr|]{} 3CR & $z$ & class & radio & optical & Cluster & [[*Chandra*]{}]{} & [[*XMM-Newton*]{}]{} & X-ray\
name & & & morph. & class & flag & detections & detections & obs.\
2.0 & 1.03658 & QSO & & & & e & & c\
6.1 & 0.8404 & RG & FRII & HERG & & h & & c\
9.0 & 2.01976 & QSO & LDQ & & & k;l & & c\
11.1 & ? & UNID & - & - & & & & x\
13.0 & 1.351 & RG & FRII & HERG & & e;h & & c-x\
14.0 & 1.469 & QSO & & & & & & c\
14.1 & ? & UNID & - & - & & & & x\
15.0 & 0.07368 & RG & FRI & LERG & & k;l & & c\
16.0 & 0.405 & RG & FRII & HERG & & e;h;l & & c-x\
17.0 & 0.219685 & QSO & & & & k & & c\
18.0 & 0.188 & RG & FRII & & & e & & c\
19.0 & 0.482 & RG & FRII & LERG & yes & h;igm & & c-x\
20.0 & 0.174 & RG & FRII & HERG & & & & c-x\
21.1 & ? & UNID & - & - & & & & x\
22.0 & 0.936 & RG & FRII & & & & & c\
27.0 & 0.184 & RG & FRII & HERG & & & & c\
28.0 & 0.19544 & RG & FRI & LERG & yes & igm & igm & c-x\
29.0 & 0.045031 & RG & FRI & LERG & yes & k;igm & & c\
31.0 & 0.017005 & RG & FRI & LERG & yes & k & igm & c-x\
33.0 & 0.0597 & RG & FRII & HERG & yes & h:l & l;igm & c-x\
33.1 & 0.181 & RG & FRII & HERG & & & & c\
33.2 & ? & UNID & - & - & & & & x\
34.0 & 0.69 & RG & FRII & HERG & yes & igm & & c\
35.0 & 0.067013 & RG & FRII & LERG & yes & e;l & l;igm & c-x\
36.0 & 1.301 & RG & FRII & HERG & & & & c\
40.0 & 0.018 & RG & FRI-WAT & LERG & yes & igm & igm & c-x\
41.0 & 0.795 & RG & FRII & HERG & & h & & c\
42.0 & 0.39598 & RG & FRII & HERG & & & & c-x\
43.0 & 1.459 & QSO & CSS & & & & & c\
44.0 & 0.66 & QSO & & & yes & & & c\
46.0 & 0.4373 & RG & FRII & HERG & yes & & igm & c-x\
47.0 & 0.425 & QSO & LDQ & & & h;l & & c\
48.0 & 0.367 & QSO & CSS & & & & & c\
49.0 & 0.23568 & RG & FRII-CSS & HERG & & & & c\
52.0 & 0.29 & RG & FRII-XS & HERG & yes & h & & c\
54.0 & 0.8274 & RG & FRII & HERG & & h & & c\
55.0 & 0.7348 & RG & FRII & HERG & & & & c-x\
61.1 & 0.18781 & RG & FRII & HERG & & h & & c-x\
63.0 & 0.175 & RG & FRII & HERG & & & & c\
65.0 & 1.176 & RG & FRII & HERG & & e;h & & c-x\
66.0A & ? & BLL & BL & - & yes & & & c-x\
66.0B & 0.021258 & RG & FRI-XS & LERG & yes & k & igm & c-x\
67.0 & 0.3102 & RG & FRII-CSS & & & & & c-x\
68.1 & 1.238 & QSO & & & & & & c\
68.2 & 1.575 & RG & FRII & HERG & & e;h & & c\
69.0 & 0.458 & RG & FRII & HERG & & & & c\
71.0 & 0.003793 & SEY & - & Sy2 & & & & c-x\
75.0 & 0.023153 & RG & FRI-WAT & LERG & yes & igm & igm & c-x\
76.1 & 0.032489 & RG & FRI & LERG & yes & igm & igm & c-x\
78.0 & 0.028653 & RG & FRI & LERG & & k & & c\
\
Col. (1): The 3CR name. Col. (2): Redshift $z$. We also verified in the literature (e.g., NED and/or SIMBAD databases) if new $z$ values were reported after the release of the 3CR catalog. Col. (3): The source classification of the sources: RG stands for radio galaxies, QSO for quasars; Sy for Seyfert galaxies and BLL for BL Lac objects. We used the acronym UNID for sources that are still unidentified; i.e., lacking of an optical spectroscopic observation. Col. (4): The radio morphological classification: FRI and FRII refer to the Fanaroff and Riley classification criterion [@fanaroff74] while LDQ and CDQ is sometimes adopted for lobe-dominated and core-dominated quasars; we also indicated if in the literature the source is classified as CSS or if presents an X-shaped radio structure (XS) or if it is a narrow or wide angle tailed radio galaxy (NAT and WAT, respectively). Col. (5): The optical spectroscopic designation: LERG, “Low Excitation Radio Galaxy”, HERG, “High Excitation Radio Galaxy”. Col. (6): The “cluster flag” as discussed in § \[sec:state\]. Col. (7): In this column we report if the source has a radio component with an X-ray counterpart in a [[*Chandra*]{}]{} observation. We used the following labels: k = jet knot; h = hotspot; l = lobe; e = extended X-ray emission around the nucleus on kpc scale and igm whenever this extended emission is associated with the hot gas present in the intergalactic medium. Col. (8): The same of the previous column (i.e. Col. 7) but for the [[*XMM-Newton*]{}]{} observations. Col. (9): The c flag indicates that at least one [[*Chandra*]{}]{} observation is present in its archive while the x flag refers to the [[*XMM-Newton*]{}]{} archive.
[|rrrrrrrrr|]{} 3CR & $z$ & class & radio & optical & Cluster & [[*Chandra*]{}]{} & [[*XMM-Newton*]{}]{} & X-ray\
name & & & morph. & class & flag & detections & detections & obs.\
79.0 & 0.2559 & RG & FRII & HERG & yes & & igm & c-x\
83.1 & 0.025137 & RG & FRI-NAT & LERG & yes & k;igm & igm & c-x\
84.0 & 0.017559 & RG & FRI & LERG & yes & igm & igm & c-x\
86.0 & ? & UNID & - & - & & & &\
88.0 & 0.030221 & RG & FRI & LERG & yes & k;igm & igm & c-x\
89.0 & 0.13981 & RG & FRI-WAT & LERG & yes & igm & & c\
91.0 & ? & UNID & - & - & & & &\
93.0 & 0.35712 & QSO & & & & e & & c\
93.1 & 0.243 & RG & FRII-CSS & HERG & yes & & & c\
98.0 & 0.030454 & RG & FRII-XS & HERG & & l & l & c-x\
99.0 & 0.426 & SEY & - & Sy2 & & & & c\
103.0 & 0.33 & RG & FRII & & & & & c\
105.0 & 0.089 & RG & FRII & HERG & & k;h & & c-x\
107.0 & 0.785 & RG & FRII & HERG & & l & & c\
109.0 & 0.3056 & RG & FRII & HERG & & h;l & & c-x\
111.0 & 0.0485 & RG & FRII & & & k;h & & c-x\
114.0 & 0.815 & RG & FRII & LERG & & & & c\
119.0 & 1.023 & QSO & CSS & & & & & c\
123.0 & 0.2177 & RG & FRII & LERG & yes & h;igm & & c\
124.0 & 1.083 & RG & FRII & HERG & & e & & c\
125.0 & ? & UNID & - & - & & & &\
129.0 & 0.0208 & RG & FRI-NAT & & yes & k;igm & igm & c-x\
129.1 & 0.0222 & RG & FRI & & yes & igm & igm & c-x\
130.0 & 0.109 & RG & FRI-WAT & & & igm & & c\
131.0 & ? & UNID & - & - & & & &\
132.0 & 0.214 & RG & FRII & LERG & yes & & & c-x\
133.0 & 0.2775 & RG & FRII & HERG & & & & c\
134.0 & ? & UNID & - & - & & & &\
135.0 & 0.12738 & RG & FRII & HERG & yes & & & c\
136.1 & 0.064 & RG & FRII-XS & HERG & & e & & c\
137.0 & ? & UNID & - & - & & & &\
138.0 & 0.759 & QSO & CSS & & & & & c\
139.2 & ? & UNID & - & - & & & &\
141.0 & ? & UNID & - & - & & & &\
142.1 & 0.4061 & RG & FRII & & & & & c\
147.0 & 0.545 & QSO & CSS & & & & & c\
152.0 & ? & UNID & - & - & & & &\
153.0 & 0.2769 & RG & FRII & LERG & yes & & & c-x\
154.0 & 0.58 & QSO & & & & e;k & & c\
158.0 & ? & UNID & - & - & & & &\
165.0 & 0.2957 & RG & FRII & LERG & & e & & c\
166.0 & 0.2449 & RG & FRII & LERG & & & & c\
169.1 & 0.633 & RG & FRII & HERG & & & & c\
171.0 & 0.2384 & RG & FRII & HERG & & e & & c-x\
172.0 & 0.5191 & RG & FRII & HERG & & & & c\
173.0 & 1.035 & QSO & CSS & HERG & & & & c\
173.1 & 0.2921 & RG & FRII & LERG & yes & h;l & & c\
175.0 & 0.77 & QSO & & & & & & c\
175.1 & 0.92 & RG & FRII & HERG & & & & c\
180.0 & 0.22 & RG & FRII & HERG & & & & c\
\
[|rrrrrrrrr|]{} 3CR & $z$ & class & radio & optical & Cluster & [[*Chandra*]{}]{} & [[*XMM-Newton*]{}]{} & X-ray\
name & & & morph. & class & flag & detections & detections & obs.\
181.0 & 1.382 & QSO & & & & h & & c\
184.0 & 0.994 & RG & FRII & HERG & yes & l & igm & c-x\
184.1 & 0.1182 & RG & FRII & HERG & yes & & & c\
186.0 & 1.06551 & QSO & CSS & & yes & igm & & c\
187.0 & 0.465 & RG & FRII & LERG & & e;l & & c\
190.0 & 0.24639 & QSO & CSS & & & & & c\
191.0 & 1.96810 & QSO & & & & k;l & & c\
192.0 & 0.05968 & RG & FRII-XS & HERG & yes & l & igm & c-x\
194.0 & 1.184 & RG & FRII & HERG & & & & c\
196.0 & 0.87063 & QSO & & & & & & c\
196.1 & 0.198 & RG & FRII & LERG & yes & igm & & c\
197.1 & 0.12825 & RG & FRII & HERG & yes & & & c\
198.0 & 0.08145 & RG & FRII & HERG & yes & & & c\
200.0 & 0.458 & RG & FRII & LERG & yes & k;l & & c\
204.0 & 1.112 & QSO & & & & & & c-x\
205.0 & 1.53154 & QSO & & & & & & c-x\
207.0 & 0.68038 & QSO & LDQ & & yes & k;l & igm & c-x\
208.0 & 1.11151 & QSO & & & & & & c\
208.1 & 1.02 & QSO & & & & & & c-x\
210.0 & 1.169 & RG & FRII & HERG & yes & e;h & igm & c-x\
212.0 & 1.04931 & QSO & LDQ & & & e;h & & c\
213.1 & 0.19405 & RG & FRI-CSS & LERG & yes & e;h & & c\
215.0 & 0.41106 & QSO & & & yes & k;l & & c-x\
216.0 & 0.67002 & QSO & & & & & & c\
217.0 & 0.8975 & RG & FRII & HERG & & & & c\
219.0 & 0.17456 & RG & FRII & HERG & yes & k;l & & c\
220.1 & 0.61 & RG & FRII & HERG & yes & igm & & c\
220.2 & 1.15610 & QSO & & & & h;l & & c\
220.3 & 0.68 & RG & FRII & & & & & c\
222.0 & 1.339 & RG & FRI & & & & & c\
223.0 & 0.13673 & RG & FRII & HERG & yes & & igm & c-x\
223.1 & 0.1075 & RG & FRII-XS & HERG & & & & c\
225.0A & 1.565 & RG & FRII & & & & & c\
225.0B & 0.58 & RG & FRII & HERG & & h & & c\
226.0 & 0.8177 & RG & FRII & & & & & c\
227.0 & 0.086272 & RG & FRII & HERG & & h & & c\
228.0 & 0.5524 & RG & FRII & HERG & & e;h & & c\
230.0 & 1.487 & RG & FRII & HERG & & & & c\
231.0 & 0.000677 & RG & FRI & LERG & & & & c-x\
234.0 & 0.184925 & RG & FRII & HERG & & h & & c-x\
236.0 & 0.1005 & RG & FRII & LERG & & & & c\
237.0 & 0.877 & RG & FRII-CSS & & & & & c\
238.0 & 1.405 & RG & FRII & HERG & & & & c\
239.0 & 1.781 & RG & FRII & & & & & x\
241.0 & 1.617 & RG & FRII & & & & & c-x\
244.1 & 0.428 & RG & FRII & HERG & yes & e & & c-x\
245.0 & 1.027872 & QSO & & & & k & & c\
247.0 & 0.7489 & RG & FRII & HERG & yes & & & c\
249.0 & 1.554 & QSO & & & & & & x\
249.1 & 0.3115 & QSO & & & & & & c-x\
250.0 & ? & UNID & - & - & & & &\
\
[|rrrrrrrrr|]{} 3CR & $z$ & class & radio & optical & Cluster & [[*Chandra*]{}]{} & [[*XMM-Newton*]{}]{} & X-ray\
name & & & morph. & class & flag & detections & detections & obs.\
252.0 & 1.1 & RG & FRII & & & & & c\
254.0 & 0.736619 & QSO & LDQ & & & e;h & & c\
255.0 & 1.355 & QSO & & & & & & c\
256.0 & 1.819 & RG & FRII & & & & & c\
257.0 & 2.474 & QSO & & & & & & x\
258.0 & 0.165 & RG & FRI-CSS & LERG & yes & & & c\
263.0 & 0.646 & QSO & LDQ & & & h & & c\
263.1 & 0.824 & RG & FRII & & & & & c\
264.0 & 0.021718 & RG & FRI & LERG & yes & k & igm & c-x\
265.0 & 0.811 & RG & FRII & & & h;l & & c\
266.0 & 1.275 & RG & FRII & & & & & c-x\
267.0 & 1.14 & RG & FRII & & & & & c\
268.1 & 0.97 & RG & FRII & & & h & & c\
268.2 & 0.362 & RG & FRII & & yes & e;h & & c-x\
268.3 & 0.37171 & RG & FRII & & & & & c\
268.4 & 1.402200 & QSO & & & & & & c-x\
270.0 & 0.007378 & RG & FRI & LERG & yes & k & igm & c-x\
270.1 & 1.528432 & QSO & & & & & & c\
272.0 & 0.944 & RG & FRII & & & & & c\
272.1 & 0.003392 & RG & FRI & LERG & yes & k & & c-x\
273.0 & 0.158339 & QSO & CDQ & & & k & & c-x\
274.0 & 0.0043 & RG & FRI & LERG & yes & k;igm & igm & c-x\
274.1 & 0.422 & RG & FRII & HERG & & e & l & c-x\
275.0 & 0.48 & RG & FRII & LERG & yes & & & c\
275.1 & 0.5551 & QSO & LDQ & & & k;h;l & & c\
277.0 & 0.414 & RG & FRII & & & & & c\
277.1 & 0.31978 & QSO & & & & & & c\
277.2 & 0.766 & RG & FRII & HERG & & & & c-x\
277.3 & 0.085336 & RG & FRII & HERG & & & & c\
280.0 & 0.996 & RG & FRII & & yes & k;h;l & & c-x\
280.1 & 1.667065 & QSO & & & & l & &\
284.0 & 0.239754 & RG & FRII & HERG & yes & & igm & c-x\
285.0 & 0.0794 & RG & FRII & HERG & & l & & c\
286.0 & 0.849934 & QSO & & & & & & c\
287.0 & 1.055 & QSO & & & & & & c-x\
287.1 & 0.2156 & RG & FRII & HERG & & h & & c\
288.0 & 0.246 & RG & FRI & LERG & yes & igm & & c\
288.1 & 0.96296 & QSO & & & & & & c\
289.0 & 0.9674 & RG & FRII & & & & & c\
292.0 & 0.71 & RG & FRII & HERG & yes & e & igm & c-x\
293.0 & 0.045034 & RG & FRI & LERG & & e & & c\
293.1 & 0.709 & RG & FRII & & & & & c\
294.0 & 1.779 & RG & FRII & & yes & h;igm & & c\
295.0 & 0.4641 & RG & FRII & LERG & yes & h;igm & & c\
296.0 & 0.024704 & RG & FRI & LERG & yes & k;igm & igm & c-x\
297.0 & 1.4061 & QSO & & & & & & c\
298.0 & 1.438120 & QSO & & & yes & & igm & c-x\
299.0 & 0.367 & RG & FRII & & yes & h & & c\
300.0 & 0.27 & RG & FRII & HERG & & & & c-x\
\
[|rrrrrrrrr|]{} 3CR & $z$ & class & radio & optical & Cluster & [[*Chandra*]{}]{} & [[*XMM-Newton*]{}]{} & X-ray\
name & & & morph. & class & flag & detections & detections & obs.\
300.1 & 1.15885 & RG & FRII & HERG & & & & c\
303.0 & 0.141186 & RG & FRII & HERG & yes & k;l & & c\
303.1 & 0.2704 & RG & FRII-CSS & HERG & & e & & c-x\
305.0 & 0.041639 & RG & FRII-CSS & HERG & & e & & c-x\
305.1 & 1.132 & RG & FRII-CSS & LERG & & & & c\
306.1 & 0.441 & RG & FRII & HERG & yes & e & & c\
309.1 & 0.905 & QSO & & & & e & & c\
310.0 & 0.0538 & RG & FRI & LERG & yes & igm & & c\
314.1 & 0.1197 & RG & FRI & LERG & yes & & & c-x\
313.0 & 0.461 & RG & FRII & HERG & yes & h;igm & & c\
315.0 & 0.1083 & RG & FRI-XS & LERG & yes & e & & c\
317.0 & 0.034457 & RG & FRI & LERG & yes & igm & igm & c-x\
318.0 & 1.574 & RG & FRII & & yes & & & c-x\
318.1 & 0.045311 & RG & FRI & & yes & igm & igm & c-x\
319.0 & 0.192 & RG & FRII & LERG & yes & & & c-x\
320.0 & 0.342 & RG & FRII & & yes & igm & & c\
321.0 & 0.0961 & RG & FRII & HERG & & h;l & & c-x\
322.0 & 1.681 & RG & FRII & & yes & & igm & x\
323.0 & 0.679 & RG & FRII & & & e & & c\
323.1 & 0.2643 & RG & FRII & HERG & yes & & & c\
324.0 & 1.2063 & RG & FRII & & yes & e;h & & c-x\
325.0 & 1.135 & RG & FRII & & & h & & c\
326.0 & 0.0895 & RG & FRII & LERG & & l & & c\
326.1 & 1.825 & RG & FRII & & & & &\
327.0 & 0.1048 & RG & FRII & HERG & yes & h & & c\
327.1 & 0.462 & RG & FRI & HERG & & k & & c-x\
330.0 & 0.55 & RG & FRII & HERG & yes & h;l & & c\
332.0 & 0.151019 & RG & FRII & HERG & yes & & & c\
334.0 & 0.5551 & QSO & LDQ & & & k;l & & c\
336.0 & 0.926542 & QSO & & & & & & c\
337.0 & 0.635 & RG & FRII & & yes & e & & c-x\
338.0 & 0.030354 & RG & FRI & LERG & yes & igm & igm & c-x\
340.0 & 0.7754 & RG & FRII & & & & & c\
341.0 & 0.448 & RG & FRII & HERG & yes & e;k & igm & c-x\
343.0 & 0.988 & QSO & & & & & & c\
343.1 & 0.75 & RG & FRII & & & & & c\
345.0 & 0.5928 & QSO & CDQ & & yes & k & igm & c-x\
346.0 & 0.162012 & RG & FRI & HERG & yes & k & & c\
348.0 & 0.155 & RG & FRI & LERG & yes & igm & igm & c-x\
349.0 & 0.205 & RG & FRII & HERG & & h & & c-x\
351.0 & 0.37194 & RG & FRII & & & h & & c\
352.0 & 0.8067 & RG & FRII & & & & & c\
353.0 & 0.030421 & RG & FRII & LERG & yes & k:l:igm & igm & c-x\
356.0 & 1.079 & RG & FRII & & & e & & c\
357.0 & 0.166148 & RG & FRII & LERG & yes & & & c\
368.0 & 1.131 & RG & FRII & & & & & c\
371.0 & 0.051 & BLL & BL & - & & k & & c\
379.1 & 0.256 & RG & FRII & HERG & & & & c\
380.0 & 0.692 & QSO & CDQ & & & k & & c\
381.0 & 0.1605 & RG & FRII & HERG & & & & c\
382.0 & 0.05787 & RG & FRII & HERG & & & & c-x\
\
[|rrrrrrrrr|]{} 3CR & $z$ & class & radio & optical & Cluster & [[*Chandra*]{}]{} & [[*XMM-Newton*]{}]{} & X-ray\
name & & & morph. & class & flag & detections & detections & obs.\
386.0 & 0.016885 & RG & FRI & LERG & yes & & igm & c-x\
388.0 & 0.0917 & RG & FRII & LERG & yes & e & & c\
389.0 & ? & UNID & - & - & & & & x\
390.0 & ? & UNID & - & - & & & &\
390.3 & 0.0561 & RG & FRII & HERG & & k;h & & c-x\
394.0 & ? & UNID & - & - & & & &\
399.1 & ? & UNID & - & - & & & &\
401.0 & 0.2011 & RG & FRII & LERG & yes & igm & & c\
402.0 & 0.025948 & RG & FRI & & yes & k & & c-x\
403.0 & 0.059 & RG & FRII & HERG & & k;h & & c\
403.1 & 0.0554 & RG & FRII & LERG & yes & & & c\
405.0 & 0.056075 & RG & FRII & & yes & h;igm & igm & c-x\
409.0 & ? & UNID & - & - & & & &\
410.0 & 0.2485 & RG & FRII & & & & & c\
411.0 & 0.467 & RG & FRII & HERG & & & & c-x\
415.2 & ? & UNID & - & - & & & &\
418.0 & 1.686 & QSO & & & & & &\
424.0 & 0.126988 & RG & FRI & LERG & yes & e & & c\
427.1 & 0.572 & RG & FRII & LERG & yes & l;igm & & c\
428.0 & ? & UNID & - & - & & & &\
430.0 & 0.055545 & RG & FRII & LERG & yes & e & & c\
431.0 & ? & UNID & - & - & & & &\
432.0 & 1.785 & QSO & & & & & & c-x\
434.0 & 0.322 & RG & FRII & & yes & & & c\
433.0 & 0.1016 & RG & FRII & HERG & & l & & c\
435.0 & 0.471 & RG & FRII & & & & & c\
436.0 & 0.2145 & RG & FRII & HERG & & e;h & & c-x\
437.0 & 1.48 & RG & FRII & & & e;h & & c\
438.0 & 0.29 & RG & FRII & HERG & yes & igm & & c\
441.0 & 0.708 & RG & FRII & & & & & c\
442.0 & 0.0263 & RG & FRI & LERG & yes & igm & & c\
445.0 & 0.055879 & RG & FRII & & yes & h & & c-x\
449.0 & 0.017085 & RG & FRI & LERG & yes & igm & igm & c-x\
452.0 & 0.0811 & RG & FRII & HERG & yes & h;l & igm & c-x\
454.0 & 1.757 & QSO & & & & & & x\
454.1 & 1.841 & RG & FRII & & yes & & &\
454.2 & ? & UNID & - & - & & & &\
454.3 & 0.859 & QSO & CDQ & & & k & & c-x\
455.0 & 0.543 & QSO & & & & & & c\
456.0 & 0.233 & RG & FRII & HERG & & & & c\
458.0 & 0.289 & RG & FRII & HERG & yes & h & & c\
459.0 & 0.22012 & RG & FRII & & & l & & c-x\
460.0 & 0.268 & RG & FRII & LERG & yes & & & c-x\
465.0 & 0.030221 & RG & FRI-WAT & LERG & yes & k;igm & igm & c-x\
468.1 & ? & UNID & - & - & & & &\
469.1 & 1.336 & RG & FRII & & & & l & c-x\
470.0 & 1.653 & RG & FRII & & & h & & c\
\
[^1]:
[^2]:
[^3]:
[^4]:
[^5]:
[^6]:
[^7]:
[^8]:
[^9]:
[^10]: <http://cxc.harvard.edu/sherpa>
[^11]: <http://cxc.harvard.edu/sherpa/ahelp/jdpileup.html>
[^12]:
[^13]: http://archive.nrao.edu/nvas/
[^14]: http://ned.ipac.caltech.edu/
[^15]: http://www.jb.man.ac.uk/atlas/
[^16]:
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Under the potential outcomes framework, we introduce matched-pair factorial designs, and propose the matched-pair estimator of the factorial effects. We also calculate the randomization-based covariance matrix of the matched-pair estimator, and provide the “Neymanian” estimator of the covariance matrix.'
author:
- |
Jiannan Lu[^1] and Alex Deng \
Analysis and Experimentation, Microsoft Corporation
bibliography:
- 'factorial\_paired\_match.bib'
title: '**On randomization-based causal inference for matched-pair factorial designs**'
---
**Keywords:** Experimental design; factorial effect; precision; potential outcome.
Introduction {#sec:intro}
============
Randomization is widely regarded as the gold standard of causal inference [@Rubin:2008]. Under the potential outcomes framework [@Neyman:1923; @Rubin:1974], for a two-level factor, we define the causal effect as the linear contrast of the potential outcomes under treatment and control. To investigate multiple factors simultaneously, $2^K$ factorial designs [@Fisher:1935; @Yates:1937] can be employed. Randomization-based casual inference for factorial designs has deep roots in the experimental design literature [e.g., @Kempthrone:1952], and was recently presented using the language of potential outcomes [@Dasgupta:2015; @Mukerjee:2016].
Pair-matching [@Cochran:1953], as a special form of stratification, has been widely adopted by researchers and practitioners [e.g., @Grossarth:2008]. For treatment-control studies (i.e., $2^1$ factorial designs), pair-matching has been extensively investigated by the causal inference community [@Rosenbaum:2002; @Imai:2008; @Imai:2009; @Ding:2016; @Fogarty:2016a; @Fogarty:2016b]. Unfortunately, similar discussion appears to be missing for general factorial designs. In this paper, we fill this theoretical gap by extending [@Imai:2008]’s analysis to matched-pair factorial designs. We restrict the experimental units to be a fixed finite population, for a two-fold reason. First, as shown in [@Imai:2008], it is straightforward to generalize the finite-population analyses to infinite populations. Second, for some practical examples, it might be unreasonable to view the experimental units as a random sample from an infinite population.
The paper proceeds as follows. Section \[sec:2k-c\] reviews the randomization-based causal inference framework for completely randomized factorial designs. Section \[sec:2k-m\] introduces matched-pair factorial designs, proposes the matched-pair estimator for the factorial effects, calculates its covariance matrix and the corresponding estimator. Section \[sec:conclusion\] briefly discusses the precision gains by pair-matching in factorial designs, and concludes.
Causal inference for completely randomized factorial designs {#sec:2k-c}
============================================================
To ensure self-containment, we first review the randomization-based causal inference framework for completely randomized factorial designs. Although most materials are adapted from [@Dasgupta:2015] and [@Lu:2016b; @Lu:2016a], some are refined for better clarity. For more detailed discussions on factorial designs, see, e.g., [@Wu:2009].
Factorial designs
-----------------
A $2^K$ factorial design consists of $K$ two-level (coded $-1$ and $+1$) factors. We represent it by the corresponding model matrix [@Wu:2009], a $2^K \times 2^K$ matrix $\bm H_K = (\bm h_0, \ldots, \bm h_{2^K-1})$ that can be constructed as follows:
1. Let $\bm h_0 = \bm 1_{2^K};$
2. For $k=1,\ldots,K$, construct $\bm h_k$ by letting its first $2^{K-k}$ entries be $-1,$ the next $2^{K-k}$ entries be $+1,$ and repeating $2^{k-1}$ times;
3. If $K \ge 2,$ order all subsets of $\{1, \ldots, K\}$ with at least two elements, first by cardinality and then lexicography. For $k = 1, \ldots 2^K-K-1,$ let $\sigma_k$ be the $k$th subset and $\bm h_{K+k} = \prod_{l \in \sigma_k} \bm h_l,$ where “$\prod$” stands for entry-wise product.
The use of the constructed $\bm H_K$ is two-fold:
1. $\bm h_0$ corresponds to the null effect; $\bm h_1$ to $\bm h_K$ correspond to the main effects of the $K$ factors; $\bm h_{K+1}$ to $\bm h_{K+\binom{K}{2}}$ correspond to the two-way interactions; $\ldots;$ $\bm h_{2^K-1}$ corresponds to the $K$-way interaction;
2. The $j$th row of $(\bm h_1, \ldots, \bm h_K)$ corresponds to the $j$th treatment combination $\bm z_j.$
For $j=1, \dots, 2^K,$ let $\bm \lambda_j$ denote the $j$th row of $\bm H_K.$
\[example:1\] For $2^2$ factorial designs, the model matrix is: $$\bm H_2 =
\bordermatrix{& \bm h_0 & \bm h_1& \bm h_2 & \bm h_3\cr
\bm \lambda_1 & +1 & -1 & -1 & +1 \cr
\bm \lambda_2 & +1 & -1 & +1 & -1 \cr
\bm \lambda_3 & +1 & +1 & -1 & -1 \cr
\bm \lambda_4 & +1 & +1 & +1 & +1}.$$ The four treatment combinations are $\bm z_1=(-1, -1),$ $\bm z_2=(-1, +1),$ $\bm z_3=(+1, -1)$ and $\bm z_4=(+1, +1).$ We represent the main effects of factors 1 and 2 by $\bm h_1 = (-1,-1,+1,+1)^\prime$ and $\bm h_2=(-1,+1,-1,+1)^\prime$ respectively, and the two-way interaction by $\bm h_3=(+1,-1,-1,+1)^\prime.$
Randomization-based causal inference
------------------------------------
We consider a $2^K$ factorial design with $N = 2^K r$ units. By invoking the Stable Unit Treatment Value Assumption [@Rubin:1980], for $i = 1, \ldots, N$ and $l=1, \ldots, 2^K,$ let the potential outcome of unit $i$ under $\bm z_l$ be $Y_i(\bm z_l),$ the average potential outcome for $\bm z_l$ be $
\bar Y(\bm z_l) = N^{-1} \sum_{i=1}^N Y_i(\bm z_l),
$ and $
\bm Y_i = \{ Y_i(\bm z_1), \ldots, Y_i(\bm z_{2^K}) \}^{\prime}.
$ Define the individual and population-level factorial effect vectors as $$\label{eq:estimand}
\bm \tau_i = \frac{1}{2^{K-1}} \bm H_K^\prime {\bm Y}_i
\quad
(i=1, \ldots, N);
\quad
\bm \tau = \frac{1}{N} \sum_{i=1}^N \bm \tau_i,$$ respectively. Our interest lies in $\bm \tau.$
We denote the treatment assignment mechanism by $$W_i(\bm z_l)
=
\begin{cases}
1, & \text{if unit $i$ is assigned treatment $\bm z_l,$ } \\
0, & \text{otherwise.} \\
\end{cases}
\quad
(i=1, \ldots, N; l = 1, \ldots, 2^K).$$ We impose the following restrictions on the treatment assignment mechanism: $$\sum_{l=1}^{2^K} W_i(\bm z_l) = 1
\quad
(i = 1, \ldots, N);
\quad
\sum_{i=1}^N W_i(\bm z_l) = r
\quad
(l = 1, \ldots, 2^K).$$ In other words, we assign $r$ units to each treatment, and one treatment to each unit. Therefore, the observed outcome of unit $i$ is $
Y_i^\textrm{obs} = \sum_{l=1}^{2^K} W_i(\bm z_l) Y_i(\bm z_l),
$ and the average observed outcome for treatment $\bm z_l$ is $
\bar Y^\textrm{obs}(\bm z_l)
=
r^{-1} \sum_{i=1}^N W_i(\bm z_l) Y_i(\bm z_l).
$ Under complete randomization, [@Dasgupta:2015] estimated $\bm \tau$ by $$\hat {\bm \tau}_{\textrm{C}} = 2^{-(K-1)} \bm H_K^\prime \bar{\bm Y}^{\mathrm{obs}},
\quad
\bar{\bm Y}^{\mathrm{obs}} = \{ \bar Y^\textrm{obs}(\bm z_1) , \ldots, \bar Y^\textrm{obs}(\bm z_{2^K}) \}^\prime.$$ The sole source of randomness of $
\hat {\bm \tau}_{\textrm{C}}
$ is the treatment assignment. [@Dasgupta:2015] and [@Lu:2016a] derived the covariance matrix of this estimator, and the “Neymanian” estimator of the covariance matrix. We summarize their main results in the following lemmas.
\[lemma:cov-c\] $\hat {\bm \tau}_\textrm{C}$ is unbiased, and its covariance matrix is $$\label{eq:cov-c}
\mathrm{Cov} ( \hat{\bm \tau}_\textrm{C} )
= \frac{1}{2^{2(K-1)}r}
\sum_{l=1}^{2^K} \bm \lambda_l^\prime \bm \lambda_l \underbrace{\frac{1}{N-1} \sum_{i=1}^N \{ Y_i(\bm z_l) - \bar Y(\bm z_l) \}^2}_{S^2(\bm z_l)}
- \frac{1}{N(N-1)} \sum_{i=1}^N (\bm \tau_i - \bm \tau)(\bm \tau_i - \bm \tau)^\prime.$$ Moreover, the “Neymanian” estimator of the covariance matirx is $$\widehat \mathrm{Cov} ( \hat{\bm \tau}_\textrm{C} ) = \frac{1}{2^{2(K-1)}r} \sum_{l=1}^{2^K} \bm \lambda_l^\prime \bm \lambda_l \underbrace{\frac{1}{r-1}\sum_{i=1}^N W_i(\bm z_l) \{ Y_i^\textrm{obs} - \bar Y^\textrm{obs}(\bm z_l) \}^2}_{s^2(\bm z_l)},$$ whose bias is $
\sum_{i=1}^N (\bm \tau_i - \bm \tau)(\bm \tau_i - \bm \tau)^\prime / (N^2-N).
$
The covariance matrix estimator $
\widehat \mathrm{Cov} ( \hat{\bm \tau}_\textrm{C} )
$ is “conservative,” because its diagonal entries, i.e., the variance estimators of the components of $\hat {\bm \tau}_\textrm{C},$ have non-negative biases.
Causal inference for matched-pair randomized factorial designs {#sec:2k-m}
==============================================================
Matched-pair designs and causal parameters
------------------------------------------
As pointed out by [@Imai:2008], they key idea behind matched-pair designs is that “experimental units are paired based on their pre-treatment characteristics and the randomization of treatment is subsequently conducted within each matched pair.” To apply this idea to factorial designs, we group the $N$ experimental units into $r$ “pairs” of $2^K$ units, and within each pair randomly assign one unit to each treatment. Let $\psi_j$ be the set of indices of the units in pair $j,$ such that $$|\psi_j| = 2^K
\quad
(j=1, \ldots, r);
\quad
\psi_j \cap \psi_{j^\prime} = \emptyset
\quad
(\forall j \neq j^\prime);
\quad
\cup_{j=1}^r \psi_j = \{1, \ldots, N\}.$$ For pair $j,$ denote the average outcomes for treatment $\bm z_l$ as $
\bar Y_{j\cdot}(\bm z_l) = 2^{-K} \sum_{i \in \psi_j} Y_i(\bm z_l),
$ and $
\bar{\bm Y}_{j\cdot} = \{ \bar Y_{j\cdot}(\bm z_1), \ldots, \bar Y_{j\cdot}(\bm z_{2^K}) \}^\prime,
$ and the factorial effect vector as $
\bm \tau_{j\cdot} = 2^{-(K-1)} \bm H_K^\prime \bar{\bm Y}_{j\cdot}.
$ It is apparent $$\frac{1}{r} \sum_{j=1}^r \bar Y_{j\cdot}(\bm z_l) = \bar Y(\bm z_l)
\quad
(l = 1, \ldots, 2^k);
\quad
\frac{1}{r} \sum_{j=1}^r \bm \tau_{j\cdot} = \bm \tau.$$
Within each pair, we randomly assign one unit to each treatment. Let the observed outcome of treatment $\bm z_l$ in pair $j$ be $
Y_j^{\textrm{obs}}(\bm z_l) = \sum_{i \in \psi_j} Y_{i}(\bm z_l) W_i(\bm z_l),
$ and $
\bm Y_j^{\textrm{obs}} = \{ Y_j^{\textrm{obs}}(\bm z_1), \ldots, Y_j^{\textrm{obs}}(\bm z_{2^K}) \}^\prime.
$ We estimate $\bm \tau_{j\cdot}$ by $
\hat{\bm \tau}_{j\cdot} = 2^{-(K-1)} \bm H_K^\prime \bm Y_j^{\textrm{obs}}.
$ The matched-pair estimator for $\bm \tau$ is $$\label{eq:est-m}
\hat {\bm \tau}_{\textrm{M}} = \frac{1}{r} \sum_{j=1}^r \hat{\bm \tau}_{j\cdot}.$$
Randomization-based inference
-----------------------------
We now present the main results of this paper.
\[prop:cov-m\] $\hat {\bm \tau}_\textrm{M}$ is an unbiased estimator of $\bm \tau,$ and its covariance matrix is $$\label{eq:cov-m}
\mathrm{Cov} ( \hat{\bm \tau}_\textrm{M} )
= \frac{1}{2^{2(K-1)}r^2}
\sum_{l=1}^{2^K} \bm \lambda_l^\prime \bm \lambda_l \Delta_l
- \frac{1}{2^K(2^K-1)r^2} \bm \Sigma,$$ where $$\Delta_l = \frac{1}{2^K-1}
\left [
(N - 1) S^2(\bm z_l) - 2^K \sum_{j=1}^r
\left\{
\bar Y_{j\cdot}(\bm z_l) - \bar Y(\bm z_l)
\right\}^2
\right ]
\quad
(l = 1, \ldots, 2^K),$$ and $$\bm \Sigma
= \sum_{i=1}^N (\bm \tau_i - \bm \tau)(\bm \tau_i - \bm \tau)^\prime
- 2^K \sum_{j=1}^r (\bm \tau_{j\cdot} - \bm \tau)(\bm \tau_{j\cdot} - \bm \tau)^\prime.$$
To prove the first part, note that $\hat{\bm \tau}_{j\cdot}$ is an unbiased estimator of $\bm \tau_{j\cdot},$ for $j=1, \ldots, r.$ This fact combined with completes the proof.
To prove the second part, let $\bm W_j = \{ W_i(\bm z_l) \}_{i\in \psi_j, l = 1, \ldots, 2^K}$ denote the treatment assignment for pair $j.$ By definition, $\bm W_j$’s are independently and identically distributed, implying the (joint) independence of $
\hat{\bm \tau}_{j\cdot}
$’s. Consequently, we can treat each pair as a completely randomized factorial design with $2^K$ units. Therefore by Lemma \[lemma:cov-c\], $$\mathrm{Cov} ( \hat{\bm \tau}_{j\cdot} )
=
\frac{1}{2^{2(K-1)}r^2}
\sum_{l=1}^{2^K} \bm \lambda_l^\prime \bm \lambda_l \underbrace{\frac{1}{2^K-1} \sum_{i \in \psi_j} \{ Y_i(\bm z_l) - \bar Y_{j\cdot}(\bm z_l) \}^2}_{S_j^2(\bm z_l)}
- \frac{1}{2^K(2^K-1)r^2} \sum_{i \in \psi_j } (\bm \tau_i - \bm \tau_{j\cdot})(\bm \tau_i - \bm \tau_{j\cdot})^\prime.$$ This implies that $$\begin{aligned}
\label{eq:cov-m-proof}
\mathrm{Cov} ( \hat{\bm \tau}_\textrm{M} )
& = & \frac{1}{r^2} \sum_{j=1}^r \mathrm{Cov} ( \hat{\bm \tau}_{j\cdot} ) \nonumber \\
& = &
\frac{1}{2^{2(K-1)}r^2}
\sum_{l=1}^{2^K} \bm \lambda_l^\prime \bm \lambda_l \sum_{j=1}^r S_j^2(\bm z_l)
- \frac{1}{2^K(2^K-1)r^2} \sum_{j=1}^r \sum_{i \in \psi_j } (\bm \tau_i - \bm \tau_{j\cdot})(\bm \tau_i - \bm \tau_{j\cdot})^\prime.\end{aligned}$$ To prove the equivalence between and , simply note that $$(2^K - 1) \sum_{j=1}^r S_j^2(\bm z_l)
+ 2^K \sum_{j=1}^r \{ \bar Y_{j\cdot}(\bm z_l) - \bar Y(\bm z_l)\}^2
= (N-1) S^2(\bm z_l)$$ and $$\sum_{j=1}^r \sum_{i \in \psi_j } (\bm \tau_i - \bm \tau_{j\cdot})(\bm \tau_i - \bm \tau_{j\cdot})^\prime
+ 2^K \sum_{j=1}^r (\bm \tau_{j\cdot} - \bm \tau)(\bm \tau_{j\cdot} - \bm \tau)^\prime
= \sum_{i=1}^N (\bm \tau_i - \bm \tau)(\bm \tau_i - \bm \tau)^\prime.$$ The proof is complete.
We discuss a special case before moving forward. When $K=1,$ we have the classic treatment-control studies, and label the treatment and control as $+1$ and $-1,$ respectively. We are interested in the difference-in-mean estimator $$\hat {\tau}_{\textrm{MP}} = \frac{1}{r} \sum_{j=1}^r \{ Y_j^{\textrm{obs}}(+1) - Y_j^{\textrm{obs}}(-1) \}.$$ Denote $\psi_j = \{j_1, j_2\}.$ [@Imai:2008] (p. 4861, Eq. (8)) derived the variance of $
\hat {\tau}_{\textrm{MP}}
$ as $$\label{eq:var-m-1}
\mathrm{Var} ( \hat {\tau}_{\textrm{MP}} ) =
\frac{1}{4r^2} \sum_{j=1}^r
\{
Y_{j_1}(+1) - Y_{j_2} (-1) - Y_{j_2} (+1) + Y_{j_1} (-1)
\}^2.$$ As a validity check, Proposition \[prop:cov-m\] reduces to when $K=1.$ We leave the proof to the readers.
We discuss the estimation of $
\mathrm{Cov} ( \hat{\bm \tau}_\textrm{M} ),
$ because Lemma \[lemma:cov-c\] does not apply for matched-pair factorial designs. Inspired by [@Imai:2008], we propose the following estimator: $$\label{eq:cov-m-est}
\widehat{\mathrm{Cov}} ( \hat{\bm \tau}_\textrm{M} )
= \frac{1}{r(r-1)}
\sum_{j=1}^r
( \hat{\bm \tau}_{j\cdot} - \hat{\bm \tau}_\textrm{M} )( \hat{\bm \tau}_{j\cdot} - \hat{\bm \tau}_\textrm{M} )^\prime.$$
\[prop:cov-m-est-bias\] The bias of the covariance estimator in is $$\mathrm{E}
\left\{
\widehat{\mathrm{Cov}} ( \hat{\bm \tau}_\textrm{M} )
\right\}
- \mathrm{Cov} ( \hat{\bm \tau}_\textrm{M} )
=
\frac{1}{r(r-1)}\sum_{j=1}^r (\bm \tau_{j\cdot} - \bm \tau) (\bm \tau_{j\cdot} - \bm \tau)^\prime.$$
The proof is a basic maneuver of the expectation and covariance operators. First, by and the joint independence of $
\hat{\bm \tau}_{j\cdot}
$’s, $$\mathrm{Cov} ( \hat{\bm \tau}_\textrm{M} )
=
r^{-2} \sum_{j=1}^r \mathrm{Cov} ( \hat{\bm \tau}_{j\cdot} ).$$ Therefore by , $$\begin{aligned}
r(r-1)
\mathrm{E}
\left\{
\widehat{\mathrm{Cov}} ( \hat{\bm \tau}_\textrm{M} )
\right\}
& = &
\sum_{j=1}^r
\mathrm{E}
(
\hat{\bm \tau}_{j\cdot} \hat{\bm \tau}_{j\cdot}^\prime
)
-
r\mathrm{E}
(
\hat{\bm \tau}_\textrm{M} \hat{\bm \tau}_\textrm{M}^\prime
)
\\
& = &
\sum_{j = 1}^r \mathrm{Cov}(\hat{\bm \tau}_{j\cdot})
+
\sum_{j=1}^r \bm \tau_{j\cdot} \bm \tau_{j\cdot}^\prime
- r\mathrm{Cov} (\hat{\bm \tau}_\textrm{M})
-
r \bm \tau \bm \tau^\prime
\\
& = &
r(r-1) \mathrm{Cov} (\hat{\bm \tau}_\textrm{M})
+
\sum_{j=1}^r
(\bm \tau_{j\cdot} - \bm \tau) (\bm \tau_{j\cdot} - \bm \tau)^\prime.\end{aligned}$$
Proposition \[prop:cov-m-est-bias\] implies that the estimator of $
\mathrm{Cov} ( \hat{\bm \tau}_\textrm{M} )
$ is also “conservative.” We leave it to the readers to prove that for treatment-control studies, Proposition \[prop:cov-m-est-bias\] reduces to the corresponding results in [@Imai:2008] (p. 4862, Prop. 2, Part 1).
Discussions and concluding remarks {#sec:conclusion}
==================================
For treatment-control studies, [@Imai:2008] compared the variance formulas for the complete-randomization and matched-pair estimators, and derived the condition under which pair-matching leads to precision gains. For general factorial designs, analogous comparisons can be made between and . However, to our best knowledge, intuitive closed-form expressions might not be available without additional assumptions on the potential outcomes.
Let $K=2$ and $N = 8.$ We define the potential outcomes in Table \[tab:1\], in a way that there are no main/interactive effects on any units.
Unit $Y_i(-1, -1)$ $Y_i(-1, +1)$ $Y_i(+1, -1)$ $Y_i(+1, +1)$
------ --------------- --------------- --------------- ---------------
a 1 1 1 1
b 2 2 2 2
c 3 3 3 3
d 4 4 4 4
e 5 5 5 5
f 6 6 6 6
g 7 7 7 7
h 8 8 8 8
: A $2^2$ factorial design with 8 units. Each row consists of the index of a specific unit, and the four corresponding potential outcomes.
\[tab:1\]
Under complete randomization with two units assigned to each treatment, the covariance matrix can be calculated using : $$\mathrm{Cov} ( \hat{\bm \tau}_\textrm{C} )
=
\left(
\begin{array}{cccc}
0 & 0 & 0 & 0 \\
0 & 3 & 0 & 0 \\
0 & 0 & 3 & 0 \\
0 & 0 & 0 & 3
\end{array}
\right).$$ We consider two pair-matching mechanisms:
1. $\psi_1 = \{ \mathrm{a, b, c, d} \},$ $\psi_2 = \{ \mathrm{e, f, g, h} \};$
2. $\psi_1 = \{\mathrm{a, b, g, h} \},$ $\psi_2 = \{ \mathrm{c, d, e, f} \}.$
For the two pair-matching mechanisms, by the corresponding covariance matrices are $$\mathrm{Cov} ( \hat{\bm \tau}_{\textrm{M}_{1}} )
=
\left(
\begin{array}{cccc}
0 & 0 & 0 & 0 \\
0 & 5/6 & 0 & 0 \\
0 & 0 & 5/6 & 0 \\
0 & 0 & 0 & 5/6
\end{array}
\right),
\quad
\mathrm{Cov} ( \hat{\bm \tau}_{\textrm{M}_{2}} )
=
\left(
\begin{array}{cccc}
0 & 0 & 0 & 0 \\
0 & 7/2 & 0 & 0 \\
0 & 0 & 7/2 & 0 \\
0 & 0 & 0 & 7/2
\end{array}
\right),$$ respectively. Compared to complete randomization, pair-matching may lead to precision gain or loss, depending on the “goodness” of the pairing.
There are multiple future directions based on our current work. First, we may compare the precisions of the complete-randomization and matched-pair estimators under certain mild restrictions on the potential outcomes. Second, it is possible to unify the randomization-based and regression-based inference frameworks, as pointed out by [@Samii:2012] and [@Lu:2016a]. Third, additional pre-treatment covariates may shed light on the pair-matching mechanism, and help sharpen our current analysis.
Acknowledgements {#acknowledgements .unnumbered}
================
The first author thanks Professor Tirthankar Dasgupta at Rutgers University and Professor Peng Ding at University California at Berkeley, for their early educations on causal inference and experimental design. We thank the Co-Editor-in-Chief and an anonymous reviewer for their thoughtful comments, which have substantially improved the presentation of this paper.
[^1]: Address for correspondence: Jiannan Lu, One Microsoft Way, Redmond, Washington 98052-6399, U.S.A. Email: `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We pose a problem called “broadcasting Holevo-information”:\
given an unknown state taken from an ensemble, the task is to generate a bipartite state transfering as much Holevo-information to each copy as possible.
We argue that upper bounds on the average information over both copies imply lower bounds on the quantum capacity required to send the ensemble without information loss. This is because a channel with zero quantum capacity has a unitary extension transfering at least as much information to its environment as it transfers to the output.
For an ensemble being the time orbit of a pure state under a Hamiltonian evolution, we derive such a bound on the required quantum capacity in terms of properties of the input and output energy distribution. Moreover, we discuss relations between the broadcasting problem and entropy power inequalities.
The broadcasting problem arises when a signal should be transmitted by a time-invariant device such that the outgoing signal has the same timing information as the incoming signal had. Based on previous results we argue that this establishes a link between quantum information theory and the theory of low power computing because the loss of timing information implies loss of free energy.
author:
- |
Dominik Janzing and Bastian Steudel\
\[1ex\] [Institut für Algorithmen und Koginitive Systeme,]{}\
[Universität Karlsruhe,]{}\
[Am Fasanengarten 5,]{}\
[76 131 Karlsruhe, Germany]{}
date: 'September 21, 2006'
title: A Quantum Broadcasting Problem in Classical Low Power Signal Processing
---
Introduction
============
Quantum information theory and the theory of low-power processing are currently quite different scientific disciplines. Even though future low power computers will operate more and more on the nanoscale and therefore in the quantum regime (e.g. single electron transistors, spintronic networks [@Awschalom]), superpositions of logically different states being crucial for quantum computing [@NC], are not intended to occur in low-power computing devices.
On the other hand, quantum computing research is little interested in issues of low power processing. The control of quantum systems involves large laboratory equipment and even power consumption rates for logical operations that are in the magnitude of usual classical chips seem currently to be out of reach.
To understand limitations of low-power information processing it is useful to construct theoretical models of computers which process information without consuming energy, i.e., the process is implemented in an energetically closed physical system. In our opinion, discussions on fundamental issues like bounds on power consumption require a [*quantum*]{} theoretical description. Interesting quantum models of computers being closed physical systems can be found in Refs. [@Benioff; @Feynman:85; @Margolus:90; @Ergodic; @ErgodicQutrits]. Remarkably, it is common to all these models that the synchronization is based upon some propagating wave or particle and that the quantum uncertainty of its position leads to an ill-defined logical state of the computer. In other words, the clock is entangled with the data register. It seems as if the clocking issue brings some aspects of [*quantum*]{} information theory into the field of low power computing. This is not surprising for the following reason: the states of a quantum mechanical system have a consistent [*classical*]{} description only if the attention is restricted to a set of mutually commuting density matrices. But the Hamiltonian dynamics automatically generated non-commuting density matrices from a given one. Hence the dynamical aspect makes it necessary to include quantum superpositions into the description. Note that this is also in the spirit of Hardy’s paper “Quantum theory from five reasonable axioms” [@Hardy], saying that every statistical theory that satisfies some very natural axioms is quantum, as soon as it makes [ *continuous reversible*]{} dynamical evolution possible.
If signal propagation in future low-power devices takes place in a system being (approximately) thermodynamically closed it must be described by a quantum Hamiltonian evolution. The idea of this article is that processing such signals leads to quantum broadcasting problems for two reasons:
First, it is a natural problem to [*distribute*]{} signals (like clock signals) to several devices. The timing information carried by a signal whose quantum state is a density operator within a family of non-commuting states cannot be considered as classical information, its distribution is therefore some kind of broadcasting problem. The results in [@clock; @viva2002] indicate that no-broadcasting theorems are expected to get relevant for the distribution when the signal energy is reduced to a scale where quantum energy-time uncertainty becomes the limiting factor for the accuracy of clocking.
The second reason why broadcasting problems come into play is more subtle. If such a clocking signal enters a device and triggers the transmission of an output signal we may desire that the output should have as much timing information as the input (in a sense that will be further specified later). Whether channels with zero quantum capacity are able to satisfy this requirement is a question that is linked to quantum broadcasting.
The intention of this article is to describe a special kind of broadcasting problem. In contrast to the usual setting [@NoBroadcast], the task is not to obtain output states that are close to the inputs. The problem is to broadcast the [*Holevo-information*]{} of an ensemble of non-commuting quantum density matrices such that each party gets almost the same amount of Holevo-information as the original ensemble possessed. The use of entropy-like information measures makes it possible to draw connections to thermodynamics.
In this paper, the ensemble of non-commuting states will always be given by the Hamiltonian time evolution of a given state. Even though the problem of broadcasting Holevo-information makes also sense for general ensembles, time evolution is the most obvious way how non-commuting ensembles occur in devices that are not designed to do quantum information processing. It seems to be hard to derive general bounds on the information loss of each copy when the Holevo-information is broadcast. Thus, we will only conjecture that it is not possible for non-commuting ensembles to get full Holevo-information for both copies. The intention of this article is therefore rather to pose the broadcasting problem and show its relevance than to solve it. However, for pure input states we will give one lower bound on the loss that depends on the energy distribution of input and output signals.
The paper is organized as follows. In Section \[TimeIn\] we introduce time-invariant signal processing devices and argue that in this setting timing information is a resource that can never be increased. In Section \[Broad\] we formally state the problem of broadcasting Holevo-information in the general case and in the case of timing information. In Section \[PureProd\] we argue that the broadcasting problem leads to the question how the Holevo-information of an ensemble of bipartite states is related to the information of the ensembles of the corresponding reduced states. We discuss this information deficit for the special case of pure product states where the problem is related to the entropy power inequalities of classical information theory. In Section \[PureEnt\] we derive a bound on the information deficit that depends on the energy spectral measure of input and output signal. In Section \[Class\] we show that the results imply lower bounds on the quantum capacity required for lossless transmission of signals having small energy uncertainty in a time-covariant way. Section \[Sec:Thermo\] derives lower bounds on the loss of free energy implied by the loss of timing information caused by a channel with too little quantum capacity. This describes an even tighter link between quantum information theory and the theory of low-power signal processing.
Quantum model of time-invariant signal processing devices {#TimeIn}
=========================================================
As already stated, the problem of transmitting non-commuting ensembles of quantum states arises most naturally for ensembles that are given by the Hamiltonian time evolution of a given state. Such an ensemble may, for instance, describe the density matrix of a propagating signal before or after it is processed by the device. If all clocking signals that enter a given device are included into the formal description, the quantum operation mapping the input onto the output is [*time-invariant*]{}. As we will describe below, such a device cannot increase the timing information. The latter is therefore considered as a resource. The idea that devices with non-zero quantum capacity seem to deal with this resource more carefully than classical channels is essential for this article.
Now we introduce the abstract description of time-invariant devices. Here a device may be a transistor, an optical element or some other system with input and output signals. The signal may, for instance, be an electric pulse, a light pulse, or an acoustical signal. We consider it as a physical system with some Hilbert space ${{\cal H}}$ and the state is a density operator $\rho$ acting on ${{\cal H}}$. For the examples mentioned above, the space ${{\cal H}}$ will typically be infinite dimensional since one may e.g. think of position degrees of freedom that are encoded into $\rho$. Before and after the signal is processed in the device its free time evolution is generated by a Hamiltonian $H$ (i.e. a densely defined self-adjoint operator on ${{\cal H}}$) and reads $$\label{Motion}
\alpha_t(\rho):=e^{-iHt} \rho\, e^{iHt}\,.$$ We assume that input state $\rho$ and its output $G(\rho)$ are related by some completely positive trace-preserving map $G$ satisfying the covariance condition $$\label{TCov}
G(\alpha_t(\rho))=\alpha_t(G(\rho)) \hspace{2cm} \forall \rho\,.$$ There are rather different situations where the covariance condition is satisfied. One example would be if the interactions between signal and device are weak. A more interesting justification is the following. Consider a signal propagating towards the device by its own autonomous Hamiltonian time evolution until it begins to interact with the latter. Then it leaves the device (as a possibly modified signal) and as soon as the interaction with the device is negligible it is again subjected to its Hamiltonian only. Such a process should be considered as a quantum stochastic analogue of a scattering process (see [@TimeCovariant] for details) and the time covariance condition (\[TCov\]) is then a generalization of the statement that the $S$-“matrix” of a scattering process commutes with the free Hamiltonian evolution of the incoming and outgoing particle [@LP67]. Note that the existence of an unitary $S$ operator would require devices which preserve the purity of the input.
In [@TimeCovariant] we have given a quite explicit description of the set of CP maps satisfying this covariance condition. Here it is more interesting to discuss the implications of covariance. We first rephrase the definition of timing information used in [@Referenz] (see also [@GroupCovariantThermo] for a more general context).
Recall that the Holevo-information of an ensemble of quantum states $\rho_x$ with probability measure $p$ (denoted by $\{p(x),\rho_x\}_x$) is defined by [@NC] $${{\cal I}}:=S\Big(\int \rho_x \,d p(x) \Big)-\int S(\rho_x) \,dp(x)\,,$$ where the measure-theoretic integral reduces to sums when $p$ is supported by a countable set of points. Here $$S(\rho)=-tr(\rho \log \rho)$$ is the von-Neumann entropy and the base of the logarithm remains unspecified. In the sequel we will measure entropy in bits or nats since sometimes one unit is more natural and sometimes the other.
Timing information refers to a specific ensemble, namely the orbit with respect to a unitary one-parameter group:
${}$\
\[Tim\] Let $\rho$ be the state of a quantum system whose Hamiltonian $H$ has discrete spectrum. Then its timing information is defined as $$\label{EqTim}
{{\cal I}}:=S(\overline{\rho})-S(\rho)\,,$$ where $\overline{\rho}$ denotes the time average $$\overline{\rho}:=\lim_{T\to\infty}
\frac{1}{T}\int_0^T \alpha_t(\rho) dt =\sum_x R_x \rho R_x \,,$$ $\alpha_t$ is defined as in eq. (\[Motion\]) and $(R_x)$ is the family of spectral projections (with eigenvalues $x$ corresponding to the system Hamiltonian. For pure states $\rho=|\psi\rangle \langle \psi|$ we have $S(\rho)=0$. Thus, ${{\cal I}}$ is the entropy of $\overline{\rho}$ which is then exactly the entropy of a classical random variable $X$ describing the distribution of energy values with $P(X=x):=\langle \psi |R_x|\psi\rangle$.
Note that it is a well-known question to what extent information on reference frames in time or space requires quantum communication or profits from it and to which degree shared reference frames are resources that are comparable to shared quantum states [@Refefficient; @Burgh; @GroverRudolph; @Enk; @GroupCovariantThermo; @Decoherence-full]. In this article we want to understand to what extent timing information should be considered as quantum information by exploring the information loss occurring when it is copied. In [@clock] we have derived lower bounds on the loss of timing information in terms of Fisher-information for the broadcasting problem. To our knowledge, no results in terms of Holevo-information can be found in the literature.
Broadcasting timing information {#Broad}
===============================
Before we pose the problem of broadcasting timing information (which we have motivated from the time-covariant transmission of signals) we first state the more general problem of “broadcasting Holevo-information”. It is defined as follows:
${}$\
Given an ensemble $\{p(x),\rho (x)\}_x$ of quantum states acting on some Hilbert space ${{\cal H}}$. Let ${{\cal I}}$ be its Holevo-information. Find an optimal broadcasting map in the following sense:
Let ${{\cal H}}_A$ and ${{\cal H}}_B$ be some arbitrary additional Hilbert spaces and $G$ be a completely positive trace-preserving operation from the density operators on ${{\cal H}}$ on the density operators acting on ${{\cal H}}_A\otimes {{\cal H}}_B$. Let ${{\cal I}}_A,{{\cal I}}_B$ denote the Holevo-information of the ensembles given by the reduced states $tr_B(G(\rho_x))$ and $tr_A(G(\rho_x))$, respectively.
Maximize the average information $$\frac{1}{2}({{\cal I}}_A+{{\cal I}}_B)$$ over all ${{\cal H}}_A\otimes {{\cal H}}_B$ and $G$ such that it gets as close to ${{\cal I}}$ as possible.
We call $$\label{IDiff}
\Delta:={{\cal I}}-\frac{1}{2}({{\cal I}}_A+{{\cal I}}_B)$$ the broadcasting loss of a given broadcasting operation. Let $\Delta_{\min}$ be the minimal loss over all broadcasting operations for a given ensemble. In the context of timing information we will also use the terminology "$\Delta_{\min}$ of a state $\rho$” when actually refering to the information loss of the ensemble $\{\alpha_t(\rho)\}_{t\in [0,\tau)}$ with uniform probability distribution over the whole time period.
Due to the monotonicity of the Holevo-information of an arbitrary ensemble with respect to CP maps [@Referenz] we certainly have ${{\cal I}}_1\leq {{\cal I}}$ and ${{\cal I}}_2\leq {{\cal I}}$. It is natural to conjecture that ${{\cal I}}={{\cal I}}_1={{\cal I}}_2$ can only be achieved if all density matrices commute, which is exactly the case where usual broadcasting is possible [@NoBroadcast]. It is furthermore obvious that there are maps that provide both parties with the [*accessible*]{} information [@NC] by applying a measurement to the input state and sending mutually orthogonal quantum states representing the results to both parties.
Now we apply the definition of broadcasting to an ensemble given by the time orbit $(\rho_t)_{t\in [0,\tau]}$ of a dynamical evolution with period $\tau$ with uniform distribution over the whole interval. Then the task is to optimally broadcast the timing information in the sense of Definition \[Tim\]. Note that information [*differences*]{} like that one in eq. (\[IDiff\]) maybe well-defined for systems with continuous spectrum where the timing information itself is infinite. By appropriate limits, one could therefore define the question on the information [*loss*]{} in broadcasting operations also for systems possessing no time average state.
To give an impression on the problem of broadcasting timing information we consider the phase-covariant cloning of an equatorial qubit state, i.e., a state $$|\psi_t\rangle:= \frac{1}{\sqrt{2}}(|0\rangle+e^{it} |1\rangle)$$ with unknown $t\in [0,2\pi)$. In the usual quantum cloning problem one tries, for instance, to obtain two copies whose states get as close to the original as possible with respect to the Fidelity. As shown in [@BrussPhase] one can generate two copies as mixed states whose Bloch vectors point in the same direction as that of the original, but are shorter than the original by the factor $1/\sqrt{2}$. Thus, the density matrices of the copies have the eigenvalues $1/2 \pm 1/(2\sqrt{2})$. The Holevo-information of each copy is then given by the entropy of the time average (which is still one bit) minus the above binary entropy when inserting the above eigenvalues: $$\begin{aligned}
{{\cal I}}&=& 1+ \frac{1}{2}(1+\frac{1}{\sqrt{2}})
\log_2 \frac{1}{2}(1+\frac{1}{\sqrt{2}})
+\frac{1}{2}(1-\frac{1}{\sqrt{2}})
\log_2 \frac{1}{2}(1-\frac{1}{\sqrt{2}})\\
&\approx& 0.399 \hbox{ bit}\,.\end{aligned}$$ The information of the original was $1$. Here, even the sum of the amount of information over both copies is less than the original amount. In other words, the average information over both copies is even smaller than it was if we had given one party the original and the other an arbitrary state that is independent from the input.
Information deficit in pure product states and entropy power inequalities {#PureProd}
=========================================================================
In the following we will not explicitly consider the broadcasting operation that generates a bipartite state from the original. Since this operation can never increase the information we focus to the following problem: Given an ensemble of bipartite states, compare the Holevo-information of the two ensembles ${{\cal I}}_A$ and ${{\cal I}}_B$ defined by the restrictions to the subsystems to the information ${{\cal I}}$ of the joint system. Call ${{\cal I}}-({{\cal I}}_A+{{\cal I}}_B)/2$ the [*information deficit*]{}. In other words, the information deficit is the broadcasting loss if the broadcasting map is the identity and the original is already a bipartite state.
Remarkably, the determination of the deficit is non-trivial even when the bipartite state is a product state. Given the state $$|\psi_A\rangle \otimes |\psi_B\rangle \in {{\cal H}}_A \otimes {{\cal H}}_B\,,$$ where each subsystem is subjected to its own Hamiltonian $H_A$ and $H_B$, respectively. We may assume without loss of generality that both Hamiltonians are diagonal and non-degenerate (since we restrict the attention to the time orbits of each state). The distribution of energy values in the state $|\psi_A\rangle\otimes |\psi_B\rangle$ defines a joint distribution of two stochastically independent classical random variables $X,Y$ by $$\label{energyMeasure}
P(X=x,Y=y):= \langle \psi_A |R_x|\psi_A\rangle
\langle\psi_B|
Q_y |\psi_B\rangle\,,$$ where $R_x$ is defined as in Definition \[Tim\] and $Q_y$ similarly. Since $H_A\otimes {\bf 1} +{\bf 1} \otimes H_B$ is the Hamiltonian of the joint system, its timing information is given by $$\label{JointI}
{{\cal I}}=S(X+Y) \,,$$ where we have decided to use the same symbol for the entropy of classical random variables as for the von-Neumann entropy of quantum states. The subsystem timing information is given by $$\label{SubIn}
{{\cal I}}_A= S(X) \,\,\,\,\,\hbox{ and } \,\,\,\,\,{{\cal I}}_B=S(Y)\,.$$ Note that it is a well-known problem in classical information theory to relate the entropy of the distributions of two independent random variables to the entropy of their sum since it addresses the question how the entropy of a real-valued signal changes when subjected to an additive noise. We rephrase the following result that applies to continuous distributions. For probability densities $P(X)$ the continuous entropy is defined by $$S(X)=-\int P(x) \ln P(x) dx +c \,,$$ with an unspecified constant $c$. For two independent random variables, i.e., when their density satisfies $P(x,y)=P(x)P(y)$, we have the entropy power inequality [@Stam] $$e^{2 S(X+Y)} \geq e^{2S(X)} + e^{2S(Y)}
\,,$$ and hence $$\begin{aligned}
2 S(X+Y)&\geq & \ln \Big(\frac{1}{2}(2e^{2S(X)}+2e^{2S(Y)})\Big)\\
&\geq& \frac{1}{2} \Big( \ln 2e^{2S(X)} +\ln 2e^{2S(Y)}\Big)\\
&=& \ln 2 + S(X)+S(Y)\,, \end{aligned}$$ where the second inequality follows from the concavity of the logarithm. Assuming that the spectral measures of $H_A$ and $H_B$ are sufficiently distributed over many energy eigenvalues we can approximate the discrete entropy with the continuous expression for appropriate densities. After using eqs. (\[JointI\]) and (\[SubIn\]) we obtain $${{\cal I}}\geq \frac{1}{2}(\ln 2 +{{\cal I}}_A +{{\cal I}}_B)
\,.$$ Note that $\ln 2$ corresponds exactly to one bit of information since the entropy power inequality refers to entropy measured in natural units. We conclude that for continuous spectrum and product states the timing information of the joint system is at least half a bit more than the average timing information over both systems.
Information deficit for pure entangled states {#PureEnt}
=============================================
To estimate the information deficit for entangled states we will also use the joint distribution of $X$ and $Y$ on ${{\mathbb{R}}}^2$ given by $$\label{JM}
P(X=x,Y=y):=tr(\rho (R_x\otimes Q_y))\,,$$ with the spectral projections $R_x$ and $Q_y$. If the bipartite system is in an entangled state, eq. (\[SubIn\]) is no longer true. Moreover, we cannot assume that both Hamiltonians are “without loss of generality” non-degenerate since the reduced states may be mixed even within a specific degenerate energy eigenspace. However, eq. (\[JointI\]) still holds for pure states. We replace eq. (\[SubIn\]) by $${{\cal I}}_A=S(\overline{\rho_A})-S(\rho_A) \,,$$ where $\rho_A$ denotes the reduced state on system $A$ and obtain ${{\cal I}}_B$ in a similar way. To derive upper bounds on the timing information of the subsystems we need the following Lemma.
${}$\
\[Ave\] Let $(R_j)_j$ be a complete family of orthogonal projections defining a measurement and $\sigma$ be an arbitrary quantum state. Let $S(p)$ be the Shannon entropy of the outcome probabilities $p_j:=tr(R_j\sigma)$. Then we have $$S(\sum_j R_j \sigma R_j) \leq S(\sigma)+S(p)\,.$$
Proof: The statement is equivalent to $$\label{MesAv}
\sum_j p_j S\Big( \frac{1}{p_j} R_j\sigma R_j \Big) \leq S(\sigma)$$ Let $\sigma=\sum_i q_i \sigma_i$ be a decomposition of $\sigma$ into pure states. We can consider $S(\sigma)$ as the Holevo-information of the ensemble $\{q_i,\sigma_i\}_i$. Then the left hand side of eq. (\[MesAv\]) is equal to the Holevo-information of the ensemble after the measurement has been applied. It can certainly be not greater than the Holevo-information of the original ensemble [@Referenz]. $\Box$
For our derivation of an upper bound on the information of the subsystems the following Lemma will be crucial.
${}$\
\[RedInf\] Let $\rho$ be a (possibly mixed) state on a bipartite system. Then the timing information of $A$ and $B$ satisfies $${{\cal I}}_A \leq S(X|Y), \hspace{1cm} {{\cal I}}_B \leq S(Y|X) \,,$$ respectively, where the joint distribution of $X$ and $Y$ is defined by eq. (\[JM\]).
Proof: By Definition \[Tim\] the timing information of system $A$ is given by $${{\cal I}}_A=S(\overline{\rho_A})-S(\rho_A)\,.$$ We decompose $\rho_A$ into $$\rho_A=\sum_y p(y) \rho_{A,y}\,,$$ where $\rho_{A,y}$ denotes the conditional state given that we had measured the energy value $y$ on system $B$. Since ${{\cal I}}_A$ is the Kullback Leibler distance between $\overline{\rho_A}$ and $\rho_A$ (see [@clockentropy]) it is convex and we get $${{\cal I}}_A \leq \sum_y p(y) (S(\overline{\rho_{A,y}})-S(\rho_{A,y}) )\,.$$ For each specific value $y$ of $Y$ $$S(\overline{\rho_{A,y}})-S(\rho_{A,y})\leq S(X|y)$$ holds due to Lemma \[Ave\]. Taking the convex sum of this inequality over all $y$ with weights $p(y)$ completes the proof. $\Box$
Note that there are conditions known [@Johnson], where the joint probability density of two [*dependent*]{} random variables satisfies the entropy power inequality $$e^{2S(X+Y)} \geq e^{2S(X|Y)}+e^{2S(Y|X)}\,.$$ Under such conditions we obtain the same lower bound on the information deficit as in Section \[PureProd\].
In the general case we have to use other methods to derive more explicit bounds from the bounds of Lemma \[RedInf\]. For doing so, we will need the following lemma.
${}$ \[QKlassischRed\] The information deficit of a bipartite system being in a pure state satisfies $$\Delta \geq \frac{1}{2}\Big(I(X:X+Y) +I(Y:X+Y)\Big)=S(X+Y)-\frac{1}{2}\Big(S(X|Y)+
S(Y|X)\Big)\,,$$ where $I(.,.)$ denotes the mutual information between classical random variables [@Cover].
Proof: Note that the equation ${{\cal I}}=S(X+Y)$ holds also for pure entangled states. Using Lemma \[RedInf\] we obtain $$\begin{aligned}
2{{\cal I}}- {{\cal I}}_A -{{\cal I}}_B &\geq& 2 S(X+Y)-S(X|Y)-S(Y|X) \\
&=&2 S(X+Y)-S(X+Y|Y)-S(X+Y|X)\\
&=&
I(X+Y:Y)+I(X+Y:X)\,.\,\,\,\,\, \Box\end{aligned}$$
It is possible to derive bounds on the information loss based on Lemma \[QKlassischRed\], since the term on the right hand vanishes only in the trivial case $S(X+Y)=0$ in which the joint system contains no timing information at all. To show this we observe that there is no joint measure where $X$ and $Y$ are both uncorrelated to $X+Y$. This is seen from $$\label{CV}
C(X,X+Y)+C(Y,X+Y)=V(X+Y)\,,$$ where $C(.,.)$ denotes the covariance and $V(.)$ the variance. However, to derive lower bounds on the mutual information based on these covariance terms requires additional assumptions on the distribution. We will deal with this point later.
In order to apply the bounds of Lemma \[QKlassischRed\] it can be convenient to relate them to other information-theoretic quantities:
${}$\
\[Con\] Let $X$ and $Y$ be two real-valued random variables and $P$ the corresponding joint distribution on ${{\mathbb{R}}}^2$ with discrete support. Let $P_{-X}$ and $P_{X+Y}$ denote the marginal distribution for $-X$ and $X+Y$, respectively. Denote the convolution of both by $P_X *P_{X+Y}$. Then we have $$\label{PY}
I(X:X+Y) \geq K(P_{Y} || P_{-X} * P_{X+Y})$$ and $$\label{PX}
I(Y:X+Y) \geq K(P_{X} || P_{-Y} * P_{X+Y})\,.$$ Moreover, we have the symmetrized statement $$\begin{aligned}
\label{PSymm}
&&I(X:X+Y)+I(Y:X+Y) \nonumber \\
&\geq &
K\Big(\frac{1}{2}(P_X +P_Y)\Big|\Big| \frac{1}{2}(P_{-X}+P_{-Y}) * P_{X+Y}\Big)\,.\end{aligned}$$
Proof: We define measures on ${{\mathbb{R}}}^2$ by $$Q(X=a,Y=b):=P(X+Y=a+b)P(Y=b)$$ and $$R(X=a,Y=b):=P(X+Y=a+b)P(X=a)\,.$$ Then we can rewrite the mutual information on the left hand side as Kullback-Leibler distances: $$I(X:X+Y)=K(P||R)$$ and $$I(Y:X+Y)=K(P||Q)\,.$$ Due to the monotonicity of relative entropy distance under marginalization [@OhyaPetz] we have $$K(P||Q) \geq K(P_X||Q_X)$$ where $P_X$ and $Q_X$ denote the marginal distribution of $X$ according to $P$ and $Q$, respectively, i.e., $Q_X(X=a):=Q(X=a)$. Similarly $$K(P||R) \geq K(P_Y||R_Y)\,.$$ We have $$\begin{aligned}
Q(X=a)&=&\sum_b Q(X=a,Y=b)\\&=&\sum_b P(X+Y=a+b)P(Y=b)\\
&=&\sum_c P(X+Y=c) P(Y=c-a)\\ &=&\sum_c P(X+Y=c) P(X=c-a)\,.\end{aligned}$$ Hence the marginal distribution $Q_X$ of $Q$ is the convolution product $P_{X+Y}* P_{-X}$ and the marginal distribution $R_Y$ of $R$ is the product $P_{X+Y}*P_{-Y}$. This proves inequalities (\[PY\]) and (\[PX\]).
We obtain the symmetrized statement from the convexity of relative entropy distance [@Cover]. $\Box$
After applying Lemma \[Con\] and and Lemma \[QKlassischRed\] we conclude:
${}$\
\[HauptE\] Given a pure state of a bipartite system $A\times B$. Let $P_X$, $P_Y$ and $P_{X+Y}$ denote the probability distributions for the energy of $A$, $B$ and $A\times B$, respectively. Then the difference between the joint timing information and the average information of the subsystems satisfies $$\begin{aligned}
\Delta &\geq& K(P_X||P_{-Y}*P_{X+Y})
+ K(P_Y||P_{-X}*P_{X+Y})\\
&\geq & K\Big(\frac{1}{2}(P_X+P_Y)\Big|\Big|
\frac{1}{2}(P_{-X}+P_{-Y})*P_{X+Y}\Big)\,.\end{aligned}$$
The intuitive content of Theorem \[HauptE\] is the following. If the energy uncertainty of $A$ and $B$ are both on the same scale as the uncertainty of $X+Y$, the convolution with $P_{X+Y}$ adds a non-negligible amount of uncertainty to $(P_X+P_{Y})/2$, which implies that the new distribution obtained by adding additional noise cannot be close to the original distribution of $X$.
It is often helpful to consider measures that are symmetric with respect to exchanging $X$ and $Y$, i.e., $P(X=x,Y=y)=P(X=y,Y=x)$. The following Lemma shows that lower bounds on $I(X:X+Y)+I(Y:X+Y)$ for symmetric joint measures automatically provide bounds for asymmetric measures:
${}$\
\[Symm\] Let $P$ be a joint distribution of $X$ and $Y$ and $\overline{P}$ its symmetrization $
\overline{P}:=(P+P')/2\,,
$ where $$P'(X=x,Y=y):=P(X=y,Y=x)\,.$$ Then we have $$I_P(X:X+Y) +I_P(Y:X+Y)\geq I_{\overline{P}}(X:X+Y)+
I_{\overline{P}}(Y:X+Y)\,,$$ where $I_P(.,.)$ refers to the mutual information induced by the measure $P$.
Proof: We write $$P(X=x,Y=y)=P(X=x|X+Y=x+y)P(X+Y=x+y)\,.$$ We obtain such a representation also for $P'$ by replacing only the conditional $P(X|X+Y)$ since the marginal distribution on $X+Y$ coincides for $P$ and $P'$. Then the Lemma follows already from the convexity of mutual information with respect to convex sums of conditionals with fixed marginals (Theorem 2.7.3 in [@Cover]). $\Box$.
A simple bound on the information deficit can be provided in terms of the fourth moments of the signal energies:
${}$\
\[E4\] Given a pure bipartite state on $A\times B$. Let $(\Delta E)^2$ denote the variance of the total energy and $\langle E^4_j \rangle$ denote the $4$th moment of the energy of system $j=A,B$ and $\langle E^4\rangle$ be the fourth moment of the total energy. Then the information deficit (measured in natural units) satisfies $$\Delta \geq
\frac{(\Delta E)^8}{64\,(\langle E_A^4\rangle + \langle E_B^4\rangle)
\langle E^4 \rangle}\,.$$
Proof: Let $P$, as above, be the discrete probability measure on ${{\mathbb{R}}}^2$ describing the energy distribution of the bipartite system. We begin by assuming that $P$ is symmetric (see Lemma \[Symm\]). Then we have $C(X,X+Y)=V(X+Y)/2$ (see eq. (\[CV\])). We define a measure $R$ as in the proof of Theorem \[HauptE\] and we can rewrite the covariance as $$C(X,X+Y)=\sum_{xy} x(x+y) (P(x,y)-R(x,y))$$ With $Z:=X+Y$ we have $$\begin{aligned}
\frac{1}{4}
V^2(X+Y)=C(X,Z)^2&=&\Big|\sum_{xz} xz \sqrt{(P(x,z-x)-R(x,z-x))} \\
&\times& \sqrt{(P(x,z-x)-R(x,z-x))}\Big|^2 \\ &\leq&
\sum_{xz} x^2z^2 |P(x,z-x)-R(x,z-x)| \\
&\times &
\sum_{xz} |P(x,z-x)-R(x,z-x)| \\ &\leq&
\sum_{xz} x^2 z^2 (P(x,z-x) +R(x,z-x))
\|P-R\|_1\\ & =&
(\langle X^2Z^2 \rangle +\langle X^2\rangle \langle Z^2\rangle
)\,\|P-R\|_1\\ &\leq&
2 \sqrt{\langle X^4\rangle \langle Z^4\rangle}\, \|P-R\|_1\,.\end{aligned}$$ From the first line to the second we have used the Cauchy Schwarz inequality which shows also $\langle X^2Z^2\rangle \leq \sqrt{\langle X^4\rangle \langle
Z^4\rangle}$ as well as $\langle X^2\rangle \leq
\sqrt{\langle X^4\rangle }$.
We recall the bound $$K(P||R)\geq \frac{1}{2}\|P-R\|^2$$ (see Lemma 12.6.1 in [@Cover]) for the relative entropy measured in natural units. Then we obtain $$\frac{1}{2}\|P-R\|^2_1 \geq
\frac{V(X+Y)^4}{128\,\langle X^4\rangle \langle (X+Y)^4\rangle} \,.$$ This implies $$I(X:X+Y)\geq
\frac{V(X+Y)^4}{128\,\langle X^4\rangle \langle (X+Y)^4\rangle}\,.$$ If we consider an asymmetric measure $P$ we have to symmetrize it first. Then we replace $\langle X^4\rangle $ with $(\langle X^4\rangle +\langle
Y^4\rangle) /2$ since the fourth moment of $Y$ with respect to the original measure $P$ coincides with the fourth moment of $X$ when calculated with respect to the reflected measure $P'(X=x,Y=y):=P(X=y,Y=x)$. Using Lemma \[QKlassischRed\] this proves the statement when replacing the statistical moments of $X, Y$, and $X+Y$ with the more physical terms $\langle E_A^4\rangle$, $\langle E_B^4\rangle$ and $\langle E^4\rangle$. $\Box$
Quantum capacity required for lossless transmission {#Class}
===================================================
In this section we will derive lower bounds on the quantum capacity required to transmit an ensemble with some fixed maximal information loss. The idea of the argument is the following. Assume that the timing information of $G(\rho)$ is exactly the same as that of $\rho$. Assume furthermore that $G$ has zero quantum capacity. This implies, roughly speaking, that $G$ can be modeled by a unitary that copies as much information to the environment as the amount of information that passes the channel. But if this would be the case we had perfect broadcast of Holevo-information, an operation that we consider unlikely to be possible for non-commuting ensembles like time orbits. To put this argument on a solid basis, we rephrase the following result of Devetak [@Devetak]. Recall that the private information capacity (see [@Devetak] for a formal definition) is the maximal number of encoded qubits per transmitted qubits, that two parties, the sender Alice and the receiver Bob, can asymptotically achieve in a protocol where a potential eavesdropper Eve, having access to the full environment of the channel, gets a vanishing amount of information. The following theorem relates the private information capacity to the information the environment obtains when the channel is represented by a unitary acting on the system and an abstract environment being in a pure state[^1].
${}$\
\[PrCC\] Let $G$ be a quantum channel mapping density operators acting on ${{\cal H}}$ to density operators acting on the same space. Let ${{\cal H}}_E$ be an additional Hilbert space thought of as the space of the environment. Moreover, let $U$ be a unitary acting on ${{\cal H}}\otimes {{\cal H}}_E$ and $|\phi\rangle\in {{\cal H}}_E$ be a state such that $$G(\rho)=tr_2(U(\rho\otimes |\phi\rangle \langle \phi|)U^\dagger)\,.$$ Let $\rho_x$ with $x\in {{\cal X}}$ be some finite family of input states (sent by Alice with probability $p(x)$) and $$\sigma_x:=U(\rho_x \otimes |\phi\rangle \langle \phi|)U^\dagger$$ be the corresponding joint states of the environment and the receiver’s (i.e Bob’s) system. Denote the restrictions to these subsystems by $\sigma_x^B$ and $\sigma_x^E$, respectively. Set $$I(X:B):= S( \sum_x p(x) \sigma^B_x) -p(x)\sum_x S(\sigma^B_x)\,.$$ and $I(X:E)$ similarly. Define the single copy private channel capacity by $$C_1(G) := \sup \{ I(X:B)-I(X:E) \}\,,$$ where the supremum is taken over all ensembles $(p(X),\rho_x)$. Let $G^{\otimes l}$ be the $l$-fold copy of $G$. Then the private channel capacity is given by $$C_p(G)=\lim_{l\to\infty} \frac{1}{l} C_1(G^{\otimes l})\,.$$
Certainly, we have $C_p(G) \geq C_1(G)$. This is seen by transmitting independently distributed product states through the copies of channels. We observe:
${}$\
\[InfLoss\] Let $\{p(x),\rho_x\}_x$ be an ensemble of quantum states with Holevo-information $$I(X:A)=S\Big(\sum_x p(x) \rho_x\Big) -\sum_x p(x) S(\rho_x)\,.$$ with minimal broadcasting loss $\Delta_{\min}$. Let $G$ be some channel with $$I(X:B)= S\Big(\sum_x p(x) G(\rho_x)\Big) -\sum_x p(x) S(G(\rho_x))\,.$$ Then the private channel capacity of $G$ can be bounded from below by $$C_p(G)\geq 2 \Big(\Delta_{\min} - (I(X:A) -I(X:B))\Big)\,.$$
Note that $I(X:A) -I(X:B)$ is the information loss caused by the channel because it is the difference between input and output Holevo-information. Given a bound for broadcasting the Holevo-information of the considerd ensemble, we have a lower bound on the quantum capacity to transmit them without loss.
Proof (of Theorem \[InfLoss\]): Given some unitary operation $U$ extending the channel $G$. We have $$\Delta := I(X:A)-\frac{1}{2}(I(X:B)+I(X:E)) \geq \Delta_{\min}$$ by definition of $\Delta_{min}$ and $$C_p(G)\geq I(X:B)-I(X:E)\,,$$ by Theorem \[PrCC\]. Then simple calculations yield the stated inequality. $\Box$
The theorem shows that for states with non-zero $\Delta_{\min}$ (which is probably every non-stationary state $\rho$) the covariant lossless transmission requires a channel with non-zero quantum capacity. Instead of deriving lower bounds on $\Delta_{\min}$, i.e., the minimum over all $\Delta$ we will use the bound from Theorem \[E4\] and only obtain bounds in terms of the fourth moments of the energy distribution. However, using this theorem is not straightforward for the following reason: Given some assumptions on the energy distribution of the input and output signals of a device we want to derive lower bounds on the quantum capacity required to transmit the signal without information loss. To this end, we use the unitary extension of the CP map formalizing the device because we have only derived bounds for [*pure*]{} bipartite states. However, the usual construction of the unitary extension uses an abstract environment Hilbert space where no “environment Hamiltonian” is specified. And, even worse, given that we had specified an arbitrary “environment Hamiltonian”, the unitary that models the channel could have lead to arbitrary energy distributions for system plus environment and we obtained no useful statements on the fourth moments.
The following Lemma shows that we can construct the unitary extension such that it is energy conserving in the constructed joint system. We have here considered a finite dimensional system for technical reasons.
${}$\
\[Extension\] Let $G$ be a completely positive trace-preserving map on the set of $d\times d$ density matrices that satisfies the covariance condition (\[TCov\]) with respect to the time evolution generated by a Hamiltonian $H$ acting on $\C^d$.
Then there is a (not necessarily finite dimensional) Hilbert space ${{\cal H}}_E$, a densely defined Hamiltonian $H_E$ on ${{\cal H}}_E$ with purely discrete spectrum and an eigenstate $|\phi\rangle$ of $H_E$ with eigenvalue $0$ such that the following condition holds:
There exists a unitary $U$ on $\C^n \otimes{{\cal H}}_E$ commuting with the extended Hamiltonian $H\otimes {\bf 1} + {\bf 1}\otimes H_E$ which satisfies $$G(\rho) = tr_2 (U (\rho\otimes |\phi \rangle \langle \phi|) U^\dagger)$$ for all density matrices $\rho$.
Proof: We assume without loss of generality that $H$ is diagonal with respect to the canonical basis. Let $$\label{Kr}
G(\rho)=\sum_{j=1}^k A_j \rho A_j^\dagger$$ be the Kraus representation of $G$ (see [@Kraus]). Define $\Sigma:=\{x-y \, | \, x,y \in {\tt spec(H)}\}$ where ${\tt spec}(H)$ denotes the spectrum of $H$. As shown in (eqs. (14) in [@TimeCovariant]) we can choose the Kraus operators such that for every $A_j$ there is some real number $\sigma_j \in \Sigma$ with $$\label{ComRel}
[H,A_j]= \sigma_j A_j\,.$$ In other words, the operator $A_j$ implements a shift of energy values by $\sigma_j$ in the sense that it maps eigenstates of $H$ with eigenvalue $\lambda$ onto states with energy $\lambda+\sigma_j$. The idea is to choose a unitary extension such that the energy shift caused by $A_j$ is compensated by the opposite shift in the environment. Thus, we define the Hamiltonian $H_E$ of the environment such that all values in $\Sigma$ occur as spectral gaps in $H_E$. Set ${{\cal H}}_E:=l^2({{\mathbb{Z}}})^{\otimes k}$ and $$H_E = \sum_{j=1}^k \sigma_j M_j\,,$$ where $M_j$ is the multiplication operator acting on the $j$th component $$M_j:={\bf 1}^{\otimes j-1} \otimes {\tt diag }(\dots,-1,0,1,\dots) \otimes {\bf 1}^{\otimes k-j}\,.$$ Let $$S_j:={\bf 1}^{\otimes j-1} \otimes S \otimes {\bf 1}^{\otimes k-j}$$ be the unitary left shift on $l^2({{\mathbb{Z}}})$ acting on the $j$th tensor component via $S|n\rangle:=|n-1\rangle$ for each $n\in {{\mathbb{Z}}}$. Define $$U:=\sum_{j=1}^k A_j \otimes S_j\,.$$ To see that $U$ is indeed unitary we consider basis states $$|l\rangle \otimes | {\bf z}\rangle\,,$$ where $l=0,\dots,d-1$ and ${\bf z}$ is in the $k$th fold cartesian product ${{\mathbb{Z}}}^{\times k}$. They are all mapped onto unit vectors because $\sum_j \langle l| A_j A_j^\dagger |l\rangle=1$. The images of different basis states are clearly mutually orthogonal whenever they correspond to different $k$-tuples ${\bf z}$. If they have ${\bf z}$ in common, they are also orthogonal since we obtain then the inner product $$\begin{aligned}
&&\sum_j \langle l|A_j A_j^\dagger |\tilde{l}\rangle
\langle z_1,\dots,z_j+1,\dots,z_k| z_1,\dots,z_j+1,\dots,z_k\rangle =\\
&&\sum_j \langle l|A_j A_j^\dagger |\tilde{l}\rangle
= \langle l| \tilde{l}\rangle =0\,.\end{aligned}$$ To see that $U$ commutes with the total Hamiltonian $H_T:=H\otimes {\bf 1} + {\bf 1} \otimes H_E$ we observe that for every eigenstate $|l\rangle$ of $H$ with eigenvalue $\lambda_l$ we have $$A_j|l\rangle =|\phi_{l,j}\rangle \,,$$ where $|\phi_{l,j}\rangle$ is some state with $$H|\phi_{\l,j}\rangle=(\lambda_l+\sigma_j)|\phi_{l,j}\rangle\,.$$ We have $$(A_j \otimes S_j) (|l\rangle \otimes |{\bf z}\rangle)=
|\phi_{l,j} \rangle \otimes |z_1,\dots,z_j-1,\dots,z_k\rangle \,,$$ which is also an eigenstate of $H_T$ for the eigenvalue $\lambda + \sum_j \sigma_j z_j$ as $|l\rangle \otimes |{\bf z}\rangle$ is. That is, $U$ maps energy basis states onto energy basis states with the same eigenvalues, i.e., it commutes with $H_T$. We can now choose $|\phi\rangle:=|{\bf 0}\rangle$ as the state of the environment. $\Box$
Note that the state $U(\rho \otimes |\phi\rangle \langle \phi |)U^\dagger$ appearing in the extension of Theorem \[Extension\] has the same energy distribution with respect to the extended Hamiltonian as $\rho$ has with respect to the original system Hamiltonian. This implies that the distribution of energy values in the joint state of system plus environment is given in terms of the distribution of input and output energies. Hence we may now apply our bounds on the information deficit to the problem of transmitting the states when only limited quantum capacity is available:
${}$\
\[ILC\] Given a covariant completely positive map $G$ with private channel capacity $C_p(G)$. Then the difference between the timing information of input and output satisfies $${{\cal I}}_{in} -{{\cal I}}_{out} \geq
\frac{(\Delta E_{in})^8}{64(9\langle E_{out}^4\rangle +
8\langle E_{in}^4\rangle)
\langle E_{in}^4 \rangle }
-\frac{1}{2} C_p(G) \,,$$ where $(\Delta E_{in})^2$ and $\langle E_{in}^4\rangle$ refer to the variance and the fourth moment of the incoming signal and similarly, $\langle E_{out}^4\rangle$ denotes the fourth moment of the outgoing signal.
Proof: Construct a unitary energy conserving extension of $G$ according to Lemma \[Extension\]. Let $E_{out}:=X$ denote the energy of the output signal and $Y$ the energy of the environment. This implies that $E_{in}:=X+Y$ is the initial energy. To get a bound for $\langle Y^4\rangle =\langle
(E_{in}-E_{out})^4\rangle$ we use $|E_{in}-E_{out}| \leq |E_{in}|+|E_{out}|$ and hence $(E_{in}-E_{out})^4 \leq 8 (E_{in}^4 +E_{out}^4)$. Then we obtain the statement using Theorem \[E4\]. $\Box$
Implications for the Energy Loss {#Sec:Thermo}
================================
In this section we want to explain why we expect the broadcasting problem to be specific to [*low-power*]{} devices. One reason is, certainly, that in current technology, information processing devices are not Hamiltonian systems. Since the system is not closed, a unitary description of the signal propagation is not justified. Furthermore, quantum broadcasting gets only relevant when the time inaccuracy of a clock signal is not dominated by classical noise of highly mixed density operators. In the latter case, the energy-time uncertainty is irrelevant. This is in agreement with the results in Ref. [@clock] showing (in terms of Fisher-information) that quantum bounds on broadcasting timing information get relevant when the signal energy times the considered timing accuracy is on the scale of $\hbar$. However, there is also another link between energy consumption of information processing devices and broadcasting problems that we have not mentioned before. The idea is that loss of [*timing information*]{} inevitably leads to loss of [*free energy*]{} in covariant devices. This is shown in [@Referenz]. We describe the relevant results.
First, we need the notion of [*passive*]{} devices, i.e., devices having no additional energy source apart from the considered incoming signal. In other words, all energy resources are explicitly included into the description.
${}$\
A device with quantum input state $\rho$ and output $G(\rho)$ is called passive if $G$ is implemented without energy supply, i.e., $$F(G(\rho)) \leq F(\rho) \hspace{2cm} \forall \rho$$ where $F(\rho):=tr(\rho H) - kT S(\rho)$ is the free energy of the system in the state $\rho$ with reference temperature $T$ and Boltzmann constant $k$.
We have shown in [@Referenz] that covariant passive channels that decrease the timing information decrease also the free energy. We rephrase this result formally.
${}$\[FreeImpl\] Let $G$ be a completely positive trace-preserving map describing a covariant passive device. The free energy loss caused by $G$ can be bounded from below by the loss of timing information: $$F(\rho)-F(G(\rho))\geq kT\,\Big( {{\cal I}}(\rho)-{{\cal I}}(G(\rho))\Big)\,.$$
This shows that the channel can only be thermodynamically reversible if it does not subject the signal to a stochastically fluctuating time delay, i.e., it has to conserve the timing information. The result is less trivial than it may seem at first sight. The increase of signal entropy caused by the additional time delay could in principle be compensated by an increase of its inner energy such that the free energy of the system is conserved. The covariance condition is indeed required to show [@Referenz] that the free energy splits up into the following two components $$F(\rho)= kT {{\cal I}}(\rho) +F(\overline{\rho})\,,$$ which cannot be converted into each other.
Together with Theorem \[FreeImpl\] we even obtain statements of the thermodynamical irreversibility of the signal transmission:
${}$\
\[FreeCl\] Let $\rho$ be a quantum state whose timing information has the broadcasting loss $\Delta_{\min}$. Then every channel $G$ satisfies $$C_p(G)\geq 2 \Big(\Delta_{\min} - \frac{1}{kT}(F(\rho)-F(G(\rho))\Big)\,.$$ In particular, for every channel with capacity $C_p(G)=0$ we have $$F(\rho)-F(G(\rho))\geq \frac{2}{kT} \Delta_{\min}\,.$$
We may combine Theorem \[FreeCl\] and Theorem \[ILC\] and obtain the following result:
${}$\
\[End\] Given a passive covariant device $G$ with private channel capacity $C_p(G)$. Let $G$ be applied to a pure input state $\rho$. Then the free energy loss caused by applying $G$ to $\rho$ satisfies $$F(\rho)-F(G(\rho)) \geq kT\,\Big(
\frac{(\Delta E_{in})^8}{64(9\langle E_{out}^4\rangle +
8\langle E_{in}^4\rangle)
\langle E_{in}^4 \rangle }
-\frac{1}{2} C_p(G) \Big)\,,$$ with $E_{in}$ and $E_{out}$ as in Theorem \[ILC\].
It would be desirable to find similar results for mixed states. However, it seems to be hard to provide general bounds. Nevertheless, Theorem \[End\] shows why time covariance brings aspects of quantum information theory into the theory of low-power signal processing. In the context of synchronization protocols we have already described in [@SynchrEntropy] why covariance gives rise to additional limitations of thermodynamically reversible information transfer with classical channels.
Conclusions
===========
We have described a quantum broadcasting problem that arises naturally in classical low power signal processing. If a time-invariant device transmits a signal such that the output signal contains the same amount of Holevo-information about an absolute time frame as the input the following two alternatives are possible: Either the channel has non-zero quantum capacity or it has internally solved a quantum broadcasting problem and copied the same amount of information to its environment. But this is not possible provided that (as we conjecture) the Holevo-information of non-commuting ensembles cannot be broadcast without loss. It is therefore likely that the time-covariant transmission of signals in a way that causes no stochastic time delay of the signal requires devices with non-zero quantum capacity. But avoiding stochastic time delays is, as we have argued a necessary requirement in order to avoid loss of free energy. Thus, we have described a link between quantum information theory and the theory of classical low-power processing.
[10]{}
D. Awschalom, D. Loss, and Samarth, editors. . Springer, Berlin, 2002.
M. Nielsen and I. Chuang. . Cambridge University Press, 2000.
P. Benioff. . , 22(5):562–591, 1980.
R. Feynman. Quantum mechanical computers. , 11:11–46, 1985.
N. Margolus. Parallel quantum computation. In W. Zurek, editor, [*Complexity, Entropy, and the Physics of Information*]{}. Addison Wesley Longman, 1990.
D. Janzing and P. Wocjan. Ergodic quantum computing. , 4(2):129–158, 2005.
D. Janzing. . .
L. Hardy. Quantum theory from five reasonable axioms. .
D. Janzing and T. Beth. Quasi-order of clocks and their synchronism and quantum bounds for copying timing information. , 49(1):230–240, 2003.
D. Janzing and Th. Beth. Are there quantum bounds on the recyclability of clock signals in low power computers? In [*Proceedings of the DFG-Kolloquium VIVA*]{}, Chemnitz, 2002. See also LANL-preprint http://xxx.lanl.gov/abs/quant-ph/0202059.
H. Barnum, C. Caves, C. Fuchs, R. Jozsa, and B. Schumacher. Noncommuting mixed states cannot be broadcast. , 76:2818–2821, 1996.
D. Janzing. Decomposition of time-covariant operations on quantum systems with continuous and/or discrete energy spectrum. , page 122107, 2005.
P. Lax and R. Phillips. . Academic Press, New York, 1967.
D. Janzing. Quantum thermodynamics with missing reference frames: Decompositions of free energy into non-increasing components.
J. Vaccaro, F. Anselmi, H. Wiseman, and K. Jacobs. Complementarity between extractable mechanical work, accessible entanglement, and ability to act as a reference frame, under arbitrary superselection rules. .
G. Chiribella, G. D’Ariano, P. Perinotti, and M. Sacchi. Efficient use of quantum resources for the transmission of a reference frame. , 93:180503, 2004.
M. de Burgh and D. Bartlett. Quantum methods for clock synchronization: Beating the standard quantum limit without entanglement. , 72:042301, 2005.
T. Rudolph and L. Grover. On the communication complexity of establishing a shared reference frame. , page 217905, 2003.
S. van Enk. Quantifying the resource of sharing a reference frame. , 71:032339, 2005.
S. Bartlett, T. Rudolph, and W. Spekkens. Decoherence-full subsystems and the cryptographic power of a private reference frame. , 70:032307, 2004.
D. Bruss, M. Cinchetti, G. D’Ariano, and C. Machiavello. Phase covariant quantum cloning. , 62:12302, 2000.
J. Stam. . , 2:102–11, 1959.
D. Janzing and T. Beth. Bounds on the entropy generated when timing information is extracted from microscopic systems. .
O. Johnson. A conditional entropy power inequality for dependent variables. , 50(8):1581–1583, 2004.
T. Cover and J. Thomas. . Wileys Series in Telecommunications, New York, 1991.
M. Ohya and D. Petz. . Springer Verlag, 1993.
I. Devetak. The private classical capacity and quantum capacity of a quantum channel. , 2003.
D. Janzing. Quantum Computing Models as a Tool Box for Controlling and Understanding the Nanoscopic World. , to appear in a special issue “Quantum Information Technology”. Springer Verlag, Berlin, 2006.
K. Kraus. . Springer-Verlag, Berlin, 1983.
D. Janzing and T. Beth. Synchronizing quantum clocks with classical one-way communication: Bounds on the generated entropy. .
[^1]: One should emphasize that the unitary extension gives only upper bounds on the information transfered to the environment. Real environments are usually in mixed states and can therefore destroy quantum superpositions without receiving information from the system (see [@Studie] for details).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the domain-wall formalism with additional Majorana mass term for the unwanted zero mode, which has recently been proposed for lattice construction of 4D ${\cal N}=1$ super Yang-Mills theory without fine-tuning. Switching off the gauge field, we study the dispersion relation of the energy eigenstates numerically, and find that the method works for reasonable values of Majorana mass. We point out, however, that a problem arises for too large Majorana mass, which can be understood in terms of the seesaw mechanism.'
address:
- 'Institute of Physics, University of Tokyo, Komaba, Meguro-ku, Tokyo 153, Japan'
- 'Institute of Physics, University of Tsukuba, Tsukuba, Ibaraki 305, Japan'
- 'Department of Physics, Nagoya University, Chikusa-ku, Nagoya 464-01, Japan'
author:
- Tomohiro Hotta
- Taku Izubuchi
- Jun Nishimura
title: |
Single massless Majorana fermion\
in the domain-wall formalism
---
lattice gauge theory; supersymmetry 11.30.Pb; 12.60.Jv
Introduction
============
There has been a remarkable progress in understanding non-perturbative aspects of supersymmetric gauge theories recently. The exact results for 4D ${\cal N} = 1$ and ${\cal N}=2$ supersymmetric gauge theories have been obtained using peculiar properties of those theories such as the non-renormalization theorem, the exact $\beta$ functions and so on. The analytic progress is, however, naturally restricted to the aspects of the theories which are strongly constrained by the supersymmetry, and the properties of the non-BPS states, for example, are not understood at all. From this point of view, lattice study of these theories would complement the recent progress and would provide us with a new insight into their nonperturbative dynamics.
Unfortunately supersymmetry is difficult to realize on the lattice. This is not so surprising since the lattice regularization breaks the translational invariance, which forms a subgroup of the supersymmetry. As the translational invariance is restored in the continuum limit, we can restore supersymmetry in the continuum limit. But the price we have to pay for the latter is that we need fine-tuning in general. For 4D ${\cal N}=1$ supersymmetric Yang-Mills theory, one can use the Wilson-Majorana fermion for the gaugino and recover supersymmetry in the continuum limit by fine-tuning the hopping parameter to the chiral limit [@CV]. Some numerical works have been started along this line [@num]. Fine-tuning is a hard task, however, as is known in the numerical studies of the chiral limit in QCD, and a method without fine-tuning is highly desired.
The overlap formalism [@NN] can be used for this purpose, since it preserves exact chiral symmetry on the lattice. The problem here is that the formalism is not suitable for numerical simulation as it stands. A practical proposal[^1] made by Ref. [@nishimura] is to use the domain-wall formalism [@kaplan; @FS], and to decouple the unwanted zero mode by adding Majorana mass term for it. We examine whether this approach really works when the gauge field is switched off as a first step. We study the dispersion relation for various values of the Majorana mass, and examine whether the model gives the desired spectrum. We confirm that the model is fine for moderate values of the Majorana mass, while for too large Majorana mass, an extra almost massless mode appears, which we must be careful of in future study of this model in a more realistic situation with dynamical gauge field.
The paper is organized as follows. In Sec. \[model\] we define our model and review the idea to obtain single massless Majorana fermion without fine-tuning. In Sec. \[hamiltonian\] we give the explicit form of the Hamiltonian of the system. We diagonalize it numerically to study the dispersion relation of the energy eigenstates for various values of the additional Majorana mass. In Sec. \[interpretation\] we interpret the appearance of the extra massless mode for too large Majorana mass in terms of the seesaw mechanism. Section \[summary\] is devoted to summary and discussions.
The model {#model}
=========
Four-dimensional ${\cal N} = 1$ super Yang-Mills theory contains the gauge boson and the gaugino. The gaugino is Majorana fermion, which is equivalent to Weyl fermion in four dimensions.
In order to avoid fine-tuning, we need to impose the chiral symmetry on the lattice. This can be done by the domain-wall formalism [@FS]. Here we have two Weyl fermions with opposite handedness, which couple to the gauge field in the vector-like way.
The idea of Ref. [@nishimura] is to apply this formalism to 4D super Yang-Mills theory, by decoupling one of the Weyl fermions by giving it mass of the order of the cutoff through the additional Majorana mass term for it. It should be noted that this can be done without violating the gauge invariance, since the fermion is in the adjoint representation, which is a real representation, for which the Majorana mass term in 4D is gauge invariant. As a first step, we examine this model by switching off the gauge field.
The action of the model consists of two parts: $$\label{action}
S = S_0 + S_{\rm mass}.$$ $S_0$ is given by $$\label{action0}
S_0 = \bar{\xi} \sigma_\mu \partial_\mu \xi + \bar{\eta}
\bar{\sigma}_\mu \partial_\mu \eta
+ \bar{\xi} {\cal M} \eta +\bar{\eta} {\cal M}^\dagger \xi,$$ where $${\cal M} = \partial_s + M + \frac{1}{2} \triangle.$$
$\xi(x_\mu, s)$ and $\eta(x_\mu, s)$ are right-handed and left-handed Weyl fermions respectively in the four-dimensional space-time, which is latticized as $\{ x_\mu \in {\cal Z};\mu=1,2,3,4 \}$. The $s$ denotes the coordinate in the fifth direction, which runs over $1, \cdots, N_s$. The boundary condition is taken to be free in the fifth direction, and to be periodic in the four space-time directions. Summation over the five-dimensional coordinates ($x_\mu,s$) is suppressed in Eq. (\[action0\]) and similar abbreviations are used in the rest of this paper. $\triangle$ is the five-dimensional lattice Laplacian. $\partial$ should be understood as the lattice derivative. $\sigma_\mu$ and $\bar{\sigma}_\mu$ are defined by $\sigma_\mu = (1, i
\sigma_i)$ and $\bar{\sigma}_\mu = (1, -i \sigma_i)$, where $\sigma_i$ ($i=1,2,3$) are the Pauli matrices. $M$ is a mass parameter, which is fixed at some value within $0 < M <
1$, when one takes the continuum limit.
With the action $S_0$, one obtains a right-handed Weyl fermion and a left-handed one localized at the boundaries of the fifth direction $s
= 1, N_s$, respectively. In the $N_s \rightarrow \infty$ limit, the chiral symmetry is exact and we end up with one massless Dirac fermion [@FS]. For finite $N_s$, the chiral symmetry is violated, but the violation vanishes exponentially with increasing $N_s$ [@Neu; @Vranas].
Let us identify the zero mode in $\xi$ as the massless Majorana fermion we want, namely the gaugino. In order to decouple the unwanted zero mode in $\eta$, we introduce the additional term $S_{\rm mass}$ in the action. There is a variety of choice for the $S_{\rm mass}$. We can, for example, take the Majorana mass term given by [@nishimura]. $$\label{majorana}
S_{\rm mass} = \displaystyle m \left. \left ( \eta^T_s \sigma_2
\eta_s + \bar{\eta}_s \sigma_2 \bar{\eta}^T_s \right ) \right|_{s =
N_s}.$$ $m$ is the parameter which we refer to as the Majorana mass. It should be kept fixed, when one takes the continuum limit, in order to give mass of the order of the cutoff to the unwanted zero mode.
The dispersion relation {#hamiltonian}
=======================
We examine the dispersion relation to see if we get the desired spectrum. This analysis has been done for the domain-wall formalism without the extra Majorana mass term in Ref. [@jansen].
The Hamiltonian of the system can be obtained from the action (\[action\]) as $$\begin{aligned}
\label{eqn:hamiltonian}
H & = & -i \xi^\dagger \sigma_i \partial_i \xi
+i \eta^\dagger \sigma_i \partial_i \eta
+ \xi^\dagger \partial_s \eta - \eta^\dagger \partial_s \xi
-M (\xi^\dagger \eta + \eta^\dagger \xi) \nonumber \\
& & - \frac{1}{2} ( \xi^\dagger \triangle_4 \eta
+ \eta^\dagger \triangle_4 \xi )
- m \left. \left ( \eta^T \sigma_2 \eta + \bar{\eta} \sigma_2
\bar{\eta}^T \right ) \right|_{s = N_s} ,\end{aligned}$$ where $\triangle_4$ represents the lattice Laplacian in the $(x,y,z,s)$ directions. Since we have switched off the gauge field, the system is translationally invariant in the $(x,y,z)$ directions, and therefore we can partially diagonalize the Hamiltonian by working in the momentum basis.
The Hamiltonian for each three-dimensional momentum $\bf p$ can be given as $$\begin{aligned}
H({\bf p}) & = & \xi^\dagger({\bf p}) \sigma_i \sin p_i \, \xi({\bf
p})
- \eta^\dagger({\bf p}) \sigma_i \sin p_i \, \eta({\bf p})
\nonumber \\
& & + \xi^\dagger({\bf p}) \left\{ \partial_s - \frac{1}{2}
\triangle_s - M - (\cos p_i - 1) \right\} \eta({\bf p}) \nonumber
\\
& & + \eta^\dagger({\bf p}) \left\{ - \partial_s - \frac{1}{2}
\triangle_s - M - (\cos p_i - 1) \right\} \xi({\bf p}) \nonumber
\\
& & + [{\bf p} \leftrightarrow - {\bf p} ] \nonumber \\
& & - 2m \left. \left\{ \eta^T(-{\bf p}) \sigma_2 \eta({\bf p})
+ \eta^\dagger({\bf p}) \sigma_2 \eta^\ast(-{\bf p})
\right\} \right|_{s = N_s} , \end{aligned}$$ where $\triangle_s$ is the lattice Laplacian in the $s$ direction. We note that the Majorana mass term mixes the fields with the momentum $\bf p$ and those with $\bf -p$.
We calculate numerically the eigenvalues of the above Hamiltonian for each three-dimensional momentum $\bf p$ for various values of the Majorana mass term. The only difference from the analysis in Ref. [@jansen] is that since the fermion number is not conserved due to the Majorana mass term, we have to make a Bogoliubov transformation to diagonalize the Hamiltonian.
For moderate Majorana mass we have one massless Weyl fermion localized at $s = 1$ as expected. Figure \[fig:dispersion1\] shows the energy of $\xi$ and $\eta$ as a function of $p_x$ with $p_y = p_z = 0$ when the Majorana mass is $0.2$. Here and henceforth, we take $N_s = 20$ and $M = 0.9$. One can see that the $\xi$ has a linear dispersion relation near the origin $p_x = 0$, while the $\eta$ has a mass gap of order one. The doublers of $\xi$ and $\eta$ are removed as in the case without the Majorana mass term.
One might think that larger Majorana mass only results in larger mass for $\eta$ without any problem, but this is not the case. Figure \[fig:dispersion2\] shows the dispersion relation for large Majorana mass $m = 1000$. One can see that although the $\xi$ remains massless and the $\eta$ massive, the doublers of $\xi$ become very light.
=7.5cm
In Fig. \[fig:doubler\] we plot the mass of the next lightest mode at $\bf p = \bf 0$ as well as that of the doublers of $\xi$ as a function of the Majorana mass. The doublers have mass of the order of the cutoff for $m <1.0$, while for $m > 1.0$, the mass decreases with increasing Majorana mass as $\sim 1/m$. The doublers with many $\pi$’s in the momentum components are heavier than those with less $\pi$’s. The mass of the next lightest mode at $\bf p = \bf 0$ grows linearly as the Majorana mass increases, but saturates for $m>0.5$.
The results in this section lead us to the conclusion that there is an appropriate range of the Majorana mass to obtain single Majorana fermion. Note, however, that this does not mean the need for fine-tuning of the parameter since we have quite a large allowed range for the Majorana mass.
Interpretation of the result for the large Majorana mass {#interpretation}
========================================================
We first confirm the behavior of the doubler mass in the large Majorana mass case by looking at the poles of the propagator, which give the masses of the intermediate states. The propagator of the fermions in the domain-wall formalism with the Majorana-type coupling has been calculated in Ref. [@aoki]. The one for $\eta$ can be written as $$< \eta_s \bar{\eta}_t > = - \sigma_\mu \partial_\mu
\left \{ A_R e^{-\alpha (s + t)} + A^m_L e^{\alpha (s + t - 2N_s)}
+ B e^{-\alpha |s - t|} \right \} ,$$ where $A_R$, $A^m_L$ and $B$ are functions of the external momentum $p$. $\alpha$ is a positive constant determined by the parameters in the action. In Ref. [@aoki], they examined $A_R$, $A^m_L$ and $B$ in the $p \rightarrow 0$ limit and showed that there exists no pole at $p^2 = 0$ when the Majorana mass is non-zero, which means that the $\eta$ has been made heavy successfully.
Similarly we can see the existence of very light doublers for the large $m$, by looking at the behavior of the propagator of $\xi$ when $p$ is near one of the corners of the Brillouin zone. We extract the masses of the almost massless doublers from the singular part of the propagator of $\xi$ as $$\label{mdoubler}
m_{\rm doubler} \sim \frac{(2n - M + 2)(2n - M)}{m},$$ for $m \gg 1$, where $n$ is the number of $\pi$’s in the momentum components of the doubler. We have checked that the masses of the doublers extracted from the Hamiltonian diagonalization as in the previous section fit exactly to this formula.
We note that the behavior of the doublers for large Majorana mass seen above can be understood intuitively in terms of the seesaw mechanism, which was originally proposed to explain the lightness of neutrino. A typical example of the mechanism is given by the case in which Dirac and Majorana mass terms coexist. When we diagonalize the mass matrix of the fermion, a very small eigenvalue appears when the Majorana mass is much larger than the Dirac mass.
In fact, the doublers have the two types of mass term in our model. The Dirac mass term comes from the Wilson term in (\[action0\]) and can be written as $$S_{\rm Dirac} = 2 n ( \bar{\xi} \eta + \bar{\eta} \xi ),$$ where $n$ is the number of $\pi$’s in the momentum components of the doubler as before. Together with the Majorana mass term which comes from (\[majorana\]), we have the following mass matrix for the doublers in the basis of the two-component Weyl fermion. $$\left(
\begin{array}{cc}
0 & 2n \\
2n & m
\end{array}
\right).$$ The eigenvalues $\lambda$ of this matrix for $m \gg n$ are given by $$\label{eigenvalue}
\lambda \simeq m , \ \frac{4n^2}{m}.$$ The second one reproduces the large $m$ behavior of the masses of the doublers.
Summary and Discussion {#summary}
======================
We examined whether the proposal for decoupling the unwanted zero-mode in the domain-wall approach by adding the Majorana mass term for it works when the gauge field is switched off. Above all, we clarified what values we should take for the Majorana mass to be added. We observed the desired dispersion relation for moderate values of the Majorana mass, which means that the approach is promising. We pointed out, however, that for too large Majorana mass, the doublers of the desired Majorana fermion become very light. We gave a natural explanation of this phenomenon in terms of the seesaw mechanism. We also confirmed our conclusion by the analysis of the fermion propagator.
There are various types of additional Majorana mass term $S_{\rm
mass}$ that can be used instead of the particular one we used above. We checked that the following alternatives can be used successfully to give $\eta$ mass of the order of the cutoff, while keeping $\xi$ massless.
1. Majorana mass term for both $\xi$ and $\eta$ localized at $s =
N_s$.
2. Majorana mass term for $\eta$ in some finite region near $s=N_s$. One can even extend the region to cover the whole extent of the fifth direction.
3. Majorana mass term for both $\xi$ and $\eta$ in some finite region near $s=N_s$. Unlike the case (ii), one [*cannot*]{} extend the region to cover the whole extent of the fifth direction in this case.
In either case, unwanted light modes appear when we take the Majorana mass too large.
Our next task is of course to switch on the gauge field. We also have to introduce additional boson fields to subtract the heavy modes in the bulk as in Ref. [@Neu; @Vranas]. We hope this approach will finally enable us to understand general nonperturbative phenomena in the super Yang-Mills theory, including the ones related to the vacuum structure such as gaugino condensation.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors would like to thank S. Aoki, K. Nagai and S.V. Zenkin for discussions. T.H. and T.I. are JSPS Research Fellow.
[99]{}
G. Curci and G. Veneziano, Nucl. Phys. B292 (1987) 555.
I. Montvay, Nucl. Phys. B466 (1996) 259; A. Donini and M. Guagnelli, Phys. Lett. B383 (1996) 301.
R. Narayanan and H. Neuberger, Nucl. Phys. B443 (1995) 305.
H. Neuberger, hep-lat/9710089.
J. Nishimura, Phys. Lett. B406 (1997) 215.
D. Kaplan, Phys. Lett. B288 (1992) 342.
V. Furman and Y. Shamir, Nucl. Phys. B439 (1995) 54.
P. Vranas, hep-lat/9705023, hep-lat/9709119.
K. Jansen, Phys. Lett. B288 (1992) 348.
S. Aoki, K. Nagai and S.V. Zenkin, hep-lat/9705001.
[^1]: While this work was being completed, a preprint [@Neu] appeared which includes an alternative proposal in this direction.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The recently proposed Hořava-Lifshitz (HL) theory of gravity is analyzed from the quantum cosmology point of view. By employing usual quantum cosmology techniques, we study the quantum Friedmann-Lemaître-Robertson-Walker (FLRW) universe filled with radiation in the context of HL gravity. We find that this universe is quantum mechanically nonsingular in two different ways: the expectation value of the scale factor $\left<a\right>(t)$ never vanishes and, if we abandon the detailed balance condition suggested by Hořava, the quantum dynamics of the universe is uniquely determined by the initial wave packet and no boundary condition at $a=0$ is indeed necessary.'
author:
- 'João Paulo M. Pitelli'
- Alberto Saa
title: 'Quantum Singularities in Hořava-Lifshitz Cosmology'
---
Introduction
============
In 2009, Hořava proposed a new theory of gravitation [@horava] based on an anisotropic scaling of space $\bf x$ and time $t$ coordinates. The resulting theory, since then dubbed Hořava-Lifshitz (HL) gravity, has proved to be power countable renormalizable. One of its key points is that, even though it does not exhibit relativistic invariance at short distances, General Relativity (GR) is indeed recovered for low-energy limits. Some interesting consequences of this theory include the existence of nonsingular bouncing universes [@wang; @calgani; @brandenberger] and the possibility that it may represent an alternative to inflation, since it might solve the flatness and horizon problem and generate scale invariant perturbations for the early universe without the need of exponential expansion [@wang2; @kiritsis; @mukohyama].
Due to the asymmetry of space and time in the HL gravity, its natural framework is the Arnowitt-Deser-Misner (ADM) formalism [@adm], where the spacetime metric $g_{\mu\nu}(t,{\bf x})$ is decomposed as usual in terms of the 3-dimensional metric $h_{ij}(t,{\bf x})$ of the spatial slices of constant $t$, the lapse function $N(t,{\bf x})$, and the shift vector $N^i(t,{\bf x})$. In his original work, Hořava made an important assumption on the lapse function to simplify the HL gravitational action, the so-called “projectablity condition”, namely $N\equiv N(t)$. There are, nevertheless, extended models where this condition is relaxed. Projectable theories give rise to a unique integrated Hamiltonian constraint, leading to great complications when compared with the GR. The so-called non-projectable theories, on the other hand, typically give rise to a local Hamiltonian constraint, as in GR. Healthy non-projectable extensions of the original HL gravitational action are discussed in details in case[@blas]. Fortunately, since FLRW spacetimes are homogeneous and isotropic, the spatial integral can be dropped from the integrated Hamiltonian constraint in our case [@sotiriou; @Bellorin:2011ff], yielding a true local constraint even for the projectable case. For our purposes here, it suffices to consider the simplest HL projectable theory in a FLRW spacetime. We notice that a modified $F(R)$ HL theory in a FLRW spacetime has been recently considered in [@Elizalde:2010ep], leading to very interesting results regarding the possible unification between primordial inflation and dark energy. Another important assumption originally introduced by Horava is the principle of “detailed balance”. This condition, which states that the potential in the gravitation action follows from the gradient flow generated by a 3-dimensional action, reduces the number of independent coupling constants. Recently, it has became clear that the detailed balance condition can be also relaxed [@sotiriou; @wang3; @Carloni:2009jc; @Carloni:2010ji; @Zhu:2011yu]. In particular, in [@Carloni:2009jc; @Carloni:2010ji], the dynamical role and the consequences for the matter couplings of the detailed balance condition in classical cosmology are detailed. In this paper we will abandon the detailed balance condition, since, as we will show, it gives rise to the most interesting quantum universes. The limit where this condition can be recovered will also be discussed.
There have been many attempts to incorporate Quantum Mechanics into GR. One of the first ones was Quantum Cosmology. In Quantum Cosmology, we work with the Hamiltonian (ADM) formulation of GR, use Dirac’s algorithm [@dirac] of quantization, [*i.e.*]{}, the substitution $\pi_q\to-i\delta/\delta q$, where $\pi_q$ is the canonical momentum associated with the variable $q$ (which can be one of the three canonical variables in GR, $h_{ij}$, $N$ or $N^i$) and the imposition that the first-class constraints of the theory should annihilate the wave function of the spacetime. GR has four constraints, three of them just tell us that the wave function of the spacetime depends only on the intrinsic geometry of the spatial slices in the ADM decomposition, while the last one is a dynamical constraint which gives the dynamical equation of quantum cosmology, the so-called Wheeler-DeWitt equation [@dewitt]. The wave function is a priori defined on the space of all 3-metrics, called superspace, which are in general very intricate infinite dimensional spaces. However, we can take advantage of the symmetries of a homogeneous universe to freeze out all but a finite number of degrees of freedom of the the metric and then quantize the remaining ones. These models are known as minisuperspace models. Quantum cosmology in FLRW minisuperspace filled with a perfect fluid has been shown to be viable and interesting in the sense that the initial big-bang singularity is not present in such model since $\left<a\right>(t)\neq 0$ for all times and the classical behavior of the universe is recovered for large times [@lemos; @alvarenga]. Moreover, in this class of models, a certain evolution parameter of the fluid gives us a measure of time and one can investigate the evolution of the scale factor as the fluid evolves.
For static spacetimes, Horowitz and Marolf [@horowitz] found an original way of classifying a spacetime as quantum mechanically nonsingular. In their work, a spacetime is said to be quantum mechanically nonsingular if the evolution of quantum particles in the classical background is uniquely determined by the initial wave packet, [*i.e.*]{}, no boundary conditions at the classical singular points are necessary. This is equivalent to say that the spatial part of the wave equation is essentially a self-adjoint operator, [*i.e.*]{}, it has a unique self-adjoint extension (for a review about the mathematical framework necessary to define quantum singularities in static spacetimes, see [@pitelli]). In the GR context, the quantization of the FLRW minisuperspace filled with a perfect fluid does require a boundary condition at $a=0$ in order to assure the self-adjointness the Wheeler-DeWitt equation, which, on the other hand, is necessary to guarantee a unitary time evolution. Mathematically, the Hamiltonian operator corresponding to the evolution equation of the universe is not essentially self-adjoint. In this way the quantum dynamics of the universe is not unique since we do not know, in principle, which boundary condition we must apply at the initial singularity. However, as we will see later, it is possible to find quantum cosmologies in the HL gravity context for which the quantum evolution of the universe is unique, and no boundary condition for the wave function is indeed necessary.
In this paper we will apply the machinery of quantum cosmology to the HL theory of gravity. In particular, we will investigate the necessity of initial boundary conditions for the Wheeler-DeWitt equation and also the behavior of the universe, by examining the time evolution of the expectation value of the scale factor. A certain evolution parameter of the radiation filling the universe will play the role of the time coordinate. The content of the universe will be introduced in the gravitational action via the Schutz formalism [@schutz1; @schutz2], demanding the recovering of the usual GR formulation in the low-energy[@saridakis]. The paper is organized as follows. Sections II, III, and IV present brief reviews of the results we need, respectively, the main definitions about quantum singularities, the HL theory of gravity, and the usual quantum cosmology in the GR context. Our main results are presented in the sections V and VI. The last section is devoted to some concluding remarks.
Quantum Singularities
=====================
Typical solutions of the Einstein field equations are known to exhibit singularities. They can be classified as [@konkowski]: quasiregular singularities, where the observer feels no physical quantity diverging, except at the moment when its worldline reaches the singularity (for instance, the conical singularity of a cosmic string); scalar curvature singularities, where every observer approaching the singularity experience diverging tidal forces (for example, the singularity inside a black hole and, more important in the present context, the big bang singularity in FLRW cosmology); non scalar singularities, where there are some curves in which the observers experience unbounded tidal forces (for example, whimper cosmologies). It is well known [@hawking] that under very reasonable conditions (the energy conditions), which basically state that gravity must be attractive, singularities are inevitable in GR. In this way, cosmological models with non-exotic fluids, as radiation or dust, present typically an initial singularity, known as the big bang singularity. Since we cannot escape this fact in GR, we hope that the quantum theory of gravitation will solve this issue, guiding us on how to deal with the singularities, or even excluding them at all. Unfortunately, we do not have such theory yet. However, there are several evidences that this theory would actually solve this problem. These evidences come with the introduction of quantum mechanics in GR in many different ways. In this paper we will highlight two distinct approaches.
The first approach is quantum field theory in curved spacetimes. In this framework, we analyze the behavior of quantum particles (or fields) in a classical curved background, which we assume to be a regular solution of the Einstein field equations. We adopt Horowitz and Marolf’s definitions[@horowitz] . In their work, they analyze the behavior of a scalar particle in singular static spacetimes possessing a timelike Killing vector field $\xi^\mu$. In such spacetimes, the wave equation can be separated into $$\frac{\partial^2\Psi}{\partial t^2}=-A\Psi,$$ where $A=-VD^i(VD_i)+V^2M^2$ and $V=-\xi^{\mu}\xi_{\mu}$, with $D_i$ being the spatial covariant derivative in a static slice $\Sigma$ not containing the singularity. In principle, the domain $\mathcal{D}(A)$ of the operator $A$ is not known, so we choose as a first attempt $\mathcal{D}(A)=C_{0}^{\infty}(\Sigma)$. In this way, our operator is symmetric and positive definite. Howerver, this domain is unnecessarily small, or in other words, the conditions on the functions are so restrictive that the operator $A$ is not self-adjoint. Its adjoint operator $A^{\ast}$ has a much larger domain $\mathcal{D(A^{\ast})}=\left\{\Psi\in L^2(\Sigma):A\Psi\in L^2(\Sigma)\right\}$. It is important to notice that we have chosen $L^2(\Sigma)$ as the Hilbert space of our quantum theory (for a discussion about this point see [@pitelli]). We must relax the conditions on the allowed functions in order to extend the domain of $A$ in such a way that $\mathcal{D}(A^{\ast})\to \mathcal{D}(A)$. If the extended operator is unique, $A$ is said essentially self adjoint and its extension is given by $(\overline{A},\mathcal{D}(\overline{A}))$, where $\overline{A}$ is the closure of $A$ (for more detais see [@reed]). The time evolution of the particle will be then given by $$\Psi(t)=\exp{\left(-i t \overline{A}^{1/2}\right)}\Psi(0),$$ and the spacetime is said to be quantum mechanically nonsingular. However, if the extension is not unique, [*i.e.*]{}, if there exists infinitely many extensions $A_{\alpha}$, with $\alpha$ being a parameter such that to each $\alpha$ there corresponds one boundary condition at the singular point, then we have a different time evolution $$\Psi_\alpha(t)=\exp{\left(-i t A_\alpha^{1/2}\right)}\Psi(0)$$ for each $\alpha$. In this case, the spacetime is said to be quantum mechanically singular. Similarly to the classical case, when a spacetime is quantum mechanically singular, an extra information (a boundary condition) must be given in order to obtain the time evolution. In GR, in particular, we do need to tell what happen to the particle when it reaches the singularity.
The second approach we exploit here is quantum cosmology in minisuperspaces. In this framework, we consider a few degrees of freedom of the system (the rest is assumed to be frozen) and quantize the constraints of the theory via Dirac’s algorithm. We impose $\left[a,p_a\right]=i$ (in units where $\hbar=1$), where $a$ is the scale factor of FLRW models, and $\left[T,p_T\right]=i$, where $T$ is a parameter associated with the evolution of the fluid filling the universe, obtaining, in this way, the Wheeler-DeWitt equation of the universe, which, as we will see, is a Schödinger like equation, from where we can define an internal product between two solutions and, therefore, evaluate expectation values of observables. In this context, we define the universe as nonsingular if $\left<a\right>(t)\neq 0$ for all times. Since the operator $\hat{a}$ is positive in $L^2(0,\infty)$, we will have $\left<a\right>(t) = 0$ if the wave function representing the universe is sharply peaked at $a=0$. Note that this criterion is different from that one originally stated by DeWitt, which says that the universe is quantum mechanically nonsingular if $\Psi(a=0,t)\neq 0$ $\forall$ $t$. In fact, it was shown that this criterion is not enough to prevent singularities in quantum cosmological models [@lemos2].
The two classifications of quantum singularities described above belong to completely different frameworks, but we can apply the mathematical machinery used in static spacetimes in order to decide if the evolution of a wave packet governed by the Wheeler-Dewitt equation is unique in a given quantum cosmology scenario.
HL gravity
==========
In order to introduce the HL theory of gravity, let us first introduce the decomposition of the metric in the ADM form $$ds^2=-N^2c^2dt^2+h_{ij}(dx^i-N^idt)(dx^j-N^jdt),$$ and then let us postulate that the dimensions of space and time are (in units of momentum) $[dx^i]=-1$ and $[dt]=-3$. This assumption assures that theory is power-countable renormalizable in four dimensions. In these units, we have $[N]=[h_{ij}]=0$, while $[N^i]=2$, leading to $[ds^2]=-2$. Notice that the volume element, defined by $$dV_4=N\sqrt{h}d^3{\bf x}dt,$$ has dimension $[dV_4]=-6$.
The extrinsic curvature tensor, which measures how the spatial slices in the ADM decomposition of spacetime curves with respect to external observers, is defined by $$K_{ij}=\frac{1}{2N}\left(\frac{\partial h_{ij}}{\partial t}-\nabla_{\left(i\right.}N_{\left.j\right)}\right).$$ It is easy to see that it has dimension $[K_{ij}]=3$. The most general term involving the extrinsic curvature tensor which is invariant under the group of diffeomorphism of the spatial slices will define the kinetic term in the action. This term depends on two coupling constants $\alpha$ and $\lambda$ and is given by $$S_K=\alpha\int{dtd^3{\bf x}\sqrt{h}N\left(K_{ij}K^{ij}-\lambda K^2\right)}.$$ Note that $[\alpha]=0$, [*i.e.*]{}, $\alpha$ is a dimensionless constant. This is the reason why we made the choice $[dt]=-3$.
The potential term for the gravitational action is given by $$S_V=-\int{dtd^3{\bf x}\sqrt{h}NV[h_{ij}]},$$ where $V[h_{ij}]$ is built out of the spatial metric and its spatial derivatives. Since $[dV_4]=-6$, we must have $[V[h_{ij}]]=6$ in order to assure that $S_V$ be a scalar. The most general action (without the detailed balance condition) containing terms with dimensions less or equal than $6$ is given by (for more details see [@sotiriou]) $$S_{HL}=S_K+S_V,$$ where $$\begin{aligned}
V[h_{ij}]&=g_0\zeta^6+g_1\zeta^4 R+g_2\zeta^2 R^2+g_3\zeta^2R_{ij}R^{ij}\\
&+g_4R^3+g_5R(R_{ij}R^{ij})+g_6R^{i}_{\phantom{i}j}R^{j}_{\phantom{j}k}R^{k}_{\phantom{k}i}\\&+g_7 R\nabla^2R+g_8\nabla_iR_{jk}\nabla^iR^{jk}.
\end{aligned}$$ Here the constant $\zeta$ has dimension $[\zeta]=1$ and ensures that all the coupling $g_a$ are dimensionless. In order to restore the units where $c=1$, [*i.e.*]{}, $[dx]=[dt]$ we need to perform the transformation $dt\to\zeta^{-2}dt$.
Since $[R^i_{jkl}]=2$, as we go to lower momenta, the dominant action is $$S_{IR}=\int{dtd^3{\bf x}N\sqrt{h}\left[\alpha\left(K_{ij}K^{ij}-\lambda K^2\right)-g_1 \zeta^4 R-g_0\zeta^6\right]}.$$ We can now re-scale time and space so that $\alpha=1$ and $g_1=-1$, and set $c=\lambda=1$, leading to $$S_{IR}=\zeta^2\int{dtd^3{\bf x}N\sqrt{h}\left[\left(K_{ij}K^{ij}- K^2\right)+ R-g_0\zeta^2\right]}.$$ Note that, by choosing $$\zeta^2\equiv\frac{1}{16\pi G}, \;\;\; \Lambda=\frac{g_0\zeta^2}{2},$$ we have the usual Einstein-Hilbert action $$\begin{aligned}
S_{GR}&=\frac{1}{16\pi G}\int{dtd^3{\bf x}N\sqrt{h}\left(K_{ij}K^{ij}- K^2+ R-2\Lambda\right)}\\
&=\frac{1}{16\pi G}\int{d^4x\sqrt{-^{(4)}g}\left(^{(4)}R-2\Lambda\right)},
\end{aligned}$$ where $^{(4)}g_{\mu\nu}$ and $^{(4)}R$ are the spacetime metric and Ricci scalar, respectively.
The full HL action we will consider hereafter is $$\begin{aligned}
S_{HL}&=\frac{M_P^2}{2}\int dt d^3{\bf x}N\sqrt{h}\left(K_{ij}K^{ij}-\lambda K^2+R-2\Lambda\right.\\
&-g2M_P^{-2}R^2-g_3M_P^{-2}R_{ij}R^{ij}-g_4M_P^{-4}R^3\\&-g_5M_{P}^{-4}R(R_{ij}R^{ij})-g_6M_P^{-4}R^{i}_{\phantom{i}j}R^{j}_{\phantom{j}k}R^{k}_{\phantom{k}i}\\&\left.-g_7M_P^{-4}R\nabla^2R-g_8M_P^{-4}\nabla_iR_{jk}\nabla^iR^{jk}\right),
\end{aligned}\label{HL action}$$ where $M_P=1/\sqrt{8\pi G}$ stands for the Planck mass in $c=1$, $\hbar=1$ units.
Quantum Cosmology in GR
=======================
In the so-called Schutz formalism [@schutz1; @schutz2] for the matter content of GR, the four-velocity of a perfect fluid is expressed in terms of six potentials in the form $$U_{\nu}=\mu^{-1}(\phi_{,\nu}+\alpha\beta_{,\nu}+\theta S_{,\nu}),
\label{four-velocity}$$ where $\mu$ and $S$ are, respectively, the specific enthalpy and the specific entropy of the fluid. The potentials $\alpha$ and $\beta$ are connected with rotations and, hence, they are not present in the FLRW universe due to its symmetry. The potentials $\phi$ and $\theta$ have no clear physical meaning. With the usual normalization $$U^{\nu}U_{\nu}=-1,$$ Schutz showed that the action for the fluid in GR is given by $$S_{f}=\int{d^4x\sqrt{-g}p},$$ where $p$ is the pressure of the fluid, which is related to the density by the equation of state $p=w\rho$. In this way, the total action for the spacetime filled with a perfect fluid is given by $$S=\frac{M_P^2}{2}\int{d^4x\sqrt{-g}\left(R-2\Lambda\right)}+\int{d^4x\sqrt{-g}p}.$$ Varying the above action with respect to the metric we get the usual Einstein equations $$G_{\mu\nu}+\Lambda g_{\mu\nu}=M_P^{-2}T_{\mu\nu},$$ with $T_{\mu\nu}$ given by $$T_{\mu\nu}=(\rho+p)U_{\mu}U_{\nu}+pg_{\mu\nu}.$$
For the FLRW universe with metric $$ds^2=-N^2dt^2+a(t)^2\left(\frac{dr^2}{1-kr^2}+r^2d\Omega^2\right),$$ where $d\Omega^2$ is the metric in the unit sphere and $k=-1,0,1$ for the open, flat and closed universe, respectively, the four velocity of the fluid is given by $U_{\mu}=N\delta^{0}_{\mu}$ so that $$\mu=(\dot{\phi}+\theta\dot{S})/N.$$ On the other hand, by thermodynamical considerations, Lapchinskii and Rubakov [@lapchinskii] found that the expression for the pressure is given in terms of the potentials by $$\begin{aligned}
p&=\frac{w\mu^{1+1/w}}{(1+w)^{1+1/w}}e^{-S/w}\\&=\frac{w}{(1+w)^{1+1/w}}\left(\frac{\dot{\phi}+\theta\dot{S}}{N}\right)^{1+1/w}e^{-S/w}.
\end{aligned}$$ For the particular case of FLRW universes, we have $$K_{ij}=\frac{\dot{a}}{Na}h_{ij},\,\,\, R_{ij}=\frac{2k}{a^2}h_{ij},$$ with $h_{ij}=\text{diag}\left(\frac{1}{1-kr^2},r^2,r^2\sin^2{\theta}\right)$, so that the total action is given by (in units where $16\pi G=1$) $$\begin{aligned}
S&=\int{dtd^3{\bf x}N\sqrt{h}\left(K_{ij}K^{ij}-K^2+R\right)}+\int{dtd^3{\bf x}N\sqrt{h}p}\\
&=\int \frac{r^2 \sin{\theta}}{\sqrt{1-kr^2}}d^3x\int dt\Bigg\{-6\frac{\dot{a}^2a}{N}+6kNa\\&+N^{-1/w}a^3\frac{w}{(1+w)^{1+1/w}}(\dot{\phi}+\theta\dot{S})^{1+1/w}e^{-S/w}\Bigg\}.
\end{aligned}$$ The spatial integration does not affect the equations of motion, and we have the following canonical momenta associated, respectively, to the dynamical variables $a$, $\phi$ and $S$ $$\begin{aligned}
&p_a=-\frac{12\dot{a}a}{N},\,\,\,p_\phi=\frac{N^{-1/w}a^3}{(1+w)^{1/w}}(\dot{\phi}+\theta\dot{S})^{1/w}e^{-S/w},\,\,\,\\&p_S=\theta p_\phi.
\end{aligned}$$ The Hamiltonian of the system will be given by $$H=p_a\dot{a}+p_{\phi}(\dot{\phi}+\theta\dot{S})-L,$$ where $$\begin{aligned}
L=&-6\frac{\dot{a}a}{N}+6kNa+N^{-1/w}a^3\frac{w}{(1+w)^{1+1/w}}\times\\&\times(\dot{\phi}+\theta\dot{S})^{1+1/w}e^{-S/w}.
\end{aligned}$$ After a tedious but straightforward calculation, we find $$H=N\left(-\frac{p_a^2}{24a}-6ka+p_{\phi}^{1+w}a^{-3w}e^{-S/w}\right).$$
Since the action does not depend on $\dot{N}$, we conclude that $N$ is actually a Lagrange multiplier of the theory. This is not surprising since the results could not depend on how the spacetime is sliced. Varying the action $$S=\int{dt \left[p_a\dot{a}+p_{\phi}(\dot{\phi}+\theta\dot{S})-H\right]}$$ with respect to $N$ leads to the super-Hamiltonian constraint $$\mathcal{H}=-\frac{p_a^2}{24a}-6ka+p_{\phi}^{1+w}a^{-3w}e^{-S/w}\approx 0.$$ Performing a canonical transformation of the form $$\begin{aligned}
&T=-p_Se^Sp_\phi^{-(1+w)},\,\,\,p_T=p_\phi^{(1+w)}e^S,\\&\bar{\phi}=\phi+(1+w)\frac{p_S}{p_\phi},\,\,\,\bar{p}_\phi=p_\phi,
\end{aligned}$$ we get $$\mathcal{H}=-\frac{p_a^2}{24a}-6ka+\frac{p_T}{a^{3w}}\approx 0.$$ Now, we proceed with Dirac’s algorithm of quantization of constrained systems by making the substitutions $p_a\to-i\partial/\partial a$, $p_T=-i\partial/\partial T$, and demand that the constraint annihilate the wave function, finding the Schrödinger-Wheeler-DeWitt equation of the universe $$\frac{\partial^2\Psi}{\partial a^2}+144ka^2\Psi+i24a^{1-3w}\frac{\partial \Psi}{\partial t}=0,$$ with $t=-T$ being the time coordinate in the gauge $N=a^{3w}$, as follows from Hamilton’s classical equations of motion [@alvarenga2]. Notice that the above equation is of the form $i\partial \Psi/\partial t=\hat{H}\Psi$. In order to the Hamiltonian operator $\hat{H}$ to be self-adjoint we define the internal product of two wave functions as $$\left<\Phi,\Psi\right>=\int_{0}^{\infty}{a^{1-3w} \Phi^{\ast}\Psi da}
\label{internal product}$$ and impose restrictive boundary conditions at $a=0$. The simplest ones are the Dirichlet and Neumann conditions: $$\begin{aligned}
&\Psi(0,t)=0,\,\,\,\text{(Dirichlet)}\\
&\frac{\partial \Psi(0,t)}{\partial a}=0,\,\,\, \text{(Neumann)}.
\end{aligned}$$ As we will see, the situation is qualitatively different in the HL theory of gravity.
Quantum Cosmology in HL
=======================
The total action we will consider here is $$S=S_{HL}+\int{dtd^3{\bf x}N\sqrt{h}p},$$ where $S_{HL}$ is given by Eq. (\[HL action\]). We choose this action basically because the GR action can be recovered in the low-energy limit. Discarding the spatial integration again, we have $$\begin{aligned}
S=&\int dt\Bigg[-3(3\lambda-1)\frac{\dot{a}^2a}{N}+6Nka-2\Lambda a^3\\&-\frac{12kN}{a}(3g_2+g_3)-\frac{24kN}{a^3}(9g_4+3g_5+g_6)\\&+N^{-1/w}a^3\frac{w}{(1+w)^{1+1/w}}(\dot{\phi}+\theta\dot{S})^{1+1/w}e^{-S/w}\Bigg]
\label{action}
\end{aligned}$$ Let us introduce the following constants (as in Ref. [@bertolami]): $$\begin{aligned}
&g_C=6k, \,\,\, g_{\Lambda}=2\Lambda, g_{r}=12k(3g_2+g_3),\\& g_S=24k(9g_4+3g_5+g_6).
\end{aligned}$$ Now, proceeding as in GR, [*i.e.*]{}, defining the momenta corresponding to each one of the dynamical variables and calculating the canonical Hamiltonian, we arrive at $$\mathcal{H}=-\frac{p_a^2}{12(3\lambda-1)a}-g_Ca+g_\Lambda a^3+\frac{g_r}{a}+\frac{g_s}{a^3}+\frac{p_T}{a^{3w}}\approx 0.$$
Specializing to the radiation case ($w=1/3$), we find the Schrödinger-Wheeler-DeWitt equation $$\frac{\partial^2\Psi}{\partial a^2}-12(3\lambda-1)\left(g_Ca^2-g_\Lambda a^3-g_r-\frac{g_s}{a^2}\right)\Psi+12(3\lambda-1)i\frac{\partial \Psi}{\partial t}=0,
\label{main equation}$$ again with $t=-T$. Note that $g_r$ just shifts the energy levels, since it is a constant in the potential. However the term $g_s$ changes dramatically the effective potential. The case $g_s=0$ corresponds to the detailed balance condition (see [@calgani] where, with the use of the detailed balance condition, Calcagni obtain an action similar to Eq. (\[action\]), but without the term proportional to $a^{-3}$).
Exact solutions
===============
Flat FLRW universe
-------------------
First, note that if we take a spatially flat universe ($k=0$) with $\Lambda=0$, we have the following equation $$-\frac{1}{12(3\lambda-1)}\frac{\partial^2\Psi}{\partial a^2}=i\frac{\partial \Psi}{\partial t},
\label{free particle}$$ which is a Schrödinger-like equation for a free particle with $\hbar=1$ and mass $m_{\lambda}=6(3\lambda-1)$, except for the requirement $a>0$. Let us consider only the case $\lambda>1/3$. In order to ensure the self-adjointness of the above equation, a boundary condition has to be chosen, as discussed in Section III. For sake of simplicity, we choose Dirichlet boundary condition. For an initial wave packet of the form $$\Psi(a,0)=\left(\frac{128\sigma^3}{\pi}\right)^{1/4}a e^{-\sigma a^2},
\label{normalized initial wave packet}$$ Eq. (\[free particle\]) can be easily solved using a specific propagator (see [@lemos]). The result is $$\begin{aligned}
\Psi(a,t)&=\left(\frac{m_\lambda}{m_\lambda+2it\sigma}\right)^{3/2}\left(\frac{128\sigma^3}{\pi}\right)^{1/4}a \\&\times\exp{\left(-\frac{-\sigma m_\lambda^2 a^2}{m_\lambda^2+4\sigma^2t^2}\right)}\exp{\left(\frac{2itm_\lambda\sigma^2a^2}{m_\lambda^2+4\sigma^2t^2}\right)}
\end{aligned}$$ We can calculate the expectation value of the operator $a$ through the formula $$\begin{aligned}
\left<a\right>(t)&=\left<\Psi,a\Psi\right>=\int_{0}^{\infty}{a|\Psi(a,t)|^2}\\
&=\frac{2}{m_{\lambda}}\sqrt{\frac{2}{\pi\sigma}}\sqrt{\frac{m_\lambda^2}{4}+\sigma^2t^2}.
\end{aligned}$$
Note that if we take $\lambda=1$, we recover the result obtained in Ref. [@lemos]. Nothing changes in HL theory in a flat FLRW universe in comparison with GR (as in the classical case, see for instance [@wang]). In particular, $\left<a\right>(t)$ is nonsingular and $\left<a\right>(t)\sim t$ as $t\to \infty$, recovering the classical behavior of the universe in the classical limit. Besides, the evolution of the wave packet is given once we choose a particular boundary condition at $a=0$. Therefore the evolution of the universe is not unique in the sense stated in Sec. II.
Closed FLRW universe
---------------------
For the spatially closed ($k=1$) FLRW spacetime, there is an effective potential in the Schrödinger-Wheeler-DeWitt equation of the universe. From Eq. (\[main equation\]), we see that this potential represents a shifted quantum harmonic oscillator with a singular perturbation. Setting the mass and frequency of the harmonic oscillator in order to have $m_\lambda=6(3\lambda-1)$ and $\omega_\lambda=\sqrt{2/(3\lambda-1)}$, respectively, we have the following Schrödinger-like equation $$-\frac{1}{2m_\lambda}\frac{\partial^2\Psi}{\partial a^2}+\left(\frac{1}{2}m_\lambda \omega_\lambda^2a^2-g_r-\frac{g_s}{a^2}\right)\Psi=i\frac{\partial \Psi}{\partial t}.
\label{without balance}$$ First of all, we will analyze the necessity of a boundary condition at $a=0$ on this equation. The first step is to separate variables in the form $\Psi(a,t)=\psi(a)e^{-i E t}$, leading to $$\left[-\frac{d^2}{da^2}+V(a)\right]\psi(a)=2m_\lambda E\psi(a),$$ with $$V(a)=m_\lambda^2\omega_\lambda^2a^2-2m_\lambda g_r-(2m_\lambda g_s)/a^2.$$ Following Ref. [@reed], we say that $V(a)$ is in the limit circle case at infinity and at zero if for all $\lambda$, all solutions of $$\left[-\frac{d^2}{da^2}+V(a)\right]\psi(a)=\lambda \psi(a)$$ are square integrable at infinity and at zero, respectively. If $V(a)$ is not in the limit circle case, it is said to be in the limit point case. We now enunciate the Theorem X.7 from Ref. [@reed], which gives us a criterion to decide if the Hamiltonian operator $\hat{H}=-d^2/da^2+V(a)$ is essentially self-adjoint, [*i.e.*]{}, if it has a unique self-adjoint extension.
Let $V(a)$ be a continuous real-valued function on $(0,\infty)$. Then $\hat{H}=-d^2/da^2+V(a)$ is essentially self-ajoint if and only if $V(a)$ is in the limit point case at both zero and infinity.
We need now a criterion to decide if $\hat{H}$ is in the limit point or limit circle case at zero and infinity. We will find this criterion in the next two theorems, extracted again from Ref. [@reed].
Let $V(a)$ be a continuous real-valued function on $(0,\infty)$ and suppose that there exists a positive differentiable function $M(a)$ so that
- $V(a)\geq -M(a)$.
- $\int_1^\infty{\sqrt{M(a)}}da=\infty.$
- $M(a)/(M(a))^{3/2}$is bounded near $\infty$.
Then $V(a)$ is in the limit point case at $\infty$. \[two\]
Let $V(a)$ be continuous and positive near $a=0$. If $V(a)\geq \frac{3}{4a^2}$ near zero then $-d^2/da^2+V(a)$ is in the limit point case at zero. If for some $\epsilon>0$, $V(a)\leq (\frac{3}{4}-\epsilon)a^{-2}$ near zero, then $-d^2/da^2+V(a)$ is in the limit circle case. \[three\]
From now on, we will consider $g_s< 0$. In the end of this section, we will return to the case $g_s>0$. Let us first use Theorem \[two\] to show that the Hamiltonian operator in Eq. (\[without balance\]) is in the limit point case at infinity. To verify this fact, note that the potential $V(a)$ has a minimum $V_{\text{min}}=2m_\lambda\left[\sqrt{-2g_s}m_\lambda^{1/2}\omega_\lambda-g_r\right]$ at $a=\left(\frac{-2gs}{m_\lambda}\right)^{1/4}\omega_\lambda^{-1/2}$. If $V_{\text{min}}\geq0$ we choose $M(a)=1$ and all the requirements of Theorem \[two\] are fulfilled. If $V_{\text{min}}\leq 0$, we take $M(a)=\left|V_{\text{min}}\right|$. In any case, we conclude that $\hat{H}$ is in the limit point case at infinity.
Note now that the potential $V(a)$ has the form $V(a)\sim -2m_\lambda g_s/a^2$ near $a=0$. Therefore, by Theorem \[three\], if $-2m_\lambda g_s\geq 3/4$ then $\hat{H}$ is in the limit point case at zero, otherwise it is in the limit circle case. We have established a range of parameter in which the operator $\hat{H}$ is essentially self-adjoint, [*i.e.*]{}, if $$-m_\lambda g_s\geq3/8,$$ then the evolution of the wave function representing the universe is uniquely determined by the initial wave packet and no boundary condition at $a=0$ is indeed necessary. Otherwise, we need to impose a boundary condition at this point. For simplicity we did not consider the case $\Lambda\neq 0$, but the previous analysis still works in this case.
It turns out that Eq. (\[without balance\]) can be exactly solved. By introducing the new variable $x=\sqrt{m_\lambda\omega_\lambda}a$ and a parameter [@bertolami] $$\alpha=\frac{1}{2}\sqrt{1-8m_\lambda g_s},$$ Eq. (\[without balance\]) becomes $$-\frac{d^2}{dx^2}\psi(x)+\left[x^2-\frac{2}{\omega_\lambda}(gr+E)+\frac{4\alpha^2-1}{4x^2}\right]\psi(x)=0.$$ By introducing the new function $y(\eta)$ given by $$\psi(x)=e^{-\frac{x^2}{2}}x^{\alpha+1/2}y(x^2),$$ it is easy to see that $y(\eta)$ satisfies the associated Laguerre equation $$\eta y''(\eta)+(1+\alpha-\eta)y'(\eta)+\lambda_Ey(\eta)=0,
\label{laguerre}$$ where $$\lambda_E=\left[\frac{(E+g_r)}{2\omega_\lambda}-\frac{(1+\alpha)}{2}\right].$$ This equation is known to be Hermitian (formally self-adjoint) with the inner product $$\left<f,g\right>_L=\int_{0}^{\infty}{e^{-\eta}\eta^{\alpha}\overline{f(\eta)}g(\eta)d\eta}.$$ The general solution of Eq. (\[laguerre\]) is given by [@abramowitz] $$y(\eta)=A_EM(-\lambda_E,1+\alpha,\eta)+B_EU(-\lambda_E,1+\alpha,\eta),$$ where $M$ and $U$ are the confluent hypergeometric functions of first and second kinds, respectively. For $\lambda_E=n=0,1,2,\dots$, both $M$ and $U$ are polynomials of degree $n$, proportional to the associated Laguerre polynomial $L_n^\alpha(\eta)$. The other linear independent solution is not square-integrable near $a=0$, so it must be excluded from the present analysis. If $\lambda_E\notin \mathbb{N}\cup \{0\}$, we have the following asymptotic behavior for $M$ and $U$, as $\eta\to\infty$: $$\begin{aligned}
&M(-\lambda_E,1+\alpha,\eta)\sim \frac{\Gamma(1+\alpha)}{\Gamma(-\lambda_E)}e^{\eta}\eta^{-1-\lambda_E-\alpha},\\
&U(-\lambda_E,1+\alpha,\eta)\sim \eta^{\lambda_E}.
\end{aligned}$$ Therefore $M$ is not square-integrable near infinity, meanwhile $U$ is. So $M$ is not an acceptable solution in this case. As $\eta\to0$, the asymptotic behavior of $U$ is given by $$U(-\lambda_E,1+\alpha,\eta)\sim \eta^{-\alpha}\Gamma(\alpha)/\Gamma(-\lambda_E),
\label{asymptotic}$$ and, hence, $U$ is square-integrable near $a=0$ only if $\alpha<1$. For $\alpha\geq1$, we do not have an acceptable solution, except in the case $\lambda_E=n=0,1,2,\dots$. Therefore, for $\alpha\geq 1$ (which corresponds to $-m_\lambda g_s\geq 3/8$), we quantize automatically the energy levels of the universe, which are given by $$E_n=(2n+1+\alpha)\omega_\lambda-g_r.$$ The corresponding normalized eigenstates are given by $$\begin{aligned}
\Psi_n(a,t)=&(4m_\lambda \omega_\lambda)^{1/4}\left[\frac{\Gamma(n+1)}{\Gamma(\alpha+n+1)}\right]^{1/2}(m_\lambda \omega_\lambda a^2)^{\frac{2\alpha+1}{4}}\\&\times\exp{\left(-\frac{m_\lambda \omega_\lambda}{2}a^2\right)}L_n^\alpha(m_\lambda \omega_\lambda a^2)e^{-i E_n t},
\label{eingenfunctions}
\end{aligned}$$ and a general solution $\Psi(a,t)$, depending on the initial wave packet $\Psi(a,0)$, is then given by $$\Psi(a,t)=\sum_{n=0}^{\infty}{c_n\Psi_n(a,t)},
\label{sum}$$ with $$c_n=\int_{0}^{\infty}{\Psi(a,0)\Psi_n^{\ast}(a,0)da}.$$
For $1/2\leq\alpha<1$, Eq. (\[asymptotic\]) shows that $U$ is square-integrable at $a=0$. We then need a boundary condition at $a=0$ in order to have a well posed Sturm-Liouville problem. They are found in [@derkach] and are given by $$\Gamma_1y=\theta \Gamma_0y,\,\,\,\,\, \theta \in \mathcal{R},$$ where $$\begin{aligned}
&\Gamma_0y=\lim_{x\to 0}{x^{\alpha+1}y'(x)},\\
&\Gamma_1y=\lim_{x\to 0}{\left[y(x)+\frac{x}{\alpha}y'(x)\right]}.
\end{aligned}$$ If $\theta=\infty$, then $\lim_{x\to 0}{x^{\alpha+1}y'(x)}=0$ and we find [@everitt] $\lambda_E=n=0,1,2,\dots$ and the corresponding associated Laguerre polynomials. For others values of $\theta$, the quantized energy levels are not so simple. In all cases, the wave function will satisfy the DeWitt condition $$\Psi(0,t)=0,$$ but we stress the fact that this is not the condition which turns the Wheeler-DeWitt equation into a self-adjoint form.
It is worth to analyze the case $\lambda\to1$, $g_r\to0$ and $\alpha\to 1/2$, where we expect to recover the usual quantum cosmology. In this limit, Eq. (\[eingenfunctions\]) becomes $$\begin{aligned}
\Psi_n(a,t)=&\left(\frac{\sqrt{48}}{\sqrt{\pi}2^{2n+1}(2n+1)!}\right)^{1/2}H_{2n+1}(\sqrt{12}a)\\&\times e^{-6a^2}e^{-i E_n t},
\label{eingenfunctions1}
\end{aligned}$$ with $$E_n=2n+3/2.$$ These same eigenstates have been already found in [@lemos], in the context of ordinary quantum cosmology. They satisfy the Dirichlet boundary condition $\Psi(0,t)=0$. Neumann boundary condition is satisfied when $\alpha\to-1/2$. We do not consider this case here, but it is trivial to generalize our results to $-1<\alpha<1/2$ (see [@derkach]). Therefore, the HL quantum cosmology tend naturally to usual quantum cosmology in the appropriate limit.
Have studied the self-adjointness of the evolution equation of the universe, let us now focus on the evolution of the expectation value of the scale factor given a solution representing the state of the universe. Obviously, if we calculate the expectation value of the scale factor in any of these eigenstates, it will be constant. But the universe evolves so that we must consider wave packets representing the state of the universe. In order to find exact solutions, we choose an initial wave packet of the form $$\Psi(a,0)=\left[\frac{2^{\nu+5/2}\sigma^{\nu+3/2}}{\Gamma\left(\nu+3/2\right)}\right]^{1/2}a^{\nu+1}e^{-\sigma a^2},
\label{initial nu}$$ where $g_s=-\nu(\nu+1)/(2m_\lambda)$. Note that $g_s< 0$ in this case, so that the potential $V(a)$ is repulsive, preventing the formation of a classical singularity. The propagator for Eq. (\[without balance\]) is given by [@efthimiou] $$\begin{aligned}
G(a,a';t;\nu)=&\frac{m_\lambda\omega_\lambda\sqrt{aa'}}{\sin{(\omega_\lambda t)}}e^{ig_r t}i^{-(\nu+3/2)}\\&\times\exp{\left[\frac{im_\lambda\omega_\lambda}{2}\cot{(\omega_\lambda t)(a^2+a'^2)}\right]}\\&\times J_{\nu+1/2}\left(\frac{m_\lambda\omega_\lambda a a'}{\sin{(\omega_\lambda t)}}\right),
\end{aligned}$$ and through the equation $$\Psi(a,t)=\int_{0}^{\infty}{G(a,a';t;\nu)\Psi(a',0)da'},$$ we find, after some tedious calculation, $$\begin{aligned}
\Psi(a,t)=&\left[\frac{2^{\nu+5/2}\sigma^{\nu+3/2}}{\Gamma\left(\nu+3/2\right)}\right]^{1/2}\left(\frac{m_\lambda \omega_\lambda}{i \sin{(\omega_\lambda t)}}\right)^{\nu+3/2}\\&\times\frac{a^{\nu+1}e^{ig_r t}}{\left[2\sigma-im_\lambda\omega_\lambda \cot{(\omega_\lambda t)}\right]^{\nu+3/2}}\\&\times\exp{\left(\frac{-m_\lambda^2\omega_\lambda^2\sigma a^2}{4\sigma^2\sin^2{(\omega_\lambda t)}+m_\lambda^2\omega_\lambda^2\cos^2{(\omega_\lambda t)}}\right)}\\&\times\exp{\Bigg[\frac{im_\lambda\omega_\lambda\sin{(\omega_\lambda t)}\cos{(\omega_\lambda t)}a^2}{2(4\sigma^2\sin^2{(\omega_\lambda t)}+m_\lambda^2\omega_\lambda^2\cos^2{(\omega_\lambda t)})}}\\&(4\sigma^2-m_\lambda^2\omega_\lambda^2)\Bigg].
\end{aligned}$$ The expectation value of the scale factor is given by $$\begin{aligned}
\left<a\right>(t)&=\frac{\Gamma(\nu+2)}{\sqrt{2\sigma}\Gamma\left(\nu+3/2\right)}\frac{1}{m_{\lambda}\omega_\lambda}\\&\times\sqrt{4\sigma^2\sin^2{(\omega_\lambda t)}+m_{\lambda}^2\omega_{\lambda}^2\cos^2{(\omega_\lambda t)}}\\&=\frac{\left<a\right>(0)}{m_\lambda\omega_\lambda}\sqrt{4\sigma^2\sin^2{(\omega_\lambda t)}+m_{\lambda}^2\omega_{\lambda}^2\cos^2{(\omega_\lambda t)}}.
\end{aligned}$$ We stress the nonsingular character of the above equation, since $\left<a\right>(t)\neq 0$ for all times. As in the usual quantum cosmology [@lemos], the singularity is not present in the quantum model. Note also the similarity with the results found in [@lemos]. The only difference here is that the parameter $\lambda$ changes the frequency of oscillation of the scale factor. The behavior of this quantum universe in HL theory is not much different from usual quantum cosmology.
We can also study the case $g_s> 0$, when the potential $V(a)$ is attractive, not preventing the formation of a classical singularity. In order to obtain exact solutions we must restrict $2m_\lambda g_s\leq 1/8$ so that $\alpha\geq 0$. We do not have a propagator in this case, but we can use expression (\[sum\]) to find numerically the evolution of the scale factor. This in done in Fig. \[fig\], where we have chosen an initial wave packet of the form $$\Psi(a,0)=2 \left(\frac{8}{\pi}\right)^{1/4}ae^{-a^2}.$$ The parameters characterizing HL gravity are $\lambda=1$, $gr=0$ and $\alpha=1/4$. Note that $\left<a\right>(t)\neq 0$ $\forall$ $t$. The singularity has been excluded.
![The evolution of the scale factor in the case of an attractive potential. Twenty terms in expansion (\[sum\]) were used in this numerical approximation.[]{data-label="fig"}](fig)
Concluding Remarks
==================
We have seen that in the HL theory of gravity, it is possible not only to exclude the initial big bang singularity, but also to determine uniquely the evolution of the wave function of the universe given an initial wave packet. An equivalent statement is that no boundary condition at $a=0$ is necessary in a quantum cosmology in the context of HL gravity. In general, theories of gravity do not tell us which boundary condition we must choose, so it is a remarkable fact that one of these theories excludes this ambiguity. Moreover, in HL quantum cosmology, the evolution of the expectation value of the scale factor resembles the evolution found in usual quantum cosmology, the only difference being the frequency of oscillation of the bouncing universe. It is interesting to notice that the quantum regime of the HL theory of gravity can also provide a viable framework for the description of the “asymptotic darkness” of the visible universe [@Elizalde:2011tx]. Our early universe results are in a certain sense complementary to the asymptotic regime described in [@Elizalde:2011tx]. It is certainly worthy to further explore the connections between the two approaches.
The authors thank FAPESP and CNPq for the financial support.
[99]{}
P. Horǎva, [*Quantum gravity at a Lifshitz point*]{}, Phys. Rev. D [**79**]{} (2009) 084008 \[hep-th:0901.3775\]
A. Wang and Y. Wu, [*Thermodynamics and classification of cosmological models in the Hořava-Lifshitz theory of gravity*]{}, JCAP [**0907**]{} (2009) 012 \[hep-th:0905.4117\]
G. Calcagni, [*Cosmology of the Lifshitz universe*]{}, JHEP [**09**]{} (2009) 112 \[hep-th:0904.0829\]
R. Brandenberger, [*Matter bounce in Hořava-Lifshitz cosmology*]{}, Phys. Rev. D [**80**]{} (2009) 043516 \[hep-th:0904.2835\]
A. Wang, D. Wands and R. Maartens, [*Scalar field perturbations in Hořava-Lifshitz cosmology*]{}, JCAP [**1003**]{} (2010) 013 \[hep-th:0909.5167\]
E. Kiritsis and G. Kofinas, [*Hořava-Lifshitz Comology*]{}, Nucl. Phys. B [**821**]{} (2009) 467 \[hep-th:0904.1334\]
S. Mukohyama, [*Scale-invariant cosmological perturbations from Hořava-Lifshitz gravity without inflation*]{} (2009) JCAP [**0906**]{}, 001 \[hep-th:0904.2190\]
R. L. Arnowitt, S. Deser, and C. W. Misner, [*The dynamics of general relativity*]{}, in [*Gravitation: An Introduction to Current Research*]{}, edited by L. Witten, John Wiley, New York, (1962) \[gr-qc:0405109\]
D. Blas, O. Pujolas and S. Sibiryakov, [*Models of non-relativistic quantum gravity: the good, the bad and the healthy*]{}, JHEP [**1104**]{} (2011) 018 \[hep-th:1007.3503\]
T.P. Sotiriou, M. Visser and S. Weinfurtner, [*Quantum gravity without Lorentz invariance*]{}, JHEP [**0910**]{} (2009) 033 \[hep-th:0905.2798\]
J. Bellorin and A. Restuccia, [*Consistency of the Hamiltonian formulation of the lowest-order effective action of the complete Horava theory*]{}, Phys. Rev. D [**84**]{}, 104037 (2011) \[arXiv:1106.5766\]. E. Elizalde, S. Nojiri, S. D. Odintsov and D. Saez-Gomez, [*Unifying inflation with dark energy in modified F(R) Horava-Lifshitz gravity*]{}, Eur. Phys. J. C [**70**]{}, 351 (2010) \[arXiv:1006.3387\]. A. Wang and R Maartens, [*Cosmological perturbations in Horava-Lifshitz theory without detailed balance*]{}, Phys. Rev. D [**81**]{} (2010) 024009 \[hep-th:0907.1748\]
S. Carloni, E. Elizalde and P. J. Silva, [*An analysis of the phase space of Horava-Lifshitz cosmologies*]{}, Class. Quant. Grav. [**27**]{}, 045004 (2010) \[arXiv:0909.2219\]. S. Carloni, E. Elizalde and P. J. Silva, [*Matter couplings in Horava-Lifshitz and their cosmological applications*]{}, Class. Quant. Grav. [**28**]{}, 195002 (2011) \[arXiv:1009.5319\]. T. Zhu, F. -W. Shu, Q. Wu and A. Wang, [*General covariant Horava-Lifshitz gravity without projectability condition and its applications to cosmology*]{}, Phys. Rev. D [**85**]{}, 044053 (2012) \[arXiv:1110.5106\].
P. A. M. Dirac, [*Lectures on Quantum Mechanics*]{}, Dover Publications, Inc., New York, (2001)
B. DeWitt, [*Quantum Theory of Gravity. I. The Canonical Theory*]{}, Phys. Rev. [**160**]{} (1967) 1113
N.A. Lemos, [*Radiation-Dominated Quantum Friedmann Models*]{}, J. Math. Phys. [**37**]{} (1996) 1449 \[gr-qc:9511082\]
F.G. Alvarenga, J.C. Fabris, N.A. Lemos and G.A. Monerat, [*Quantum cosmological perfect fluid models*]{}, Gen. Rel. Grav. [**34**]{} (2002) 651 \[gr-qc:0106051\]
G.T. Horowitz and D. Marolf, [*Quantum Probes of Spacetime Singularities*]{}, Phys. Rev. D [**52**]{} (1995) 5670 \[gr-qc:9504028\]
J.P.M. Pitelli and P.S. Letelier, [*Quantum Singularities in Static Spacetimes*]{}, Int. J. Mod. Phys. D [**20**]{} (2011) 729 \[gr-qc:1010.3052\] B.F. Schutz, [*Perfect Fluids in General Relativity: Velocity Potentials and a Variational Principle*]{}, Phys. Rev. D [**2**]{} (1970) 2762
B.F. Schutz, [*Hamiltonian Theory of a Relativistic Perfect Fluid*]{}, Phys. Rev. D [**4**]{} (1971) 3559
E. Saridakis, [*Aspects of Horava-Lifshitz cosmology*]{}, \[astro-ph:1101.0300\]
T. M. Helliwell, D. A. Konkowski and V. Arndt, Gen. Rel. Grav. [*Quantum Singularity in Quasiregular Spacetimes, as Indicated by Klein-Gordon, Maxwell and Dirac Fields*]{} [**35**]{} (2003) 79
S. W. Hawking and G. F. R. Ellis, [*The Large Scale Structure of Universe*]{}, Cambridge University Press, Cambridge U.K., (1974)
M. Reed and B. Simon, [*Fourier Analysis and Self-Adjointness*]{}, Academic Press, New York, (1972)
N.A. Lemos, [*Conservation of probability and quantum cosmological singularities*]{}, Phys. Rev. D [**36**]{} (1987) 2364
V.G. Lapchinskii and V.A. Rubakov, [*Quantum Gravitation: Quantization of the Friedmann Model*]{}, Theor. Math. Phys. [**33**]{} (1977) 1076
F.G. Alvarenga and N.A. Lemos, [*Dynamical Vacuum in Quantum Cosmology*]{}, Gen. Rel. Grav. [**30**]{} (1998) 681 \[gr-qc:9802029\]
O. Bertolami, [*Hořava-Lifshitz quantum cosmology*]{}, Phys. Rev. D [**84**]{} (2011) 044042 \[hep-th:1106.0126\]
I.S. Gradshteyn and I.M. Ryzhik, [*Table of Integrals, Series and Products*]{}, Academic Press, San Diego, (2007)
M. Abramowitz and Irene A. Stegun, [*Handbook of Mathematical Functions*]{}, Dover Publications, New York, (1972)
M. Hajmirzaahmad, [*Laguerre polynomial expansions*]{}, J. Comp. Appl. Math. [**59**]{} (1995) 25
V. Derkach, [*Extensions of Laguerre operators in indefinite inner product spaces*]{}, Math. Notes, [**63**]{} (1998) 449
W.N. Everitt, [*A Catalogue of Sturm-Liouville Diferrential Equations*]{}, in [*Sturm-Liouville Theory, Past and Present*]{}, edited by W.O. Amrein, A.M. Hinz and D. B. Pearson, Birkhäuser Verlag, Basel, (2005)
C.J. Efthimiou, [*Remark on the propagator of the radial oscillator*]{}, Phys. Rev. A [**48**]{} (1993) 4758
E. Elizalde and P. J. Silva, [*On Asymptotic Darkness in Hořava-Lifshitz Gravity*]{}, JHEP [**1206**]{}, 049 (2012) \[arXiv:1110.0036\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study a hypothesis testing problem in which data is compressed distributively and sent to a detector that seeks to decide between two possible distributions for the data. The aim is to characterize all achievable encoding rates and exponents of the type 2 error probability when the type 1 error probability is at most a fixed value. For related problems in distributed source coding, schemes based on random binning perform well and often optimal. For distributed hypothesis testing, however, the use of binning is hindered by the fact that the overall error probability may be dominated by errors in binning process. We show that despite this complication, binning is optimal for a class of problems in which the goal is to “test against conditional independence.” We then use this optimality result to give an outer bound for a more general class of instances of the problem.'
author:
- 'Md. Saifur Rahman and Aaron B. Wagner$^{\footnote{Both authors are with the School of Electrical and Computer Engineering, Cornell University, Ithaca, NY, 14853, USA. (Email: [email protected], [email protected].)}}$'
title: ' **Optimality of Binning for Distributed Hypothesis Testing** '
---
[**Keywords:**]{} distributed hypothesis testing, binning, test against conditional independence, Quantize-Bin-Test scheme, Gaussian many-help-one hypothesis testing against independence, Gel‘fand and Pinsker hypothesis testing against independence, rate-exponent region, outer bound.
Introduction
============
Consider the problem of measuring the traffic on two links in a communication network and inferring whether the two links are carrying any common traffic [@He; @Ameya]. Evidently, this inference cannot be made by inspecting the measurements from one of the links alone, except in the extreme situation in which that link carries no traffic at all. Thus it is necessary to transport the measurements from one of the links to the other, or to transport both measurements to a third location. The measured data is potentially high-rate, however, so this transportion may require that the data be compressed. This raises the question of how to compress data when the goal is not to reproduce it *per se*, but rather to perform inference. A similar problem arises when inferring the speed of a moving vehicle from the times that it passes certain waypoints.
These problems can be modeled mathematically by the setup depicted in Fig. \[fig:Fig1\], which we call the $L$-encoder general hypothesis testing problem. A vector source $(X_1,\dots,X_L, Y)$ has different joint distributions $P_{X_1,\dots,X_L,Y}$ and $Q_{X_1,\dots,X_L,Y}$ under two hypotheses $H_0$ and $H_1$, respectively. Encoder $l$ observes an i.i.d. string distributed according to $X_l$ and sends a message to the detector at a finite rate of $R_l$ bits per observation using a noiseless channel. The detector, which has access to an i.i.d. string distributed according to $Y$, makes a decision between the hypotheses. The detector may make two types of error: the type 1 error ($H_0$ is true but the detector decides otherwise) and the type 2 error ($H_1$ is true but the detector decides otherwise). The type 1 error probability is upper bounded by a fixed value. The type 2 error probability decreases exponentially fast, say with an exponent $E$, as the length of the i.i.d. strings increases. The goal is to characterize the rate-exponent region of the problem, which is the set of all achievable rate-exponent vectors $(R_1,\dots,R_L,E)$, in the regime in which the type 1 error probability is small. This problem was first introduced by Berger [@Berger] (see also [@Han]) and arises naturally in many applications. Yet despite these applications, the theoretical understanding of this problem is far from complete, especially when compared with its sibling, distributed source coding, where random binning has been shown to be a key ingredient in many optimal schemes.
![$L$-encoder general hypothesis testing[]{data-label="fig:Fig1"}](Fig8.pdf){width="3.5in"}
Note that if one of the variables in the set $(X_1,\ldots,X_L,Y)$ has a different marginal distribution under $P_{X_1,\dots,X_L,Y}$ and $Q_{X_1,\dots,X_L,Y}$, then one of the terminals can detect the underlying hypothesis with an exponentially-decaying type 2 error probability, even without receiving any information from the other terminals, and could communicate this decision to other terminals by broadcasting a single bit. Motivated by the applications mentioned above, we shall focus our attention on the case in which the variables $X_1,\ldots,X_L,Y$ have the same marginal distibutions under both hypotheses.
Ahlswede and Csiszár [@Ahl] studied a special case of this problem in which $L=1$. They presented a scheme in which the encoder sends a quantized value of $X_1$ to the detector which uses it to perform the test with the help of $Y$. They showed that their scheme is optimal for a “test against independence.” Their scheme was later improved by Han [@Han1] and Shimokawa-Han-Amari [@Han2]. In the latter improvement, the encoder first quantizes $X_1$, then bins the quantized value using a Slepian and Wolf encoder [@Slepian]. The detector first decodes the quantized value with the help of $Y$ and then performs a likelihood ratio test. In this scheme, type 2 errors can occur in two different ways: the binning can fail so that the receiver decodes the wrong codeword and therefore makes an incorrect decision, or the true codeword can be decoded correctly yet be atypically distributed with $Y$, again resulting in an incorrect decision. Moreover, there is a tension between these two forms of error. If the codeword is a high fidelity representation of $X_1$, then binning errors are likely, yet the detector is relatively unlikely to make an incorrect decision if it decodes the codeword correctly. If the codeword is a low fidelity representation, then binning errors are unlikely, but the detector is more likely to make an incorrect decision when it decodes correctly.
![Shimokawa-Han-Amari achievable region for a fixed channel $P_{U_1|X_1}$[]{data-label="fig:Fig2"}](Fig13.pdf){width="3.8in"}
Fig. \[fig:Fig2\] illustrates this tradeoff for a fixed test channel $P_{U_1|X_1}$ used for quantization. All mutual information quantities are computed with respect to $P$. $\rho_2^{*}(U_1)$ and $\rho_1^{*}(U_1)$ are the exponents associated with type 2 errors due to binning errors and assuming correct decoding of the codeword, respectively. Formulas for each are available in [@Han]. For low rates, binning errors are common and $\rho_2^{*}(U_1)$ dominates the overall exponent. For high rates, binning errors are uncommon and $\rho_1^{*}(U_1)$ dominates the overall exponent. To achieve the overall performance, the test channel should be chosen so that these two exponents are equal; if they are not, then making the test channel slightly more or less noisy will yield better performance. A similar tradeoff arises in the analysis of error exponents of binning-based schemes for the Wyner-Ziv problem [@Ben; @Ben2; @Ben3; @Kochman] and in the design of short block-length codes for Wyner-Ziv or joint source-channel coding. Evidently the benefit accrued from binning is reduced when one considers error exponents, as opposed to when the design criterion is vanishing error probability or average distortion, because the error exponent associated with the binning process itself may dominate the overall performance.
The Shimokawa-Han-Amari scheme uses random, unstructured binning. It is known from the lossless source coding literature that structured binning schemes can strictly improve upon unstructured binning schemes in terms of the error exponents [@Csiszar; @Ben1; @Csiszar1]. Thus, two questions naturally arise:
1. Is the tradeoff depicted in Fig. \[fig:Fig2\] fundamental to the problem or an artifact of a suboptimal scheme?
2. Can the scheme be improved by using structured binning?
We conclusively answer both questions and show that unstructured binning is optimal in several important cases. We begin by considering a special case of the problem that we call $L$-encoder hypothesis testing against conditional independence. Here $Y$ is replaced by a three-source $(X_{L+1},Y,Z)$ such that $Z$ induces conditional independence between $(X_1,\dots,X_L,X_{L+1})$ and $Y$ under $H_1$. In addition, $(X_1,\dots,X_L,X_{L+1},Z)$ and $(Y,Z)$ have the same distributions under both hypotheses. This problem is a generalization of the single-encoder test against independence studied by Ahlswede and Csiszár [@Ahl],
For this problem we provide an achievable region, based on a scheme we call Quantize-Bin-Test, that reduces to the Shimokawa-Han-Amari region for $L = 1$ yet is significantly simpler. We also introduce an outer bound similar to the outer bound for the distributed rate-distortion problem given by Wagner and Anantharam [@Wagner3]. The idea is to introduce an auxiliary random variable that induces conditional independence between the sources. This technique of obtaining an outer bound has been used to prove results in many distributed source coding problems [@Wagner3; @Wagner1; @Wagner; @Wagner2; @Ozarow; @Wang].
The inner (achievable) and outer bounds are shown to match in three examples. The first is the case in which there is only one encoder ($L = 1$). Although this problem is simply the conditional version of the test against independence studied by Ahlswede and Csiszár [@Ahl], the conditional version is much more complicated due to the necessary introduction of binning. It follows that the Shimokawa-Han-Amari scheme is optimal for $L = 1$, providing what appears to be the first nontrivial optimality result for this scheme. This problem arises in detecting network flows in the presence of common cross-traffic that is known to the detector. Here $X_1$ represents the network traffic measured at a remote location, $Y$ is the traffic measured at the detector, and $Z$ represents the cross-traffic. The goal is to detect the presence of common traffic beyond $Z$, i.e., to determine whether $Z$ captures all of the dependence between $X_1$ and $Y$.
The second is a problem inspired by a result of Gel‘fand and Pinsker [@Gelfand]. We refer to this as the Gel‘fand and Pinsker hypothesis testing against independence problem, the setup of which is shown in Fig. \[fig:Fig3\]. Here $X_{L+1}$ and $Z$ are deterministic and there is a source $X$ which under $H_0$ is the minimum sufficient statistic for $Y$ given $(X_1,\dots,X_L)$ such that $X_1,\dots,X_L, Y$ are conditionally independent given $X$. We characterize the set of rate vectors $(R_1,\dots,R_L)$ that achieve the centralized exponent $I(X;Y)$. We show that the Quantize-Bin-Test scheme is optimal for this problem.
![Gel‘fand and Pinsker hypothesis testing against independence[]{data-label="fig:Fig3"}](Fig7.pdf){width="3.5in"}
The third is the Gaussian many-help-one hypothesis testing against independence problem, the setup of which is shown in Fig. \[fig:Fig4\]. Here the sources are jointly Gaussian and there is another scalar Gaussian source ${X}$ observed by the main encoder which sends a message to the detector at a rate $R$. The encoder observing $X_l$ is now referred to as the helper $l$. We characterize the rate-exponent region of this problem in a special case when $X_1,\dots,X_L,Y$ are conditionally independent given $X$. We use results on related source coding problem by Oohama [@Oohama2005] and Prabhakaran *et al.* [@Vinod] to obtain an outer bound, which we show is achieved by the Quantize-Bin-Test scheme.
![Gaussian many-help-one hypothesis testing against independence[]{data-label="fig:Fig4"}](Fig6.pdf){width="3.5in"}
For all three examples, we obtain the solution by observing that the relevant error exponent takes the form of a mutual information, and thereby relate the problem to a source-coding problem. This correspondence was first observed by Ahlswede and Csiszár [@Ahl]. Tian and Chen later applied it in the context of successive refinement [@Tian]. These three conclusive results enable us to answer both of the above questions. Because the Shimokawa-Han-Amari scheme is optimal for $L=1$, the tradeoff that it entails, depicted in Fig. \[fig:Fig2\], must be fundamental to the problem. Moreover, as both the Shimokawa-Han-Amari and Quantize-Bin-Test schemes do not use structured binning, we conclude that it is not necessary for this problem, at least in the special case considered here.
As a byproduct of our results, we obtain an outer bound for a more general class of instances of the distributed hypothesis testing problem. This is the first nontrivial outer bound for the problem, and numerical experiments show that it is quite close to the existing achievable regions in many cases.
The rest of the paper is organized as follows. In Section 2, we introduce the notation used in the paper. We give the mathematical formulation of the $L$-encoder general hypothesis testing problem in Section 3. Section 4 is devoted to the $L$-encoder hypothesis testing against conditional independence problem. Section 5 is on the special case in which there is only one encoder. The Gel‘fand and Pinsker hypothesis testing against independence problem is studied in Section 6. The Gaussian many-help-one hypothesis testing against independence problem is studied on Section 7. Finally, we present the outer bound for a class of the general problem in Section 8.
Notation
========
We use upper case to denote random variables and vectors. Boldface is used to distinguish vectors from scalars. Arbitrary realizations of random variables and vectors are denoted in lower case. For a random variable $X$, $X^n$ denotes an i.i.d. vector of length $n$, $X^n(i)$ denotes its *i*th component, $X^n(i : j)$ denotes the *i*th through *j*th components, and $X^n(i^c)$ denotes all but the *i*th component. For random variables $X$ and $Y$, we use $\sigma^2_X$ and $\sigma^2_{X|Y}$ to denote the variance of $X$ and the conditional variance of $X$ given $Y$, respectively. The closure of a set $\mathcal{A}$ is denoted by $\overline{\mathcal{A}}$. $|f|$ denotes the cardinality of the range of a function $f$. $1_A$ denotes the indicator function of an event $A$. The determinant of a matrix ${\mathbf}{K}$ is denoted by $\det ({\mathbf}{K})$. The notation $x^{+}$ denotes $\max(x, 0)$. All logarithms are to the base 2. $\mathbb{R}_{+}^L$ is used to denote the positive orthant in $L$-dimensional Euclidean space. The notation $X \leftrightarrow Y \leftrightarrow Z$ means that $X, Y,$ and $Z$ form a Markov chain in this order. For $0 \le p \le 1$, $H_b(p)$ denotes the binary entropy function defined as $$H_b(p) \triangleq - p \log p - (1-p) \log (1-p).$$ All entropy and mutual information quantities are under the null hypothesis, $H_0$, unless otherwise stated.
$L$-Encoder General Hypothesis Testing
======================================
Problem Formulation
-------------------
Let ${\left (}X_1, \dots, X_L,Y{\right )}$ be a generic source taking values in $\prod_{l=1}^{L} \mathcal{X}_l \times \mathcal{Y}$, where $\mathcal{X}_1, \dots, \mathcal{X}_L, $ and $\mathcal{Y}$ are alphabet sets of ${X}_1, \dots, {X}_L, $ and ${Y}$, respectively. The distribution of the source is $P_{{X_1\dots X_LY}}$ under the null hypothesis $H_0$ and is $Q_{{X_1\dots X_LY}}$ under the alternate hypothesis $H_1$, i.e., $$\begin{aligned}
H_0: \hspace{0.05in}& P_{{X_1\dots X_LY}} \\
H_1: \hspace{0.05in}& Q_{{X_1\dots X_LY}}.\end{aligned}$$ Let ${\left \{}
\newcommand{\rmb}{\right \}}{\left (}{X}^n_1(i), \dots, {X}^n_L(i),{Y}^n(i){\right )}\rmb_{i=1}^{n}$ be an i.i.d. sequence of random vectors with the distribution at a single stage same as that of $(X_{1}, \dots, X_{L}, Y)$. We use $\mathcal{L}$ to denote the set $\{1,\dots,L\}$. For $S \subseteq \mathcal{L}$, $S^c$ denotes the complement set $\mathcal{L} \setminus S$ and ${\mathbf}{X}^n_{S}(i)$ denotes $({X}^n_l(i))_{l \in S}$. When $S = \mathcal{L}$, we simply write ${\mathbf}{X}^n_{\mathcal{L}}(i)$ as ${\mathbf}{X}^n(i)$. Likewise when $S = \{l\}$, we write ${\mathbf}{X}^n_{\{l\}}(i)$ and ${\mathbf}{X}^n_{\{l\}^c}(i)$ as $X^n_l(i)$ and ${\mathbf}{X}^n_{l^c}(i)$, respectively. Similar notation will be used for other collections of random variables.
As depicted in Fig. \[fig:Fig1\], the encoder $l$ observes $X^n_l$, then sends a message to the detector using an encoding function $$\begin{aligned}
f_{l}^{(n)} : \mathcal{X}_l^{n} \mapsto \left \{1,\dots,M_{l}^{(n)} \right \}.\end{aligned}$$ $Y^n$ is available at the detector, which uses it and the messages from the encoders to make a decision between the hypotheses based on a decision rule $$g^{(n)}{\left (}m_1, \dots, m_L,{y}^n {\right )}= \left\{
\begin{array}{l l}
H_0 & \quad \mbox{if ${\left (}m_1, \dots, m_L,{y}^n {\right )}\mbox{is in } A$}\\
H_1 & \quad \mbox{otherwise,}\\ \end{array} \right.$$ where $$A \subseteq \prod_{l=1}^L {\left \{}
\newcommand{\rmb}{\right \}}1,\dots,M_{l}^{(n)} \rmb \times \mathcal{Y}^{n}$$ is the acceptance region for $H_0$. The encoders $f_l^{(n)}$ and the detector $g^{(n)}$ are such that the type 1 error probability does not exceed a fixed $\epsilon$ in $(0,1)$, i.e., $$P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n}(A^c) \leq \epsilon, \\$$ and the type 2 error probability does not exceed $\eta$, i.e., $$Q_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}} {Y}^n}(A) \leq \eta.$$
A rate-exponent vector $$({\mathbf}{R},E) = (R_1,\dots,R_L,E)$$ is *achievable* for a fixed $\epsilon$ if for any positive $\delta$ and sufficiently large $n$, there exists encoders $f^{(n)}_l$ and a detector $g^{(n)}$ such that $$\begin{aligned}
\frac{1}{n}\log M^{(n)}_l &\leq R_l + \delta \hspace {0.05in} \textrm{for all} \hspace {0.05in} l \hspace {0.05in} \textrm{in} \hspace {0.05in} \mathcal{L}, \hspace {0.05in} \textrm{and}\\
-\frac{1}{n}\log \eta &\geq E- \delta.\end{aligned}$$ Let $\mathcal{R}_{\epsilon}$ be the set of all achievable rate-exponent vectors for a fixed $\epsilon$. The *rate-exponent region* $\mathcal{R}$ is defined as $$\begin{aligned}
\mathcal{R} \triangleq \bigcap_{\epsilon > 0} \mathcal{R}_{\epsilon}.\end{aligned}$$
Our goal is to characterize the region $\mathcal{R}$.
Entropy Characterization of the Rate-Exponent Region
----------------------------------------------------
We start with the entropy characterization of the rate-exponent region. We shall use it later in the paper to obtain inner and outer bounds. Define the set $$\begin{aligned}
\mathcal{R}_{*} \triangleq \bigcup_n \bigcup_{\left(f_l^{(n)}\right)_{l \in \mathcal{L}}} \mathcal{R}_{*}\left(n,\left(f_l^{(n)}\right)_{l \in \mathcal{L}}\right),\end{aligned}$$ where $$\begin{aligned}
\mathcal{R}_{*}\left(n,\left(f_l^{(n)}\right)_{l \in \mathcal{L}}\right) \triangleq \Biggr \{ ({\mathbf}{R},E) :
R_l &\ge \frac{1}{n} \log \left|f_l^{(n)}\left({X}_l^n\right)\right| \hspace {0.05 in} \textrm{for all \emph{l} in} \hspace{0.05in} \mathcal{L}, \hspace {0.05 in} \textrm{and} \nonumber\\
E &\le \frac{1}{n} D\biggr(P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n} \Bigr\| Q_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n}\biggr) \Biggr \}.\end{aligned}$$ We have the following Proposition.
$\mathcal{R} = \overline{\mathcal{R}_{*}}$.
The proof of Proposition 1 is a straight-forward generalization of that of Theorem 1 in [@Ahl] and is hence omitted. Ahlswede and Csiszár [@Ahl] showed that for $L=1$, the strong converse holds, i.e., $\mathcal{R}_{\epsilon}$ is independent of ${\epsilon}$. Thus, $\overline{\mathcal{R}_{*}}$ is essentially a characterization for both $\mathcal{R}$ and $\mathcal{R}_{\epsilon}$. While we expect this to hold for the problem under investigation too, we shall not investigate it here. We next study a class of instances of the problem before returning to the general problem in Section 8.
$L$-Encoder Hypothesis Testing against Conditional Independence
===============================================================
We consider a class of instances of the general problem, referred to as the $L$-encoder hypothesis testing against conditional independence problem, and obtain inner and outer bounds to the rate-exponent region. These bounds coincide and characterize the region completely in some cases. Moreover, the outer bound for this problem can be used to give an outer bound for a more general class of problems, as we shall see later.
Let $X_{L+1}$ and $Z$ be two generic sources taking values in alphabet sets $\mathcal{X}_{L+1}$ and $\mathcal{Z}$, respectively such that $({\mathbf}{X},X_{L+1})$ and $Y$ are conditionally independent given $Z$ under $H_1$, and the distributions of $({\mathbf}{X},X_{L+1},Z)$ and $(Y,Z)$ are the same under both hypotheses, i.e., $$\begin{aligned}
H_0: \hspace{0.05in}& P_{{{\mathbf}{X}X_{L+1}Y|Z}}P_Z \\
H_1: \hspace{0.05in}& P_{{{\mathbf}{X}X_{L+1}|Z}}P_{Y|Z}P_Z.\end{aligned}$$ The problem formulation is the same as before with $Y$ replaced by $(X_{L+1},Z,Y)$ in it. The reason for focusing on this special case is that the relative entropy in (1) becomes a mutual information, which simplifies the analysis. Let $\mathcal{R}^{CI}$ be the rate-exponent region of this problem. Here “$\emph{CI}$" stands for conditional independence. Let $$\begin{aligned}
\mathcal{R}_{*}^{CI} \triangleq \bigcup_n \bigcup_{\left(f_l^{(n)}\right)_{l \in \mathcal{L}}} \mathcal{R}_{*}^{CI}\left(n,\left(f_l^{(n)}\right)_{l \in \mathcal{L}}\right),\end{aligned}$$ where $$\begin{aligned}
\mathcal{R}_{*}^{CI}\left(n,\left(f_l^{(n)}\right)_{l \in \mathcal{L}}\right) \triangleq \Biggr \{ ({\mathbf}{R},E) :
R_l &\ge \frac{1}{n} \log \left|f_l^{(n)}\left({X}_l^n\right)\right| \hspace{0.05in} \textrm{for all \emph{l} in} \hspace{0.05in} \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
E &\le \frac{1}{n} I\biggr(\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, {X}_{L+1}^n;{Y}^n\Bigr|Z^n\biggr) \Biggr \}.\nonumber\end{aligned}$$ We have the following corollary as a consequence of Proposition 1.
$\mathcal{R}^{CI} = \overline{\mathcal{R}_{*}^{CI}}$.
With mutual information replacing relative entropy, the problem can be analyzed using techniques from distributed rate-distortion. In particular, both inner and outer bounds for that problem can be applied here.
Quantize-Bin-Test Inner Bound
-----------------------------
Our inner bound is based on a simple scheme which we call the Quantize-Bin-Test scheme. In this scheme, encoders, as in the Shimokawa-Han-Amari scheme, quantize and then bin their observations, but the detector now performs the test directly using the bins. The inner bound obtained is similar to the generalized Berger-Tung inner bound for distributed source coding [@Berger1; @Tung; @Gastper]. Let $\Lambda_i$ be the set of finite-alphabet random variables $\lambda_i = (U_1,\dots, U_L, T)$ satisfying
1. $T$ is independent of $({\mathbf}{X},X_{L+1},Y,Z)$, and
2. ${U}_l \leftrightarrow ({X}_l,T) \leftrightarrow ({\mathbf}{U}_{l^c}, {\mathbf}{X}_{l^c},X_{L+1},Y,Z)$ for all $l$ in $\mathcal{L}$.
Define the set $$\begin{aligned}
\mathcal{R}_i^{CI}(\lambda_i) \triangleq \biggr \{ ({\mathbf}{R},E) :
\sum_{l \in S}R_l &\ge I\left({\mathbf}{X}_S; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},X_{L+1},Z,T \right)\hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
E &\le I \left(Y;{\mathbf}{U},X_{L+1}|Z,T \right) \biggr \}\nonumber\end{aligned}$$ and let $$\begin{aligned}
\mathcal{R}_i^{CI} \triangleq \bigcup_{\lambda_i \in \Lambda_i} \mathcal{R}_i^{CI}(\lambda_i).\nonumber\end{aligned}$$ The following lemma asserts that $\mathcal{R}^{CI}_i$ is computable and closed.
1. $\mathcal{R}_i^{CI}$ remains unchanged if we impose the following cardinality bound on $({\mathbf}{U},T)$ in $\Lambda_i$ $$\begin{aligned}
|\mathcal{U}_l| &\le |\mathcal{X}_l| + 2^{L} - 1 \hspace{0.05in}\textrm{for all l in} \hspace{0.05in} \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
|\mathcal{T}| &\le 2^{L}.\end{aligned}$$
2. $\mathcal{R}_i^{CI}$ is closed.
The proof of Lemma 1 is presented in Appendix A. Although the cardinality bound is exponential in the number of encoders, one can obtain an improved bound by exploiting the contra-polymatroid structure of $\mathcal{R}_i^{CI}$ [@Chen; @Viswanath]. We do not do so here because it is technically involved and we just want to prove that $\mathcal{R}_i^{CI}$ is closed. The following theorem gives an inner bound to the rate-exponent region.
$\mathcal{R}_i^{CI} \subseteq \mathcal{R}^{CI}.$
Theorem 1 is proved in Appendix B.
*Remark 1:* Although our inner bound is stated for the special case of the test against conditional independence, it can be extended to the general case. But, the inner bound thus obtained will be quite complicated, with competing exponents, and it is not needed in this paper.
It is worth pointing out that the Quantize-Bin-Test scheme is in general suboptimal for problems in which encoders’ observations have common randomness, i.e., there exists deterministic functions of encoders’ observations that is common to encoders. However, it is straightforward to generalize this scheme by using the idea from the common-component scheme for distributed source coding problems [@Wagner4].
Outer Bound
-----------
The outer bound is similar to the outer bound for the distributed rate-distortion problem given by Wagner and Anantharam [@Wagner3]. Let $\Lambda_o$ be the set of finite-alphabet random variables $\lambda_o = ({\mathbf}{U}, W, T)$ satisfying
1. $(W,T)$ is independent of $({\mathbf}{X},X_{L+1},Y,Z)$, and
2. ${U}_l \leftrightarrow ({X}_l,W,T) \leftrightarrow ({\mathbf}{U}_{l^c}, {\mathbf}{X}_{l^c},X_{L+1},Y,Z)$ for all $l$ in $\mathcal{L}$,
and let $\chi$ be the set of finite-alphabet random variable $X$ such that $X_1, \dots, X_L, X_{L+1}, Y$ are conditionally independent given $(X, Z)$. Note that $\chi$ is nonempty because it contains $({\mathbf}{X},X_{L+1})$. For a given $X$ in $\chi$ and $\lambda_o$ in $\Lambda_o$, the joint distribution of $X$, $({\mathbf}{X},X_{L+1},Y,Z)$, and $\lambda_o$ satisfy the Markov condition $$X \leftrightarrow ({\mathbf}{X},X_{L+1},Y,Z) \leftrightarrow \lambda_o.$$ Wagner and Anantharam [@Wagner3] refer to this condition as “[Markov coupling]{}" between $X$ and $\lambda_o$. Define the set $$\begin{aligned}
\mathcal{R}_o^{CI}(X,\lambda_o) \triangleq \biggr \{ ({\mathbf}{R},E) :
\sum_{l \in S}R_l &\ge I \left({X}; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},X_{L+1},Z,T\right) + \sum_{l \in S} I\left(X_l ; U_l | X, W,X_{L+1}, Z,T\right) \hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
E &\le I\left(Y;{\mathbf}{U},X_{L+1}|Z,T\right) \biggr \}.\nonumber\end{aligned}$$ Also let $$\begin{aligned}
\mathcal{R}_o^{CI} \triangleq \bigcap_{X \in \chi} \bigcup_{\lambda_o \in \Lambda_o} \mathcal{R}_o^{CI}(X,\lambda_o).\nonumber\end{aligned}$$ We have the following outer bound to the rate-exponent region.
$\mathcal{R}_{*}^{CI} \subseteq \mathcal{R}_o^{CI}$ and therefore $\mathcal{R}^{CI} \subseteq \overline{\mathcal{R}_o^{CI}}.$
The proof of the first inclusion is presented in Appendix C. The first inclusion and Corollary 1 imply the second inclusion. The next three sections provide examples in which the inner and outer bounds coincide. In Section 8, we will see how to extend the outer bound to the general problem.
$1$-Encoder Hypothesis Testing against Conditional Independence
===============================================================
In this section, we study a special case in which $L=1$. We prove that the Quantize-Bin-Test scheme is optimal for this problem. We also prove that the Shimokawa-Han-Amari inner bound coincides with the Quantize-Bin-Test inner bound, establishing that the Shimokawa-Han-Amari scheme is also optimal.
Rate-Exponent Region
--------------------
For this problem, the rate-exponent region $$\begin{aligned}
\mathcal{R}^{CI} = \overline{\mathcal{R}^{CI}_o} &= \mathcal{R}^{CI}_i \\
&= \tilde {\mathcal{R}}^{CI} \triangleq \Bigr \{(R_1,E): \hspace{0.05in}\textrm{there exists} \hspace{0.05in}U_1 \hspace{0.05in}\textrm{such that} \nonumber\\
&\hspace{1in}R_1 \ge I(X_1 ; U_1 | {X}_2, Z), \nonumber\\
&\hspace{1in}E \le I(Y;{U_1},{X}_2|Z), \\
&\hspace{1in}| {\mathcal{U}}_1| \le |\mathcal{X}_1| + 1, \hspace{0.05in} \textrm{and} \nonumber\\
&\hspace{1in} U_1 \leftrightarrow X_1 \leftrightarrow ({X}_2, Y, Z)\Bigr \}.\nonumber\end{aligned}$$
To show (2), it suffices to show that $$\mathcal{R}^{CI}_o \subseteq \mathcal{R}^{CI}_i,$$ because $\mathcal{R}^{CI}_i$ is closed from Lemma 1(b). Consider $(R_1,E)$ in $\mathcal{R}^{CI}_o$. Take $X={X}_2$. It is evident that ${X}_2$ is in $\chi$. Then there exists $\lambda_o = (U_1,W,T)$ in $\Lambda_o$ such that $(R_1,E)$ is in $\mathcal{R}^{CI}_o({X}_2,\lambda_o)$, i.e., $$\begin{aligned}
R_1 &\ge I({X}_2; {U}_1| {X}_2,Z,T) + I(X_1 ; U_1 | {X}_2, Z,W,T)\\
&=I(X_1 ; U_1 | {X}_2, Z,W,T),\end{aligned}$$ and $$\begin{aligned}
E &\le I(Y;{U_1},{X}_2|Z,T) \nonumber\\
&=H(Y|Z,T) - H(Y|U_1,{X}_2,Z,T) \nonumber\\
&\le H(Y|Z,W,T) - H(Y|U_1,{X}_2,Z,W,T)\\
&= I(Y;{U_1},{X}_2|Z,W,T),\nonumber\end{aligned}$$ where (4) follows from conditioning reduces entropy and the fact that $(Y,Z)$ is independent of $(W,T)$. If we set $\tilde T = (W,T)$, then it is easy to verify that $\lambda_i = (U_1,\tilde T)$ is in $\Lambda_i$ and we have $$\begin{aligned}
R_1 &\ge I(X_1 ; U_1 | {X}_2, Z,\tilde T), \hspace{0.05in} \textrm{and}\\
E &\le I(Y;{U_1},{X}_2|Z,\tilde T).\end{aligned}$$ Therefore, $(R_1,E)$ is in $\mathcal{R}^{CI}_i(\lambda_i)$, which implies that $(R_1,E)$ is in $\mathcal{R}^{CI}_i$. This completes the proof of (2).
To prove (3), it suffices to show that $$\mathcal{R}^{CI}_i \subseteq \tilde {\mathcal{R}}^{CI}.$$ The reverse containment immediately follows if we restrict $T$ to be deterministic in the definition of $\mathcal{R}^{CI}_i$. Continuing from the proof of (2), let $\tilde U_1 = (U_1, \tilde T)$. Since $(U_1,\tilde T)$ is in $\Lambda_i$, we have that $\tilde T$ is independent of $({X}_1,X_2,Y,Z)$ and that $$U_1 \leftrightarrow (\tilde T, X_1) \leftrightarrow ({X}_2,Y,Z).$$ Both together imply that $$\begin{aligned}
\tilde U \leftrightarrow X_1 \leftrightarrow ({X}_2,Y,Z).\end{aligned}$$ We next have from (5) that $$\begin{aligned}
R_1 &\ge I(X_1 ; U_1 | {X}_2, Z,\tilde T) \nonumber\\
&= I(X_1 ; U_1 | {X}_2, Z,\tilde T) + I(X_1 ; \tilde T | {X}_2, Z)\\
&=I(X_1 ; U_1,\tilde T| {X}_2, Z) \nonumber\\
&=I(X_1 ; \tilde U_1| {X}_2, Z),\nonumber\end{aligned}$$ where (7) follows because $\tilde T$ is independent of $({X}_1,X_2,Y,Z)$. And (6) similarly yields $$\begin{aligned}
E &\le I(Y;\tilde {U}_1,{X}_2|Z).\nonumber\end{aligned}$$ Using the support lemma [@csiszar Lemma 3.4, pp. 310] as in the proof of Lemma 1(a), we can obtain the cardinality bound $$|\tilde {\mathcal{U}}_1| \le |\mathcal{X}_1| + 1 \nonumber.$$ We thus conclude that $(R_1,E)$ is in $\tilde {\mathcal{R}}^{CI}$.
Optimality of Shimokawa-Han-Amari Scheme
----------------------------------------
![Shimokawa-Han-Amari achievable region for a fixed $P_{U_1|X_1}$[]{data-label="fig:Fig5"}](Fig9.pdf){width="3.7in"}
The Shimokawa-Han-Amari scheme operates as follows. Consider a test channel $P_{U_1|X_1}$, a sufficiently large block length $n$, and $\alpha > 0$. Let $\bar R_1 = I(X_1;U_1) + \alpha$. To construct the codebook, we first generate $2^{n \bar R_1}$ independent codewords $U_1^n$, each according to $\prod_{i=1}^n P_{U_1}(u_{1i})$, and then distribute them uniformly into $2^{n R_1}$ bins. The codebook and the bin assignment are revealed to the encoder and the detector. The encoder first quantizes $X_1^n$ by selecting a codeword $U_1^n$ that is jointly typical with it. With high probability, there will be at least one such codeword. The encoder then sends to the detector the index of the bin to which the codeword $U_1^n$ belongs. The joint type of $(X_1^n,U_1^n)$ is also sent to the detector, which requires zero additional rate asymptotically. The detector finds a codeword $\hat {U}_1^n$ in the bin that minimizes the empirical entropy $H({U}^n_1,Y^n)$. It then performs the test and declares $H_0$ if and only if both $(X_1^n,U_1^n)$ and $(Y^n,\hat{U}_1^n)$ are jointly typical under $H_0$. The inner bound thus obtained is as follows. Define $$\begin{aligned}
\mathcal{A}(R_1) &\triangleq \Bigr \{U_1: R_1 \ge I(X_1;U_1|{X}_2,Y,Z), \hspace{0.1in} U_1 \leftrightarrow X_1 \leftrightarrow ({X}_2,Y,Z), \hspace {0.05 in} \textrm{and} \hspace {0.05 in} |\mathcal{U}_1| \le |\mathcal{X}_1| + 1 \Bigr \} \\
\mathcal{B}(U_1) &\triangleq \Bigr \{P_{\tilde U_1 \tilde X_1 \tilde X_2 \tilde Y \tilde Z}: P_{\tilde U_1 \tilde {X}_1} = P_{U_1X_1}\hspace {0.05 in} \textrm{and}\hspace {0.05 in} P_{\tilde U \tilde {{X}}_{2} \tilde Y \tilde Z } = P_{U_1 {X}_2YZ} \Bigr\}\\
\mathcal{C}(U_1) &\triangleq \Bigr \{P_{\tilde U_1 \tilde X_1 \tilde X_2 \tilde Y \tilde Z }: P_{\tilde U_1 \tilde {X}_1} = P_{U_1X_1}, \hspace {0.05 in} P_{\tilde {{X}}_{2} \tilde Y \tilde Z} = P_{{X}_2YZ}, \hspace {0.05 in} \textrm{and}\hspace {0.05 in} H (\tilde U_1 | \tilde {{X}}_{2}, \tilde Y,\tilde Z ) \ge H\left (U_1|{X}_2,Y,Z\right) \Bigr\}.\end{aligned}$$ In addition, define the exponents $$\begin{aligned}
\rho_1^{*}(U_1) &\triangleq \min_{P_{\tilde U_1 \tilde X_1 \tilde X_2 \tilde Y \tilde Z} \in \mathcal{B}(U_1)} D \bigr(P_{\tilde U_1\tilde X_1 \tilde X_2 \tilde Y \tilde Z } \|P_{U_1|X_1} P_{X_1 X_2|Z}P_{Y|Z}P_Z\bigr) \\
\rho_2^{*}(U_1) &\triangleq \left\{
\begin{array}{l l}
+ \infty & \quad \mbox{if $R_1 \ge I(U_1;X_1)$}\\
\rho_2(U_1) & \quad \mbox{otherwise}\\ \end{array} \right. \\
\rho_2(U_1) &\triangleq [R_1 - I(X_1;U_1|{X}_2,Y,Z)]^{+} + \min_{P_{\tilde U_1 \tilde X_1 \tilde X_2 \tilde Y \tilde Z} \in \mathcal{C}(U_1)} D \bigr(P_{\tilde U_1 \tilde X_1 \tilde X_2 \tilde Y \tilde Z } \|P_{U_1|X_1} P_{X_1 X_2|Z}P_{Y|Z}P_Z\bigr).\end{aligned}$$ Finally, define $$\begin{aligned}
E_{SHA}(R_1) \triangleq \max_{U_1 \in \mathcal{A}(R_1)} \min \hspace {0.05 in} \left ( \rho_1^{*}(U_1), \rho_2^{*}(U_1) \right).\end{aligned}$$ Recall that $\rho_2^{*}(U_1)$ and $\rho_1^{*}(U_1)$ are the exponents associated with type 2 errors due to binning errors and assuming correct decoding of the codeword, respectively.
[@Han2] $(R_1,E)$ is in the rate-exponent region if $$E \le E_{SHA}(R_1).$$
Fig. \[fig:Fig5\] shows the Shimokawa-Han-Amari achievable exponent as a function of the rate assuming a fixed channel $P_{U_1|X_1}$ is used for quantization. This is simply Fig. \[fig:Fig2\] particularized to the $1$-encoder hypothesis testing against conditional independence problem. For rates $R_1 \ge I(X_1,U_1|{X}_2,Z)$, $\rho_1^{*}(U_1)$ dominates $\rho_2^{*}(U_1)$ and there is no penalty for binning at these rates as the exponent stays the same. Therefore, we can bin all the way down to the rate $R_1 = I(X_1,U_1|{X}_2,Z)$ without any loss in the exponent. However, if we bin further at rates $R_1$ in $[I(X_1,U_1|{X}_2,Y,Z),I(X_1,U_1|{X}_2,Z))$, then $\rho_2^{*}(U_1)$ dominates $\rho_1^{*}(U_1)$, the exponent decreases linearly with $R_1$, and the performance deteriorates all the way down to a point at which the message from the encoder is useless. At this point, the binning rate $R_1$ equals $I(X_1,U_1|{X}_2,Y,Z)$ and the exponent equals $I(Y;{X}_2|Z)$, which is the exponent when the detector ignores the encoder’s message. This competition between the exponents makes the optimality of the Shimokawa-Han-Amari scheme unclear. We prove that it is indeed optimal by showing that the Shimokawa-Han-Amari inner bound simplifies to the Quantize-Bin-Test inner bound, which by Theorem 3 is tight. Let us define $$\mathcal{A}^{*}(R_1) \triangleq \Bigr \{U_1: R_1 \ge I(X_1;U_1|{X}_2,Z), \hspace{0.1in} U_1 \leftrightarrow X_1 \leftrightarrow ({X}_2,Y,Z), \hspace {0.05 in} \textrm{and}\hspace {0.05 in}|\mathcal{U}_1| \le |\mathcal{X}_1| + 1 \Bigr \}$$ and $$E_{QBT}(R_1) \triangleq \max_{U_1 \in \mathcal{A}^{*}(R_1)} I(Y;{U_1},{X}_2|Z).$$ We have the following theorem.
If $(R_1,E)$ is in the rate-exponent region, then $$E \le E_{QBT}(R_1) = E_{SHA}(R_1).$$
The inequality follows from Theorem 3. To prove the equality, it is sufficient to show that $$E_{SHA}(R_1) \ge E_{QBT}(R_1).$$ The reverse inequality follows from Theorem 3 and 4. Since conditioning reduces entropy and any $U_1$ in $\mathcal{A}^{*}(R_1)$ satisfies the Markov chain $$U_1 \leftrightarrow X_1 \leftrightarrow ({X}_2,Y,Z),$$ we have $$\begin{aligned}
R_1 &\ge I(X_1;U_1|{X}_2,Z) \\
&= H(U_1|{X}_2,Z) - H(U_1|{X}_1{X}_2,Z)\\
&\ge H(U_1|{X}_2,Y,Z) - H(U_1|{X}_1{X}_2,Y,Z)\\
&= I(X_1;U_1|{X}_2,Y,Z),\end{aligned}$$ which means that $U_1$ is in $\mathcal{A}(R_1)$. Hence, $\mathcal{A}^{*}(R_1) \subseteq \mathcal{A}(R_1)$. This implies that $$\begin{aligned}
E_{SHA}(R_1) &\triangleq \max_{U_1 \in \mathcal{A}(R_1)} \min \hspace {0.05 in} \left ( \rho_1^{*}(U_1), \rho_2^{*}(U_1) \right) \nonumber\\
&\ge \max_{U_1 \in \mathcal{A}^{*}(R_1)} \min \hspace {0.05 in} \left ( \rho_1^{*}(U_1), \rho_2^{*}(U_1) \right).\end{aligned}$$ Now the objective of the optimization problem in the definition of $\rho_1^{*}(U_1)$ can be lower bounded as $$\begin{aligned}
D \bigr(P_{\tilde U_1\tilde {{X}}_1 \tilde {{X}}_2 \tilde Y \tilde Z } \|P_{U_1|X_1} P_{{{X}}_1 {{X}}_2|Z}P_{Y|Z}P_Z\bigr) &\ge D \bigr(P_{\tilde U_1 \tilde {{X}}_{2} \tilde Y \tilde Z } \|P_{U_1{X}_2|Z}P_{Y|Z}P_Z\bigr) \\
&= D \bigr(P_{ U_1 {{X}}_{2} Y Z } \|P_{U_1{X}_2|Z}P_{Y|Z}P_Z\bigr) \\
&= I(Y; {U}_1,{X}_2|Z).\end{aligned}$$ The lower bound is achieved by the distribution $P_{ U_1 {{X}}_{2} Y Z} P_{X_1|U_1 {{X}}_{2} Z}$ in $\mathcal{B}(U_1)$. Therefore, $$\begin{aligned}
\rho_1^{*}(U_1) = I(Y; {U}_1,{X}_2|Z).\end{aligned}$$ Similarly, we can lower bound the optimization problem in the definition of $\rho_2(U_1)$ as $$\begin{aligned}
D \bigr(P_{\tilde U_1\tilde {{X}}_1 \tilde {{X}}_2 \tilde Y \tilde Z } \|P_{U_1|X_1} P_{{X}_1 X_2|Z}P_{Y|Z}P_Z\bigr) &\ge D \bigr(P_{ \tilde {{X}}_{2} \tilde Y \tilde Z } \|P_{{X}_2|Z}P_{Y|Z}P_Z\bigr) \\
&= D \bigr(P_{ {{X}}_{2} Y Z } \|P_{{X}_2|Z}P_{Y|Z}P_Z\bigr) \\
&= I(Y; {X}_2|Z),\end{aligned}$$ and the lower bound is achieved by the distribution $P_{{{X}}_{2} Y Z} P_{U_1 X_1| {{X}}_{2} Z}$ in $\mathcal{C}(U_1)$. Therefore, $$\begin{aligned}
\rho_2(U_1) = [R_1 - I(X_1;U_1|{X}_2,Y,Z)]^{+} + I(Y; {X}_2|Z).\end{aligned}$$ Consider any $U_1$ in $\mathcal{A}^{*}(R_1).$ If $R_1 \ge I(X_1;U_1)$, then $$\begin{aligned}
\min \hspace {0.05 in} \left ( \rho_1^{*}(U_1), \rho_2^{*}(U_1) \right) &= \rho_1^{*}(U_1) \nonumber\\
&= I(Y; {U}_1,{X}_2|Z).\end{aligned}$$ And if $I(X_1;U_1) > R_1 \ge I(X_1;U_1|{X}_2,Z)$, then $$\begin{aligned}
\min \hspace {0.05 in} \left ( \rho_1^{*}(U_1), \rho_2^{*}(U_1) \right) &= \min \hspace {0.05 in} \bigr ( I(Y; {U}_1,{X}_2|Z), R_1 - I(X_1;U_1|{X}_2,Y,Z) + I(Y; {X}_2|Z) \bigr ) \nonumber\\
&\ge \min \hspace {0.05 in} \bigr ( I(Y; {U}_1,{X}_2|Z), I(X_1;U_1|{X}_2,Z) - I(X_1;U_1|{X}_2,Y,Z) + I(Y; {X}_2|Z) \bigr)\nonumber\\
&= \min \hspace {0.05 in} \bigr ( I(Y; {U}_1,{X}_2|Z), I(Y;U_1|{X}_2,Z)+ I(Y; {X}_2|Z) \bigr)\nonumber\\
&= \min \hspace {0.05 in} \bigr ( I(Y; {U}_1,{X}_2|Z), I(Y; {U}_1,{X}_2|Z) \bigr)\nonumber\\
&= I(Y; {U}_1,{X}_2|Z).\end{aligned}$$ Now (8) through (10) imply $$\begin{aligned}
E_{SHA}(R_1) &\ge \max_{U_1 \in \mathcal{A}^{*}(R_1)} I(Y; {U}_1,{X}_2|Z) \nonumber\\
&= E_{QBT} (R_1).\nonumber\end{aligned}$$ Theorem 5 is thus proved.
Gel‘fand and Pinsker Hypothesis Testing against Independence
============================================================
We now consider another special case, which we call the Gel‘fand and Pinsker hypothesis testing against independence problem, because it is related to the source coding problem studied by Gel‘fand and Pinsker [@Gelfand].
Suppose that $X_{L+1}$ and $Z$ are deterministic and suppose there exists a function of $X_1$, …, $X_L$, say $X$, such that under $H_0$,
1. $ X_1,..,X_L, Y$ are conditionally independent given $X$, and
2. for any finite-alphabet random variable $U$ such that $Y \leftrightarrow X \leftrightarrow U$ and $Y \leftrightarrow U \leftrightarrow X$, we have $H(X|U) = 0.$
Conditions (C5) and (C6) imply that under $H_0$, $X$ is a minimal sufficient statistic for $Y$ given ${\mathbf}{X}$ such that $ X_1,\dots,X_L, Y$ are conditionally independent given $X$. We shall characterize the centralized rate region, the set of rate vectors that achieve the centralized type 2 error exponent $I({\mathbf}{X};Y) = I(X;Y)$. More precisely, we shall characterize the set $$\left \{{\mathbf}{R} : ({\mathbf}{R},I(X;Y)) \in \mathcal{R}^{CI} \right\},$$ denoted by $\mathcal{R}^{CI}\bigr(I(X;Y)\bigr)$. We define $\mathcal{R}^{CI}_i\bigr(I(X;Y)\bigr)$ and $\overline{\mathcal{R}^{CI}_o}\bigr(I(X;Y)\bigr)$ similarly. We need the following lemma.
Condition (C6) is equivalent to
1. For any positive $\epsilon$, there exists a positive $\delta$ such that for all finite-alphabet random variables $U$ such that $Y \leftrightarrow X \leftrightarrow U$ and $I(X;Y|U) \le \delta$, we have $H(X|U) \le \epsilon.$
The proof of Lemma 2 is presented in Appendix D. Let us define a function $$\begin{aligned}
\phi(\delta) \triangleq \inf \Bigr \{ \epsilon : \hspace {0.05 in} \textrm{for all finite-alphabet} \hspace {0.05 in} U \hspace {0.05 in} \textrm{such that}\hspace {0.05 in} &Y \leftrightarrow X \leftrightarrow U \hspace {0.05 in} \textrm{and} \hspace {0.05 in}I(X;Y|U) \le \delta, \hspace {0.05 in} \textrm{we have}\hspace {0.05 in} H(X|U) \le \epsilon \Bigr\}.\end{aligned}$$ It is clear that $\phi$ is continuous at zero with the value $ \phi (0) = 0.$ We have the following theorem.
For this problem, the centralized rate region $$\mathcal{R}^{CI}\bigr(I(X;Y)\bigr) = \mathcal{R}^{CI}_i\bigr(I(X;Y)\bigr) = \overline{\mathcal{R}^{CI}_o}\bigr(I(X;Y)\bigr).$$
It suffices to show that $$\overline{\mathcal{R}^{CI}_o}\bigr(I(X;Y)\bigr) \subseteq \mathcal{R}^{CI}_i \bigr (I(X;Y) \bigr).$$ Consider any ${\mathbf}{R}$ in $\overline{\mathcal{R}^{CI}_o}\bigr(I(X;Y)\bigr)$, any positive $\delta$, and $X$ defined as above. Then there exists $\lambda_o = ({\mathbf}{U},W,T)$ in $\Lambda_o$ such that $(R_1+\delta,\dots,R_L+\delta,I(X;Y)-\delta)$ is in $\mathcal{R}^{CI}_o(X,\lambda_o)$, i.e., $$\begin{aligned}
\sum_{l \in S}(R_l+\delta) &\ge I({X}; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},T) + \sum_{l \in S} I(X_l ; U_l | X, W, T) \hspace {0.05 in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
I(X;Y) -\delta &\le I(Y;{\mathbf}{U}|T).\end{aligned}$$ We have the Markov chain $$\begin{aligned}
Y \leftrightarrow X \leftrightarrow ({\mathbf}{U},T),\end{aligned}$$ which implies $$\begin{aligned}
I(X;Y|{\mathbf}{U},T) &= H(Y|{\mathbf}{U},T) - H (Y|X,{\mathbf}{U},T) \\
&= H(Y|{\mathbf}{U},T) - H (Y|X) \\
&= I(X;Y) - I(Y;{\mathbf}{U}|T)\\
&\le \delta,\end{aligned}$$ where the last inequality follows from (12). Therefore, by the definition of $\phi$ function $$\begin{aligned}
H(X|{\mathbf}{U},T) \le \phi(\delta).\end{aligned}$$ Now $$\begin{aligned}
I(X;{\mathbf}{U}_S|{\mathbf}{U}_{S^c},T) &= H(X|{\mathbf}{U}_{S^c},T) - H(X|{\mathbf}{U},T) \nonumber \\
&\ge H(X|{\mathbf}{U}_{S^c},W,T) - \phi(\delta) \\
&\ge I(X;{\mathbf}{U}_S|{\mathbf}{U}_{S^c},W,T) - \phi(\delta),\nonumber\end{aligned}$$ where (14) follows from (13) and the fact that conditioning reduces entropy. This together with (11) implies $$\begin{aligned}
\sum_{l \in S}(R_l+\delta+\phi(\delta)) &\ge I({X}; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},W,T) + \sum_{l \in S} I(X_l ; U_l | X, W, T) \nonumber\\
&= I({X}; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},W,T) + I({\mathbf}{X}_S; {\mathbf}{U}_S | {\mathbf}{U}_{S^c}, X,W, T) \nonumber\\
&=I({X},{\mathbf}{X}_S; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},W,T) \nonumber\\
&\ge I({\mathbf}{X}_S; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},W,T).\nonumber\end{aligned}$$ Again since conditioning reduces entropy and $Y$ is independent of $(W,T)$, we obtain from (12) that $$\begin{aligned}
I(X;Y) - \delta &\le I(Y;{\mathbf}{U}|T) \\
&= H(Y|T) - H(Y|{\mathbf}{U},T) \\
& \le H(Y|W,T) - H(Y|{\mathbf}{U},W,T)\\
&= I(Y;{\mathbf}{U}|W,T).\end{aligned}$$ Define $\tilde T = (W,T)$. It is then clear that $\lambda_i = ({\mathbf}{U},\tilde T)$ is in $\Lambda_i$, $$\begin{aligned}
\sum_{l \in S}(R_l+\delta+\phi(\delta)) &\ge I({\mathbf}{X}_S; {\mathbf}{U}_S|{\mathbf}{U}_{S^c}, \tilde T) \hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L},\hspace{0.05in} \textrm{and} \\
I(X;Y) -\delta &\le I(Y;{\mathbf}{U}|\tilde T).\end{aligned}$$ Hence, $(R_1+\delta+\phi(\delta),\dots,R_L+\delta+\phi(\delta),I(X;Y)-\delta)$ is in $\mathcal{R}^{CI}_i(\lambda_i),$ which implies that $({\mathbf}{R},I(X;Y))$ is in $\mathcal{R}^{CI}_i$ because $\mathcal{R}^{CI}_i$ is closed from Lemma 1(b). Therefore, ${\mathbf}{R}$ is in $\mathcal{R}^{CI}_i \bigr ( I(X;Y)\bigr).$
Gaussian Many-Help-One Hypothesis Testing against Independence
==============================================================
We now turn to a continuous example of the problem studied in Section 4. This problem is related to the quadratic Gaussian many-help-one source coding problem [@Wagner; @Oohama2005; @Vinod]. We first obtain an outer bound similar to the one in Theorem 2 and then show that it is achieved by the Quantize-Bin-Test scheme.
Let ${\left (}X, Y, X_1, \dots, X_L{\right )}$ be a zero-mean Gaussian random vector such that $$X_l = X + N_l$$ for each $l$ in $\mathcal{L}$. ${X}$ and ${Y}$ are correlated under the null hypothesis $H_0$ and are independent under the alternate hypothesis $H_1$, i.e., $$\begin{aligned}
H_0: \hspace{0.05in} &Y = X + N \\
H_1: \hspace{0.05in} &Y {\protect\mathpalette{\protect\independenT}{\perp}}X.\end{aligned}$$ We assume that $X,N, N_1, N_2, \dots, N_L$ are mutually independent, and that $\sigma^2_N$ and $\sigma^2_{N_l}$ are positive. The setup of the problem is shown in Fig. \[fig:Fig4\]. Unlike the previous problem, we now allow $X$ to be observed by an encoder, which sends a message to the detector at a finite rate $R$. We use $f^{(n)}$ to denote the corresponding encoding function. In order to be consistent with the source coding terminology, we call this the main encoder. The encoder observing $X_l$ is now called helper $l$. We assume that $X_{L+1}$ and $Z$ are deterministic. The rest of the problem formulation is the same as the one in Section 3.1. Let $\mathcal{R}^{MHO}$ be the rate-exponent region of this problem. We need the entropy characterization of $\mathcal{R}^{MHO}$. For that, define $$\begin{aligned}
\mathcal{R}_{*}^{MHO} \triangleq \bigcup_n \bigcup_{f^{(n)}, \left(f_l^{(n)}\right)_{l \in \mathcal{L}}} \mathcal{R}_{*}^{MHO}\left(n,\left(f_l^{(n)}\right)_{l \in \mathcal{L}}\right),\end{aligned}$$ where $$\begin{aligned}
\mathcal{R}_{*}^{MHO}\left(n,\left(f_l^{(n)}\right)_{l \in \mathcal{L}}\right) \triangleq \Biggr \{ \bigr(R, {\mathbf}{R},E \bigr) :
R &\ge \frac{1}{n} \log \left|f^{(n)}\left({X}^n\right)\right|, \nonumber\\
R_l &\ge \frac{1}{n} \log \left|f_l^{(n)}\left({X}_l^n\right)\right| \hspace {0.05 in} \textrm{for all} \hspace {0.05 in} l \hspace {0.05 in} \textrm{in} \hspace {0.05 in}\mathcal{L}, \hspace {0.05 in} \textrm{and} \\
E &\le \frac{1}{n} I\left(Y^n; f^{(n)} (X^n),\left(f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}\right) \Biggr \}.\nonumber\end{aligned}$$
$\mathcal{R}^{MHO} = \overline{\mathcal{R}_{*}^{MHO}}$.
The proof of this result is almost identical to that of Proposition 1. Define the set $$\begin{aligned}
\tilde{\mathcal{R}}^{MHO} \triangleq \Biggr \{&(R, R_1,\dots,R_L, E) : \hspace{0.05in} \textrm{there exists} \hspace{0.05in} (r_1,\dots,r_L) \in \mathbb{R}_{+}^L \hspace{0.05in} \textrm{such that} \\
&R_l \ge r_l \hspace{0.05in}\textrm{for all} \hspace{0.05in} l \hspace{0.05in} \textrm{in} \hspace{0.05in} \mathcal{L}, \hspace{0.05in} \textrm{and}\\
&R + \sum_{l \in S} R_l \ge \frac{1}{2} \log^{+} \left [ \frac{1}{D } \left (\frac{1}{\sigma^2_X} + \sum_{l \in S^c} \frac{1- 2^{- 2 r_l}}{\sigma^2_{N_l}} \right )^{-1}\right ] + \sum_{l \in S} r_l \hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L} \Biggr\},\end{aligned}$$ where $$D = (\sigma^2_X+\sigma^2_N) 2^{-2E} - \sigma^2_N.$$
The rate-exponent region of this problem $$\mathcal{R}^{MHO} = \tilde{\mathcal{R}}^{MHO}.$$
The proof of inclusion $\mathcal{R}^{MHO} \subseteq \tilde{\mathcal{R}}^{MHO}$ is similar to the converse proof of the Gaussian many-help-one source coding problem by Oohama [@Oohama2005] and Prabhakaran *et al.* [@Vinod] (see also [@Wagner3]). Their proofs continue to work if we replace the original mean square error distortion constraint with the mutual information constraint that we have here. It is noteworthy though that Wang *et al.*’s [@wang] approach does not work here because it relies on the distortion constraint.
We start with the continuous extension of Theorem 2. Let $\Lambda_o$ be the set of random variables $\lambda_o = (U,{\mathbf}{U}, W, T)$ such that each take values in a finite-dimensional Euclidean space and collectively they satisfy
1. $(W,T)$ is independent of $(X,{\mathbf}{X},Y)$,
2. ${U} \leftrightarrow ({X},W,T) \leftrightarrow ({\mathbf}{U}, {\mathbf}{X},Y)$,
3. ${U}_l \leftrightarrow ({X}_l,W,T) \leftrightarrow (U,{\mathbf}{U}_{l^c}, X,{\mathbf}{X}_{l^c},Y)$ for all $l$ in $\mathcal{L}$, and
4. the conditional distribution of $U_l$ given $(W,T)$ is discrete for each $l$.
Define the set $$\begin{aligned}
\mathcal{R}_o^{MHO}(\lambda_o) \triangleq \biggr \{ (R,{\mathbf}{R},E) :
R_l &\ge I(X_l ; U_l | X, W, T) \hspace{0.05in} \textrm{for all} \hspace{0.05in} l \hspace{0.05in} \textrm{in} \hspace{0.05in} \mathcal{L}, \hspace {0.05 in} \\
R+\sum_{l \in S}R_l &\ge I({X}; U,{\mathbf}{U}_S|{\mathbf}{U}_{S^c},T) + \sum_{l \in S} I(X_l ; U_l | X, W, T) \hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
E &\le I(Y;U,{\mathbf}{U}|T) \biggr \}.\end{aligned}$$ Finally, let $$\begin{aligned}
\mathcal{R}_o^{MHO} \triangleq \bigcup_{\lambda_o \in \Lambda_o} \mathcal{R}_o^{CI}(\lambda_o).\nonumber\end{aligned}$$ We have the following lemma.
$\mathcal{R}_{*}^{MHO} \subseteq \mathcal{R}_o^{MHO}.$
The inequalities (16) and (17) can be established as in the proof of Theorem 2. In particular, we obtain (16) by considering only those constraints on the sum of rate combinations that include $R$. The inequality (15) is not present in Theorem 2. However, it can be derived easily. We need the following lemma.
[@Wagner3 Lemma 9] If $\lambda_o$ is in $\Lambda_o$, then for all $S \subseteq \mathcal{L}$, $$2^{2I(X;{\mathbf}{U}_S|W,T)} \le 1 + \sum_{l \in S} \frac{1 - 2^{-2I(X_l;U_l|X,W,T)}}{\sigma^2_{N_l}/\sigma^2_X}.$$
Consider any $(R,{\mathbf}{R},E)$ in $\mathcal{R}_o^{MHO}$. Then there exists $\lambda_o$ in $\Lambda_o$ such that for all $S \subseteq \mathcal{L}$, $$\begin{aligned}
R+\sum_{l \in S}R_l &\ge I({X}; U,{\mathbf}{U}_S|{\mathbf}{U}_{S^c},T) + \sum_{l \in S} I(X_l ; U_l | X, W, T) \nonumber\\
&=I({X}; U,{\mathbf}{U}|T) - I({X}; {\mathbf}{U}_{S^c}|T) + \sum_{l \in S} I(X_l ; U_l | X, W, T),\end{aligned}$$ and $$\begin{aligned}
E &\le I(Y;U,{\mathbf}{U}|T).\end{aligned}$$ We can lower bound the first term in (18) by applying the entropy power inequality [@Cover] and obtain $$\begin{aligned}
2^{2 h (Y|U,{\mathbf}{U},T)} &= 2^{2 h (X + N|U,{\mathbf}{U},T)} \\
&\ge 2^{2 h ({X}|U,{\mathbf}{U},T)} + 2^{2 h ({N})} \\
&= 2^{2 h ({X}|U,{\mathbf}{U},T)} + 2 \pi e \sigma^2_N,\end{aligned}$$ which simplifies to $$\begin{aligned}
h (Y|U,{\mathbf}{U},T) \ge \frac{1}{2} \log \left (2^{2 h ({X}|U,{\mathbf}{U},T)} + 2 \pi e \sigma^2_N \right ).\end{aligned}$$ Now (19) and (20) together imply $$\begin{aligned}
I(X; U,{\mathbf}{U}|T) \ge \frac{1}{2} \log \frac{\sigma^2_X}{(\sigma^2_X+\sigma^2_N)2^{-2E}-\sigma^2_N}.\end{aligned}$$ We next upper bound the second term in (18). Since conditioning reduces entropy and $X$ is independent of $(W,T)$, we have $$\begin{aligned}
I({X}; {\mathbf}{U}_{S^c}|T) &= h(X|T) - h(X|{\mathbf}{U}_{S^c},T) \nonumber\\
&\le h(X|W,T) - h(X|{\mathbf}{U}_{S^c},W,T) \nonumber\\
&=I({X}; {\mathbf}{U}_{S^c}|W,T).\end{aligned}$$ Define $$r_l \triangleq I(X_l ; U_l | X, W, T).$$ Then we have from (18), (21), (22), and Lemma 4 that $$\begin{aligned}
R+\sum_{l \in S}R_l &\ge \frac{1}{2} \log^{+} \left [ \frac{1}{\Bigr ((\sigma^2_X+\sigma^2_N)2^{-2E}-\sigma^2_N \Bigr)} \left (\frac{1}{\sigma^2_X} + \sum_{l \in S^c} \frac{1- 2^{- 2 r_l}}{\sigma^2_{N_l}} \right )^{-1}\right ] + \sum_{l \in S} r_l.\end{aligned}$$ On applying Lemma 3 and Corollary 2, we obtain $\mathcal{R}^{MHO} \subseteq \tilde {\mathcal{R}}^{MHO}.$
![Gaussian many-help-one source coding problem[]{data-label="fig:Fig6"}](Fig15.pdf){width="3.5in"}
We use the Quantize-Bin-Test scheme to prove the reverse inclusion. Consider $(R, {\mathbf}{R}, E)$ in $\tilde{\mathcal{R}}^{MHO}$. Then there exists ${\mathbf}{r} \in \mathbb{R}^L_{+}$ such that $$\begin{aligned}
&R_l \ge r_l \hspace{0.05in } \textrm{for all \emph{l} in} \hspace{0.05in }\mathcal{L}, \hspace{0.05in } \textrm{and} \\
&R + \sum_{l \in S} R_l \ge \frac{1}{2} \log^{+} \left [ \frac{1}{D} \left (\frac{1}{\sigma^2_X} + \sum_{l \in S^c} \frac{1- 2^{- 2 r_l}}{\sigma^2_{N_l}} \right )^{-1}\right ] + \sum_{l \in S} r_l \hspace{0.05in } \textrm{for all} \hspace{0.05in } S \subseteq \mathcal{L}.\end{aligned}$$ We therefore have from Oohama’s result [@Oohama2005] that $(R, {\mathbf}{R}, D)$ is achievable for the quadratic Gaussian many-help-one source coding problem, the setup of which is shown in Fig. \[fig:Fig6\]. In this problem, the main encoder and helpers operate as before. The decoder however uses all available information to estimate $X$ such that the mean square error of the estimate is no more than a fixed positive number $D$. Since $(R, {\mathbf}{R}, D)$ is achievable, it follows by Oohama’s achievability proof that for any positive $\delta$ and sufficiently large $n$, there exists quantize and bin encoders $f^{(n)}, f^{(n)}_1, \dots, f^{(n)}_L$, and a decoder $\psi^{(n)}$ such that $$\begin{aligned}
R+\delta &\ge \frac{1}{n} \log \left |f^{(n)}(X^n) \right|, \\
R_l+\delta &\ge \frac{1}{n} \log \left|f^{(n)}_l(X_l^n)\right| \hspace{0.05in } \textrm{for all \emph{l} in} \hspace{0.05in } \mathcal{L},\hspace{0.05 in} \textrm{and}\\
D+\delta &\ge \frac{1}{n} \sum_{i=1}^n E\left [\left (X^n(i) - \hat X^n(i)\right)^2\right],\end{aligned}$$ where $$\begin{aligned}
\hat X^n &= \psi^{(n)} \left (f^{(n)}(X^n), \left (f^{(n)}_l(X_l^n) \right)_{l \in \mathcal{L}} \right).\end{aligned}$$ For each $i$, we have $$\begin{aligned}
E \left [\left(Y^n(i) - \hat X^n(i)\right)^2\right] &= E \left [\left(Y^n(i) - X^n(i) + X^n(i)- \hat X^n(i)\right)^2\right] \nonumber\\
&= E \left [\left(N^n(i) + X^n(i)- \hat X^n(i)\right)^2\right] \nonumber\\
&= \sigma^2_N+E \left [\left(X^n(i)- \hat X^n(i)\right)^2\right],\nonumber\end{aligned}$$ where the last equality follows because $$Y^n(i) \leftrightarrow X^n(i) \leftrightarrow \hat X^n(i).$$ By averaging over time, we obtain $$\begin{aligned}
\frac{1}{n} \sum_{i=1}^n E \left [\left(Y^n(i) - \hat X^n(i)\right)^2\right] &= \sigma^2_N+ \frac{1}{n} \sum_{i=1}^n E \left [\left(X^n(i)- \hat X^n(i)\right)^2\right] \nonumber\\
&\le \sigma^2_N + D + \delta,\nonumber\end{aligned}$$ where the last inequality follows from (25). Therefore, the code achieves a distortion $\sigma^2_N + D + \delta$ in $Y$. Hence, $$\frac{1}{n} I \left (Y^n; f^{(n)}(X^n), \left (f^{(n)}_l(X_l^n) \right)_{l \in \mathcal{L}} \right )$$ must be no less than the rate-distortion function of $Y$ at a distortion $\sigma^2_N + D + \delta$, i.e., $$\begin{aligned}
\frac{1}{n} I \left (Y^n; f^{(n)}(X^n), \left (f^{(n)}_l(X_l^n) \right)_{l \in \mathcal{L}} \right ) &\ge \frac{1}{2} \log \frac{\sigma^2_X+\sigma^2_N}{\sigma^2_N + D + \delta} \nonumber\\
&=\frac{1}{2} \log \frac{\sigma^2_X+\sigma^2_N}{ (\sigma^2_X+\sigma^2_N) 2^{-2E}+ \delta} \nonumber\\
&\ge\frac{1}{2} \log \frac{\sigma^2_X+\sigma^2_N}{ (\sigma^2_X+\sigma^2_N) 2^{-2(E - \bar {\delta})}} \\
&= E - \bar {\delta},\end{aligned}$$ where (26) follows for a positive $\bar {\delta}$ such that $\bar {\delta} \rightarrow 0$ as ${\delta}\rightarrow 0$. We now have from (23), (24), and (27) that $(R, {\mathbf}{R}, E)$ is in $\overline{\mathcal{R}_{*}^{MHO}}$. Hence by Corollary 2, $\tilde {\mathcal{R}}^{MHO} \subseteq {\mathcal{R}}^{MHO}$.
Special Cases
-------------
Consider the following special cases. We continue to use the terminology from the source coding literature.
1. *Gaussian CEO hypothesis testing against independence:* When $R=0$, the problem reduces to the Gaussian CEO hypothesis testing against independence problem. Let $\mathcal{R}^{CEO}$ be the rate-exponent region of this problem. Define the set $$\begin{aligned}
\tilde{\mathcal{R}}^{CEO} \triangleq \Biggr \{&(R_1, \dots, R_L, E) : \hspace{0.05in}\textrm{there exists} \hspace{0.05in} {\mathbf}{r} \in \mathbb{R}_{+}^L \hspace{0.05in} \textrm{such that} \\
&\sum_{l \in S} R_l \ge \frac{1}{2} \log^{+} \left [ \frac{1}{D } \left (\frac{1}{\sigma^2_X} + \sum_{l \in S^c} \frac{1- 2^{- 2 r_l}}{\sigma^2_{N_l}} \right )^{-1}\right ] + \sum_{l \in S} r_l \hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L} \Biggr\}.\end{aligned}$$ We immediately have the following corollary as a consequence of Theorem 7.
${\mathcal{R}}^{CEO} = \tilde{\mathcal{R}}^{CEO}.$
2. *Gaussian one-helper hypothesis testing against independence:* When $L=1$, the problem reduces to the Gaussian one-helper hypothesis testing against independence problem. Let $\mathcal{R}^{OH}$ be the rate-exponent region of this problem. Define the sets $$\begin{aligned}
\tilde{\mathcal{R}}^{OH} \triangleq \Biggr \{&(R, R_1, E) : \hspace{0.05in}\textrm{there exists} \hspace{0.05in} r_1 \in \mathbb{R}_{+} \hspace{0.05in} \textrm{such that} \\
&R_1 \ge r_1, \\
&R+R_1 \ge \frac{1}{2} \log^{+} \left [ \frac{\sigma^2_X}{D} \right ] + r_1, \hspace{0.05in} \textrm{and}\\
&R \ge \frac{1}{2} \log^{+} \left [ \frac{1}{D} \left (\frac{1}{\sigma^2_X} + \frac{1- 2^{- 2 r_1}}{\sigma^2_{N_1}} \right )^{-1} \right ] \Biggr\},\end{aligned}$$ and $$\begin{aligned}
\bar{\mathcal{R}}^{OH} \triangleq \Biggr\{&(R, R_1, E) : R \ge \frac{1}{2} \log^{+} \left [ \frac{\sigma^2_X}{ D} \left ( 1- \rho^2 +\rho^2 2^{-2 R_1}\right)\right ] \Biggr\},\end{aligned}$$ where $$\rho^2 = \frac{\sigma^2_X}{\sigma^2_X+\sigma^2_{N_1}}.$$
${\mathcal{R}}^{OH} = \tilde{\mathcal{R}}^{OH} = \bar{\mathcal{R}}^{OH}.$
The first equality follows from Theorem 7. Consider any $(R, R_1, E)$ in $\tilde{\mathcal{R}}^{OH}$. It must satisfy $$\begin{aligned}
R &\ge \min_{0 \le r_1 \le R_1} \max \left \{\frac{1}{2} \log^{+} \left [ \frac{1}{D } \left (\frac{1}{\sigma^2_X} + \frac{1- 2^{- 2 r_1}}{\sigma^2_{N_1}} \right )^{-1} \right ],\hspace{0.05in} \frac{1}{2} \log^{+} \left [ \frac{\sigma^2_X}{D} \right ] + r_1 - R_1 \right \}\nonumber\\
&= \frac{1}{2} \log^{+} \left [ \frac{\sigma^2_X}{ D } \left ( 1- \rho^2 +\rho^2 2^{-2 R_1}\right)\right ],\end{aligned}$$ where the equality is achieved by $$\begin{aligned}
r_1 = R_1 + \frac{1}{2} \log\left( 1- \rho^2 +\rho^2 2^{-2 R_1}\right).\end{aligned}$$ We therefore have that $(R, R_1, E)$ is in $\bar{\mathcal{R}}^{OH}$, and hence $\tilde{\mathcal{R}}^{OH} \subseteq \bar{\mathcal{R}}^{OH}$. The proof of the reverse containment follows by noticing that for any $(R, R_1, E)$ in $\bar{\mathcal{R}}^{OH}$, there exists $r_1$ as in (28) such that all inequalities in the definition of $\tilde{\mathcal{R}}^{OH}$ are satisfied.
A General Outer Bound
=====================
We return to the general problem formulated in Section 3. The problem remains open till date. Several inner bounds are known for $L=1$ [@Han; @Ahl; @Han1; @Han2]. But even for $L=1$, there is no nontrivial outer bound with which to compare the inner bounds. We give an outer bound for a class of instances of the general problem.
Consider the class of instances such that $P_{\mathbf}{X} = Q_{\mathbf}{X}$, i.e., the marginal distributions of ${\mathbf}{X}$ are the same under both hypotheses. Stein’s lemma [@Cover] asserts that the centralized type 2 error exponent for this class of problems is $$E_C \triangleq D\left(P_{{\mathbf}{X}Y} \| Q_{{\mathbf}{X}Y}\right),$$ which is achieved when ${\mathbf}{X}$ and $Y$ both are available at the detector. Let $$\mathcal{R}_{C} \triangleq \{({\mathbf}{R},E) : E \le E_C\}.$$ We have the following trivial centralized outer bound.
$\mathcal{R} \subseteq \mathcal{R}_{C}.$
Let $\Xi$ be the set of random variables $Z$ such that there exists two joint distributions $P_{{\mathbf}{X}YZ}$ and $Q_{{\mathbf}{X}YZ}$ satisfying
1. $\sum_{\mathcal{Z}} P_{{\mathbf}{X}YZ}=P_{{\mathbf}{X}Y},$ the distribution under $H_0$,
2. $\sum_{\mathcal{Z}} Q_{{\mathbf}{X}YZ}=Q_{{\mathbf}{X}Y},$ the distribution under $H_1$,
3. $Q_{{\mathbf}{X}YZ} = Q_{{\mathbf}{X}|Z} Q_{Y|Z}Q_Z,$ i.e., ${\mathbf}{X}$ and $Y$ are conditionally independent given $Z$ under the $Q$ distribution, and
4. $P_{{\mathbf}{X}Z} = Q_{{\mathbf}{X}Z},$ i.e., the joint distributions of $({\mathbf}{X},Z)$ are the same under both distributions.
Note that the joint distributions of $(Y,Z)$ need not be the same under the two distributions. If $P_{{\mathbf}{X}YZ}$ and $Q_{{\mathbf}{X}YZ}$ are the joint distributions of ${\mathbf}{X}$, $Y$, and $Z$ under $H_0$ and $H_1$, respectively and $Z$ is available to the detector, then the problem can be related to the $L$-encoder hypothesis testing against conditional independence. Now $Z$ is not present in the original problem, but we can augment the sample space by introducing $Z$ and supplying it to the decoder. The outer bound for this new problem is then an outer bound for the original problem. Moreover, we can then optimize over $Z$ to obtain the best possible bound.
Let $\chi$ and $\Lambda_o$ be defined as in Section 4.2 with $X_{L+1}$ restricted to be deterministic. If $\Xi$ is nonempty, then for any $(Z,X,\lambda_o)$ in $\Xi \times \chi \times \lambda_o$, define the set $$\begin{aligned}
\mathcal{R}_o(Z,X,\lambda_o) \triangleq \biggr \{ ({\mathbf}{R},E) :
\sum_{l \in S}R_l &\ge I({X}; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},Z,T) + \sum_{l \in S} I(X_l ; U_l | X, W, Z,T) \hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
E &\le I(Y;{\mathbf}{U}|Z,T)+D\left(P_{Y|Z}\|Q_{Y|Z}|Z\right) \biggr \}.\nonumber\end{aligned}$$ Finally, let $$\mathcal{R}_o \triangleq \left\{
\begin{array}{l l}
\bigcap_{Z \in \Xi} \bigcap_{X \in \chi} \bigcup_{\lambda_o \in \Lambda_o} \mathcal{R}_o(Z,X,\lambda_o) & \quad \mbox{if $\Xi$ is nonempty}\\
\mathbb{R}^{L+1}_{+} & \quad \mbox{otherwise.}\\ \end{array} \right.$$ We have the following outer bound to the rate-exponent region of this class of problems.
$\mathcal{R} \subseteq \overline{\mathcal{R}_o} \cap \mathcal{R}_{C}.$
In light of Proposition 1 and Lemma 5, it suffices to show that $$\mathcal{R}_{*} \subseteq \mathcal{R}_o.$$ Consider $({\mathbf}{R},E)$ in $\mathcal{R}_{*}$.Then there exists a block length $n$ and encoders $f_l^{(n)}$ such that $$\begin{aligned}
R_l &\ge \frac{1}{n}\log \left|f_l^{(n)}\left({X}_l^n\right)\right|\hspace{0.05in} \textrm{for all \emph{l} in} \hspace{0.05in} \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
E &\le \frac{1}{n} D\biggr(P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n} \Bigr\| Q_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n}\biggr).\end{aligned}$$ Consider any $Z$ in $\Xi$. Then $$\begin{aligned}
D&\biggr(P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n} \Bigr\| Q_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n}\biggr) \\
&\le D\biggr(P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^nZ^n} \Bigr\| Q_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^nZ^n}\biggr) \nonumber\\
&= D\biggr(P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n\bigr|Z^n} \Bigr\| Q_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n\bigr|Z^n} \Bigr| Z^n\biggr) \nonumber\\
&= D\biggr(P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n\bigr|Z^n} \Bigr\| P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}\bigr|Z^n}Q_{{Y}^n|Z^n} \Bigr| Z^n\biggr) \nonumber\\
&= D\biggr(P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}{Y}^n\bigr|Z^n} \Bigr\| P_{\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}\bigr|Z^n}P_{{Y}^n|Z^n} \Bigr| Z^n\biggr) + D\Bigr(P_{{Y}^n|Z^n} \bigr\| Q_{{Y}^n|Z^n} \bigr| Z^n\Bigr) \nonumber\\
&= I\biggr(\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}};{Y}^n \Bigr|Z^n\biggr)+ nD\left(P_{{Y}|Z} \| Q_{{Y}|Z} | Z\right), \nonumber\end{aligned}$$ which together with (30) implies $$\begin{aligned}
E \le \frac{1}{n}I\biggr(\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}};{Y}^n \Bigr|Z^n\biggr) + D\left(P_{{Y}|Z} \| Q_{{Y}|Z} | Z\right).\end{aligned}$$ It now follows from (29), (31), and Corollary 1 that $\left({\mathbf}{R},\left(E-D\left(P_{{Y}|Z} \| Q_{{Y}|Z} | Z\right)\right)^{+}\right)$ is in $\mathcal{R}_{*}^{CI}$. Therefore from Theorem 2, it must also be in $\mathcal{R}^{CI}_o$. Hence for any $X$ in $\chi$, there exists $\lambda_o$ in $\Lambda_o$ such that $\left({\mathbf}{R},\left(E-D\left(P_{{Y}|Z} \| Q_{{Y}|Z} | Z\right)\right)^{+}\right)$ is in $\mathcal{R}_o^{CI}(X,\lambda_o)$, i.e., $$\begin{aligned}
\sum_{l \in S}R_l &\ge I({X}; {\mathbf}{U}_S|{\mathbf}{U}_{S^c},Z,T) + \sum_{l \in S} I(X_l ; U_l | X, W, Z,T)\hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
\left(E-D\left(P_{{Y}|Z} \| Q_{{Y}|Z} | Z\right)\right)^{+} &\le I(Y;{\mathbf}{U}|Z,T).\nonumber\end{aligned}$$ This means that $({\mathbf}{R},E)$ is in $\mathcal{R}_o(Z,X,\lambda_o)$, and hence $\mathcal{R}_{*} \subseteq \mathcal{R}_o$.
Although the outer bound above is not computable in general, it simplifies to the following computable form for the special case in which $L=1$. Let $$\begin{aligned}
\tilde {\mathcal{R}} &\triangleq \bigcap_{Z \in \Xi}\Bigr \{(R_1,E): \hspace{0.05in}\textrm{there exists} \hspace{0.05in} U_1 \hspace{0.05in}\textrm{such that} \nonumber\\
&\hspace{1in}R_1 \ge I(X_1 ; U_1 | Z), \nonumber\\
&\hspace{1in}E \le I(Y;{U_1}|Z)+D\left(P_{{Y}|Z} \| Q_{{Y}|Z} | Z\right), \\
&\hspace{1in}| {\mathcal{U}}_1| \le |\mathcal{X}_1| + 1, \hspace{0.05in} \textrm{and} \nonumber\\
&\hspace{1in} U_1 \leftrightarrow X_1 \leftrightarrow (Y, Z)\Bigr \}.\nonumber\end{aligned}$$
For $1$-encoder general hypothesis testing, $\overline{\mathcal{R}_o}=\tilde {\mathcal{R}}$ and hence $\mathcal{R} \subseteq \tilde {\mathcal{R}} \cap \mathcal{R}_{C}.$
It suffices to show that $\overline{\mathcal{R}_o}=\tilde {\mathcal{R}}$. This immediately follows by noticing that given any $Z$ in $\Xi$, the outer bound can be related to the rate-exponent region of the $1$-encoder hypothesis testing against conditional independence problem. The result then follows from Theorem 3.
It is easy to see that the outer bound is tight for the test against independence.
(Test against independence, [@Ahl]) If $Q_{X_1Y} = P_{X_1} P_Y$, then $$\mathcal{R} = \tilde{\mathcal{R}}.$$
This follows by choosing $Z$ to be deterministic in the outer bound and then invoking the result of Ahlswede and Csiszár [@Ahl].
*Remark 2:* The outer bound is not always better than the centralized outer bound. In particular, if $$D\left(P_{{Y}|Z} \| Q_{{Y}|Z} | Z\right) \ge E_C$$ for all $Z$ in $\Xi$, then the outer bound is no better than the centralized outer bound.
Gaussian Case
-------------
To illustrate this bound, let us consider a Gaussian example in which $X_1$ and $Y$ are zero-mean unit-variance jointly Gaussian sources with the correlation coefficients $\rho_0$ and $\rho_1$ under $H_0$ and $H_1$, respectively, where $\rho_0 \neq \rho_1$, $\rho_0^2 < 1,$ and $\rho_1^2 < 1$. We can assume without loss of generality that $0 \le \rho_1 < 1$ because the case $-1 < \rho_1 \le 0$ can be handled by multiplying $Y$ by $-1$. We use lowercase $p$ and $q$ to denote appropriate Gaussian densities under hypotheses $H_0$ and $H_1$, respectively. Let $\mathcal{R}^{G}$ be the rate-exponent region of this problem. We focus on the following three regions (Fig. \[fig:Fig7\]) for which the outer bound is nontrivial. $$\begin{aligned}
\mathcal{D}_1 &\triangleq \{(\rho_0,\rho_1) : 0 \le \rho_1 < \rho_0 < 1\}, \\
\mathcal{D}_2 &\triangleq \{(\rho_0,\rho_1) : 0 \le \rho_1 \hspace{0.05in}\textrm{and}\hspace{0.05in} 2 \rho_1 - 1 \le \rho_0 < \rho_1\}, \\
\mathcal{D}_3 &\triangleq \left \{(\rho_0,\rho_1) : -1 < \rho_0 \le 2 \rho_1 - 1 \hspace{0.05in}\textrm{and}\hspace{0.05in} \frac{ 2(\log e) \rho_1}{1-\rho_1} \le \frac{1}{2} \log \left(\frac{1-\rho_1^2}{1-\rho_0^2}\right) - \frac{(\log e) \rho_1(\rho_0-\rho_1)}{1-\rho_1^2} \right\}.\end{aligned}$$
![Regions of pair $(\rho_0,\rho_1)$ for which the outer bound is nontrivial[]{data-label="fig:Fig7"}](Region1.pdf){width="3.6in"}
### Outer Bound
Let us define $$\begin{aligned}
\rho &\triangleq \left\{
\begin{array}{l l}
\frac{\rho_0 - \rho_1}{1 - \rho_1} & \quad \mbox{if $(\rho_0,\rho_1)$ is in $\mathcal{D}_1 \cup \mathcal{D}_2$}\\
\frac{\rho_0 + \rho_1}{1 - \rho_1} & \quad \mbox{if $(\rho_0,\rho_1)$ is in $\mathcal{D}_3$}. \\ \end{array} \right.\end{aligned}$$ and $$\begin{aligned}
C &\triangleq \left\{
\begin{array}{l l}
0 & \quad \mbox{if $(\rho_0,\rho_1)$ is in $\mathcal{D}_1 \cup \mathcal{D}_2$}\\
\frac{2 (\log e) \rho_1}{1 - \rho_1} & \quad \mbox{if $(\rho_0,\rho_1)$ is in $\mathcal{D}_3$}. \\ \end{array} \right.\end{aligned}$$ The centralized type 2 error exponent is $$\begin{aligned}
E^G_C &\triangleq D\left(p_{{X}_1Y} \| q_{{X}_1Y}\right) \\
&= \frac{1}{2} \log \left(\frac{1-\rho_1^2}{1-\rho_0^2}\right) - \frac{(\log e) \rho_1(\rho_0-\rho_1)}{1-\rho_1^2}.\end{aligned}$$ Define the sets $$\begin{aligned}
{\mathcal{R}}^{G}_o \triangleq & \Biggr \{(R_1,E) : E \le \frac{1}{2} \log \left ( \frac{1}{1 - {\rho}^2 + {\rho}^2 2^{-2R_1}}\right) + C\Biggr\}\end{aligned}$$ and $$\mathcal{R}^G_{C} \triangleq \left\{({R_1},E) : E \le E^G_C\right\}.$$ We have the following outer bound.
If $(\rho_0,\rho_1)$ is in $\mathcal{D}_1 \cup \mathcal{D}_2 \cup \mathcal{D}_3$, then $$\mathcal{R}^{G} \subseteq {\mathcal{R}}^{G}_o \cap \mathcal{R}^{G}_C.$$
The proof is in two steps: obtain a single letter outer bound similar to the one in Corollary 5 and then use it to obtain the desired outer bound. Consider $(\rho_0,\rho_1)$ in $\mathcal{D}_1$. Let $Z$, $Z^{'}$, $W$, and $V$ be standard normal random variables independent of each other. ${X}_1$ and ${Y}$ can be expressed as $$\begin{aligned}
X_1 &= \sqrt{\rho_1} Z + \sqrt{\rho_0-\rho_1} Z^{'} + \sqrt{1-\rho_0} W \\
Y &= \sqrt{\rho_1} Z + \sqrt{\rho_0-\rho_1} Z^{'}+ \sqrt{1-\rho_0} V\end{aligned}$$ under $H_0$ and as $$\begin{aligned}
X_1 &= \sqrt{\rho_1} Z + \sqrt{1-\rho_1} W \\
Y &= \sqrt{\rho_1} Z + \sqrt{1-\rho_1} V\end{aligned}$$ under $H_1$. It is easy to verify that conditions (C12) through (C15) are satisfied if we replace the distributions by the corresponding Gaussian densities. Therefore, $Z$ is in $\Xi$. Define the set $$\begin{aligned}
\tilde {\mathcal{R}}^{G} \triangleq \Bigr \{(R_1,E) : \hspace{0.05in}&\textrm{there exists} \hspace{0.05in} U_1 \hspace{0.05in}\textrm{such that} \nonumber\\
&R_1 \ge I({X}_1; U_1|{Z}), \\
&E \le I({Y} ;U_1 |{Z})+D(p_{Y|Z}\|q_{Y|Z}|Z),\hspace{0.05in} \textrm{and}\\
&({Y,Z}) \leftrightarrow {X}_1\leftrightarrow U_1 \Bigr \}.\end{aligned}$$
$\mathcal{R}^{G} \subseteq \overline{\tilde {\mathcal{R}}^{G}} \cap \mathcal{R}^{G}_C$.
The proof is immediate as a continuous extension of Corollary 5. From Corollary 7, it suffices to show that $$\tilde {\mathcal{R}}^{G} \subseteq {\mathcal{R}}^{G}_o.$$ Note first that $$\begin{aligned}
D(p_{Y|Z}\|q_{Y|Z}|Z) = 0\end{aligned}$$ here because the joint densities of $(Y,Z)$ are the same under both hypotheses. Consider any $(R_1,E)$ in $\tilde {\mathcal{R}}^{G}$. Then there exists a random variable $U_1$ such that $({Y,Z}) \leftrightarrow {X}_1\leftrightarrow U_1$, $$\begin{aligned}
R_1 &\ge I({X}_1; U_1|{Z}), \hspace{0.05in} \textrm{and}\\
E &\le I({Y} ;U_1 |{Z}).\end{aligned}$$ Since $X_1, Y$, and $Z$ are jointly Gaussian under $H_0$, we can write that $$Y = {\rho} X_1 + \sqrt{\rho_1} (1 - {\rho}) Z + B,$$ where $B$ is a zero-mean Gaussian random variable with the variance $$\sigma^2_{Y|X_1Z} = (1 - \rho_1)\left (1-{\rho}^2\right),$$ and is independent of $X_1$ and $Z$. We now have $$\begin{aligned}
h(Y|U_1,Z) &= h\left({\rho} X_1 + \sqrt{\rho_1 } (1 - {\rho}) Z + B|U_1,Z\right) \nonumber\\
&= h\left({\rho} X_1 + B|U_1,Z\right) \nonumber\\
&\ge \frac{1}{2} \log \left( 2^{2h({\rho}X_1|U_1,Z)} + 2^{2h(B)} \right) \\
&= \frac{1}{2} \log \left( {\rho}^2 2^{2h(X_1|U_1,Z)} + 2^{2h(B)} \right) \nonumber\\
&= \frac{1}{2} \log \left( {\rho}^2 2^{2\left(h(X_1|Z) - I(X_1;U_1|Z)\right)} + 2^{2h(B)} \right) \nonumber\\
&= \frac{1}{2} \log \left( {\rho}^2 (1 - \rho_1) 2^{-2 I(X_1;U_1|Z)} + (1 - \rho_1)\left (1-{\rho}^2 \right) \right) + \frac{1}{2} \log (2 \pi e)\nonumber\\
&\ge \frac{1}{2} \log \left( {\rho}^2 (1 - \rho_1) 2^{-2 R_1} + (1 - \rho_1)\left (1-{\rho}^2 \right)\right) + \frac{1}{2} \log (2 \pi e),\end{aligned}$$ where
1. follows from the entropy power inequality [@Cover] because $X_1$ and $B$ are independent given $(U_1,Z)$, and
2. follows because function $$f(x) = \frac{1}{2} \log \left( p 2^{-2x} + q \right)$$ is monotonically decreasing in $x$ for $p > 0$, and we have the rate constraint in (32).
Now (33) and (35) imply $$\begin{aligned}
E &\le \frac{1}{2}\log \left (\frac{\sigma^2_{Y|Z}}{{\rho}^2 (1 - \rho_1) 2^{-2 R_1} + (1 - \rho_1)\left (1-{\rho}^2 \right)} \right ) \nonumber\\
&=\frac{1}{2} \log \left ( \frac{1}{1 - {\rho}^2 + {\rho}^2 2^{-2R_1}}\right),\end{aligned}$$ which proves that $(R_1,E)$ is in ${\mathcal{R}}^{G}_o$. This completes the proof for the region $\mathcal{D}_1$.
The proof is analogous for $(\rho_0,\rho_1)$ in the region $\mathcal{D}_2$. The only difference is that under $H_0$, ${X}_1$ and ${Y}$ can now be expressed as $$\begin{aligned}
X_1 &= \sqrt{\rho_1} Z + \sqrt{\rho_1-\rho_0} Z^{'} + \sqrt{1-2 \rho_1+\rho_0} W \\
Y &= \sqrt{\rho_1} Z - \sqrt{\rho_1-\rho_0} Z^{'}+ \sqrt{1-2 \rho_1+\rho_0} V.\end{aligned}$$
Suppose now that $(\rho_0,\rho_1)$ is in $\mathcal{D}_3$. One can verify that $-\rho_0-\rho_1 > 0$ here. Hence, $X_1$ and $Y$ can be expressed as $$\begin{aligned}
X_1 &= \sqrt{\rho_1} Z + \sqrt{-\rho_0-\rho_1} Z^{'} + \sqrt{1+\rho_0} W \\
Y &= -\sqrt{\rho_1} Z - \sqrt{-\rho_0-\rho_1} Z^{'}+ \sqrt{1+\rho_0} V\end{aligned}$$ under $H_0$. Their expressions under $H_1$ are the same as before. It is evident that $Z$ is in $\Xi$. Therefore, the outer bound in Corollary 7 is valid for this case, which implies that it suffices to show that $$\tilde{\mathcal{R}}^{G} \subseteq {\mathcal{R}}^{G}_o.$$ Under $H_0$, the conditional distribution of $Y$ given $Z=z$ is Gaussian with the mean $-\sqrt{\rho_1}z$ and the variance $1-\rho_1$. Similarly under $H_1$, it is Gaussian with the mean $\sqrt{\rho_1}z$ and the variance $1-\rho_1$. We therefore obtain $$\begin{aligned}
D(p_{Y|Z}\|q_{Y|Z}|Z) &= \int_{z\in \mathbb{R}} p_Z(z) dz\int_{y\in \mathbb{R}} p_{Y|Z}(y|z) \log \frac{p_{Y|Z}(y|z)}{q_{Y|Z}(y|z)} dy \\
&= \int_{z\in \mathbb{R}} p_Z(z) dz\int_{y\in \mathbb{R}} p_{Y|Z}(y|z) \log \left[\exp\left(\frac{(y-\sqrt{\rho_1}z)^2}{2(1-\rho_1)}-\frac{(y+\sqrt{\rho_1}z)^2}{2(1-\rho_1)}\right) \right] dy\\
&= \int_{z\in \mathbb{R}} p_Z(z) dz\int_{y\in \mathbb{R}} p_{Y|Z}(y|z) \left[-\frac{2(\log e)\sqrt{\rho_1}yz}{1-\rho_1}\right] dy\\
&= -\frac{2(\log e)\sqrt{\rho_1}}{1-\rho_1} \int_{z\in \mathbb{R}} z p_Z(z) dz\int_{y\in \mathbb{R}} yp_{Y|Z}(y|z) dy\\
&= -\frac{2(\log e)\sqrt{\rho_1}}{1-\rho_1} \int_{z\in \mathbb{R}} z p_Z(z) dz\left(-\sqrt{\rho_1}z\right)\\
&= \frac{2(\log e)\rho_1}{1-\rho_1} \int_{z\in \mathbb{R}} z^2 p_Z dz\\
&=\frac{2 (\log e) \rho_1}{1 - \rho_1}.\end{aligned}$$ Again, since $X_1, Y,$ and $Z$ are jointly Gaussian under $H_0$, we can write $$Y = {\rho} X_1 - \sqrt{\rho_1} (1 + {\rho}) Z + B,$$ where $B$ is defined as before. The rest of the proof is identical to the region $\mathcal{D}_1$ case.
### Ahlswede and Csiszár’s Inner Bound
We next compare the outer bound with Ahlswede and Csiszár’s inner bound, which is obtained by using a Gaussian test channel to quantize $X_1$. One can use better inner bounds [@Han1; @Han2], but they are quite complicated and for the Gaussian case considered here, Ahlswede and Csiszár’s bound itself is quite close to our outer bound in some cases. Let $$\begin{aligned}
{\mathcal{R}}^{G}_i \triangleq & \Biggr \{(R_1,E) : E \le \frac{1}{2} \log {\left (}\frac{1 - \rho^2_1 {\left (}1 - 2^{-2 R_1} {\right )}}{1 - \rho^2_0 {\left (}1 - 2^{-2 R_1} {\right )}}{\right )}- \frac{(\log e)\rho_1 {\left (}\rho_0 - \rho_1 {\right )}{\left (}1 - 2^{-2 R_1} {\right )}}{1 - \rho^2_1 {\left (}1 - 2^{-2 R_1} {\right )}} \Biggr\}.\end{aligned}$$
[@Ahl] ${\mathcal{R}}^{G}_i \subseteq {\mathcal{R}}^{G}.$
Fix any $(R_1,E)$ in ${\mathcal{R}}^{G}_i$. Let $U_1 = X_1 + P$, where $P$ is a zero-mean Gaussian random variable independent of $(X_1,Y)$ such that $$I(X_1;U_1) = R_1,$$ which implies that the variance of $P$ $$\sigma^2_P = \frac{1}{2^{2R_1}-1}.$$ The covariance matrix of $({U}_1, {Y})$ is $${\mathbf}{K}_0 = {\left [ \begin{array}{cc} 1+\sigma^2_P & \rho_0 \\
\rho_0 & 1 \end{array} \right]}$$ under $H_0$ and is $${\mathbf}{K}_1 = {\left [ \begin{array}{cc} 1+\sigma^2_P & \rho_1 \\
\rho_1 & 1 \end{array} \right]}.$$ under $H_1$. It now follows from Ahlswede and Csiszár’s scheme [@Ahl Theorem 5] that the achievable exponent is $$\begin{aligned}
E_{AC} &= D(p_{U_1Y}\| q_{U_1Y}) \\
&= \int_{{\mathbf}{z} \in \mathbb{R}^2} p_{U_1Y} ({\mathbf}{z}) \log \frac{p_{U_1Y}({\mathbf}{z})}{q_{U_1Y}({\mathbf}{z})} d {\mathbf}{z}\\
&= -\frac{1}{2} \log \left((2 \pi e)^2 \det({\mathbf}{K}_0)\right) - \int_{{\mathbf}{z} \in \mathbb{R}^2} p_{U_1Y} ({\mathbf}{z}) \log {q_{U_1Y} ({\mathbf}{z})} d {\mathbf}{z}\\
&= -\frac{1}{2} \log \left((2 \pi e)^2 \det({\mathbf}{K}_0)\right) - \int_{{\mathbf}{z} \in \mathbb{R}^2} p_{U_1Y} ({\mathbf}{z}) \left [-\frac{(\log e)}{2} {\mathbf}{z}^T {\mathbf}{K}_1^{-1} {\mathbf}{z} -\frac{1}{2} \log \left((2 \pi)^2 \det({\mathbf}{K}_1)\right) \right ] d {\mathbf}{z}\\
&= \frac{1}{2} \log \frac{\det({\mathbf}{K}_1)}{\det({\mathbf}{K}_0)} - (\log e) + \frac{(\log e)}{2} \int_{{\mathbf}{z} \in \mathbb{R}^2} p_{U_1Y} ({\mathbf}{z}) \left({\mathbf}{z}^T {\mathbf}{K}_1^{-1} {\mathbf}{z}\right) d {\mathbf}{z}\\
&= \frac{1}{2} \log \frac{\det({\mathbf}{K}_1)}{\det({\mathbf}{K}_0)} - (\log e) + \frac{(\log e)(1+\sigma^2_P-\rho_0 \rho_1)}{\det({\mathbf}{K}_1)} \\
&= \frac{1}{2} \log \frac{(1+\sigma^2_P- \rho_1^2)}{(1+\sigma^2_P- \rho_0^2)} - \log e + \frac{(\log e)(1+\sigma^2_P-\rho_0 \rho_1)}{(1+\sigma^2_P- \rho_1^2)} \\
&=\frac{1}{2} \log {\left (}\frac{1 - \rho^2_1 {\left (}1 - 2^{-2 R_1} {\right )}}{1 - \rho^2_0 {\left (}1 - 2^{-2 R_1} {\right )}}{\right )}- \frac{(\log e)\rho_1 {\left (}\rho_0 - \rho_1 {\right )}{\left (}1 - 2^{-2 R_1}{\right )}}{1 - \rho^2_1 {\left (}1 - 2^{-2 R_1}{\right )}}.\end{aligned}$$ This proves that $(R_1,E)$ is in $\mathcal{R}^G$.
The inner and outer bounds coincide for the test against independence.
(Test against independence, [@Ahl; @Oohama]) If ${X}_1$ and ${Y}$ are independent under $H_1$, i.e., $\rho_1 = 0$, then $$\mathcal{R}^{G} = {\mathcal{R}}^{G}_o = {\mathcal{R}}^{G}_i.$$
### Numerical Results
![Outer and inner bounds for four examples[]{data-label="fig:Fig8"}](Fig_Outerbound.pdf){width="6.5in"}
Fig. \[fig:Fig8\] shows the inner and outer bounds for four examples. Fig. \[fig:Fig8\](a)-(c) are the examples when $(\rho_0,\rho_1)$ is in $\mathcal{D}_1 \cup \mathcal{D}_2$. Observe that the two bounds are quite close near zero and at all large rates. Fig. \[fig:Fig8\](d) is an example when $(\rho_0,\rho_1)$ is in $\mathcal{D}_3$. For this example, there is a gap between the inner and outer bounds at zero rate. This is due to the fact that in our outer bound, the joint densities of $(Y,Z)$ are different under the two hypotheses. Numerical results suggest that for a fixed $\rho_0$, the maximum gap between the inner and outer bounds decreases as we decrease $\rho_1$ and finally becomes zero at $\rho_1 = 0$, which is the test against independence.
*Remark 3:* The outer bound can be extended to the vector Gaussian case. One can obtain a single letter outer bound similar to the one in Corollary 7. Then the outer bound can be optimized over all choices of $U_1$ by using an invertible transformation [@Chao; @Globerson] and the scalar solution obtained above. It follows from our earlier work that the outer bound is tight for the test against independence [@Rahman].
Acknowledgment {#acknowledgment .unnumbered}
==============
This research was supported by the Air Force Office of Scientific Research (AFOSR) under grant FA9550-08-1-0060.\
\
[**Appendix A:** **Proof of Lemma 1**]{}\
The proof is rather well known and appears in source coding literature quite often. For instance, the similar proof can be found in [@Wagner3]. Let us define $$\begin{aligned}
\bar {\Lambda}_i \triangleq \Bigr \{ \lambda_i = ({\mathbf}{U},T) \in {\Lambda}_i : |\mathcal{U}_l| &\le |\mathcal{X}_l| + 2^{L} - 1 \hspace{0.05in}\textrm{for all} \hspace{0.05in} l \in \mathcal{L}, \hspace{0.05in}\textrm{and} \\
|\mathcal{T}| &\le 2^{L} \Bigr \},\end{aligned}$$ and $$\begin{aligned}
\bar {\mathcal{R}}^{CI}_i \triangleq \bigcup_{\lambda_i \in \bar {\Lambda}_i} \mathcal{R}_i^{CI}(\lambda_i).\nonumber\end{aligned}$$ We want to show that ${\mathcal{R}}_i^{CI}=\bar {\mathcal{R}}_i^{CI}$. We start with the deterministic $T$ case. Consider $\lambda_i = ({\mathbf}{U},T)$ in $\Lambda_i$, where $T$ is deterministic. For any $S \subseteq \mathcal{L}$ containing 1, we have $$I({\mathbf}{X}_S;{\mathbf}{U}_S | {\mathbf}{U}_{S^c},X_{L+1},Z) = H({\mathbf}{X}_S | {\mathbf}{U}_{S^c},X_{L+1},Z) - H({\mathbf}{X}_S | {\mathbf}{U}_{1^c}, U_1, X_{L+1},Z),$$ and for any nonempty $S$ not containing 1, we have $$I({\mathbf}{X}_S;{\mathbf}{U}_S | {\mathbf}{U}_{S^c},X_{L+1},Z) = I({\mathbf}{X}_S;{\mathbf}{U}_S | {\mathbf}{U}_{S^c \setminus \{1\}}, U_1 , X_{L+1},Z).$$ Moreover, $$I(Y; {\mathbf}{U},X_{L+1}|Z) = H(Y|X_{L+1},Z) - H(Y | {\mathbf}{U}_{1^c}, U_1,X_{L+1},Z).$$ It follows from the support lemma [@csiszar Lemma 3.4, pp. 310] that there exists $\bar {U}_1$ with $\bar {\mathcal{U}}_1 \subseteq {\mathcal{U}}_1$ such that $$|\bar {\mathcal{U}}_1| \le |\mathcal{X}_1| + 2^{L} - 1,$$ $$\sum_{u_1 \in \bar {\mathcal{U}}_1} \textrm{Pr}(X_1=x_1| U_1 = u_1) \textrm{Pr}(\bar U_1 = u_1) = \textrm{Pr}(X_1=x_1) \hspace{0.05in} \textrm{for all} \hspace{0.05in} x_1 \hspace{0.05in}\textrm{in} \hspace{0.05in}\mathcal{X}_1\hspace{0.05in} \textrm{but one},$$ $$\begin{aligned}
H({\mathbf}{X}_S | {\mathbf}{U}_{1^c}, U_1, X_{L+1},Z) = H({\mathbf}{X}_S | {\mathbf}{U}_{1^c}, \bar U_1, X_{L+1},Z) \hspace{0.05in} \textrm{for all} \hspace{0.05in} S \hspace{0.05in}\textrm{containing 1},\end{aligned}$$ $$\begin{aligned}
I({\mathbf}{X}_S;{\mathbf}{U}_S | {\mathbf}{U}_{S^c \setminus \{1\}}, U_1 , X_{L+1},Z) = I({\mathbf}{X}_S;{\mathbf}{U}_S | {\mathbf}{U}_{S^c \setminus \{1\}}, \bar U_1 , X_{L+1},Z) \hspace{0.05in} \textrm{for all nonempty} \hspace{0.05in} S \hspace{0.05in}\textrm{not containing 1},\end{aligned}$$ and $$H(Y | {\mathbf}{U}_{1^c}, U_1,X_{L+1},Z) = H(Y | {\mathbf}{U}_{1^c}, \bar U_1,X_{L+1},Z).$$ Since $$U_1 \leftrightarrow X_1 \leftrightarrow ({\mathbf}{U}_{1^c},{\mathbf}{X}_{1^c},X_{L+1},Y,Z),$$ if we replace $U_1$ by $\bar U_1$ then the resulting $\lambda_i$ is in $\Lambda_i$ and $\mathcal{R}_i^{CI}(\lambda_i)$ remains unchanged. By repeating this procedure for $U_2, \dots, U_L$, we conclude that there exists $\bar {\lambda}_i = (\bar {{\mathbf}{U}}, \bar T)$ in $\bar {\Lambda}_i$ such that $\bar T$ is deterministic and $\mathcal{R}_i^{CI}(\lambda_i) = {\mathcal{R}}_i^{CI}(\bar {\lambda}_i)$.
We now turn to general $T$. Consider $\lambda_i = ({\mathbf}{U},T)$ in $\Lambda_i$. Let $({\mathbf}{U},t)$ denote the joint distribution of $({\mathbf}{U},T)$ conditioned on $\{T=t\}$. It follows from the deterministic $T$ case that for each $t$ in $\mathcal{T}$, there exists $\bar {{\mathbf}{U}}$ such that $(\bar {{\mathbf}{U}}, t)$ is in $\bar {\Lambda}_i$ and $\mathcal{R}_i^{CI}({\mathbf}{U},t) = {\mathcal{R}}_i^{CI}(\bar {{\mathbf}{U}}, t)$. Hence, on replacing ${\mathbf}{U}$ by $\bar {{\mathbf}{U}}$ for each $t$ in $\mathcal{T}$, we obtain $(\bar {{\mathbf}{U}}, T)$ in $\Lambda_i$ such that $|\bar {\mathcal{U}}_l| \le |\mathcal{X}_l| + 2^{L} - 1$ for all $l$ in $\mathcal{L}$ and $\mathcal{R}_i^{CI}({\mathbf}{U},T) = {\mathcal{R}}_i^{CI}(\bar {{\mathbf}{U}}, T)$. Now ${\mathcal{R}}_i^{CI}(\bar {{\mathbf}{U}}, T)$ is the set of vectors $({\mathbf}{R},E)$ such that $$\begin{aligned}
\sum_{l \in S} R_l &\ge I({\mathbf}{X}_S;{\mathbf}{U}_S|{\mathbf}{U}_{S^c},X_{L+1},Z,T) \hspace{0.05in}\textrm{for all} \hspace{0.05in} S, \hspace{0.05in} \textrm{and}\\
E &\le I(Y;{\mathbf}{U},X_{L+1}|Z,T).\end{aligned}$$ It again follows from the support lemma that there exists $\bar {T}$ with $\bar {\mathcal{T}} \subseteq \mathcal{T}$ such that $$\begin{aligned}
|\bar {\mathcal{T}}| &\le 2^L, \\
I({\mathbf}{X}_S;{\mathbf}{U}_S|{\mathbf}{U}_{S^c},X_{L+1},Z,T) &= I({\mathbf}{X}_S;{\mathbf}{U}_S|{\mathbf}{U}_{S^c},X_{L+1},Z,\bar T), \hspace{0.05in} \textrm{and}\\
I(Y;{\mathbf}{U},X_{L+1}|Z,T) &= I(Y;{\mathbf}{U},X_{L+1}|Z,\bar T).\end{aligned}$$ We therefore have that $\bar {\lambda}_i = (\bar {{\mathbf}{U}}, \bar T)$ is in $\bar {\Lambda}_i$ and $\mathcal{R}_i^{CI}(\lambda_i) = \bar {\mathcal{R}}_i^{CI}(\bar {\lambda}_i).$ This proves $\mathcal{R}_i^{CI} \subseteq \bar {\mathcal{R}}_i^{CI}$, and hence $\mathcal{R}_i^{CI} = \bar {\mathcal{R}}_i^{CI}$ because the reverse containment trivially holds.
For part (b), it suffices to show that $\bar {\mathcal{R}}_i^{CI}$ is closed. Consider any sequence $\left({\mathbf}{R}^{(n)},E^{(n)}\right)$ in $\bar {\mathcal{R}}_i^{CI}$ that converges to $({\mathbf}{R},E)$. Since conditional mutual information is a continuous function, $\bar {\Lambda}_i$ is a compact set. Hence, there exists a sequence $\lambda_i^{(n)} = \left({\mathbf}{U}^{(n)},T^{(n)}\right)$ in $\bar {\Lambda}_i$ that converges to $\lambda_i = \left({\mathbf}{U},T\right)$ in $\bar {\Lambda}_i$ such that $\left({\mathbf}{R}^{(n)},E^{(n)}\right)$ is in $\bar {\mathcal{R}}_i^{CI}\left (\lambda_i^{(n)}\right)$, i.e., $$\begin{aligned}
\sum_{l \in S} R_l^{(n)} &\ge I\left({\mathbf}{X}_S;{\mathbf}{U}^{(n)}_S \bigr|{\mathbf}{U}^{(n)}_{S^c},X_{L+1},Z,T^{(n)}\right) \hspace{0.05in}\textrm{for all} \hspace{0.05in} S, \hspace{0.05in} \textrm{and}\\
E^{(n)} &\le I\left(Y;{\mathbf}{U}^{(n)},X_{L+1}\bigr|Z,T^{(n)}\right).\end{aligned}$$ Again, by the continuity of conditional mutual information, this implies that $$\begin{aligned}
\sum_{l \in S} R_l &\ge I({\mathbf}{X}_S;{\mathbf}{U}_S|{\mathbf}{U}_{S^c},X_{L+1},Z,T) \hspace{0.05in}\textrm{for all} \hspace{0.05in} S, \hspace{0.05in} \textrm{and}\\
E &\le I(Y;{\mathbf}{U},X_{L+1}|Z,T).\end{aligned}$$ We thus have that $({\mathbf}{R},E)$ is in $\bar {\mathcal{R}}_i^{CI}$.\
\
[**Appendix B:** **Proof of Theorem 1**]{}\
We prove the deterministic $T$ case. The general case follows by time sharing. Consider any $\lambda_i = ({\mathbf}{U},T)$ in $\Lambda_i$ with $T$ being deterministic. Consider $({\mathbf}{R},E)$ such that $$\begin{aligned}
\sum_{l \in S} R_l &\ge I({\mathbf}{X}_S;{\mathbf}{U}_S|{\mathbf}{U}_{S^c},X_{L+1},Z) \hspace{0.05in} \textrm{for all} \hspace{0.05in} S \subseteq \mathcal{L}, \hspace{0.05in} \textrm{and} \\
E &\le I(Y;{\mathbf}{U},X_{L+1}|Z).\end{aligned}$$ It suffices to show that $({\mathbf}{R},E)$ belongs to the rate-exponent region $\mathcal{R}^{CI}$.
Consider a sufficiently large block length $n$, $\epsilon > 0$, and $\mu > 0$. For each $l$ in $\mathcal{L}$, let $\bar R_l = I(X_l;U_l) + \alpha$, where $\alpha > 0$. To construct the codebook of encoder $l$, we first generate $2^{n \bar R_l}$ independent codewords $U_l^n$, each according to $\prod_{i=1}^n P_{U_l}(u_{li})$, and then distribute them uniformly into $2^{n (R_l+\epsilon)}$ bins. The codebooks and the bin assignments are revealed to the encoders and the detector. The encoding is done in two steps: quantization and binning. The encoder $l$ first quantizes $X_l^n$ by selecting a codeword $U_l^n$ that is jointly $\mu$-typical with it. We adopt the typicality notion of Han [@Han1]. If there is more than one such codeword, then the encoder $l$ selects one of them arbitrarily. If there is no such codeword, it selects an arbitrary codeword. The encoder then sends to the detector the index of the bin to which the codeword $U_l^n$ belongs. In order to be consistent with our earlier notation, we denote this encoding function by $f_l^{(n)}$. It is clear that the rate constraints are satisfied, i.e., $$\begin{aligned}
\frac{1}{n} \log \left|f_l^{(n)}(X_l^n)\right| = R_l +\epsilon \hspace{0.05in} \textrm{for all} \hspace{0.05in} l \hspace{0.05in} \textrm{in} \hspace{0.05in} \mathcal{L}.\end{aligned}$$
The next lemma is a standard achievability result in distributed source coding.
For any $\delta > 0, \epsilon >0, \mu > 0,$ and all sufficiently large $n$, there exists a function $$\varphi^{(n)} : \prod_{l=1}^L \left \{1,\dots, 2^{n(R_l+\epsilon)} \right \} \times \mathcal{X}^n_{L+1} \times \mathcal{Z}^n \mapsto \prod_{l=1}^L \mathcal{U}_l^n$$ such that (a) if $$V \triangleq \left \{{\mathbf}{U}^n, X_{L+1}^n,Y^n, Z^n \hspace{0.05in} \textrm{are jointly $\mu$-typical under $H_0$} \right \},$$ then $P(V) \ge 1 - \delta$; and (b) $$\begin{aligned}
p_e \triangleq P\left (\varphi^{(n)} \left ((f_l^{(n)}(X_l^n))_{l \in \mathcal{L}},
X_{L+1}^n, Z^n \right) \neq {\mathbf}{U}^n \right) \le \delta.\end{aligned}$$
One can prove this lemma using standard random coding arguments. See [@Berger1; @Tung; @Gastper] for proofs of similar results. Applying this lemma to the hypothesis testing problem at hand, we have $$\begin{aligned}
\frac{1}{n} I &\left(\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}},X_{L+1}^n;{Y}^n \Bigr|Z^n\right) \nonumber\\
&= \frac{1}{n}H\left (Y^n|Z^n \right) - \frac{1}{n}H \left({Y}^n \Bigr|\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}},X_{L+1}^n,Z^n\right)\nonumber\\
&= H\left (Y|Z \right) + \frac{1}{n} H\left (\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}} \Bigr|X_{L+1}^n,Z^n \right) - \frac{1}{n}H \left(\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}},{Y}^n \Bigr |X_{L+1}^n,Z^n\right).\end{aligned}$$ We can lower bound the second term in (39) as $$\begin{aligned}
\frac{1}{n} H \left (\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}} \Bigr|X_{L+1}^n,Z^n \right)\nonumber &=\frac{1}{n} I\left (\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}};{\mathbf}{U}^n \Bigr|X_{L+1}^n,Z^n \right) \nonumber\\
&=\frac{1}{n} H\left ({\mathbf}{U}^n|X_{L+1}^n,Z^n \right) - \frac{1}{n} H \left({\mathbf}{U}^n \Bigr| \left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, X_{L+1}^n,Z^n\right ) \nonumber\\
&\ge \frac{1}{n} H\left ({\mathbf}{U}^n|X_{L+1}^n,Z^n \right) - \frac{1}{n} H \left({\mathbf}{U}^n \Bigr|\varphi^{(n)} \left ( \left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, X_{L+1}^n,Z^n \right )\right ) \\
&\ge \frac{1}{n} H\left ({\mathbf}{U}^n|X_{L+1}^n,Z^n \right) - \frac{1}{n} H_b(p_e) - p_e \sum_{l=1}^L\log |\mathcal{U}_l| \\
&\ge \frac{1}{n} H\left ({\mathbf}{U}^n|X_{L+1}^n,Z^n \right) - \frac{1}{n} - \delta \sum_{l=1}^L\log |\mathcal{U}_l|,\end{aligned}$$ where
1. follows from data processing inequality [@Cover Theorem 2.8.1],
2. follows from Fano’s inequality [@Cover Theorem 2.10.1], and
3. follows Lemma 6(b) and the fact that $H_b(p_e) \le 1$.
The third term in (39) can be upper bounded as $$\begin{aligned}
\frac{1}{n}H \left(\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}},{Y}^n \Bigr|X_{L+1}^n,Z^n\right) &\le \frac{1}{n}H \left({\mathbf}{U}^n, \left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}},{Y}^n \Bigr|X_{L+1}^n,Z^n\right) \nonumber\\
&= \frac{1}{n}H \left({\mathbf}{U}^n,{Y}^n|X_{L+1}^n,Z^n\right).\end{aligned}$$ On applying bounds (42) and (43) into (39), we obtain $$\begin{aligned}
\frac{1}{n} I &\left(\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}},X_{L+1}^n;{Y}^n \Bigr|Z^n\right) \nonumber\\
&\ge H\left (Y|Z \right) + \frac{1}{n} H\left ({\mathbf}{U}^n|X_{L+1}^n,Z^n \right) - \frac{1}{n}H \left({\mathbf}{U}^n,{Y}^n|X_{L+1}^n,Z^n\right) - \frac{1}{n} - \delta \sum_{l=1}^L\log |\mathcal{U}_l|\nonumber\\
&= H\left (Y|Z \right) - \frac{1}{n}H \left({Y}^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n\right) - \frac{1}{n} - \delta \sum_{l=1}^L\log |\mathcal{U}_l| \nonumber\\
&= H\left (Y|Z \right) - \frac{1}{n}H \left({Y}^n,1_V|{\mathbf}{U}^n,X_{L+1}^n,Z^n\right) - \frac{1}{n} - \delta \sum_{l=1}^L\log |\mathcal{U}_l| \nonumber\\
&= H\left (Y|Z \right) - \frac{1}{n}H \left(1_V|{\mathbf}{U}^n,X_{L+1}^n,Z^n\right) - \frac{1}{n}H \left(Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n, 1_V\right) - \frac{1}{n} - \delta \sum_{l=1}^L\log |\mathcal{U}_l| \nonumber\\
&\ge H\left (Y|Z \right) - \frac{1}{n} - \frac{1}{n}H \left(Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n, 1_V=1\right) P(V) \nonumber\\
&\hspace{0.5in}- \frac{1}{n}H \left(Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n, 1_V=0\right) P(V^c) - \frac{1}{n} - \delta \sum_{l=1}^L\log |\mathcal{U}_l|\\
&\ge H\left (Y|Z \right) - \frac{1}{n}H \left(Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n, 1_V=1\right) - \frac{2}{n} -\delta \log |\mathcal{Y}|-\delta \sum_{l=1}^L\log |\mathcal{U}_l|,\end{aligned}$$ where
1. follows from the fact that $H\left(1_V|{\mathbf}{U}^n,X_{L+1}^n,Z^n\right) \le 1$, and
2. follows from Lemma 6(a) and the facts that $$\begin{aligned}
\frac{1}{n}H \left(Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n, 1_V=0\right) &\le \log |\mathcal{Y}|, \\
P(V) &\le 1.\end{aligned}$$
We now proceed to upper bound the second term in (45). Let ${T}^n_{\mu}({\mathbf}{U}X_{L+1}YZ)$ be the set of all jointly $\mu$-typical $({\mathbf}{u}^n,x_{L+1}^n,y^n,z^n)$ sequences. We need the following lemma.
[@Han1 Lemma 1(d)] If $n$ is sufficiently large, then for any $({\mathbf}{u}^n,x_{L+1}^n,y^n,z^n)$ in ${T}^n_{\mu}({\mathbf}{U}X_{L+1}YZ)$, we have $$\begin{aligned}
P_{Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n}(y^n|{\mathbf}{u}^n,x_{L+1}^n,z^n) \ge \exp \left[ -n \left (H \left(Y|{\mathbf}{U},X_{L+1},Z\right) + 2\mu \right)\right] .\end{aligned}$$
Using this lemma, we obtain $$\begin{aligned}
\frac{1}{n}H \left(Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n, 1_V=1\right) &= - \frac{1}{n} \sum_{{T}^n_{\mu}({\mathbf}{U}X_{L+1}YZ)} P_{{\mathbf}{U}^n,X_{L+1}^n,Y^n,Z^n| 1_V=1} \log P_{Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n, 1_V=1} \nonumber\\
&=- \frac{1}{n} \sum_{{T}^n_{\mu}({\mathbf}{U}X_{L+1}YZ)} P_{{\mathbf}{U}^n,X_{L+1}^n,Y^n,Z^n| 1_V=1} \log \frac{P_{Y^n|{\mathbf}{U}^n,X_{L+1}^n,Z^n}}{P_{1_V=1|{\mathbf}{U}^n,X_{L+1}^n,Z^n}} \nonumber\\
&\le \sum_{{T}^n_{\mu}({\mathbf}{U}X_{L+1}YZ)} P_{{\mathbf}{U}^n,X_{L+1}^n,Y^n,Z^n| 1_V=1} \left (H \left(Y|{\mathbf}{U},X_{L+1},Z\right) + 2\mu \right) \nonumber\\
&= H \left(Y|{\mathbf}{U},X_{L+1},Z\right) + 2\mu.\end{aligned}$$ Substituting (46) into (45) gives $$\begin{aligned}
\frac{1}{n} I \left(\left (f_l^{(n)} \left ({X}_l^n \right) \right)_{l \in \mathcal{L}},X_{L+1}^n;{Y}^n \Bigr|Z^n\right)
&\ge I \left(Y;{\mathbf}{U},X_{L+1}|Z\right) - \frac{2}{n} - 2 \mu - \delta \log |\mathcal{Y}| - \delta \sum_{l=1}^L\log |\mathcal{U}_l| \nonumber\\
&\ge E - 3 \mu - \delta \log |\mathcal{Y}| - \delta \sum_{l=1}^L\log |\mathcal{U}_l|,\end{aligned}$$ where the last inequality follows from (37) and the fact that $n$ can be made arbitrarily large. We conclude from (38) and (47) that $$\left(R_1+\epsilon,\dots,R_L+\epsilon,E-3 \mu - \delta \log |\mathcal{Y}| - \delta \sum_{l=1}^L\log |\mathcal{U}_l|\right)$$ is in ${\mathcal{R}}_{*}^{CI}$. Since this is true for any $\delta > 0, \epsilon >0,$ and $\mu > 0,$ we have that $({\mathbf}{R},E)$ is in $\overline{{\mathcal{R}}_{*}^{CI}}$. This together with Corollary 1 implies that $({\mathbf}{R},E)$ is in ${\mathcal{R}}^{CI}.$\
\
[**Appendix C:** **Proof of Theorem 2**]{}\
Suppose $({\mathbf}{R},E)$ is in $\mathcal{R}_{*}^{CI}$. Then there exists a block length $n$ and encoders $f_l^{(n)}$ such that $$\begin{aligned}
R_l &\ge \frac{1}{n}\log \left|f_l^{(n)}\left({X}_l^n\right)\right| \hspace{0.05in} \textrm{for all \emph{l} in} \hspace{0.05in} \mathcal{L}, \hspace {0.05 in} \textrm{and} \\
E &\le \frac{1}{n}I\biggr(\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, X_{L+1}^n;{Y}^n \Bigr|Z^n\biggr).\end{aligned}$$ Consider any $X$ in $\chi$. Let $T$ be a time sharing random variable uniformly distributed over $\{1,\dots,n\}$ and independent of $({\mathbf}{X}^n, X_{L+1}^n,X^n, Y^n, Z^n)$. Define $$\begin{aligned}
X_l &= X_l^n(T) \hspace {0.05 in} \textrm{for each} \hspace {0.05 in} l \hspace {0.05 in}\textrm{in} \hspace {0.05 in} \mathcal{L} \cup \{L+1\},\\
X &= X^n(T), \\
Y &= Y^n(T), \\
Z &= Z^n(T), \\
U_l & = \left(f_l^{(n)}\left({X}_l^n\right), X^n(1 : T-1), X^n_{L+1}(T^c), Z^n(T^c) \right) \hspace {0.05 in} \textrm{for each} \hspace {0.05 in} l \hspace {0.05 in}\textrm{in} \hspace {0.05 in} \mathcal{L}, \hspace {0.05 in}\textrm{and} \\
W &= \left (X^n(T^c),X^n_{L+1}(T^c),Z^n(T^c) \right ).\end{aligned}$$ It is easy to verify that $\lambda_o = ({\mathbf}{U}, W, T)$ is in $\Lambda_o$ and $$X \leftrightarrow ({\mathbf}{X},X_{l+1}, Y,Z) \leftrightarrow \lambda_o.$$ It suffices to show that $({\mathbf}{R}, E)$ is in $\mathcal{R}^{CL}_o(X,\lambda_o)$. We obtain the following from (49) $$\begin{aligned}
E &\le \frac{1}{n} I\left( \left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, X_{L+1}^n;{Y}^n \Bigr|Z^n\right) \nonumber\\
&= \frac{1}{n} \left [H(Y^n|Z^n) - H\left (Y^n\Bigr |\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, X_{L+1}^n,Z^n \right) \right ]\nonumber\\
&=\frac{1}{n}\sum_{i=1}^n \left [ H(Y^n(i)|Z^n(i)) - H\left (Y^n(i)\Bigr |\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, Y^n(1:i-1), X_{L+1}^n,Z^n \right)\right ] \nonumber\\
&\le \frac{1}{n}\sum_{i=1}^n \left [ H(Y^n(i)|Z^n(i)) - H\left (Y^n(i)\Bigr |\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, Y^n(1:i-1), X^n(1:i-1), X_{L+1}^n,Z^n \right)\right ] \\
&= \frac{1}{n}\sum_{i=1}^n \left [ H(Y^n(i)|Z^n(i)) - H\left (Y^n(i)\Bigr |\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, X^n(1:i-1), X_{L+1}^n,Z^n \right)\right ] \\
&= \frac{1}{n}\sum_{i=1}^n I \left (Y^n(i) ; \left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}}, X^n(1:i-1),X^n_{L+1}(i^c),Z^n(i^c), X_{L+1}^n(i) \Bigr| Z^n(i) \right) \nonumber\\
&= I \left (Y^n(T) ; {\mathbf}{U},X^n_{L+1}(T) | Z^n(T), T \right) \nonumber\\
&= I \left (Y ; {\mathbf}{U} , X_{L+1}| Z, T \right),\nonumber\end{aligned}$$ where
1. follows from conditioning reduces entropy, and
2. follows because of the Markov chain $$Y^n(1:i-1) \leftrightarrow \left (X^n(1:i-1),Z^n(1:i-1)\right) \leftrightarrow \left (\left (f_l^{(n)} \left ({X}_l^n \right)\right)_{l \in \mathcal{L}},X_{L+1}^n, Y^n(i), Z^n(i:n) \right).$$
Now let $S \subseteq \mathcal{L}$. Then (48) implies $$\begin{aligned}
n \sum_{l \in S} R_l &\ge \sum_{l \in S} \log \left|f_l^{(n)}\left({X}_l^n\right)\right| \nonumber\\
&\ge \sum_{l \in S} H \left (f_l^{(n)}\left({X}_l^n\right)\right) \nonumber\\
&\ge H \left (\left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S}\right) \nonumber\\
&\ge H \left (\left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S} \Bigr | \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S^c}, X_{L+1}^n,Z^n\right) \\
&= I \left (X^n, {\mathbf}{X}^n_S ; \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S} \Bigr | \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S^c}, X_{L+1}^n,Z^n\right) \nonumber\\
&= I \left (X^n ; \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S} \Bigr | \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S^c}, X_{L+1}^n,Z^n\right) \nonumber\\
&\hspace{0.5in}+ I \left ({\mathbf}{X}^n_S ; \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S} \Bigr | \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S^c}, X^n, X_{L+1}^n,Z^n\right) \nonumber\\
&= \sum_{i=1}^n I \left (X^n(i) ; \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S} \Bigr | \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S^c}, X^n(1:i-1), X_{L+1}^n,Z^n\right) \nonumber\\
&\hspace{0.5in} + \sum_{l \in S}I \left ({X}^n_l ; f_l^{(n)}\left({X}_l^n\right) \Bigr | X^n, X_{L+1}^n,Z^n\right),\end{aligned}$$ where
1. follows from conditioning reduces entropy, and
2. follows because $X$ is in $\chi$.
We next lower bound the second sum in (53). $$\begin{aligned}
I &\left ({X}^n_l ; f_l^{(n)}\left({X}_l^n\right) \Bigr | X^n, X_{L+1}^n,Z^n\right) \nonumber\\
&= \sum_{i=1}^n I \left ({X}^n_l(i) ; f_l^{(n)}\left({X}_l^n\right) \Bigr | X^n, {X}^n_l(1:i-1),X_{L+1}^n,Z^n\right)\nonumber \\
&=\sum_{i=1}^n \left [H\left ({X}^n_l(i) \Bigr | X^n, {X}^n_l(1:i-1),X_{L+1}^n,Z^n\right) - H\left ({X}^n_l(i) \Bigr | f_l^{(n)}\left({X}_l^n\right) , X^n, {X}^n_l(1:i-1),X_{L+1}^n,Z^n\right) \right ]\nonumber \\
&\ge \sum_{i=1}^n \left [H\left ({X}^n_l(i) \Bigr | X^n, X_{L+1}^n,Z^n\right) - H\left ({X}^n_l(i) \Bigr | f_l^{(n)}\left({X}_l^n\right) , X^n, X_{L+1}^n,Z^n\right) \right ] \\
&= \sum_{i=1}^n I\left ({X}^n_l(i) ; f_l^{(n)}\left({X}_l^n\right) \Bigr | X^n, X_{L+1}^n,Z^n\right),\end{aligned}$$ where (54) again follows from conditioning reduces entropy. On applying (55) in (53), we obtain $$\begin{aligned}
\sum_{l \in S} R_l &\ge\frac{1}{n}\sum_{i=1}^n \biggr [ I \left (X^n(i) ; \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S} \Bigr | \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S^c}, X^n(1:i-1), X_{L+1}^n,Z^n\right) \nonumber\\
&\hspace{0.4in} + \sum_{l \in S}I \left ({X}^n_l(i) ; f_l^{(n)}\left({X}_l^n\right) \Bigr | X^n, X_{L+1}^n,Z^n\right) \biggr ].\end{aligned}$$ If $S^c$ is nonempty, then continuing from (56) gives $$\begin{aligned}
\sum_{l \in S} R_l &\ge I \left (X^n(T) ; {\mathbf}{U}_S \bigr | {\mathbf}{U}_{S^c}, X_{L+1}^n(T),Z^n(T),T\right) \\
&\hspace{0.3in}+ \sum_{l \in S}I \left ({X}^n_l(T) ; U_l \bigr | X^n(T), X_{L+1}^n(T),Z^n(T), X^n(T^c),X_{L+1}^n(T^c), Z^n(T^c),T\right) \\
&= I \left (X ; {\mathbf}{U}_S \bigr | {\mathbf}{U}_{S^c}, X_{L+1},Z,T\right) + \sum_{l \in S}I \left ({X}_l ; U_l \bigr | X, W, X_{L+1},Z,T\right).\end{aligned}$$ Finally if $S = \mathcal{L}$, then $$\begin{aligned}
I &\left (X^n(i) ; \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S} \Bigr | \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S^c}, X^n(1:i-1), X_{L+1}^n,Z^n\right) \nonumber \\
&= I \left (X^n(i) ; \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S} \Bigr | X^n(1:i-1), X_{L+1}^n,Z^n\right) \nonumber\\
&= I \left (X^n(i) ; \left (f_l^{(n)}\left({X}_l^n\right) \right )_{l \in S}, X^n(1:i-1), X_{L+1}^n(i^c),Z^n(i^c) \Bigr | X_{L+1}^n(i), Z^n(i)\right).\end{aligned}$$ Substituting (57) into (56) yields $$\begin{aligned}
\sum_{l \in \mathcal{L}} R_l &\ge I \left (X ; {\mathbf}{U} \bigr | X_{L+1},Z,T\right) + \sum_{l \in \mathcal{L}}I \left ({X}_l ; U_l \bigr | X, W,X_{L+1},Z,T\right).\end{aligned}$$ This completes the proof of Theorem 2.\
\
[**Appendix D:** **Proof of Lemma 2**]{}\
It suffices to show that (C6) implies (C7). The other direction immediately follows by letting $\epsilon \rightarrow 0.$ We can assume without loss of generality that $|\mathcal{X}| \ge 2$ because the lemma trivially holds otherwise. Let $\mathcal{X} = \{1,2,\dots,|\mathcal{X}|\}$ be the alphabet set of $X$. Let $P_i$ be the $i$th row of the stochastic matrix $P_{Y|X}$ corresponding to $X = i$. We need the following lemma.
If (C6) holds, then rows $P_i$ corresponding to positive $P_X(i)$ are distinct.
The proof is by contradiction. Suppose that $P_X(1)$ and $P_X(2)$ are positive and $P_1 = P_2.$ Let us define a random variable $U$ as $$U \triangleq \left\{
\begin{array}{l l}
2 & \quad \mbox{if $X = 1,2$}\\
X & \quad \mbox{otherwise.}\\ \end{array} \right.$$ The stochastic matrix $P_{X|U}$ has $$\begin{aligned}
P_{X|U}(1|2) &= \frac{P_X(1)}{P_X(1)+P_X(2)}, \\
P_{X|U}(2|2) &= \frac{P_X(2)}{P_X(1)+P_X(2)}, \hspace{0.05in} \textrm{and} \\
P_{X|U}(i|i) &= 1 \hspace{0.05in} \textrm{ for all} \hspace{0.05in} i \hspace{0.05in} \textrm{in} \hspace{0.05in} \left \{3,4,\dots,|\mathcal{X}| \right\}.\end{aligned}$$ It is easy to see that $Y, X,$ and $U$ form a Markov chain $$\begin{aligned}
Y \leftrightarrow X \leftrightarrow U.\end{aligned}$$ We now have $$\begin{aligned}
H(Y|U) &= \sum_{i = 2}^{|\mathcal{X}|} H(Y|U=i) P_U(i) \nonumber\\
&= H(Y|U=2) P_U(2)+\sum_{i = 3}^{|\mathcal{X}|} H(Y|U=i) P_U(i) \nonumber\\
&= H \left (\sum_{j=1}^{|\mathcal{X}|} P_j P_{X|U}(j|2)\right) P_U(2)+\sum_{i = 3}^{|\mathcal{X}|} H \left (\sum_{j=1}^{|\mathcal{X}|} P_j P_{X|U}(j|i)\right) P_U(i) \nonumber\\
&= H \left (P_2\right) P_U(2)+\sum_{i = 3}^{|\mathcal{X}|} H \left (P_i\right) P_U(i) \nonumber\\
&= \sum_{i = 2}^{|\mathcal{X}|} H \left (P_i \right) P_U(i),\end{aligned}$$ and $$\begin{aligned}
H(Y|X) &= \sum_{j = 1}^{|\mathcal{X}|} H(P_j) P_X(j) \nonumber\\
&= \sum_{j = 1}^{|\mathcal{X}|} H(P_j) \left(\sum_{i=2}^{|\mathcal{X}|}P_{X|U}(j|i)P_U(i)\right) \nonumber\\
&= \sum_{i = 2}^{|\mathcal{X}|} P_U(i) \sum_{j = 1}^{|\mathcal{X}|} P_{X|U}(j|i) H \left (P_j \right) \nonumber\\
&= P_U(2) \sum_{j = 1}^{|\mathcal{X}|} P_{X|U}(j|2) H \left (P_j \right) + \sum_{i = 3}^{|\mathcal{X}|} P_U(i) \sum_{j = 1}^{|\mathcal{X}|} P_{X|U}(j|i) H \left (P_j \right) \nonumber\\
&= P_U(2) H \left (P_2 \right) + \sum_{i = 3}^{|\mathcal{X}|} P_U(i) H \left (P_i \right) \nonumber\\
&= \sum_{i = 2}^{|\mathcal{X}|}P_U(i) H \left (P_i \right).\end{aligned}$$ Now (58) through (60) together imply that $I(X;Y|U) = 0$, and hence $Y \leftrightarrow U \leftrightarrow X.$ However, $$\begin{aligned}
H(X|U) &= \sum_{i = 2}^{|\mathcal{X}|} H(X|U=i) P_U(i) \nonumber\\
&= H(X|U=2) P_U(2) \nonumber\\
&= H_b\left(\frac{P_X(1)}{P_X(1)+P_X(2)}\right) (P_X(1)+P_X(2)) \nonumber\\
&> 0,\end{aligned}$$ which contradicts our assumption that (C6) holds.
Consider any $U$ that satisfies the Markov chain $$U \leftrightarrow X \leftrightarrow Y.$$ We can assume without loss of generality that $P_U(u)$ is positive for all $u$ in $\mathcal{U}$ because only positive $P_U(u)$ contributes to $H(X|U)$ and $I(X;Y|U)$ in conditions (C6) and (C7). Then $$\begin{aligned}
I(X;Y|U) &= H(Y|U) - H(Y|X) \nonumber\\
&= \sum_{u \in \mathcal{U}} H(Y|U=u) P_U(u) - \sum_{i = 1}^{|\mathcal{X}|} P_X(i) H(P_i) \nonumber\\
&= \sum_{u \in \mathcal{U}} H \left (\sum_{i=1}^{|\mathcal{X}|} P_i P_{X|U}(i|u) \right) P_U(u) - \sum_{i = 1}^{|\mathcal{X}|} \left(\sum_{u \in \mathcal{U}} P_{X|U}(i|u)P_U(u)\right) H(P_i) \nonumber\\
&= \sum_{u \in \mathcal{U}} P_U(u) \left [H \left (\sum_{i=1}^{|\mathcal{X}|} P_i P_{X|U}(i|u) \right)- \sum_{i = 1}^{|\mathcal{X}|} P_{X|U}(i|u)H(P_i) \right ] \nonumber\\
&= \sum_{u \in \mathcal{U}} P_U(u) T\left(P_{X|U}(.|u)\right),\end{aligned}$$ where (61) follows by setting $$T\left(P_{X|U}(.|u)\right) \triangleq H \left (\sum_{i=1}^{|\mathcal{X}|} P_i P_{X|U}(i|u) \right)- \sum_{i = 1}^{|\mathcal{X}|} P_{X|U}(i|u)H(P_i).$$ Since entropy is a strictly concave and continuous function, $T$ is a nonnegative continuous function of $P_{X|U}(.|u)$. Moreover, for any $u$ in $\mathcal{U}$, $P_{X|U}(i|u)=0$ for all $i$ in $\mathcal{X}$ such that $P_X(i)=0$. Let $\mathcal{P}$ denote the set of all such $P_{X|U}(.|u)$. Define $$\gamma(\delta) \triangleq \sup_{P \in \mathcal{P}} \{H(P) : T(P) \le \delta \}.$$ It now follows from Lemma 8 that if $T(P)=0$ for some $P$ in $\mathcal{P}$, then $P$ must be a point mass and hence $H(P)=0$. Therefore, $\gamma(0)=0$. We next show that $\gamma$ is continuous at $0$. Consider a nonnegative sequence $\delta_n \rightarrow 0$. Then there exists a sequence of distributions $P_n$ in $\mathcal{P}$ such that $$\begin{aligned}
T(P_n) &\le \delta_n \\
H(P_n) &\ge \frac{\gamma(\delta_n)}{2}.\end{aligned}$$ Now, since the set of all distributions on $\mathcal{X}$ is a compact set, by considering a subsequence, we can assume without loss of generality that $P_n$ converges to $P$ in $\mathcal{P}$. By letting $n \rightarrow \infty$ in (62), we obtain that $T(P) = 0$, i.e., $P$ is a point mass. Therefore, $H(P)=0$. It now follows from (63) that $\gamma(\delta_n)\rightarrow 0=\gamma(0)$ as $n \rightarrow \infty$. Hence, $\gamma$ is continuous at $0$.
Fix $0 < \epsilon < \log |\mathcal{X}|$ (condition (C7) is always true for $\epsilon \ge \log |\mathcal{X}|$). Choose $\epsilon_1 > 0$ such that $\gamma\left (\epsilon_1/{\log |\mathcal{X}|}\right)+\epsilon_1 = \epsilon$. Set $\delta = \left(\epsilon_1/{\log |\mathcal{X}|}\right)^2$. Let $I(X;Y|U) \le \delta$. Define the sets $$\begin{aligned}
\mathcal{U}_1 &\triangleq\left \{u \in \mathcal{U} : T(u) \le \sqrt{\delta}\right\} \hspace{0.05in} \textrm{and} \hspace{0.05in} \mathcal{U}_2 \triangleq \mathcal{U} \setminus \mathcal{U}_1.\end{aligned}$$ Note that $\mathcal{U}_1$ is nonempty because $\delta < 1$. We now have $$\begin{aligned}
\delta &\ge I(X;Y|U) \\
&= \sum_{\mathcal{U}} P_U(u) T(u) \\
&\ge \sum_{\mathcal{U}_2} P_U(u) T(u) \\
&> \sqrt{\delta} \sum_{\mathcal{U}_2} P_U(u),\end{aligned}$$ which implies $$\sum_{\mathcal{U}_2} P_U(u) < \sqrt{\delta}.$$ Hence, $$\begin{aligned}
H(X|U) &= \sum_{\mathcal{U}_1} H(X|U=u) P_U(u) + \sum_{\mathcal{U}_2} H(X|U=u) P_U(u) \nonumber\\
&< \gamma \bigr(\sqrt{\delta} \bigr) + \sqrt{\delta} \log |\mathcal{X}|\\
&= \gamma\left (\epsilon_1/{\log |\mathcal{X}|}\right)+\epsilon_1\\
&= \epsilon.\end{aligned}$$
[10]{} \[1\][\#1]{} url@rmstyle \[2\][\#2]{}
T. He and L. Tong, “Detection of information flows," *IEEE Trans. Inf. Theory*, vol. 54, no. 11, pp. 4925-4945, Nov. 2008. A. Agaskar, T. He, and L. Tong, “Distributed detection of multi-hop information flows with fusion capacity constraints," *IEEE Trans. Sig. Proc.*, vol. 58, no. 6, pp. 3373-3383, June 2010. T. Berger, “Decentralized estimation and decision theory," in *IEEE 7th Spring Workshop on Inf. Theory*, Mt. Kisco, NY, Sept. 1979. T. S. Han and S. Amari, “Statistical inference under multiterminal data compression," *IEEE Trans. Inf. Theory*, vol. 44, no. 6, pp. 2300-2324, Oct. 1998. R. Ahlswede and I. Csiszár, “Hypothesis testing with communication constraints," *IEEE Trans. Inf. Theory*, vol. 32, no. 4, pp. 533-542, July 1986. T. S. Han, “Hypothesis testing with multiterminal data compression," *IEEE Trans. Inf. Theory*, vol. 33, no. 6, pp. 759-772, Nov. 1987. H. Shimokawa, T. S. Han, and S. Amari, “Error bound of hypothesis testing with data compression," in *IEEE Int. Symp. Inf. Theor. Proc.*, 1994, p. 29. D. Slepian and J. K. Wolf, “Noiseless coding of correlated information sources," *IEEE Trans. Inf. Theory*, vol. 19, no. 4, pp. 471-480, July 1973. B. G. Kelly and A. B. Wagner, “Reliability in source coding with side information," preprint. B. G. Kelly, A. B. Wagner, and A. Vamvatsikos, “Error exponents and test channel optimization for the [G]{}aussian [W]{}yner-[Z]{}iv problem," in *IEEE Int. Symp. Inf. Theor. Proc.*, 2008, pp. 414-418. B. G. Kelly and A. B. Wagner, “Error exponents and test channel optimization for the [W]{}yner-[Z]{}iv problem," in *Proc. 45th Annual Allerton Conference*, 2007. Y. Kochman and G. W. Wornell, “On the excess distortion exponent of the quadratic-[G]{}aussian [W]{}yner-[Z]{}iv problem," in *IEEE Int. Symp. Inf. Theor. Proc.*, 2010, p. 36-40. I. Csiszár, “Linear codes for sources and source networks: error exponents, universal coding," *IEEE Trans. Inf. Theory*, vol. 28, no. 4, pp. 585-592, July 1982. B. G. Kelly and A. B. Wagner, “Improved source coding exponents via Witsenhausen’s rate," preprint. I. Csiszár and J. Körner, “Graph decomposition: a new key to coding theorems," *IEEE Trans. Inf. Theory*, vol. 27, no. 1, pp. 5-12, Jan. 1981. A. B. Wagner and V. Anantharam, “An improved outer bound for multiterminal source coding," *IEEE Trans. Inf. Theory*, vol. 54, no. 5, pp. 1919-1937, May 2008. S. Tavildar, P. Viswanath, and A. B. Wagner, “The Gaussian many-help-one distributed source coding problem," *IEEE Trans. Inf. Theory*, vol. 56, no. 1, pp. 564-581, Jan. 2010. A. B. Wagner, S. Tavildar, and P. Viswanath, “Rate region of the quadratic Gaussian two-encoder source-coding problem," *IEEE Trans. Inf. Theory*, vol. 54, no. 5, pp. 1938-1961, May 2008. A. B. Wagner, “On distributed compression of linear functions," in *Proc. 46th Annual Allerton Conference*, 2008, pp. 1546-1553. Ozarow, “On a source-coding problem with two channels and three receivers," *Bell Sys. Tech. J.*, vol. 59, no. 10, pp. 1909-1921, 1980. H. Wang and P. Viswanath, “Vector Gaussian multiple description with two levels of receivers," *IEEE Trans. Inf. Theory*, vol. 55, no. 1, pp. 401-410, Jan. 2009. S. I. Gel‘fand and M. S. Pinsker, “Coding of sources on the basis of observations with incomplete information," (in Russian), *Problemy Peredachi Informatsii*, vol. 15, no. 2, pp. 45-57, April-June 1979. Y. Oohama, “Rate-distortion theory for Gaussian multiterminal source coding systems with several side informations at the decoder," *IEEE Trans. Inf. Theory*, vol. 51, no. 7, pp. 2577-2593, July 2005. V. Prabhakaran, D. Tse, and K. Ramchandran, “Rate region of the quadratic Gaussian CEO problem," in *IEEE Int. Symp. Inf. Theor. Proc.*, 2004, p. 117. C. Tian and J. Chen, “Successive refinement for hypothesis testing and lossless one-helper problem," *IEEE Trans. Inf. Theory*, vol. 54, no. 10, pp. 4666-4681, Oct. 2008. T. Berger, “Multiterminal source coding," in *The Information Theory Approach to Communications*, ser. CISM Courses and Lectures, G. Longo, Ed. Springer-Verlag, 1978, vol. 229, pp. 171-231. S. Y. Tung, “Multiterminal source coding," Ph.D. dissertation, School of Electrical Engineering, Cornell University, Ithaca, NY, May 1978. M. Gastpar, “The Wyner-Ziv problem with multiple sources," *IEEE Trans. Inf. Theory*, vol. 50, no. 11, pp. 2762-2768, Nov. 2004. J. Chen, X. Zhang, T. Berger, and S. B. Wicker, “An upper bound on the sum-rate distortion function and its corresponding rate allocation schemes for the CEO problem," *IEEE J. Select. Areas Commun.*, vol. 22, no. 6, pp. 977-987, Aug. 2004. P. Viswanath, “Sum rate of a class of Gaussian multiterminal source coding problems," in *Advances in Network Information Theory*, ser. DIMACS in Discrete Mathematics and Theoretical Computer Science, P. Gupta, G. Kramer, and A. J. van Wijngaarden, Eds. AMS, 2004, vol. 66, pp. 43-60. A. B. Wagner, B. G. Kelly, and Y. Altuğ, “The lossy one-helper conjecture is false," in *Proc. 47th Annual Allerton Conference*, 2009, pp. 716-723. I. Csiszár and J. Körner, *Information Theory: Coding Theorems for Discrete Memoryless Systems,* 1st ed. Academic, New York, 1981. J. Wang, J. Chen, and X. Wu, “On the minimum sum rate of Gaussian multiterminal source coding: new proofs," in *IEEE Int. Symp. Inf. Theor. Proc.*, 2009, pp. 1463-1467. T. M. Cover and J. A. Thomas, *Elements of Information Theory*, 2nd ed. New York: John Wiley & Sons, 2005. Y. Oohama, “Gaussian multiterminal source coding," *IEEE Trans. Inf. Theory*, vol. 43, no. 6, pp. 1912-1923, Nov. 1997. C. Tian and J. Chen, “Remote vector Gaussian source coding with decoder side information under mutual information and distortion constraints," *IEEE Trans. Inf. Theory*, vol. 55, no. 10, pp. 4676-4680, Oct. 2009. A. Globerson and N. Tishby, “On the optimality of the Gaussian information bottleneck curve," in *Hebrew Univ. Tech. Report*, 2004. Md. S. Rahman and A. B. Wagner, “Vector Gaussian hypothesis testing and lossy one-helper problem," in *IEEE Int. Symp. Inf. Theor. Proc.*, 2009, pp. 968-972.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'K. Hulek and J. Spandaw'
title: Degenerations of abelian surfaces and Hodge structures
---
\[cla\][Conjecture]{} \[cla\][Corollary]{} \[cla\][Lemma]{} \[cla\][Proposition]{} \[cla\][Theorem]{} \[cla\][Remark]{} \[cla\][Remarks]{} \[cla\][Definition]{} \[cla\][Question]{}
Introduction
============
The starting point of this note is twofold: In [@HKW] the first author together with Kahn and Weintraub constructed and described a torodial compactification of the moduli space of $(1,p)$-polarized abelian surfaces with a (canonical) level structure. Moreover, Mumford’s construction was used to associate to each boundary point a degenerate abelian surface. On the other hand Carlson, Cattani and Kaplan gave in [@CCK] an interpretation of torodial compactifications of moduli spaces of abelian varieties in terms of mixed Hodge structures. Here we want to discuss a connection between [@HKW] and [@CCK]. More precisely, we restrict ourselves to corank 1 degenerations in the sense of [@HKW], or equivalently to type II degenerations, i.e. to cycles of elliptic ruled surfaces. Our main result says that a degenerate abelian surface associated to a boundary point is (almost) completely determined by the boundary point (for a precise formulation see theorem \[T\]). The crucial ingredient in the proof is the Local Invariant Cycle theorem which relates the variation of Hodge structure (VHS) to the mixed Hodge structure (MHS) on the singular surface.
In section \[section2\] we collect some basic facts about semi-stable degenerations of abelian surfaces, some of which are well known. The MHS of a cycle of elliptic ruled surfaces is computed in section \[section3\]. Section \[section4\] starts with a review of corank 1 boundary points in moduli spaces of $(1,p)$-polarized abelian surfaces with a level structure. Next the VHS of a specific family associated to such a boundary point is described. Using the Local Invariant Cycle theorem, this and the result from section \[section3\] give theorem \[T\].
We would like to thank J. Steenbrink for helpful discussions.
Both authors are grateful to the DFG for financial support under grant HU 337/2-4.
Preliminaries on degenerations {#section2}
==============================
In this section we want to summarize basic facts on degenerations of abelian surfaces some of which are well known ([@FM; @P]).
Cycles of elliptic surfaces
---------------------------
Let $\Delta=\{z\in{{\Bbb C}}:
|z|<1\}$ be the unit disk. We consider proper, flat families $$\begin{CD}
{{X}}\\
@VV{p}V\\
\Delta
\end{CD}$$ with $X_t$ smooth abelian for $t\neq0$. We shall always assume the total space ${{X}}$ to be smooth and Kähler, and the components of the singular fibre $X_0$ to be algebraic. We also assume $X$ to be relatively minimal.
If $X_0$ has global normal crossings and no triple points then Persson [@P proposition 3.3.1] or [@FM p. 11 and 17] has shown, that $X_0$ is smooth abelian or a cycle of elliptic ruled surfaces, i.e. $$X_0=Y_1\cup\ldots\cup Y_N.$$ The $Y_i$ are smooth ruled elliptic surfaces. $Y_i$ and $Y_{i+1}$ intersect transversally along a smooth curve which is a section of both $Y_i$ and $Y_{i+1}$. In particular, the $Y_i$ are all ruled surfaces over the same base curve $C$ with two disjoint sections. The situation can be envisaged as in figure 1.
(38,76)(10,5) (22,14)
(22,14) (-2,12)[(1,0)[22]{}]{} (-2,0)[(1,0)[22]{}]{} (0,-1)(3,0)[7]{}[(0,1)[14]{}]{}
(12,62) (18,74) (18,62) (18,56)[(0,1)[4]{}]{} (18,56)[(0,-1)[0]{}]{} (18,52)(0,1)[3]{} (18,46)[(0,1)[4]{}]{} (18,46)[(0,-1)[0]{}]{} (12,32) (18,44) (18,32) (18,26)[(0,1)[4]{}]{} (18,26)[(0,-1)[0]{}]{} (12,12) (18,24) (19.5,25.8)[(0,0)\[bl\][$x$]{}]{} (24,12) (23,10)[(0,0)\[tr\][$x{+}s$]{}]{} (33,11.5)[(0,0)\[l\][$C$]{}]{} (24,8)[(0,1)[2]{}]{} (36,8)[(24,6)\[b\]]{} (48,8)[(0,1)[70]{}]{} (33,78)[(30,6)\[t\]]{} (18,78)[(0,-1)[2]{}]{}
We take $Y_1$ to be the bottom ruled surface and $Y_n$ the top ruled surface, and call $s$ the gluing parameter or shift of $X_0$. A careful analysis of Persson’s proof shows that, if one replaces the hypothesis of global normal crossings by local normal crossings, than the only additional case is the following: $X_0$ is irreducible, its normalization is an elliptic ruled surface with two disjoint sections and $X_0$ arises from this by gluing these two sections with a shift $s$. We shall always refer to this situation (including the case $N=1$) as a cycle of elliptic ruled surfaces.
The central fibre of such a degeneration is determined by the following data:
1. the number $N$ of components
2. the base curve $C$
3. line bundles ${{\cal L}}_i$ with ${{\Bbb P}}({{\cal O}}\oplus{{\cal L}}_i)\cong {Y_{{i}}}$
4. the gluing parameter $s$.
We first want to show that these data are not independent of each other. Before we can do this, we fix some more notation. First assume that $N\ge 2$. We denote the intersection of ${Y_{{i}}}$ and ${Y_{{i+1}}}$ by $C_i$. Furthermore we normalize the line bundle ${{\cal L}}_i$ in such a way that ${{\cal O}}_{{Y_{{i}}}}(C_{i-1})|_{C_{i-1}}={{\cal L}}_i$ and hence ${{\cal O}}_{{Y_{{i}}}}(C_{i})|_{C_{i}}={{\cal L}}_i^{-1}$. Under this assumption the line bundle ${{\cal L}}_i$ is uniquely determined. All indices have to be read modulo $N$. The above discussion also makes sense in case $N=1$, if we replace ${Y_{{1}}}$ by its normalization. The next two propositions are special cases of results proved by Persson [@P].
$K_{{{X}}}={{\cal O}}_{{{X}}}$.
Since $K_{{{X}}}|_{X_t}={{\cal O}}_{X_t}$ for $t\neq0$ it follows that $$K_X=\sum_{i=1}^N m_i {Y_{{i}}}$$ where the $m_i$ are uniquely defined up to a common summand. After possibly relabelling the components of $X_0$ we can assume $m_1$ to be maximal. Moreover by adding multiples of a fibre of $p$ we may assume that $m_1=0$. This already gives the result for $N=1$. Now assume $N\geq2$.
Adjunction gives $$(K_{{{X}}}+{Y_{{i}}})|_{{Y_{{i}}}}=K_{{Y_{{i}}}}.$$ Since $$K_{{Y_{{i}}}}=-C_{i-1}-C_{i}+a_i f_{P_i}$$ for a suitable ruling $f_{P_i}$ over a point $P_i\in C$ and $${Y_{{i}}}|_{{Y_{{i}}}} = -C_{i-1}-C_i$$ this implies $$(m_{i-1}-m_i) C_{i-1} +(m_{i+1}-m_i) C_i -a_i f_{P_i} =0.$$ Since $m_1=0$ we find $$m_0 C_0 +m_2 C_1 = a_1 f_{P_1}.$$ Since, moreover, all $m_i\le 0$ this shows that $m_0=m_2=a_1=0$. Continuing in this way we find $m_i=a_i=0$ for all $i$.
$\operatorname{(i)}$ All line bundles ${{\cal L}}_i$ are isomorphic.
$\operatorname{(ii)}$ $\deg {{\cal L}}_i=0$ for all $i$.
\(i) For the normal bundle of $C_i$ in ${{X}}$ we have $$\begin{aligned}
{{\cal N}}_{C_i/{{X}}} &= {{\cal N}}_{C_i/{Y_{{i}}}} \oplus {{\cal N}}_{C_i/{Y_{{i+1}}}}\\
&= {{\cal L}}_i^{-1} \oplus {{\cal L}}_{i+1}.\end{aligned}$$ On the other hand, since $K_{{{X}}}={{\cal O}}_{{{X}}}$ and since $C_i$ is an elliptic curve, we find again by adjunction that $\det
{{\cal N}}_{C_i/{{X}}}={{\cal O}}_{C_i}$. This shows (i).
\(ii) Assume that $\deg{{\cal L}}_i=m\neq0$. Then $|(C_i)_{{Y_{{i}}}}^2|=m$ and $(C_i)_{{Y_{{i+1}}}}^2=-(C_i)_{{Y_{{i}}}}^2$ for all $i$. But then a topological argument (see [@P p. 94]) shows that $$H_1(X_t,{{\Bbb Z}}) ={{\Bbb Z}}^3 \oplus {{\Bbb Z}}_m$$ for general $t$. This contradicts the fact that $X_t$ is abelian.
Additional structures
---------------------
We shall now consider further structures on the family $p:X\to\Delta$. First of all we consider degenerations of [*polarized*]{} abelian surfaces, i.e. we assume that a line bundle ${{\cal O}}_X(1)$ exists on $X$ such that ${{\cal O}}_{X_t}(1)={{\cal O}}(1)|_{X_t}$ for $t\neq0$ is a polarization on the smooth abelian surface $X_t$, and that ${{\cal O}}_{X_0}(1)$ is ample. We are particulary interested in the case where ${{\cal O}}_{X_t}(1)$ represents a polarization of type $(1,p)$ where $p\geq3$ is a prime number.
We also want to assume that $p:X\to\Delta$ is a degeneration of polarized abelian surfaces with a (canonical) level structure. For the concept of (canonical) level structure on $(1,p)$-polarized abelian surfaces see [@HKW I.1]. Before we give a formal definition recall the following: Let $Y={{\Bbb P}}(\cal O\oplus \cal L)$ be a ${{\Bbb P}}^1$-bundle over an elliptic curve $C$ with $\deg\cal L=0$. Let $Y^0$ be the open part of $Y$ which is given by removing the two sections defined by line bundles $\cal O$, resp. $\cal L$. Then $Y^0$ is a ${{\Bbb C}}^*$-bundle over the base curve $C$. More precisely $Y^0$ carries the structure of a commutative complex Lie group, and as such it is an extension of the form $$1\longrightarrow{{\Bbb C}}^*\longrightarrow Y^0\longrightarrow
C\longrightarrow0.$$ In fact $Y^0$ is a semi-abelian surface of rank 1.
Let $p:X\to\Delta$ be a degeneration of $(1,p)$-polarized abelian surfaces. We say that this is a degeneration of $(1,p)$-polarized abelian surfaces with a ([*canonical*]{}) [*level structure*]{} if the following holds:
\(i) There exists an open subset $X^0\subset X$ such that $X^0\to\Delta$ is a family of abelian Lie groups with the following property: $X_t^0=X_t$ for $t\neq0$ and $X_0^0$ is the smooth part of a component of $X_0$ and as such carries the structure of a semi-abelian surface.
\(ii) There exists an action of ${{\Bbb Z}}_p\times{{\Bbb Z}}_p$ on $X$ over $\Delta$ with the following properties: It leaves ${{\cal O}}_X(1)$ invariant and defines a (canonical) level structure on $X_t$ for $t\neq0$ (in particular it operates on $X_t$ by translation by elements of order $p$). Moreover, the subgroup of ${{\Bbb Z}}_p\times{{\Bbb Z}}_p$ which stabilies $X_0^0$ acts on $X_0^0$ as a subgroup.
The two next results show that the presence of a polarization and a level structure imposes strong conditions on the singular fibre $X_0$.
Let $p:{{X}}\to\Delta$ be a degeneration of abelian surfaces with $(1,p)$-polarization and a (canonical) level-structure. Then there are only two possibilities:
$\operatorname{(i)}$ The central fibre consists of one component $X_0$. If $\tilde{X_0}$ is its normalization, then ${{\cal O}}_{\tilde{X_0}}(1)={{\cal O}}_{\tilde{X_0}}(C_0 + p f_P)$ for a suitable point $P$ in the base curve $C$.
$\operatorname{(ii)}$ The central fibre consists of $p$ components ${Y_{{i}}}$, $i=1,\ldots,p$. In this case ${{\cal O}}_{{Y_{{i}}}}(1)={{\cal O}}_{{Y_{{i}}}}(C_{i-1} +
f_{P_i})$ for suitable points $P_i$.
We have $${{\cal O}}_{{Y_{{i}}}}(1) ={{\cal O}}_{{Y_{{i}}}}(a_i C_{i-1} +b_i f_{P_i})$$ for suitable integers $a_i$ and $b_i$. Since these line bundles glue together to give a line bundle on $X_0$ it follows immediately that all $b_i$ are equal. We denote this number by $b$. Since ${{\cal O}}_{X_0}(1)$ is ample, and since we are dealing with a degeneration of abelian surfaces with a $(1,p)$-polarization it follows that $$\label{1}
2b\sum_{i=1}^N a_i = 2p$$ with all $a_i>0$.
We now consider the action of ${{\Bbb Z}}_p\times{{\Bbb Z}}_p$ on the set $\{{Y_{{1}}},\ldots,{Y_{{N}}}\}$. Let $G$ be the stabilizer of ${Y_{{1}}}$. Since $p$ is a prime number, there are three possibilities:
\(i) $G=\{1\}$. Then $N\ge p^2$ and this contradicts formula (\[1\]).
\(ii) $G={{\Bbb Z}}_p$. In this case $N\ge p$. It follows from formula (\[1\]) that $N=p$ and $a_i=b=1$.
\(iii) $G={{\Bbb Z}}_p\times {{\Bbb Z}}_p$. The group ${{\Bbb Z}}_p\times {{\Bbb Z}}_p$ acts on $X_0$ as a group of automorphisms, hence it must leave its singular locus invariant. Since $p$ is an odd prime number the group $G$ must leave the curve $C_0$ invariant.
As a subgroup of $X_0^0$ the group $G$ acts by translation on $C_0$. The multiplicative group ${{\Bbb C}}^\ast$ contains no subgroup isomorphic to ${{\Bbb Z}}_p\times{{\Bbb Z}}_p$, hence the group $G$ must contain at least a subgroup ${{\Bbb Z}}_p$ which acts non-trivially on $C_0$. Since ${{\cal O}}_X(1)$ is invariant under $G$ it follows that the degree of ${{\cal O}}_X(1)$ restricted to $C_0$ must be divisible by $p$. I.e. $b$ must be divisible by $p$. By formula (1) it follows that $b=p$, $N=1$ and $a_1=1$.
It is easy to construct degenerations $p:X\to\Delta$ of $(1,p)$-polarized abelian surfaces without a level structure, such that the polarization on different components $Y_i$ is numerically different.
In lemma 4 we have seen that the central fibre of a degeneration $p:{{X}}\to\Delta$ of $(1,p)$-polarized abelian surfaces with a level-structure has either 1 or $p$ components. Recall that ${Y_{{i}}}={{\Bbb P}}({{\cal O}}\oplus{{\cal L}}_i)$ and that all ${{\cal L}}_i$ are isomorphic and of degree 0 (proposition 2). We shall denote this line bundle by ${{\cal L}}$. Moreover, assume that we have chosen an origin $O$ on the base curve $C$. Then we can consider the shift $s$ as a point on $C$.
Let $p:X\to\Delta$ be a degeneration of $(1,p)$-polarized abelian surfaceswith a (canonical) level structure, and denote the shift of $X_0$ by $s$. Then there are two possibilities:
$\operatorname{(i)}$ $N=1$ and $\cal L={{\cal O}}_C(ps-pO)$
$\operatorname{(ii)}$ $N=p$ and there exist a point $s'\in
C$ with $s=ps'$ such that $\cal L={{\cal O}}(s'-O)$.
\(i) Assume $N=1$. Then ${{\cal O}}_{\tilde{X_0}}(1)={{\cal O}}_{\tilde{X_0}}(C_0+pf_P)$ and hence $${{\cal O}}_{\tilde{X_0}}(1)|_{C_0}={{\cal L}}\otimes {{\cal O}}_C(pP),\qquad
{{\cal O}}_{\tilde{X_0}}(1)|_{C_1}={{\cal O}}_C(pP).$$ Since $C_1$ and $C_0$ are glued with the shift $s$ a necessary and sufficient condition for ${{\cal O}}_{\tilde{X_0}}(1)$ to descend to a line bundle on $X_0$ is $${{\cal L}}\otimes{{\cal O}}_C(pP)={{\cal O}}_C(pP)\otimes{{\cal O}}_C(s-O)^{\otimes p}$$ which gives the claim.
\(ii) Assume $N=p$. Then ${{\cal O}}_{{Y_{{i}}}}(1)={{\cal O}}_{{Y_{{i}}}}(C_{i-1}+f_{P_i})$ and hence $${{\cal O}}_{{Y_{{1}}}}(1)|_{C_1} ={{\cal O}}_C(P_1),\qquad
{{\cal O}}_{{Y_{{2}}}}(1)|_{C_1}={{\cal L}}\otimes {{\cal O}}_C(P_2).$$ From gluing ${Y_{{1}}}$ and ${Y_{{2}}}$ we obtain $${{\cal O}}_C(P_2)={{\cal L}}^{-1}\otimes {{\cal O}}_C(P_1).$$ Continuing in this way, we get $${{\cal O}}_C(P_p)={{\cal L}}^{-(p-1)}\otimes {{\cal O}}_C(P_1).$$ Finally gluing $C^p$ and $C_0$ with a shift $s$ gives the condition $${{\cal L}}\otimes {{\cal O}}_C(P_1) =
{{\cal L}}^{-(p-1)}\otimes {{\cal O}}_C(P_1)\otimes {{\cal O}}_C(s-O)$$ i.e. $${{\cal L}}^p = {{\cal O}}_C(s-O)$$ as claimed.
In [@HKW part II] a number of explicit examples of degenerations of $(1,p)$-polarized abelian surfaces with a level structure were constructed. It was shown that both types of degenerations which were discussed above actually occur. Moreover, all elliptic curves $C$ and all shifts $s$ can be realised.
The mixed Hodge structure on $H^1(X_0)$ {#section3}
=======================================
Let $X_0=\cup_{i\in{{\Bbb Z}}_N} Y_i$ be a cycle of elliptic ruled surfaces as in the previous section. In this section we calculate the MHS on $H^1(X_0)$ and show that the base curve $C$ and the shift $s$ can be recovered from it.
The spectral sequence associated to $X_0$
-----------------------------------------
For technical reasons we want that $X_0$ has global normal crossings. Therefore, we shall first assume $N\geq2$. However, this is not an essential hypothesis (see remark (\[R\])).
The MHS on $H^q(X_0)$ can be computed as follows (see [@G p.103]). Let $Y^0= \bigsqcup_{i\in{{{\Bbb Z}}_N}} Y_i$ and $Y^1=\bigsqcup_{i\in{{{\Bbb Z}}_N}} C_i$, where $C_i=Y_i\cap Y_{i+1}$. The maps $\alpha_i$ and $\beta_i$ are defined as the inclusions $$\alpha_i:C_i{\hookrightarrow}Y_i\text{ and }\beta_i: C_i{\hookrightarrow}Y_{i+1}.$$ Consider the double complex $$A^{pq}=A^q(Y^p)$$ of global $C^{\infty}$ differential forms, where $d:A^{pq}\to
A^{p,q+1}$ is the exterior derivative and $\delta: A^{0q}=\oplus_{i\in{{{\Bbb Z}}_N}} A^q(Y_i) \to A^{1q}
=\oplus_{i\in{{{\Bbb Z}}_N}} A^q(C_i)$ on $A^q(Y_i)$ is given by $(-\beta_{i-1}^{\ast},\alpha_i^{\ast}):A^q(Y_i)\to A^q(C_{i-1})\oplus
A^q(C_i)$. These coboundary maps satisfy the relations $d^2=0$, $d\delta=\delta d$ and $\delta^2=0$. There is a single complex $(A^{\bullet}, D)$ associated to $(A^{\bullet\bullet},d,\delta)$: $$A^k=\oplus_{p+q=k} A^{pq}\quad\text{$D=(-1)^p d+\delta$ on
$A^{pq}$.}$$ We call $
{{\Bbb H}}^{k}:=H_D^{k}(A^{\bullet})
$ the hypercohomology of the double complex.
\[apq\] The hypercohomology ${{\Bbb H}}^k$ is canonically isomorphic to $H^k(X_0)$.
Let $i_p:Y^p\to X_0$ be the natural map. Let ${{\cal A}}^{pq}$ be the sheaf on $X_0$ defined by $${{\cal A}}^{pq}(U)=A^q(i_p^{-1}U).$$ Set ${{\cal A}}^k:=\oplus_{p+q=k}{{\cal A}}^{pq}$ and let $D$ be the sheafified version of the coboundary operator $D$ above. It is shown in [@GS lemma 4.6] that the complex $({{\cal A}}^{\bullet},D)$ is an acyclic resolution of the constant sheaf ${{\Bbb C}}_{X_0}$, hence $H^k(X_0,{{\Bbb C}})=H^k(\Gamma(X_0,{{\cal A}}^{\bullet}))=H_D^k(A^{\bullet})={{\Bbb H}}^k$.
There exists a spectral sequence $E_r^{pq}$ with $E_0^{pq}=A^{pq}$ and $d_0=d$. The map $d_1$ is induced by $\delta$ and, more generally, $d_r:E_r^{pq}\to E_r^{p+r,q-r+1}$ is induced by $D$. Notice that $E_1^{pq}=H^q(Y^p)$. Since $A^{pq}=0$ for $p<0$ and $p>1$, the spectral sequence degenerates at $E_2$, i.e. $E_2^{pq}=E_{\infty}^{pq}$.
The weight filtration $W_{\bullet}$ on ${{\Bbb H}}$ is defined to be the filtration induced by $W_k(A^{\bullet})=\oplus_{q\le k} A^{\ast,q}$. By the theory of spectral sequences we have $E_{\infty}^{pq}={{\operatorname{Gr}}}_{q}^W {{\Bbb H}}^{p+q}$. Since $A^{pq}=0$ for $p<0$ and $p>1$, this boils down to a commutative diagram with exact rows $$\begin{CD}
0 @>>> W_0 @>>> H^1(X_0) @>>> {{\operatorname{Gr}}}_1^W @>>> 0\\
@. @| @| @|\\
0 @>>> E_{\infty}^{10} @>{i}>> {{\Bbb H}}^1 @>{\pi}>> E_{\infty}^{01}
@>>> 0.
\end{CD}$$ The map $i$ is induced by the inclusion $A^{10}\to A^{10}\oplus
A^{01}$; the map $\pi$ is induced by the projection $A^{10}\oplus
A^{01}\to A^{01}$.
The Hodge filtration on ${{\Bbb H}}^k$ is induced by $F^p(A^k)=\oplus_{a+b=k} F^p(A^{a,b})$. In our case we have $$F^1{{\Bbb H}}^1={{\operatorname{Ker}}}(D|_{F^1(Y^0)})={{\operatorname{Ker}}}(\delta: F^1(Y^0)\to F^1(Y^1)).$$
\[wf\] Let $X_0$ be a cycle of ruled surfaces over a smooth curve $C$. The weight filtration on $H^1(X_0)$ takes the form $$0 \subset W_0 \subset W_1=H^1(X_0).$$ Furthermore, $W_0$ is the unique 1-dimensional Hodge structure ${T\langle 0 \rangle}$ of type $(0,0)$ and ${{\operatorname{Gr}}}_1^W$ is canonically isomorphic to $H^1(C)$ (as a Hodge structure).
Since $E_1^{pq}=H^q(Y^p)$ we have $W_0=E_{\infty}^{10}=E_2^{10}={{\operatorname{Coker}}}(\delta: H^0(Y^0)\to H^0(Y^1))$. Now it is easy to see that $$\begin{CD}
H^0(Y^0) @>{\delta}>> H^0(Y^1) @>>> {{\Bbb Z}}@>>> 0\\
@. @| @AA{S}A\\
@. \oplus H^0(X_i) @= \oplus {{\Bbb Z}},
\end{CD}$$ where $S$ is the summation map, is commutative and has an exact top row, hence $W_0$ is indeed the trivial Hodge structure ${T\langle 0 \rangle}$.
To show that ${{\operatorname{Gr}}}_1^W=E_{\infty}^{01}=E_2^{01}={{\operatorname{Ker}}}(\delta: H^1(Y^0)\to H^1(Y^1))$ is isomorphic to $H^1(C)$, we introduce the following maps: Let $$\rho_i: Y_i\to C_i\text{ and }\sigma_i: Y_{i+1}\to C_i$$ be the projections and let $\tau_i=\sigma_0 {{\scriptstyle\circ}}(\alpha_1{{\scriptstyle\circ}}\sigma_1)
{{\scriptstyle\circ}}(\alpha_2{{\scriptstyle\circ}}\sigma_2){{\scriptstyle\circ}}\cdots{{\scriptstyle\circ}}(\alpha_{i-1}{{\scriptstyle\circ}}\sigma_{i-1}):
Y_i\to C_0=C$. Then one easily shows that the following sequence is exact $$\begin{CD}
0 @>>> H^1(C) @>{\tau^{\ast}}>> H^1(Y^0) @>{\delta}>> H^1(Y^1),
\end{CD}$$ where $\tau^{\ast}:=(\tau_1^\ast,\ldots,\tau_N^\ast): H^1(C) \to
H^1(Y^0)=H^1(Y_1)\oplus\cdots\oplus H^1(Y_N)$.
The base curve and the extension class {#subsection22}
--------------------------------------
The mixed Hodge structure $H^1(X_0)$ can be regarded as an extension of the pure Hodge structures $W_0$ and $\operatorname{Gr}^W_1$. We want to show how one can recover the base curve $C$ and the shift $s$ from this extension. In [@C] an object $\operatorname{Ext}(\operatorname{Gr}^W_1,W_0)$ was introduced which parametrizes all such extensions. In our case it is an abelian variety, namely the base curve $C$.
\[lemma14\] Let $W_{\bullet}$ be as above. Then there exists a canonical isomorphism between ${{\operatorname{Ext}}}({{\operatorname{Gr}}}_1^W,W_0)$ and ${{\operatorname{Alb}}}(C)$.
For any Hodge structure $H$ we define $J^0 H =
H_{{{\Bbb C}}}/H_{{{\Bbb Z}}}+ F^0 H$. If $H_1$ and $H_2$ are two Hodge structures, then we can apply this to $H={{\operatorname{Hom}}}(H_1,H_2)$ and we get ${{\operatorname{Ext}}}(H_1,H_2)=J^0 {{\operatorname{Hom}}}(H_1,H_2)$ (see [@C prop. 2]). Now if $H_1=H^1(X)$, where $X$ is a smooth projective variety and $H_2={T\langle 0 \rangle}$, then ${{\operatorname{Hom}}}(H_1,H_2)$ is the dual Hodge structure of $H^1(X)$ and has weight $-1$. Its Hodge filtration has the form $$0=F^1\subset F^0\subset F^{-1}=H^1(X)^\ast,$$ where $F^0(H^1(X)^\ast)=(H^{0,1}(X))^\ast$. Hence $H^1(X)^{\ast}/F^0(H^1(X)^{\ast})=H^{1,0}(X)^{\ast}$ and $J^0(H^1(X)^{\ast})=H^{1,0}(X)^{\ast}/H_1(X,{{\Bbb Z}})={{\operatorname{Alb}}}(X)$.
In [@C] we find the following algorithm to calculate the extension class $e\in J^0{{\operatorname{Hom}}}(B,A)$ of an extension $$\begin{CD}
0 @>>> A @>{i}>> H @>{\pi}>> B @>>> 0.
\end{CD}$$
1. Choose an integral retraction $r:H \to A$, i.e. a map defined over ${{\Bbb Z}}$ satisfying $r{{\scriptstyle\circ}}i ={{\bf 1}}_A$.
2. Define two Hodge filtrations on $A\oplus B$: $$\tilde{F}_{\infty}^{\bullet}:=(r,\pi)(F^{\bullet}H)\text{ and }
\tilde{F}_0^{\bullet}:=F^{\bullet}A \oplus F^{\bullet}B.$$
3. Find a $\psi\in{{\operatorname{Hom}}}(B,A)_{{{\Bbb C}}}$ such that $${\pmatrix {{\bf 1}}_{A} & \psi\\ 0 & {{\bf 1}}_{B}
\endpmatrix} \tilde{F}_0^{\bullet} =\tilde{F}_{\infty}^{\bullet}.$$
Then $e=[\psi]\in J^0{{\operatorname{Hom}}}(B,A)$.
Let $X_0$ be a cycle of ruled surfaces over the elliptic curve $C=V/\Lambda$ with shift $s\in {{\operatorname{Alb}}}(C)$. We now describe the ${\Bbb
Z}$-splitting of the exact sequence $0\to W_0\to {{\Bbb H}}^1\to{{\operatorname{Gr}}}_1^W\to 0$, as required in the first step of this algorithm, explicitely in terms of $C$ and $s$. Let $\{\lambda_1,\lambda_2\}$ be a ${{\Bbb Z}}$-basis for $\Lambda$ and let $x_i:V\to{{\Bbb R}}$ be the corresponding coordinate functions ($i=1$, 2). Let $s_i=x_i(s)$, i.e. $$s=s_1\lambda_1+s_2\lambda_2.$$ (The numbers $s_i$ are defined only up to integers.) Let $\phi\in
H^1(C)$. Since $\{dx_1,dx_2\}$ is a basis for $H^1(C)$, we may represent $\phi$ by $\xi_1 dx_1 + \xi_2 dx_2\in A^1(C)$, which we will also denote by $\phi$. Consider $$\tau^{\ast}\phi:=(\tau_1^{\ast},\ldots,\tau_N^{\ast})\phi\in
E_0^{01}.$$ Recall that $\tau^{\ast}\phi$ represents an element in $E_{\infty}^{01}$. Set $$f=(f_1,\ldots,f_N)\in E_0^{10},$$ where $$f_i=\begin{cases} 0 & 1\le i<N\\
- \xi_1 s_1 - \xi_2 s_2 & i=N
\end{cases}$$ is a constant function on $C_i$.
The pair $(\tau^{\ast}\phi,f)\in A^{01}\oplus A^{10}$ represents a class in ${{\Bbb H}}^1$.
First of all, $d(\tau^{\ast}\phi)=0$ since $d\phi=0$. Furthermore, $\delta(\tau^{\ast}\phi)=(0,\ldots,0,(t_{-s}-{{\bf 1}}_C)^{\ast}\phi)$, where $t_{-s}:x\mapsto x-s: C\to C$. Let $g=\xi_1 x_1+\xi_2
x_2\in{{\operatorname{Hom}}}_{{{\Bbb R}}}(V,{{\Bbb R}})$. Then $\phi=dg$ and $f_N=(t_{-s}-{{\bf 1}}_C)^{\ast}g$.
We write $(\tau^{\ast}\phi,f)=\sigma(\phi)$. By the lemma above, the map $\sigma: H^1(C)\to E_0^{01}\oplus E_0^{10}$ induces a map $$\bar{\sigma}: {{\operatorname{Gr}}}_1^W=H^1(C)\to {{\Bbb H}}^1.$$ Since $\pi:{{\Bbb H}}^1\to {{\operatorname{Gr}}}_1^W$ is induced by the projection $A^{01}\oplus A^{10}\to A^{01}$, $\pi{{\scriptstyle\circ}}\bar{\sigma}={{\bf 1}}_{{{\operatorname{Gr}}}_1^W}$, i.e. $\bar{\sigma}$ is a splitting over ${{\Bbb R}}$. In fact, $\bar{\sigma}$ is defined over ${{\Bbb Z}}$:
\[fl\] The map $\bar{\sigma}$ splits the sequence $$\begin{CD}
0 @>>> W_0 @>{i}>> {{\Bbb H}}^1 @>{\pi}>> {{\operatorname{Gr}}}_1^W @>>> 0
\end{CD}$$ over ${{\Bbb Z}}$.
Assuming this for the moment, we show that the extension class is identified with the shift.
\[cor17\] The natural isomorphism ${{\operatorname{Ext}}}({{\operatorname{Gr}}}_1^W,W_0)\cong{{\operatorname{Alb}}}(C)$ identifies the extension class $e=[\psi]$ with the shift $s$.
Let $r:{{\Bbb H}}^1\to W_0$ be the map such that $$(r,\pi)=(i,\bar{\sigma})^{-1}:{{\Bbb H}}^1 \to W_0 \oplus {{\operatorname{Gr}}}_1^W.$$ As before we set $$\tilde{F}_{\infty}^1=(r,\pi)F^1{{\Bbb H}}^1\text{ and }
\tilde{F}_0^1=F^1(W_0)\oplus F^1 ({{\operatorname{Gr}}}_1^W)$$ in $W_0 \oplus {{\operatorname{Gr}}}_1^W$. The extension class is represented by a map $\psi: {{\operatorname{Gr}}}_1^W\to W_0$ such that $$\begin{pmatrix} {{\bf 1}}_{W_0} &- \psi\\
0 & {{\bf 1}}_{{{\operatorname{Gr}}}_1^W} \end{pmatrix}
\tilde{F}_{\infty}^1=\tilde{F}_0^1.$$ Since $\tilde{F}_0^1\cap W_0=F^1(W_0)=0$ we get $$r(\omega)-\psi(\pi(\omega))=0\quad\text{for all } \omega\in
F^1{{\Bbb H}}^1.$$
Now assume that $C={{\Bbb C}}/({{\Bbb Z}}\tau+{{\Bbb Z}})$ and let $\phi=\tau dx_1+dx_2$, where $x_1$, $x_2$ are the coordinates dual to the basis $\lambda_1=\tau$, $\lambda_2=1$. Then $F^1H^1(C)={{\Bbb C}}\phi$ (see §4.2). Recall that $(\tau^{\ast}\phi,0)\in
A^{01}\oplus A^{10}$ represents an element $\omega\in F^1{{\Bbb H}}^1$ such that $\pi(\omega)=\phi$. Consider $\sigma(\phi)=(\tau^{\ast}\phi,f)$. Since $i:W_0\to {{\Bbb H}}^1$ is induced by the inclusion $A^{10}\to
A^{01}\oplus A^{10}$, $$\omega=i(-f)+\sigma(\phi),$$ hence $$\begin{aligned}
\psi(\tau dx_1 + dx_2) &=\psi(\pi(\omega))\\
& = r(\omega)\\
&=-f\end{aligned}$$ in $W_0$. Identifying $W_0$ with ${{\Bbb Z}}$ as in the proof of lemma \[wf\] we find $$\begin{aligned}
\psi(\tau dx_1+dx_2) &=- \sum f_i\\
&=-f_N\\
&=\tau s_1 + s_2\\
&=\int_{0}^s \tau dx_1+dx_2.\end{aligned}$$ Hence $\psi=\int_0^s$ on $F^1H^1(C)$, i.e. ${{\operatorname{Ext}}}({{\operatorname{Gr}}}_1^W,W_0)\cong{{\operatorname{Alb}}}(C)$ identifies the extension class $e=[\psi]$ with the shift $s$.
\[R\]The above result is also true for $N=1$. In this case let $Y_0$ be the normalization of $X_0$. Let $X_0'$ be the cycle of elliptic ruled surfaces consisting of two copies of $Y_0$ and glued with the same shift $s$ as $X_0$. There exists a projection map $X_0'\to X_0$ given by contracting one of the components of $X_0'$. By functoriality, the isomorphism between the vector spaces $H^1(X_0)$ and $H^1(X_0')$ induced by this map is in fact an isomorphism of mixed Hodge structures. Hence we can work with $X_0'$ instead of $X_0$ .
Instead of cycles of elliptic ruled surfaces we can also consider cycles of ${{\Bbb P}}^1$-bundles over an abelian variety $Z$. Again the Hodge filtration takes on the form $0=W_{-1}\subset W_0\subset W_1=H^1(X_0)$ where $W_0$ is the 1-dimensional Hodge structure of type $(0,0)$. As before $\operatorname{Gr}^W_1=H^1(Z)$ and hence $\operatorname{Ext}(\operatorname{Gr}_1^W,W_0)=
\operatorname{Alb}(Z)$. The same proof as before also shows that the extension class in $\operatorname{Ext}(\operatorname{Gr}_1^W,W_0)$ again coincides with the shift $s$.
Proof of lemma \[fl\]
---------------------
Let $c:[0,1]\to X_0$ be the loop described in figure 2 [@Gre 23.8].
\[f2\]
(12,5) (1,1)[(1,0)[5]{}]{} (6.4,1)[(1,0)[0.2]{}]{} (6.9,1)[(1,0)[0.2]{}]{} (7.4,1)[(1,0)[0.2]{}]{} (8,1)[(1,0)[3]{}]{} (1,5)[(1,0)[5]{}]{} (6.4,5)[(1,0)[0.2]{}]{} (6.9,5)[(1,0)[0.2]{}]{} (7.4,5)[(1,0)[0.2]{}]{} (8,5)[(1,0)[3]{}]{} (1,1)[(0,1)[4]{}]{} (3,1)[(0,1)[4]{}]{} (5,1)[(0,1)[4]{}]{} (9,1)[(0,1)[4]{}]{} (11,1)[(0,1)[4]{}]{} (1,2)[(1,0)[5]{}]{} (6.4,2)[(1,0)[0.2]{}]{} (6.9,2)[(1,0)[0.2]{}]{} (7.4,2)[(1,0)[0.2]{}]{} (8,2)[(1,0)[1]{}]{} (9,2)[(2,1)[2.0]{}]{} (9.2,2)[(1,0)[0.1]{}]{} (9.6,2)[(1,0)[0.1]{}]{} (10.0,2)[(1,0)[0.1]{}]{} (10.4,2)[(1,0)[0.1]{}]{} (10.8,2)[(1,0)[0.1]{}]{} (1,2) (3,2) (5,2) (9,2) (11,2) (11,3) (0,2.2)[(0.8,0.8)\[rb\][$p_0$]{}]{} (2,2.2)[(0.8,0.8)\[rb\][$p_1$]{}]{} (4,2.2)[(0.8,0.8)\[rb\][$p_2$]{}]{} (8,2.2)[(0.8,0.8)\[rb\][$p_{N-1}$]{}]{} (11.2,3.2)[(0.8,0.8)\[lb\][$p_N=p_0$]{}]{} (11.2,2.2)[(0.8,0.8)\[lb\][$p_N+s$]{}]{} (1,0)[(2,0.8)\[t\][$Y_1$]{}]{} (3,0)[(2,0.8)\[t\][$Y_2$]{}]{} (9,0)[(2,0.8)\[t\][$Y_N$]{}]{}
Clearly, $H_1(X_0,{{\Bbb Z}})=H_1(C,{{\Bbb Z}})\oplus {{\Bbb Z}}c$. Using the Kronecker product (see [@Gre]) the loop $c$ determines a map $$c: H^1(X_0,{{\Bbb Z}})\to{{\Bbb Z}}.$$ Since $\bar{\sigma}$ splits the sequence over ${{\Bbb R}}$, all we have to do is show that $c(\bar{\sigma}(dx_i))\in{{\Bbb Z}}$ for $i=1$, 2. In fact, we will prove that $c{{\scriptstyle\circ}}\bar{\sigma}=0$.
In order to understand how the loop $c$ operates on ${{\Bbb H}}^1$, we want to identify ${{\Bbb H}}^1=H^1(\Gamma(X_0,{{\cal A}}^{\bullet}))$ with the simplicial cohomology $H_{\nabla}^1(X_0)$. In [@GM p.95] we find a canonical isomorphism $$H^k(\Gamma(X_0,{{\cal B}}^{\bullet}))\to H_{\nabla}^k(X_0),$$ induced by integration, where $${{\cal B}}^k:={{\operatorname{Ker}}}(\delta: {{\cal A}}^{0k}\to{{\cal A}}^{1k}).$$ One easily checks that the complex $({{\cal B}}^{\bullet},d)$ is also an acyclic resolution of the constant sheaf ${{\Bbb C}}_{X_0}$. The inclusions $j_k:{{\cal B}}^k{\hookrightarrow}{{\cal A}}^{0k}{\hookrightarrow}{{\cal A}}^k$ commute with the coboundary operators, hence the canonical isomorphism between $H^k(\Gamma(X_0,{{\cal B}}^{\bullet}))$ and $H^k(\Gamma(X_0,{{\cal A}}^{\bullet}))$ is induced by $j_k$ (see [@W 5.24]).
Let $\phi=\xi_1 dx_1+ \xi_2 dx_2$ and $(\omega,f)=\bar{\sigma}(\phi)$, i.e. $$\begin{aligned}
\omega
&=(\omega_1,\ldots,\omega_N)\qquad\omega_i=\tau_i^{\ast}\phi\\
f&=(0,\ldots,0,f_N)\qquad f_N=-\xi_1 s_1- \xi_2 s_2\end{aligned}$$ Let $g=(g_1,\ldots,g_N)\in\Gamma(X_0,{{\cal A}}^0)$ be such that $$g_i|_{C_j}=\begin{cases} -f_N &(i,j)=(N,0)\\
0 &\text{otherwise}
\end{cases}$$ Since $\delta g= -f$, we have that $\omega'=\omega+dg\in \Gamma(X_0,{{\cal B}}^1)$. Since $(\omega',0)=(\omega,f)+(df,\delta f)$ we have $j_1[\omega']=\bar{\sigma}(\phi)$ in ${{\Bbb H}}^1$.
Now we view the loop $c$ as an element of ${{\operatorname{Hom}}}_{{{\Bbb Z}}}(H_{\nabla}^1(X_0,{{\Bbb Z}}),{{\Bbb Z}})$. Using the defining properties of $g_i$ and the fact that integrals of $\omega_i$ along the rulings vanish, we get $$\begin{aligned}
c(\bar{\sigma}(\phi)) &=\sum_{i=1}^N \int_{p_{i-1}}^{p_i}
\omega_i'\\
&=\int_{p_{N-1}}^{p_N} \omega_N'\\
&=g_N(p_N)-g_N(p_{N-1})+\int_{p_N+s}^{p_N}
\omega_N\\
&=(s_1\xi_1 +s_2 \xi_2) - (s_1\xi_1 +s_2
\xi_2)\\
&=0.\end{aligned}$$ $\Box$
Variation of Hodge structure {#section4}
============================
In this section we want to compute the VHS associated to boundary points of the moduli space of $(1,p)$-polarized abelian surfaces with a (canonical) level structure (cf. definition \[D\] and theorem \[T\]).
$D$-polarized abelian varieties
-------------------------------
Our main reference is [@LB]. Fix a type $D$, i.e. an ordered sequence $(d_1,\ldots,d_g)$ of positive integers satisfying $d_i|d_{i+1}$ ($i=1,\ldots,g-1$). We will often write $D={{\operatorname{diag}}}(d_1,\ldots,d_g)$. Let ${{\Bbb H}}_g$ be the [*Siegel space of degree*]{} $g$. To $\tau\in{{\Bbb H}}_g$ we associate an abelian variety $X_{\tau,D}$ of dimension $g$ and a polarization $E_{\tau,D}$ of type $D$ as follows. First, $$X_{\tau,D}={{\Bbb C}}^g/\Lambda_{\tau,D},$$ where $$\Lambda_{\tau,D}=(\tau, D){{\Bbb Z}}^{2g}.$$ Let $\lambda_i$ ($i=1,\ldots,2g$) be the $i$-th column of $(\tau,D)$. Then we define $E_{\tau,D}$ to be the map $\Lambda\times\Lambda\to{{\Bbb Z}}$ given by the matrix $$E_{\tau,D}=\begin{pmatrix} 0 & D\\ -D & 0 \end{pmatrix}$$ with respect to the basis $\{\lambda_1,\ldots,\lambda_{2g}\}$ of $\Lambda_{\tau,D}$. For this reason we will refer to this basis as the symplectic basis. As an alternating map $E$ can be regarded as an element in $H^2(X_{\tau,D},{{\Bbb Z}})$. It is standard knowledge that (up to sign) it is in fact the first Chern class of an ample line bundle on $X_{\tau,D}$, i.e. a polarization.
Every $D$-polarized abelian variety is isomorphic to $(X_{\tau,D},E_{\tau,D})$ for some $\tau\in{{\Bbb H}}_g$. Furthermore, we can construct a universal $D$-polarized abelian varietyover ${{\Bbb H}}_g$ as follows: $${{X}}_{D}={{\Bbb Z}}^{2g}\backslash ({{\Bbb C}}^g\times{{\Bbb H}}_g),$$ where ${{\Bbb Z}}^{2g}$ operates on ${{\Bbb C}}^g\times{{\Bbb H}}_g$ by $$l(v,\tau)=(v+(\tau,D)l,\tau)$$ for $l\in{{\Bbb Z}}^{2g}$ and $(v,\tau)\in{{\Bbb C}}^g\times{{\Bbb H}}_g$. The projection ${{\Bbb C}}^g\times{{\Bbb H}}_g\to{{\Bbb H}}_g$ induces a projection $$\pi:{{X}}_D\to{{\Bbb H}}_g,$$ such that $\pi^{-1}(\tau)=X_{\tau,D}$. $X_{D}$ is a complex manifold and carries a universal polarization ${{\cal L}}$ (see [@LB lemma 8.7.1]).
We are only interested in $D=(1,p)$, $D=(1)$ and $D=(p)$ and will often omit $D$ if confusion seems unlikely.
The polarized Hodge structure of $(X_{\tau,D},E_{\tau,D})$
----------------------------------------------------------
We can describe the Hodge structure of the abelian variety $X_{\tau,D}$ very explicitly. For any abelian variety $X=V/\Lambda$ we have $$\matrix
F^1(H^1(X))=H^{1,0}(X) & \subset & H^1(X,{{\Bbb C}})\\
\| && \| \\
V^{\ast}={{\operatorname{Hom}}}_{{{\Bbb C}}}(V,{{\Bbb C}}) & \subset & {{\operatorname{Hom}}}_{{{\Bbb R}}}(V,{{\Bbb C}}).
\endmatrix$$ Now consider $X=X_{\tau,D}$, let $\{e_1,\ldots,e_g\}$ be the standard basis of $V={{\Bbb C}}^g$ and let $\{\lambda_1,\ldots,\lambda_{2g}\}$ be the symplectic basis of $\Lambda=\Lambda_{\tau,D}$. Then $\{\lambda_1^{\ast},\ldots,\lambda_{2g}^{\ast}\}$ is a ${{\Bbb Z}}$-basis of $H^1(X,{{\Bbb Z}})$ and $\{e_1^{\ast},\ldots,e_g^{\ast}\}$ is a ${{\Bbb C}}$-basis of $F^1(H^1(X))$. Using the coordinates of $H^1(X,{{\Bbb C}})=H^1(X,{{\Bbb Z}})\otimes{{\Bbb C}}$ determined by $\{\lambda_1^{\ast},\ldots,\lambda_{2g}^{\ast}\}$, $e_i^{\ast}$ is given by the $i$-th row of the period matrix $(\tau,D)$.
A polarization $\omega$ on a compact complex manifold $X$ induces a bilinear form $Q$ on $H^n(X)$: $$Q(\phi,\psi)=(-1)^{n(n-1)/2} \int_{X}
\phi\wedge\psi\wedge\omega^{d-n},$$ where $d=\dim X$. The pair $(H^n(X), Q)$ is a polarized Hodge structure of weight $n$ (see [@G p.7]).
We now calculate the polarization on $H^1(X_{\tau,D})$ induced by $E_{\tau,D}$. Let $x_i:V\to {{\Bbb R}}$ ($i=1,\ldots,2g$) be the coordinates with respect to the real basis $\{\lambda_1,\ldots,\lambda_{2g}\}$ of $V$. Then $dx_i$ corresponds to $\lambda_i^{\ast}$ under the natural isomorphism $H_{\hbox{de Rham}}^1(X) \cong {{\operatorname{Hom}}}(\Lambda,{{\Bbb Z}})$. Furthermore $E_{\tau,D}\in H^2(X,{{\Bbb Z}})$ is represented by the 2-form $$\omega=-\sum_{i=1}^g d_i dx_i\wedge dx_{i+g}$$ and $$\int_X dx_1\wedge dx_{g+1} \wedge\cdots\wedge dx_{g}\wedge
dx_{2g} =1$$ (see [@LB lemmas 3.6.4 and 3.6.5]). It follows that $Q$ is given by the matrix $$(g-1)!
\left(\begin{array}{cc} 0 & -\hat{D}\\ \hat{D} &
0\end{array}\right)
\qquad \hat{D}=\Big(\prod_{i=1}^{2g} d_i\Big) D^{-1}$$ with respect to the basis $\{\lambda_1^{\ast},\cdots,\lambda_{2g}^{\ast}\}$ of $H^1(X,{{\Bbb Z}})={{\operatorname{Hom}}}(\Lambda,{{\Bbb Z}})$.
Boundary points of ${\cal A}^*(1,p)$
------------------------------------
We have to recall briefly some facts about compactifications of moduli spaces of abelian surfaces. All relevant details can be found in [@HKW]. By $\cal A(1,p)$ we denote the moduli space of $(1,p)$-polarized abelian surfaces with a (canonical) level structure ($p\geq3$, prime). Recall that $$\cal A(1,p)={{\Bbb H}}_2/\Gamma_{1,p}$$ where $$\Gamma_{1,p}=\left\{g\in\operatorname{Sp}(4,{{\Bbb Z}})\ ;\
g-\mbox{\boldmath$1$}\in
\begin{pmatrix}
{{\Bbb Z}}&{{\Bbb Z}}&{{\Bbb Z}}&p{{\Bbb Z}}\\ p{{\Bbb Z}}&p{{\Bbb Z}}&p{{\Bbb Z}}&p^2{{\Bbb Z}}\\
{{\Bbb Z}}&{{\Bbb Z}}&{{\Bbb Z}}&p{{\Bbb Z}}\\
{{\Bbb Z}}&{{\Bbb Z}}&{{\Bbb Z}}&p{{\Bbb Z}}\end{pmatrix}\right\}$$ acts on Siegel space ${{\Bbb H}}_2$ by $$g=\begin{pmatrix} A&B\\C&D\end{pmatrix}
:\tau\longmapsto(A\tau+B)(C\tau+D)^{-1}.$$ In [@HKW] a torodial compactification ${\cal A}^*(1,p)$ of $\cal A(1,p)$ was constructed. To compactify $\cal A(1,p)$ one has to add (non-compact) boundary surfaces (corank 1 boundary points) and boundary curves (corank 2 boundary points). Here we shall restrict ourselves exclusively to the boundary surfaces. These surfaces are indexed by the vertices of the Tits building of $\Gamma_{1,p}$ which correspond to lines $l\subset{{\Bbb Q}}^4$. According to [@HKW] there are two types of boundary surfaces, namely one [*central boundary surface*]{} $D(l_0)$ and $p(p-1)/2$ [*peripheral boundary surfaces*]{} $D(l_{(a,b)})$ where $(a,b)\in({{\Bbb Z}}_p\times{{\Bbb Z}}_p\setminus\{0\})/(\pm1)$. The group $\operatorname{SL}(2,{{\Bbb Z}}_p)$ acts on ${\cal A}^*(1,p)$ and permutes the peripheral boundary surfaces of ${\cal A}^*(1,p)$ transitively. Therefore it is enough to consider one of them, namely $D(l_{(0,1)})$.
From now on let $l=l_0$ or $l_{(0,1)}$. The stabilizer $P(l)$ of $l$ in $\Gamma_{1,p}$ is an extension of the form $$1\longrightarrow P'(l)\longrightarrow P(l)\longrightarrow
P''(l)\longrightarrow 1$$ where $P'(l)$ is a rank 1 lattice. The compactification procedure requires that one first takes the partial quotient of ${{\Bbb H}}_2$ with respect to $P'(l)$. For $l=l_0$ $$P'(l)=\left\{\left(
\begin{array}{c|c}
\begin{array}{ccc} && \\ &\mbox{\boldmath$1$}&\\ && \end{array} &
\begin{array}{ccc} n&&0\\ && \\ 0&&0\end{array} \\ \hline
\begin{array}{ccc} && \\ &0& \\ &&\end{array} &
\begin{array}{ccc} && \\ &\mbox{\boldmath$1$}& \\ &&\end{array}
\end{array}\right)\ ;\ n\in{{\Bbb Z}}\right\}$$ and for $l=l_{(0,1)}$: $$P'(l)=\left\{\left(
\begin{array}{c|c}
\begin{array}{ccc} && \\ &\mbox{\boldmath$1$}&\\ && \end{array} &
\begin{array}{ccc} 0&&0\\ && \\ 0&&np^2\end{array} \\ \hline
\begin{array}{ccc} && \\ &0& \\ &&\end{array} &
\begin{array}{ccc} && \\ &\mbox{\boldmath$1$}& \\ &&\end{array}
\end{array}\right)\ ;\ n\in{{\Bbb Z}}\right\}$$ The partial quotient map $e(l):{{\Bbb H}}_2\to{{\Bbb H}}_2/P'(l)$ is then given by $$e(l_0):{{\Bbb H}}_2\longrightarrow{{\Bbb C}}^*\times{{\Bbb C}}\times{{\Bbb H}}_1,\
\begin{pmatrix}\tau_1&\tau_2\\ \tau_2&\tau_3
\end{pmatrix}\longmapsto(e^{2\pi i\tau_1},\tau_2,\tau_3)$$ resp. $$e(l_{(0,1)}):{{\Bbb H}}_2\longrightarrow{{\Bbb H}}_1\times{{\Bbb C}}\times{{\Bbb C}}^*,\
\begin{pmatrix}\tau_1&\tau_2\\ \tau_2&\tau_3
\end{pmatrix}\longmapsto(\tau_1,\tau_2,e^{2\pi i\tau_3/p^2}).$$ Here ${{\Bbb H}}_1$ denotes the usual upper half plane. This maps ${{\Bbb H}}_2$ to an interior neighbourhood of $\{0\}\times{{\Bbb C}}\times{{\Bbb H}}_1$, resp. ${{\Bbb H}}_1\times{{\Bbb C}}\times\{0\}$, i.e. the interior of the closure of the image, $X(l)=(\overline{e({{\Bbb H}}_2)})^o$ is an open neighbourhood of $\{0\}\times{{\Bbb C}}\times{{\Bbb H}}_1$, resp. ${{\Bbb H}}_1\times{{\Bbb C}}\times\{0\}$. The partial compactification in the direction of $l$ then consists of adding the set $\{0\}\times{{\Bbb C}}\times{{\Bbb H}}_1$, resp. ${{\Bbb H}}_1\times{{\Bbb C}}\times\{0\}$. There is a natural map $X(l)\to{\cal A}^*(1,p)$ given by dividing out the extended action of $P''(l_0)$ on $X(l)$, which maps $\{0\}\times{{\Bbb C}}\times{{\Bbb H}}_1$, resp. ${{\Bbb H}}_1\times{{\Bbb C}}\times\{0\}$ to the boundary surface $D(l_0)$, resp. $D(l_{(0,1)})$.
Two 1-parameter families {#subsection34}
------------------------
We consider two 1-parameter families of $(1,p)$-polarized abelian surfaces which are closely related to the central, resp. peripheral boundary components (see proof of theorem \[T\]). For $M$ a positive integer we set $${{\Bbb H}}_1(M):=\{\tau\in{{\Bbb H}}_1\ ;\ \operatorname{Im}\tau>M\}$$ and $$\Delta^*(M):=\{t\in{{\Bbb C}}\ ;\ 0<|t|<e^{-2\pi iM}\}.$$ First we fix a pair $(\tau_2,\tau_3)\in{{\Bbb C}}\times{{\Bbb H}}_1$. For sufficiently large $M$ we have a map $${{\Bbb H}}_1(M)\longrightarrow{{\Bbb H}}_2,\ \tau\longmapsto\begin{pmatrix}
\tau&\tau_2\\\tau_2&\tau_3\end{pmatrix}.$$ The first 1-parameter family which we want to consider is the pull back of the universal family $\pi:X_D\to{{\Bbb H}}_2$ to a family $\pi_1:X_D\to{{\Bbb H}}_1(M)$ via this map. For the second family we fix a pair $(\tau_1,\tau_2)\in{{\Bbb H}}_1\times{{\Bbb C}}$ and pull back the universal family $\pi:X_D\to{{\Bbb H}}_2$ to a family $\pi_2:X_D\to{{\Bbb H}}_1(M)$ via the map $${{\Bbb H}}_1(M)\longrightarrow{{\Bbb H}}_2,\
\tau\longmapsto\begin{pmatrix}\tau_1&\tau_2\\
\tau_2&p^2\tau\end{pmatrix}$$ where $M$ is again chosen sufficiently large.
We denote the polarized abelian surface associated to $\tau\in{{\Bbb H}}_1(M)$ by $(X_{\tau},E_{\tau})$. Since in both cases $(X_{\tau},E_{\tau})$ depends only on $\tau\mod{{\Bbb Z}}$, the family $\pi_i$ is the pull back of a family $p_i: {{X}}_D\to\Delta^{\ast}(M)$ via the map $${{\Bbb H}}_1(M)\longrightarrow\Delta^*(M),\ \tau\longmapsto t=e^{2\pi i\tau}.$$
Consider the maps $$\tilde{\phi}_1:{{\Bbb H}}_1(M)\to{{\operatorname{Grass}}}(2,4),\
\tau\mapsto{\bmatrix {\tau} & \tau_2 & 1 & 0\\
\tau_2 & {\tau_3} & 0 & p \endbmatrix},$$ and $$\tilde{\phi}_2:
{{\Bbb H}}_1(M)\to{{\operatorname{Grass}}}(2,4),\ \tau\mapsto{\bmatrix {\tau_1} & \tau_2 & 1 & 0\\
\tau_2 & {p^2\tau} & 0 & p \endbmatrix}.$$ There image is contained in Griffiths’ period domain ${{\cal D}}_2\subset{{\operatorname{Grass}}}(2,4)$. The monodromy $T_i$ is defined by $\sideset{^t}{}{\tilde{\phi}}_i(\tau+1)=T_i
\sideset{^t}{}{\tilde{\phi}}_i(\tau)$. One easily checks that $T_i=\left(\smallmatrix {{\bf 1}}_2 & \eta_i\\ 0 &
{{\bf 1}}_2\endsmallmatrix\right)$, where $\eta_1=\left( \smallmatrix 1 &
0\\ 0 & 0 \endsmallmatrix \right)$ and $\eta_2=\left( \smallmatrix 0
& 0\\ 0 & p \endsmallmatrix \right)$ By construction $\tilde{\phi}_i$ descends to a map $\phi_i: \Delta^\ast(M) \to \langle T_i \rangle \backslash {{\cal D}}_2$: $$\begin{CD}
{{\Bbb H}}_1 @>{\tilde{\phi_i}}>> {{\cal D}}_2\\
@V{t}VV @VVV\\
\Delta^\ast(M) @>>{\phi_i}>\langle T_i \rangle \backslash {{\cal D}}_2
\end{CD}$$ If we identify $H^1(X,{{\Bbb C}})$ with ${{\Bbb C}}^4$ using the coordinates induced by the basis $\{\lambda_1^{\ast},\ldots,\lambda_4^{\ast}\}$ of $H^1(X,{{\Bbb Z}})$), then by the description of the Hodge structure on the first cohomology of an abelian variety given in the previous section, $\phi_i$ is the VHS associated to the family $p_i:{{X}}_D\to\Delta^{\ast}(M)$.
The limit mixed Hodge structure {#subsection35}
-------------------------------
The family $p_i:{{X}}_D\to\Delta^{\ast}(M)$ ($i=1$,2) induces a MHS on $H^1(X_t)$ ($t\in\Delta^{\ast}(M)$ arbitrary but fixed; see [@G Chapter IV]). Its weight filtration is of the form $0=W_{-1}\subset W_0 \subset W_1 \subset W_2=H^1(X_t)$. It is is given by $$W_0 = {{\operatorname{Im}}}N\text{ and }
W_1 = {{\operatorname{Ker}}}N,$$ where $N=\log T=T-\mbox{\boldmath$1$}=\left(\smallmatrix 0 &
\eta\\ 0 & 0\endsmallmatrix\right)$. (We have written $T$ instead of $T_i$ etc.) The limit Hodge filtration $F_{\infty}^{\bullet}\subset
H^1(X_t)$ is calculated as follows: first define $\tilde{\psi}:
{{\Bbb H}}_1(M) \to \check{{{\cal D}}}_2\subset{{\operatorname{Grass}}}(2,4)$ (where $\check{{{\cal D}}}_2$ is the compact dual of ${{\cal D}}_2$) by $$\sideset{^t}{}{\tilde{\psi}}(\tau)
=T^{-\tau}\sideset{^t}{}{\tilde{\phi}}(\tau)=
(\mbox{\boldmath$1$}-\tau N)
\sideset{^t}{}{\tilde{\phi}}(\tau).$$ By construction $\tilde{\psi}(\tau+1)=\tilde{\psi}(\tau)$, hence $\tilde{\psi}$ descends to a map $\psi: \Delta^\ast(M)\to
\check{{{\cal D}}}_2.$ By the work of Griffiths and Schmid $\psi$ extends to $\Delta$ and $F_{\infty}^{\bullet}:=\psi(0)$ together with the monodromy weight filtration defines a MHS on $H^1(X_t)$.
From now on we will write $e_i$ instead of $\lambda_i^{\ast}$ for the basis vectors of $H^1(X,{{\Bbb Z}})$. In our case, an immediate calculation shows that $\tilde{\psi}$ is constant and $$W_0=[e_1],\quad
W_1=[e_1, e_2, e_4],\quad
F_{\infty}^1={\bmatrix {0} & \tau_2 & 1 & 0\\
\tau_2 & {\tau_3} & 0 & p \endbmatrix}.$$ for the family $p_1:{{X}}_D\to{{\Bbb H}}_1(M)$ and $$W_0=[e_2],\quad
W_1=[e_1, e_2, e_3],\quad
F_{\infty}^1={\bmatrix {\tau_1} & \tau_2 & 1 & 0\\
\tau_2 & {0} & 0 & p \endbmatrix}.$$ for the family $p_2:{{X}}_D\to{{\Bbb H}}_1(M)$.
We shall now restrict our attention to $W_1$. Here the Hodge filtration is given by $$F_{\infty}^1\cap W_1 =
[\tau_2 e_1 + \tau_3 e_2 + pe_4]$$ resp. $$F_{\infty}^1\cap W_1 =
[\tau_1 e_1 + \tau_2 e_2 + e_3].$$ By abuse of notation we will denote this by $F_{\infty}^1$, too. Finally, recall that all elements of the nilpotent orbit $\{T^{\tau}F_{\infty}^1\}$ of Hodge filtrations on $W_2=H^1(X_t)$ induce the same MHS on $W_1$ (cf. [@G p.84]).
Calculation of $\mbox{Gr}_1^W$ and the extension class {#subsection36}
------------------------------------------------------
In the previous section we computed the limit MHS on $H^1(X_t)$ determined by $p_i:X\to\Delta^*(M)$ ($i=1,2$). We now want to compute $\operatorname{Gr}_1^W$ and the extension class.
For any integer $n$ and $\tau\in{{\Bbb H}}_1$ we define $$X_{\tau,n}=V/\Lambda={{\Bbb C}}f/({{\Bbb Z}}\mu_1+{{\Bbb Z}}\mu_2)$$ where $\mu_1=\tau f$ and $\mu_2=nf$.
\[P1\] $\operatorname{(i)}$ In the central case ($i=1$) $\operatorname{Gr}_1^W=H^1(C)$ where $C=X_{\tau_3,p}$.
$\operatorname{(ii)}$ In the peripheral case ($i=2$) $\operatorname{Gr}_1^W=H^1(C)$ where $C=X_{\tau_1,1}$.
\(i) Setting $\mu_1^*=e_2$ and $\mu_2^*=e_4$ we have $$F^1=[f^*]=[\tau_3\mu_1^*+p\mu_2^*]=[\tau_3e_2+pe_4]$$ and the claim follows immediately from the calculations of \[subsection35\].
\(ii) Setting $\mu_1^*=e_1$ and $\mu_2^*=e_3$ we have $$F^1=[f^*]=[\tau_1\mu_1^*+\mu_2^*]=[\tau_1e_1+e_3]$$ and the result follows as before.
Next we want to calculate the class of the extensions $$\label{2} 0\longrightarrow
W_0\stackrel{i}{\longrightarrow}W_1\stackrel{\pi}{\longrightarrow}
\operatorname{Gr}_1^W\longrightarrow 0.$$
The recipe for finding the extension class of (\[2\]) demands that we choose an integral retraction $r: W_1\to W_0$ and then transfer $F_{\infty}^1$ to $W_0\oplus {{\operatorname{Gr}}}_1^W$ via $(r,\pi)$. In the central case we take $r(e_1)=e_1$ and $r(e_2)=r(e_4)=0$ and in the peripheral case $r(e_2)=e_2$ and $r(e_1)=r(e_3)=0$. Then, in both cases, $(r,\pi)={{\bf 1}}_{{{\operatorname{Gr}}}_1^W}$, hence $\tilde{F}_{\infty}^1=F_{\infty}^1$. Next we define the trivial extension $$\tilde{F}_0^1 = F^1(W_0)\oplus F^1({{\operatorname{Gr}}}_1^W)
= F^1({{\operatorname{Gr}}}_1^W) = [\tau_3 e_2 +pe_4]$$ for the first family and similarly $$\tilde{F}_0^1 =[\tau_1 e_1 + e_3]$$ for the second family. The extension class $e\in
J^0{{\operatorname{Hom}}}({{\operatorname{Gr}}}_1^W,W_0)$ is represented by any map $\psi: {{\operatorname{Gr}}}_1^W \to W_0$ such that $${\pmatrix {{\bf 1}}_{W_0} & \psi\\ 0 & {{\bf 1}}_{{{\operatorname{Gr}}}_1^W}
\endpmatrix} \tilde{F}_0^1 =\tilde{F}_{\infty}^1.$$ For the first family we can take $\psi=(\tau_2/p)
e_4^{\ast}\otimes e_1$ and for the second family $\psi=\tau_2
e_3^{\ast} \otimes e_2$.
After identifying $W_0={{\Bbb Z}}$ and ${{\operatorname{Gr}}}_1^W=H^1(C)$, where $C$ is as in proposition \[P1\], the extension class $e=[\psi]$ is well defined in $$\begin{aligned}
\operatorname{Ext}(\operatorname{Gr}_1^W,W_0)
&= J^0 {{\operatorname{Hom}}}({{\operatorname{Gr}}}_1^W,W_0)\\
&= H^1(C)^\ast / (H^{0,1}(C))^\ast + H^1(C,{{\Bbb Z}})^\ast\\
&= H^{1,0}(C)^\ast / H_1(C,{{\Bbb Z}})\\
&={{\operatorname{Alb}}}(C)\end{aligned}$$ and is represented by $\psi|_{H^{1,0}}$.
\[P2\] Under the identification $\operatorname{Ext}(\operatorname{Gr}_1^W,W_0)=\operatorname{Alb}(C)$ the extension class of (\[2\]) corresponds to $[\tau_2f]\in\operatorname{Alb}(C)$.
Notice that if $C=V/\Lambda$, then $H^1(C,{{\Bbb C}})={{\operatorname{Hom}}}_{{{\Bbb R}}}(V,{{\Bbb C}})$ and $$\begin{CD}
H^1(C,{{\Bbb C}}) @>{\int_{0}^{[v]}}>> {{\Bbb C}}/{\operatorname{periods}}\\
@| @AAA\\
{{\operatorname{Hom}}}_{{{\Bbb R}}}(V,{{\Bbb C}}) @>{\operatorname{ev}(v)}>> {{\Bbb C}}\end{CD}$$ commutes.
The extension class $e\in{{\operatorname{Ext}}}({{\operatorname{Gr}}}_1^W,W_0)$ corresponds to $[\tau_2
f]\in{{\operatorname{Alb}}}(C)$ since $$\begin{aligned}
\int_0^{[\tau_2 f]} (\tau_3 e_2 + p e_4)
&= \int_0^{[\tau_2 f]} (f^{\ast})\\
&=\tau_2\\
&=\psi(\tau_3 e_2 + p e_4)\end{aligned}$$ in the central case and $$\begin{aligned}
\int_0^{[\tau_2 f]} (\tau_1 e_1 + e_3)
&= \int_0^{[\tau_2 f]} (f^{\ast})\\
&=\tau_2\\
&=\psi(\tau_1 e_2 + e_3)\end{aligned}$$ in the peripheral case.
The number of components {#sec47}
------------------------
In this section we will proof the following proposition.
\[tnoc\] Let $\pi\colon X \to \Delta$ be a semi-stable degeneration of $(1,p)$-polarized abelian surfaces with (canonical) level structure. Assume that the central fibre $X_0$ is a cycle of $N$ ruled surfaces. Then the number $N$ is determined by the polarized VHS (PVHS) on $\Delta^\ast$.
The polarization on the VHS is induced by the polarization ${{\cal L}}$ in the following way: For $Y\subset X$ define $$c_Y = c_1({{\cal L}}|_Y) \in H^2(Y)$$ and $$Q_Y \colon (a,b) \mapsto a \cup b \cup c_Y\colon
H^1(Y) \times H^1(Y) \to H^4(Y).$$ We are mostly interested in the case where $Y$ is $X$, $X_0$ or $X_t$. We write $c_0$ for $c_{X_0}$ and $Q_0$ for $Q_{X_0}$. Similarly, we write $c_t$ for $c_{X_t}$ and $Q_t$ for $Q_{X_t}$.
The existence of a global level structure is needed to insure that $\deg({{\cal L}}|_{F_k})$, where $F_k$ is the fibre of $Y_k$, is independent of $k$. Indeed, by lemma 4 of §2.2 it is always 1.
For the rest of the section $\pi\colon X \to \Delta$ is as in proposition \[tnoc\]. As in §3.1 we identify $C$ with $C_0=Y_0\cap
Y_1$. Let $Y_k$ and $\tau_k\colon Y_k \to C$ be as in §3.1 and let $i\colon C{\hookrightarrow}X_0$ and $j_k\colon Y_k {\hookrightarrow}X_0$ be the inclusion maps.
\[cpt\] Under the isomorphism $(\tau_k^\ast)\colon H^1(C) \to {{\operatorname{Gr}}}_1^W$, the induced map $(j_k^\ast)\colon W_1 \to {{\operatorname{Gr}}}_1^W$ corresponds to the map $i^\ast\colon H^1(X_0)\to H^1(C)$ induced by the inclusion $i\colon C{\hookrightarrow}X_0$.
This follows immediately from $j_k^\ast=\tau_k^\ast i^\ast\colon H^1(X_0) \to H^1(Y_k)$, which in turn follows from the following facts: $\rho_k^\ast$ is the inverse of $\alpha_k^\ast$, $\sigma_k^\ast$ is the inverse of $\beta_k^\ast$, $i=i_0=j_1\beta_0$, and $j_{k+1}\beta_k= j_k\alpha_k$.
The PVHS on $\Delta^\ast$ determines the bilinear form $Q_t$ on $W_2=H^1(X_t)$. Since the maps on cohomology induced by an inclusion $Y'\subset Y$ are compatible with $Q_Y$ and $Q_{Y'}$, $Q_t$ induces $Q_0$ on $H^1(X_0)=W_1\subset W_2$, i.e., we have a commutative diagram $$\begin{CD}
H^1(X_t) \otimes H^1(X_t) @>{Q_t}>> H^4(X_t)\\
@A{k_t^\ast\times k_t^\ast}AA @AA{k_t^\ast}A\\
H^1(X) \otimes H^1(X) @>{Q}>> H^4(X)\\
@V{k_0^\ast\times k_0^\ast}VV @VV{k_0^\ast}V\\
H^1(X_0) \otimes H^1(X_0) @>{Q_0}>> H^4(X_0).
\end{CD}$$ For $t\in\Delta$ let $k_t\colon X_t{\hookrightarrow}X$ be the inclusion map and let $$\delta\colon H^4(X_0) \to H^4(X_t)$$ be $k_t^\ast (k_0^\ast)^{-1}$ (recall that $k_0^\ast$ is an isomorphism). Since the monodromy is compatible with $Q_t$, i.e., $
Q_t(Ta,Tb)= Q_t(a,b),
$ so is $N=T-{{\bf 1}}$, i.e. $
Q_t(Na,b)=Q_t(a,Nb).
$ Since $W_1=$ Ker $N$ and $W_0=$ Im $N$ $
Q_t(W_1,W_0)=0.
$ In other words, there exists a bilinear form $q$ on $H^1(C)$ making the diagram
(190,70) (0,0)[(90,10)[$H^1(C)\times H^1(C)$]{}]{} (0,50)[(90,10)[$H^1(X_0)\times H^1(X_0)$]{}]{} (130,50)[(60,10)[$H^4(X_t)={{\Bbb Z}}$]{}]{} (45,40)[(0,-1)[20]{}]{} (5,20)[(40,20)[$i^\ast\times i^\ast$]{}]{} (100,55)[(1,0)[20]{}]{} (75,20)[(2,1)[40]{}]{} (100,18)[$q$]{} (100,55)[(20,20)[$\delta Q_0$]{}]{}
commute. Furthermore, since $i^\ast\colon H^1(X_0) \to
H^1(C)$ is surjective, $q$ is determined by $\epsilon Q_0$ where $$\varepsilon: H^2(C)\longrightarrow H^4(X_t)$$ is the isomorphism which sends the generator of $H^2(C)$ corresponding to the canonical orientation of $C$ to the same thing on $X_t$. Let $$\gamma:H^1(X)\times H^1(C)\longrightarrow H^2(C)$$ be the intersection form. We will show in corollary \[mcc\] that $q=N\varepsilon\gamma$. This means that the data over $\Delta^*$ determine $N$.
For convenience, we shall first assume that $N\ge2$. The case $N=1$ requires minor modifications and is dealt with later.
Since the line bundle determining $Y_k$ has degree $0$, there exists a continuous map $q_k\colon Y_k \to {{\Bbb P}}^1$ such that $$(\tau_k,q_k)\colon Y_k \to C \times {{\Bbb P}}^1$$ is a homeomorphism. Let $f\in H^2({{\Bbb P}}^1)$ be the canonical generator, determined by the complex structure. We define $\beta$ by demanding that the diagram $$\begin{CD} H^2(C) @>{\beta}>> H^4(X_0)\\ @V{({{\bf 1}},\cdots,{{\bf 1}})\otimes
f}VV @VV{(j_k^\ast)}V\\
\oplus H^2(C)\otimes H^2({{\Bbb P}}^1) @>>{(\tau_k^\ast \cup q_k^\ast)}>
\oplus H^4(Y_k)
\end{CD}$$ be commutative. (Notice that $(j_k^\ast)$ is an isomorphism.)
\[mcl1\] The diagram $$\begin{CD}
H^1(X_0)\times H^1(X_0) @>{Q_0}>> H^4(X_0)\\
@V{i^\ast\times i^\ast}VV @AA{\beta}A\\
H^1(C) \times H^1(C) @>>{\gamma}> H^2(C)
\end{CD}$$ commutes.
By definition of $\beta$ we have to show that $$\begin{CD}
H^1(X_0)\times H^1(X_0) @>{Q_0}>> H^4(X_0) @>{(j_k^\ast)}>>
\oplus H^4(Y_k)\\
@V{i^\ast\times i^\ast}VV @. @AA{(\tau_k^\ast\cup q_k^\ast)}A\\
H^1(C) \times H^1(C) @>>{(\gamma,\ldots,\gamma)}>
\oplus H^2(C) @>>{{}\otimes f}>
\oplus H^2(C) \otimes H^2({{\Bbb P}}^1)
\end{CD}$$ commutes. In lemma 4 of §2 we showed that $\deg({{\cal L}}|_{F_k})=1$ for all $k$, where $F_k$ is the ruling of $Y_k$. Hence we can write $
j_k^\ast (c_0) = \tau_k^\ast (d_k) + q_k^\ast (f)
$ for some $d_k\in H^2(C)$. Since $j_k^\ast=\tau_k^\ast i^\ast$ on $H^1(X_0)$ (see proof of lemma \[cpt\]) $$j_k^\ast Q_0(a,b) = \tau_k^\ast i^\ast (a\cup b) \cup
(\tau_k^\ast (d_k) + q_k^\ast (f))
= \tau_k^\ast i^\ast (a \cup b) \cup q_k^\ast (f),$$ because $i^\ast(a\cup b)\cup d_k\in H^4(C)=0$.
The orientations of the components of the singular fibre are compatible with the orientation of the general fibre in the following sense:
\[mcl2\] If $N$ is the number of components of $X_0$, then $\delta\beta=N\epsilon$.
By definition of $\beta$, it suffices to show that the diagram $$\begin{CD}
H^4(X_0) @<{k_0^\ast}<{\sim}< H^4(X)\\
@V{(j_k^\ast)}V{\sim}V @VV{k_t^\ast}V\\
\oplus H^4(Y_k) @>>{({{\bf 1}},\ldots,{{\bf 1}})}> H^4(X_t)
\end{CD}$$ commutes. The inverse of $k_0^\ast$ is induced by a retraction $r\colon X \to X_0$, which exhibits $X_0$ as a strong deformation retract of $X$ [@Cl; @P; @St]. This retraction restricts to $r_t\colon X_t\to X_0$. Let $Z_k\subset X_t$ be the inverse image of $Y_k$. Then $S_k:=Z_k\cap Z_{k+1}$ is the inverse image of $T_k:=Y_k\cap Y_{k+1}$. Pick any ${\kappa}\in\{1,\ldots,N\}$. Let $Y'=Y_{{\kappa}}$, $Y''=\cup_{k\neq {\kappa}} Y_k$, $Z'=Z_{{\kappa}}$ and $Z''=\cup_{k\neq {\kappa}} Z_k$. Let $S=Z'\cap
Z''=S_{{\kappa}-1}\cup S_{{\kappa}}$ and $T=Y'\cap Y''=T_{{\kappa}-1}\cup T_{{\kappa}}$.
We will need below that $H^4(Z')=H^4(Z'')=0$. To see this, we use that the retraction $r_t\colon Z_k\to Y_k$ is the real oriented blow up along $T_{k-1}\cup T_k$ [@P p. 36]. The exceptional divisors $S_k=r_t^{-1}(T_k)$ are $S^1$-bundles over $T_k$. They are trivial, because the triple $(Y_k, T_{k-1}, T_k)$ is homeomorphic to $(B\times S^2, B\times\{0\}, B\times\{\infty\})$, where $B=S^1\times
S^1$. It follows that the triple $(Z_k,S_{k-1},S_k)$ is homeomorphic to $(B\times S^1\times [0,1], B\times S^1\times \{0\}, B\times S^1\times
\{1\})$. In particular, $H^4(Z_k)=0$ and $H^3(Z_k)\to H^3(S_k)$ and $H^3(Z_k)\to H^3(S_{k-1})$ are surjective for all $k$. Mayer-Vietoris now implies that $H^4(Z'')=\oplus_{k\neq {\kappa}} H^4(Z_k)$ and this proves the claim.
Consider the commutative diagram $$\begin{CD}
H^4(Y') @<<< H^4(X_0) @>{r_t^\ast}>> H^4(X_t)\\
@A{\sim}AA @AAA @AA{\sim}A\\
H^4(Y',T) @<{\sim}<< H^4(X_0,Y'') @>{\sim}>{r_t^\ast}>
H^4(X_t,Z'')
\end{CD}$$ The map $H^4(X_0,Y'')\to H^4(X_t,Z'')$ is an isomorphism by Alexander duality since $r_t$ is an homeomorphism $X_0\setminus Y''\to X_t\setminus Z''$ (cf. [@La p.23]). Similarly, $ H^4(X_0,Y'')\to H^4(Y',T)$ is an isomorphism. The left vertical map is an isomorphism because $T$ has topological dimension 2. The right vertical map is surjective because $H^4(Z'')=0$, as we have seen above. But a surjective map form ${{\Bbb Z}}$ to ${{\Bbb Z}}$ is automatically injective.
It follows from [@Gre p.99] that the map $H^4(Y')\to H^4(X_0)$ via $H^4(Y',T)$ and $H^4(X_0,Y'')$ composed with the Mayer-Vietoris isomorphism $H^4(Y')\oplus H^4(Y'') \cong H^4(X_0)$ is just the inclusion of $H^4(Y')$ into $H^4(Y')\oplus H^4(Y'')$. We have to show that the composition of this map with $r_t^\ast\colon H^4(X_0)\to H^4(X_t)$ maps the orientation class to the orientation class. As before, we may identify the relative cohomology groups with the cohomology with compact supports of the respective complements. But $r_t\colon
X_t\setminus Z''\to Y'\setminus T$ is a complex isomorphism. In particular, it respects the orientation.
\[mcc\] $q=N\epsilon\gamma$.
By lemma \[mcl1\] $q=\delta\beta\gamma$. Now the result follows immediately from the previous lemma \[mcl2\].
Finally, we consider the case $N=1$. We proceed as in the case $N\ge 2$. First we have to define $\beta\colon H^2(C)\to H^4(X_0)$. To do this, notice that we have an exact Mayer-Vietoris sequence $$H^{q-1}(\tilde{C})\to
H^q(X_0)\to H^q(\tilde{X_0})\oplus H^q(C)\to H^q(\tilde{C}),$$ where $\tilde{X_0}$ is the normalization of $X_0$ and $\tilde{C}=\pi^{-1}(C)$ (see [@C p.120]). Hence the normalization map $\pi\colon \tilde{X_0}\to X_0$ induces an isomorphism $H^4(X_0)\to H^4(\tilde{X_0})$. We identify $H^4(X_0)$ with $H^4(\tilde{X_0})$ via $\pi^\ast$ and define $\beta$ by demanding that it maps the orientation class of $C$ to the orientation class of $\tilde{X_0}$. With this definition of $\beta$ and using that the degree $\deg(\pi^\ast({{\cal L}})|_F)$ of ${{\cal L}}$ along the fibre $F$ in the normalization is 1 again by lemma 4 of §2, the diagram in lemma \[mcl1\] again commutes. The proof is an easy adaption of the proof given for $N\ge2$ and is left to the reader.
To prove lemma \[mcl2\] for $N=1$, one argues exactly as in the case $N\ge2$ using the commutative diagram $$\begin{CD}
H^4(\tilde{X_0}) @<<< H^4(X_0) @>{r_t^\ast}>> H^4(X_t)\\
@AAA @AAA @AAA\\
H^4(\tilde{X_0},\tilde{C}) @<<< H^4(X_0,C)
@>{r_t^\ast}>> H^4(X_t,S),
\end{CD}$$ where $S=r_t^{-1}(C)$. (This time, all maps are isomorphisms.)
A uniqueness result {#section5}
===================
Here we combine the results of sections \[section3\] and \[section4\] to prove a uniqueness result for degenerate abelian surfaces.
In (\[subsection34\]) we discussed central and peripheral corank 1 boundary points of ${\cal A}^*(1,p)$. Note that the universal family $\pi:X_D\to{{\Bbb H}}_2$ descends to a family $\pi':X_D\to e({{\Bbb H}}_2)$. By [@HKW part II.4] this family can be extended to a family $\bar\pi:\bar{X}_D\to X(l)$. The fibres of $\bar{X}_D$ over the ”boundary” $\{0\}\times{{\Bbb C}}\times{{\Bbb H}}_1$, resp. ${{\Bbb H}}_1\times{{\Bbb C}}\times\{0\}$ are cycles of elliptic ruled surfaces with $N=1$, resp. $N=p$.
\[D\] Let $[q]\in{\cal A}^*(1,p)$ be a corank 1 boundary point. We say that a surface $X_0$ is a [*degenerate abelian surface associated to*]{} $[q]$ if the following holds: There exists a 1-parameter degeneration of polarized abelian surfaces with (canonical) level structure $p:X\to\Delta$ as in section 2.2 with central fibre $p^{-1}(0)=X_0$ and an embedding $f:\Delta\to X(l)$ such that:
\(i) $f(\Delta)$ meets $\{0\}\times{{\Bbb C}}\times{{\Bbb H}}_1$ resp. ${{\Bbb H}}_1\times{{\Bbb C}}\times\{0\}$ transversely in the point $q=f(0)$ which is mapped to $[q]$.
\(ii) The restriction of $p:X\to\Delta$ to $\Delta^*$ is the pull back of the family $\pi':X_D\to e({{\Bbb H}}_2)$ via $f$.
The reason why we work with $X(l)$ rather than with the compactification ${\cal A}^*(1,p)$ itself is that the map ${{\Bbb H}}_2\to\cal A(1,p)$ is branched over two Humbert surfaces $H_1$ and $H_2$. Near these surfaces, resp. their closure in ${\cal
A}^*(1,p)$ one has no universal family.
\[lem25\] If $X_0$ is a degeneration with only local normal crossing singularities associated to a corank 1 boundary point $[q]$, then the MHS on $H^1(X_0)$ is determined by this point.
This follows from the Local Invariant Cycle theorem [@G Ch. VI] (which also holds when we have local normal crossing rather than global normal crossing) together with the existence of the extended family $\bar{\pi}:\bar{X}_D\to X(l)$: Indeed, compare the family $p:X\to\Delta$ as in definition (\[D\]) with the family $f^{\ast}\bar{X}_D\to\Delta$. On $\Delta^{\ast}$ these two families agree (they are both the pull back of the universal family), hence they determine the same VHS and thus the same limit MHS on $H^1(X_t)$ ($t\in\Delta^{\ast}$ arbitrary but fixed). But by the Local Invariant Cycle theorem the MHS on the central fibre is determined by this limit MHS at least over ${{\Bbb Q}}$: more precisely, $H^1(X_0)=W_1(H^1(X_t))=H^1(\bar{\pi}^{-1}(q))$ as ${{\Bbb Q}}-$MHS. It remains to show that $$0\longrightarrow H^1(X)\longrightarrow
H^1(X_t)\stackrel{N}{\longrightarrow} H^1(X_t)$$ is exact over ${{\Bbb Z}}$. For $n\gg0$ the zero locus $Y\subset X$ of a general element of $\Gamma (X,{{\cal L}}^n)$ is a semi-stable degeneration of the smooth curve $Y_t=Y\cap X_t$. By the Picard-Lefschetz theorem [@BPV theorem III. 14.1] $$0\longrightarrow H^1(Y)\longrightarrow
H^1(Y_t)\stackrel{N}{\longrightarrow} H^1(Y_t)$$ is exact over ${{\Bbb Z}}$. By the Lefschetz hyperplane theorem $H^1(X_t)\longrightarrow H^1(Y_t)$ is injective, hence so is $H^1(X)\longrightarrow H^1(Y)$. We have now reduced our problem to showing that $H^1(Y)/H^1(X)$ is torsion free. Since $H^1(Y)/H^1(X)\hookrightarrow H^2(X,Y)$ it is sufficient to show that $H^2(X,Y)$ or, equivalently $H_1(X,Y)=\text{Coker }
(H_1(Y)\longrightarrow H_1(X))$ is torsion free. Since dim ker $N=3$ the central fibre $X_0$ is not smooth abelian. Hence by Persson’s theorem $X_0$ is a cycle of elliptic ruled surfaces. Assume that $X_0$ has at least two components. (The remaining case is analogous.) Let $C$ be one of the double curves and $X_i$ one of the components of $X_0$. Then $$H_1(X)=H_1(X_0)\cong H_1(C)\oplus {{\Bbb Z}}c$$ where $c\in H_1(X_0)$ is as in 3.3. Let $Y_i=X_i\cap Y$. Then $X_i$ and $Y_i$ are smooth and irreducible. We have a commutative diagram $$\begin{CD}
H_1(C)= @. H_1(X_i) @>>> H_1(X)\\
@. @AAA @AAA\\
@. H_1(Y_i) @>>> H_1(Y).
\end{CD}$$ Since $c$ is clearly induced from $Y$ up to $H_1(C)$, it is sufficient to prove that $$H_1(Y_i)\longrightarrow H_1(X_i)$$ is surjective. But this follows from the Lefschetz hyperplane theorem since $Y_i$ is ample on $X_i$. This shows that $H^1(X_0)=W_1(H^1(X_t))$ depends only on the point$[q]$.
\[T\] Let $X_0$ be a degeneration associated to a corank 1 boundary point $[q]$ of ${\cal A}^*(1,p)$ with only local normal crossing singularities and no triple points. Then $X_0$ is a cycle of ruled surfaces (possibly with only one component) over an elliptic curve $C$. There exists a ${{\cal L}}\in{{\operatorname{Pic}}}^0(C)$ such that all components are isomorphic to ${{\Bbb P}}^1({{\cal O}}\oplus{{\cal L}})$. If $q$ is a central boundary point then $X_0$ is completely determined by $q$. If $q$ is peripheral, then the number $N$ of components, the base curve $C$ andthe shift $s$ are uniquely determined by the point $[q]$. The line bundle ${{\cal L}}\in{{\operatorname{Pic}}}^0(C)$ determining the elliptic ruled surfaces of the cycle is determined (at least) up to a $p$-torsion point of ${{\operatorname{Pic}}}^0(C)$. More precisely:
1. If $[q]=[\tau_2,\tau_3]$ is a central boundary point, then $N=1$, $C=X_{\tau_3,p}$, $s=[\tau_2]$ and ${{\cal L}}={{\cal O}}(s-O)$.
2. If $[q]=[\tau_1,\tau_2]$ is a peripheral boundary point, then $N=p$, $C=X_{\tau_1,1}$, $s=[\tau_2]$ and there exists a point $s'\in C$ such that ${{\cal L}}={{\cal O}}(s'-O)$.
We first note that the pair $(X_{\tau_3,p},[\tau_2])$, resp. $(X_{\tau_1,1},[\tau_2])$ only depends on the point $[q]$ but not on its representative [@HKW part I.3].
If $X_0$ has only global normal crossings then Persson’s result [@P proposition 3.3.1] says that $X_0$ is a cycle of elliptic ruled surfaces.
The previous lemma asserts that the MHS on $H^1(X_0)$ is determined by $q$. Furthermore, this MHS is the weight 1 part of the limit MHS on $H^1(X_t)$ of the family $p_1$ (resp. $p_2$) defined in §4.4 in the central (resp. peripheral) case. Indeed, $p_1$ (resp. $p_2$) is the restriction of the universal family to a small disk $\Delta\times (\tau_2,\tau_3)$ (resp. $(\tau_1,\tau_2)\times\Delta$). By lemma \[wf\] the base curve $C$ is determined by the equation $H^1(C)=
{{\operatorname{Gr}}}_1^W(H^1(X_0))={{\operatorname{Gr}}}_1^W(H^1(X_t))$ and this equation was solved in proposition \[P1\]. Corollary \[cor17\] shows how to recover the shift $s$ from $H^1(X_0)=W_1(H^1(X_t))$ and the result for the families $p_1$ and $p_2$ is given in proposition \[P2\].
Applying proposition \[tnoc\] of §\[sec47\] to the family $X$ defining $X_0$ and to the family constructed in [@HKW], one sees that the number of components of $X_0$ is independent of the choice of the disk $\Delta{\hookrightarrow}X(l)$ (cf. proof of lemma \[lem25\]). But it follows from our calculations in §4 that the polarization induces $\left(\begin{array}{cc} 0& -1\\1&0\end{array}\right)$ (with respect to the basis $\{e_2,e_4\}\bmod e_1$ of §\[subsection35\]) in the central case and $\left(\begin{array}{cc} 0& -p\\p&0\end{array}\right)$ (with respect to the basis $\{e_1,e_3\}\bmod e_2$) in the peripheral case. It follows from proposition \[tnoc\] of §\[sec47\] that the number of components is 1 and $p$ respectively.
The statement about the form of ${{\cal L}}\in{{\operatorname{Pic}}}^0(C)$ now follows directly from proposition 6 of §2.2
Note that in the proof of theorem \[T\] the level structure is not needed to determine $C$ and $s$.
It seems reasonable to expect that also in the peripheral case ${{\cal L}}$ is completely determined by the boundary point $q$. To show this in the spirit of this paper, one probably has to translate the notion of (canonical) level structure to Hodge structures.
This result is compatible with [@HKW]. Notice that we only used the existence of the degeneration $\bar{X}_D$ as constructed there in order to prove our uniqueness result, but not the precise description of its singular fibres.
[CCK]{} W. Barth, C. Peters, A. Van de Ven, [*Compact complex surfaces*]{}, Springer Verlag Berlin, (1984) J. Carlson, [*Extensions of mixed Hodge structures*]{}, Journées de géométrie algébriques d’Angers, (1979), Sijthoff & Noordhoff, (1980), pp. 107–127. J. Carlson, E. Cattani, A. Kaplan, [*Mixed Hodge structures and compactifications of Siegel’s space*]{}, Journées de géométrie algébriques d’Angers, (1979), Sijthoff & Noordhoff, (1980), pp. 77–105. C.H. Clemens, [*Degeneration of Kähler manifolds*]{}, Duke Math. J. [**44**]{}, 215–290, (1977). R. Friedman, D. Morrison, [*The birational geometry degenerations*]{}, Progress in Math. [**29**]{}, Birkhäuser (1983). M. J. Greenberg, J. R. Harper, [*Algebraic topology: A first course*]{}, The Benjamin/Cummings Publishing Company, Reading, Massachusetts (1981). Ph. Griffiths, [*Topics in transcendental algebraic geometry*]{}, Annals of math. studies, 106, Princeton University Press, (1984). Ph. Griffiths, J. Morgan, [*Rational homotopy theory and differential forms*]{}, Progress in Math. [**16**]{}, Birkhäuser, Boston, (1981). Ph. Griffiths, W. Schmid, [*Recent developments in Hodge theory: a discussion of techniques and results*]{}, in Discrete subgroups of Lie groups, Bombay, Oxford Univ. Press (1973), pp. 31–127. K. Hulek, C. Kahn, S.H. Weintraub, [*Moduli spaces of abelian surfaces: compactification, degenerations, and theta functions*]{}, De Gruyter Exposition in Mathematics [**12**]{}, Walter de Gruyter, Berlin, New York (1993). K. Lamotke, [*The topology of complex projective varieties after S. Lefschetz*]{}, Topology [**20**]{}, 15–51, (1981). H. Lange, Ch. Birkenhake, [*Complex Abelian Varieties*]{}, Grundlehren der mathematischen Wissenschaften [**302**]{}, Springer-Verlag, Berlin (1992). U. Persson, [*On degenerations of algebraic surfaces*]{}, Memoirs of the AMS [**189**]{}, (1977). J.H.M. Steenbrink, [*Limits of Hodge structures*]{}, Inventiones math. [**31**]{}, 229–257, (1976). F. Warner, [*Foundations of differentiable manifolds and Lie groups*]{}, Graduate texts in maths. [**94**]{}, Springer-Verlag, New York (1983).
| {
"pile_set_name": "ArXiv"
} |
[**Unbounded perturbations of two-dimensional diffusion processes with nonlocal boundary conditions**]{}
Pavel Gurevich[^1]
[**1.**]{} A general form of a generator of a strongly continuous contractive nonnegative semigroup (Feller semigroup) of operators acting between the spaces of continuous functions on an interval was investigated in [@Feller2]. In the multidimensional case, it was proved that the generator of a Feller semigroup is an elliptic differential operator (possibly with degeneration) whose domain of definition consists of continuous functions satisfying nonlocal conditions which involve an integral over the closure of the region with respect to a nonnegative Borel measure [@Ventsel]. The inverse problem remains open: given an elliptic integro-differential operator whose domain of definition is described by nonlocal boundary conditions, whether or not the closure of this operator is a generator of a Feller semigroup. In [@SU; @BCP; @Taira3; @Ishikawa; @GalSkJDE], the authors considered the [*transversal*]{} case, where the order of nonlocal terms is less than the order of local terms. The [*nontransversal*]{} case, where these orders coincide, was studied in [@SkubDAN89; @GalSkJDE] (see also the bibliography in [@GalSkJDE]).
It was assumed in [@GalSkJDE] that the coefficients at the nonlocal terms are less than one. In this paper, we consider nontransversal nonlocal conditions on the boundary of a plane domain and investigate the “limit case,” where the coefficients at nonlocal terms may equal one at some points (the Borel measure is assumed to be atomic at these points). We formulate sufficient conditions on unbounded perturbations of elliptic operator and on the Borel measure in nonlocal conditions which guarantee that the corresponding nonlocal operator is a generator of a Feller semigroup.
Let $G\subset{{\mathbb R}}^2$ be a bounded domain with boundary ${{\partial G}}$. Consider a set ${{{\mathcal K}}}\subset\partial G$ consisting of finitely many points. Let $\partial G\setminus{\mathcal
K}=\bigcup_{i=1}^{N}\Gamma_i$, where $\Gamma_i$ are open (in the topology of $\partial G$) $C^\infty$ curves. Assume that the domain $G$ is a plane angle in some neighborhood of each point $g\in{\mathcal K}$.
For any closed sets $Q\subset{{\overline G}}$ and $K\subset{{\overline G}}$ such that $Q\cap K\ne\varnothing$, we introduce the space $ C_K(Q)=\{u\in C(Q): u(y)=0,\ y\in Q\cap K\} $ with the maximum-norm. Introduce the space $H_a^k(G)$ as a completion of the set of infinitely differentiable functions vanishing near $\mathcal K$ with respect to the norm $$\|u\|_{H_a^k(G)}=\Bigg(
\sum_{|\alpha|\le k}\int_G \rho^{2(a+|\alpha|-k)} |D^\alpha u(y)|^2 dy
\Bigg)^{1/2},$$ where $a\in \mathbb R$, $k\ge 0$ is an integer, and $\rho=\rho(y)={{\rm dist}}(y,\mathcal K)$. For an integer $k\ge1$, denote by $H_a^{k-1/2}(\Gamma)$ the space of traces of the functions from $H_a^k(G)$ on a smooth curve $\Gamma\subset\overline{G}$. We also introduce the norms in weighted spaces depending on the parameter $q>0$. Set $
\|v\|_{H_{a}^0(\Gamma_i)}=\big(\,\int_{\Gamma_i}\rho^{2a}|v(y)|^2d\Gamma\big)^2,
$ $${{|\!|\!|}}u {{|\!|\!|}}_{H_a^k(G)}=
\left(\|u\|_{H_{a}^k(G)}^2+q^k\|u\|^2_{H_a^0(G)}\right)^{1/2},\qquad
k\ge0,$$ $${{|\!|\!|}}v {{|\!|\!|}}_{H_a^{k-1/2}(\Gamma_i)}=
\left(\|v\|_{H_a^{k-1/2}(\Gamma_i)}^2+q^{k-1/2}\|
v\|_{H_{a}^0(\Gamma_i)}^2\right)^{1/2},\qquad k\ge1.$$
[**2.**]{} Consider the differential operator $
P_0u=\sum_{i,j=1}^{2}p_{ij}(y)u_{y_iy_j}(y)+
\sum_{i=1}^2p_i(y)u_{y_i}(y)+p_0(y)u(y), $ where $p_{ij},p_i\in
C^\infty({{\mathbb R}}^2)$ are real-valued functions, $p_0\ge0$, and $p_{ij}=p_{ji}$, $i,j=1,2$. We assume that there is a constant $c>0$ such that $\sum_{i,j=1}^{2}p_{ij}(y)\xi_i\xi_j\ge c|\xi|^2$ for $y\in\overline{G}$ and $\xi=(\xi_1,\xi_2)\in{{\mathbb R}}^2.$
Introduce operators corresponding to nonlocal terms supported near the set $\mathcal K$. For any set $\mathcal M$, we denote its $\varepsilon$-neighborhood by $\mathcal
O_{\varepsilon}(\mathcal M)$. Let $\Omega_{is}$ ($i=1, \dots, N;$ $s=1, \dots, S_i$) be $C^\infty$ diffeomorphisms taking some neighborhood ${\mathcal O}_i$ of the curve $\overline{\Gamma_i\cap\mathcal O_{{\varepsilon}}(\mathcal K)}$ to the set $\Omega_{is}({\mathcal O}_i)$ in such a way that $\Omega_{is}(\Gamma_i\cap\mathcal
O_{{\varepsilon}}(\mathcal K))\subset G$ and $ \Omega_{is}(g)\in\mathcal K$ for $
g\in\overline{\Gamma_i}\cap\mathcal K. $ Thus, the transformations $\Omega_{is}$ take the curves $\Gamma_i\cap\mathcal O_{{\varepsilon}}(\mathcal K)$ inside the domain $G$ and the set of their end points $\overline{\Gamma_i}\cap\mathcal K$ to itself.
Let us specify the structure of the transformations $\Omega_{is}$. Denote by $\Omega_{is}^{+1}$ the transformation $\Omega_{is}:{\mathcal O}_i\to\Omega_{is}({\mathcal
O}_i)$ and by $\Omega_{is}^{-1}:\Omega_{is}({\mathcal
O}_i)\to{\mathcal O}_i$ the inverse transformation. The set of points $\Omega_{i_qs_q}^{\pm1}(\dots\Omega_{i_1s_1}^{\pm1}(g))\in{\mathcal
K}$ ($1\le s_j\le S_{i_j},\ j=1, \dots, q$) is said to be an [*orbit*]{} of the point $g\in{\mathcal K}$. In other words, the orbit of a point $g\in{{\mathcal K}}$ is formed by the points (of the set $\mathcal
K$) that can be obtained by consecutively applying the transformations $\Omega_{i_js_j}^{\pm1}$ to the point $g$. The set $\mathcal K$ consists of finitely many disjoint orbits, which we denote by $\mathcal K_\nu$, $\nu=1,\dots,N_0$.
Take a sufficiently small number $\varepsilon>0$ such that there exist neighborhoods $\mathcal O_{\varepsilon_1}(g_j)$, $ \mathcal
O_{\varepsilon_1}(g_j)\supset\mathcal O_{\varepsilon}(g_j) $, satisfying the following conditions: 1. the domain $G$ is a plane angle in the neighborhood $\mathcal O_{\varepsilon_1}(g_j)$; 2. $\overline{\mathcal O_{\varepsilon_1}(g)}\cap\overline{\mathcal
O_{\varepsilon_1}(h)}=\varnothing$ for any $g,h\in\mathcal K$, $g\ne h$; 3. if $g_j\in\overline{\Gamma_i}$ and $\Omega_{is}(g_j)=g_k,$ then ${\mathcal
O}_{\varepsilon}(g_j)\subset\mathcal
O_i$ and $\Omega_{is}\big({\mathcal
O}_{\varepsilon}(g_j)\big)\subset{\mathcal
O}_{\varepsilon_1}(g_k).$
For each point $g_j\in\overline{\Gamma_i}\cap\mathcal K_\nu$, we fix a linear transformation $Y_j: y\mapsto y'(g_j)$ (the composition of the shift by the vector $-\overrightarrow{Og_j}$ and rotation) mapping the point $g_j$ to the origin in such a way that $ Y_j({\mathcal O}_{\varepsilon_1}(g_j))={\mathcal
O}_{\varepsilon_1}(0),\ Y_j(G\cap{\mathcal
O}_{\varepsilon_1}(g_j))=K_j\cap{\mathcal O}_{\varepsilon_1}(0), $ $ Y_j(\Gamma_i\cap{\mathcal
O}_{\varepsilon_1}(g_j))=\gamma_{j\sigma}\cap{\mathcal
O}_{\varepsilon_1}(0)\ (\sigma=1\ \text{or}\ 2), $ where $
K_j$ is a plane angle of nonzero opening and $\gamma_{j\sigma}$ its sides.
\[condK1\] Let $g_j\in\overline{\Gamma_i}\cap\mathcal K_\nu$ and $\Omega_{is}(g_j)=g_k\in\mathcal K_\nu;$ then the transformation $
Y_k\circ\Omega_{is}\circ Y_j^{-1}:{\mathcal
O}_{\varepsilon}(0)\to{\mathcal O}_{\varepsilon_1}(0) $ is the composition of rotation and homothety.
Introduce the nonlocal operators: $
\mathbf B_{i}u=\sum_{s=1}^{S_i}
b_{is}(y) u(\Omega_{is}(y))$, $
y\in\Gamma_i\cap\mathcal O_{\varepsilon}(\mathcal K),$ and $
\mathbf B_{i}u=0,$ $ y\in\Gamma_i\setminus\mathcal
O_{\varepsilon}(\mathcal K),$ where $b_{is}\in C^\infty(\mathbb
R^2)$ are real-valued functions, ${{\rm supp\,}}b_{is}\subset\mathcal
O_{{\varepsilon}}(\mathcal K)$.
\[cond1.2\]
1. $ b_{is}(y)\ge0,\quad \sum_{s=1}^{S_i} b_{is}(y)\le 1,\quad
y\in\overline{\Gamma_i};
$
2. $
\sum_{s=1}^{S_i}
b_{is}(g)+\sum_{s=1}^{S_j} b_{js}(g)<2,\
g\in\overline{\Gamma_i}\cap\overline{\Gamma_j}\subset{{\mathcal K}},\quad\text{if}\
i\ne j\ \text{and}\
\overline{\Gamma_i}\cap\overline{\Gamma_j}\ne\varnothing.
$
\[th\] Let Conditions $\ref{condK1}$ and $\ref{cond1.2}$ be fulfilled. Then there is a number $\delta_0>0$ such that, for $k=0,1,2,\dots$, $\delta\in(0,\delta_0)$, $q> q_0(\delta)\ge 0$, and $\psi_i\in C_{{\mathcal K}}(\overline{\Gamma_i})\cap
H_{a}^{k+3/2}(\Gamma_i)$, $a=k+1-\delta$, the problem $$P_0u-q u =0, \ y\in G;\qquad u|_{\Gamma_i}-{{\mathbf B}}_i u=\psi_i(y), \
y\in\Gamma_i,\ i=1,\dots,N,$$ has a unique solution $u\in C^\infty(G)\cap C_{{\mathcal K}}(\overline G)\cap H_{a}^{k+2}(G);$ moreover, $$\|u\|_{C({{\overline G}})}+{{|\!|\!|}}u{{|\!|\!|}}_{H_{a}^{k+2}(G)}\le c
\sum_{i=1}^N\big(\|\psi_i\|_{C(\overline{\Gamma_i})}+{{|\!|\!|}}\psi_i{{|\!|\!|}}_{H_{a}^{k+3/2}(\Gamma_i)}\big),$$ where $c>0$ does not depend on $q$, $\psi_i$, or $u$.
From now on, we fix $\delta\in(0,\delta_0)$, an integer $k\ge2$, and $a=k+1-\delta$.
[**3.**]{} Consider a linear bounded operator $P_1: H_a^{k+2}(G)\to
H_{a-1}^k(G)$.
\[cond3.1\]
1. If a function $u\in H_a^{k+2}(G)$ achieves its positive maximum at a point $y^0\in G$, then $P_1 u(y^0)\le0$.
2. If $u\in C({{\overline G}})\cap H_a^{k+2}(G)$, then the function $P_1u$ is bounded on $G$.
3. For all sufficiently small $\varrho>0$, we have $
P_1=P_{1\varrho}^1+P_{1\varrho}^2, $ where the operators $P_{1\varrho}^1,P_{1\varrho}^2:H_a^{k+2}(G)\to H_{a-1}^k(G)$ are such that
1. $\| P_{1\varrho}^1u\|_{H_{a-1}^k(G)}\le c(\varrho)\|
u\|_{H_a^{k+2}(G)}$, $c(\varrho)>0$ does not depend on $u$ and $c(\varrho)\to0$ as $\varrho\to 0$,
2. the operator $P_{1\varrho}^2$ is compact.
Note that ${{\rm D}}(P_1)\subset C^k({{\overline G}}\setminus{{\mathcal K}})\subset C^2(G)$ and ${{\mathcal R}}(P_1)\subset C^{k-2}({{\overline G}}\setminus{{\mathcal K}})\subset C(G)$. However, ${{\mathcal R}}(P_1)\not\subset C({{\overline G}})$ in general.
The prototype of the abstract operators $P_1$, $P_{1\varrho}^1$, and $P_{1\varrho}^2$ are integral operators of the form (cf. [@GalSkJDE; @Taira3; @GarMenaldi]) $$\begin{aligned}
P_{1}u(y)&=\int_{F} [u(y+z(y,\eta))-u(y)-(\nabla
u(y),z(y,\eta))]m(y,\eta)\pi(d\eta),\\
P_{1\varrho}^1u(y)&=\int_{Z\le\varrho} [u(y+z(y,\eta))-u(y)-(\nabla
u(y),z(y,\eta))]m(y,\eta)\pi(d\eta),\\
P_{1\varrho}^2u(y)&=\int_{Z>\varrho} [u(y+z(y,\eta))-u(y)-(\nabla
u(y),z(y,\eta))]m(y,\eta)\pi(d\eta),
\end{aligned}$$ where $F$ is a space with a $\sigma$-algebra ${{\mathcal F}}$ and a Borel measure $\pi$, $ y+ z(y,\eta)\in{{\overline G}}$ and $|D^\alpha_y z(y,\eta)|\le Z(\eta)$ for $y\in{{\overline G}},\ \eta\in F$, $|\alpha|\le k$, $Z(\eta)$ is a nonnegative $\pi$-measurable bounded function, and $m(y,\eta)\ge0$ (some additional restrictions on the functions $z(y,\eta)$, $Z(\eta)$, and $m(y,\eta)$ should also be imposed).
[**4.**]{} In this paper, we consider the nonlocal conditions in the [*nontransversal*]{} case (see, e.g., [@Taira3] for the probabilistic interpretation): $$\label{eq57}
u(y)-\int_{{\overline G}}u(\eta)\mu_i(y,d\eta)=0,\ y\in\Gamma_i,\
i=1,\dots,N;\qquad u(y) =0,\ y\in{{\mathcal K}},$$ where $\mu_i(y,\cdot)$ is a nonnegative Borel measure on ${{\overline G}}$ such that $\mu_i(y,{{\overline G}})\le 1,\ y\in\Gamma_i$.
Introduce the measures $\delta_{is}(y,\cdot)$ as follows: $
\delta_{is}(y,Q)=b_{is}(y)\chi_Q(\Omega_{is}(y)),$ $y\in\Gamma_i\cap{{\mathcal O}}_\varepsilon({{\mathcal K}}),$ and $\delta_{is}(y,Q)=0$, $
y\in\Gamma_i\setminus{{\mathcal O}}_\varepsilon({{\mathcal K}})$, where $Q\subset{{\overline G}}$ is an arbitrary Borel set and $\chi_Q(\cdot)$ a characteristic function of the set $Q$.
We study those measures $\mu_i(y,\cdot)$ which can be represented in the form $$\mu_i(y,\cdot)=\sum_{s=1}^{S_i}\delta_{is}(y,\cdot)+\alpha_i(y,\cdot)+\beta_i(y,\cdot),\qquad
y\in\Gamma_i,$$ where $\alpha_i(y,\cdot)$ and $\beta_i(y, \cdot)$ are nonnegative Borel measures to be specified below.
Denote $ {{\rm spt\,}}\alpha_i(y,\cdot)={{\overline G}}\setminus\bigcup_{V\in T}\{V\in
T: \alpha_i(y,V\cap{{\overline G}})=0\} $ ($T$ is the set of all open sets in ${{\mathbb R}}^2$).
Set $ {{\mathbf B}}_{\alpha i}u(y)=\int_{{\overline G}}u(\eta)\alpha_i(y, d\eta),\
{{\mathbf B}}_{\beta i}u(y)=\int_{{\overline G}}u(\eta)\beta_i(y, d\eta),\
y\in\Gamma_i. $
We assume that the measures $\alpha_i(y,\cdot)$ and $\beta_i(y,\cdot)$ satisfy the following conditions (cf. [@GalSkJDE]).
\[cond2.3\] There exist numbers $\varkappa_1>\varkappa_2>0$ and $\sigma>0$ such that
1. ${{\rm spt\,}}\alpha_i(y,\cdot)\subset{{\overline G}}\setminus{{\mathcal O}}_{\varkappa_1}({{\mathcal K}})$, $y\in\Gamma_i$, and $ {{|\!|\!|}}{{\mathbf B}}_{\alpha i}u{{|\!|\!|}}_{H_a^{k+3/2}(\Gamma_i)}\le c{{|\!|\!|}}u{{|\!|\!|}}_{H_a^{k+2}(G\setminus\overline{{{\mathcal O}}_{\varkappa_1}({{\mathcal K}})})}, $
2. ${{\rm spt\,}}\alpha_i(y,\cdot)\subset\overline{G_\sigma}$, $y\in\Gamma_i\setminus{{\mathcal O}}_{\varkappa_2}({{\mathcal K}})$, and $ {{|\!|\!|}}{{\mathbf B}}_{\alpha
i}u{{|\!|\!|}}_{H_a^{k+3/2}(\Gamma_i\setminus\overline{{{\mathcal O}}_{\varkappa_2}({{\mathcal K}})})}\le
c{{|\!|\!|}}u {{|\!|\!|}}_{H_a^{k+2}(G_\sigma)}, $
where $G_\sigma=\{y\in G:{{\rm dist}}(y,{{\partial G}})<\sigma\}.$
Let $
{{\mathcal N}}=\bigcup_{i=1}^N\{y\in\Gamma_i: \mu_i(y,{{\overline G}})=0\}\cup{{\mathcal K}}$ and ${{\mathcal M}}={{\partial G}}\setminus {{\mathcal N}}.$ Assume that ${{\mathcal N}}$ and ${{\mathcal M}}$ are Borel sets.
\[cond2.4\] $\beta_i(y,{{\mathcal M}})<1$ for $y\in\Gamma_i\cap{{\mathcal M}}$, $i=1,\dots, N$.
\[cond2.5\] For any function $u\in C_{{\mathcal N}}(\overline G)$, the functions ${{\mathbf B}}_{\alpha i}u$ and ${{\mathbf B}}_{\beta i}u$ can be extended to $\overline{\Gamma_i}$ in such a way that the extended functions [(]{}which we also denote by ${{\mathbf B}}_{\alpha i}u$ and ${{\mathbf B}}_{\beta
i}u$, respectively[)]{} belong to $C_{{\mathcal N}}(\overline{\Gamma_i})$.
We represent the measures $\beta_i(y,\cdot)$ in the form $
\beta_i(y,\cdot)=\beta_i^1(y,\cdot)+\beta_i^2(y,\cdot), $ where $\beta_i^1(y,\cdot)$ and $\beta_i^2(y,\cdot)$ are nonnegative Borel measures. Let us specify them. Set ${{\mathcal M}}_p={{\mathcal O}}_p(\overline{{{\mathcal M}}})$, $p>0$. Consider a cut-off function $\hat\zeta_p\in C^\infty({{\mathbb R}}^2)$ such that $0\le\hat\zeta_p(y)\le
1$, $\hat\zeta_p(y)=1$ for $y\in{{\mathcal M}}_{p/2}$, and $\hat\zeta_p(y)=0$ for $y\notin{{\mathcal M}}_{p}$. Set $\tilde\zeta_p=1-\hat\zeta_p$. Introduce the operators $$\hat{{\mathbf B}}_{\beta i}^1
u(y)=\int_{{{\overline G}}}\hat\zeta_p(\eta)u(\eta)\beta_i^1(y,d\eta),\qquad
\tilde{{\mathbf B}}_{\beta i}^1 u(y)=\int_{{{\overline G}}}\tilde\zeta_p(\eta)
u(\eta)\beta_i^1(y,d\eta),$$ $${{\mathbf B}}_{\beta i}^2
u(y)=\int_{{{\overline G}}}u(\eta)\beta_i^2(y,d\eta).$$
Consider the Banach spaces with the norms depending on the parameter $q>0$:
1. $H_{{{\mathcal N}},a}^{k+2}(G)=C_{{\mathcal N}}({{\overline G}})\cap
H_a^{k+2}(G)$, $ {{|\!|\!|}}u
{{|\!|\!|}}_{H_{{{\mathcal N}},a}^{k+2}(G)}=\|u\|_{C_{{\mathcal N}}({{\overline G}})}+{{|\!|\!|}}u {{|\!|\!|}}_{H_a^{k+2}(G)},
$
2. $H_{{{\mathcal N}},a}^{k+3/2}(\Gamma_i)=C_{{\mathcal N}}(\overline{\Gamma_i})\cap
H_a^{k+3/2}(\Gamma_i)$, $ {{|\!|\!|}}v
{{|\!|\!|}}_{H_{{{\mathcal N}},a}^{k+3/2}(\Gamma_i)}=\|v\|_{C_{{\mathcal N}}(\overline{\Gamma_i})}+{{|\!|\!|}}v {{|\!|\!|}}_{H_a^{k+3/2}(\Gamma_i)}.$
\[cond3.3\] For each sufficiently small $p>0$ the following assertions are true[:]{}
1. ${{|\!|\!|}}\hat{{\mathbf B}}_{\beta i}^1{{|\!|\!|}}_{H_{{{\mathcal N}},a}^{k+2}(G)\to
H_{{{\mathcal N}},a}^{k+3/2}(\Gamma_i)}\to 0,\ p\to0,$ uniformly with respect to $q;$
2. the norms ${{|\!|\!|}}\tilde{{\mathbf B}}_{\beta i}^1{{|\!|\!|}}_{H_{{{\mathcal N}},a}^{k+2}(G)\to
H_{{{\mathcal N}},a}^{k+3/2}(\Gamma_i)}$ are bounded uniformly with respect to $q$.
\[cond3.4\] The operators ${{\mathbf B}}_{\beta i}^2:H_{{{\mathcal N}},a}^{k+2}(G)\to
H_{{{\mathcal N}},a}^{k+3/2}(\Gamma_i)$, $i=1,\dots,N$, are compact.
Introduce the space $ C_B({{\overline G}})=\{u\in C({{\overline G}}): u\ \text{satisfy
nonlocal conditions \eqref{eq57}}\}. $ We consider the unbounded operator ${{\mathbf P}}: {{\rm D}}({{\mathbf P}})\subset C_B(\overline G)\to
C_B(\overline G)$ given by $${{\mathbf P}}u=P_0u+P_1u,\qquad u\in {{\rm D}}({{\mathbf P}})=\{u\in C_B({{\overline G}})\cap
H_{a}^{k+2}(G): P_0u+P_1u\in C_B(\overline G)\}.$$ Note that ${{\rm D}}({{\mathbf P}})\subset C^2(G)\cap C_B({{\overline G}})$ due to the relation $k\ge 2$ and the Sobolev embedding theorem. The proof of the following main result is based on Theorem \[th\], the Hille–Iosida theorem, and the maximum principle.
\[th3.1\] Let Conditions $\ref{condK1}$–$\ref{cond3.4}$ hold. Then the operator ${{\mathbf P}}$ admits the closure $\overline{{{\mathbf P}}}:{{\rm D}}(\overline{{{\mathbf P}}})\subset C_B({{\overline G}})\to C_B({{\overline G}})$ and the operator $\overline{{{\mathbf P}}}$ is a generator of a Feller semigroup.
[**6.**]{} Consider an example of nonlocal conditions in which the assumptions of the paper hold. Let ${{\partial G}}=\Gamma_1\cup\Gamma_2\cup{{\mathcal K}}$, where $\Gamma_1$ and $\Gamma_2$ are $C^\infty$ curves open and connected in the topology of ${{\partial G}}$, $\Gamma_1\cap\Gamma_2=\varnothing$, and $\overline{\Gamma_1}\cap\overline{\Gamma_2}={{\mathcal K}}$; the set ${{\mathcal K}}$ consists of two points $g_1$ and $g_2$. Let $\Omega_j$, $j=1,\dots,4$, be nondegenerate transformations of class $C^{k+2}$ defined on a neighborhood of $\overline{\Gamma_1}$ and satisfying the following conditions (see Fig. \[figEx4-4\]):
[ ]{}
1. $\Omega_1({{\mathcal K}})={{\mathcal K}}$, $\Omega_1(\Gamma_1\cap{{\mathcal O}}_\varepsilon({{\mathcal K}}))\subset G$, $\Omega_1(\Gamma_1\setminus{{\mathcal O}}_\varepsilon({{\mathcal K}}))\subset
G\cup\Gamma_2$, and $\Omega_1(y)$ is a composition of a shift of the argument, rotation, and homothety for $y\in\overline{\Gamma_1}\cap{{\mathcal O}}_\varepsilon({{\mathcal K}})$;
2. there exist numbers $\varkappa_1>\varkappa_2>0$ and $\sigma>0$ such that $\Omega_2(\overline{\Gamma_1})\subset{{\overline G}}\setminus{{\mathcal O}}_{\varkappa_1}({{\mathcal K}})$ and $\Omega_2(\overline{\Gamma_1}\setminus{{\mathcal O}}_{\varkappa_2}({{\mathcal K}}))\subset
\overline{G_\sigma}$; moreover, $\Omega_2(g_1)\in\Gamma_1$ and $\Omega_2(g_2)\in G$;
3. $\Omega_3(\overline{\Gamma_1})\subset G\cup{\Gamma_2}$ and $\Omega_3({{\mathcal K}})\subset\Gamma_2$;
4. $\Omega_4(\overline{\Gamma_1})\subset G\cup\overline{\Gamma_2}$ and $\Omega_4({{\mathcal K}})={{\mathcal K}}$; the angle between the rays tangent to $\Gamma_1$ and $\Omega_4(\Gamma_1)$ at the point $g_j$ is nonzero.
Let $b_j\in C^{k+2}(\overline{\Gamma_1})$ and $b_j\ge0$, $j=1,\dots,4$.
Consider the following nonlocal conditions: $$\begin{aligned}
u(y)-\sum_{j=1}^4 b_j(y)u(\Omega_j(y)) &=0, &
&y\in\Gamma_1,\\
u(y) &=0, & &y\in\overline{\Gamma_2}.
\end{aligned}$$
We assume that $$\begin{gathered}
\sum_{j=1}^4 b_j(y)\le 1,\ y\in\Gamma_1;\\
b_2(g_1)=0\
\text{or}\ \sum_{j=1}^4 b_j(\Omega_2(g_1))=0;\quad b_2(g_2)=0;\quad b_4(g_j)=0.
\end{gathered}$$
Introduce a cut-off function $\zeta\in C^\infty({{\mathbb R}}^2)$ supported in ${{\mathcal O}}_\varepsilon({{\mathcal K}})$, equal to $1$ on ${{\mathcal O}}_{\varepsilon/2}({{\mathcal K}})$, and such that $0\le\zeta(y)\le 1$, $y\in{{\mathbb R}}^2$. Let $y\in {\Gamma_1}$, and let $Q\subset{{\overline G}}$ be an arbitrary Borel set; then the measures $$\begin{gathered}
\delta(y,Q)=\zeta(y)b_1(y)\chi_Q(\Omega_1(y)),\qquad
\alpha(y,Q)=b_2(y)\chi_Q(\Omega_2(y)),\\
\beta^1(y,Q)=\big(1-\zeta(y)\big)b_1(y)\chi_Q(\Omega_1(y))+\sum_{j=3,4}
b_j(y)\chi_Q(\Omega_j(y)),\qquad
\beta^2(y,Q)=0
\end{gathered}$$ (for simplicity, we have omitted the subscript “1” in the notation of the measures) satisfy Conditions \[condK1\], \[cond1.2\], and \[cond2.3\]–\[cond3.4\].
The author is grateful to Prof. A.L. Skubachevskii for attention to this work.
[99]{}
, [*Trans. Amer. Math. Soc.*]{}, [**77**]{}, 1–30 (1954).
A. D. Ventsel, [*Teor. Veroyatnost. i Primen., **4***]{}, 172–185 (1959); English transl.: [*Theory Probab. Appl.*]{}, [**4**]{} (1959).
K. Sato, T. Ueno, [*J. Math. Kyoto Univ.*]{} [**4**]{}, 529–605 (1965).
J. M. Bony, P. Courrege, P. Priouret, [*Ann. Inst. Fourier [(]{}Grenoble[)]{}*]{} [**18**]{}, (1968) 369–521.
K. Taira, Boundary Value Problems, Feller Semigroups, and Markov Processes, [*in*]{} “Interaction between Functional Analysis, Harmonic Analysis, and Probability”, Vol. 40, Marcel Dekker, New York–Basel–Hong Kong, 1996.
Y. Ishikawa, [*J. Math. Soc. Japan*]{}, [**42**]{}, 171–184 (1990).
E. I. Galakhov, A. L. Skubachevskii, [*J. Differ. Equ.*]{}, [**176**]{}, 315–355 (2001).
A. L. Skubachevskii, [*Dokl. Akad. Nauk SSSR*]{}, [**307**]{}, 287–291 (1989); English transl. in [*Soviet Math. Dokl.*]{}, [**40**]{} (1990).
M. G. Garroni, J. L. Menaldi, [*Second Order Elliptic Integro-Differential Problems*]{}, Chapman $\&$ Hall/CRC, London–New York–Washington, D.C., 2002.
[^1]: This research was supported by Russian Foundation for Basic Research (project No. 07-01-00268) and the Alexander von Humboldt Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we propose corrections to the likelihood ratio test and John’s test for sphericity in large-dimensions. New formulas for the limiting parameters in the CLT for linear spectral statistics of sample covariance matrices with general fourth moments are first established. Using these formulas, we derive the asymptotic distribution of the two proposed test statistics under the null. These asymptotics are valid for general population, i.e. not necessarily Gaussian, provided a finite fourth-moment. Extensive Monte-Carlo experiments are conducted to assess the quality of these tests with a comparison to several existing methods from the literature. Moreover, we also obtain their asymptotic power functions under the alternative of a spiked population model as a specific alternative.'
address:
- |
Qinwen Wang\
Department of Mathematics\
Zhejiang University\
- |
Jianfeng Yao\
Department of Statistics and Actuarial Science\
The University of Hong Kong\
Pokfulam, Hong Kong\
author:
-
-
title: 'On the sphericity test with large-dimensional observations'
---
Introduction
============
Consider a sample $Y_1,\ldots,Y_n$ from a $p$-dimensional multivariate distribution with covariance matrix $\Sigma_p$. An important problem in multivariate analysis is to test the sphericity, namely the hypothesis $H_{0}: \Sigma_p = \sigma^{2}I_p$ where $\sigma^{2}$ is [*unspecified*]{}. If the observations represent a multivariate error with $p$ components, the null hypothesis expresses the fact that the error is cross-sectionally uncorrelated (independent if in addition they are normal) and have a same variance (homoscedasticity).
Much of the existing theory about this test has been exposed first in details in [@r17] about Gaussian likelihood ratio test and later in [@r10], [@r11], [@r27] and also in textbooks like [@r18 Chapter 8] and [@r1 Chapter 10]. Assume for a moment that the sample has a normal distribution with mean zero and covariance matrix $\Sigma_p$. Let $S_n =n^{-1} \sum_i Y_iY_i^*$ be the sample covariance matrix and denote its eigenvalues by $\{\ell_i\}_{1\le i\le p}$. Two well established procedures for testing the sphericity are the likelihood ratio test ([*[LRT]{}*]{}) and a test devised in [@r10]. The likelihood ratio statistic is, see e.g. [@r1 §10.7.2], $$\label{Ln}
L_n = \left( \frac{ (\ell_1\cdots \ell_p)^{1/p}} {\frac1p (\ell_1+\cdots+\ell_p)}
\right)^{\frac12pn}~,$$ which is a power of the ratio of the geometric mean of the sample eigenvalues to the arithmetic mean. It is here noticed that in this formula it is necessary to assume that $p\le n$ to avoid null eigenvalues in (the numerator of) $L_n$. If we let $n\to\infty$ while keeping $p$ fixed, classical asymptotic theory indicates that under the null hypothesis, $-2\log L_n \Longrightarrow \chi^{2}_f $, a chi-square distribution with degree of freedom $f= \frac{1}{2}p(p+1)-1$. This asymptotic distribution is further refined by the following Box-Bartlett correction (referred as [*[BBLRT]{}*]{}): $$\begin{aligned}
\label{exp-Ln}
P(-2\rho\log L_n\le x) = P_f ( x) +
\omega_2\left\{ P_{f+4} (x ) - P_f(x) \right\} +O(n^{-3})~,\end{aligned}$$ where $P_k(x)= P(\chi^2_k \le x)$ and $$\rho= 1- \frac{2p^2+p+2}{6pn},\qquad
\omega_2=\frac{ (p+2)(p-1)(p-2)(2p^3+6p^2+3p+2)}{288p^2n^2\rho^2}~.$$
By observing that the asymptotic variance of $-2\log L_n$ is proportional to ${\mathop{\text{tr}}}\{
\Sigma({\mathop{\text{tr}}}\Sigma)^{-1}-p^{-1}I_p \}^2$, [@r10] proposed to use the statistic $$T_2 = \frac{ p^2n}{2} {\mathop{\text{tr}}}\left\{ S_n({\mathop{\text{tr}}}S_n)^{-1} - p^{-1}I_p
\right\}^2~
\label{Nagao}$$ for testing sphericity. [[When $p$ is fixed and $n\to\infty$, under the null hypothesis, it also holds that $T_2 \Longrightarrow \chi^{2}_f $, which we referred to as *[John’s test]{}.*]{}]{} It is observed that $T_2$ is proportional to the square of the coefficient of variation of the sample eigenvalues, namely $$T_2 = \frac{np}{2} \cdot
\frac{ p^{-1} \sum (\ell_i -\overline \ell)^2}{{\overline \ell}^2}~,
\qquad \text{with~~} \overline \ell = \frac1n\sum_i \ell_i~.$$ Following the idea of the Box-Bartlett correction, [@r19] established an expansion for the distribution function of the statistics $T_2$ (referred as [*[Nagao’s test]{}*]{}), $$\begin{aligned}
P(T_2\le x) &=& P_f(x) + \frac1n \left\{a_p P_{f+6} (x) + b_p P_{f+4}(x)+c_p
P_{f+2}(x) +d_p P_f(x)
\right\}\nonumber \\
&&+ O(n^{-2}),
\label{exp-Nagao}\end{aligned}$$ where $$\begin{aligned}
&& a_p = \frac1{12}(p^3+3p^2-12-200p^{-1}),~
b_p =
\frac1{8}(-2p^3-5p^2+7p-12-420p^{-1})~,
\\
&&
c_p = \frac1{4}(p^3+2p^2-p-2-216p^{-1})~,
d_p = \frac1{24}(-2p^3-3p^2+p+436p^{-1})~.\end{aligned}$$
It has been well known that classical multivariate procedures are in general challenged by large-dimensional data. A small simulation experiment is conducted to explore the performance of the BBLRT and Nagao’s test (two corrections) with growing dimension $p$. The sample size is set to $n=64$ while dimension $p$ increases from 4 to 60 (we have also run other experiments with larger sample sizes $n$ but conclusions are very similar), and the nominal level is set to be $\alpha=0.05$. The samples come from normal vectors with mean zero and identity covariance matrix, and each pair of $(p,n)$ is assessed with 10000 independent replications.
Table \[tradsize\] gives the empirical sizes of BBLRT and Nagao’s test. It is found here that when the dimension to sample size ratio $p/n$ is below $1/2$, both tests have an empirical size close to the nominal test level $0.05$. Then when the ratio grows up, the BBLRT becomes quickly biased while Nagao’s test still has a correct empirical size. It is striking that although Nagao’s test is derived under classical “$p$ fixed, $n\rightarrow \infty$” regime, it is remarkably robust against dimension inflation.
$(p,n)$ (4,64) (8,64) (16,64) (32,64) (48,64) (56,64) (60,64)
-------------- -------- -------- --------- --------- --------- --------- ---------
BBLRT 0.0483 0.0523 0.0491 0.0554 0.1262 0.3989 0.7605
Nagao’s test 0.0485 0.0495 0.0478 0.0518 0.0518 0.0513 0.0495
: \[tradsize\] Empirical sizes of BBLRT and Nagao’s test at $5\%$ significance level based on $10000$ independent replications using normal vectors $\mathcal{N}(0,I_p)$ for $n=64$ and different values of $p$.
Therefore, the goal of this paper is to propose novel corrections to both LRT and John’s test to cope with the large-dimensional context. Similar works have already been done in [@r15], which confirms the robustness of John’s test in large-dimensions; however, these results assume a Gaussian population. In this paper, we remove such a Gaussian restriction, and prove that the robustness of John’s test is in fact general. Following the idea of [@r15], [@r9] proposed to use a family of well selected U-statistics to test the sphericity; however, as showed in our simulation study in Section \[sec:simul\], the powers of our corrected John’s test are slightly higher than this test in most cases. More recently, [@r26] examined the performance of $T_1$ (a statistic first put forward in [@r24]) under non-normality, but with the moment condition $\gamma=3+O(p^{-\epsilon})$, which essentially matches the Gaussian case ($\gamma=3$) asymptotically. We have also removed this moment restriction in our setting. In short, we have unveiled two corrections that have a better performance and removed the Gaussian or nearly Gaussian restriction found in the existing literature.
From the technical point of view, our approach differs from [@r15] and follows the one devised in [@r4] and [@r6]. The central tool is a CLT for linear spectral statistics of sample covariance matrices established in [@r2] and later refined in [@r21]. The paper also contains an original contribution on this CLT reported in the Appendix: new formulas for the limiting parameters in the CLT. Since such CLT’s are increasingly important in large-dimensional statistics, we believe that these new formulas will be of independent interest for applications other than those considered in this paper.
The remaining of the paper is organized as follows. Large-dimensional corrections to LRT and John’s test are introduced in Section \[sec:corrected\]. Section \[sec:simul\] reports a detailed Monte-Carlo study to analyze finite-sample sizes and powers of these two corrections under both normal and non-normal distributed data. Next, Section \[powerspike\] gives the theoretical analysis of their asymptotic power under the alternative of a spiked population model. Section \[ge\] generalizes our test procedures to populations with an unknown mean. Technical proofs and calculations are relegated to Section \[sec:proofs\]. The last Section contains some concluding remarks.
Large-dimensional corrections {#sec:corrected}
=============================
From now on, we assume that the observations $Y_1,\ldots,Y_n$ have the representation $Y_j=\Sigma_p^{1/2} X_j$ where the $p\times n $ table $\{X_1,\ldots,X_n\}=\{x_{ij}\}_{1 \le i\le p,1\le j\le n}$ are made with an array of i.i.d. standardized random variables (mean 0 and variance 1). This setting is motivated by the random matrix theory and it is generic enough for a precise analysis of the sphericity test. Furthermore, under the null hypothesis $H_0: \Sigma_p=\sigma^2I_p$ ($\sigma^2$ is unspecified), we notice that both LRT and John’s test are independent from the scale parameter $\sigma^2$ under the null. Therefore, we can assume w.l.o.g. $\sigma^2=1$ when dealing with the null distributions of these test statistics. This will be assumed in all the sections.
Throughout the paper we will use an indicator $\kappa$ set to 2 when $\{x_{ij}\}$ are real and to 1 when they are complex as defined in [@r3]. Also, we define the kurtosis coefficient $\beta=E|x_{ij}|^{4}-1-\kappa$ for both cases and note that for normal variables, $\beta=0$ (recall that for a standard complex-valued normal random variable, its real and imaginary parts are two iid. $N(0, \frac12)$ real random variables).
The corrected likelihood ratio test (CLRT) {#ssec: CLRT}
------------------------------------------
For the correction of LRT, let $\mathcal L_n=-2n^{-1}\log L_n~$ be the test statistic for $n\ge 1$. Our first main result is the following.
\[clt\] Assume $\{x_{ij}\}$ are iid, satisfying $Ex_{ij}=0, E|x_{ij}|^{2}=1,
E|x_{ij}|^{4}< \infty$. Then under $H_{0}$ and when $\frac{p}{n}=y_{n}\rightarrow y\in (0,1)$, $$\begin{aligned}
\label{a1}
&&\mathcal L_n + (p-n)\cdot\log(1-\frac{p}{n})-p\nonumber\\
&&\Longrightarrow
N\{-\frac{\kappa-1}{2}\log(1-y) + \frac12\beta y,
-\kappa\log(1-y)-\kappa y \}~.
\end{aligned}$$
The test based on this asymptotic normal distribution will be hereafter referred as the [*corrected likelihood-ratio test*]{} ([*[CLRT]{}*]{}). One may observe that the limiting distribution of the test crucially depends on the limiting dimension-to-sample ratio $y$ through the factor $-\log(1-y)$. In particular, the asymptotic variance will blow up quickly when $y$ approaches 1, so it is expected that the power will seriously break down. Monte-Carlo experiments in Section \[sec:simul\] will provide more details on this behavior.
The proof of Theorem \[clt\] is based on the following lemma. In all the following, $F^y$ denotes the Marčenko-Pastur distribution of index $y$ ($>0$) which is introduced and discussed in the Appendix. [[And $F^{y}(f(x))=\int f(x) F^y(dx)$ denotes the integral of function $f(x)$ with respect to $F^y$.]{}]{}
\[lem:joint\] Let $\{\ell_i\}_{1\le i\le p} $ be the eigenvalues of the sample covariance matrix $S_{n}=n^{-1}\sum_i
Y_iY_i^*$. Then under $H_{0}$ and the conditions of Theorem \[clt\], we have $$\left(
\begin{array}{cc}
\sum_{i=1}^{p}\log \ell_i -pF^{y_{n}}(\log x)
\\
\sum_{i=1}^{p}\ell_i -pF^{y_{n}}(x)
\\
\end{array}
\right) \Longrightarrow
N(\mu_1,V_1),$$ with $$\mu_1=
\begin{pmatrix}
\frac{\kappa-1}{2}\log (1-y) - \frac12 \beta y
\\
0
\end{pmatrix} ~,$$ and $$V_1 =
\begin{pmatrix}
-\kappa\log(1-y)+ \beta y
&
(\beta+\kappa)y
\\
(\beta+\kappa)y
& (\beta+\kappa)y
\end{pmatrix} ~.$$
The proof of this lemma is postponed to Section \[sec:proofs\].
(of Theorem \[clt\]). Let $A_{n}\equiv \sum_{i=1}^{p}\log \ell_i -pF^{y_{n}}(\log x)$ and $B_{n}\equiv \sum_{i=1}^{p}\ell_i -pF^{y_{n}}(x)$. By Lemma \[lem:joint\], $$\left(
\begin{array}{cc}
A_{n}
\\
B_{n}
\\
\end{array}
\right) \Longrightarrow
N(\mu_1,V_1).$$ Consequently, $-A_n+B_n$ is asymptotically normal with mean $-\frac{\kappa-1}{2}\log(1-y)+\frac{1}{2}\beta y$ and variance $$V_1(1,1)+V_1(2,2)-2V_1(1,2)=-\kappa\log(1-y)-\kappa y.$$ Besides, $$\begin{aligned}
\mathcal L_n&=&-\Sigma_{i=1}^{p}\log \ell_i +p\log(\frac{1}{p}\Sigma_{i=1}^{p}\ell_i )\\
&=&-(A_{n}+pF^{y_{n}}(\log x))+p\log (\frac{1}{p}(B_{n}+p))\\
&=&-A_{n}-pF^{y_{n}}(\log x)+p\log (1+\frac{B_{n}}{p}).\\\end{aligned}$$ Since $ B_{n} \Rightarrow N(0,y(\beta+\kappa))$, $B_{n}=O_{p}(1)$ and $ \log(1+{B_{n}}/{p})={B_{n}}/{p}+O_{p}({1}/{p^{2}}).$ Therefore, $$\mathcal L_n =-A_n-pF^{y_{n}}(\log x)+B_n+O_{p}(\frac{1}{p})~.$$ The conclusion follows with the following well-known integrals w.r.t. the Marčenko-Pastur distribution $F^y$ ($y<1$) [[in [@r2]]{}]{}, $$F^{y}(\log x)=\frac{y-1}{y}\log(1-y)-1~, \quad
F^{y}(x)=1.$$ The proof of Theorem \[clt\] is complete.
The corrected John’s test (CJ) {#ssec:CN}
------------------------------
Earlier than the asymptotic expansion given in [@r19], [@r10] proved that when the observations are normal, the sphericity test based on $T_2$ is a locally most powerful invariant test. It is also established in [@r11] that under these conditions, the limiting distribution of $T_2$ under $H_0$ is $\chi^2_f$ with degree of freedom $f=\frac12p(p+1)-1 $, or equivalently, $$\begin{aligned}
n U-p\Longrightarrow \frac{2}{p}\chi^2_f
-p~,\end{aligned}$$ where for convenience, we have let $U=2(np)^{-1}T_2$. Clearly, this limit has been established for $n\to\infty$ and a fixed dimension $p$. However, if we now let $p\to\infty$ in the right-hand side of the above result, it is not hard to see that $\frac{2}{p}\chi^2_f -p$ will tend to the normal distribution $N(1,4)$. It then seems “natural” to conjecture that when both $p$ and $n$ grow to infinity in some “proper” way, it may happen that $$\begin{aligned}
\label{LW}
n U-p\Longrightarrow N(1,4)~.\end{aligned}$$ This is indeed the main result of [@r15] where this asymptotic distribution was established assuming that data are normal-distributed and $p$ and $n$ grow to infinity in a proportional way (i.e. $p/n\to y>0$).
In this section, we provide a more general result using our own approach. In particular, the distribution of the observation is arbitrary provided a finite fourth moment exists.
\[cnclt\] Assume $\{x_{ij}\}$ are iid, satisfying $Ex_{ij}=0, E|x_{ij}|^{2}=1,
E|x_{ij}|^{4}< \infty$, and let $U=2(np)^{-1}T_2$ be the test statistic. Then under $H_{0}$ and when $p\rightarrow \infty, n\rightarrow \infty,
\frac{p}{n}=y_{n}\rightarrow y\in (0,\infty)$, $$\begin{aligned}
\label{a2}
n U-p\Longrightarrow N(\kappa+\beta-1,2\kappa)~.\end{aligned}$$
The test based on the asymptotic normal distribution given in equation will be hereafter referred as the [*corrected John’s test (CJ)*]{}.
A striking fact in this theorem is that as in the normal case, the limiting distribution of CJ is [*independent*]{} of the dimension-to-sample ratio $y=\lim p/n$. In particular, the limiting distribution derived under classical scheme ($p$ fixed, $n\rightarrow\infty$), e.g. the distribution $\frac2p \chi_{f}^2-p$ in the normal case, when used for large $p$, stays very close to this limiting distribution derived for large-dimensional scheme ($p\rightarrow\infty, n\rightarrow\infty, p/n\rightarrow y \in (0,\infty)$). In this sense, Theorem \[cnclt\] gives a theoretic explanation to the widely observed robustness of John’s test against the dimension inflation. Moreover, CJ is also valid for the $p$ larger (or much larger) than $n$ case in contrast to the CLRT where this ratio should be kept smaller than 1 to avoid null eigenvalues.
It is also worth noticing that for real normal data, we have $\kappa=2$ and $\beta=0$ so that the theorem above reduces to $n U-p\Rightarrow
N(1,4)$. This is exactly the result discussed in [@r15]. Besides, if the data has a non-normal distribution but has the same first four moments as the normal distribution, we have again $n U-p\Rightarrow N(1,4)$, which turns out to have a universality property.
The proof of Theorem \[cnclt\] is based on the following lemma.
\[cnlemma:joint\] Let $\{\ell_i\}_{1\le i\le p} $ be the eigenvalues of the sample covariance matrix $S_{n}=n^{-1}\sum_i
Y_iY_i^*$. Then under $H_{0}$ and the conditions of Theorem \[cnclt\], we have $$\left(
\begin{array}{cc}
\sum_{i=1}^{p}\ell_i ^2-p(1+y_n)
\\
\sum_{i=1}^{p}\ell_i -p
\\
\end{array}
\right) \Longrightarrow
N(\mu_2,V_2),$$ with $$\mu_2=
\begin{pmatrix}
(\kappa-1+\beta)y
\\
0
\end{pmatrix} ~,$$ and $$V_2 =
\begin{pmatrix}
2\kappa y^2+4(\kappa+\beta)(y+2y^2+y^3)
&
2(\kappa+\beta)(y+y^2)
\\
2(\kappa+\beta)(y+y^2)
& (\kappa+\beta)y
\end{pmatrix} ~.$$
The proof of this lemma is postponed to Section \[sec:proofs\].
(of Theorem \[cnclt\]).The result of Lemma \[cnlemma:joint\] can be rewritten as:\
$$n
\left(
\begin{array}{cc}
p^{-1}\sum_{i=1}^p \ell_i ^2-1-\frac{p}{n}-\frac{(\kappa+\beta-1)y}{p} \\
p^{-1}\sum_{i=1}^p \ell_i -1
\end{array}
\right)
\Longrightarrow
N\Bigg(
\left(
\begin{array}{cc}
0 \\ 0
\end{array}
\right), \
\frac{1}{y^2}\cdot V_2
\Bigg).$$ Define the function $f(x,y)=\frac{x}{y^2}-1$, then $U=f(p^{-1}\Sigma_{i=1}^{p}\ell_i ^2,~p^{-1}\Sigma_{i=1}^{p}\ell_i )$.
We have $$\begin{aligned}
&&\frac{\partial f}{\partial x}(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}, 1)=1~,\\
&&\frac{\partial f}{\partial y}(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}, 1)=-2(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p})~,\\
&&f(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}, 1)=\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}~.\\\end{aligned}$$ By the delta method, $$n\Big(U-f(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}, 1)\Big)\Longrightarrow
N(0,\lim C),$$ where\
$C=
$$
\left(
\begin{array}{cc}
\frac{\partial f}{\partial x}(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}, 1)
\\
\frac{\partial f}{\partial y}(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}, 1)
\\
\end{array}
\right)^T \
$$\cdot
\Big(\frac{1}{y^2} V_2 \Big)\cdot
$$
\left(
\begin{array}{cc}
\frac{\partial f}{\partial x}(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}, 1)
\\
\frac{\partial f}{\partial y}(1+\frac{p}{n}+\frac{(\kappa+\beta-1)y}{p}, 1)
\\
\end{array}
\right) \
$$
$\
$\longrightarrow 2\kappa~.$\
Therefore, $$\begin{aligned}
n(U- \frac{p}{n}-\frac{(\kappa+\beta-1)y}{p})\Longrightarrow N(0, 2\kappa)~,\\\end{aligned}$$ that is, $$\begin{aligned}
n U-p\Longrightarrow N(\kappa+\beta-1,2\kappa)~.\\
\end{aligned}$$ The proof of Theorem \[cnclt\] is complete.
Note that in [[Theorems]{}]{} \[clt\] and \[cnclt\] appears the parameter $\beta$, which is in practice unknown with real data. So we may estimate the parameter $\beta$ using the fourth-order sample moment: $$\widehat\beta= \frac{1}{np} \sum_{i=1}^p\sum_{j=1}^n|Y_j(i)|^4
- \kappa -1~.$$ According to the law of large numbers, $\hat{\beta}=\beta+o_p(1)$, so substituting $\hat{\beta}$ for $\beta$ in [[Theorems]{}]{} \[clt\] and \[cnclt\] does not modify the limiting distribution.
Monte Carlo study {#sec:simul}
=================
Monte Carlo simulations are conducted to find empirical sizes and powers of CLRT and CJ. In particular, here we want to examine the following questions: how robust are the tests against non-normal distributed data and what is the range of the dimension to sample ratio $p/n$ where the tests are applicable.
For comparison, we show both the performance of the LW test using the asymptotic $N(1,4)$ distribution in (Notice however this is the CJ test under normal distribution) and the Chen’s test (denoted as [*[C]{}*]{} for short) using the asymptotic $N(0,4)$ distribution derived in [@r9]. The nominal test level is set to be $\alpha=0.05$, and for each pair of $(p,n)$, we run 10000 independent replications.
We consider two scenarios with respect to the random vectors $Y_i$ :
1. $Y_i$ is $p$-dimensional real random vector from the multivariate normal population ${\mathcal{N}}(0, I_p)$. In this case, $\kappa=2$ and $\beta=0$.
2. $Y_i$ consists of iid real random variables with distribution $Gamma(4,2)-2$ so that $y_{ij}$ satisfies $E y_{ij}=0, E y_{ij}^4=4.5$. In this case, $\kappa=2$ and $\beta=1.5$.
----------- -- -------- -------- -------- -- -- -------- -------- -------- --------
LW/CJ CLRT C LW CLRT CJ C
(4,64) 0.0498 0.0553 0.0523 0.1396 0.074 0.0698 0.0717
(8,64) 0.0545 0.061 0.0572 0.1757 0.0721 0.0804 0.078
(16,64) 0.0539 0.0547 0.0577 0.1854 0.0614 0.078 0.0756
(32,64) 0.0558 0.0531 0.0612 0.1943 0.0564 0.0703 0.0682
(48,64) 0.0551 0.0522 0.0602 0.1956 0.0568 0.0685 0.0652
(56,64) 0.0547 0.0505 0.0596 0.1942 0.0549 0.0615 0.0603
(60,64) 0.0523 0.0587 0.0585 0.194 0.0582 0.0615 0.0603
(8,128) 0.0539 0.0546 0.0569 0.1732 0.0701 0.075 0.0754
(16,128) 0.0523 0.0534 0.0548 0.1859 0.0673 0.0724 0.0694
(32,128) 0.051 0.0545 0.0523 0.1951 0.0615 0.0695 0.0693
(64,128) 0.0538 0.0528 0.0552 0.1867 0.0485 0.0603 0.0597
(96,128) 0.055 0.0568 0.0581 0.1892 0.0539 0.0577 0.0579
(112,128) 0.0543 0.0522 0.0591 0.1875 0.0534 0.0591 0.0593
(120,128) 0.0545 0.0541 0.0561 0.1849 0.051 0.0598 0.0596
(16,256) 0.0544 0.055 0.0574 0.1898 0.0694 0.0719 0.0716
(32,256) 0.0534 0.0515 0.0553 0.1865 0.0574 0.0634 0.0614
(64,256) 0.0519 0.0537 0.0522 0.1869 0.0534 0.0598 0.0608
(128,256) 0.0507 0.0505 0.0498 0.1858 0.051 0.0555 0.0552
(192,256) 0.0507 0.054 0.0518 0.1862 0.0464 0.052 0.0535
(224,256) 0.0503 0.0541 0.0516 0.1837 0.0469 0.0541 0.0538
(240,256) 0.0494 0.053 0.0521 0.1831 0.049 0.0533 0.0559
(32,512) 0.0542 0.0543 0.0554 0.1884 0.0571 0.0606 0.059
(64,512) 0.0512 0.0497 0.0513 0.1816 0.0567 0.0579 0.0557
(128,512) 0.0519 0.0567 0.0533 0.1832 0.0491 0.0507 0.0504
(256,512) 0.0491 0.0503 0.0501 0.1801 0.0504 0.0495 0.0492
(384,512) 0.0487 0.0505 0.0499 0.1826 0.051 0.0502 0.0507
(448,512) 0.0496 0.0495 0.0503 0.1881 0.0526 0.0482 0.0485
(480,512) 0.0488 0.0511 0.0505 0.1801 0.0523 0.053 0.0516
----------- -- -------- -------- -------- -- -- -------- -------- -------- --------
: \[SIZE\]Empirical sizes of LW, CJ, CLRT and C test at $5\%$ significance level based on 10000 independent applications with real $N(0,1)$ random variables and with real Gamma(4,2)-2 random variables.
Table \[SIZE\] reports the sizes of the four tests in these two scenarios for different values of $(p, n)$. We see that when $\{y_{ij}\}$ are normal, LW (=CJ), CLRT and C all have similar empirical sizes tending to the nominal level 0.05 as either $p$ or $n$ increases. But when $\{y_{ij}\}$ are Gamma-distributed, the sizes of LW are higher than 0.1 no matter how large the values of $p$ and $n$ are while the sizes of CLRT and CJ all converge to the nominal level 0.05 as either $p$ or $n$ gets larger. This empirically confirms that normal assumptions are needed for the result of [@r15] while our corrected statistics CLRT and CJ (also the C test) have no such distributional restriction.
As for empirical powers, we consider two alternatives (here, the limiting spectral distributions of $\Sigma_p$ under these two alternatives differs from that under $H_0$):
1. $\Sigma_p$ is diagonal with half of its diagonal elements 0.5 and half 1. We denote its power by Power 1 ;
2. $\Sigma_p$ is diagonal with 1/4 of the elements equal 0.5 and 3/4 equal 1. We denote its power by Power 2 .
Table \[POWER\] reports the powers of LW(=CJ), CLRT and C when $\{y_{ij}\}$ are distributed as $N(0,1)$, and of CJ, CLRT and C when $\{y_{ij}\}$ are distributed as Gamma(4,2)-2, for the situation when $n$ equals $64$ or $128$, with varying values of $p$ and under the above mentioned two alternatives. For $n=256$ and $p$ varying from 16 to 240, all the tests have powers around 1 under both alternatives so that these values are omitted. And in order to find the trend of these powers, we also present the results when $n=128$ in Figure \[normal128\] and Figure \[gamma128\].
[cccccccccc]{}\
& & & & &\
& & LW/CJ & CLRT & C & & & LW/CJ & CLRT & C\
(4,64) & & 0.7754 & 0.7919 & 0.772 & & & 0.4694 & 0.6052 & 0.4716\
(8,64) & & 0.8662 & 0.8729 & 0.8582 & & & 0.5313 & 0.6756 & 0.5308\
(16,64) & & 0.912 & 0.9075 & 0.9029 & & & 0.5732 & 0.6889 & 0.5671\
(32,64) & & 0.9384 & 0.8791 & 0.931 & & & 0.5868 & 0.6238 & 0.5775\
(48,64) & & 0.9471 & 0.7767 & 0.9389 & & & 0.6035 & 0.5036 & 0.5982\
(56,64) & & 0.949 & 0.6663 & 0.9411 & & & 0.6025 & 0.4055 & 0.5982\
(60,64) & & 0.9501 & 0.5575 & 0.941 & & & 0.6048 & 0.3328 & 0.5989\
(8,128) & & 0.9984 & 0.9989 & 0.9986 & & & 0.9424 & 0.9776 & 0.9391\
(16,128) & & 0.9998 & 1 & 0.9998 & & & 0.9698 & 0.9926 & 0.9676\
(32,128) & & 1 & 1 & 1 & & & 0.9781 & 0.9956 & 0.9747\
(64,128) & & 1 & 1 & 1 & & & 0.9823 & 0.9897 & 0.9788\
(96,128) & & 1 & 0.9996 & 1 & & & 0.9824 & 0.9532 & 0.9804\
(112,128) & & 1 & 0.9943 & 1 & & & 0.9841 & 0.881 & 0.9808\
(120,128) & & 1 & 0.9746 & 1 & & & 0.9844 & 0.7953 & 0.9817\
\
\
\
& & & & &\
& & CJ & CLRT & C & & & CJ & CLRT & C\
(4,64) & & 0.6517 & 0.6826 & 0.6628 & & & 0.3998 & 0.5188 & 0.4204\
(8,64) & & 0.7693 & 0.7916 & 0.781 & & & 0.4757 & 0.5927 & 0.4889\
(16,64) & & 0.8464 & 0.8439 & 0.846 & & & 0.5327 & 0.633 & 0.5318\
(32,64) & & 0.9041 & 0.848 & 0.9032 & & & 0.5805 & 0.5966 & 0.5667\
(48,64) & & 0.9245 & 0.7606 & 0.924 & & & 0.5817 & 0.4914 & 0.5804\
(56,64) & & 0.9267 & 0.6516 & 0.9247 & & & 0.5882 & 0.4078 & 0.583\
(60,64) & & 0.9288 & 0.5547 & 0.9257 & & & 0.5919 & 0.3372 & 0.5848\
(8,128) & & 0.9859 & 0.9875 & 0.9873 & & & 0.8704 & 0.9294 & 0.8748\
(16,128) & & 0.999 & 0.999 & 0.9987 & & & 0.9276 & 0.9699 & 0.9311\
(32,128) & & 0.9999 & 1 & 0.9999 & & & 0.9582 & 0.9873 & 0.9587\
(64,128) & & 1 & 0.9998 & 1 & & & 0.9729 & 0.984 & 0.9727\
(96,128) & & 1 & 0.999 & 1 & & & 0.9771 & 0.9482 & 0.9763\
(112,128) & & 1 & 0.9924 & 1 & & & 0.9781 & 0.8747 & 0.9763\
(120,128) & & 1 & 0.9728 & 1 & & & 0.9786 & 0.7864 & 0.977\
The behavior of Power 1 and Power 2 in each figure related to the three statistics are similar, except that Power 1 is much higher compared with Power 2 for a given dimension design $(p,n)$ and any given test for the reason that the first alternative differs more from the null than the second one. The powers of LW (in the normal case), CJ (in the Gamma case) and C are all monotonically increasing in $p$ for a fixed value of $n$. But for CLRT, when $n$ is fixed, the powers first increase in $p$ and then become decreasing when $p$ is getting close to $n$. This can be explained by the fact that when $p$ is close to $n$, some of the eigenvalues of $S_n$ are getting close to zero, causing the CLRT nearly degenerate and losing power.
![Empirical powers of LW/CJ, CLRT and C test at $5\%$ significance level based on 10000 independent applications with real N(0,1) random variables for fixed $n=128$ under two alternatives Power 1 and 2 (see the text for details). []{data-label="normal128"}](normal128power1.eps){width="6.5cm"}
![Empirical powers of LW/CJ, CLRT and C test at $5\%$ significance level based on 10000 independent applications with real N(0,1) random variables for fixed $n=128$ under two alternatives Power 1 and 2 (see the text for details). []{data-label="normal128"}](normal128power2.eps){width="6.5cm"}
![Empirical powers of CJ, CLRT and C test at $5\%$ significance level based on 10000 independent applications with real Gamma(4,2)-2 random variables for fixed $n=128$ under two alternatives Power 1 and 2 (see the text for details). []{data-label="gamma128"}](gamma128power1.eps){width="6.5cm"}
![Empirical powers of CJ, CLRT and C test at $5\%$ significance level based on 10000 independent applications with real Gamma(4,2)-2 random variables for fixed $n=128$ under two alternatives Power 1 and 2 (see the text for details). []{data-label="gamma128"}](gamma128power2.eps){width="6.5cm"}
Besides, we find that in the normal case the trend of C’s power is very much alike of those of LW while in the Gamma case it is similar with those of CJ under both alternatives. And in most of the cases (especially in large $p$ case), the power of C test is slightly lower than LW (in the normal case) and CJ (in the Gamma case).
Lastly, we examine the performance of CJ and C when $p$ is larger than $n$. Empirical sizes and powers are presented in Table \[cnsizepower\]. We choose the variables to be distributed as Gamma(4,2)-2 since CJ reduces to LW in the normal case, and [@r15] has already reported the performance of LW when $p$ is larger than $n$. From the table, we see that when $p$ is larger than $n$, the size of CJ is still correct and it is always around the nominal level 0.05 as the dimension $p$ increases and the same phenomenon exists for C test.
----------- -- -------- --------- --------- -- -- -------- --------- ---------
Size Power 1 Power 2 Size Power 1 Power 2
(64,64) 0.0624 0.9282 0.5897 0.0624 0.9257 0.5821
(320,64) 0.0577 0.9526 0.612 0.0576 0.9472 0.6059
(640,64) 0.0558 0.959 0.6273 0.0562 0.9541 0.6105
(960,64) 0.0543 0.9631 0.6259 0.0551 0.955 0.6153
(1280,64) 0.0555 0.9607 0.6192 0.0577 0.9544 0.6067
----------- -- -------- --------- --------- -- -- -------- --------- ---------
: \[cnsizepower\]Empirical sizes and powers (Power 1 and 2) of CJ test and C test at $5\%$ significance level based on 10000 independent applications with real Gamma(4,2)-2 random variables when $p\geq n$.
When we evaluate the power, the same two alternatives Power 1 and Power 2 as above are considered. The sample size is fixed to $n=64$ and the ratio $p/n$ varies from 1 to 20. We see that Power 1 are much higher than Power 2 for the same reason that the first alternative is easier to be distinguished from $H_0$. Besides, the powers under both alternatives all increase monotonically for $1\leq \frac{p}{n}\leq 15$. However, when $p/n$ is getting larger, say $p/n=20$, we can observe that its size is a little larger and powers a little drop (compared with $p/n=15$) but overall, it still behaves well, which can be considered as free from the assumption constraint “$p/n\rightarrow y$”. Besides, the powers of CJ are always slightly higher than those of C in this “large $p$ small $n$” setting.
Since the asymptotic distribution for the CLRT and CJ are both derived under the “Marcenko-Pasture scheme” (i.e $p/n\rightarrow y \in (0,\infty)$), if $p/n$ is getting too large ($p \gg n$), it seems that the limiting results provided in this paper will loose accuracy. It is worth noticing that [@r7] has extended the LW test to such a scheme ($p \gg n$) for multivariate normal distribution.
Summarizing all these findings from this Monte-Carlo study, the overall figure is the following: when the ratio $p/n$ is much lower than 1 (say smaller than 1/2), it is preferable to employ CLRT (than CJ, LW or C); while this ratio is higher, CJ (or LW for normal data) becomes more powerful (slightly more powerful than C).
Asymptotic powers: under the spiked population alternative {#powerspike}
==========================================================
In this section, we give an analysis of the powers of the two corrections: CLRT and CJ. To this end, we consider an alternative model that has attracted lots of attention since its introduction by [@r13], namely, the spiked population model. This model can be described as follows: the eigenvalues of $\Sigma_p$ are all one except for a few fixed number of them. Thus, we restrict our sphericity testing problem to the following: $$H_0:\Sigma_p=I_p~~~vs~~~H_1^{*}: \Sigma_p=diag(\underbrace{a_1,\cdots,a_1}_{n_1},
\cdots,
\underbrace{a_k,\cdots,a_k}_{n_k},
\underbrace{1,\cdots,1}_{p-M}),$$ where the multiplicity numbers $n_i$’s are fixed and satisfying $n_1+\cdots+n_k=M$. We derive the explicit [[expressions]{}]{} of the power [[functions]{}]{} of CLRT and CJ in this section.
Under $H_1^{*}$, the empirical spectral distribution of $\Sigma_p$ is $$\begin{aligned}
\label{converge}
H_n=\frac{1}{p}\sum_{i=1}^kn_i\delta_{a_i}+\frac{p-M}{p}\delta_1~,\end{aligned}$$ and it will converge to $\delta_1$, a Dirac mass at 1, which is the same as the limit under the null hypothesis $H_0: \Sigma_p=I_p$. From this point of view, anything related to the limiting spectral distribution remains the same whenever under $H_0$ or $H_1^{*}$. Then recall the CLT for LSS of the sample covariance matrix, [[as provided in [@r2]]{}]{}, is of the form: $$p\int f(x)d(F^{S_n}(x)-F^{y_n}(x))\Longrightarrow N(\mu, \sigma^2)~,$$ where the right side of this equation is determined only by the limiting spectral distribution. So we can conclude that the limiting parameters $\mu$ and $\sigma^2$ remain the same under $H_0$ and $H_1^{*}$, only the centering term $p\int f(x)dF^{y_n}(x)$ possibly makes a difference. Since there’s a $p$ in front of $\int f(x)dF^{y_n}(x)$, which tends to infinity as assumed, so knowing the convergence $H_n\rightarrow \delta_1$ is not enough and more details about the convergence are needed. In [@r28], we have established an asymptotic expansion for the centering parameter when the population has a spiked structure. We will use these formulas like [[equations]{}]{} , and in the following to derive the powers of the CLRT and CJ.
Lemma \[lem:joint\] remains the same under $H_1^{*}$, except that this time the centering terms become (see formulas (4.11) and (4.12) in [@r28]): $$\begin{aligned}
&&F^{y_n}(\log x)=\frac{1}{p}\sum_{i=1}^k n_i\log a_i-1+(1-\frac{1}{y_n})\log(1-y_n)+O(\frac{1}{p^2})~,\label{logx}\\
&&F^{y_n}(x)=1+\frac{1}{p}\sum_{i=1}^k n_i a_i-\frac{M}{p}+O(\frac{1}{p^2})~.\label{x}\end{aligned}$$ Repeating the proof of Theorem \[clt\], we can get: $$\begin{aligned}
\label{spike1clt}
&&\mathcal{L}_n-p+(p-n)\log(1-\frac{p}{n})+\sum_{i=1}^{k}n_i\big(\log a_i-a_i+1\big)\nonumber\\
&&\Longrightarrow N\Big(-\frac{\kappa-1}{2}\log(1-y)+\frac12\beta y~, ~-\kappa\log(1-y)-\kappa y\Big)\end{aligned}$$ under $H_1^{*}$. As a result, the power of CLRT for testing $H_0$ against $H_1^{*}$ can be expressed as: $$\begin{aligned}
\label{b1}
\beta_1(\alpha)=1-\Phi\bigg(\Phi^{-1}(1-\alpha)-\frac{\sum_{i=1}^k n_i\big(a_i-\log a_i-1\big)}{\sqrt{-\kappa \log(1-y)-\kappa y}}\bigg)~\end{aligned}$$ for a pre-given significance level $\alpha$.
[[It is worth noticing here that if the alternative has only one simple spike, i.e. $k=1,~n_k=1$, and assuming the real Gaussian variable case, i.e. $\kappa=2$, reduces to a result provided in [@r20]. However, our formula is valid for a general number of spikes with eventual multiplicities. Besides, these authors use some more sophisticated tools of asymptotic contiguity and Le Cam’s first and third lemmas, which are totally different from ours.]{}]{}
In order to calculate the power function of CJ, we restated Lemma \[cnlemma:joint\] as follows: $$\left(
\begin{array}{cc}
\sum_{i=1}^{p} l_{i}^2-pF^{y_{n}}( x^2)
\\
\sum_{i=1}^{p}l_{i}-pF^{y_{n}}(x)
\\
\end{array}
\right) \Longrightarrow
N(\mu_2,V_2)~,$$ where $$\begin{aligned}
&&F^{y_n}(x^2)=\frac{2}{n}\sum_{i=1}^k n_ia_i-\frac{2}{n} M+1+y_n-\frac{M}{p}+\frac1p\sum_{i=1}^k n_i a_i^2+O(\frac{1}{p^2}),\label{xx}\\
&&F^{y_n}(x)=1+\frac{1}{p}\sum_{i=1}^k n_i a_i-\frac{M}{p}+O(\frac{1}{p^2})~.\nonumber\end{aligned}$$ Using the delta method as in the proof of Theorem \[cnclt\], this time, we have $$\begin{aligned}
\label{spike2clt}
nU-p-\frac np \sum_{i=1}^{k}n_i(a_i-1)^2\Longrightarrow N(\kappa+\beta-1, 2\kappa)\end{aligned}$$ under $H_1^{*}$.
Power function of CJ can be expressed as $$\begin{aligned}
\label{b2}
\beta_2(\alpha)=1-\Phi\bigg(\Phi^{-1}(1-\alpha)-\frac np\cdot\frac{\sum_{i=1}^k n_i(a_i-1)^2}{\sqrt{2\kappa}}\bigg)~\end{aligned}$$ for a pre-given significance level $\alpha$.
Now consider the [[functions]{}]{} $a_i-\log a_i-1$ and $(a_i-1)^2$ appearing in the [[expressions]{}]{} and , they will achieve their minimum value $0$ at $a_i=1$, which is to say, once $a_i$’s going away from 1, the powers $\beta_1(\alpha)$ and $\beta_2(\alpha)$ will both increase. This phenomenon agrees with our intuition, since the more $a_i$’s deviate away from 1, the easier to distinguish $H_0$ from $H_1^{*}$. Therefore, the powers should naturally grow higher.
![Theoretical powers of CLRT ($\beta_1(\alpha)$) and CJ ($\beta_2(\alpha)$) under the spiked population alternative.[]{data-label="beta"}](beta.eps){width="7.5cm"}
Then, we consider the power functions $\beta_1(\alpha)$ and $\beta_2(\alpha)$ as functions of $y$, and see how they are going along with $y's$ changing. we see that in expression , $\sqrt{-\log(1-y)-y}$ is increasing when $y \in (0, 1)$, so $\beta_1(\alpha)$ is decreasing as the function of $y$, which attains its maximum value $1$ when $y\rightarrow0_{+}$ and minimum value $\alpha$ when $y\rightarrow1_{-}$. Also, expression is obviously a decreasing function of $y$, attaining its maximum value $1$ when $y\rightarrow0_{+}$ and minimum value $\alpha$ when $y\rightarrow +\infty$. We present the trends of $\beta_1(\alpha)$ and $\beta_2(\alpha)$ (corresponding to the power of CLRT and CJ) in Figure \[beta\] when only one spike $a=2.5$ exists. It is however a little different from the non-spiked case as showed in the simulation in Section \[sec:simul\] (both Figure \[normal128\] and Figure \[gamma128\]), where the power of CLRT first increases then decreases while the power of CJ is always increasing along with the increase of the value of $p$. These power drops are due to the fact that when $p$ increases, since only one spike eigenvalue is considered, it becomes harder to distinguish both hypotheses. Besides, an interesting finding here is that these power functions give a new confirmation of the fact that CLRT behaves quite badly when $y\rightarrow 1_{-}$, while CJ test has a reasonable power for a significant range of $y>1$.
Generalization to the case when the population mean is unknown {#ge}
==============================================================
So far, we have assumed the observation $(Y_i)$ are centered. However, this is hardly true in practical situations when $\mu=EY_i$ is usually unknown. Therefore, the sample covariance matrix should be taken as $$S_n^{*}=\frac{1}{n-1}\sum_{i=1}^{n}(Y_i-\overline{Y})(Y_i-\overline{Y})^{*}~,$$ where $\overline{Y}=\frac{1}{n}\sum_{i=1}^{n}Y_i$ is the sample mean. Because $\overline{Y}\overline{Y}^{*}$ is a rank one matrix, substituting $S_n$ for $S_n^{*}$ when $\mu$ is unknown will not affect the limiting distribution in the CLT for LSS; while it is not the case for the centering parameter, for it has a $p$ in front.
Recently, [@r22] shows that if we use $S_n^{*}$ in the CLT for LSS when $\mu$ is unknown, the limiting variance remains the same as we use $S_n$; while the limiting mean has a shift which can be expressed as a complex contour integral. Later, [@r30] looks into this shift and finally derives a concise conclusion on the CLT corresponding to $S_n^{*}$: the random vector $\big(X_n^{*}(f_1), \cdots, X_n^{*}(f_k)\big)$ converges weakly to a Gaussian vector with the same mean and covariance function as given in Theorem \[t1\], where this time, $X_n^{*}(f)=p\int f(x)d(F^{S_n^{*}}-F^{y_{n-1}})$. It is here important to pay attention that the only difference is in the centering term, where we use the new ratio $y_{n-1}=\frac{p}{n-1}$ instead of the previous $y_{n}=\frac{p}{n}$, while leaving all the other terms unchanged.
Using this result, we can modify our Theorems \[clt\] and \[cnclt\] to get the CLT of CLRT and CJ under $H_0$ when $\mu$ is unknown only by considering the eigenvalues of $S_n^{*}$ and substituting $n-1$ for $n$ in the centering terms. More precisely, now [[equations]{}]{} and in [[Theorems]{}]{} \[clt\] and \[cnclt\] become $$\begin{aligned}
&&\mathcal L_n + (p-n+1)\cdot\log(1-\frac{p}{n-1})-p\nonumber\\
&&\Longrightarrow
N\{-\frac{\kappa-1}{2}\log(1-y) + \frac12\beta y,
-\kappa\log(1-y)-\kappa y \}~
\end{aligned}$$ and $$\begin{aligned}
n U-\frac{np}{n-1}\Longrightarrow N(\kappa+\beta-1,2\kappa)~.
\end{aligned}$$
The same procedures can be applied to get the CLT of CLRT and CJ under $H_1^{*}$ when $\mu$ is unknown. This time, equations and become $$\begin{aligned}
&&\mathcal{L}_n-p+(p-n+1)\log(1-\frac{p}{n-1})+\sum_{i=1}^{k}n_i\big(\log a_i-a_i+1\big)\nonumber\\
&&\Longrightarrow N\Big(-\frac{\kappa-1}{2}\log(1-y)+\frac12\beta y~, ~-\kappa\log(1-y)-\kappa y\Big)\end{aligned}$$ and $$\begin{aligned}
nU-\frac{np}{n-1}-\frac {n}{p} \sum_{i=1}^{k}n_i(a_i-1)^2\Longrightarrow N(\kappa+\beta-1, 2\kappa)~,\end{aligned}$$ and therefore the powers of CLRT and CJ under the spiked alternative remain unchanged as expressed in and .
.
Additional proofs {#sec:proofs}
=================
[[We recall these two important formulas which appear in the Appendix as and here for the convenience of reading: $$\begin{aligned}
{\mathbb{E}}[X_f]& =& (\kappa-1) I_1(f) + \beta I_2(f) ~,\\
{\mathop{\text{Cov}}}(X_{f},X_{g}) &=& \kappa J_1(f,g) + \beta
J_2(f,g)~.
\end{aligned}$$ ]{}]{}
Proof of Lemma \[lem:joint\]
----------------------------
Let for $x>0$, $f(x)=\log x$ and $g(x)=x$. Define $A_n$ and $B_n$ by the decompositions $$\begin{aligned}
\sum_{i=1}^{p}\log \ell_i & = & p\int f(x)
d(F_{n}(x)-F^{y_{n}}(x))+pF^{y_{n}}(f) = A_n+pF^{y_{n}}(f)~ , \\
\sum_{i=1}^{p}\ell_i & = &
p\int g(x) d(F_{n}(x)-F^{y_{n}}(x))+pF^{y_{n}}(g)= B_n + pF^{y_{n}}(g) ~.\end{aligned}$$ Applying [[Theorem \[t1\]]{}]{} given in the Appendix to the pair $(f,g)$, we have $$\left(
\begin{array}{cc}
A_n \\ B_n
\end{array}
\right)
\Longrightarrow
N\Bigg(
\left(
\begin{array}{cc}
EX_{f} \\ EX_{g}
\end{array}
\right), \
\left(
\begin{array}{cc}
{\mathop{\text{Cov}}}(X_f,X_f)
&
{\mathop{\text{Cov}}}(X_f, X_g)
\\
{\mathop{\text{Cov}}}(X_g,X_f)
& {\mathop{\text{Cov}}}(X_g,X_g)
\end{array}
\right) \
\Bigg).$$ It remains to evaluate the limiting parameters and this results from the following calculations [[where $h$ is denoted as $\sqrt{y}$]{}]{}: $$\begin{aligned}
I_1(f,r)&=& \frac12\log\left( 1-h^2/r^2\right)~,\label{I1f}\\
I_1(g,r) &=& 0 ~, \label{I1g} \\
I_2(f)&=& -\frac12 h^2~, \label{I2f} \\
I_2(g)&=& 0~, \label{I2g} \\
J_1(f,g,r) &=& \frac {h^2}{r^2}~, \label{J1fg}\\
J_1(f,f,r) &=& -\frac{1}{r} \log(1-h^2/r) ~, \label{J1ff}\\
J_1(g,g,r) &=& \frac {h^2}{r^2}~, \label{J1gg}\\
J_2(f,g) &=& {h^2} ~, \label{J2fg}\\
J_2(f,f) &=& {h^2} ~, \label{J2ff}\\
J_2(g,g) &=& {h^2} ~. \label{J2gg}\end{aligned}$$
We now detail these calculations to complete the proof. They are all based on the formula given in Proposition \[t2\] in the Appendix and repeated use of the residue theorem.
[*Proof of* ]{}:We have $$\begin{aligned}
I_1(f,r)&=& \frac{1}{2\pi i}\oint_{|\xi|=1}f(|1+h\xi|^{2})\Big[\frac{\xi}{\xi^{2}-r^{-2}}-\frac{1}{\xi}\Big]d\xi\\
&=& \frac{1}{2\pi i}\oint_{|\xi|=1}\log(|1+h\xi|^{2})\Big[\frac{\xi}{\xi^{2}-r^{-2}}-\frac{1}{\xi}\Big]d\xi\\
&=& \frac{1}{2\pi i}\oint_{|\xi|=1}(\frac{1}{2}\log((1+h\xi)^{2})+\frac{1}{2}\log((1+h\xi^{-1})^{2})\Big[\frac{\xi}{\xi^{2}-r^{-2}}-\frac{1}{\xi}\Big]d\xi\\
&=& \frac{1}{2\pi i}\Big[\oint_{|\xi|=1}\log(1+h\xi)\frac{\xi}{\xi^{2}-r^{-2}}d\xi-\oint_{|\xi|=1}\log(1+h\xi)\frac{1}{\xi}d\xi\\
&&+\oint_{|\xi|=1}\log(1+h\xi^{-1})\frac{\xi}{\xi^{2}-r^{-2}}d\xi-\oint_{|\xi|=1}\log(1+h\xi^{-1})\frac{1}{\xi}d\xi\Big]\\\end{aligned}$$ For the first integral, note that as $r>1$, the poles are $\pm \frac{1}{r}$ and we have by the residue theorem, $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi|=1}\log(1+h\xi)\frac{\xi}{\xi^{2}-r^{-2}}d\xi\\
&=&\frac{\log(1+h\xi)\cdot\xi}{\xi-r^{-1}}\Bigg|_{\xi=-r^{-1}}+\frac{\log(1+h\xi)\cdot\xi}{\xi+r^{-1}}\Bigg|_{\xi=r^{-1}}\\
&=&\frac12 \log(1-\frac{h^{2}}{r^{2}})~.\end{aligned}$$ For the second integral, $$\begin{aligned}
{{\frac1{2\pi i}}}\oint_{|\xi|=1}\log(1+h\xi)\frac{1}{\xi}d\xi= \log(1+h\xi)\big|_{\xi=0}=0~.\end{aligned}$$ The third one is $$\begin{aligned}
&&{{{\frac1{2\pi i}}}\oint_{|\xi|=1}\log(1+h\xi^{-1})\frac{\xi}{\xi^{2}-r^{-2}}d\xi}\\
&&=-{{\frac1{2\pi i}}}\oint_{|z|=1}\log(1+hz)\frac{z^{-1}}{z^{-2}-r^{-2}}\cdot\frac{-1}{z^{2}}dz\\
&&={{\frac1{2\pi i}}}\oint_{|z|=1}\frac{\log(1+hz)r^{2}}{z(z+r)(z-r)}dz=\frac{\log(1+hz)r^{2}}{(z+r)(z-r)}\Bigg|_{z=0}=0~,\end{aligned}$$ where the first equality results from the change of variable $z=\frac{1}{\xi}$, and the third equality holds because $r >1$, so the only pole is $z=0$.\
The fourth one equals $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi|=1}\log(1+h\xi^{-1})\frac{1}{\xi}d\xi=-{{\frac1{2\pi i}}}\oint_{|z|=1}\log(1+hz)\frac{-z}{z^{2}}dz\\
&&=\log(1+hz)\big|_{z=0}=0~.\end{aligned}$$ Collecting the four integrals leads to the desired formula for $I_1(f,r)$.
[*Proof of* ]{}:We have
$$\begin{aligned}
I_1(g,r)&=& \frac{1}{2\pi i}\oint_{|\xi|=1}g(|1+h\xi|^{2})\cdot[\frac{\xi}{\xi^{2}-r^{-2}}-\frac{1}{\xi}]d\xi\\
&=& \frac{1}{2\pi i}\oint_{|\xi|=1}|1+h\xi|^{2}\cdot\Big[\frac{\xi}{\xi^{2}-r^{-2}}-\frac{1}{\xi}\Big]d\xi\\
&=& \frac{1}{2\pi i}\oint_{|\xi|=1}\frac{\xi+h+h\xi^2+h^2\xi}{\xi}\cdot\Big[\frac{\xi}{\xi^{2}-r^{-2}}-\frac{1}{\xi}\Big]d\xi\\
&=& \frac{1}{2\pi i}\oint_{|\xi|=1}\frac{\xi+h+h\xi^2+h^2\xi}{\xi^2-r^{-2}}d\xi\\
&&-\frac{1}{2\pi i}\oint_{|\xi|=1}\frac{\xi+h+h\xi^2+h^2\xi}{\xi^2}d\xi~.\\\end{aligned}$$
These two integrals are calculated as follows: $$\begin{aligned}
&&\frac{1}{2\pi i}\oint_{|\xi|=1}\frac{\xi+h+h\xi^2+h^2\xi}{\xi^2-r^{-2}}d\xi\\
&=&\frac{\xi+h+h\xi^2+h^2\xi}{\xi-r^{-1}}\Big|_{\xi=-r^{-1}}+\frac{\xi+h+h\xi^2+h^2\xi}{\xi+r^{-1}}\Big|_{\xi=r^{-1}}\\
&=&1+h^2~;\end{aligned}$$ $$\begin{aligned}
&&\frac{1}{2\pi i}\oint_{|\xi|=1}\frac{\xi+h+h\xi^2+h^2\xi}{\xi^2}d\xi
=\frac{\partial}{\partial \xi}(\xi+h+h\xi^2+h^2\xi)\Big|_{\xi=0}\\
&&=1+h^2~.\end{aligned}$$
Collecting the two terms leads to $I_1(g,r)=0$.
[*Proof of* ]{}: $$\begin{aligned}
I_2(f)&=&\frac{1}{2\pi i}\oint_{|\xi|=1}\log(|1+h\xi|^{2})\frac{1}{\xi^{3}}d\xi\\
&=&\frac{1}{2\pi i}\Bigg[\oint_{|\xi|=1}\frac{\log(1+h\xi)}{\xi^{3}}d\xi+\oint_{|\xi|=1}\frac{\log(1+h\xi^{-1})}{\xi^{3}}d\xi\Bigg]~.\\\end{aligned}$$ We have $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi|=1}\frac{\log(1+h\xi)}{\xi^{3}}d\xi=\frac12\frac{\partial^{2}}{\partial
\xi^{2}}\log(1+h\xi)\Big|_{\xi=0} =-\frac12 h^{2}~;\\
&&{{\frac1{2\pi i}}}\oint_{|\xi|=1}\frac{\log(1+h\xi^{-1})}{\xi^{3}}d\xi=-{{\frac1{2\pi i}}}\oint_{|z|=1}\frac{\log(1+hz)}{\frac{1}{z^{3}}}\cdot \frac{-1}{z^{2}}dz
=0~.\end{aligned}$$ Combining the two leads to $I_2(f)=-\frac12 h^{2}$.
[*Proof of* ]{}: $$\begin{aligned}
&& I_2(g)
=\frac{1}{2\pi i}\oint_{|\xi|=1}\frac{(1+h\overline{\xi})(1+h\xi)}{\xi^{3}}d\xi
=\frac{1}{2\pi i}\oint_{|\xi|=1}\frac{\xi+h\xi^{2}+h+h^{2}\xi}{\xi^{4}}d\xi
=0~.\end{aligned}$$
[*Proof of* ]{}: $$\begin{aligned}
J_{1}(f,g,r)
&=&\frac{1}{2\pi i}\oint_{|\xi_{2}|=1}|1+h\xi_{2}|^{2}\cdot {{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1} \frac{\log(|1+h\xi_{1}|^{2})}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}d\xi_{2}~.\\\end{aligned}$$ We have, $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1} \frac{\log(|1+h\xi_{1}|^{2})}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}\\
&=&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1} \frac{\log(1+h\xi_{1})}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}+{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1} \frac{\log(1+h\xi_{1}^{-1})}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}~.\end{aligned}$$ The first term $$\begin{aligned}
{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1} \frac{\log(1+h\xi_{1})}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}=0,\end{aligned}$$ because for fixed $|\xi_2|=1$, $|r\xi_{2}|=|r|> 1$, so $r\xi_{2}$ is not a pole.\
The second term is $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1} \frac{\log(1+h\xi_{1}^{-1})}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}=-{{\frac1{2\pi i}}}\oint_{|z|=1} \frac{\log(1+hz)}{(\frac{1}{z}-r\xi_{2})^{2}}\cdot \frac{-1}{z^{2}}dz\\
&&={{\frac1{2\pi i}}}\cdot\frac{1}{(r\xi_2)^{2}}\oint_{|z|=1} \frac{\log(1+hz)}{(z-\frac{1}{r\xi_{2}})^{2}}dz=\frac{1}{(r\xi_2)^{2}}\cdot \frac{\partial}{\partial z}\log(1+hz)\Big|_{z=\frac{1}{r\xi_2}}\\
&&=\frac{h}{r\xi_2(r\xi_2+h)}~,\end{aligned}$$ where the first equality results from the change of variable $z=\frac{1}{\xi_1}$, and the third equality holds because for fixed $|\xi_2|=1$, $|\frac{1}{r\xi_2}|=\frac{1}{|r|}<1$, so $\frac{1}{r\xi_2}$ is a pole of second order.\
Therefore, $$\begin{aligned}
&&J_1(f,g,r)\\
&=&\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{(1+h\xi_2)(1+h\overline{\xi_2})}{\xi_{2}(\xi_{2}+\frac{h}{r})}d\xi_{2}\\
&=&\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{\xi_2+h\xi_2^{2}+h+h^{2}\xi_2}{\xi_{2}^{2}(\xi_{2}+\frac{h}{r})}d\xi_{2}\\
&=&\frac{h}{2\pi ir^{2}}\Bigg[\oint_{|\xi_{2}|=1}\frac{1+h^2}{\xi_{2}(\xi_{2}+\frac{h}{r})}d\xi_{2}+
\oint_{|\xi_{2}|=1}\frac{h}{\xi_{2}+\frac{h}{r}}d\xi_{2}
+\oint_{|\xi_{2}|=1}\frac{h}{\xi_{2}^2(\xi_{2}+\frac{h}{r})}d\xi_{2}\Bigg]~.\end{aligned}$$ Finally we find $J_1(f,g,r)=\frac{h^2}{r^2}$ since $$\begin{aligned}
&&\frac{h}{2\pi
ir^{2}}\oint_{|\xi_{2}|=1}\frac{1+h^2}{\xi_{2}(\xi_{2}+\frac{h}{r})}d\xi_{2}=0~,\quad
\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{h}{\xi_{2}+\frac{h}{r}}d\xi_{2}=\frac{h^2}{r^2}~,\\
&&\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{h}{\xi_{2}^2(\xi_{2}+\frac{h}{r})}d\xi_{2}=0~.\end{aligned}$$
[*Proof of* ]{}: $$\begin{aligned}
J_1(f,f,r)&=&{{\frac1{2\pi i}}}\oint_{|\xi_{2}|=1}f(|1+h\xi_{2}|^{2})\cdot {{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1} \frac{f(|1+h\xi_{1}|^{2})}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}d\xi_{2}\\
&=& {{\frac1{2\pi i}}}\oint_{|\xi_{2}|=1}f(|1+h\xi_{2}|^{2})\frac{h}{r\xi_2(r\xi_2+h)}d\xi_{2}\\
&=&\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{\log(1+h\xi_{2})}{\xi_{2}(\frac{h}{r}+\xi_{2})}d\xi_{2}+\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{\log(1+h\xi_{2}^{-1})}{\xi_{2}(\frac{h}{r}+\xi_{2})}d\xi_{2}~.\\\end{aligned}$$ We have $$\begin{aligned}
&&\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{\log(1+h\xi_{2})}{\xi_{2}(\frac{h}{r}+\xi_{2})}d\xi_{2}\\
&=&\frac{h}{r^{2}}\Bigg[\frac{\log(1+h\xi_{2})}{\frac{h}{r}+\xi_{2}}\Bigg|_{\xi_{2}=0}+\frac{\log(1+h\xi_{2})}{\xi_{2}}\Bigg|_{\xi_{2}=-\frac{h}{r}}\Bigg]\\
&=&-\frac{1}{r}\log(1-\frac{h^{2}}{r})~,\\end{aligned}$$ and $$\begin{aligned}
&&\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{\log(1+h\xi_{2}^{-1})}{\xi_{2}(\frac{h}{r}+\xi_{2})}d\xi_{2}=\frac{-h}{2\pi ir^{2}}\oint_{|z|=1}\frac{\log(1+hz)}{\frac{1}{z}(\frac{h}{r}+\frac{1}{z})}\cdot \frac{-1}{z^2}dz\\
&&=\frac{1}{2\pi ir}\oint_{|z|=1}\frac{\log(1+hz)}{z+\frac{r}{h}}dz
=0~,\end{aligned}$$ where the first equality results from the change of variable $z=\frac{1}{\xi_2}$, and the third equality holds because $|\frac{r}{h}|> 1$, so $\frac{r}{h}$ is not a pole.
Finally, we find $J_1(f,f,r)=-\frac{1}{r}\log(1-\frac{h^{2}}{r})$ .
[*Proof of* ]{}: $$\begin{aligned}
J_1(g,g,r)&=&{{\frac1{2\pi i}}}\oint_{|\xi_{2}|=1}|1+h\xi_{2}|^{2}\cdot{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{|1+h\xi_{1}|^{2}}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}d\xi_{2}~.\end{aligned}$$ We have $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{|1+h\xi_{1}|^{2}}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}
={{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{\xi_1+h\xi_1^{2}+h+h^{2}\xi_1}{\xi_1(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}\\
&&={{\frac1{2\pi i}}}\Bigg[\oint_{|\xi_{1}|=1}\frac{1+h^2}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}+\oint_{|\xi_{1}|=1}\frac{h\xi_1}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}\\
&&+\oint_{|\xi_{1}|=1}\frac{h}{\xi_1(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}\Bigg]\\
&&=\frac{h}{r^2\xi_2^{2}}~,\end{aligned}$$ since $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{1+h^2}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}=0~,\quad
{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{h\xi_1}{(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}=0~,\\
&&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{h}{\xi_1(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}=\frac{h}{(\xi_{1}-r\xi_{2})^{2}}\Bigg|_{\xi_1=0}=\frac{h}{r^2\xi_2^{2}}~,\\\end{aligned}$$ where the equality above holds because for fixed $|\xi_2|=1$, $|r\xi_{2}|=|r|> 1$, so $r\xi_{2}$ is not a pole. Therefore, $$\begin{aligned}
J_1(g,g,r)
&=&\frac{h}{2\pi ir^{2}}\oint_{|\xi_{2}|=1}\frac{\xi_2+h\xi_2^{2}+h+h^{2}\xi_2}{\xi_{2}^{3}}d\xi_{2}\\
&=&\frac{h}{2\pi
ir^{2}}\Bigg[\oint_{|\xi_{2}|=1}\frac{1+h^2}{\xi_{2}^{2}}d\xi_{2}+\oint_{|\xi_{2}|=1}\frac{h}{\xi_{2}}d\xi_{2}+\oint_{|\xi_{2}|=1}\frac{h}{\xi_{2}^{3}}d\xi_{2}\Bigg]~,\\
&=&\frac{h^2}{r^2}~.\end{aligned}$$
[*Proof of *]{}: We have $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{f(|1+h\xi_{1}|^{2})}{\xi_{1}^{2}}d\xi_{1}
={{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{\log(|1+h\xi_{1}|^{2})}{\xi_{1}^{2}}d\xi_{1}\\
&&={{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{\log(1+h\xi_{1})+\log(1+h\xi_{1}^{-1})}{\xi_{1}^{2}}d\xi_{1}~=h~,\end{aligned}$$ since $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{\log(1+h\xi_{1})}{\xi_{1}^{2}}d\xi_{1}= \frac{\partial}{\partial\xi_{1}}\Big(\log(1+h\xi_{1})\Big)\Bigg|_{\xi_{1}=0}=h~,\\
&&{{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{\log(1+h\xi_{1}^{-1})}{\xi_{1}^{2}}d\xi_{1}=-{{\frac1{2\pi i}}}\oint_{|z|=1}\frac{\log(1+hz)}{\frac{1}{z^{2}}}\cdot (-\frac{1}{z^{2}}dz)\\
&&={{\frac1{2\pi i}}}\oint_{|z|=1}\log(1+hz)dz
=0~.\end{aligned}$$ Similarly, $$\begin{aligned}
&&{{\frac1{2\pi i}}}\oint_{|\xi_{2}|=1}\frac{g(|1+h\xi_{2}|^{2})}{\xi_{2}^{2}}d\xi_{2}
={{\frac1{2\pi i}}}\oint_{|\xi_{2}|=1}\frac{\xi_2+h\xi_2^{2}+h+h^2\xi_2}{\xi_{2}^{3}}d\xi_{2}
=h.\end{aligned}$$ Therefore, $$\begin{aligned}
&&J_2(f,g)={{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{f(|1+h\xi_{1}|^{2})}{\xi_{1}^{2}}d\xi_{1}\cdot {{\frac1{2\pi i}}}\oint_{|\xi_{2}|=1}\frac{g(|1+h\xi_{2}|^{2})}{\xi_{2}^{2}}d\xi_{2}=h^2~,\\
&&J_2(f,f)={{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{f(|1+h\xi_{1}|^{2})}{\xi_{1}^{2}}d\xi_{1}\cdot {{\frac1{2\pi i}}}\oint_{|\xi_{2}|=1}\frac{f(|1+h\xi_{2}|^{2})}{\xi_{2}^{2}}d\xi_{2}=h^2~,\\
&&J_2(g,g)={{\frac1{2\pi i}}}\oint_{|\xi_{1}|=1}\frac{g(|1+h\xi_{1}|^{2})}{\xi_{1}^{2}}d\xi_{1}\cdot {{\frac1{2\pi i}}}\oint_{|\xi_{2}|=1}\frac{g(|1+h\xi_{2}|^{2})}{\xi_{2}^{2}}d\xi_{2}=h^2~.\\\end{aligned}$$
The proof of Lemma \[lem:joint\] is complete.\
Proof of Lemma \[cnlemma:joint\]
--------------------------------
Let $f(x)=x^2$ and $g(x)=x$. Define $C_n$ and $B_n$ by the decompositions $$\begin{aligned}
\sum_{i=1}^{p} \ell_i ^2& = & p\int f(x)
d(F_{n}(x)-F^{y_{n}}(x))+pF^{y_{n}}(f) = C_n+pF^{y_{n}}(f)~ , \\
\sum_{i=1}^{p}\ell_i & = &
p\int g(x) d(F_{n}(x)-F^{y_{n}}(x))+pF^{y_{n}}(g)= B_n + pF^{y_{n}}(g) ~.\end{aligned}$$ Applying [[Theorem \[t1\]]{}]{} given in the Appendix to the pair $(f,g)$, we have $$\left(
\begin{array}{cc}
C_n \\ B_n
\end{array}
\right)
\Longrightarrow
N\Bigg(
\left(
\begin{array}{cc}
EX_{f} \\ EX_{g}
\end{array}
\right), \
\left(
\begin{array}{cc}
{\mathop{\text{Cov}}}(X_f,X_f)
&
{\mathop{\text{Cov}}}(X_f, X_g)
\\
{\mathop{\text{Cov}}}(X_g,X_f)
& {\mathop{\text{Cov}}}(X_g,X_g)
\end{array}
\right) \
\Bigg).$$ It remains to evaluate the limiting parameters and this results from the following calculations: $$\begin{aligned}
I_1(f,r)&=& \frac{h^2}{r^2}~,\label{2I1f}\\
I_1(g,r) &=& 0 ~, \label{2I1g} \\
I_2(f)&=& h^2~, \label{2I2f} \\
I_2(g)&=& 0~, \label{2I2g} \\
J_1(f,g,r) &=& \frac {2h^2+2h^4}{r^2}~, \label{2J1fg}\\
J_1(f,f,r) &=& \frac{2h^4+(2h+2h^3)^2r}{r^3} ~, \label{2J1ff}\\
J_1(g,g,r) &=& \frac {h^2}{r^2}~, \label{2J1gg}\\
J_2(f,g) &=& {2h^2+2h^4} ~, \label{2J2fg}\\
J_2(f,f) &=& {(2h+2h^3)^2} ~, \label{2J2ff}\\
J_2(g,g) &=& {h^2} ~. \label{2J2gg}\end{aligned}$$
The results , , and are exactly the same as those found in , , and in the proof of Lemma \[lem:joint\]. The remaining results are found by similar calculations using Proposition \[t2\] in the Appendix and their details are omitted.
Concluding remarks
==================
Using recent central limit theorems for eigenvalues of large sample covariance matrices, we are able to find new asymptotic distributions for two major procedures to test the sphericity of a large-dimensional distribution. Although the theory is developed under the scheme $p\rightarrow \infty$, $n\rightarrow \infty$ and $p/n \rightarrow y >0$, our Monte-Carlo study has proved that: on the one hand, both CLRT and CJ are already very efficient for middle dimension such as $(p,n)=(96,128)$ both in size and power, see Table \[SIZE\] and Table \[POWER\]; and on the other hand, CJ also behaves very well in most of “large $p$, small $n$” situation, see Table \[cnsizepower\].
Three characteristic features emerge from our findings:
1. These asymptotic distributions are universal in the sense that they depend on the distribution of the observations only through its first four moments;
2. The new test procedures improve quickly when either the dimension $p$ or the sample size $n$ gets large. In particular, for a given sample size $n$, within a wide range of values of $p/n$, higher dimensions $p$ lead to better performance of these corrected test statistics.
3. CJ is particularly robust against the dimension inflation. Our Monte-Carlo study shows that for a small sample size $n=64$, the test is effective for $0<p/n\le 20$.
In a sense, these new procedures have benefited from the “blessings of dimensionality”.
Acknowledgements {#acknowledgements .unnumbered}
================
We are grateful to the associate editor and two referees for their numerous comments that have lead to important modifications of the paper.
[9]{}
<span style="font-variant:small-caps;">Anderson, T.W.</span> (1984). *An introduction to Multivariate Statistical Analysis* (2nd edition). Wiley, New York.
<span style="font-variant:small-caps;">Bai, Z.D. and Silverstein, J.W.</span> (2004). [CLT]{} for linear spectral statistics of large-dimensional sample covariance matrices. *Ann. Probab.*, [**32**]{}, 553–605.
<span style="font-variant:small-caps;">Bai, Z.D. and Yao, J.F.</span> (2005). On the convergence of the spectral empirical process of Wigner matrices. *Bernoulli*, [**11**]{}(6), 1059–1092.
<span style="font-variant:small-caps;">Bai, Z.D., Jiang, D.D., Yao, J.F. and Zheng, S.R.</span> (2009). Corrections to [LRT]{} on large-dimensional covariance matrix by [RMT]{}. *Ann. Statist.*, [**37**]{}, 3822–3840.
<span style="font-variant:small-caps;">Bai, Z.D. and Silverstein, J.W.</span> (2010). *Spectral Analysis of Large Dimensional Random Matrices* (2nd edition). Springer, 20.
<span style="font-variant:small-caps;">Bai, Z.D., Jiang, D.D., Yao, J.F. and Zheng, S.R.</span> (2012). Testing linear hypotheses in high-dimensional regressions. *Statistics*, doi:10.1080/02331888.2012.708031.
<span style="font-variant:small-caps;">Birke, M. and Dette, H.</span> (2005). A note on testing the covariance matrix for large dimension. *Statistics and Probability Letters*, [**74**]{}, 281–289.
<span style="font-variant:small-caps;">Chen, S.X. and Qin, Y.L.</span> (2010). A two-sample test for high-dimensional data with applications to gene-set testing. *Ann. Statist.*, [**38**]{}, 808–835.
<span style="font-variant:small-caps;">Chen, S.X., Zhang, L.X. and Zhong, P.S.</span> (2010). Tests for high-dimensional covariance matrices. *J. Amer. Statist. Assoc.*, [**105**]{}, 810–819.
<span style="font-variant:small-caps;">John, S.</span> (1971). Some optimal multivariate tests. *Biometrika,* [**58**]{}, 123–127.
<span style="font-variant:small-caps;">John, S.</span> (1972). The distribution of a statistic used for testing sphericity of normal distributions. *Biometrika,* [**59**]{}, 169–173.
<span style="font-variant:small-caps;">Jonsson, D.</span> (1982). Some limit theorems for the eigenvalues of a sample covariance matrix. *J. Multivariate Anal.*, [**12**]{}(1), 1–38.
<span style="font-variant:small-caps;">Johnstone, I.M.</span> (2001). On the distribution of the largest eigenvalue in principal components analysis. *Ann. Statist.*, [**29**]{}(2), 295–327.
<span style="font-variant:small-caps;">Johnstone, I.M. and Titterington, D.M.</span> (2009). Statistical challenges of high-dimensional data. *Philos. Trans. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci.*, [**367**]{}, [4237–4253]{}.
<span style="font-variant:small-caps;">Ledoit, O. and Wolf, M.</span> (2002). Some hypothesis tests for the covariance matrix when the dimension is large compared to the sample size. *Ann. Statist.*, [**30**]{}, 1081–1102.
<span style="font-variant:small-caps;">Lytova, A. and Pastur, L.</span> (2009). Central limit theorem for linear eigenvalue statistics of the Wigner and the sample covariance random matrices. *Metrika*, [**69**]{}, 153–172.
<span style="font-variant:small-caps;">Mauchly, J.W.</span> (1940). Test for sphericity of a normal n-variate distribution. *Ann. Mathe. Statist.*, [**11**]{}, 204–209.
<span style="font-variant:small-caps;">Muirhead, R.J.</span> (1982). *Aspects of Multivariate Statistical Theory*. Wiley, New York.
<span style="font-variant:small-caps;">Nagao, H.</span> (1973). On some test criteria for covariance matrix. *Ann. Statist.*, [**1**]{}, 700–709.
<span style="font-variant:small-caps;">Onatski, A., Moreira, M.J. and Hallin, M.</span> (2013). Asymptotic power of sphericity tests for high-dimensional data. *Ann. Statist.*, [**41**]{}(3), 1204–1231.
<span style="font-variant:small-caps;">Pan, G.M. and Zhou, W.</span> (2008). Central limit theorem for signal-to-interference ratio of reduced rank linear receiver. *Ann. Appl. Probab.*, [**18**]{}, 1232-1270.
<span style="font-variant:small-caps;">Pan, G.M.</span> (2012). Comparision between two types of large sample covariance matrices. To appear in *Ann. Institut Henri Poincaré*.
<span style="font-variant:small-caps;">Schott, J.R.</span> (2007). Some high-dimensional tests for a one-way [MANOVA]{}. *J. Multivariate Anal.*, [**98**]{}, 1825–1839.
<span style="font-variant:small-caps;">Srivastava, M.S.</span> (2005). Some tests concerning the covariance matrix in high dimensional data. *J. Japan Statist. Soc.*, [**35**]{}, 251–272.
<span style="font-variant:small-caps;">Srivastava, M.S. and Fujikoshi, Y.</span> (2006). Multivariate analysis of variance with fewer observations than the dimension. *J. Multiv. Anal.*, [**97**]{}, 1927–1940.
<span style="font-variant:small-caps;">Srivastava, M.S., Kollo, T. and Rosen, D.</span> (2011). Some tests for the covariance matrix with fewer observations than the dimension under non-normality. *J. Multiv. Anal.*, [**102**]{}, 1090–1103.
<span style="font-variant:small-caps;">Sugiura, N.</span> (1972). Locally best invariant test for sphericity and the limiting distributions. *Ann. Mathe. Statist.*, [**43**]{}, 1312-1316.
<span style="font-variant:small-caps;">Wang, Q.W., Silverstein, J.W. and Yao, J.F.</span> (2013). A note on the CLT of the LSS for sample covariance matrix from a spiked population model. *Preprint*, available at `arXiv:1304.6164`.
<span style="font-variant:small-caps;">Zheng, S.R.</span> (2012). Central limit theorem for linear spectral statistics of large dimensional $F$ matrix. *Ann. Institut Henri Poincaré Probab. Statist.*, [**48**]{}, 444–476.
<span style="font-variant:small-caps;">Zheng, S.R. and Bai, Z.D.</span> (2013). A note on central limit theorems for linear spectral statistics of large dimensional $F$ Matrix. *Preprint*, available at `arXiv:1305.0339`.
Formula for limiting parameters in the CLT for eigenvalues of a sample covariance matrix with general fourth moments
====================================================================================================================
Given a sample covariance matrix $S_n$ of dimension $p$ with eigenvalues $\lambda_1,\ldots,\lambda_p$, linear spectral statistics of the form $F_n(g)=p^{-1} \sum_{i=1}^{p} g(\lambda_j)$ for suitable functions $g$ are of central importance in multivariate analysis. Such CLT’s have been successively developed since the pioneering work of [@r12], see [@r2] and [@r16] for a recent account on the subject.
The CLT in [@r2] (see also an improved version in [@r5]) has been widely used in applications as this CLT also provides, for the first time, explicit formula for the mean and covariance parameters of the normal limiting distribution. In the special case with an array $\{x_{ij}\}$ of independent variables, this CLT assumes the following moment conditions:
1. For each $n$, $x_{ij}=x_{ij}^{n},i\leqslant p,j\leqslant n$ are independent.
2. $Ex_{ij}=0, E|x_{ij}|^{2}=1, \max_{i,j,n}E|x_{ij}|^{4} <\infty$.
3. If $\{x_{ij}\}$’s are real, then $Ex_{ij}^{4}=3$; otherwise (complex variables), $Ex^{2}_{ij}=0$ and $E|x_{ij}|^{4}=2$.
In Condition (c), the fourth moments of the entries are set to the values 3 or 2 matching the normal case. This is indeed a quite demanding and restrictive condition since in the real case for example, it is incredibly hard to find a non-normal distribution with mean 0, variance 1 and fourth moment equaling 3. As a consequence, most of if not all applications published in the literature using this CLT assumes a normal distribution for the observations. Recently, effort have been made in [@r21], [@r16] and [@r29] to overcome these moment restrictions. We present below such a CLT with general forth moments that will be used for the sphericity test.
In all the following, we use an indicator $\kappa$ set to 2 when $\{x_{ij}\}$ are real and to 1 when they are complex. Define $\beta=E|x_{ij}|^{4}-1-\kappa$ for both cases and $h=\sqrt y$.
For the presentation of the results, let be the sample covariance matrix $S_n =\frac1n \sum_{i=1}^n X_i X_i^*$ where $X_i=
(x_{ki})_{1\le k\le p}$ is the $i$-th observed vector. It is then well-known that when $p\to\infty$, $n\to\infty$ and $p/n\to y >0~,$ the distribution of its eigenvalues converges to a nonrandom distribution, namely the Marčenko-Pastur distribution $F^y$ with support $[a,b]=[(1\pm\sqrt y)^2]$ (an additional mass at the origin when $y>1$). Moreover, the Stieltjes transform ${\underline{m}}$ of a companion distribution defined by $\underline F^y = (1-y)\delta_0+yF_c$ satisfies an inverse equation for $z\in \mathbb{C}^+$, $$\label{eq:z}
z = -\frac1{\underline{m}}+ \frac{y}{1+{\underline{m}}}.$$
The following CLT is a particular instance of Theorem 1.4 in [@r21].
\[t1\]\[[@r21]\] Assume that for each $n$, the variables $x_{ij}=x_{ij}^{n},i\leqslant p,j\leqslant n$ are independent and identically distributed satisfying $Ex_{ij}=0$, $E|x_{ij}|^{2}=1$, $E|x_{ij}|^{4}=\beta+1+\kappa <\infty$ and in case of they are complex, $Ex^{2}_{ij}=0$. Assume moreover, $$p\to\infty, ~~n\to\infty, ~~ p/n\to y >0~.$$ Let $f_{1},\cdots f_{k}$ be functions analytic on an open region containing the support of $F^y$. The random vector $\{ X_{n}(f_{1}),\cdots X_{n}(f_{k})\}$ where $$X_n(f) = p\left\{ F_n (f) - F^{y_n}(f) \right\}$$ converges weakly to a normal vector $(X_{f_{1}},\cdots X_{f_{k}})$ with mean function and covariance function: $$\begin{aligned}
{\mathbb{E}}[X_f]& =& (\kappa-1) I_1(f) + \beta I_2(f) ~,
\label{eq:E}\\
{\mathop{\text{Cov}}}(X_{f},X_{g}) &=& \kappa J_1(f,g) + \beta
J_2(f,g)~,
\label{eq:cov}
\end{aligned}$$ where $$\begin{aligned}
I_1(f) &=& - {{\frac1{2\pi i}}}\oint
\frac{y\left\{{\underline{m}}/(1+{\underline{m}})\right\}^3(z)f(z)}
{ \left[1-y \left\{ {\underline{m}}/(1+{\underline{m}})\right\}^2 \right]^2}dz~,
\\
I_2(f) &=&
-\frac{1}{2\pi i}\oint
\frac{y\left\{{\underline{m}}/(1+{\underline{m}})\right\}^3(z)f(z)}
{ 1-y \left\{ {\underline{m}}/(1+{\underline{m}})\right\}^2 }dz~, \\
J_1(f,g) &=&
-\frac{1}{4\pi^{2}}
\oint\oint\frac{f(z_{1})g(z_{2})}{(\underline{m}(z_{1})-\underline{m}(z_{2}))^{2}}\underline{m}'(z_{1})\underline{m}'(z_{2})dz_{1}dz_{2}~,\\
J_2(f,g) &=& \frac{-y}{4\pi^2}
\oint f(z_1) \frac{\partial}{\partial z_1}
\left\{ \frac{{\underline{m}}}{1+{\underline{m}}} (z_1)\right\} dz_1
\cdot
\oint g(z_2) \frac{\partial}{\partial z_2}
\left\{ \frac{{\underline{m}}}{1+{\underline{m}}} (z_2)\right\} dz_2~,
\end{aligned}$$ where the integrals are along contours (non overlapping in $J_1$) enclosing the support of $F^y$.
However, concrete applications of this CLT are not easy since the limiting parameters are given through those integrals on contours that are only vaguely defined. The purpose of this appendix is to go a step further by providing alternative formula for these limiting parameters. These new formulas, presented in the following Proposition convert all these integral along the unit circle; they are much easier to use for concrete applications, see for instance the proofs of Lemma \[lem:joint\] and \[cnlemma:joint\] in the paper. Furthermore, these formulas will be of independent interest for applications other than those developed in this paper.
\[t2\] The limiting parameters in Theorem \[t1\] can be expressed as following: $$\begin{aligned}
I_1(f) &=& \lim_{r\downarrow 1} I_1(f,r)~,\label{I1}\\
I_2(f) &=& \frac{1}{2\pi i}\oint_{|\xi|=1}f(|1+h\xi|^{2})\frac{1}{\xi^{3}}d\xi~,\label{I2}\\
J_1(f,g) &=& \lim_{r\downarrow 1} J_1(f,g,r)~, \label{J1}\\
J_2(f,g)
&=&-\frac{1}{4\pi^{2}}\oint_{|\xi_{1}|=1}\frac{f(|1+h\xi_{1}|^{2})}{\xi_{1}^{2}}d\xi_{1}\oint_{|\xi_{2}|=1}\frac{g(|1+h\xi_{2}|^{2})}{\xi_{2}^{2}}d\xi_{2}~,\label{J2}
\end{aligned}$$ with $$\begin{aligned}
I_1(f,r) &=& \frac1{2\pi i}\oint_{|\xi|=1}f(|1+h\xi|^{2})[\frac{\xi}{\xi^{2}-r^{-2}}-\frac{1}{\xi}]d\xi
~,\\
J_1(f,g,r) &=&
-\frac{1}{4\pi^{2}}\oint_{|\xi_{1}|=1}\oint_{|\xi_{2}|=1}
\frac{f(|1+h\xi_{1}|^{2})g(|1+h\xi_{2}|^{2})} {(\xi_{1}-r\xi_{2})^{2}}d\xi_{1}d\xi_{2}~.
\end{aligned}$$
We start with the simplest formula $I_2(f)$ to explain the main argument and indeed, the other formula are obtained similarly. The idea is to introduce the change of the variable $z=1+hr\xi+h r^{-1}\overline\xi+h^2$ with $r>1$ but close to 1 and $|\xi|=1$ (recall $h=\sqrt y$). Note that this idea has been employed in [@r29]. It can be readily checked that when $\xi$ runs counterclockwisely the unit circle, $z$ will run a contour $\cal C$ that encloses closely the support interval $[a,b]=[(1\pm h)^2]$ (recall $h=\sqrt y$). Moreover, by the Eq. , we have on $\cal C$ $${\underline{m}}= -\frac{1}{1+hr\xi},
\quad \text{and~~} dz = h(r-r^{-1}\xi^{-2})d\xi~.$$ Applying this variable change to the formula of $I_2(f)$ given in Theorem \[t1\], we have $$\begin{aligned}
I_2(f) &=& \lim_{r\downarrow 1}
{{\frac1{2\pi i}}}\oint_{|\xi|=1} f(z) \frac{1}{\xi^3}\frac{r\xi^2-r^{-1}}{r(r^2\xi^2-1)}d\xi\\
&=& {{\frac1{2\pi i}}}\oint_{|\xi|=1} f(|1+h\xi|^2) \frac{1}{\xi^3}d\xi~.
\end{aligned}$$ This proves the formula . For , we have similarly $$\begin{aligned}
I_1(f) &=& \lim_{r\downarrow 1}
{{\frac1{2\pi i}}}\oint_{|\xi|=1} f(z)
\frac{1}{\xi^3}\frac{r\xi^2-r^{-1}}{r(r^2\xi^2-1)} \frac1{1-r^{-2}\xi^{-2}}d\xi\\
&=& \lim_{r\downarrow 1}
{{\frac1{2\pi i}}}\oint_{|\xi|=1} f(|1+h\xi|^2) \frac1{\xi(\xi^2-r^{-2})}\\
&=& \lim_{r\downarrow 1} I_1(f,r)~.
\end{aligned}$$ Formula for $J_2(f,g)$ is calculated in a same fashion by observing that we have $$\frac{\partial}{\partial z}
\left\{ \frac{{\underline{m}}}{1+{\underline{m}}} (z)\right\} dz
= \frac{\partial}{\partial \xi}
\left\{ \frac{{\underline{m}}}{1+{\underline{m}}} (\xi)\right\} d\xi
= \frac{\partial}{\partial \xi}
\left\{ \frac{1}{-hr\xi} \right\} d\xi
= \frac{1}{hr\xi^2}d\xi~.$$ Finally for , we use two non-overlapping contours defined by $z_j= 1+hr_j\xi_j+h r_j^{-1}\overline\xi_j+h^2 $, $j=1,2$ where $r_2>r_1>1$. By observing that $${\underline{m}}'(z_j)dz_j = \left(\frac{\partial}{\partial
\xi_j}{\underline{m}}\right)d\xi_j =
\frac{hr_j}{(1+hr_j\xi_j)^2} d\xi_j~,$$ we find $$\begin{aligned}
J_1(f,g) &=&
\lim_{ {\footnotesize
\begin{array}{c}r_2>r_1>1 \\[-1.5mm] r_2\downarrow 1
\end{array}}
}
-\frac1{4\pi^2}
\oint_{|\xi_1|=1}\!\oint_{|\xi_2|=1}
\frac{f(z_1)g(z_2)} { \left\{ {\underline{m}}(z_1)-{\underline{m}}(z_2) \right\}^2 }\\
&&\cdot \frac{hr_1}{(1+hr_1\xi_1)^2}
\cdot \frac{hr_2}{(1+hr_2\xi_2)^2} d\xi_1d\xi_2 \\
&=&
\lim_{ {\footnotesize
\begin{array}{c}r_2>r_1>1, \\[-1.5mm] r_2\downarrow 1
\end{array}}
}
-\frac1{4\pi^2}
\oint_{|\xi_1|=1}\!\oint_{|\xi_2|=1}
\frac{f(z_1)g(z_2)}{\left\{ r_1\xi_1-r_2\xi_2\right\}^2 } d\xi_1d\xi_2
\\
&=& \lim_{r\downarrow 1}
-\frac1{4\pi^2}
\oint_{|\xi_1|=1}\!\oint_{|\xi_2|=1}
\frac{f(|1+h\xi_1|^2)g(|1+h\xi_2|^2)}{\left\{ \xi_1-r\xi_2\right\}^2 } d\xi_1d\xi_2~.
\end{aligned}$$ The proof is complete.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study partition functions of low-energy effective theories of M2-branes, whose type IIB brane constructions include orientifolds. We mainly focus on circular quiver superconformal Chern-Simons theory on $S^3$, whose gauge group is $O(2N+1)\times USp(2N)\times \cdots \times O(2N+1)\times USp(2N)$. This theory is the natural generalization of the $\mathcal{N}=5$ ABJM theory with the gauge group $O(2N+1)_{2k} \times USp(2N)_{-k}$. We find that the partition function of this type of theory has a simple relation to the one of the M2-brane theory without the orientifolds, whose gauge group is $U(N)\times \cdots \times U(N)$. By using this relation, we determine an exact form of the grand partition function of the $O(2N+1)_{2} \times USp(2N)_{-1}$ ABJM theory, where its supersymmetry is expected to be enhanced to $\mathcal{N}=6$. As another interesting application, we discuss that our result gives a natural physical interpretation of a relation between the grand partition functions of the $U(N+1)_4 \times U(N)_{-4}$ ABJ theory and $U(N)_2 \times U(N)_{-2}$ ABJM theory, recently conjectured by Grassi-Hatsuda-Mariño. We also argue that partition functions of $\hat{A}_3$ quiver theories have representations in terms of an ideal Fermi gas systems associated with $\hat{D}$-type quiver theories and this leads an interesting relation between certain $U(N)$ and $USp(2N)$ supersymmetric gauge theories.'
author:
- |
Masazumi Honda[^1]\
\
\
\
title: 'Exact relations between M2-brane theories with and without Orientifolds'
---
WIS/10/15-NOV-DPPA
Introduction
============
In a couple of years, there is remarkable progress in understanding non-perturbative effects in M-theory through gauge/gravity duality. Most important tools in this progress are the supersymmetry localization [@Pestun:2007rz; @Kapustin:2009kz] and Fermi gas approach [@Marino:2011eh]. These are applied to partition functions in a class of low-energy effective theories of $N$ M2-branes on $S^3$ and it has turned out that the partition functions are described by an ideal Fermi gas system:
Z(N) =\_[S\_N]{} (-1)\^d\^N x \_[j=1]{}\^N (x\_j ,x\_[(j)]{}) , \[eq:Fermi\]
where $\rho$ plays an role of the density matrix in the Fermi gas system. Thanks to these techniques, now we know detailed structures of the non-perturbative effects in M-theory on $AdS_4\times S^7/\mathbb{Z}_k$ [@Hatsuda:2013oxa; @Matsumoto:2013nya; @Honda:2014npa; @Hatsuda:2013yua], which is dual to the 3d $\mathcal{N}=6$ superconformal Chern-Simons (CS) theory known as the ABJ(M) theory [@Aharony:2008ug; @Aharony:2008gk] via AdS/CFT correspondence (see also important earlier works [@Marino:2009jd; @Drukker:2009hy; @Herzog:2010hf; @Fuji:2011km; @Okuyama:2011su; @Hanada:2012si; @Hatsuda:2012hm; @Hatsuda:2012dt; @Awata:2012jb; @Hatsuda:2013gj; @Honda:2013pea]).
On the other hand, we still do not have detailed understanding of the non-perturbative effects “beyond ABJ(M) theory”, namely more general M2-brane theories with less supersymmetry[^2] (SUSY). For instance, it is unclear whether many attractive features found in the ABJ(M) theory such as the Airy functional behavior [@Fuji:2011km; @Marino:2011eh], pole cancellation [@Hatsuda:2012dt; @Hatsuda:2013oxa] and correspondence to topological string [@Marino:2009jd; @Hatsuda:2012hm; @Hatsuda:2013oxa] are universal for general M2-brane theories or accidental for the ABJ(M) theory. While the Airy functional behavior has been found for a broad class of M2-brane theories [@Marino:2011eh; @Marino:2012az; @Mezei:2013gqa; @Assel:2015hsa; @Moriyama:2015jsa] and seems universal [@Bhattacharyya:2012ye] (see also [@Dabholkar:2014wpa]), the other features have been found in few examples. This problem has been addressed in special cases of Imamura-Kimura type theory [@Imamura:2008nn], whose type IIB brane construction consists of NS5-branes and $(1,k)$-5 branes connected by $N$ D3-branes. Especially the orbifold ABJM theory and $(p,q)$ model [@Gaiotto:2008sd] have been studied well in [@Mezei:2013gqa; @Grassi:2014vwa; @Hatsuda:2014vsa; @Moriyama:2014gxa; @Moriyama:2014nca; @Hatsuda:2015lpa]. Also $\hat{D}$-type quiver theories [@Assel:2015hsa; @Moriyama:2015jsa] and $O$ or $USp$ gauge theories with single node [@Mezei:2013gqa; @Okuyama:2015auc] have been studied (see also [@Gulotta:2012yd]). In order to understand the non-perturbative effects in more detail, it is very important to investigate the non-perturbative effects in various theories of M2-branes.
In this paper we consider a generalization along a different direction. We study partition functions of low-energy effective theories of M2-branes on $S^3$, whose type IIB brane constructions include orientifolds. We mainly focus on 3d superconformal CS theory of circular quiver type with the gauge group[^3] $O(2N+1)\times USp(2N)\times \cdots \times O(2N+1)\times USp(2N)$. This theory is a natural generalization of the $O(2N+1)_{2k} \times USp(2N)_{-k}$ ABJM theory with $\mathcal{N}=5$ SUSY [@Hosomichi:2008jb; @Aharony:2008gk]. We show that the $S^3$ partition function of this type of theory is also described by an ideal Fermi gas system as in and its density matrix $\rho_{O(2N+1)\times USp(2N)}$ takes the following form
\_[O(2N+1)USp(2N)]{} (x,y) =\_[U(N)]{}\^[(-)]{} (x,y) , \[eq:rho\_project\]
where
\^[()]{} (x,y) = .
Here $\rho_{U(N)}$ is the density matrix associated with the M2-brane theories without the orientifolds, which are obtained by the replacement $O(2N+1),USp(2N) \rightarrow U(N)$ in the orientifold theories. This indicates that the density matrix for the orientifold theory is the projection of the one without the orientifolds. Introducing the grand canonical partition function by
= \_N Z(N) e\^[N]{} = [Det]{}(1+e\^) ,
the relation indicates that the grand partition function of the orientifold theory is related to the one of the non-orientifold theory by
\_[O(2N+1)USp(2N)]{} \[\] =\_[U(N)]{}\^[(-)]{} \[\] , \[eq:main\]
where $\Xi^{(\pm )} [\mu ]$ denotes the grand canonical partition function defined by $\rho^{(\pm )}$. This relation implies that we can obtain non-perturbative information on the orientifold theory from the non-orientifold theory.
Here we present two interesting applications of our main result . One of them is to determine an exact form of the grand partition function of the $O(2N+1)_{2k} \times USp(2N)_{-k}$ ABJM theory with $k=1$, whose SUSY is expected to be enhanced to $\mathcal{N}=6$ from $\mathcal{N}=5$. This is achieved by combining our result with recent results of [@Grassi:2014uua; @Codesido:2014oua] and we obtain
\_[O(2N+1)\_2 USp(2N)\_[-1]{}]{} () = \_[U(N)\_1 U(N)\_[-1]{}]{} ( /2 +i/2 )\_[U(N)\_1 U(N)\_[-1]{}]{} ( /2 -i/2 ) . \[eq:exactk1\]
Here $\Xi_{U(N)_1 \times U(N)_{-1}}$ is the grand partition function of the $U(N)_1 \times U(N)_{-1}$ ABJM theory, whose exact form is conjectured as [@Codesido:2014oua]
&&\_[U(N)\_1 U(N)\_[-1]{}]{} ()\
&&=\
&&( \_2 (|/4 ,|/4 ) +i\_1 (|/4 ,|/4 ) ) , \[eq:k1ABJM\]
where several definitions will be given in sec. \[sec:exact\].
The other application of is to give a natural physical interpretation of a mysterious relation recently conjectured by Grassi-Hatsuda-Mariño [@Grassi:2014uua]. They conjectured a relation between the grand partition functions of the $U(N+1)_4 \times U(N)_{-4}$ ABJ theory and $U(N)_2 \times U(N)_{-2}$ ABJM theory as
\_[U(N)\_4 U(N+1)\_[-4]{}ABJ]{} \[\] =\_[U(N)\_2 U(N)\_[-2]{}ABJM]{}\^[(-)]{} \[\] . \[eq:GHM\]
This should be compared with our result for the $\mathcal{N}=5$ ABJM theory:
\_[O(2N+1)\_[2k]{}USp(2N)\_[-k]{}]{} \[\] =\_[U(N)\_[2k]{}U(N)\_[-2k]{}]{}\^[(-)]{} \[\] . \[eq:grandABJM\]
Combining with , we find
\_[U(N)\_4 U(N+1)\_[-4]{}ABJ]{} \[\] =\_[O(2N+1)\_[2]{}USp(2N)\_[-1]{}]{} \[\] . \[eq:enhance\]
Remarkably this relation is indeed equivalent to the conjecture in [@Aharony:2008gk]. The $\mathcal{N}=5$ ABJM theory is expected to be low energy effective theories of $N$ M2-branes probing $\mathbb{C}^4 /\hat{D}_k$ with the binary dihedral group $\hat{D}_k$ defined in . Since $\mathbb{C}^4 /\hat{D}_k$ for $k=1$ is $\mathbb{C}^4 /\mathbb{Z}_4$, moduli of the $O(2N+1)_{2} \times USp(2N)_{-1}$ ABJM theory become the same as the one of the $U(N+M)_k \times U(N)_{-k}$ ABJ(M) with $k=4$. Therefore the work [@Aharony:2008gk] conjectured that the $O(2N+1)_{2} \times USp(2N)_{-1}$ ABJM theory has the enhanced $\mathcal{N}=6$ SUSY and equivalent to the $U(N+1)_4 \times U(N)_{-4}$ ABJ theory[^4]:
O(2N+1)\_[2]{} USp(2N)\_[-1]{} U(N+1)\_4 U(N)\_[-4]{} , \[eq:isomorphism\]
which gives[^5] . If we assume this, then our result leads us to the Grassi-Hatsuda-Mariño relation , while if we assume , then our result indicates the conjectural equivalence .
We also discuss that partition functions of $\hat{A}_3$ quiver theories have representations in terms of ideal Fermi gas systems associated with $\hat{D}$-type quivers[^6] and this leads an interesting relation between certain $U(N)$ and $USp(2N)$ SUSY gauge theories with single node. The $U(N)$ gauge theory under consideration is $\mathcal{N}=4$ vector multiplet with one adjoint hyper multiplet and $N_f$ fundamental hyper multiplets, while the $USp(2N)$ gauge theory is $\mathcal{N}=4$ vector multiplet with one anti-symmetric hyper multiplet and $N_f$-fundamental hyper multiples. Regarding these theories, the work [@Okuyama:2015auc] has proposed the equivalence
Z\_[U(N)+adj.]{} (N,N\_f =4 ) =Z\_[USp(2N)+A]{}(N,N\_f =3 ) . \[UequalUSp\]
This relation is expected from 3d mirror symmetry[^7] [@Intriligator:1996ex; @deBoer:1996mp]. It is known that the $U(N)$ and $USp(2N)$ theories are equivalent to $\hat{A}_{N_f -1}$ and $\hat{D}_{N_f }$ quiver theories without CS terms, where only one of the vector multiplets is coupled to one fundamental hyper multiplet. Since $\hat{A}_3 =\hat{D}_3$, should hold via the 3d mirror symmetries. In appendix we explicitly prove this relation by using the technique in [@Assel:2015hsa; @Moriyama:2015jsa].
This paper is organized as follows. In sec. \[sec:ABJM\], we consider the $\mathcal{N}=5$ ABJM theory with the gauge group $O(2N+1)_{2k}\times USp(2N)_{-k}$. In sec. \[sec:general\], we generalize our analysis in sec. \[sec:ABJM\] to more general quiver gauge theories. We also identify quantum mechanical operators in ideal Fermi gas systems naturally corresponding to orientifolds in type IIB brane constructions. As interesting examples, we deal with orientifold projections of the $(p,q)$ model and orbifold ABJM theory. Section \[sec:conclusion\] is devoted to conclusion and discussions. In appendix, we explicitly prove the equivalence .
$O(2N+1)_{2k} \times USp(2N)_{-k}$ ABJM theory {#sec:ABJM}
==============================================
In this section we consider the $\mathcal{N}=5$ ABJM theory with the gauge group $O(2N+1)_{2k}\times USp(2N)_{-k}$. We will generalize our analysis in this section to more general theory in next section.
Orientifold ABJM theory as a Fermi gas {#sec:rhoABJM}
--------------------------------------
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Multiplet One-loop determinant
----------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------
$\mathcal{N}=2$ $O(2N+1)$ vector multiplet $ \prod_{i<j} \Bigl[ 2\sinh{\frac{\mu_i -\mu_j}{2}}\cdot 2\sinh{\frac{\mu_i + \mu_j}{2}} \Bigr]^2 \prod_{j=1}^N 4\sinh^2{\frac{\mu_j}{2}}$
$\mathcal{N}=2$ $USp(2N)$ vector multiplet $\prod_{i<j} \Bigl[ 2\sinh{\frac{\nu_i -\nu_j}{2}}\cdot 2\sinh{\frac{\nu_i + \nu_j}{2}} \Bigr]^2 \prod_{j=1}^N 4\sinh^2{\nu_j}$
$O(2N+1)\times USp(2N)$ bi-fund. chiral mult. $\left( \prod_{i,j}\Bigl[ 2\cosh{\frac{\mu_i -\nu_j}{2}}\cdot 2\cosh{\frac{\mu_i + \nu_j}{2}}
\Bigr] \prod_j 2\cosh{\frac{\nu_j}{2}} \right)^{-1}$
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: One-loop determinant of each multiplet in the localization of the $O(2N+1)_{2k} \times USp(2N)_{-k}$ ABJM theory on $S^3$. []{data-label="tab:1loop"}
Thanks to the localization [@Kapustin:2009kz], the partition function of the $O(2N+1)_{2k} \times USp(2N)_{-k}$ ABJM theory on $S^3$ can be written as[^8] (see tab. \[tab:1loop\] for detail) $$\begin{aligned}
Z_{\mathcal{N}=5 \rm ABJM}(N) \nonumber
&=& \frac{1}{2^{2N}N!^2} \int \frac{d^N \mu}{(2\pi )^N}
\frac{d^N \nu}{(2\pi )^N}
e^{\frac{ik}{2\pi}\sum_{j=1}^N (\mu_j^2 -\nu_j^2 )}
\prod_{j=1}^N 4\sinh^2{\frac{\mu_j}{2}}\cdot 4\sinh^2{\nu_j} {\nonumber}\\
&&\times
\frac{\prod_{i<j} \Bigl[
2\sinh{\frac{\mu_i -\mu_j}{2}}\cdot 2\sinh{\frac{\mu_i + \mu_j}{2}}\cdot
2\sinh{\frac{\nu_i -\nu_j}{2}}\cdot 2\sinh{\frac{\nu_i + \nu_j}{2}}
\Bigr]^2}
{\prod_{i,j}\Bigl[
2\cosh{\frac{\mu_i -\nu_j}{2}}\cdot 2\cosh{\frac{\mu_i + \nu_j}{2}} \Bigr]^2
\prod_j 4\cosh^2{\frac{\nu_j}{2}}} .\end{aligned}$$ Now we write $Z_{\mathcal{N}=5 \rm ABJM}$ in terms of an ideal Fermi gas as in circular quiver $U(N)$ SUSY gauge theories [@Marino:2011eh; @Okuyama:2011su]. For this purpose we use the Cauchy determinant-like formula [@Mezei:2013gqa]
&&\
&& = \_[S\_N]{} (-1)\^\_j , \[eq:Cauchy\]
and rewrite the partition function as $$\begin{aligned}
&&Z_{\mathcal{N}=5 \rm ABJM}(N) \\\nonumber
&&= \frac{1}{2^{2N}N!} \sum_{\sigma\in S_N} (-1)^\sigma
\int \frac{d^N \mu}{(2\pi )^N} \frac{d^N \nu}{(2\pi )^N}
e^{\frac{ik}{2\pi}\sum_{j=1}^N (\mu_j^2 -\nu_j^2 )} \prod_{j=1}^N 4\sinh^2{\frac{\mu_j}{2}}\cdot 4\sinh^2{\nu_j} {\nonumber}\\
&&\times
\prod_j \frac{1}{2\cosh{\left( \frac{\mu_j -\nu_j}{2} \right) } \cdot
2\cosh{\left( \frac{\mu_j +\nu_j}{2} \right) } \cdot
2\cosh{\left( \frac{\mu_j -\nu_{\sigma (j)}}{2} \right) } \cdot
2\cosh{\left( \frac{\mu_j +\nu_{\sigma (j)}}{2} \right) } 4\cosh^2{\frac{\nu_j}{2}} } {\nonumber}\\
&=& \frac{1}{N!} \sum_{\sigma\in S_N} (-1)^\sigma \int d^N \mu \prod_j \rho_{\mathcal{N}=5 \rm ABJM} (\mu_j ,\mu_{\sigma (j)} ) ,\end{aligned}$$ where
\_[=5 ABJM]{} (x,y) &=&\
&& ,\
with $k'=2k$. This equation tells us that the partition function of the $\mathcal{N}=5$ ABJM theory is described by the ideal Fermi gas system with the density matrix $\rho_{\mathcal{N}=5 \rm ABJM} (x,y)$. We regard $\rho_{\mathcal{N}=5 \rm ABJM} (x,y)$ as the matrix element of a quantum mechanical operator as in [@Marino:2011eh],
\_[=5 ABJM]{} (x,y) = x| \_[=5 ABJM]{}( ,) |y,
where
\[ , \] = i,=2k’ =4k . \[eq:hbar\]
The operator $\hat{\rho}_{\mathcal{N}=5 \rm ABJM} $ is defined as
\_[=5 ABJM]{} ( ,) &=& e\^[\^2]{} e\^[-\^2]{}\
&=&e\^[\^2]{} e\^[-\^2]{} , \[eq:rhoABJM\]
where $\hat{R}|x\rangle = |-x\rangle$ and we have used
=x| |y. \[eq:minus\]
By using the operator equations[^9] $e^{\frac{i}{2\hbar}\hat{Q}^2} f(\hat{P})e^{-\frac{i}{2\hbar}\hat{Q}^2}=f(\hat{P}-\hat{Q})$ and $e^{\frac{i}{2\hbar}\hat{P}^2} g(\hat{Q})e^{-\frac{i}{2\hbar}\hat{P}^2}=g(\hat{Q}+\hat{P})$, we simplify $\hat{\rho}_{\mathcal{N}=5 \rm ABJM}$ as
\_[=5 ABJM]{} ( ,) = = e\^[\^2]{} e\^[-\^2]{} .
Performing the similarity transformation
\_[=5 ABJM]{} ( ,) ( e\^[\^2]{}) \_[=5 ABJM]{} ( ,) ( e\^[\^2]{})\^[-1]{} , \[eq:rhoN5ABJ\]
we obtain the following highly simplified expression
\_[=5 ABJM]{}( ,) = .
Recalling that $\hat{\rho}$ for the $\mathcal{N}=6$ ABJM theory with the gauge group $U(N)_{2k} \times U(N)_{-2k}$ is given by[^10]
\_[=6ABJM]{}( ,) = ,
we finally obtain
\_[=5 ABJM]{}( ,) = \_[=6 ABJM]{}( ,) . \[eq:rho\_final\]
This indicates that the density matrix operator $\hat{\rho}_{\mathcal{N}=5 \rm ABJM}$ of the $\mathcal{N}=5$ ABJM theory is the projection of the one of the $\mathcal{N}=6$ ABJM theory. Since the $\mathcal{N}=5$ ABJM theory is the orientifold projection of the $\mathcal{N}=6$ ABJM theory, presumably the operation of $(1-\hat{R})/2$ to $\hat{\rho}_{\mathcal{N}=6 \rm ABJM}$ corresponds to the orientifold projection. It is interesting if one can understand this relation more precisely.
### Remarks {#remarks .unnumbered}
1. The representation of $\hat{\rho}_{\mathcal{N}=5 \rm ABJM}$ gives the matrix element
\_[=5 ABJM]{} (x,y) = .
This gives the following representation of the partition function
Z\_[=5 ABJM]{} (N,k) = \_[j=1]{}\^N , \[eq:N5ABJM2\]
where we have rescaled as $x\rightarrow 2x$. Let us compare this with the partition function of the $USp(2N)$ gauge theory with $\mathcal{N}=4$ vector multiplet, one symmetric hyper multiplet and $N_f$-fundamental hyper multiples[^11] (called $USp+S$ theory in [@Mezei:2013gqa]):
&&Z\_[USp+S]{}(N, N\_f )\
&&= \_[j=1]{}\^N\
&&= \_[j=1]{}\^N .\
Comparing this with , we easily see that the $\mathcal{N}=5$ ABJM theory with $k=1$ agrees with[^12] the $USp+S$ theory with $N_f =1$:
Z\_[=5 ABJM]{} (N,k=1) = Z\_[USp+S]{}(N, N\_f =1 ) . \[eq:USp+S\]
It is interesting if one can understand this relation by the brane constructions. Note that this result is essentially the same as the recent result in [@Okuyama:2015auc], which has shown the equivalence between the grand partition function of the $USp+S$ theory with $N_f =1$ and $\Xi^{(-)}$ part of the $U(N)_2 \times U(N)_{-2}$ ABJM theory. Because of , our result is equivalent to this result: $\Xi_{O(2N+1)_2\times USp(2N)_{-1}} =\Xi_{U(N)_2 \times U(N)_{-2}}^{(-)} =\Xi_{USp+S} (N_f =1)$.
2. When we identify the quantum mechanical operator associated with $\rho_{\mathcal{N}=5 \rm ABJM} (x,y)$, we could use the following identity once or twice instead of ,
=x| |y. \[eq:idplus\]
Then the partition function $Z_{\mathcal{N}=5 \rm ABJM}$ is described by different representations of $\hat{\rho}$. If we use this identity and just once by once, then we find
\_[=5 ABJM]{}’ &=&\
&=& ,
while if we use twice, then we get
\_[=5 ABJM]{}” = .
To summarize, we have four different representations of $\hat{\rho}$ to describe the same partition function $Z_{\mathcal{N}=5 \rm ABJM}$:
\_[=5 ABJM]{} =e\^[\^2]{} f\_() f\_() e\^[-\^2]{} f\_() f\_() ,
where we can freely choose “$+$” or “$-$” at every “$f_\pm (1\pm \hat{R})f_\pm$” and $f_\pm$ is given by
f\_+ (Q) = , f\_- (Q) = .
In this paper we always choose “$-$” since taking “$-$” seems technically simpler.
Exact grand partition function for $k=1$ {#sec:exact}
----------------------------------------
Here we find the exact form of the grand partition function of the $O(2N+1)_{2k} \times USp(2N)_{-k}$ ABJM theory for $k=1$. Grassi, Hatsuda and Mariño conjectured [@Grassi:2014uua]
\_[U(N)\_2 U(N)\_[-2]{}]{}\^[(-)]{} () = \_[U(N)\_1 U(N)\_[-1]{}]{} ( /2 +i/2 )\_[U(N)\_1 U(N)\_[-1]{}]{} ( /2 -i/2 ) .
Combining this with our result , we immediately find
\_[O(2N+1)\_2 USp(2N)\_[-1]{}]{} () &=& \_[U(N)\_2 U(N)\_[-2]{}]{}\^[(-)]{} ()\
&=& \_[U(N)\_1 U(N)\_[-1]{}]{} ( /2 +i/2 ) \_[U(N)\_1 U(N)\_[-1]{}]{} ( /2 -i/2 ) .
The exact form of the grand partition function $\Xi_{U(N)_1 \times U(N)_{-1}}$ was proposed as [@Codesido:2014oua]
&&\_[U(N)\_1 U(N)\_[-1]{}]{} ()\
&&=\
&& ( \_2 (|/4 ,|/4 ) +i\_1 (|/4 ,|/4 ) ) ,
where $\vartheta_{1,2}$ is the Jacobi theta function[^13] and[^14]
&& = \_3 F\_2 ( ,,; 1,; - ) ,=e\^\
&&\_\_0 () = G\_[3,3]{}\^[2,3]{} ( .
, & , & 0,&0,& -
| -) + \_3 F\_2 ( ,,; 1,; - ) ,\
&& \_\^2 \_0 () = -8\^3 i| ,| = ( \_\^2 \_0 () -\_\_0 () ) ,\
&& \_1 = - -,F\_1\^[NS]{} = -.
In terms of , we can explicitly write the exact form of the grand partition function of the $O(2N+1)_{2} \times USp(2N)_{-1}$ ABJM theory.
Grassi-Hatsuda-Mariño exact functional relation from geometry
-------------------------------------------------------------
Grassi, Hatsuda and Mariño conjectured the relation on the grand canonical partition function of the ABJ theory [@Grassi:2014uua]: $$\Xi_{U(N+1)_4 \times U(N)_{-4} } (\mu ) = \Xi_{U(N)_2 \times U(N)_{-2}}^{(-)} (\mu) .$$ Physical interpretation of this relation has been unclear and therefore this relation has been considered as accidental. Now we give a physical interpretation on this relation. Let us compare this result with our result : $$\Xi_{O(2N+1)_{2k}\times USp(2N)_{-k}} [\mu ] =\Xi_{U(N)_{2k}\times U(N)_{-2k}}^{(-)} [\mu ] .$$ Plugging into leads us to $$\Xi_{U(N)_4 \times U(N+1)_{-4}\rm ABJ} [\mu ]
=\Xi_{O(2N+1)_{2}\times USp(2N)_{-1}} [\mu ] .$$ This relation is equivalent to the conjecture in [@Aharony:2008gk]. The $O(2N+1)_{2k}\times USp(2N)_{-k}$ ABJM theory is expected to be low energy effective theories of $N$ M2-branes probing $\mathbb{C}^4 /\hat{D}_k$ with the binary dihedral group $\hat{D}_k$, whose action to the complex coordinate $(z_1 ,z_2 ,z_3 , z_4)$ of $\mathbb{C}^4$ is
(z\_1 ,z\_2 ,z\_3 ,z\_4) \~e\^(z\_1 ,z\_2 ,z\_3 ,z\_4) \~(iz\_2\^, -iz\_1\^,iz\_4\^,-iz\_3\^) . \[eq:dihedral\]
Since $\mathbb{C}^4 /\hat{D}_k$ for $k=1$ is $\mathbb{C}^4 /\mathbb{Z}_4$, the moduli of the $O(2N+1)_{2} \times USp(2N)_{-1}$ ABJM theory become the same as the one of the $U(N+M)_4 \times U(N)_{-4}$ ABJ(M) theory. Therefore the work [@Aharony:2008gk] conjectured that the $O(2N+1)_{2} \times USp(2N)_{-1}$ ABJM theory has $\mathcal{N}=6$ SUSY and equivalent to the $U(N+1)_4 \times U(N)_{-4}$ ABJ theory (see [@Cheon:2012be] for the test by superconformal index): $$O(2N+1)_{2} \times USp(2N)_{-1}\ \leftrightarrow \ U(N+1)_4 \times U(N)_{-4} .$$ If we assume this, then our result leads us to the Grassi-Hatsuda-Mariño relation , while if we assume the Grassi-Hatsuda-Mariño relation , then our result indicates the conjecture .
Generalization {#sec:general}
==============
In this section we generalize our analysis in sec. \[sec:ABJM\] to a class of CS theory, which is circular quiver with the gauge group $[O(2N+1 ) \times USp(2N) ]^r$ and bi-fundamental chiral multiplets one by one between nearest neighboring pairs of the gauge groups.
Fermi gas formalism {#sec:generalFermi}
-------------------
Let us consider the circular quiver CS theory with the gauge group $O(2N+1)_{2k_1}\times USp (2N)_{k_1'} \times \cdots \times O(2N+1)_{2k_r}\times USp (2N)_{k_r'}$, where $O(2N+1)_{2k_a}$ and $USp(2N+1)_{2k_a'}$ are coupled to $N_f^{(a)}$ and $N_f^{\prime (a)}$ fundamental hyper multiplets, respectively. We parametrize the CS levels $k_a ,k_a'$ as $k_a = k n_a , k_a' = k n_a'$ with rational numbers $n_a$ and $n_a'$. Applying the localization, the partition function becomes [@Kapustin:2009kz] $$\begin{aligned}
&& Z_{O(2N+1)\times USp(2N)}(N) \nonumber \\
&&= \frac{1}{2^{2rN}N!^{2r}} \int \prod_{a=1}^r \frac{d^N \mu^{(a)}}{(2\pi )^N} \frac{d^N \nu^{(a)}}{(2\pi )^N}
\prod_{j=1}^N 4\sinh^2{\frac{\mu_j^{(a)}}{2}} f^{(a)}(\mu_j^{(a)} ) \cdot \frac{\sinh^2{\nu_j^{(a)}}}{\cosh^2{\frac{\nu_j^{(a)}}{2}}}
f^{\prime (a)}(\nu_j^{(a)} ) {\nonumber}\\
&&\ \ \ \times
\frac{\prod_{i<j} \Bigl[
2\sinh{\frac{\mu_i^{(a)} -\mu_j^{(a)}}{2}}\cdot 2\sinh{\frac{\mu_i^{(a)} + \mu_j^{(a)}}{2}}\cdot
2\sinh{\frac{\nu_i^{(a)} -\nu_j^{(a)}}{2}}\cdot 2\sinh{\frac{\nu_i^{(a)} + \nu_j^{(a)}}{2}}
\Bigr]^2}
{\prod_{i,j}
2\cosh{\frac{\mu_i^{(a)} -\nu_j^{(a)}}{2}}\cdot 2\cosh{\frac{\mu_i^{(a)} + \nu_j^{(a)}}{2}} \cdot
2\cosh{\frac{\mu_i^{(a+1)} -\nu_j^{(a)}}{2}}\cdot 2\cosh{\frac{\mu_i^{(a+1)} + \nu_j^{(a)}}{2}}
} ,\end{aligned}$$ where $\mu_i^{(r+1)}=\mu_i^{(1)}$ and[^15]
f\^[(a)]{}(x ) = ,f\^[(a)]{}(x ) = .
By similar arguments to sec. \[sec:rhoABJM\], we rewrite the partition function as $$\begin{aligned}
Z_{O(2N+1)\times USp(2N)} (N) \nonumber
&=& \frac{1}{N!}
\sum_{\sigma\in S_N} (-1)^\sigma
\int d^N \mu \prod_j \rho (\mu_j ,\mu_{\sigma (j)} ) .\end{aligned}$$ Here the function $\rho (x,y)$ is defined by
&&\_[O(2N+1)USp(2N)]{} (x,y)\
&&= ( \_[a=2]{}\^r f\^[(a)]{}(\^[(a)]{} ) ) ( \_[b=1]{}\^r )\
&&\
&& .
By appropriate similarity transformations, we obtain
\_[O(2N+1)USp(2N)]{} ( ,) &=& \_[a=1]{}\^r f\^[(a)]{}( ) f\^[(a)]{}( )\
&=& \_[U(N)]{}( ,) , \[eq:rho\_general\]
where $\hat{\rho}_{U(N)}$ is the density matrix operator associated with the non-orientifold theory, which is obtained by the replacement $O(2N+1),USp(2N) \rightarrow U(N)$ in the orientifold theories. This relation shows that $\hat{\rho}$ for the orientifold theory is the projection of the one without the orientifolds.
Identification of operators corresponding to orientifolds
---------------------------------------------------------
![ The type IIB brane construction of the $(p,q)$ model for $(p,q)=(2,3)$. []{data-label="fig:brane23"}](brane_23model.eps){width="7.4cm"}
Here we identify quantum mechanical operators, which naturally correspond to the orientifolds[^16] $\widetilde{O3}^\pm$ in type IIB brane construction. First, it is known that D5-brane, NS5-brane and $(1,k)$-5 brane naturally correspond to[^17] (see e.g. [@Assel:2014awa; @Drukker:2015awa])
\_[D5]{}= ,\_[NS5]{}= , \_[(1,k)5]{}= e\^ e\^[-]{} = .
This is actually consistent with $\hat{\rho}$ of $\mathcal{N}=3$ circular quiver CS theory with $U(N)$ gauge group and $SL(2,\mathbb{Z})$ symmetry in type IIB string. For example, let us consider the $(p,q)$-model, whose IIB brane construction consists of $p$ NS5-branes and $q$ $(1,k)$-5 branes connected by $N$ D3-branes on a circle (see fig. \[fig:brane23\]). This theory is $\mathcal{N}=3$ circular quiver superconformal CS theory with the gauge group $U(N)_k \times U(N)_0^{q-1} \times U(N)_{-k}^{p-1}$ and $\hat{\rho}_{(p,q)}$ associated with this theory is
\_[(p,q)]{} = \_[(1,k)5]{}\^q \_[NS5]{}\^p = ( )\^q ( )\^p ,
which is equivalent to $\hat{\rho}$ of the $(p,q)$-model by an appropriate canonical transformation.
![ The type IIB brane construction of the $\mathcal{N}=5$ ABJM theory with the gauge group $O(2N+1)_{2k}\times USp(2N)_{-k}$. []{data-label="fig:brane"}](brane_ABJM.eps){width="7.4cm"}
Similarly let us consider the $O(2N+1)_{2k}\times USp(2N)_{-k}$ ABJM theory, whose brane construction is given by $(\widetilde{O3}^+ -D3)-(1,2k)-(\widetilde{O3}^- -D3) -(NS5)$ on a circle (see fig. \[fig:brane\]). As discussed in sec. \[sec:rhoABJM\], $\hat{\rho}$ for the $\mathcal{N}=5$ ABJM theory is
\_[=5ABJM]{} =e\^[\^2]{} e\^[-\^2]{} = \_[(1,2k)]{} \_[NS5]{} .
If we assume that this can be rewritten as
\_[=5ABJM]{} =\_[\^+]{} \_[(1,2k)]{} \_[\^-]{} \_[NS5]{} ,
where $\hat{\mathcal{O}}_{\widetilde{O3}^\pm}$ corresponds to $\widetilde{O3}^\pm$, then it is natural to identify[^18]
\_[\^-]{} = ,\_[\^+]{} = .
This identification is consistent for more general quiver gauge theories described in sec. \[sec:generalFermi\].
Orientifold projection of $(p,q)$-model
---------------------------------------
![ The type IIB brane construction of the orientifold projection of the $(2,4)$ model. []{data-label="fig:brane24P"}](brane_24modelP.eps){width="7.4cm"}
As an interesting example, we consider orientifold projection of the $(p,q)$-model analyzed well in [@Mezei:2013gqa; @Grassi:2014vwa; @Hatsuda:2014vsa; @Moriyama:2014gxa; @Moriyama:2014nca; @Hatsuda:2015lpa]. The $(p,q)$-model is the circular quiver theory with the gauge group $U(N)_k \times U(N)_0^{q-1}\times U(N)_{-k}\times U(N)_0^{p-1}$, whose type IIB brane construction is $[(D3)-(NS5)]^p -[(D3)-(1,k)]^q$. Then let us consider a circular quiver theory with the brane construction (see fig. \[fig:brane24P\]) $$[(\widetilde{O3}^- -D3) -(NS5)-(\widetilde{O3}^+ -D3) -(NS5)]^p - [(\widetilde{O3}^- -D3)-(1,2k)-(\widetilde{O3}^+ -D3)-(1,2k)]^q .$$ Then corresponding $\hat{\rho}$ is
&=& ( \_[\^-]{}\^[(- )]{} \_[NS5]{} \_[\^+]{}\^[(- )]{} \_[NS5]{} )\^p ( \_[\^-]{}\^[(- )]{} \_[(1,2k)5]{} \_[\^+]{}\^[(- )]{} \_[(1,2k)5]{} )\^q\
&=& ( )\^[2p]{} ( )\^[2q]{} =\_[(2p,2q)]{} .
This can be understood as the projection of the $(2p,2q)$-model.
We can also consider the orientifold projection of the $(p,q)$-model with odd $p$ and $q$. For example suppose the brane construction $$[(\widetilde{O3}^- -D3) -(1,2k)-(\widetilde{O3}^+ -D3) -(NS5)] - [(\widetilde{O3}^- -D3)-(NS5)-(\widetilde{O3}^+ -D3)-(NS5)]^m ,$$ which gives
&=& \_[\^-]{}\^[(- )]{} \_[(1,2k)5]{} \_[\^+]{}\^[(- )]{} \_[NS5]{} ( \_[\^-]{}\^[(- )]{} \_[NS5]{} \_[\^+]{}\^[(- )]{} \_[NS5]{} )\^m\
&= & ( )\^[2m+1]{} =\_[(1,2m+1)]{} .
This is the projection of the $(1,2m+1)$-model.
Orientifold projection of orbifold ABJM theory
-----------------------------------------------
![ \[Left\] The type IIB brane construction of the orbifold ABJM theory for $r=2$. \[Right\] Its orientifold projection. []{data-label="fig:brane_orbi"}](brane_orbi2.eps "fig:"){width="7.4cm"} ![ \[Left\] The type IIB brane construction of the orbifold ABJM theory for $r=2$. \[Right\] Its orientifold projection. []{data-label="fig:brane_orbi"}](brane_orbi2P.eps "fig:"){width="7.4cm"}
Next we consider the orientifold projection of the orbifold ABJM theory. Recalling that the brane construction of the orbifold ABJM theory is $[(D3) -(NS5)-(D3) -(1,k)]^r$, let us take the following brane construction (see fig. \[fig:brane\_orbi\]) $$[(\widetilde{O3}^- -D3) -(NS5)-(\widetilde{O3}^+ -D3) -(1,2k)]^r ,$$ which gives the $[O(2N+1)_{2k}\times USp(2N)_{-k}]^r$ circular quiver superconformal CS theory. Then corresponding $\hat{\rho}$ is
\_[\[O(2N+1)\_[2k]{}USp(2N)\_[-k]{}\]\^r]{} &=& \^r\
&= & ( )\^r =( \_[=5ABJM]{} )\^r , \[eq:orbifoldABJM\]
which is the projection of the orbifold ABJM theory. We can express the grand partition function of the orientifold projected orbifold ABJM theory in terms of the one of the $\mathcal{N}=5$ ABJM theory by using the argument in [@Honda:2014ica]. Namely, when the density matrix operator $\hat{\rho}$ satisfies $\hat{\rho}=(\hat{\rho}')^r$, the grand partition function becomes
[Det]{}( 1+e\^) =\_[j=-]{}\^ [Det]{}( 1+’ e\^) ,
independent of detail form of $\hat{\rho}'$. Hence, the relation immediately leads us to[^19]
\_[\[O(2N+1)\_[2k]{}USp(2N)\_[-k]{}\]\^r]{} () =\_[j=-]{}\^ \_[=5ABJM]{} ( ) . \[Jrelation\]
Since we already know the exact form of the grand partition function for the $O(2N+1)_2 \times USp(2N)_{-1}$ by and , we can also explicitly write the one of the orientifold projected orbifold ABJM theory with $k=1$ in terms of .
Comments on $O(2N) \times USp(2N)$ type {#sec:O2NUSp}
---------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------
Multiplet One-loop determinant
--------------------------------------------- -----------------------------------------------------------------------------------------------------
$\mathcal{N}=2$ $O(2N)$ vector multiplet $ \prod_{i<j} \Bigl[ 2\sinh{\frac{\mu_i -\mu_j}{2}}\cdot 2\sinh{\frac{\mu_i + \mu_j}{2}} \Bigr]^2 $
$O(2N)\times USp(2N)$ bi-fund. chiral mult. $\left( \prod_{i,j}\Bigl[ 2\cosh{\frac{\mu_i -\nu_j}{2}}\cdot 2\cosh{\frac{\mu_i + \nu_j}{2}}
\Bigr] \right)^{-1}$
---------------------------------------------------------------------------------------------------------------------------------------------------
: One-loop determinant of each multiplet in the localization of the $O(2N)_{2k} \times USp(2N)_{-k}$ ABJM theory on $S^3$. []{data-label="tab:1loop_2"}
In this section we give some comments on partition functions of $O(2N)_{2k} \times USp(2N)_{-k}\times \cdots \times O(2N)_{2k} \times USp(2N)_{-k}$ type theories, recently studied well in [@Moriyama:2015asx]. The $S^3$ partition function of this theory is technically equivalent to redefinition of $f^{(a)}(x)$ and $f^{\prime (a)}(x)$ in our analysis presented in sec. \[sec:generalFermi\]. For simplicity, let us consider the $\mathcal{N}=5$ ABJM theory with the gauge group $O(2N)_{2k} \times USp(2N)_{-k}$. Applying the localization, the partition function of this theory becomes $$\begin{aligned}
Z_{O(2N)_{2k}\times USp(2N)_{-k}} \nonumber
&=& \frac{1}{2^{2N}N!^2} \int \frac{d^N \mu}{(2\pi )^N} \frac{d^N \nu}{(2\pi )^N}
e^{\frac{ik'}{2\pi}\sum_{j=1}^N (\mu_j^2 -\nu_j^2 )} \prod_{j=1}^N 4\sinh^2{\nu_j} {\nonumber}\\
&&\times
\frac{\prod_{i<j} \Bigl[
2\sinh{\frac{\mu_i -\mu_j}{2}}\cdot 2\sinh{\frac{\mu_i + \mu_j}{2}}\cdot
2\sinh{\frac{\nu_i -\nu_j}{2}}\cdot 2\sinh{\frac{\nu_i + \nu_j}{2}}
\Bigr]^2}
{\prod_{i,j}\Bigl[
2\cosh{\frac{\mu_i -\nu_j}{2}}\cdot 2\cosh{\frac{\mu_i + \nu_j}{2}} \Bigr]^2} .{\nonumber}\\\end{aligned}$$ By similar arguments to sec. \[sec:generalFermi\], we find $$\begin{aligned}
&&Z_{O(2N)_{2k}\times USp(2N)_{-k}} \nonumber \\
&&= \frac{1}{2^{2N}N!}
\sum_\sigma (-1)^\sigma \int \frac{d^N \mu}{(2\pi )^N} \frac{d^N \nu}{(2\pi )^N}
e^{\frac{ik}{2\pi}\sum_{j=1}^N (\mu_j^2 -\nu_j^2 )} \prod_{j=1}^N 4\sinh^2{\nu_j} {\nonumber}\\
&&\ \ \ \times
\prod_j \frac{1}{2\cosh{\left( \frac{\mu_j -\nu_j}{2} \right) } \cdot
2\cosh{\left( \frac{\mu_j +\nu_j}{2} \right) } \cdot
2\cosh{\left( \frac{\mu_j -\nu_{\sigma (j)}}{2} \right) } \cdot
2\cosh{\left( \frac{\mu_j +\nu_{\sigma (j)}}{2} \right) } } {\nonumber}\\
&&=
\sum_\sigma (-1)^\sigma \int d^N \mu
\prod_j \rho_{O(2N)_{2k}\times USp(2N)_{-k}} (\mu_j ,\mu_{\sigma (j)} ) ,\end{aligned}$$ where
\_[O(2N)\_[2k]{}USp(2N)\_[-k]{}]{} (x,y) = .
The quantum mechanical operator $\hat{\rho}_{O(2N)_{2k}\times USp(2N)_{-k}}$ associated with this is
\_[O(2N)\_[2k]{}USp(2N)\_[-k]{}]{} = e\^[\^2]{} f\_() f\_() e\^[-\^2]{}\^2 f\_() f\_() , \[eq:O2NABJM\]
which is of course the same as the result of [@Moriyama:2015asx].
Next we consider operators corresponding to orientifolds $O3^\pm$. Let us recall that the brane construction of the $O(2N)_{2k'}\times USp(2N)_{-k'}$ ABJM is given by $(O3^- -D3) -(NS5)- (O3^+ -D3)-(1,2k')$. The $\hat{\rho}$ for the $O(2N)_{2k}\times USp(2N)_{-k} $ ABJM theory can be rewritten as
\_[O(2N)\_[2k]{}USp(2N)\_[-k]{}]{} = f\_() \_[(1,2k)]{} f\_() 4\^2 f\_() \_[NS5]{} f\_() .
Assuming $\hat{\rho}_{O(2N)_{2k}\times USp(2N)_{-k}}=\mathcal{O}_{O3^+} \mathcal{O}_{(1,2k)} \mathcal{O}_{O3^-} \mathcal{O}_{NS5} $, we arrive at the following identification
\_[O3\^+]{}\^[()]{} = 4\^2 f\_\^2 () ,\_[O3\^-]{}\^[()]{} = f\_\^2 () .
Conclusion and Discussions {#sec:conclusion}
==========================
In this paper we have studied the partition functions of the low-energy effective theories of M2-branes, whose type IIB brane constructions include the orientifolds. We have mainly focused on the circular quiver superconformal CS theory on $S^3$ with the gauge group $O(2N+1)\times USp(2N)\times \cdots \times O(2N+1)\times USp(2N)$, which is the natural generalization of the $O(2N+1)_{2k} \times USp(2N)_{-k}$ $\mathcal{N}=5$ ABJM theory. We have found that the partition function of this type of theory have the simple relation to the one of the M2-brane theories without the orientifolds with the gauge group $U(N)\times \cdots \times U(N)$. By using this relation and the recent results in [@Grassi:2014uua; @Codesido:2014oua], we have found the exact form of the grand partition function of the $O(2N+1)_{2} \times USp(2N)_{-1}$ ABJM theory, where its SUSY is expected to be enhanced to $\mathcal{N}=6$ [@Aharony:2008gk]. As another application, we discussed that our result gives the natural physical interpretation of the relation conjectured by Grassi-Hatsuda-Mariño. We also argued in appendix that the partition function of $\hat{A}_3$ quiver theory has the representation in terms of an ideal Fermi gas system of $\hat{D}$-type quiver theory and this leads the relation between the $U(N)$ and $USp(2N)$ SUSY gauge theories.
Our result , shows that the density matrix operator for the orientifold theory is the projection of the non-orientifold theory by the operator $(1-\hat{R})/2$. It is nice if we can understand this relation more precisely. Our result also implies that one can systematically study the partition function of the orientifold theory by using techniques developed in the studies of the non-orientifold theory. For instance the technique introduced in [@Okuyama:2015auc] allows us to compute WKB expansion of ${\rm Tr}(\hat{\rho}^\ell \hat{R})$ systematically[^20] in terms of information on Wigner transformation of $\hat{\rho}^\ell $. It is interesting to determine non-perturbative effects in the orientifold theories by such techniques.
Recalling that the $U(N)_k \times U(N)_{-k}$ $\mathcal{N}=6$ ABJM theory is described by topological string on local $\mathbb{P}^1 \times \mathbb{P}^1$, this relation would imply that the $O(2N+1)_{2k}\times USp(2N)_{-k}$ ABJM theory is described by certain projection in the topological string. There should be a physical meaning of $(1-\hat{R})/2$ in the context of the topological string.
Although we have found the physical interpretation of one of relations conjectured by Grassi-Hatsuda-Mariño [@Grassi:2014uua], they also conjectured other relations among the grand partition functions of the ABJ(M) theory with specific values of the parameters:
&&\_[U(N)\_4 U(N)\_[-4]{}]{} ()= \_[U(N+1)\_2 U(N)\_[-2]{}]{}\^[(+)]{} () ,\_[U(N+2)\_4 U(N)\_[-4]{}]{} ()= \_[U(N+1)\_2 U(N)\_[-2]{}]{}\^[(-)]{} () ,\
&& \_[U(N+2)\_8 U(N)\_[-8]{}]{} () = \_[U(N+2)\_4 U(N)\_[-4]{}]{}\^[(-)]{}() .
Although these relations might be accidental coincidences, it would be illuminating if we can find some physical interpretations.
One of immediate extensions of our analysis is to consider the gauge group $O(2N_1 +1)\times USp(2N_2 ) \times \cdots \times O(2N_1 +1)\times USp(2N_2 )$. Probably this can be done by combining the technique in [@Honda:2013pea; @Matsumoto:2013nya] with the Cauchy determinant-like formula . If this is the case, $\hat{\rho}$ for the $O(2N_1 +1)_{2k}\times USp(2N_2 )_{-2k}$ $\mathcal{N}=5$ ABJ(M) theory would be projection[^21] of the one of the $U(N_1 )_{2k} \times U(N_2 )_{-2k}$ $\mathcal{N}=6$ ABJ(M) theory by $(1-\hat{R})/2$. Another interesting direction is to study other supersymmetric observables such as supersymmetric Wilson loops. Then the techniques established in [@Hatsuda:2013yua] would be efficient.
Acknowledgment {#acknowledgment .unnumbered}
--------------
We are grateful to Kazumi Okuyama for his early collaboration and many valuable discussions. We thank Benjamin Assel and Sanefumi Moriyama for helpful discussions.
An exact relation between $USp(2N)$ and $U(N)$ gauge theories {#sec:A3D3}
=============================================================
In this appendix we show the exact relation between the SUSY gauge theories with $U(N)$ and $USp(2N)$ gauge groups. The $U(N)$ gauge theory, which we consider here, is $\mathcal{N}=4$ vector multiplet with one adjoint hyper multiplet and $N_f$ fundamental hyper multiplets, whose partition function is described by so-called $N_f$-matrix model [@Grassi:2014vwa]: $$\begin{aligned}
Z_{U+adj.}(N,N_f )
= \frac{1}{N!} \int \frac{d^N \mu}{(2\pi )^N}
\prod_{j=1}^N \frac{1}{\left( 2\cosh{\frac{\mu_j}{2}} \right)^{N_f}}
\prod_{i<j} \tanh^2{\frac{\mu_i -\mu_j}{2}} .
\label{eq:Nf_matrix}\end{aligned}$$ This matrix model has been studied well in [@Mezei:2013gqa; @Grassi:2014vwa; @Moriyama:2014gxa; @Moriyama:2014nca; @Hatsuda:2015lpa]. The $USp(2N)$ gauge theory is $\mathcal{N}=4$ vector multiplet with one anti-symmetric hyper multiplet and $N_f$-fundamental hyper multiples and its partition function is $$\begin{aligned}
Z_{USp+A}(N, N_f )
= \frac{1}{2^{2N}N!} \int \frac{d^N \mu}{(2\pi )^N}
\prod_{j=1}^N \frac{4\sinh^2{\mu_j}}{\left( 4\cosh^2{\frac{\mu_j}{2}} \right)^{N_f}}
\prod_{i<j} \Biggl[ \frac{ \sinh{\frac{\mu_i -\mu_j}{2}}\cdot \sinh{\frac{\mu_i + \mu_j}{2}} }
{ \cosh{\frac{\mu_i -\mu_j}{2}}\cdot \cosh{\frac{\mu_i + \mu_j}{2}} } \Biggr]^2 ,\end{aligned}$$ which has been analyzed in [@Mezei:2013gqa; @Assel:2015hsa; @Okuyama:2015auc]. Regarding these theories, the work [@Okuyama:2015auc] has proposed the following equivalence[^22] $$Z_{U(N)+adj.} (N,N_f =4 ) =Z_{USp(2N)+A}(N,N_f =3 ) .$$ This relation is expected from 3d mirror symmetry [@Intriligator:1996ex; @deBoer:1996mp]. It is known that the $U(N)$ and $USp(2N)$ theories are equivalent to $\hat{A}_{N_f -1}$ and $\hat{D}_{N_f }$ quiver theories without CS levels, where only one of the vector multiples is coupled to one fundamental hyper multiplet, respectively. Since $\hat{A}_3 =\hat{D}_3$, the equation should hold. In this section we explicitly prove this relation by using the technique in [@Assel:2015hsa; @Moriyama:2015jsa].
$\hat{A}_3 =\hat{D}_3$ {#sec:A3D3proof}
----------------------
Here we show that partition function of $\hat{A}_3$ quiver theory has a representation in terms of an ideal Fermi gas system of $\hat{D}$-type quiver theory. Although this may be already proven in [@Assel:2015hsa; @Moriyama:2015jsa], it is unclear to us whether their derivation includes our analysis in this section or not and therefore we explicitly prove it.
First we precisely explain what we would like to prove. Suppose the SUSY CS theory with $\hat{A}_n$ quiver, namely the circular quiver with the gauge group $U(N)_{k_1} \times \cdots U(N)_{k_{n+1}}$, which is coupled to $N_f^{(a)}$ fundamental hyper multiplets. The partition function of the $\hat{A}_n$ quiver theory can be denoted by [@Kapustin:2009kz] $$\begin{aligned}
Z_{\hat{A}_n} \nonumber
&=& \frac{1}{N!^{n+1}} \int \prod_{a=1}^{n+1} \frac{d^N \mu^{(a)}}{(2\pi )^N}
\prod_{j=1}^N f^{(a)}(\mu_j^{(a)} )
\frac{\prod_{i<j} \Bigl[
2\sinh{\frac{\mu_i^{(a)} -\mu_j^{(a)}}{2}}\cdot 2\sinh{\frac{\mu_i^{(a)} + \mu_j^{(a)}}{2}}\Bigr]^2}
{\prod_{i,j} 2\cosh{\frac{\mu_i^{(a)} -\mu_j^{(a+1)}}{2}}} ,\end{aligned}$$ where $\mu_j^{(n+2)}=\mu_j^{(1)}$ and
f\^[(a)]{}(x ) = .
It is known that one can rewrite the partition function of the $\hat{A}_n$ theory as [@Marino:2011eh]
Z\_[\_n]{}(N) =\_[S\_N]{} (-1)\^d\^N x \_[j=1]{}\^N \_[\_n]{} (x\_j ,x\_[(j)]{}) ,
where $\rho_{\hat{A}_n}(x,y)$ is the density matrix of the ideal Fermi gas system associated with the quantum mechanical operator
\_[\_n]{} ( ,) = \_[a=1]{}\^[n+1]{} f\^[(a)]{}( ) .
![ The $\hat{D}_{L+2}$ quiver diagram. []{data-label="fig:Dquiver"}](Dquiver.eps){width="7.4cm"}
Next let us consider the $\hat{D}_{L+2}$ quiver CS theory with the gauge group $U(N)_{k_0} \times U(N)_{k_0'}\times U(2N)_{k_1} \times \cdots U(2N)_{k_{L-1}}\times U(N)_{k_L} \times U(N)_{k_L'}$ (see fig. \[fig:Dquiver\]) . The partition function of this theory is given by [@Kapustin:2009kz] $$\begin{aligned}
Z_{\hat{D}_{L+2}}
&=& \frac{1}{N!^4 (2N!)^{L-1}} \int \frac{d^N \mu^{(0)}}{(2\pi )^N}\frac{d^N \mu^{\prime (0)}}{(2\pi )^N}
\frac{d^N \mu^{(L)}}{(2\pi )^N}\frac{d^N \mu^{\prime (L)}}{(2\pi )^N}
\prod_{a=1}^{L-1} \frac{d^{2N} \mu^{(a)}}{(2\pi )^{2N}}
\prod_{a=1}^{L-1} \prod_{J=1}^{2N} F^{(a)}(\mu_J^{(a)} ) {\nonumber}\\
&& \prod_{j=1}^N F^{(0)}(\mu_j^{(0)} ) F^{\prime (0)}(\mu_j^{\prime (0)} )
F^{(L)}(\mu_j^{(L)} ) F^{\prime (L)}(\mu_j^{\prime (L)} )
\frac{\prod_{a=1}^{L-1} \prod_{I\neq J} 2\sinh{\frac{\mu_I^{(a)} -\mu_J^{(a)}}{2}} }
{\prod_{a=1}^{L-2} \prod_{I,J} 2\cosh{\frac{\mu_I^{(a)} -\mu_J^{(a+1)}}{2}} } {\nonumber}\\
&& \frac{\prod_{i\neq j} 2\sinh{\frac{\mu_i^{(0)} -\mu_j^{(0)}}{2}} \cdot 2\sinh{\frac{\mu_i^{\prime (0)} -\mu_j^{\prime (0)}}{2}} \cdot
2\sinh{\frac{\mu_i^{(L)} -\mu_j^{(L)}}{2}} \cdot 2\sinh{\frac{\mu_i^{\prime (L)} -\mu_j^{\prime (L)}}{2}} }
{\prod_{i=1}^N \prod_{J=1}^{2N} 2\cosh{\frac{\mu_i^{(0)} -\mu_J^{(1)}}{2}} \cdot 2\cosh{\frac{\mu_i^{\prime (0)} -\mu_J^{(1)}}{2}} \cdot
2\cosh{\frac{\mu_i^{(L)} -\mu_J^{(L-1)}}{2}} \cdot 2\cosh{\frac{\mu_i^{\prime (L)} -\mu_J^{(L-1)}}{2}} } , {\nonumber}\\\end{aligned}$$ where
F\^[(a)]{}(x ) = ,F\^[(a)]{}(x ) = .
It is also known that the partition function of the $\hat{D}_{L+2}$ quiver theory is described by an ideal Fermi gas system [@Assel:2015hsa; @Moriyama:2015jsa]:
Z\_[\_[L+2]{}]{} = \_[S\_[N]{}]{} (-1)\^ \_[j=1]{}\^N \_[\_[L+2]{}]{}\^[()]{} (x\_j ,x\_[(j)]{}) ,
where
&&\_[\_[L+2]{}]{}\
&&= ( F\^[(0)]{}()F\^[(0)]{}() +F\^[(0)]{}()F\^[(0)]{}() ) ( \_[a=1]{}\^[L-1]{} F\^[(a)]{}() )\
&& ( F\^[(L)]{}()F\^[(L)]{}() +F\^[(L)]{}()F\^[(L)]{}() ) ( \_[a=1]{}\^[L-1]{} F\^[(L-a)]{}() ) .\
In this section we prove
Z\_[\_3]{} = \_[S\_[N]{}]{} (-1)\^ \_[j=1]{}\^N , \[eq:D3Fermi\]
where
\_[L1]{} \_[\_[L+2]{}]{} &=& ( F\^[(0)]{}()F\^[(0)]{}() +F\^[(0)]{}()F\^[(0)]{}() )\
&& ( F\^[(1)]{}()F\^[(1)]{}() +F\^[(1)]{}()F\^[(1)]{}() ) .
As mentioned above, this may be already proven in [@Assel:2015hsa; @Moriyama:2015jsa]. However their derivation apparently seems to take $L\geq 2$, where at least one $U(2N)$ node is present, and it is unclear to us whether their derivation includes $\hat{D}_3$ ($L=1$) case or not. Therefore we explicitly prove this relation.
Now let us consider the $A_3$ quiver theory: $$\begin{aligned}
Z_{\hat{A}_3} \nonumber
&=& \frac{1}{N!^4} \int \prod_{a=1}^4 \frac{d^N \mu^{(a)}}{(2\pi )^N}
\prod_{j=1}^N f^{(a)}(\mu_j^{(a)} )
\frac{\prod_{i<j} \Bigl[
2\sinh{\frac{\mu_i^{(a)} -\mu_j^{(a)}}{2}}\cdot 2\sinh{\frac{\mu_i^{(a)} + \mu_j^{(a)}}{2}}\Bigr]^2}
{\prod_{i,j} 2\cosh{\frac{\mu_i^{(a)} -\mu_j^{(a+1)}}{2}}} .\end{aligned}$$ Let us redefine the variables as in $\hat{D}_3$-quiver language:
&&\_j\^[(1)]{} = x\_j ,\_j\^[(2)]{} = y\_[N+j]{} ,\_j\^[(3)]{} = x\_[N+j]{} ,\_j\^[(4)]{} = y\_j ,\
&& F\^[(0)]{}(x) = f\^[(1)]{}(x),F\^[(0)]{}(x) = f\^[(3)]{}(x),F\^[(1)]{}(x) = f\^[(4)]{}(x),F\^[(1)]{}(x) = f\^[(2)]{}(x) .\
Then the partition function becomes
Z\_[\_3]{} &=& \_[j=1]{}\^N F\^[(0)]{}(x\_j ) F\^[(0)]{}(x\_[N+j]{} ) F\^[(1)]{}(y\_j ) F\^[(1)]{}(y\_[N+j]{} )\
&& ,
where $I,J =1,\cdots ,2N$. By inserting
1 = ,
to the integrand and using the Cauchy determinant formula, we find
Z\_[\_3]{} &=& \_[S\_[2N]{}]{} (-1)\^ \_[j=1]{}\^N F\^[(0)]{}(x\_j ) F\^[(0)]{}(x\_[N+j]{} ) F\^[(1)]{}(y\_j ) F\^[(1)]{}(y\_[N+j]{} )\
&& .
Below in this subsection we just repeat the argument of [@Assel:2015hsa]. According to [@Assel:2015hsa], we introduce
R(j)=N+j,R(N+j) =j .
Now we would like to rewrite the integral in terms of a kernel acting on set of $N$ eigenvalues $\mathcal{K}(\sigma )$ among $x_J$’s, which is dependent on the permutation $\sigma$. More precisely, we take $\mathcal{K}(\sigma )$ such that $R\tau^{-1}R\tau (j) \in\mathcal{K}(\sigma ) $ for given $j\in \mathcal{K}(\sigma )$. Then we rewrite the partition function as
Z\_[\_3]{} &=& \_[S\_[2N]{}]{} (-1)\^ \_[j=1]{}\^N F\^[(0)]{}(x\_j ) F\^[(0)]{}(x\_[N+j]{} ) F\^[(1)]{}(y\_j ) F\^[(1)]{}(y\_[N+j]{} )\
&& \_[j()]{} ,\
\[eq:exp1\]
where
s(j) = {
0 & [for]{}& j=1,,N 1 & [for]{}& j=N+1,,2N
. .
Note that we can also write this as
Z\_[\_3]{} &=& \_[S\_[2N]{}]{} (-1)\^[R]{} \_[j=1]{}\^N F\^[(0)]{}(x\_j ) F\^[(0)]{}(x\_[N+j]{} ) F\^[(1)]{}(y\_j ) F\^[(1)]{}(y\_[N+j]{} )\
&& \_[j()]{} .\
\[eq:exp2\]
Averaging over these, we obtain
Z\_[\_3]{} &=& \_[S\_[2N]{}]{} (-1)\^ \_[j=1]{}\^N F\^[(0)]{}(x\_j ) F\^[(0)]{}(x\_[N+j]{} ) F\^[(1)]{}(y\_j ) F\^[(1)]{}(y\_[N+j]{} )\
&&\_[j()]{} (-1)\^[s((j)+s(j) +1]{}\
&=& \_[S\_[2N]{}]{} (-1)\^ \_[j()]{} (-1)\^[s((j)+s(j) ]{} (x\_j ,x\_[R\^[-1]{}R(j)]{})\
&=& \_[S\_[N]{}]{} (-1)\^ \_[j=1]{}\^N \_[\_3]{}\^[()]{} (x\_j ,x\_[(j)]{}) ,
where
&&\_[\_3]{} (x\_1 ,x\_2 )\
&&= -2 ( F\^[(0)]{}(y) F\^[(0)]{}(y’ ) +F\^[(0)]{}(y ) F\^[(0)]{}(y’) )\
&& ( F\^[(1)]{}(x’) F\^[(1)]{}(x ) +F\^[(1)]{}(x’ ) F\^[(1)]{}(x) ) .
Hence corresponding operator $\hat{\rho}_{\hat{D}_3}$ is
\_[\_3]{} &=& ( F\^[(0)]{}()F\^[(0)]{}() +F\^[(0)]{}()F\^[(0)]{}() )\
&& ( F\^[(1)]{}()F\^[(1)]{}() +F\^[(1)]{}()F\^[(1)]{}() )\
&=&\_[L1]{} \_[\_[L+2]{}]{} .
$\hat{A}_n \rightarrow U(N)+adj.$ {#sec:mirrorA}
---------------------------------
Suppose the $\hat{A}_n$ quiver theories without CS terms, where only one of the $U(N)$ vector multiples is coupled to one fundamental hyper multiplet. This theory is related to the $U(N)$ gauge theory with $\mathcal{N}=4$ vector multiplet, one adjoint hyper multiplet and $n+1$ fundamental hyper multiplets. We can easily show this for the partition functions [@Kapustin:2010xq; @Marino:2011eh; @Drukker:2015awa]. To be self contained, here we repeat its derivation. The density matrix operator $\hat{\rho}$ of the $\hat{A}_n$ theory is
\_[\_n]{} =
By the canonical transformation $(Q,P)\rightarrow (Q,-P)$, we get
\_[\_n]{} = .
This $\hat{\rho}$ gives the $N_f$ matrix model with $N_f =n+1$.
$\hat{D}_n \rightarrow USp +A$ {#sec:mirrorUSp}
------------------------------
We also review the proof of the 3d mirror symmetry between the partition functions on $S^3$ of the $\hat{D}_n$ quiver and $USp+A$ theories. The gauge group of the $\hat{D}_n$ quiver theory consists of four $U(N)$ nodes and $(n-3)$ $U(2N)$ nodes, where one of $U(N)$ nodes associates one fundamental hypermultiplet. The partition function of this theory is given by $$\begin{aligned}
Z_{\hat{D}_n}
&=& \frac{1}{N!^2 (2N!)^{n-3}} \int \frac{d^N \mu}{(2\pi )^N} \frac{d^N \mu'}{(2\pi )^N} \frac{d^N \nu}{(2\pi )^N} \frac{d^N \nu'}{(2\pi )^N}
\frac{d^{2N} \lambda^{(1)}}{(2\pi )^{2N}} \cdots \frac{d^{2N} \lambda^{(n-3)}}{(2\pi )^{2N}} {\nonumber}\\
&& \frac{ \prod_{i\neq j} 2\sinh{\frac{\mu_i -\mu_j}{2}}\cdot 2\sinh{\frac{\mu_i' - \mu_j'}{2}} }
{\prod_j 2\cosh{\frac{\mu_j}{2}} \prod_{i,J} 2\cosh{\frac{\mu_i -\lambda^{(1)}_J}{2}}\cdot 2\cosh{\frac{\mu'_i -\lambda^{(1)}_J}{2}} } {\nonumber}\\
&& \prod_{\alpha =1}^{n-4}\Biggl[ \frac{ \prod_{I\neq J} 2\sinh{\frac{\lambda_I^{(\alpha )} -\lambda_J^{(\alpha )}}{2}} \cdot
2\sinh{\frac{\lambda_I^{(\alpha +1)} -\lambda_J^{(\alpha +1)}}{2}} }
{ \prod_{I,J} 2\cosh{\frac{\lambda_I^{(\alpha )} -\lambda_J^{(\alpha +1)}}{2}} } \Biggr]
\frac{ \prod_{i\neq j} 2\sinh{\frac{\nu_i -\nu_j}{2}}\cdot 2\sinh{\frac{\nu_i' - \nu_j'}{2}} }
{\prod_{i,J} 2\cosh{\frac{\nu_i -\lambda^{(n-3)}_J}{2}}\cdot 2\cosh{\frac{\nu'_i -\lambda^{(n-3)}_J}{2}} } . {\nonumber}\\\end{aligned}$$ Corresponding $\hat{\rho}$ is
\_[\_n]{} &=& { , } ( )\^[n-2]{} ( )\^[n-2]{} . \[eq:Dquiver\]
By using
{ , } = ( + ) ,\
we find
\_[\_n]{} &=& ( 2 2 +2 2 )\
&& ( )\^[n-1]{} ( )\^[n-2]{} .
Then the similarity transformation
\_[\_n]{} 2 ,
leads us to
\_[\_n]{} = (2 2 +22 ) .
By the canonical transformation
(P,Q) (Q,-P) ,
we obtain
\_[\_n]{} = ( 2 2 + 2 2 ) .
Indeed this gives the same partition function as the $USp+A$ theory with $N_f =n$ because $\hat{\rho}$ of the $USp+A$ theory:
\_[USp+A]{} ( ,) = ,
satisfies
\_[\_n]{} &=& ( 2 2 + 2 )\
&=& ( \_[USP+A]{} ) 2 .
Thus, combining the results in app. \[sec:A3D3proof\], app. \[sec:mirrorA\] and app. \[sec:mirrorUSp\], we prove .
[10]{}
V. Pestun, [*[Localization of gauge theory on a four-sphere and supersymmetric Wilson loops]{}*]{}, [*Commun.Math.Phys.*]{} [**313**]{} (2012) 71–129, \[[[arXiv:0712.2824]{}](http://xxx.lanl.gov/abs/0712.2824)\].
A. Kapustin, B. Willett, and I. Yaakov, [*[Exact Results for Wilson Loops in Superconformal Chern-Simons Theories with Matter]{}*]{}, [*JHEP*]{} [**1003**]{} (2010) 089, \[[[ arXiv:0909.4559]{}](http://xxx.lanl.gov/abs/0909.4559)\]; D. L. Jafferis, [*[The Exact Superconformal R-Symmetry Extremizes Z]{}*]{}, [ *JHEP*]{} [**1205**]{} (2012) 159, \[[[ arXiv:1012.3210]{}](http://xxx.lanl.gov/abs/1012.3210)\]; N. Hama, K. Hosomichi, and S. Lee, [*[Notes on SUSY Gauge Theories on Three-Sphere]{}*]{}, [*JHEP*]{} [**1103**]{} (2011) 127, \[[[arXiv:1012.3512]{}](http://xxx.lanl.gov/abs/1012.3512)\].
M. Marino and P. Putrov, [*[ABJM theory as a Fermi gas]{}*]{}, [ *J.Stat.Mech.*]{} [**1203**]{} (2012) P03001, \[[[arXiv:1110.4066]{}](http://xxx.lanl.gov/abs/1110.4066)\].
Y. Hatsuda, M. Marino, S. Moriyama, and K. Okuyama, [*[Non-perturbative effects and the refined topological string]{}*]{}, [[arXiv:1306.1734]{}](http://xxx.lanl.gov/abs/1306.1734).
S. Matsumoto and S. Moriyama, [*[ABJ Fractional Brane from ABJM Wilson Loop]{}*]{}, [*JHEP*]{} [**1403**]{} (2014) 079, \[[[arXiv:1310.8051]{}](http://xxx.lanl.gov/abs/1310.8051)\].
M. Honda and K. Okuyama, [*[Exact results on ABJ theory and the refined topological string]{}*]{}, [*JHEP*]{} [**1408**]{} (2014) 148, \[[[arXiv:1405.3653]{}](http://xxx.lanl.gov/abs/1405.3653)\].
Y. Hatsuda, M. Honda, S. Moriyama, and K. Okuyama, [*[ABJM Wilson Loops in Arbitrary Representations]{}*]{}, [*JHEP*]{} [**1310**]{} (2013) 168, \[[[arXiv:1306.4297]{}](http://xxx.lanl.gov/abs/1306.4297)\].
O. Aharony, O. Bergman, D. L. Jafferis, and J. Maldacena, [*[N=6 superconformal Chern-Simons-matter theories, M2-branes and their gravity duals]{}*]{}, [*JHEP*]{} [**0810**]{} (2008) 091, \[[[arXiv:0806.1218]{}](http://xxx.lanl.gov/abs/0806.1218)\].
O. Aharony, O. Bergman, and D. L. Jafferis, [*[Fractional M2-branes]{}*]{}, [ *JHEP*]{} [**0811**]{} (2008) 043, \[[[ arXiv:0807.4924]{}](http://xxx.lanl.gov/abs/0807.4924)\].
M. Marino and P. Putrov, [*[Exact Results in ABJM Theory from Topological Strings]{}*]{}, [*JHEP*]{} [**1006**]{} (2010) 011, \[[[arXiv:0912.3074]{}](http://xxx.lanl.gov/abs/0912.3074)\]; N. Drukker, M. Marino, and P. Putrov, [*[From weak to strong coupling in ABJM theory]{}*]{}, [*Commun.Math.Phys.*]{} [**306**]{} (2011) 511–563, \[[[arXiv:1007.3837]{}](http://xxx.lanl.gov/abs/1007.3837)\].
N. Drukker and D. Trancanelli, [*[A Supermatrix model for N=6 super Chern-Simons-matter theory]{}*]{}, [*JHEP*]{} [**1002**]{} (2010) 058, \[[[arXiv:0912.3006]{}](http://xxx.lanl.gov/abs/0912.3006)\]; A. Klemm, M. Marino, M. Schiereck, and M. Soroush, [*[ABJM Wilson loops in the Fermi gas approach]{}*]{}, [[ arXiv:1207.0611]{}](http://xxx.lanl.gov/abs/1207.0611); A. Grassi, J. Kallen, and M. Marino, [*[The topological open string wavefunction]{}*]{}, [*Commun.Math.Phys.*]{} [**338**]{} (2015), no. 2 533–561, \[[[arXiv:1304.6097]{}](http://xxx.lanl.gov/abs/1304.6097)\].
C. P. Herzog, I. R. Klebanov, S. S. Pufu, and T. Tesileanu, [*[Multi-Matrix Models and Tri-Sasaki Einstein Spaces]{}*]{}, [*Phys.Rev.*]{} [**D83**]{} (2011) 046001, \[[[arXiv:1011.5487]{}](http://xxx.lanl.gov/abs/1011.5487)\].
H. Fuji, S. Hirano, and S. Moriyama, [*[Summing Up All Genus Free Energy of ABJM Matrix Model]{}*]{}, [*JHEP*]{} [**1108**]{} (2011) 001, \[[[arXiv:1106.4631]{}](http://xxx.lanl.gov/abs/1106.4631)\].
K. Okuyama, [*[A Note on the Partition Function of ABJM theory on $S^3$]{}*]{}, [*Prog.Theor.Phys.*]{} [**127**]{} (2012) 229–242, \[[[arXiv:1110.3555]{}](http://xxx.lanl.gov/abs/1110.3555)\].
M. Hanada, M. Honda, Y. Honma, J. Nishimura, S. Shiba, and Y. Yoshida, [ *[Numerical studies of the ABJM theory for arbitrary N at arbitrary coupling constant]{}*]{}, [*JHEP*]{} [**1205**]{} (2012) 121, \[[[arXiv:1202.5300]{}](http://xxx.lanl.gov/abs/1202.5300)\].
Y. Hatsuda, S. Moriyama, and K. Okuyama, [*[Exact Results on the ABJM Fermi Gas]{}*]{}, [*JHEP*]{} [**1210**]{} (2012) 020, \[[[arXiv:1207.4283]{}](http://xxx.lanl.gov/abs/1207.4283)\]; P. Putrov and M. Yamazaki, [*[Exact ABJM Partition Function from TBA]{}*]{}, [*Mod.Phys.Lett.*]{} [**A27**]{} (2012) 1250200, \[[[arXiv:1207.5066]{}](http://xxx.lanl.gov/abs/1207.5066)\].
Y. Hatsuda, S. Moriyama, and K. Okuyama, [*[Instanton Effects in ABJM Theory from Fermi Gas Approach]{}*]{}, [*JHEP*]{} [**1301**]{} (2013) 158, \[[[arXiv:1211.1251]{}](http://xxx.lanl.gov/abs/1211.1251)\]; F. Calvo and M. Marino, [*[Membrane instantons from a semiclassical TBA]{}*]{}, [*JHEP*]{} [**1305**]{} (2013) 006, \[[[arXiv:1212.5118]{}](http://xxx.lanl.gov/abs/1212.5118)\].
H. Awata, S. Hirano, and M. Shigemori, [*[The Partition Function of ABJ Theory]{}*]{}, [*Prog. Theor. Exp. Phys.*]{} (2013) 053B04, \[[[arXiv:1212.2966]{}](http://xxx.lanl.gov/abs/1212.2966)\].
Y. Hatsuda, S. Moriyama, and K. Okuyama, [*[Instanton Bound States in ABJM Theory]{}*]{}, [*JHEP*]{} [**1305**]{} (2013) 054, \[[[arXiv:1301.5184]{}](http://xxx.lanl.gov/abs/1301.5184)\].
M. Honda, [*[Direct derivation of “mirror” ABJ partition function]{}*]{}, [ *JHEP*]{} [**1312**]{} (2013) 046, \[[[ arXiv:1310.3126]{}](http://xxx.lanl.gov/abs/1310.3126)\].
M. Honda and S. Moriyama, [*[Instanton Effects in Orbifold ABJM Theory]{}*]{}, [*JHEP*]{} [**08**]{} (2014) 091, \[[[arXiv:1404.0676]{}](http://xxx.lanl.gov/abs/1404.0676)\].
S. Moriyama and T. Nosaka, [*[Partition Functions of Superconformal Chern-Simons Theories from Fermi Gas Approach]{}*]{}, [*JHEP*]{} [**11**]{} (2014) 164, \[[[arXiv:1407.4268]{}](http://xxx.lanl.gov/abs/1407.4268)\]; S. Moriyama and T. Nosaka, [*[ABJM membrane instanton from a pole cancellation mechanism]{}*]{}, [*Phys. Rev.*]{} [**D92**]{} (2015), no. 2 026003, \[[[arXiv:1410.4918]{}](http://xxx.lanl.gov/abs/1410.4918)\].
S. Moriyama and T. Nosaka, [*[Exact Instanton Expansion of Superconformal Chern-Simons Theories from Topological Strings]{}*]{}, [*JHEP*]{} [**05**]{} (2015) 022, \[[[arXiv:1412.6243]{}](http://xxx.lanl.gov/abs/1412.6243)\].
Y. Hatsuda, M. Honda, and K. Okuyama, [*[Large N non-perturbative effects in $\mathcal{N}=4$ superconformal Chern-Simons theories]{}*]{}, [*JHEP*]{} [**09**]{} (2015) 046, \[[[arXiv:1505.07120]{}](http://arxiv.org/abs/1505.07120)\].
M. Marino and P. Putrov, [*[Interacting fermions and N=2 Chern-Simons-matter theories]{}*]{}, [[arXiv:1206.6346]{}](http://xxx.lanl.gov/abs/1206.6346).
M. Mezei and S. S. Pufu, [*[Three-sphere free energy for classical gauge groups]{}*]{}, [*JHEP*]{} [**02**]{} (2014) 037, \[[[arXiv:1312.0920]{}](http://xxx.lanl.gov/abs/1312.0920)\].
B. Assel, N. Drukker, and J. Felix, [*[Partition functions of 3d $\hat
D$-quivers and their mirror duals from 1d free fermions]{}*]{}, [*JHEP*]{} [ **08**]{} (2015) 071, \[[[ arXiv:1504.07636]{}](http://arxiv.org/abs/1504.07636)\].
S. Moriyama and T. Nosaka, [*[Superconformal Chern-Simons Partition Functions of Affine D-type Quiver from Fermi Gas]{}*]{}, [*JHEP*]{} [**09**]{} (2015) 054, \[[[arXiv:1504.07710]{}](http://arxiv.org/abs/1504.07710)\].
S. Bhattacharyya, A. Grassi, M. Marino, and A. Sen, [*[A One-Loop Test of Quantum Supergravity]{}*]{}, [[ arXiv:1210.6057]{}](http://xxx.lanl.gov/abs/1210.6057).
A. Dabholkar, N. Drukker, and J. Gomes, [*[Localization in supergravity and quantum $AdS_4/CFT_3$ holography]{}*]{}, [*JHEP*]{} [**10**]{} (2014) 90, \[[[arXiv:1406.0505]{}](http://xxx.lanl.gov/abs/1406.0505)\].
Y. Imamura and K. Kimura, [*[On the moduli space of elliptic Maxwell-Chern-Simons theories]{}*]{}, [*Prog. Theor. Phys.*]{} [**120**]{} (2008) 509–523, \[[[arXiv:0806.3727]{}](http://xxx.lanl.gov/abs/0806.3727)\].
D. Gaiotto and E. Witten, [*[Janus Configurations, Chern-Simons Couplings, And The theta-Angle in N=4 Super Yang-Mills Theory]{}*]{}, [*JHEP*]{} [**06**]{} (2010) 097, \[[[ arXiv:0804.2907]{}](http://xxx.lanl.gov/abs/0804.2907)\]; K. Hosomichi, K.-M. Lee, S. Lee, S. Lee, and J. Park, [*[N=4 Superconformal Chern-Simons Theories with Hyper and Twisted Hyper Multiplets]{}*]{}, [*JHEP*]{} [**07**]{} (2008) 091, \[[[ arXiv:0805.3662]{}](http://xxx.lanl.gov/abs/0805.3662)\].
A. Grassi and M. Marino, [*[M-theoretic matrix models]{}*]{}, [*JHEP*]{} [ **02**]{} (2015) 115, \[[[ arXiv:1403.4276]{}](http://xxx.lanl.gov/abs/1403.4276)\].
Y. Hatsuda and K. Okuyama, [*[Probing non-perturbative effects in M-theory]{}*]{}, [*JHEP*]{} [**10**]{} (2014) 158, \[[[arXiv:1407.3786]{}](http://xxx.lanl.gov/abs/1407.3786)\].
K. Okuyama, [*[Probing non-perturbative effects in M-theory on orientifolds]{}*]{}, [[ arXiv:1511.02635]{}](http://arxiv.org/abs/1511.02635).
D. R. Gulotta, J. Ang, and C. P. Herzog, [*[Matrix Models for Supersymmetric Chern-Simons Theories with an ADE Classification]{}*]{}, [*JHEP*]{} [**1201**]{} (2012) 132, \[[[ arXiv:1111.1744]{}](http://xxx.lanl.gov/abs/1111.1744)\]; D. R. Gulotta, C. P. Herzog, and T. Nishioka, [*[The ABCDEF’s of Matrix Models for Supersymmetric Chern-Simons Theories]{}*]{}, [*JHEP*]{} [**1204**]{} (2012) 138, \[[[ arXiv:1201.6360]{}](http://xxx.lanl.gov/abs/1201.6360)\]; P. M. Crichigno, C. P. Herzog, and D. Jain, [*[Free Energy of $D_n$ Quiver Chern-Simons Theories]{}*]{}, [*JHEP*]{} [**03**]{} (2013) 039, \[[[arXiv:1211.1388]{}](http://xxx.lanl.gov/abs/1211.1388)\].
S. Moriyama and T. Suyama, [*[Instanton Effects in Orientifold ABJM Theory]{}*]{}, [[arXiv:1511.01660]{}](http://arxiv.org/abs/1511.01660).
K. Hosomichi, K.-M. Lee, S. Lee, S. Lee, and J. Park, [*[N=5,6 Superconformal Chern-Simons Theories and M2-branes on Orbifolds]{}*]{}, [ *JHEP*]{} [**09**]{} (2008) 002, \[[[ arXiv:0806.4977]{}](http://xxx.lanl.gov/abs/0806.4977)\].
A. Grassi, Y. Hatsuda, and M. Marino, [*[Quantization conditions and functional equations in ABJ(M) theories]{}*]{}, [[arXiv:1410.7658]{}](http://xxx.lanl.gov/abs/1410.7658).
S. Codesido, A. Grassi, and M. Marino, [*[Exact results in $ \mathcal{N}=8 $ Chern-Simons-matter theories and quantum geometry]{}*]{}, [*JHEP*]{} [**07**]{} (2015) 011, \[[[ arXiv:1409.1799]{}](http://xxx.lanl.gov/abs/1409.1799)\].
S. Cheon, D. Gang, C. Hwang, S. Nagaoka, and J. Park, [*[Duality between N=5 and N=6 Chern-Simons matter theory]{}*]{}, [*JHEP*]{} [**11**]{} (2012) 009, \[[[arXiv:1208.6085]{}](http://xxx.lanl.gov/abs/1208.6085)\].
K. A. Intriligator and N. Seiberg, [*[Mirror symmetry in three-dimensional gauge theories]{}*]{}, [*Phys.Lett.*]{} [**B387**]{} (1996) 513–519, \[[[hep-th/9607207]{}](http://xxx.lanl.gov/abs/hep-th/9607207)\]; A. Hanany and E. Witten, [*[Type IIB superstrings, BPS monopoles, and three-dimensional gauge dynamics]{}*]{}, [*Nucl.Phys.*]{} [**B492**]{} (1997) 152–190, \[[[ hep-th/9611230]{}](http://xxx.lanl.gov/abs/hep-th/9611230)\].
J. de Boer, K. Hori, H. Ooguri, and Y. Oz, [*[Mirror symmetry in three-dimensional gauge theories, quivers and D-branes]{}*]{}, [*Nucl. Phys.*]{} [**B493**]{} (1997) 101–147, \[[[hep-th/9611063]{}](http://xxx.lanl.gov/abs/hep-th/9611063)\].
B. Assel, [*[Hanany-Witten effect and SL(2, $\mathbb{Z}$) dualities in matrix models]{}*]{}, [*JHEP*]{} [**10**]{} (2014) 117, \[[[arXiv:1406.5194]{}](http://xxx.lanl.gov/abs/1406.5194)\].
N. Drukker and J. Felix, [*[3d mirror symmetry as a canonical transformation]{}*]{}, [*JHEP*]{} [**05**]{} (2015) 004, \[[[arXiv:1501.02268]{}](http://arxiv.org/abs/1501.02268)\].
S. Moriyama and T. Suyama, [*[Orthosymplectic Chern-Simons Matrix Model and Chirality Projection]{}*]{}, [[ arXiv:1601.03846]{}](http://arxiv.org/abs/1601.03846).
A. Kapustin, B. Willett, and I. Yaakov, [*[Nonperturbative Tests of Three-Dimensional Dualities]{}*]{}, [*JHEP*]{} [**1010**]{} (2010) 013, \[[[arXiv:1003.5694]{}](http://xxx.lanl.gov/abs/1003.5694)\].
[^1]: masazumi.hondaATweizmann.ac.il
[^2]: Only exceptions so far are the orbifold ABJM theory and $(2,2)$ model analyzed in [@Honda:2014ica] and [@Moriyama:2014gxa; @Moriyama:2014nca; @Hatsuda:2015lpa], respectively. The grand potential for the orbifold ABJM theory has a simple relation to the one of the ABJM [@Honda:2014ica] and the $(2,2)$ model is expected to be described by topological string on local $D_5$ del Pezzo [@Moriyama:2014nca; @Hatsuda:2015lpa].
[^3]: Recently there appeared a paper [@Moriyama:2015asx] on arXiv considering a similar physical setup. This reference mainly considers CS theories of $O(2N)\times USp(2N)$ type, which differs from our setup of $O(2N+1)\times USp(2N)$ type. But we also give some comments on the $O(2N)\times USp(2N)$ type in sec. \[sec:O2NUSp\].
[^4]: In order to fix the value of $M$, we should compare not only the moduli but also discrete torsion [@Aharony:2008gk].
[^5]: This statement has been partially checked by using superconformal index [@Cheon:2012be].
[^6]: The papers [@Assel:2015hsa; @Moriyama:2015jsa] have written partition functions of $\hat{D}$-type quiver theories in terms of ideal Fermi gas systems. it is unclear to us whether their derivation includes $\hat{D}_3$ case. However their derivation apparently seems to consider $\hat{D}_{n\geq 4}$ and it is unclear to us whether their derivation includes the $\hat{D}_3$ case or not. Hence we explicitly prove this for the $\hat{D}_3$ case. Even if [@Assel:2015hsa; @Moriyama:2015jsa] did not prove it for the $\hat{D}_3$ case, our derivation is not essentially new.
[^7]: We thank Kazumi Okuyama for useful discussions on this point.
[^8]: Note that the $O(2N+1)_{2k} \times USp(2N)_{-k}$ $\mathcal{N}=5$ ABJM theory has only one bi-fundamental hyper multiplet since one of two bi-fundamental hyper multiplets in the $\mathcal{N}=6$ ABJM theory is projected out by the orientifold projection [@Hosomichi:2008jb; @Aharony:2008gk]. In the localization formula, the $O\times USp$ bi-fundamental chiral multiplet (with R-charge 1/2) behaves like “half” of the hyper multiplet because of the group structure.
[^9]: Note also $\hat{R}f(\hat{Q})=f(-\hat{Q})\hat{R}$, $\hat{R}f(\hat{P})=f(-\hat{P})\hat{R}$ and $((1-\hat{R})/2)^2 =(1-\hat{R})/2$.
[^10]: Note that the definition of $\hbar$ in is slightly different from the one usually used in Fermi gas systems associated with $U(N)$ CS theories.
[^11]: When we go to the last line from the second line, we have used $\sinh^2{\mu_j} = 4\sinh^2{\frac{\mu_j}{2}} \cosh^2{\frac{\mu_j}{2}} $.
[^12]: We can also compare this with the $O(2N+1)$ gauge theory with $\mathcal{N}=4$ vector multiplet, one symmetric hyper multiplet and $N_f$-fundamental hyper multiples ($O(2N+1)+S$ theory). Because of $Z_{O(2N+1)+S}(N, N_f )=Z_{USp+S}(N, N_f -2 )$, the relation also shows $Z_{\mathcal{N}=5 \rm ABJM} (N,k=1) = Z_{O(2N+1)+S}(N, N_f =3 )$.
[^13]: Their definitions are $$\vartheta_1 (v,\tau )
=\sum_{n\in\mathbb{Z}} (-1)^{n-1/2} e^{\pi i(n+1/2)^2 \tau +2\pi i (n+1/2)v} ,\quad
\vartheta_2 (v,\tau )
=\sum_{n\in\mathbb{Z}} e^{\pi i(n+1/2)^2 \tau +2\pi i (n+1/2)v} .$$
[^14]: See [@Codesido:2014oua] for details.
[^15]: We could also include masses and FI-terms. Then $f^{(a)}(x)$ and $f^{\prime (a)}(x )$ are modified but the result in this section does not essentially change up to this modification.
[^16]: $\widetilde{O3}^-$ can be regarded as $O3^-$ plane with a half D3-brane while $\widetilde{O3}^+$ is perturbatively the same as $O3^+$ plane but different non-perturbatively.
[^17]: We could also consider $(1,\tilde{k})$-5 brane with $\tilde{k}=nk$, whose corresponding operator is $\hat{\mathcal{O}}_{(1,\tilde{k})5}= e^{\frac{i\tilde{n}Q^2}{2\hbar}} \frac{1}{2\cosh{\frac{P}{2}}} e^{-\frac{i\tilde{n}Q^2}{2\hbar}} = \frac{1}{2\cosh{\frac{P-\tilde{n}Q}{2}}} .$
[^18]: As mentioned in remark 2 of sec. \[sec:rhoABJM\], we have multiple representations of $\hat{\rho}$ to describe the same partition function. Then identifications of $\mathcal{O}_{\tilde{O3}^\pm}$ are more generally $$\mathcal{O}_{\widetilde{O3}^-} = 4\sinh^2{\frac{\hat{Q}}{2k}} f_\pm^2 (\hat{Q}) \frac{1\pm\hat{R}}{2} ,\quad
\mathcal{O}_{\widetilde{O3}^+} = \frac{\sinh^2{\frac{\hat{Q}}{k}}}{\cosh^2{\frac{\hat{Q}}{2k}}} f_\pm^2 (\hat{Q}) \frac{1\pm\hat{R}}{2} .$$
[^19]: Using the result of [@Honda:2014ica], we can also write “modified grand potential” of the orientifold projected orbifold ABJM theory in terms of the one of the $\mathcal{N}=5$ ABJM theory.
[^20]: We thank Kazumi Okuyama for discussions on this point.
[^21]: After this paper appeared in arXiv, this statement is proven in [@Moriyama:2016xin].
[^22]: One can also compare this with the $O(2N+1)$ gauge theory with $\mathcal{N}=4$ vector multiplet, one anti-symmetric hyper multiplet and $N_f$-fundamental hyper multiples ($O(2N+1)+A$ theory). Then because of $Z_{O(2N+1)+A}(N, N_f )=Z_{USp+A}(N, N_f -2 )$, the relation also indicates $Z_{U(N)+adj.} (N,N_f =4 ) =Z_{O(2N+1)+A}(N,N_f =5 )$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Penalized regression approaches are standard tools in quantitative genetics. It is known that the fit of an *ordinary least squares* (OLS) regression is independent of certain transformations of the coding of the predictor variables, and that the standard mixed model *ridge regression best linear unbiased prediction* (RRBLUP) is neither affected by translations of the variable coding, nor by global scaling. However, it has been reported that an extended version of this mixed model, which incorporates interactions by products of markers as additional predictor variables is affected by translations of the marker coding. In this work, we identify the cause of this loss of invariance in a general context of penalized regression on polynomials in the predictor variables. We show that in most cases, translating the coding of the predictor variables has an impact on effect estimates, with an exception when only the size of the coefficients of monomials of highest total degree are penalized. The invariance of RRBLUP can thus be considered as a special case of this setting, with a polynomial of total degree 1, where the size of the fixed effect (total degree 0) is not penalized but all coefficients of monomials of total degree 1 are. The extended RRBLUP, which includes interactions, is not invariant to translations because it does not only penalize interactions (total degree 2), but also additive effects (total degree 1). Our observations are not restricted to ridge regression, but generally valid for penalized regressions, for instance also for the $\ell_1$ penalty of LASSO. The fact that coding translations alter the estimates of interaction effects, provides an additional reason for interpreting the biological meaning of estimated interaction effects with caution. Moreover, this problem does not only apply to gene by gene interactions, but also to other types of interactions whose covariance is modeled with Hadamard products of covariance matrices (for instance gene by environment interactions).'
author:
- 'Johannes W R Martini$^{1,2}$'
- 'Francisco Rosales$^{3}$'
- 'Ngoc-Thuy Ha$^2$'
- Thomas Kneib$^4$
- Johannes Heise$^5$
- Valentin Wimmer$^1$
date: |
$^1$ KWS SAAT SE, Einbeck, Germany\
$^2$ Department of Animal Breeding and Genetics, University of Goettingen, Göttingen, Germany\
$^3$ Department of Finance, Universidad del Pacífico, Lima, Peru\
$^4$ Chairs of Statistics and Econometrics, University of Goettingen, Göttingen, Germany\
$^5$ IT Solutions for Animal Production (vit), Verden, Germany
title: 'Lost in translation: On the impact of data coding on penalized regression with interactions'
---
Background {#background .unnumbered}
==========
Genomic prediction is the prediction of properties of individuals from their genomic data. It is a crucial ingredient of modern breeding programs [@meuwissen01; @schaeffer06; @habier07; @hayes09a; @hayes13]. The traditional quantitative genetics theory is built upon linear models in which allele effects are mostly modeled additively [@falconer96]. In more detail, the standard model to represent the effect of the genotype on the phenotype is given by $$\label{eq:01}
\mathbf{y}=\mathbf{1}_n \mu + {\mathbf{M}}\bm{\beta} + \bm{\epsilon},$$ where $\mathbf{y}$ is the $n \times 1$ vector of the phenotypic observations of $n$ individuals and $\mathbf{1}_n$ an $n \times 1$ vector with each entry equal to $1$. Moreover, $\mu$ is the $y$-intercept, and ${\mathbf{M}}$ the $n \times p$ matrix describing the marker states of $n$ individuals at $p$ loci. Dealing with single nucleotide polymorphisms (SNPs) and a diploid species, the entries $M_{i,j}$ can for instance be coded as $0$ $(\mathbf{aa})$, $1$ $(\mathbf{aA}$ or $\mathbf{Aa})$ or $2$ $(\mathbf{AA})$ counting the occurrence of the reference allele $\mathbf{A}$. The $p \times 1$ vector $\bm{\beta}$ represents the allele substitution effects of the $p$ loci, and $\bm{\epsilon}$ the $n \times 1$ error vector. For single marker regression, which may for instance be used in genome-wide association studies (GWAS), we could apply ordinary least squares regression to estimate (or to predict) $\hat{\beta}$. However, in approaches of genomic prediction, we model the effects of many different loci simultaneously and the number of markers $p$ is usually much larger than the number of observations $n$. To reduce overfitting and to deal with a large number of predictor variables, different methods have been applied in the last decades, among which *ridge regression best linear unbiased prediction* (RRBLUP) is the most popular [@schaeffer2004application]. RRBLUP penalizes the squared $\ell_2$ norm of $\bm{\beta}$ and has been built on the additional model specifications of $\mu$ being a fixed unknown parameter, $\bm{\beta}\sim \mathcal{N}(\bm{0},\sigma_\beta^2 \mathbf{I}_p)$ and $\bm{\epsilon}\sim \mathcal{N}(\bm{0},\sigma_{\bm{\epsilon}}^2 \mathbf{I}_n)$. With an approach of maximizing a certain density, these assumptions allow to derive the optimal penalty factor as the ratio of the variance components $\lambda:= \frac{\sigma_{\bm{\epsilon}}^2}{\sigma_\beta^2}$ [@henderson75; @henderson76; @henderson77]. Please note that RRBLUP is not a pure ridge regression, but actually a mixed model in which the size of $\mu$ is not penalized, but the entries of $\bm{\beta}$ are. This mixed model RRBLUP is also called *genomic best linear unbiased prediction* (GBLUP) when it is reformulated with $\mathbf{g}:={\mathbf{M}}\bm{\beta}$, and thus $\mathbf{g}\sim \mathcal{N}(0,\sigma_\beta^2 {\mathbf{M}}' {\mathbf{M}})$.\
It is known that translations of the marker coding, that is, subtracting a constant $m_i$ from the $i$-th column of ${\mathbf{M}}$, does not change the predictions $\hat{{\mathbf{y}}}$ of an OLS regression (provided it is well-defined). This invariance also holds for RRBLUP, when the penalty factor remains fixed. Also when modeling interactions by products of two predictor variables, that is when fitting the coefficients of a polynomial of total degree 2 to the data, OLS predictions are not affected by translations of the marker coding. Contrarily, the predictions of its penalized regression analogue *extended genomic best linear unbiased prediction* (eGBLUP) are sensitive to a translation of the coding [@he2016does; @Martini17].\
In this work, we address the question of why the penalized regression method is affected by translations of the marker coding when a polynomial function of higher total degree is used. We start with a short summary of the different methods.
Theory: Specification of regression methods {#theory-specification-of-regression-methods .unnumbered}
===========================================
If an expression includes an inverse of a matrix, we implicitly assume that the matrix is invertible for the respective statement, also if not mentioned explicitly. Analogously, some statements for OLS may implicitly assume that a unique estimate exists, which in particular restricts to cases in which the number of observations is at least the same as the number of parameters that have to be determined.\
[**Additive effect regression**]{}\
The additive effect model has already been presented in Eq. (\[cor:01\]).\
[**OLS**]{} The ordinary least squares approach determines $\hat{\bm{\beta}}$ by minimizing the sum of squared residuals (SSR): $$\label{eq:03}
\begin{pmatrix}
\hat{\mu} \\
\hat{\bm{\beta}}
\end{pmatrix}_{\mbox{\small OLS}} := \operatorname*{arg\,min}_{(\mu,\bm{\beta}) \in \mathbb{R}^{p+1}} \sum\limits_{i=1}^n (y_i - {\mathbf{M}}_{i,\bullet}\bm{\beta} - \mu)^2$$ ${\mathbf{M}}_{i,\bullet}$ denotes here the $i$-th row of ${\mathbf{M}}$ representing the genomic data of individual $i$. The solution to the minimization problem of Eq. (\[eq:03\]) is given by the well-known OLS estimate $$\label{eq:04}
\begin{pmatrix}
\hat{\mu} \\
\hat{\bm{\beta}}
\end{pmatrix}_{\mbox{\small OLS}} =
\left(
\begin{pmatrix}
\mathbf{1}_n & \hspace{-0.2cm}{\mathbf{M}}\end{pmatrix}^{t}
\begin{pmatrix}
\mathbf{1}_n & \hspace{-0.2cm} {\mathbf{M}}\end{pmatrix} \right)^{-1}
\begin{pmatrix}
\mathbf{1}_n & \hspace{-0.2cm} {\mathbf{M}}\end{pmatrix}^{t}
\mathbf{y}$$ provided that the required inverse exists, which in particular also means that $n$ has to be greater than $p$.\
In problems of statistical genetics, we often deal with a high number of loci and a relatively low number of observations. In this situation of $p \geq n$, the solution to Eq. (\[eq:03\]) is not unique but a vector subspace of which each point minimizes Eq. (\[eq:03\]) to zero (“overfitting”). Using an arbitrary value of this subspace, predictions $\hat{\mathbf{y}}$ for genotypes which have not been used to estimate the parameters $(\hat{\mu},\hat{\beta})$ usually have a low correlation with the corresponding realized phenotypes. An approach to prevent overfitting is RRBLUP.\
[**RRBLUP / GBLUP**]{} minimizes $$\label{eq:05}
\begin{pmatrix}
\hat{\mu} \\
\hat{\bm{\beta}}
\end{pmatrix}_{\mbox{\small RR}_\lambda} :=\operatorname*{arg\,min}_{(\mu,\bm{\beta}) \in \mathbb{R}^{p+1}} \sum\limits_{i=1}^n (y_i - {\mathbf{M}}_{i,\bullet}\bm{\beta} - \mu)^2 + \lambda \sum\limits_{j=1}^p {\beta}_j^2$$ for a penalty factor $\lambda > 0$. Using an approach of maximizing the density of the joint distribution of $(\mathbf{y},\bm{\beta})$, the model specifications of ${\beta_i}\stackrel{i.i.d.}{\sim} \mathcal{N}(0,\sigma_\beta^2)$ and ${\bm{\epsilon}_i}\stackrel{i.i.d.}{\sim} \mathcal{N}(0,\sigma_{\bm{\epsilon}}^2)$ allow to determine the penalty factor as ratio of the variance components as $\lambda := \frac{\sigma_{\bm{\epsilon}}^2}{\sigma_\beta^2}$. We stress again that Eq. (\[eq:05\]) is not a pure ridge regression, as the name RRBLUP might suggest, but a mixed model which treats $\mu$ and ${\bm{\beta}}$ differently by not penalizing the size of $\mu$. This is the version, which is most frequently used in the context of genomic prediction (often with additional fixed effects).\
The corresponding solution is given by $$\begin{pmatrix}
\hat{\mu} \\
\hat{\bm{\beta}}
\end{pmatrix}_{\mbox{\small RR}_\lambda} =
\left(
\begin{pmatrix}
\mathbf{1}_n & \hspace{-0.2cm}{\mathbf{M}}\end{pmatrix}^{t}
\begin{pmatrix}
\mathbf{1}_n & \hspace{-0.2cm} {\mathbf{M}}\end{pmatrix} + \lambda \begin{pmatrix}
0 & \mathbf{0 }_p^t \\
\mathbf{0 }_p & \mathbf{I}_p
\end{pmatrix}
\right)^{-1}
\begin{pmatrix}
\mathbf{1}_n & \hspace{-0.2cm} {\mathbf{M}}\end{pmatrix}^{t}
\mathbf{y}. \label{eq:06}$$ where $\mathbf{0 }_p$ denotes the $p \times 1$ vector of zeros. The effect of the introduction of the penalization term $\lambda \sum\limits_{j=1}^p {\beta}_j^2$ is that for the minimization of Eq. (\[eq:05\]), we have a trade-off between fitting the data optimally and shrinking the squared effects to $0$. The method will only “decide” to increase the estimate $\hat{\beta}_j$, if the gain from improving the fit is greater than the penalized loss generated by the increase of $\hat{\beta}_j$.\
[**First order epistasis: Polynomials of total degree two**]{}\
An extension of the additive model of Eq. (\[eq:01\]) is a first order epistasis model given by a polynomial of total degree two in the marker data [@ober15; @jiang15; @Martini16] $$\label{eq:07}
y_i =\mathbf{1}_n \mu + {\mathbf{M}}_{i,\bullet}\bm{\beta} + \sum\limits_{k=1}^p\sum\limits_{j=k+1}^p h_{j,k}M_{i,j}M_{i,k} + \bm{\epsilon},$$ Please note that there is a variant of this model, in which also $j=k$ is included. This interaction of a locus with itself allows to model dominance [@Martini16].\
We recapitulate some terms which are important in the context of polynomials in multiple variables. Each product of the variables $M_{i,1},M_{i,2},\dots,M_{i,p}$ is called a monomial. For instance $M_{i,1}$, $M_{i,2}$, $M_{i,1} M_{i,2}$ and $M_{i,1}^2$ are four different monomials. Since the product is commutative, $M_{i,1} M_{i,2}$ and $M_{i,2} M_{i,1}$ are the same monomial (and their coefficients are assumed to be summed up in any polynomial which we will address later). Moreover, the total degree of a monomial is the sum of the powers of the variables in the respective monomial. For instance, $M_{i,1}$ and $M_{i,2}$ are monomials of total degree 1, whereas $M_{i,1}$$M_{i,2}$, and $M_{i,1}^2$ are monomials of total degree 2. Moreover, $M_{i,1}M_{i,2}$ is a monomial of degree 1 in each variable $M_{i,1}$ and $M_{i,2}$ and $M_{i,1}^2$ is a monomial of degree 2 in $M_{i,1}$ and of degree $0$ in $M_{i,2}$. Since a polynomial model is also linear in the coefficients, the regression equations are only slightly modified.\
[**OLS**]{} Eq. (\[eq:04\]) with a modified matrix ${\mathbf{M}}$ including the products of markers as additional predictor variables represents the OLS solver for model (\[eq:07\]).\
[**eRRBLUP**]{} The extended RRBLUP is based on Eq. (\[eq:07\]) and the assumptions of $\mu$ being fixed, ${\beta_i}\stackrel{i.i.d.}{\sim} \mathcal{N}(0,\sigma_\beta^2)$, $h_{j,k }\stackrel{i.i.d.}{\sim} \mathcal{N}(0,\sigma_h^2)$ and ${{\bm{\epsilon}}_i}\stackrel{i.i.d.}{\sim} \mathcal{N}(0,\sigma_{\bm{\epsilon}}^2)$. In this case, the solution is also given by an analogue of Eq. (\[eq:06\]), but with two different penalty factors, $\lambda_1:=\frac{\sigma_{\bm{\epsilon}}^2}{\sigma_\beta^2}$ for additive effects and $\lambda_2:=\frac{\sigma_{\bm{\epsilon}}^2}{\sigma_h^2}$ for interaction effects.\
[**Translations of the marker coding**]{}\
In quantitative genetics, column means are often subtracted from the original $0$, $1$, $2$ coding of ${\mathbf{M}}$ to use $\tilde{{\mathbf{M}}}:={\mathbf{M}}-\mathbf{1}_n{\mathbf{P}}^t$ with ${\mathbf{P}}$ the vector of column means of ${\mathbf{M}}$ [@vanraden2008efficient] such that $$\sum_{i=1}^n \tilde{M}_{i,j} = 0\quad \forall j =1,\dots,p.$$ However, other types of translations, for instance a symmetric $\{-1,0,1\}$ coding or a genotype-frequency centered coding [@alvarez2007unified; @vitezica] can also be found in quantitative genetics’ literature. Thus, the question occurs whether this has an impact on the estimates of the marker effects or on the prediction of genetic values of genotypes.\
The answer is that for the additive setup of Eq. (\[eq:01\]), a shift from ${\mathbf{M}}$ to $\tilde{{\mathbf{M}}}$ will change $\hat{\mu}$ but not $\hat{\bm{\beta}}$ and any prediction $\hat{\mathbf{y}}$ will not be affected, neither for OLS, nor for RRBLUP (provided that $\lambda$ is not changed). This invariance of the additive model does not hold for the extended RRBLUP method.\
We give an example and discuss the effect of translations of the marker coding in a more general way afterwards.
\[ex:06\] Let the marker data of five individuals with two markers be given: $$\mathbf{y}= (-0.72,2.34,0.08,-0.89,0.86)^t \qquad
{\mathbf{M}}= \begin{pmatrix}
2 & 2 \\
1 & 2 \\
2 & 0 \\
2 & 1 \\
1 & 0 \\
\end{pmatrix}$$ Moreover, let us use the original matrix ${\mathbf{M}}$, and the column mean centered matrix $\tilde{{\mathbf{M}}}:= {\mathbf{M}}- \mathbf{1}_5 \underbrace{(1.6,1.0)}_{=:{\mathbf{P}}^t}$. We consider the first order epistasis model $$y_i := \mu + \beta_1 M_{i,1} + \beta_2 M_{i,2} + h_{1,2} M_{i,1}M_{i,2} + \bm{\epsilon}_i$$ and estimate the corresponding parameters with i) an OLS regression, ii) a mixed model regression eRRBLUP-1 with $\lambda_1=\lambda_2=1$, and iii) a mixed model regression eRRBLUP-2 with $\lambda_1=0$ and $\lambda_2=1$. The difference between eRRBLUP-1 and eRRBLUP-2 is that the first method penalizes the additive effects and the interaction effect, whereas the latter method only penalizes the interaction effect. The results are reported in Table \[table:example1\]. We summarize our observations from the reported results as follows:
- Comparing the centered and non-centered versions of OLS, the estimates for $\mu$, $\beta_1$ and $\beta_2$ change, but the estimated interaction $\hat{h}_{1,2}$ as well as the prediction of ${\mathbf{y}}$ remains unchanged.
- Comparing the centered and non-centered versions of eRRBLUP-1, both codings give different estimates for all the parameters and these solutions produce different predictions for ${\mathbf{y}}$.
- Comparing the centered and non-centered versions of eRRBLUP-2, both codings give different estimates for $\mu$, $\beta_1$ and $\beta_2$, but the same for $h_{1,2}$ and the same predictions for ${\mathbf{y}}$.
---------------------- ------- ------- ------- ------- ------- -------
Estimates nc c nc c nc c
$\hat{\mu}$ 1.83 0.33 1.81 0.33 2.69 0.33
$\hat{\beta_1}$ -0.97 -2.11 -0.89 -1.15 -1.54 -2.11
$\hat{\beta_2}$ 1.88 0.06 0.71 0.09 1.03 0.11
$\hat{h}_{1,2}$ -1.14 -1.14 -0.48 -0.57 -0.57 -0.57
$\hat{{\mathbf{y}}}$
-0.91 -0.91 -0.46 -0.27 -0.63 -0.63
2.34 2.34 1.39 1.46 2.06 2.06
-0.11 -0.11 0.03 0.01 -0.40 -0.40
-0.51 -0.51 -0.21 -0.13 -0.51 -0.51
0.86 0.86 0.92 0.59 1.15 1.15
---------------------- ------- ------- ------- ------- ------- -------
: Results from Example 1. “nc” denotes the use of the non-centered matrix ${\mathbf{M}}$ and “c” indicates the use of the centered matrix $\tilde{\mathbf{M}}$.[]{data-label="table:example1"}
The different cases presented in Example \[ex:06\] have a certain systematic pattern, which we discuss in the following section.
Results {#results .unnumbered}
=======
The observations made in Example \[ex:06\] are explained by the following proposition which has several interesting implications.
\[prop:01\] Let ${\mathbf{M}}_{i,\bullet}$ be the $p$ vector of the marker values of individual $i$ and let $f({\mathbf{M}}_{i,\bullet}): \mathbb{R}^p \rightarrow \mathbb{R}$ be a polynomial of total degree $D$ in the marker data. Moreover, let $\tilde{{\mathbf{M}}}:= {\mathbf{M}}- \mathbf{1}_n {\mathbf{P}}^t$ be a translation of the marker coding and let us define a polynomial $\tilde{f}$ in the translated variables $\tilde{{\mathbf{M}}}$ by $\tilde{f}(\tilde{{\mathbf{M}}}_{i,\bullet}):= f(\tilde{{\mathbf{M}}}_{i,\bullet} + {\mathbf{P}}^t)=f({\mathbf{M}}_{i,\bullet})$. Then for any data ${\mathbf{y}}$, the sum of squared residuals (SSR) will be identical $$\sum_{i=1,...,n} (y_i - f({\mathbf{M}}_{i,\bullet}))^2 = \sum_{i=1,...,n} (y_i - \tilde{f}(\tilde{{\mathbf{M}}}_{i,\bullet}))^2$$ and for any monomial $m$ of highest total degree $D$, the corresponding coefficient $a_m$ of $f({\mathbf{M}}_{i,\bullet})$ and $\tilde{a}_m$ of $\tilde{f}(\tilde{{\mathbf{M}}}_{i,\bullet})$ will be identical: $$a_m = \tilde{a}_m.$$
The fact that the SSR remains the same, results from the definition of the polynomials. To see that the coefficients of monomials of highest total degree D are identical, choose a monomial $m(M_{l_1},M_{l_2},...,M_{l_D})$ of the loci $l_1,...,l_D$ of total degree $D$ of $f$. Multiplying the factors of $m(\tilde{M}_{l_1}+P_{l_1},\tilde{M}_{l_2}+P_{l_2},...,\tilde{M}_{l_D}+P_{l_D})$ gives the same monomial $m(\tilde{M}_{l_1},\tilde{M}_{l_2},...,\tilde{M}_{l_D})$ as a summand of highest total degree, plus additional monomials of lower total degree. Thus, the coefficients of monomials of total degree $D$ remain the same.
Proposition \[prop:01\] has the very simple statement that if we have a certain fit $f$ based on ${\mathbf{M}}$, and we use the translated marker coding $\tilde{{\mathbf{M}}}$, the polynomial $\tilde{f}$ will fit the data with the same SSR and with the same predictions $\hat{{\mathbf{y}}}$ (due to the definition of $\tilde{f}$). Moreover, the coefficients of monomials of highest total degree will be the same.\
Since OLS is defined only by the minimal SSR, this also means that it is invariant to any translation of the coding, provided that $\tilde{f}$ of Proposition \[prop:01\] is a valid fit. To make sure that $\tilde{f}$ is a valid fit, the possibility to adapt coefficients of monomials of lower total degrees is required. We cannot adapt the regression completely if certain coefficients are forced to zero by the model structure. If a coefficient is equal to zero in $f$, it may be different from zero in $\tilde{f}$. We illustrate this with an example.
\[ex:07\] Let us consider the data ${\mathbf{M}}$ and ${\mathbf{y}}$ of Example \[ex:06\] but with the assumption that marker $2$ does not have an additive effect. Then $$\begin{pmatrix}
\hat{\mu} \\
\hat{\beta}_1 \\
\hat{h}_{1,2}
\end{pmatrix}_{\mbox{\small OLS}} = \begin{pmatrix}
3.71 \\
-2.098\\
-0.012\\
\end{pmatrix}
\qquad \mbox{and} \qquad \begin{pmatrix}
\tilde{\mu} \\
\tilde{\beta}_1 \\
\tilde{h}_{1,2}
\end{pmatrix}_{\mbox{\small OLS}}= \begin{pmatrix}
0.334\\
-2.11 \\
-1.162
\end{pmatrix}$$ and also the estimates $\hat{{\mathbf{y}}}$ and $\tilde{{\mathbf{y}}}$ are different.
Example \[ex:07\] illustrates that “completeness" of the model is required to have the possibility to adapt to translations of the coding. We define this property more precisely.
\[def01\] Let ${\mathbf{M}}_{i,\bullet}$ be the $p$ vector of the marker values of individual $i$ and let $f({\mathbf{M}}_{i,\bullet}): \mathbb{R}^p \rightarrow \mathbb{R}$ be a polynomial of total degree $D$ in the marker data. The polynomial model $f$ is called complete if for any monomial ${\mathbf{M}}_{i,j_1}^{d_1} {\mathbf{M}}_{i,j_2}^{d_2}\cdot \cdot \cdot {\mathbf{M}}_{i,j_m}^{d_m}$ of $f$, all monomials $${\mathbf{M}}_{i,j_1}^{\delta_1} {\mathbf{M}}_{i,j_2}^{\delta_2}\cdot \cdot \cdot {\mathbf{M}}_{i,j_m}^{\delta_m} \qquad \forall \; 0 \leq {\delta_1} \leq d_1, \; \forall \; 0 \leq {\delta_2} \leq d_2, \; ...\; , \forall \; 0 \leq {\delta_m} \leq d_m$$ are included with an coefficient to be estimated.
Although, this definition seems rather abstract, its meaning can be understood easily by an example. Let us consider Eq. (\[eq:07\]). Its monomials are of shape $M_{i,k}$ or $M_{i,k}M_{i,l}$. For $M_{i,k}$, Definition \[def01\] states that $M_{i,k}^0=1$ and $M_{i,k}^1$ have to be included, which is obviously the case. For $M_{i,k}M_{i,l}$, $M_{i,k}^0=1$, $M_{i,k}^1$ and $M_{i,k}^1M_{i,j}^1$ have to be included, which is also true. Thus, the model is complete. Analogously, if we also include the interactions $M_{i,k}^2$, the model remains complete. Contrarily, Example \[ex:07\] is based on the model $$y_i = \mu+ \beta_1 M_{i,1} + h_{1,2}M_{i,1}M_{i,2} + \epsilon_i.$$ Since $M_{i,1}M_{i,2}$ is included with a coefficient to be estimated, $M_{i,1}$ and $M_{i,2}$ have to be included to make the model complete. Since $M_{i,2}$ is not included, the polynomial is not complete.\
Given that the model is complete, Proposition \[prop:01\] has various implications. The following corollaries explain the results observed in our examples and highlight some additional properties of penalized regression methods in general. For all statements, it is assumed that penalty factors remain unchanged and that the model is complete.
\[cor:01\] For a model of any total degree $D$, the OLS estimates of the coefficients of highest total degree as well as the predictions $\hat{{\mathbf{y}}}$ are invariant with respect to translations of the marker coding.
Corollary \[cor:01\] is a result of the OLS method being defined only by the SSR, and $f$ and the corresponding $\tilde{f}$ of Proposition \[prop:01\] fitting the data with the same SSR when their respective coding is used. The statement of Corrolary \[cor:01\] has been observed in Example \[ex:06\], where the OLS fits for $\hat{{\mathbf{y}}}$ are identical when the coding is translated, and where the estimated coefficients $\hat{h}_{1,2}$ of highest total degree remain unchanged.
\[cor:02\] For a polynomial model of total degree $D$, and a penalized regression which only penalizes the coefficients of monomials of total degree $D$, the estimates of the coefficients of monomials of total degree $D$, as well as the predictions $\hat{{\mathbf{y}}}$ are invariant with respect to translations of the marker coding.
Corollary \[cor:02\] is a result of the following observation: for each $f$, its corresponding $\tilde{f}$ will have the same SSR (each polynomial with its respective coding), and the same coefficients of highest total degree. Thus, it will have the same value for the target function which we aim to minimize (The target function is the analogue of Eq. (\[eq:05\]) with a penalty on only the coefficients of monomials of highest total degree). Because this is true for any polynomial $f$, it is in particular true for the solution minimizing the target function. A central point of Corollary \[cor:02\] is that it is valid for any penalty on the size of the estimated coefficients of highest total degree. The sufficient condition is that only these coefficients of highest total degree are penalized.
\[cor:03\] RRBLUP predictions $\hat{{\mathbf{y}}}$ are invariant with respect to translations of the marker coding.
Corollary \[cor:02\] applied to complete models of total degree $1$ gives the result of Corollary \[cor:03\], that is RRBLUP being invariant to translations of the marker coding. This fact has been previously proven using a marginal likelihood setup [@stranden11], or the mixed model equations [@Martini17].
\[cor:04\] An additive least absolute shrinkage and selection operator (LASSO) regression [@tibshirani96] based on a polynomial model of total degree 1 and $\ell_1$ penalizing the additive marker effects but not the intercept, is invariant to translations of the marker coding.
Corollary \[cor:04\] is a special case of Corollary \[cor:02\].\
We give a small example, highlighting cases which are not invariant to translations of the marker coding. We recommend to use the data of Example \[ex:06\] to validate the statements.\
- Pure ridge regression of an additive model of Eq. (\[eq:01\]) with a penalty on the size of $\mu$ (“random intercept”) is not invariant to translations.
- RRBLUP with the fixed effect forced to zero is not invariant to translations of the marker coding.
- An extended LASSO $\ell_1$ penalizing additive effects and interactions is not invariant to translations of the coding.
Proposition \[prop:01\] stated that the coefficients of monomials of highest total degree $D$ of $f$ and $\tilde{f}$ will be identical. This statement can even be generalized for some situations. Consider for instance the model $$y_i = f(M_{i,1},M_{i,2},M_{i,3}) = \mu + \beta_1 M_{i,1} + \beta_2 M_{i,2} + \beta_3 M_{i,3} + h_{2,3} M_{i,2}M_{i,3} + \epsilon_i$$ The model is a polynomial $f$ of total degree 2. Thus, Proposition \[prop:01\] states that the coefficient of monomial $M_{i,2}M_{i,3}$ will be identical for $f$ and $\tilde{f}$. However, since $M_{i,1}$ is not included in any other monomial, its coefficient will also be identical for both polynomials. We did not generalize Proposition \[prop:01\] into this direction to make the manuscript not more technical than necessary. The statement we made in Proposition \[prop:01\] is sufficient to explain the observations related to genomic prediction models.
Discussion {#discussion .unnumbered}
==========
The illustrated problem of the coding having an impact on the estimates of interactions in penalized regressions is essential for quantitative genetics, where Hadamard products are often used to model interaction such as epistasis or gene by environment interaction [@perez17]. Hadamard products of covariance matrices represent exact reformulations of certain interaction effect models [@jiang15; @Martini16]. In particular, our observations illustrate once more that the size of interaction effect estimates should be interpreted with caution because a biological meaning is not necessarily given.\
It should be highlighted, that the problem does not seem to be a consequence of non-orthogonality of the predictor variables (marker values and their products), since these problems would not appear in an OLS regression (provided that a unique solution exists), where the variables have the same coding and thus the same angle.\
Finally, note that it has been reported that a Gaussian reproducing kernel regression [@Morota14] can be interpreted as a limit of a polynomial regression with increasing total degree (and all possible monomials) [@jiang15]. Being a limit case of a method which is affected by translations of the coding, the question appears why the Gaussian kernel regression is invariant to translations of the marker coding. It may be interesting to reconsider the limit behavior from a theoretical point of view.
Conclusion {#conclusion .unnumbered}
==========
We identified the cause of the coding-dependent performance of epistasis effects models. Our results were motivated by ridge regression, but do equally hold for many other types of penalized regressions, for instance for the $\ell_1$ penalized LASSO. The fact that the estimated effect sizes depend on the coding highlights once more that estimated interaction effect sizes should be interpreted with caution with regard to their biological, mechanistic meaning. Moreover, the problem of coding is not only present for marker by marker interaction, but for any mixed model in which interactions are modeled by Hadamard products of covariance matrices, in particular also for gene by environment (G x E) models.
[99]{}
Meuwissen, T. H. E., Hayes, B. J., and Goddard, M. E. (2001) Prediction of total genetic value using genome-wide dense marker maps. *Genetics* 157: 1819–1829.
Schaeffer, L. R. (2006) Strategy for applying genome-wide selection in dairy cattle. *J Anim Breed Genet* 123: 218–223.
Habier, D., Fernando, R. L., and Dekkers, J. C. M. (2007) The impact of genetic relationship information on genome-assisted breeding values. *Genetics* 177: 2389–2397.
Hayes, B. J., Visscher, P. M., and Goddard, M. E. (2009) Increased accuracy of artificial selection by using the realized relationship matrix. *Genet Res* 91: 47–60.
Hayes, B. J., Cogan, N. O. I., Pembleton, L. W., Goddard, M. E., Wang, J., Spangenberg, G. C., and Forster, J. W. (2013) Prospects for genomic selection in forage plant species. *Plant Breeding* 132: 133–143.
Falconer, D. S., and Mackay, T. F. C. (1996) Introduction to Quantitative Genetics. Pearson Education, London, 978-0582243026.
Schaeffer, L. R. (2004) Application of random regression models in animal breeding. *Livest Prod Sci* 86(1): 35–45.
Henderson, C. R. (1975) Best linear unbiased estimation and prediction under a selection model. *Biometrics* 31: 423–447.
Henderson, C. R., and Quaas, R. L. (1976) Multiple trait evaluation using relatives’ records. *J Anim Sci* 43: 1188–1197.
Henderson, C. R. (1977) Best linear unbiased prediction of breeding values not in the model for records. *J Dairy Sci* 60: 783–787.
He, D., and Parida, L. (2016) Does encoding matter? a novel view on the quantitative genetic trait prediction problem. *BMC Bioinformatics* 17(9): 272.
Martini, J. W. R. Gao, N., Cardoso, D. F., Wimmer, V., Erbe, M., Cantet, R. J. C., and Simianer, H. (2017) Genomic prediction with epistasis models: on the marker-coding-dependent performance of the extended [GBLUP]{} and properties of the categorical epistasis model ([CE]{}). *BMC Bioinformatics* 18: 3.
Ober, U., Huang, W., Magwire, M., Schlather, M., Simianer, H., and Mackay, T. F. C. (2015). Accounting for genetic architecture improves sequence based genomic prediction for a [*rosophila*]{} fitness trait. *PLOS ONE* 10: e0126880
Jiang, Y., and Reif, J. C. (2015) Modeling epistasis in genomic selection. *Genetics* 201: 759–768.
Martini, J. W. R., Wimmer, V., Erbe, M., and Simianer, H. (2016). Epistasis and covariance: how gene interaction translates into genomic relationship. *Theor Appl Genet* 129: 963–976.
VanRaden, P. M. (2008). Efficient methods to compute genomic predictions. *J Dairy Sci* 91 (11): 4414–4423.
lvarez-Castro, J. M., and Carlborg, [Ö]{}. (2007). A unified model for functional and statistical epistasis and its application in quantitative trait loci analysis. *Genetics* 176(2): 1151–1167.
Vitezica, Z. G., Legarra, A., Toro, M. A., Varona, L. (2017) Orthogonal estimates of variances for additive, dominance, and epistatic effects in populations. *Genetics* 206(3): 1297–1307.
Strand[é]{}n, I., and Christensen, O. F. (2011) Allele coding in genomic evaluation. *Genet Sel Evol* 43(1): 25.
Tibshirani, R. (1996) Regression shrinkage and selection via the [L]{}asso. *J Roy Stat Soc B Met* 58(1): 267–288.
P[é]{}rez-Rodr[í]{}guez, P., Crossa, J., Rutkoski, J., Poland, J., Singh, R., Legarra, A., Autrique, E., de los Campos, G., Burgue[ñ]{}o, J. and Dreisigacker, S. (2017) Single-step genomic and pedigree genotype$\times$ environment interaction models for predicting wheat lines in international environments. *The plant genome* 10(2).
Morota, G., and Gianola, D. (2014) Kernel-based whole-genome prediction of complex traits: a review. *Front Genet* 5: 363.
Author’s contribution {#authors-contribution .unnumbered}
=====================
JWRM: Proposed to consider the topic; derived the theoretical results; JWRM and FR wrote the manuscript; All authors: Discussed the research
Acknowledgements {#acknowledgements .unnumbered}
================
JWRM thanks KWS SAAT SE for financial support during his PhD. FR thanks Universidad del Pacífico for financial support.
| {
"pile_set_name": "ArXiv"
} |
[Some Analytics for Steiner Minimal Tree Problem for Four Terminals]{}
[Alexei Yu. Uteshev[^1]]{}
[*St. Petersburg State University, St. Petersburg, Russia*]{}
Given the coordinates of four terminals in the Euclidean plane we present explicit formulas for Steiner point coordinates for Steiner minimal tree problem. We utilize the obtained formulas for evaluation of the influence of terminal coordinates on the loci of Steiner points.
Introduction {#SIntro}
============
**Problem.** Given set of points $ \mathbb P= \{P_j\}_{j=1}^n $ in the Euclidean plane, find a system of line segments such that their union forms a connected set $ \mathbb U $ containing $ \mathbb P $, and such that the total length of the line segments is minimized.
This problem of construction the shortest possible network interconnecting the points of the set $ \mathbb P $ is known as the (**Euclidean**) **Steiner minimal tree problem** (**SMT** problem).
The stated problem in its particular case of $ n=3 $ noncollinear points is known since 1643 as (classical) Fermat-Torricelli problem. It has a unique solution which
- either coincides with the system of three segments connecting the points $ P_1,P_2,P_3 $ with the so-called *Fermat* or *Fermat-Torricelli point* of the triangle $ P_1P_2P_3 $; this is the case where every angle of the triangle $ P_1P_2P_3 $ is less than $ 2\pi/3 $;
- or, in case where there exists a vertex $ P_j $ with the triangle angle equal to or greater than $ 2\pi/3 $, it coincides with the system of two triangle sides meeting at $ P_j $.
For the sake of concordance with the foregoing definition, we will refer to the Fermat-Torricelli point of the triangle as to its **Steiner point**.
The case of $ n>3 $ points is much harder in treatment. First of all, the Sreiner minimal tree problem should be distinguished from another problem frequently called the *Fermat-Torricelli problem*, the latter consists in finding a junction point $ S $ which minimize $ \sum_{j=1}^n |SP_j| $. However the solution for this problem even for the case of $ n=4 $ points located at the vertices of a square does not provide the shortest network. It turns out that two junction points are necessary . Thus, for the general case, the problem is reduced to determining of the set[^2] $ \mathbb S=\{S_1,\dots,S_k\} $ of $ k\ge 0 $ extra junction points and the line segments connecting them with the points of the set $ \mathbb P $. We refer to [@Brazil_2014] for a comprehensive and intriguing historical review of the theory.
The general results for the problem are formulated in terms of Graph Theory [@Cock70]. A tree $ \mathbb U $ with vertices $ \mathbb P \cup \mathbb S $ and rectilinear edges linking certain pairs of vertices is a **Steiner tree on** $ \mathbb P $ iff it has the following properties
- $ \mathbb U $ is non-self-intersecting.
- The *valency*[^3] of every $ S_i $ equals $ 3 $.
- The valency of every $ P_j $ is $ \le 3 $.
- Each $ S_j $ is the Steiner point of the triangle formed by the points which directly connect $ S_j $ in $ \mathbb U $.
- $ 0 \le k \le n-2 $.
The points $ P_j $ are usually called **terminals** while the *junction* points $ S_i $ are referred to as **Steiner points** of a Steiner tree.
For any given $ \mathbb P $, there are finitely many Steiner trees and at least one of them is the Steiner minimal tree [@b1]. A **full Steiner tree** on $ \mathbb P $ is a tree which satisfies **(P1)**-**(P4)** and also has $ k=n-2 $. Any Steiner minimal tree is a union of full Steiner subtrees.
The present paper is devoted to the four terminals SMT problem. We restrict ourselves here with the cases of full Steiner trees, i.e. the trees with exactly two Steiner points. We aim at finding the conditions for existence of such trees as well as the coordinates of Steiner points; both problems are to be solved in terms of terminal coordinates. The paper can be treated as a continuation of the previous work [@Uteshev14] on the generalized Fermat-Torricelli (three terminals) problem.
In comparison with the admirable elegance of geometric approaches for the problem, the representation of its solution in final analytical form (Section \[Stein4\]) looks like prosaic, cumbersome and boring. The only arguments which might excuse the present author for their creation are the following:
- these formulas are universal and yield the exact result (i.e., free of truncation errors);
- they provide one with an opportunity to evaluate the influence of parameters involved in the problem on its solution.
The latter problem is briefly touched in Section \[Scor\] where we investigate the dynamics of Steiner points under the variation of a certain terminal.
**Notation.** We denote the coordinates of the terminal $ P_j $ by $ (x_j,y_j) $, $$r_{jk} = |P_jP_k|= \sqrt{(x_j-x_k)^2+(y_j-y_k)^2} \ ;$$ and by $$\langle \overrightarrow{A_1A_2}, \overrightarrow{A_3A_4} \rangle$$ the inner product of the vectors.
For the terminals $ \{P_j\}_{j=1}^4 $ and Steiner points $ S_1, S_2 $, the representation $$\begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array}$$ has the meaning that terminals $ P_1 $ and $ P_2 $ are both connected to Steiner point $ S_1 $, that $ P_3 $ and $ P_4 $ are both connected to Steiner point $ S_2 $; finally $ S_1 $ and $ S_2 $ are connected with each other. We will refer to such a representation as to **topology of Steiner tree**.
Three terminals {#Stein3}
===============
We first recall the geometrical construction of Steiner point for three terminals $ P_1, P_2 $ and $ P_3 $ . The algorithm outlined in the following example is a combination of Torricelli and Simpson approaches for the construction of Fermat-Torricelli point for the triangle $P_1P_2P_3 $.
\[Ex0\] For the terminals $$P_1=(4,4), \ P_2= (2,1), P_4=(7,1)$$ construct the Steiner minimal tree.
**Solution.** First construct the equilateral triangle on the segment $ P_1P_2 $ and outside the triangle $ P_1P_2P_3 $ (Fig. 1).
{width="50.00000%"}
Denote its third vertex by $ Q_1 $. Next, draw the circle circumscribing this triangle; we denote it as $ \mathfrak{C} $. Finally draw the segment through $ Q_1 $ and $ P_3 $.
{width="50.00000%"}
The intersection point of the segment with the circle $ \mathfrak{C} $ is Steiner points for the tree (Fig. 2). The length of SMT equals $|Q_1P_3| $.
Analytical solution for the SMT problem is given by the following result[^4] [@Uteshev14]:
\[Steiner3\] Let all the angles of the triangle $ P_1P_2P_3 $ be less than $ 2\pi/3 $, or, equivalently: $$r_{12}^2+r_{13}^2+r_{12}r_{13}-r_{23}^2>0,\ r_{23}^2+r_{12}^2+r_{12}r_{23}-r_{13}^2>0,\ r_{13}^2+r_{23}^2+r_{13}r_{23}-r_{12}^2>0 \ .$$ The coordinates of Steiner point for the terminals $ P_1,P_2,P_3 $ are as follows: $$x_{\ast}=\frac{\kappa_1\kappa_2\kappa_3}{2 \sqrt{3} |\mathfrak S| d^2} \left(\frac{x_1}{\kappa_1}+\frac{x_2}{\kappa_2}+\frac{x_3}{\kappa_3} \right), \
y_{\ast}=\frac{\kappa_1\kappa_2\kappa_3}{2 \sqrt{3} |\mathfrak S| d^2} \left(\frac{y_1}{\kappa_1}+\frac{y_2}{\kappa_2}+\frac{y_3}{\kappa_3} \right)
\label{FTp}$$ with the length of the SMT equal to $$d=\sqrt{\frac{\kappa_1+\kappa_2+\kappa_3}{\sqrt{3}}}=\sqrt{\frac{r_{12}^2+r_{13}^2+r_{23}^2}{2}+ \sqrt{3}\, |\mathfrak S|} .
\label{Stein-dist3}$$ Here $$\mathfrak S=\left|\begin{array}{ccc} 1 & 1 & 1 \\ x_1 & x_2 & x_3 \\ y_1 & y_2 & y_3 \end{array} \right| \, ,
\label{S}$$ (and therefore $ | \mathfrak S| $ equals the doubled area of the triangle $ P_1P_2P_3 $), while $$\kappa_1=\frac{\sqrt{3}}{2}(r_{12}^2+r_{13}^2-r_{23}^2)+|\mathfrak S| , \ \kappa_2=\frac{\sqrt{3}}{2}(r_{23}^2+r_{12}^2-r_{13}^2)+|\mathfrak S| , \
\kappa_3=\frac{\sqrt{3}}{2}(r_{13}^2+r_{23}^2-r_{12}^2)+|\mathfrak S | .$$
\[cor1\] Let the loci of terminals $ P_1 $ and $ P_2 $ be fixed in the $ (x,y) $-plane while the locus of the terminal $ P_3 $ be variable with the restrictions that the points $ P_1, P_2 $ and $ P_3 $ are numbered counterclockwise and satisfy the conditions of Theorem \[Steiner3\]. The locus of Steiner point for terminals $ P_1,P_2,P_3 $ coincides with the arc of the circle $ \mathfrak{C} $ constructed via the algorithm outlined in solution of Example \[Ex0\]. It has its center at $$C=\left(\frac{1}{2}x_1+\frac{1}{2}x_2-\frac{1}{2\sqrt{3}}y_1+\frac{1}{2\sqrt{3}}y_2\, ,\ \frac{1}{2\sqrt{3}}x_1-\frac{1}{2\sqrt{3}}x_2+\frac{1}{2}y_1+\frac{1}{2}y_2 \right)
\label{circle_c}$$ and its radius equal to $ r_{12}/\sqrt{3} $. This circle is the circumscribing one for the equilateral triangle $ P_1P_2Q_1 $ with $$Q_1=\left(\frac{1}{2}x_1+\frac{1}{2}x_2-\frac{\sqrt{3}}{2}y_1+\frac{\sqrt{3}}{2}y_2\, ,\ \frac{\sqrt{3}}{2}x_1-\frac{\sqrt{3}}{2}x_2+\frac{1}{2}y_1+\frac{1}{2}y_2 \right) \, .
\label{Q1}$$
Four terminals {#Stein4}
==============
We will start the treatment of this case with recalling the geometrical algorithm for Steiner tree construction .
\[Ex1\] For the terminals $$P_1=(2,6), \ P_2= (1,1), P_3=(9,2), P_4=(6,7)$$ construct Steiner tree of the topology $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $.
**Solution.** First construct two equilateral triangles on the sides $ P_1P_2 $ and $ P_3P_4 $ and outside the quadrilateral $ P_1P_2P_3P_4 $ (Fig.3).
{width="70.00000%"}
Their third vertices are denoted by $ Q_1 $ and $ Q_2 $ correspondingly. Next, draw two circles circumscribing these triangles (Fig.4).
{width="70.00000%"}
Finally draw the segment through $ Q_1 $ and $ Q_2 $ (Fig.5).
{width="70.00000%"}
The intersection points of the segment with the circles are Steiner points for the tree. The length of the obtained tree (Fig.6) equals $ |Q_1Q_2| $.
{width="70.00000%"}
As for the general case, the problem of applicability of the suggested algorithm for Steiner tree construction is in question: one can easily imagine such a quadrilateral for which the algorithm results in points $ S_1 $ and $ S_2 $ located *outside* it; thus they do not satisfy the condition **(P4)** mentioned in Section \[SIntro\].
**Assumption.** Hereinafter we will treat the case where the terminals $ \{P_j\}_{j=1}^4 $, while counted counterclockwise, compose a convex quadrilateral $P_1P_2P_3P_4$.
The condition of convexity is a necessary one for the existence of a full Steiner tree for the given set of terminals [@Pollak].
\[Steiner4\] Let the preceding assumption be fulfilled. Set $$\begin{aligned}
\tau_1 &= & 2\, x_1 - x_2 -2\, x_3 + x_4 +\sqrt{3} (y_2- y_4), \label{tau1} \\
\tau_2 & = & - x_1 +2\, x_2 + x_3 -2\, x_4 +\sqrt{3} (y_3- y_1), \label{tau2}\end{aligned}$$ and $$\eta_1=-\frac{1}{\sqrt{3}}(\tau_1+2\,\tau_2),\ \eta_2 = \frac{1}{\sqrt{3}}(2\,\tau_1+\tau_2) \label{eta12} \ .$$ If all the values $$\delta = -(x_1-x_3) \eta_1 + (y_1-y_3) \tau_1 ,
\label{delta}$$ $$\left\{\begin{array}{cc}
\begin{array}{c}
\delta_1 = (x_1-x_2) \eta_2 - (y_1-y_2) \tau_2, \\
\delta_2 = (x_1-x_2) \eta_1 - (y_1-y_2) \tau_1,
\end{array}
&
\begin{array}{c}
\delta_3 = -(x_3-x_4) \eta_2 + (y_3-y_4) \tau_2, \\
\delta_4 = -(x_3-x_4) \eta_1 + (y_3-y_4) \tau_1
\end{array}
\end{array}
\right.
\label{delta1234}$$ are positive then there exists a Steiner tree of the topology $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $. The coordinates of Steiner point $ S_1 $ are as follows: $$x_{\ast}=x_1 - \frac{\sqrt{3}}{2} \cdot \frac{\delta_1 \tau_1}{\tau_1^2+\tau_1 \tau_2+ \tau_2^2}, \quad y_{\ast}=y_1 - \frac{\sqrt{3}}{2} \cdot \frac{\delta_1 \eta_1}{\tau_1^2+\tau_1 \tau_2+ \tau_2^2}
\label{Steiner pt1}$$ and those of $ S_2 $: $$x_{\ast \ast}=x_3 + \frac{\sqrt{3}}{2} \cdot \frac{\delta_3 \tau_1}{\tau_1^2+\tau_1 \tau_2+ \tau_2^2}, \quad
y_{\ast \ast}=y_3 + \frac{\sqrt{3}}{2} \cdot \frac{\delta_3 \eta_1}{\tau_1^2+\tau_1 \tau_2+ \tau_2^2} \ .
\label{Steiner pt2}$$ The length of the tree equals to $$d= \sqrt{\frac{\tau_1^2+\tau_1 \tau_2+ \tau_2^2}{3}} \, . \label{Stein-dist4}$$
**Proof.** Consider the objective function corresponding to the assumed topology of the tree $$F(x_{\ast},y_{\ast},x_{\ast \ast},y_{\ast \ast})=|S_1P_1|+|S_1P_2|+|S_1S_2|+|S_2P_3|+|S_2P_4| \ .
\label{obj_fun}$$ The system providing its stationary points is as follows $$\left\{
\begin{array}{ccc}
\displaystyle \frac{x_{\ast}-x_1}{|S_1P_1|} + & \displaystyle \frac{x_{\ast}-x_2}{|S_1P_2|} + & \displaystyle \frac{x_{\ast}-x_{\ast \ast}}{|S_1S_2|} = 0, \\
\displaystyle \frac{y_{\ast}-y_1}{|S_1P_1|} + & \displaystyle \frac{y_{\ast}-y_2}{|S_1P_2|} + & \displaystyle \frac{y_{\ast}-y_{\ast \ast}}{|S_1S_2|} = 0, \\
\displaystyle \frac{x_{\ast \ast}-x_3}{|S_2P_3|} + & \displaystyle \frac{x_{\ast \ast}-x_4}{|S_2P_4|} + & \displaystyle \frac{x_{\ast \ast}-x_{\ast}}{|S_1S_2|} = 0, \\
\displaystyle \frac{y_{\ast \ast}-y_3}{|S_2P_3|} + & \displaystyle \frac{y_{\ast \ast}-y_4}{|S_2P_4|} + & \displaystyle \frac{y_{\ast \ast}-y_{\ast}}{|S_1S_2|} = 0.
\end{array}
\right.
\label{stat_points}$$ To prove the first of these formulas for the values of $ x_{\ast},y_{\ast}, x_{\ast \ast},y_{\ast \ast} $ given by (\[Steiner pt1\]) and (\[Steiner pt2\]), we will first deduce the alternative representation for these values: $$x_{\ast}=x_2 - \frac{\sqrt{3}}{2} \cdot \frac{\delta_2 \tau_2}{\tau_1^2+\tau_1 \tau_2+ \tau_2^2}, \quad y_{\ast}=y_2 - \frac{\sqrt{3}}{2} \cdot \frac{\delta_2 \eta_2}{\tau_1^2+\tau_1 \tau_2+ \tau_2^2} \, .
\label{Steiner pt10}$$ The equivalence of the first formula to its counterpart from (\[Steiner pt1\]) stems from $$\delta_1 \tau_1 - \delta_2 \tau_2 \stackrel{(\ref{delta1234})}{=} \left[(x_1-x_2) \eta_2 - (y_1-y_2) \tau_2\right] \tau_1 -
\left[(x_1-x_2) \eta_1 - (y_1-y_2) \tau_1\right] \tau_2$$ $$=(x_1-x_2)\left[\eta_2 \tau_1-\eta_1 \tau_2\right] \stackrel{(\ref{eta12})}{=} (x_1-x_2) \left[ \frac{1}{\sqrt{3}}(2\,\tau_1+\tau_2) \tau_1 +\frac{1}{\sqrt{3}}(\tau_1+2\,\tau_2) \tau_2 \right]$$ $$= \frac{2}{\sqrt{3}} (x_1-x_2) \left[\tau_1^2+ \tau_1 \tau_2 + \tau_2^2 \right] \ .$$ The second formula from (\[Steiner pt10\]) can be proven similarly.
Next, with the aid of representations (\[Steiner pt1\]) and (\[Steiner pt10\]), one can prove that $$|S_1P_1|$$ $$= \sqrt{(x_{\ast}-x_1)^2+(y_{\ast}-y_1)^2}\stackrel{(\ref{Steiner pt1})}{=}\sqrt{\frac{3}{4} \cdot \frac{\delta_1^2 \tau_1^2 + \delta_1^2 \eta_1^2}{(\tau_1^2+ \tau_1 \tau_2 + \tau_2^2)^2}} \stackrel{(\ref{eta12})}{=} \sqrt{\frac{\delta_1^2}{\tau_1^2+ \tau_1 \tau_2 + \tau_2^2}}$$ $$= \frac{\delta_1}{\sqrt{\tau_1^2+ \tau_1 \tau_2 + \tau_2^2}}
\label{S1P1}$$ (the last equality follows from the assumption $ \delta_1 > 0 $ stated in the theorem) and, similarly, that $$|S_1P_2|= \frac{\delta_2}{\sqrt{\tau_1^2+ \tau_1 \tau_2 + \tau_2^2}} \ .
\label{S1P2}$$
Finally we intend to deduce the equalities $$x_{\ast}-x_{\ast \ast}= \frac{\sqrt{3}}{2} \cdot \frac{(\tau_1+\tau_2) \delta}{\tau_1^2+ \tau_1 \tau_2 + \tau_2^2}, \quad
y_{\ast}-y_{\ast \ast}= \frac{1}{2} \cdot \frac{ (\tau_1-\tau_2) \delta}{\tau_1^2+ \tau_1 \tau_2 + \tau_2^2}
\label{minus}$$ from which evidently follows[^5] $$|S_1S_2|= \frac{\delta}{\sqrt{\tau_1^2+ \tau_1 \tau_2 + \tau_2^2}} \ .
\label{S1S2}$$ The proof of the first of the formulas (\[minus\]) is based on the equality $$\delta_1+\delta_3 = (x_1-x_3)(\eta_1+\eta_2)-(y_1-y_3)(\tau_1+\tau_2)
\label{eq_w0}$$ the validity of which can be established as follows: $$\delta_1+\delta_3 \stackrel{(\ref{delta1234})}{=} (x_1-x_2-x_3+x_4) \eta_2+(-y_1+y_2+y_3-y_4) \tau_2$$ $$=\frac{1}{\sqrt{3}} (x_1-x_2-x_3+x_4) (2\tau_1+\tau_2) +(-y_1+y_2+y_3-y_4) \tau_2$$ $$=\frac{2}{\sqrt{3}}(x_1-x_2-x_3+x_4)\tau_1+\left[ \frac{1}{\sqrt{3}} (x_1-x_2-x_3+x_4)+(-y_1+y_2+y_3-y_4) \right] \tau_2$$ $$=\left[ \frac{1}{\sqrt{3}} (x_1-x_3)-y_1+y_3 \right] \tau_1 +\left[ \frac{1}{\sqrt{3}} (x_1-2\,x_2-x_3+2\,x_4)+y_1-y_3 \right] \tau_1$$ $$+
\left[ \frac{1}{\sqrt{3}} (2\,x_1-x_2-2\,x_3+x_4)+y_2-y_4 \right] \tau_2 + \left[ \frac{1}{\sqrt{3}} (-x_1+x_3)-y_1+y_3 \right]\tau_2$$ $$=\frac{1}{\sqrt{3}}(x_1-x_3)(\tau_1-\tau_2)-(y_1-y_3)(\tau_1+\tau_2)=(x_1-x_3)(\eta_1+\eta_2)-(y_1-y_3)(\tau_1+\tau_2) \, .$$ We now start deducing the first formula from (\[minus\]): $$x_{\ast}-x_{\ast \ast} \stackrel{(\ref{Steiner pt1}), (\ref{Steiner pt2})}{=} x_1-x_3-
\frac{\sqrt{3}}{2} \cdot \frac{(\delta_1+\delta_3) \tau_1}{\tau_1^2+\tau_1 \tau_2+ \tau_2^2} \, .
\label{eq_w1}$$ Let us manipulate with the numerator of the fraction from the last formula: $$(\delta_1+\delta_3)\tau_1 \stackrel{(\ref{eq_w0})}{=}\left[(x_1-x_3)(\eta_1+\eta_2)-(y_1-y_3)(\tau_1+\tau_2) \right]\tau_1$$ $$=\left[(x_1-x_3)(\eta_1+\eta_2)\tau_1-(x_1-x_3)(\tau_1+\tau_2) \eta_1 \right] +
\left[(x_1-x_3)(\tau_1+\tau_2)\eta_1-(y_1-y_3)(\tau_1+\tau_2) \tau_1 \right]$$ $$=(x_1-x_3)\left[ (\eta_1+\eta_2)\tau_1- (\tau_1+\tau_2)\eta_1 \right] +(\tau_1+\tau_2)\left[(x_1-x_3) \eta_1 - (y_1-y_3) \tau_1 \right]$$ $$\stackrel{(\ref{eta12}),(\ref{delta})}{=}(x_1-x_3)\left[ \frac{1}{\sqrt{3}}(\tau_1-\tau_2)\tau_1 + \frac{1}{\sqrt{3}}(\tau_1+\tau_2)(\tau_1+2\, \tau_2)
\right]- (\tau_1+\tau_2) \delta$$ $$=\frac{2}{\sqrt{3}} (x_1-x_3) (\tau_1^2+ \tau_1 \tau_2 + \tau_2^2) - (\tau_1+\tau_2) \delta \ .$$ Substitution of the last expression into the right-hand side of (\[eq\_w1\]) results in the first formula from (\[minus\]). The second formula from (\[minus\]) can be deduced similarly.
Applying now the formulas (\[Steiner pt1\]) and (\[S1P1\]) for the first summand from the left-hand side of the first formula from (\[stat\_points\]), the formulas (\[Steiner pt10\]) and (\[S1P2\]) for the second summand, (\[minus\]) and (\[S1S2\]) for the third summand, one gets the true equality. The validity of the remained equalities can be established similarly.
We collect now all the formulas for the segment lengths $$|P_1S_1|=\frac{\delta_1}{\sqrt{3} d},\ |P_2S_1|=\frac{\delta_2}{\sqrt{3} d},\ |P_3S_2|=\frac{\delta_3}{\sqrt{3} d},\ |P_4S_2|=\frac{\delta_4}{\sqrt{3} d},\ |S_1S_2|=\frac{\delta}{\sqrt{3} d}$$ and use them for finding the critical value of the objective function (\[obj\_fun\]). One has $$\delta_1+\delta_2+\delta_3+\delta_4+ \delta$$ $$\stackrel{(\ref{delta1234})}{=} (x_1-x_2-x_3+x_4) (\eta_1+\eta_2)+ (-y_1+y_2+y_3-y_4) (\tau_1+\tau_2) + \delta$$ $$=(x_1-x_3) \eta_1 + (-y_1+y_3) \tau_1 + (-x_2+x_4) \eta_1 + (x_1-x_2-x_3+x_4) \eta_2 + (y_2-y_4) \tau_1 +
(-y_1+y_2+y_3-y_4) \tau_2 + \delta$$ $$\stackrel{(\ref{delta}),(\ref{eta12})}{=} -(-x_2+x_4) \frac{(\tau_1+2\tau_2)}{\sqrt{3}} + (x_1-x_2-x_3+x_4) \frac{(2\tau_1+\tau_2)}{\sqrt{3}} + (y_2-y_4) \tau_1 + (-y_1+y_2+y_3-y_4) \tau_2$$ $$=\frac{\tau_1}{\sqrt{3}} \left(2\, x_1-x_2-2\,x_3+x_4+\sqrt{3}(y_2-y_4) \right) + \frac{\tau_2}{\sqrt{3}} \left(
x_1+x_2-x_3-x_4+\sqrt{3}(-y_1+y_2+y_3-y_4) \right)$$ $$\stackrel{(\ref{tau1}),(\ref{tau2})}{=} \frac{1}{\sqrt{3}} \tau_1^2 + \frac{1}{\sqrt{3}} \tau_2 (\tau_1+\tau_2) = \frac{1}{\sqrt{3}} (\tau_1^2+\tau_1 \tau_2 + \tau_2^2) \, .$$ This yields the claimed critical value (\[Stein-dist4\]) for the objective function (\[obj\_fun\]).
We do not prove here that this value is indeed the minimal one. Instead of this we will discuss the meaning of the positivity restrictions imposed in the theorem on the values $ \delta, \delta_1,\dots,\delta_4 $. If we alter simultaneously their sign to negative then substitution of formulas (\[Steiner pt1\]) and (\[Steiner pt2\]) into (\[stat\_points\]) still keep them valid. What reason underlines then the choice of the sign for deltas mentioned in the theorem? To answer this question let us determine the relative position of the points $ S_1, S_2 $ with respect to the quadrilateral $ P_1P_2P_3P_4 $. The validity of the following equalities: $$\left|
\begin{array}{lll}
1 & 1 & 1 \\
x_1 & x_2 & x_{\ast} \\
y_1 & y_2 & y_{\ast}
\end{array}
\right| = \frac{\sqrt{3}}{2} \cdot \frac{\delta_1 \delta_2 }{\tau_1^2+\tau_1 \tau_2 + \tau_2^2}\, , \quad
\left|
\begin{array}{lll}
1 & 1 & 1 \\
x_2 & x_3 & x_{\ast} \\
y_2 & y_3 & y_{\ast}
\end{array}
\right| = \frac{\sqrt{3}}{2} \cdot \frac{\delta \delta_2 + \delta_2 \delta_3}{\tau_1^2+\tau_1 \tau_2 + \tau_2^2}\, , $$ $$\left|
\begin{array}{lll}
1 & 1 & 1 \\
x_3 & x_4 & x_{\ast} \\
y_3 & y_4 & y_{\ast}
\end{array}
\right| = \frac{\sqrt{3}}{2} \cdot \frac{\delta_3 \delta_4+ \delta_3 \delta+\delta_4 \delta}{\tau_1^2+\tau_1 \tau_2 + \tau_2^2}\, , \
\left|
\begin{array}{lll}
1 & 1 & 1 \\
x_4 & x_1 & x_{\ast} \\
y_4 & y_1 & y_{\ast}
\end{array}
\right|=\frac{\sqrt{3}}{2} \cdot \frac{\delta \delta_1 + \delta_1 \delta_4}{\tau_1^2+\tau_1 \tau_2 + \tau_2^2}$$ can be verified by direct substitution of the formulas for $ x_{\ast}, y_{\ast}, x_{\ast \ast}, y_{\ast \ast} $ and deltas from the statement of the theorem. Due to assumption of the theorem, all the left-hand side determinants are positive. Therefore the point $ S_1 $ lies inside the quadrilateral $ P_1P_2P_3P_4 $. Similar arguments lead to the same claim for the point $ S_2 $.
For the final check, let us evaluate the angles $ P_1S_1P_2 $ and $ P_1S_1S_2 $ $$\frac{\left<\overrightarrow{S_1P_1}, \overrightarrow{S_1P_2}\right>}{|S_1P_1|\cdot |S_1P_2|}
=\frac{(x_{\ast}-x_1)(x_{\ast}-x_2)+(y_{\ast}-y_1)(y_{\ast}-y_2)}{|S_1P_1|\cdot |S_1P_2|}
\stackrel{(\ref{Steiner pt1}),(\ref{Steiner pt10}) \atop (\ref{S1P1}),(\ref{S1P2})}{=}
\frac{3/4(\tau_1 \tau_2+ \eta_1 \eta_2)}{\tau_1^2+\tau_1 \tau_2 + \tau_2^2} \stackrel{(\ref{eta12})}{=}- \frac{1}{2} \, ;$$ $$\frac{\left<\overrightarrow{S_1P_1}, \overrightarrow{S_1S_2}\right>}{|S_1P_1|\cdot |S_1S_2|}
\stackrel{(\ref{Steiner pt1}),(\ref{Steiner pt2}) \atop (\ref{S1P1}),(\ref{S1P2})}{=}\frac{-\frac{\sqrt{3}}{2}\left( \tau_1(x_1-x_3)+\eta_1(y_1-y_3)\right)+ \delta_1+ \delta_3}{\delta}$$ $$\stackrel{(\ref{delta}),(\ref{delta1234})}{=} \frac{-\frac{\sqrt{3}}{2}\left( \tau_1(x_1-x_3)+\eta_1(y_1-y_3)\right)- \delta+(x_1-x_3) \eta_2+(-y_1+y_3)\tau_2}{\delta}\stackrel{(\ref{eta12}),(\ref{delta})}{=}-\frac{1}{2}\, .$$ Therefore both angles in question equals $ 2\pi/3 $. This agrees with the property **(P4)** of Steiner points mentioned in Section 1.
\[Ex2\] Find the coordinates for Steiner points for topology from Example \[Ex1\].
**Solution.** Here $$\tau_1=-9-6\sqrt{3},\ \tau_2=-3-4\sqrt{3},\ \eta_1= 14+5\sqrt{3},\ \eta_2= -16-7\sqrt{3} \, .$$ The conditions of Theorem \[Steiner4\] are fulfilled: the values $$\delta=62+11\sqrt{3},\ \delta_1=-1+13\sqrt{3},\ \delta_2=59+35\sqrt{3},\ \delta_3=63+ 41\sqrt{3},\ \delta_4=3+ 15\sqrt{3}$$ are positive. Further, $$\tau_1^2+\tau_1 \tau_2 + \tau_2^2 = 345+186\sqrt{3}$$ and therefore the length of Steiner tree equals $$d=\sqrt{115+62\sqrt{3}} \approx 14.912651 \, .$$ Formulas (\[Steiner pt1\]) and (\[Steiner pt2\]) yield the coordinates of Steiner points $$\begin{aligned}
x_{\ast}&=& \displaystyle \frac{571+323\sqrt{3}}{2(115+62\sqrt{3})}=\frac{5587}{3386}+\frac{1743}{3386}\sqrt{3} \approx 2.541631, \\
y_{\ast}&=& \displaystyle \frac{3609+2051\sqrt{3}}{6(115+62\sqrt{3})}=\frac{11183}{3386}+\frac{12107}{10158}\sqrt{3} \approx 5.367094\end{aligned}$$ and $$\begin{aligned}
x_{\ast \ast}&=& \displaystyle \frac{3(441+227\sqrt{3})}{2(115+62\sqrt{3})}=\frac{25479}{3386}-\frac{3711}{3386}\sqrt{3} \approx 5.626509, \\
y_{\ast \ast}&=& \displaystyle \frac{1349+747\sqrt{3}}{2(115+62\sqrt{3})}= \frac{16193}{3386}+\frac{2267}{3386}\sqrt{3}\approx 5.941984.\end{aligned}$$
**Remark 1.** Comparing the formulas for Steiner points (\[Steiner pt1\]) and (\[Steiner pt2\]) with their counterparts (\[FTp\]) for the three terminals problem, one may watch the following property: The denominators of all the formulas for Steiner point coordinates contain an explicit expression for the length of the corresponding tree. It looks like every Steiner point “knows” the length of the tree which this point is a part of[^6].
**Remark 2.** The condition $ \delta \ne 0 $ with $ \delta $ defined by (\[delta\]) guarantees that the Steiner points $ S_1 $ and $ S_2 $ do not coincide. The restriction $ \delta > 0 $ can be reformulated in terms of geometry of the quadrilateral $ P_1P_2P_3P_4 $. Indeed, $$\delta \stackrel{(\ref{delta})}{=} -(x_1-x_3) \eta_1 + (y_1-y_3) \tau_1$$ $$\stackrel{(\ref{tau1}),(\ref{eta12})}{=}
-(x_1-x_3)(-\sqrt{3}x_2+\sqrt{3}x_4+y_2-y_4+2\,y_3-2\, y_1)+(y_1-y_3)(2\,x_1-x_2-2\,x_3+x_4+\sqrt{3}y_2-\sqrt{3}y_4)$$ $$=-(x_1-x_3)(-\sqrt{3}x_2+\sqrt{3}x_4+y_2-y_4)+(y_1-y_3)(-x_2+x_4+\sqrt{3}y_2-\sqrt{3}y_4)$$ $$=2(x_3-x_1,y_3-y_1) \left(\begin{array}{rr} \sqrt{3}/2 & 1/2 \\ -1/2 & \sqrt{3}/2 \end{array} \right) \left(\begin{array}{c} x_4-x_2 \\ y_4-y_2 \end{array} \right) \, .$$ This value is positive iff the angle between the diagonal $ \overrightarrow{P_1P_3} $ of the quadrilateral and the other diagonal $ \overrightarrow{P_2P_4} $ turned through by $ \pi/6 $ clockwise is acute. Equivalently, if we denote by $ \psi $ the angle between diagonal vectors $ \overrightarrow{P_1P_3} $ and $ \overrightarrow{P_2P_4} $ then $ \delta $ is positive iff $ \psi < \pi/2 +\pi/6=2\, \pi/3 $. This confirms the known condition for the existence of a full Sreiner tree .
Corollaries {#Scor}
===========
In the present section some statements are presented without proofs: their computer verification has been performed using the formulas from the previous section.
Length of a tree
----------------
We first transform the formula (\[Stein-dist4\]) for the Steiner tree length: $$d=\frac{1}{2}\sqrt{\frac{1}{3}(\tau_1-\tau_2)^2+(\tau_1+\tau_2)^2} =\frac{1}{2}\sqrt{A^2+B^2}
\label{dist_alt}$$ for $$\begin{aligned}
A&=&\frac{1}{\sqrt{3}} (\tau_1-\tau_2)= \sqrt{3}(x_1-x_2-x_3+x_4)+(y_1+y_2-y_3-y_4), \\
B&=&\tau_1+\tau_2=(x_1+x_2-x_3-x_4)+\sqrt{3}(-y_1+y_2+y_3-y_4)\, .\end{aligned}$$ **Remark 3.** This formula can be additionally checked with the aid of geometric algorithm of Steiner tree construction outlined in Example \[Ex1\]. Indeed, for the general case, the coordinates of the counterpart of the point $ Q_1 $ from that construction are given by (\[Q1\]), while those for $ Q_2 $ can be obtained from the latter via substitution $ 1 \rightarrow 3, 2 \rightarrow 4 $ for terminal coordinate subscripts: $$Q_2=\left(\frac{1}{2}x_3+\frac{1}{2}x_4-\frac{\sqrt{3}}{2}y_3+\frac{\sqrt{3}}{2}y_4\,,\ \frac{\sqrt{3}}{2}x_3-\frac{\sqrt{3}}{2}x_4+\frac{1}{2}y_3+\frac{1}{2}y_4 \right) \, .$$ One can easily verify that $ |Q_1 Q_2| $ equals (\[dist\_alt\]).
On expanding the radicand of (\[dist\_alt\]) further one gets $$d^2=(x_1-x_3)^2+(y_1-y_3)^2+(x_2-x_4)^2+(y_2-y_4)^2$$ $$-\left[ (x_1-x_3)(x_2-x_4)+(y_1-y_3)(y_2-y_4) \right]+ \sqrt{3} \left[(x_1-x_3)(y_2-y_4)-(y_1-y_3)(x_2-x_4) \right]$$ $$=r_{13}^2+r_{24}^2-r_{13}r_{24} \cos \psi + \sqrt{3} r_{13}r_{24} \sin \psi=r_{13}^2+r_{24}^2+2\, r_{13}r_{24} \cos \left(\frac{2\pi}{3}-\psi \right)$$ where $ \psi $ denotes the angle between diagonal vectors $ \overrightarrow{P_1P_3} $ and $ \overrightarrow{P_2P_4} $. This result coincides with the one presented in [@Booth]. Transforming it further to $$= r_{13}^2+r_{24}^2-2\, r_{13}r_{24} \cos \left(\psi+ \frac{\pi}{3} \right)$$ we are able to provide it with the following geometric meaning which follows from the law of cosines:
The length of Steiner tree of the topology $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $ equals the length of the third side of the triangle constructed on two other sides coinciding with the diagonals of the quadrilateral $ P_1P_2P_3P_4 $ and with the angle between them chosen to be equal to $ \displaystyle \psi+ \pi/3 $.
For the case of four terminals, there might exist a topology of full Steiner tree alternative to the one given in Theorem \[Steiner4\], namely $ \begin{array}{c} P_4 \\ P_1 \end{array} \tilde S_1 \tilde S_2 \begin{array}{c} P_3 \\ P_2 \end{array} $. To obtain the condition for its existence and the coordinates of Steiner points $ \tilde S_1 $ and $ \tilde S_2 $, one should make the cyclic substitution $$\left(
\begin{array}{cccc}
1 & 2 & 3 & 4 \\
2 & 3 & 4 & 1
\end{array}
\right)
\label{substitut}
\ .$$ for the terminal subscripts in the formulas of that theorem.
\[cor2\] The necessary condition for the existence of Steiner trees of both topologies $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $ and $ \begin{array}{c} P_4 \\ P_1 \end{array} \tilde S_1 \tilde S_2 \begin{array}{c} P_3 \\ P_2 \end{array} $ is that the angles between the diagonals of the quadrilateral $ P_1P_2P_3P_4 $ are less than $ 2\pi/3 $ . If these trees exist with their lengths equal to correspondingly $ d $ and $ \tilde d $ then $$d^2-\tilde d^2= -2\left\{(x_3-x_1)(x_4-x_2)+(y_3-y_1)(y_4-y_2) \right\}=-2 \left<\overrightarrow{P_1P_3}, \overrightarrow{P_2P_4}\right> \, .$$ This means: If the diagonals of the quadrilateral are normal then both topologies give the minimal tree. Otherwise the SMT coincides with the tree with Steiner points lying inside the acute vertical angles formed by the quadrilateral diagonals.
\[Ex3\] For terminals from Example \[Ex1\], find Steiner points for the topology $ \begin{array}{c} P_4 \\ P_1 \end{array} \tilde S_1 \tilde S_2 \begin{array}{c} P_3 \\ P_2 \end{array} $ .
**Solution.** The conditions of Theorem \[Steiner4\] are fulfilled. One has $$\tilde S_1 = \left( \frac{34837}{7237}-\frac{14353}{43422}\sqrt{3} \ , \frac{21203}{14474}+\frac{28342}{21711}\sqrt{3} \right)
\approx (4.241211, 3.725958) \, ,$$ $$\tilde S_2 = \left(\frac{29648}{7237}-\frac{1109}{14474}\sqrt{3} \ , \ \frac{94173}{14474}-\frac{5092}{7237}\sqrt{3} \right)
\approx (3.964015, 5.287674 ) \, .$$ and $$\tilde d= \sqrt{137+62\sqrt{3}} \approx 15.632887 \, .$$ Thus, Steiner tree of the topology $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $ drawn in Fig.6 and with coordinates for Steiner points presented in Example \[Ex2\] is the SMT for the given set of terminals.
For terminals $$P_1=(1,6), \ P_2= (2,1 ), P_3=(6,1), P_4=(8,7) \, ,$$ find Steiner points for the topologies $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $ and $ \begin{array}{c} P_4 \\ P_1 \end{array} \tilde S_1 \tilde S_2 \begin{array}{c} P_3 \\ P_2 \end{array} $.
**Solution.** Here $ \left<\overrightarrow{P_1P_3},\overrightarrow{P_2P_4} \right>=0 $. For the topology $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $ one has $$\begin{aligned}
S_1&=& \displaystyle \left(\frac{5997+ 1687\sqrt{3}}{3063},\ \frac{13014- 3103\sqrt{3}}{3063} \right) \approx (2.911841,\ 2.494106), \\
S_2& =& \displaystyle \left(\frac{7114}{1021}-\frac{3098}{3063}\sqrt{3},\ \frac{2973}{1021}-\frac{838}{3063}\sqrt{3}\right) \approx
(5.215836,\ 2.437983)\end{aligned}$$ while for $ \begin{array}{c} P_4 \\ P_1 \end{array} \tilde S_1 \tilde S_2 \begin{array}{c} P_3 \\ P_2 \end{array} $ one has $$\begin{aligned}
\tilde S_1& =& \displaystyle \left(\frac{3379}{1021}+\frac{923}{3063}\sqrt{3},\ \frac{6378}{1021}-\frac{3167}{3063}\sqrt{3} \right) \approx
(3.831434,\ 4.455956), \\
\tilde S_2&=& \displaystyle \left(\frac{4744}{1021}-\frac{1342}{3063}\sqrt{3},\ \frac{1263}{1021}+ \frac{1618}{3063}\sqrt{3} \right) \approx (3.887557,\ 2.151962) \, .\end{aligned}$$ The common value for tree length for these topologies equals $$\sqrt{122+60\sqrt{3}} \approx 15.030737 \, .$$
**Remark.** If the diagonals of $ P_1P_2P_3P_4 $ are normal then $ |S_1S_2| = |\tilde S_1\tilde S_2| $.
Wandering terminal
------------------
\[th5\] Let the loci of terminals $ P_1,P_2 $ and $ P_4 $ be fixed in the $(x,y)$-plane while the locus of the terminal $ P_3 $ be variable with the only restriction that the coordinates of terminals $ \{ P_j\}_{j=1}^4 $ meet the conditions of Theorem \[Steiner4\]. For the topology $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $, the locus of $ S_1 $ coincides with the arc of the circle $ \mathfrak{C} $ mentioned in Corollary \[cor1\] while the locus of $ S_2 $ coincides with the arc of the circle $ \widehat{\mathfrak{C}} $ with its center at $$\widehat C= \left(\frac{1}{2}x_1+\frac{1}{2}x_4+\frac{1}{2\sqrt{3}}(-y_1+2\,y_2-y_4),\ \frac{1}{2}y_1+\frac{1}{2}y_4+\frac{1}{2\sqrt{3}}(x_1-2\,x_2+x_4) \right)
\label{hatC}$$ and its radius equal to $$\widehat r_c =\sqrt{ \frac{1}{3} \left(\frac{r_{12}^2+r_{14}^2+r_{24}^2}{2}+\sqrt{3} \mathfrak S_{124} \right) } \, .$$ Here $$\mathfrak S_{124}=\left|\begin{array}{ccc} 1 & 1 & 1 \\ x_1 & x_2 & x_4 \\ y_1 & y_2 & y_4 \end{array} \right| \, ,$$ (and therefore $ \mathfrak S_{124} $ equals the doubled area of the triangle $ P_1P_2P_4 $).
The circle $ \widehat{\mathfrak{C}} $ is the circumscribing circle for the equilateral triangle constructed on the side $ Q_1 P_4 $ with point $ Q_1 $ introduced in Section \[Stein3\] (i.e. it denotes the third vertex of the equilateral triangle built on the segment $ P_1 P_2 $ with its coordinates given by (\[Q1\])). The circle $ \widehat{\mathfrak{C}} $ passes through the point $$I=\left(x_1+ \frac{\mathfrak S_{124}}{\sqrt{3}r_{24}^2} (\sqrt{3}(y_4-y_2)+x_2-x_4) ,\ y_1+ \frac{\mathfrak S_{124}}{\sqrt{3}r_{24}^2} (\sqrt{3}(x_2-x_4)+y_2-y_4) \right)\, .$$ of intersection of the circle $ \mathfrak{C} $ with the diagonal $ P_2P_4 $.
For the terminals $$P_1=(5,8),P_2=(1,1),P_5=(10,7)$$ and for $ P_3 $ moving towards $ P_2 $ from the starting position at $ (11,3) $, the loci of Steiner points for the topology $ \begin{array}{c} P_1 \\ P_2 \end{array} S_1 S_2 \begin{array}{c} P_4 \\ P_3 \end{array} $ are displayed in Fig.7.
{width="70.00000%"}
It should be noted that the trajectory of $ P_3 $ does not influence the trajectories of $ S_1 $ and $ S_2 $. The latter move along the corresponding arcs $ \mathfrak{C} $ and $ \widehat{\mathfrak{C}} $ mentioned in Theorem \[th5\] until $ P_3 $ meets the line $ P_1I $.
Conclusions
===========
We presented some computational formulas for the Steiner minimal tree problem. The obtained solution is not complete since we have restricted ourselves with the case of full Steiner trees. It is also of potential interest to find analytical solution for $ n=5 $ terminals — just to satisfy the author’s curiosity whether the observation mentioned in Remark 1 keeps to be fulfilled...
[**Acknowledgments.**]{} The author appreciate the courtesy of Elisabeth Semenova for designing Fig. 7. This research was supported by the St.Petersburg State University research grant **9.38.674.2013**.
[99]{} Booth R.S. Analytic formulas for full Steiner trees. *Discr. Comput. Geom.* 1991. V. 6, No. 1, pp. 69–-82
Brazil M., Graham R.I., Thomas D.A., Zachariasen M. On the history of the Euclidean Steiner tree problem. *Arch. Hist. Exact Sci.* 2014, V. 68, P.327–354
Cockayne E.J. On the efficiency of the algorithm for Steiner minimal trees. *SIAM J. Appl. Math.*, 1970, V. 18, pp. 150–-159
Courant R., Robbins H. *What is Mathematics?* London. Oxford University Press. 1941
Gilbert E.N., Pollak H.O. Steiner minimal trees. *SIAM J. Appl. Math.* 1968. V. 16, pp. 1–29
Pollak H.O. Some remarks on the Steiner problem. *J. Combin. Theory Ser.* A. 1978. V. 24, pp.278–295
Uteshev A.Yu. Analytical solution for the generalized Fermat-Torricelli problem. *Amer. Math. Monthly* 2014. V. 121, No. 4, pp.318–331
[^1]: [email protected]
[^2]: Possibly empty!
[^3]: Or *degree*
[^4]: It is presented here with some changes in notation: in [@Uteshev14] $ d $ denotes length and $ S $ stands for $ \mathfrak S $!
[^5]: Under the assumption $ \delta > 0 $ stated in the theorem.
[^6]: The dependency of the numerators of the fractions (\[FTp\]) from an extra parameter, namely the area of the triangle $ P_1P_2P_3 $, is not essential: one can also extract this factor from the numerators of the formulas and therefore eliminate it from the denominator [@Uteshev14].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $L^1_{\omega}(G)$ be a Beurling algebra on a locally compact abelian group $G$. We look for general conditions on the weight which allows the vanishing of continuous derivations of $L^1_{\omega}(G)$. This leads us to introducing vector-valued Beurling algebras and considering the translation of operators on them. This is then used to connect the augmentation ideal to the behavior of derivation space. We apply these results to give examples of various classes of Beurling algebras which are weakly amenable, 2-weakly amenable or fail to be even 2-weakly amenable.'
address: 'Ebrahim Samei, Department of Pure Mathematics, Faculty of Mathematics, University of Waterloo, Waterloo, ON, Canada, N2L 3G1'
author:
- Ebrahim Samei
title: 'weak amenability and 2-weak amenability of Beurling algebras'
---
Let $A$ be a Banach algebra, let $n\geq 0$ be an integer, and let $A^{(n)}$ be the $n$-th dual module of $A$ when $n>0$, and be $A$ itself when $n=0$. The algebra $A$ is said to be [*weakly amenable*]{} if bounded derivations $D{\colon}A \to A^*$ are inner, and it is said to be [*n-weakly amenable*]{} if bounded derivations $D{\colon}A \to A^{(n)}$ are inner. The algebra $A$ is [*permanently weakly amenable*]{} if it is $n$-weakly amenable for all $n\geq1$.
The concept of weak amenability was first introduced by Bade, Curtis and Dales [@BCD] for commutative Banach algebras, and was extended to the non-commutative case by B. E. Johnson [@J1] and it has been the object of many studies since (see for example, [@FM], [@G3], and [@YZ] and references therein). Dales, Ghahramani and Gr[ø]{}nb[æ]{}k initiated the study of $n$-weakly amenable Banach algebras in [@DGG], where they revealed many important properties of these algebras and presented some examples of them. For instance, they showed that $C^*$-algebras are permanently weakly amenable; the fact that was known for weakly amenable commutative Banach algebras [@BCD Theorem 1.5]. They also showed that group algebras are $2n+1$-weakly amenable for all $n>0$ (for more example see [@FM], [@J3], and [@YZ]).
Let $L^1_{\omega}(G)$ be a Beurling algebra on a locally compact abelian group $G$. One can pose the question of whether $L^1_{\omega}(G)$ is $n$-weakly amenable; in our case it means that each derivation from $L^1_{\omega}(G)$ into $L^1_{\omega}(G)^{(n)}$ is zero. The case of weak amenability has been studied in [@BCD] and [@G1]. One major result states that $l^1_{\omega}({\mathbb{Z}})$ is weakly amenable if and only if $\inf_n \frac{{\omega}(n){\omega}(-n)}{n}=0$ [@G1]. From this, it can be easily deduced that $l^1_{\omega}(G)$ is weakly amenable if $\inf_n \frac{{\omega}(nt){\omega}(-nt)}{n}=0$ for all $t\in G$.
Now it is natural to ask on what condition on the weight ${\omega}$, $L^1_{\omega}(G)$ is 2-weakly amenable. In their recent memoir [@DL], Dales and Lau have addressed this question. They show that if ${\omega}\geq 1$ is almost invariant and satisfies $\inf_n \frac{{\omega}(nt)}{n}=0$ for all $t\in G$, then $L^1_{\omega}(G)$ is 2-weakly amenable [@DL Theorem 13.8], and conjecture that $L^1_{\omega}(G)$ is 2-weakly amenable if one only assumes that $\inf_n \frac{{\omega}(nt)}{n}=0$ ($t\in G$). Ghahramani and Zabandan proved the conjecture with an additional condition which is weaker than being almost invariant [@GZ].
The central goal of this paper is to study systematically, for $G$ abelian, the behavior of the derivation space of $L^1_{\omega}(G)$ into the dual of an arbitrary symmetric Banach $L^1_{\omega}(G)$-module $X$ and to see when it vanishes. From the fundumental work of Gr[ø]{}nb[æ]{}k [@G1], this question can be reduced to studying the kernel of the multiplication map from $L^1_{\omega}(G)\widehat{\otimes} X$ into $X$. In the case of group algebras i.e. when ${\omega}=1$, it is well-known that this can be done by transfering the properties of the augmentation ideal of $L^1(G)$ into the multiplication ideal of $L^1(G)\widehat{\otimes} L^1(G) \cong L^1(G \times G)$ through the isometric isomorphism $$\gamma {\colon}L^1(G)\widehat{\otimes} L^1(G) \to L^1(G)\widehat{\otimes} L^1(G) \ \ , \ \gamma(m)(s,t)=m(s,st) \eqno{(*)},$$ and then deduce it for any module ([@S Theorem 1.8] and [@D Theorem 2.9.65]). However, in general, this idea can not be applied in its present form to the Beurling algebra $L^1_{\omega}(G)$ because the map $\gamma$ may not be well-defined if we replace $L^1(G)$ with $L^1_{\omega}(G)$ in $(*)$. Our approach is to consider a translation map similar to $\gamma$ but on $L^1_{\omega}(G)\widehat{\otimes} X$. This will allow us to look directly at the kernel of the multiplication map on $L^1_{\omega}(G)\widehat{\otimes} X$ instead of relying on $L^1_{\omega}(G)\widehat{\otimes} L^1_{\omega}(G)$. However, in order to do this, we need to consider vector-valued integration for Beurling algebras.
In Section \[S: Vec-BA\], we introduce the concept of a vector-valued Beurling algebra $L^1_{\omega}(G,A)$ and the module $L^1_{\omega}(G,X)$, where $A$ is an arbitrary Banach algebra and $X$ is a Banach left $A$-module. We show that $L^1_{\omega}(G)\widehat{\otimes} X\cong L^1_{\omega}(G,X)$ can be regarded isometrically as a module over $L^1_{\omega}(G,A)$. When ${\omega}=1$, this concept has been thoroughly developed in [@FD Chapter VIII] for cross-sectional algebras to construct examples of cross-products of C$^*$-algebras and C$^*$-algebra bundles.
For the rest of the paper, we restrict ourselves to the case when $G$ is abelian. In Sections \[S:Trans-op\] and \[S:der space-aug ideal\], we use the idea in Section \[S: Vec-BA\] for the case where $A$ is the Beurling measure algebra on a weight ${\sigma}$ and $X$ is symmetric, to define a translation map such as $(*)$ from $L^1_{{\omega}{\tilde{\sigma}}}(G)\widehat{\otimes} X$ into $L^1_{\omega}(G)\widehat{\otimes} X$, where ${\tilde{\sigma}}(t)={\sigma}(-t)$. This, in most of the desirable cases, connects the augmentation ideal of $L^1_{{\omega}{\tilde{\sigma}}}(G)$ to the kernel of the multiplication map on $L^1_{\omega}(G)\widehat{\otimes} X$ (Theorem \[T: Kernel-Agumentation\]). Eventually we demonstrate that if the augmentation ideal of $L^1_{{\omega}{\tilde{\sigma}}}(G)$ is essential, or equivalently, if there is no non-zero, continuous point derivation on the augmentation character, then the derivation space from $L^1_{\omega}(G)$ into $X$ vanishes (Theorem \[T:Beurling alg-der\]).
The reminder of this paper is devoted to investigating the weak amenability and 2-weak amenability of $L^1_{\omega}(G)$ by applying the preceding results.
In Section \[S:weak amen\], we shall show that $L^1_{\omega}(G)$ is weakly amenable if $\inf \{ \Omega (nt)/n \mid n\in {\mathbb{N}}\}=0$ for all $t\in G$, where $\Omega(t):=\omega (t)\omega (-t)$. This follows from the observation that the above assumption implies that there is no non-zero, continuous point derivation on the augmentation character of $L^1_\Omega(G)$. This result extends the result of Gr[ø]{}nb[æ]{}k and provides an alternative proof of it.
For a weight ${\omega}\geq 1$, let ${\omega}_1(s)={\displaystyle}\limsup_{t\to\infty} \frac{{\omega}(t+s)}{{\omega}(t)}$. In [@GZ], it is shown that $L^1_{\omega}(G)$ is 2-weakly amenable if $\inf \{ \omega (nt)/n \mid n\in {\mathbb{N}}\}=0$ and ${\omega}_1$ is bounded. In Section \[S:2-weak amen\], we first show that there a weight ${\sigma}_{\omega}$ on $G$ which is closely related to ${\omega}_1$. We then show that the result in [@GZ] mentioned above is a particular case of the fact that 2-weak amenablility of $L^1_{\omega}(G)$ follows if there is no non-zero, continuous point derivation on the augmentation character of $L^1_{\omega\tilde{{\sigma}}_{\omega}}(G)$. Moreover, when ${\sigma}_{\omega}$ is bounded, there is a precise correspondence between the essentiality of the augmentation ideal of $L^1_{\omega}(G)$ and 2-weak amenability of $L^1_{\omega}(G)$ (Theorem \[T:2-weak amenable\]). This fact allows us to classify various classes of weights for which their corresponding Beurling algebras are 2-weakly amenable or fail to be 2-weakly amenable. These weights, which are defined on compactely generated abelian groups, include polynomial weights, exponential weights, and certain weights satisfying condition (S) (Section \[S:compactly generated groups\]). For instance, for non-compact groups, we show that a Beurling algebra of a polynomial weight of degree $\alpha$ is 2-weakly amenable if and only if $0\leq \alpha <1$, whereas a Beurling algebra of an exponential weight of degree $\alpha$ is never 2-weakly amenable if $0< \alpha <1$. We extend the later result to a much larger class of symmetric weights for which the growth is exponential. However, we give examples of families of non-symmetric weights with sharp exponential growth, for which the Beurling algebras are 2-weakly amenable.
Preliminaries {#P}
=============
Let $A$ be a Banach algebra, and let $X$ be a Banach $A$-bimodule. An operator $D{\colon}A \to X$ is a [*derivation*]{} if for all $a,b\in A$, $D(ab)=aD(b)+D(a)b$. For each $x\in X$, the operator $ad_x\in B(A,X)$ defined by $ad_x(a)=ax-xa$ is a bounded derivation, called an [*inner derivation*]{}. Let $\mathcal{Z}^1(A,X)$ be the linear space of all bounded derivations from $A$ into $X$. When $A$ is commutative, a Banach $A$-bimodule $X$ is [*symmetric*]{} if for all $a\in A$ and $x\in X$, $ax=xa$. In this case, we say simply that $X$ is a Banach $A$-module.
Let $G$ be a locally compact group with a fixed left Haar measure $\lambda$. The measure algebra $M(G)$ is the Banach space of complex-valued, regular Borel measures on $G$. The space $M(G)$ is identified with the (dual) space of all continuous linear functionals on the Banach space $C_0(G)$, with the duality specified by setting $${\langle}\mu \ , \ f {\rangle}=\int_G f(t)d\mu(t) \ \ \ (f\in C_0(G), \mu\in
M(G)).$$ The convolution multiplication $*$ on $M(G)$ defined by setting $${\langle}\mu*\nu \ , \ f {\rangle}=\int_G\int_G f(st)d\mu(s)d\nu(t) \ \ \ (f\in
C_0(G),\ \mu,\nu \in M(G)).$$ We write $\delta_s$ for the point mass at $s\in G$; the element $\delta_e$ is the identity of $M(G)$, and $l^1(G)$ is the closed subalgebra of $M(G)$ generated by the point masses. Then $M(G)$ is a unital Banach algebra and $L^1(G)$, the group algebra on $G$, is a closed ideal in $M(G)$ [@D Theorem 3.3.36]. Moreover, the dual of $L^1(G)$ can be identified with $L^\infty(G)$, the Banach space of Borel measuarable essentially bounded functions on $G$. We let $LUC(G)$ denote the closed subspace of $L^\infty(G)$ consisting of the (equivalence classes of) bounded left uniformly continuous functions on $G$.
Let $G$ be a locally compact group with identity $e$. A weight on $G$ is a continuous function ${\omega}{\colon}G \to (0,\infty )$ such that $${\omega}(st)\leqslant {\omega}(s){\omega}(t) \ \ \ (s,t \in G), \ \ \ {\omega}(e)=1.$$
Let $X$ be a Banach space of measures or of equivalence classes of functions on a locally compact group $G$, and let ${\omega}{\colon}G \to (0,\infty )$ be a continuous function. We define the Banach space $$X({\omega}):=\{f \mid {\omega}f \in X \},$$ where the norm of $X({\omega})$ is defined so that the map $f \mapsto {\omega}f$ from $X({\omega})$ onto $X$ is a linear isometry. In particular, we let $M_{\omega}(G):=M(G)({\omega})$, $L^1_{\omega}(G):=L^1(G)({\omega})$, $l^1_{\omega}(G):=l^1(G)({\omega})$, $L^\infty_{1/{\omega}}(G):=L^\infty(G)(1/{\omega})$, $LUC_{1/{\omega}}(G):=LUC(G)(1/{\omega})$, and $C_{0,1/{\omega}}(G):=C_0(G)(1/{\omega})$. When ${\omega}$ is a weight, with the convolution multiplication of measures, $M_{\omega}(G)$ becomes a Banach algebra, having $L^1_{\omega}(G)$ as a closed two-sided ideal and $l^1_{\omega}(G)$ as a closed subalgebra. Moreover, $M_{\omega}(G)=L_{\omega}^1(G)=l^1_{\omega}(G)$ if and only if $G$ is discrete. Also $L^\infty_{1/{\omega}}(G)$ is the dual of $L^1_{\omega}(G)$, having $LUC_{1/{\omega}}(G)$ and $C_{0,1/{\omega}}(G)$ as Banach $L^1_{\omega}(G)$-submodules. The algebras $L^1_{\omega}(G)$ are the [*Beurling algebras*]{} on $G$. For more details see [@DL Chapter 7].
Vector-valued Beurling algebras {#S: Vec-BA}
===============================
Let $G$ be a locally compact group, let $\omega$ be a weight on $G$, and let ${\omega}\lambda$ be the positive regular Borel measure on $G$ defined by $${\omega}\lambda(E)=\int_E {\omega}(t)d\lambda(t),$$ where $E$ is an arbitrary $\lambda$-measurable set. It is well-known that ${\omega}\lambda$ is well-definded since ${\omega}$ is continuous and positive. Moreover, $E\subseteq G$ is ${\omega}\lambda$-measurable if and only if $E$ is $\lambda$-measurable.
Our references for vector-valued integration theory is [@DF] and [@DU]. Let $(X, ||\cdot||_X)$ be a Banach space, and let $\mathfrak{L}^1_{\omega}(G,X)$ be the set of all $\lambda$-measurable (or equivalently, ${\omega}\lambda$-measurable) vector-valued functions ${\mathfrak{f}}{\colon}G \to X$ such that $\int_G ||{\mathfrak{f}}(t)||{\omega}(t)dt < \infty $ (see [@DF Appendix B.11] or [@DU Definition II.1.1] for the definition of vector-valued measurable functions). The functions in $\mathfrak{L}^1_{\omega}(G,X)$ are called [*Bochner ${\omega}\lambda$-integrable*]{} since for them the Bochner integral exists. It is clear that $\mathfrak{L}^1_{\omega}(G,X)$ is a vector space with the standard addition and scalar multiplication. Let $L^1_{\omega}(G,X)$ be the equivalent classes of elements in $\mathfrak{L}^1_{\omega}(G,X)$ with respect to the semi-norm $\Vert \cdot \Vert=\int_G ||\cdot||_X{\omega}(t)dt$, i.e. $L^1_{\omega}(G,X):=\mathfrak{L}^1_{\omega}(G,X) / \sim$ where ${\mathfrak{f}}\sim {\mathfrak{g}}$ if and only if $\Vert {\mathfrak{f}}-{\mathfrak{g}}\Vert=0$. Then $L^1_{\omega}(G,X)$ is a Banach space with the above norm; it is called the Banach space of Bochner ${\omega}\lambda$-integrable functions from $G$ into $X$ (see [@DF Appendix B.12] or [@DU Section II.2] for the details). It follows from the definition of Bochner integrable functions that the vector space of the equivalent classes of integrable simple functions from $G$ into $X$, i.e. all the elements ${\mathfrak{f}}{\colon}G \to X$ of the form $${\mathfrak{f}}(\cdot) \sim {\displaystyle}\sum_{i=1}^n \chi_{A_i}(\cdot) x_i, \ \ \ A_i \ \text{is} \ \lambda-\text{integrable}, \ \ \ x_i\in X$$ is dense in $L^1_{\omega}(G,X)$ [@DF Appendix B.12] or [@DU Section II.2].
Let $C_{00}(G,X)$ denotes the set of all continuous functions from $G$ into $X$ with compact support. Then, for each $f\in C_{00}(G,X)$, ${\text{im}}f$ is a compact metric space, and so, it is separable. Hence $f$ is $\lambda$-measurable from [@DF Appendix B.11(c)] or [@DU Theorem II.1.2]. Moreover, $C_{00}(G,X)$ approximates $X$-valued simple functions in norm, and so, it is norm-dense in $L^1_{\omega}(G,X)$ (see [@DF Theorem B.11(d)] for the details).
The following proposition shows that, in certain cases, there is a convolution multiplication or action on $L^1_{\omega}(G,X)$.
\[P: Bochner-Vector Alg\] Let $G$ be a locally compact group, let ${\omega}$ be a weight on $G$, let $A$ be a Banach algebra, and let $X$ be a Banach left $A$-module. Then:\
[(i)]{} $L^1_{\omega}(G,A)$ becomes a Banach algebra with the convolution multiplication $*$ specified by $$({{\bf f}}*{{\bf g}})(s)=\int_G {{\bf f}}(t){{\bf g}}(s-t)dt \ \ \ ({{\bf f}},{{\bf g}}\in C_{00}(G,A));$$ [(ii)]{} $L^1_{\omega}(G,X)$ becomes a Banach left $L^1_{\omega}(G,A)$-module with the action specified by $$({{\bf f}}*{{\bf g}})(s)=\int_G {{\bf f}}(t){{\bf g}}(s-t)dt \ \ \ ({{\bf f}}\in C_{00}(G,A), {{\bf g}}\in C_{00}(G,X)).$$
\(i) Let ${{\bf f}},{{\bf g}}\in C_{00}(G,A)$. It is clear that, for every $s\in G$, the map $t\mapsto {{\bf f}}(t){{\bf g}}(s-t)$ belongs to $C_{00}(G,A)$. Hence it is $\lambda$-measurable. Thus, by [@DF Appendix B.13], the Bochner integral $\int_G {{\bf f}}(t){{\bf g}}(\cdot-t)dt$ exists. Moreover, similar to [@RS Section 3.5], it can be shown that the map $s\mapsto ({{\bf f}}*{{\bf g}})(s)$ is continuous and
$$\begin{aligned}
\int_G \Vert ({{\bf f}}*{{\bf g}})(s)\Vert_A {\omega}(s) ds &\leqslant & \int_G \int_G \Vert {{\bf f}}(t){{\bf g}}(s-t)\Vert_A {\omega}(s) dtds
\\ &\leqslant & \int_G \int_G \Vert {{\bf f}}(t) \Vert_A \Vert {{\bf g}}(s-t)\Vert_A {\omega}(t){\omega}(s-t) dsdt
\\ & = & \Vert {{\bf f}}\Vert \Vert {{\bf g}}\Vert.\end{aligned}$$
Therefore ${{\bf f}}*{{\bf g}}\in L^1_{\omega}(G,A)$ and, by [@DF Appendix B.13], $\Vert {{\bf f}}*{{\bf g}}\Vert \leqslant \Vert {{\bf f}}\Vert \Vert {{\bf g}}\Vert$. The final results follows from that of continuity and the fact that $C_{00}(G,A)$ is dense in $L^1_{\omega}(G,A)$.
The proof of (ii) is similar to (i).
Let $(X, ||\cdot||_X)$ be a Banach space, let $\omega$ be a weight on $G$, and let $L^1_{\omega}(G)\widehat{\otimes} X$ be the projective tensor product of $L^1_{\omega}(G)$ and $X$. Let $f\in C_{00}(G)$ and $x\in X$, and consider the function ${{\bf f}}_x {\colon}G \to X$ defined by $${{\bf f}}_x (t)=f(t)x \ \ (t\in G).$$ Clearly $ {{\bf f}}_x\in C_{00}(G,X)$ and $\Vert {{\bf f}}_x\Vert = \Vert f \Vert \Vert x \Vert $. Hence the map $\alpha_X {\colon}C_{00}(G)\times X \to C_{00}(G,X)$ definded by $$\alpha_X(f\otimes x)={{\bf f}}_x \ \ \ (f\in C_{00}(G), x\in X),$$ is well-definded, bilinear, and $\Vert \alpha_X(f\otimes x) \Vert= \Vert f \Vert \Vert x \Vert $. Therefore, by continuity, there is a unique operator, denoted also by $\alpha_X$, from $L^1_{\omega}(G)\widehat{\otimes} X$ into $L^1_{\omega}(G,X)$ such that $$\alpha_X(f\otimes x)(t)=f(t)x \ \ \ (f\in C_{00}(G), t\in G, x\in X).$$ It is shown in [@DF Proposition 3.3, page 29] that $\alpha_X$ is a isometric linear isomorphism.
Now suppose that $A$ is a Banach algebra and that $X$ is a Banach left $A$-module. It is well-known that $L^1_{\omega}(G)\widehat{\otimes} A$ turns into a Banach algebra along with the action specified by $$(f\otimes a)(g\otimes b)=f*g\otimes ab \ \ \ (f,g\in L^1_{\omega}(G), a,b \in A),$$ and $L^1_{\omega}(G)\widehat{\otimes} X$ turns into a Banach left $L^1_{\omega}(G)\widehat{\otimes} A$ -module with the action specified by $$(f\otimes a)(g\otimes x)=f*g\otimes ax \ \ \ (f,g\in L^1_{\omega}(G), a \in A, x\in X).$$
The following theorem shows that, through $\alpha_X$, the preceding actions coincide with their corresponding vector-valued convolution ones.
\[T: Alpha-algebraic\] Let $G$ be a locally compact group, let $\omega$ be a weight on $G$, let $A$ be a Banach algebra, and let $X$ be a Banach $A$-module. Then:\
[(i)]{} for every $u\in L^1_{\omega}(G)\widehat{\otimes} A$ and $v\in L^1_{\omega}(G)\widehat{\otimes} X$, $$\alpha_X (uv)=\alpha_A(u)* \alpha_X(v);$$\
[(ii)]{} $\alpha_A$ is an isometric algebraic isomorphism from $L^1_{\omega}(G)\widehat{\otimes} A$ onto $L^1_{\omega}(G,A)$.
\(i) It suffices to show that, for every $f,g \in
C_{00}(G)$, $a\in A$ and $x\in X$, $$\alpha_X (f*g\otimes ax)=\alpha_A(f\otimes a)* \alpha_X(g\otimes x).$$ Let $s\in G$. Then
$$\begin{aligned}
\alpha_X (f*g\otimes ax)(s) &=& (f*g)(s)ax \\
&=& ax\int_G f(t)g(s-t)dt \\
&=& \int_G [f(t)a][g(s-t)x] dt \\
&=& \int_G \alpha_A(f\otimes a)(t) \alpha_X(g\otimes x)(s-t) dt
\\ & = & [\alpha_A(f\otimes a)* \alpha_X(g\otimes x)](s).\end{aligned}$$
This completes the proof.\
(ii) Follows from (i) by replacing $X$ with $A$.
Let $X$ be a Banach left $L^1_{\omega}(G)$-module, and let ${\omega}\geqslant 1$. Let $\pi^X_{\omega}{\colon}L^1_{\omega}(G)\widehat{\otimes} X \to X$ and $\phi ^X_{\omega}{\colon}L^1_{\omega}(G)\widehat{\otimes} X \to X$ be the normed-decreasing operators specified by $$\pi^X_{\omega}(f\otimes x)=fx, \ \ \ \phi^X_{\omega}(f\otimes x)=[\int_G f(t)dt]x \ \ \ (f\in L^1_{\omega}(G), x\in X).$$ When there is no risk of ambiguity, we write $\pi$ instead of $\pi^X_{\omega}$ and $\phi$ instead of $\phi ^X_{\omega}$.
There is a vector-valued analogous of the above maps. However we need some introduction before defining them.
We recall that, if $A$ is a Banach algebra, then a Banach left $A$-module $X$ is *essential* if it is the closure of $AX=\text{span} \{ax \mid
a\in A, x\in X \}$. Suppose that $A$ has a bounded approximate identity. Then, by Cohen’s factorization theorem [@D Corollary 2.9.26], $X=AX$, and so, $A$ has a bounded left approximate identity for $X$.
Let $X$ be an essential Banach left $L^1_{\omega}(G)$-module. Then the action of $L^1_{\omega}(G)$ on $X$ can be extended to $M_{\omega}(G)$ [@DL Theorem 7.14] so that for every $x\in X$, the mapping $t\mapsto \delta_tx$ from $G$ into $X$ is continuous. The following lemma allows us to construct the vector-valued version of $\pi^X_{\omega}$.
Let $X$ be an essential Banach left $L^1_{\omega}(G)$-module. Then, for every ${{\bf f}}\in L^1_{\omega}(G,X)$, the mapping $t\mapsto \delta_t[{{\bf f}}(t)]$ from $G$ into $X$ is Bochner $\lambda$-integrable.
Since there is $M>0$ such that $||\delta_t[{{\bf f}}(t)]||\leq M{\omega}(t)||{{\bf f}}(t)||$ $(t\in G)$ and ${{\bf f}}\in L^1_{\omega}(G,X)$, it suffices to show that the mapping $t\mapsto \delta_t[{{\bf f}}(t)]$ from $G$ into $X$ is $\lambda$-measuarble. By [@DU Definition II.2.1], there is a sequence of integrable simple functions $\{{{\bf f}}_n\}$ from $G$ into $X$ such that $$\lim_{n\to \infty} \int_G||{{\bf f}}_n(t)-{{\bf f}}(t)||d\lambda=0.$$ Without the loss of generality (by going to subsequences), this implies that $\lim_{n\to \infty} ||{{\bf f}}_n(t)-{{\bf f}}(t)||=0$ $\lambda$-almost everywhere. Therefore $$\lim_{n\to \infty} ||\delta_t[{{\bf f}}_n(t)]-\delta_t[{{\bf f}}(t)]||\leq M{\omega}(t)\lim_{n\to \infty} ||{{\bf f}}_n(t)-{{\bf f}}(t)||=0$$ $\lambda$-almost everywhere. Hence we have the result if we show that for every $x\in X$ and a $\lambda$-measurable set $E$ with $\lambda(E)<\infty$, the mapping $$t\mapsto \delta_t[\chi_E(t)x]=\chi_E(t)\delta_tx$$ from $G$ into $X$ is $\lambda$-measuarble. Since $\lambda(E)$ is finite, there is a sequence of compact sets $\{K_n\}$ such that $K_n\subset E$ for all $n\in {\mathbb{N}}$ and $\lambda(E\setminus \bigcup_{n=1}^\infty K_n)=0$. On the other hand, $X$ is essential so that the mapping $t\mapsto \delta_tx$ is continuous. Hence each $\{\delta_tx \mid t\in K_n\}$ is a compact metric space, and so, it is seperable. Thus $$\{\delta_tx \mid t\in \bigcup_{n=1}^\infty K_n\}$$ is separable. Hence, by [@DU Theorem II.1.2], the mapping $t\mapsto \delta_t[\chi_E(t)x]$ is $\lambda$-measuarble.
Let $X$ be an essential Banach left $L^1_{\omega}(G)$-module, and let $\Pi^X_{\omega}{\colon}L^1_{\omega}(G,X) \to X$ and $\Phi^X_{\omega}{\colon}L^1_{\omega}(G,X) \to X$ be the norm-decreasing operators defined by $$\Pi^X_{\omega}({{\bf f}})=\int_G \delta_t[{{\bf f}}(t)] dt, \ \ \ \ \Phi^X_{\omega}({{\bf f}})=\int_G {{\bf f}}(t) dt$$ for every ${{\bf f}}\in L^1_{\omega}(G,X)$. When there is no risk of ambiguity, we write $\Pi$ instead of $\Pi^X_{\omega}$ and $\Phi$ instead of $\Phi ^X_{\omega}$.
The following proposition establishes the relationship between the above maps.
\[P: pi\_phi\] Let $G$ be a locally compact group, let $\omega, {\sigma}\geq 1$ be weights on $G$, and let $X$ be an essential Banach left $L^1_{\sigma}(G)$-module. Then:\
[(i)]{} $\pi = \Pi \circ\alpha_X$;\
[(ii)]{} $\phi =\Phi \circ \alpha_X$;\
[(iii)]{} for every ${{\bf f}}\in L^1_{\omega}(G,L^1_{\sigma}(G))$ and ${{\bf g}}\in L^1_{\omega}(G,X)$, $$\Pi ^X_{\omega}({{\bf f}}*{{\bf g}})=\Pi ^{L^1_{\sigma}(G)}_{\omega}({{\bf f}})\Pi ^X_{\omega}({{\bf g}}) \ \ \ \ \ \text{and} \ \ \ \ \Phi ^X_{\omega}({{\bf f}}*{{\bf g}})=\Phi ^{L^1_{\sigma}(G)}_{\omega}({{\bf f}})\Phi ^X_{\omega}({{\bf g}}).$$
\(i) Since $X$ is essential, the action of $L^1_{\sigma}(G)$ on $X$ can be extended to $M_{\sigma}(G)$ and $X$ becomes a unital Banach $M_{\sigma}(G)$-module [@DL Theorem 7.14]. Now to prove our result, it suffices to show that, for every $f \in C_{00}(G)$ and $x\in X$, $$\pi(f\otimes x)=\Pi (\alpha_X(f\otimes x)).$$ We have $$\begin{aligned}
\pi(f\otimes x) &=& \int_G f(t)\delta_tx dt \\
&=& \int_G \delta_t[f(t)x]dt \\
&=& \int_G \delta_t [\alpha_X(f\otimes x)(t)]dt \\
& = & \Pi (\alpha_X(f\otimes x)).\end{aligned}$$ (ii) Let $f \in C_{00}(G)$ and $x\in X$. Then $$\phi(f\otimes x)=[\int_G f(t)dt]x=\int_G \alpha_X(f\otimes x)(t)dt=\Phi (\alpha_X(f\otimes x)).$$ The final result follows from the continuity.\
(iii) It is straightforward to check that, for every $u\in L^1_{\omega}(G)\widehat{\otimes} L^1_{\sigma}(G)$ and $v\in L^1_{\omega}(G)\widehat{\otimes} X$, $$\pi ^X_{\omega}(uv)=\pi ^{L^1_{\sigma}(G)}_{\omega}(u)\pi ^X_{\omega}(v) \ \ \ \ \ \text{and} \ \ \ \ \phi ^X_{\omega}(uv)=\phi ^{L^1_{\sigma}(G)}_{\omega}(u)\phi ^X_{\omega}(v).$$ Thus the result follows from parts (i), (ii) and Theorem \[T: Alpha-algebraic\](i).
The following corollary is an immediate consequence of Proposition \[P: pi\_phi\].
\[C: classification of kernel\] Let $G$ be a locally compact group, let $\omega \geqslant 1$ be a weight on $G$, and let $X$ be an essential Banach left $L^1_{\omega}(G)$-module. Then:\
$($i$)$ $\alpha_X(\ker \pi)=\ker \Pi$;\
$($ii$)$ $\alpha_X(\ker \phi)=\ker \Phi.$
Translation of operators {#S:Trans-op}
========================
Throughout the rest of this paper, we let $G$ be a locally compact abelian group and all modules be symmetric. Let ${\omega}$ be a weight on $G$. We shall consider the following auxiliary weight on $G$: $${\tilde{\omega}}(t)={\omega}(-t) \ \ \ (t\in G).$$ In this section, we show that how can we transfer information from $\ker \Phi ^X_{{\omega}{\tilde{\sigma}}}$ into $\ker \Pi ^X_{\omega}$.
\[T: Transform TH\] Let $\omega$ and ${\sigma}$ be weights on $G$, and let $X$ be an essential Banach $L^1_{\sigma}(G)$-module. Then there is an operator $\Lambda_X {\colon}L^1_{{\omega}{\tilde{\sigma}}}(G,X) \to L^1_{{\omega}}(G,X)$ such that:\
$($i$)$ for every ${{\bf f}}\in C_{00}(G,X)$, $\Lambda_X({{\bf f}})(t)=\delta_{-t}[{{\bf f}}(t)]$;\
$($ii$)$ $\Lambda_X$ is norm-decreasing linear map with the dense range;\
$($iii$)$ if ${\sigma}=1$, then $\Lambda_X$ is an isometric isomorphism on $L^1_{{\omega}}(G,X)$;\
$($iv$)$ for every ${{\bf f}}\in L^1_{{\omega}{\tilde{\sigma}}}(G,L^1_{\sigma}(G))$ and $g\in L^1_{{\omega}{\tilde{\sigma}}}(G,X)$, $$\Lambda_X({{\bf f}}*{{\bf g}})=\Lambda_{L^1_{\sigma}(G)}({{\bf f}})*\Lambda_X({{\bf g}}).$$
By [@DL Theorem 7.44], we can assume that ${\sigma}\geq 1$. Let ${{\bf f}}\in C_{00}(G,X)$ and define $\Lambda_X({{\bf f}})(t):=\delta_{-t}[{{\bf f}}(t)]$ for every $t\in G$. Clearly $\Lambda_X({{\bf f}}) \in C_{00}(G,X)$. Let $\Vert \cdot \Vert_{{\omega}{\tilde{\sigma}}}$ and $\Vert \cdot \Vert_{\omega}$ denote the norms on $L^1_{{\omega}{\tilde{\sigma}}}(G,X)$ and $L^1_{{\omega}}(G,X)$, respectively. Then $$\begin{aligned}
\Vert\Lambda_X({{\bf f}}) \Vert_{\omega}&=& \int_G \Vert\Lambda_X({{\bf f}})(t) \Vert {\omega}(t) dt \\
&=& \int_G\Vert \delta_{-t}[{{\bf f}}(t)] \Vert {\omega}(t) dt \\
&\leqslant & \int_G \Vert {{\bf f}}(t) \Vert \Vert \delta_{-t}\Vert {\omega}(t) dt \\
&=& \int_G \Vert {{\bf f}}(t) \Vert {\sigma}(-t) {\omega}(t) dt \\
& = & \Vert {{\bf f}}\Vert_{{\omega}{\tilde{\sigma}}}.\end{aligned}$$ Hence $\Lambda_X$ can be extended to a norm-decreasing linear operator from $L^1_{{\omega}{\tilde{\sigma}}}(G,X)$ into $L^1_{{\omega}}(G,X)$. Moreover, it is easy to see that $\Lambda_X(C_{00}(G,X))=C_{00}(G,X)$. Hence the image of $\Lambda_X$ is dense in $L^1_{{\omega}}(G,X)$ i.e. (i) and (ii) hold.
For (iv), let ${{\bf f}}\in C_{00}(G,L^1_{\sigma}(G))$, ${{\bf g}}\in C_{00}(G,X)$, and $s\in G$. Then $$\begin{aligned}
[\Lambda_{L^1_{\sigma}(G)}({{\bf f}})*\Lambda_X({{\bf g}})](s) &=& \int_G \Lambda_{L^1_{\sigma}(G)}({{\bf f}})(t)\Lambda_X({{\bf g}})(s-t) dt \\
&=& \int_G [\delta_{-t}*{{\bf f}}(t)][ \delta_{t-s}{{\bf g}}(s-t)] dt \\
&=& \int_G [\delta_{-t}*{{\bf f}}(t)*\delta_{t-s}]{{\bf g}}(s-t) dt \\
&=& \int_G [\delta_{-s}*{{\bf f}}(t)]{{\bf g}}(s-t) dt \\
&=& \int_G \delta_{-s}[{{\bf f}}(t){{\bf g}}(s-t)] dt \\
&=& \delta_{-s} \int_G {{\bf f}}(t){{\bf g}}(s-t) dt \\
& = & \delta_{-s} [({{\bf f}}*{{\bf g}})(s)] \\
& = & \Lambda_X({{\bf f}}*{{\bf g}})(s).\end{aligned}$$ The final result follows from continuity.
Finally, (iii) follows since $\Vert\Lambda_X({{\bf f}}) \Vert_{\omega}=\Vert {{\bf f}}\Vert_{\omega}$ if ${\sigma}=1$.
\[T: Kernel-Agumentation\] Let $\omega$ and ${\sigma}$ be weights on $G$ such that ${\omega}\geqslant 1$, ${\omega}\geqslant {\sigma}$, ${\omega}{\tilde{\sigma}}\geqslant 1$, and let $X$ be an essential Banach $L^1_{\sigma}(G)$-module. Then $\Lambda_X(\ker \Phi ^X_{{\omega}{\tilde{\sigma}}})$ is dense in $\ker \Pi^X_{\omega}$. If, in addition, ${\sigma}=1$, then $\Lambda_X(\ker \Phi ^X_{\omega})=\ker \Pi^X_{\omega}$.
For simplicity, we set $\Pi:=\Pi^X_{\omega}$ and $\Phi:=\Phi ^X_{{\omega}{\tilde{\sigma}}}$. We first note that, by [@DL Theorem 7.44], we can assume that ${\sigma}\geq 1$. Thus $X$ becomes a unital Banach $M_{\sigma}(G)$-module [@DL Theorem 7.14] or [@D]. Also, from ${\sigma}\leqslant {\omega}$, it follows easily that $X$ is both a unital Banach $M_{\omega}(G)$-module and an essential Banach $L^1_{\omega}(G)$-module. Now let ${{\bf f}}\in \ker \Phi$. Then $$\begin{aligned}
\Pi [(\Lambda_X)({{\bf f}})] &=& \int_G \delta_t[\Lambda_X({{\bf f}})(t)] dt \\
&=& \int_G \delta_t [\delta_{-t}{{\bf f}}(t)]dt \\
&=& \int_G [\delta_t*\delta_{-t}]{{\bf f}}(t)dt \\
&=& \int_G \delta_e{{\bf f}}(t)dt \\
&=& \int_G {{\bf f}}(t)dt \\
&=& 0.\end{aligned}$$ Hence $\Lambda_X(\ker \Phi) \subseteq \ker \Pi$.
Now suppose that ${{\bf f}}\in \ker \Pi$ and $\epsilon >0$. There is ${{\bf g}}\in C_{00}(G,X)$ such that $\Vert {{\bf f}}- {{\bf g}}\Vert < \epsilon$. Let $U$ be a compact neighborhood of $e$, and let $\{ e_i \}_{i\in I}$ be a bounded approximate identity in $L^1_{\omega}(G)$ for $X$. It is well-known that $\{ e_i \}_{i\in I}$ can be chosen such that $$\{ e_i \}_{i\in I} \subset C_{00}(G) \ \ \ \ \text{and} \ \ \ \ {\text{supp}}\ e_i \subseteq U \ \ \ (i\in I). \eqno{(1)}$$ Let $M>0$ be an upper bound for $\{ e_i \}_{i\in I}$. Define, by induction on $n$, $x_n\in X$ and $e_n\in \{ e_i \}_{i\in I}$ such that $$x_0:=\Pi ({{\bf g}}), \ \ \Vert x_n-e_nx_n \Vert < 2^{-n}\epsilon \ \ \ \ \text{and} \ \ \ \ \ x_{n+1}:=x_n-e_nx_n . \eqno{(2)}$$ We first observe that, for each $n\in {\mathbb{N}}$, $\Vert x_n \Vert < 2^{-n+1}\epsilon$. Also $$\Vert x_0 \Vert = \Vert \Pi ({{\bf g}}) \Vert= \Vert \Pi ({{\bf f}}-{{\bf g}}) \Vert \leqslant \Vert {{\bf f}}-{{\bf g}}\Vert< \epsilon .$$ Hence $${\displaystyle}\sum_{n=0}^\infty \Vert e_n \Vert \Vert x_n \Vert \leqslant M\epsilon +
M {\displaystyle}\sum_{n=1}^\infty 2^{-n+1}\epsilon= 3M\epsilon.$$ Thus, if we put ${{\bf h}}=\alpha_X\left\{{\displaystyle}\sum_{n=0}^\infty e_n \otimes x_n \right \} $, then, from (1), $${{\bf h}}\in C_{00}(G,X), \ \ \ \ \text{and} \ \ \ \ \Vert {{\bf h}}\Vert < 3M\epsilon. \eqno{(3)}$$ Moreover, from Proposition \[P: pi\_phi\](i) and (2), $$\begin{aligned}
\Pi ({{\bf h}})
&=& \Pi [\alpha_X({\displaystyle}\sum_{n=0}^\infty e_n \otimes x_n)] \\
&=& {\displaystyle}\sum_{n=0}^\infty e_nx_n \\
&=& {\displaystyle}\sum_{n=0}^\infty (x_n-x_{n+1}) \\
&=& x_0 \\
&=& \Pi ({{\bf g}}).\end{aligned}$$ Hence ${{\bf g}}-{{\bf h}}\in \ker \Pi \cap C_{00}(G,X) $. Thus if we put ${\bf m}(t)=\delta_t({{\bf g}}-{{\bf h}})(t)$, then it follows that ${\bf m} \in \ker \Phi$ and $\Lambda_X({\bf m})={{\bf g}}-{{\bf h}}$. Therefore $${{\bf g}}-{{\bf h}}\in \Lambda_X(\ker \Phi).$$ Finally, from (3), we have $$\begin{aligned}
\Vert {{\bf f}}- ({{\bf g}}-{{\bf h}}) \Vert
&\leqslant & \Vert {{\bf f}}-{{\bf g}}\Vert + \Vert {{\bf h}}\Vert \\
&\leqslant & \epsilon + 3M\epsilon \\
&=& (1+ 3M)\epsilon .\end{aligned}$$ Hence $\Lambda_X(\ker \Phi)$ is dense in $\ker \Pi$.
When ${\sigma}=1$, by Theorem \[T: Transform TH\](iii), $\Lambda_X$ is an isometry on $L^1_{\omega}(G,X)$. Thus $\Lambda_X(\ker \Phi ^X_{\omega})$ is norm-closed. Therefore it is the same as $\ker \Pi ^X_{\omega}$.
the space of derivations and the augmentation ideal {#S:der space-aug ideal}
===================================================
Let $A$ be a Banach algebra, and let $\varphi$ be a character on the Banach algebra $A$ (i.e. $\varphi$ is a non-zero multiplicative linear functional on $A$). Then ${\mathbb{C}}$ is a Banach $A$-module for the product defined by $$a\cdot z= z \cdot a= \varphi (a)z \ \ \ (a\in A, z\in {\mathbb{C}});$$ this one-dimensional module is denoted by ${\mathbb{C}}_\varphi$. A derivation from $A$ into ${\mathbb{C}}_\varphi$ is called a [*point derivation at $\varphi$*]{}; it is a linear functional $d$ on $A$ such that $$d(ab)=d(a)\varphi(b)+\varphi(a)d(b) \ \ (a,b\in A).$$
It is well-known that if $A$ is weakly amenable, then there is no non-zero continuous point derivation on $A$ [@D Theorem 2.8.63(ii)], i.e. $\mathcal{Z}^1(A, {\mathbb{C}}_\varphi)=\{0\}$ for every non-zero multiplicative linear functional $\varphi$ on $A$. However the converse is not true! For example, if $\mathfrak{S}$ is the group of rotation of $\mathbb{R}^3$, then, by [@J3 Corollary 7.3], the Fourier algebra $A(\mathfrak{S})$ is not weakly amenable but we know that it has no non-zero continuous point derivation [@F Proposition 1]. Other examples include suitable lipschitz algebra and Beurling algebras [@BCD].
Our purpose in this section is to show that for Beurling algebras a weaker version of the converse is true. We start by considering the following well-known character on $M_{\omega}(G)$, and the point derivations on it.
Let $\omega$ be a weight on $G$ such that ${\omega}\geqslant 1$. Then the map $\varphi^{\omega}_0 {\colon}M_{\omega}(G) \to {\mathbb{C}}$ defined by $$\varphi^{\omega}_0(\mu)=\mu(G) ,$$ is [*the augmentation character on*]{} $M_{\omega}(G)$ and its kernel in $L^1_{\omega}(G)$ is the [*augmentation ideal of*]{} $L^1_{\omega}(G)$.
We will show that not having continuous non-zero point derivations on the augmentation character will determine derivation spaces for a large cases of modules. The following lemma indicates the relationship between the augmentation ideal and the kernel of $\phi^X_{\omega}$.
\[L:augmentation-Kernel-module\] Let $\omega \geqslant 1$ and ${\sigma}$ be weights on $G$, let $X$ be a Banach $L^1_{\sigma}(G)$-module, and let $I_0$ be the augmentation ideal of $L^1_{\omega}(G)$. Let $\iota\otimes id_X {\colon}I_0 \widehat{\otimes} X \to L^1_{{\omega}}(G) \widehat{\otimes} X$ be the norm-decreasing linear operator specified by $$\iota\otimes id_X(f\otimes x)=f\otimes x \ \ \ (f\in I_0, x\in X).$$ Then $\iota\otimes id_X$ is a bi-continuous algebraic isomorphism from $I_0 \widehat{\otimes} X$ onto $\ker \phi^X_{\omega}$.
It is clear that $\iota\otimes id_X( I_0 \widehat{\otimes} X) \subseteq \ker \phi ^X_{{\omega}}$. On the other hand, let $u={\displaystyle}\sum_{n=1}^\infty f_n \otimes x_n \in \ker \phi ^X_{{\omega}}$. Fix $g\in C_{00}(G)$ with $\int_G g(t)dt=1$, and consider the continuious projection $P {\colon}L^1_{\omega}(G) \to I_0$ defined by $$P(f)=f-g \int_G f dt \ \ (f\in L^1_{\omega}(G)).$$ Since $\phi ^X_{{\omega}}(u)=0$, we have $$\begin{aligned}
u
&=& {\displaystyle}\sum_{n=1}^\infty P(f_n) \otimes x_n + {\displaystyle}\sum_{n=1}^\infty (g\int_G f_n(t)dt \otimes x_n)\\
&=& {\displaystyle}\sum_{n=1}^\infty P(f_n) \otimes x_n+ {\displaystyle}\sum_{n=1}^\infty (g\otimes [\int_G f_n(t)dt] x_n)\\
&=& {\displaystyle}\sum_{n=1}^\infty P(f_n) \otimes x_n+ g \otimes {\displaystyle}\sum_{n=1}^\infty [\int_G f_n(t)dt] x_n\\
&=& {\displaystyle}\sum_{n=1}^\infty P(f_n) \otimes x_n+ g \otimes \phi ^X_{{\omega}}(u) \\
&=& {\displaystyle}\sum_{n=1}^\infty P(f_n) \otimes x_n.\end{aligned}$$ Hence $u\in \iota\otimes id_X(I_0 \widehat{\otimes} X)$. Thus $\iota\otimes id_X( I_0 \widehat{\otimes} X)=\ker \phi ^X_{{\omega}}$. Moreover, $P\otimes id_X$ is the right inverse of $\iota\otimes id_X$, i.e. $\iota\otimes id_X$ is one-to-one. The bi-continity follows from the open mapping theorem.
\[T: Kernel-module\] Let $\omega$ and ${\sigma}$ be weights on $G$ such that ${\omega}\geqslant 1$, ${\omega}\geqslant {\sigma}$, ${\omega}{\tilde{\sigma}}\geqslant 1$, and let $X$ be an essential Banach $L^1_{\sigma}(G)$-module. Suppose that $\mathcal{Z}^1(L^1_{{\omega}{\tilde{\sigma}}}(G), {\mathbb{C}}_{\varphi^{{\omega}{\tilde{\sigma}}}_0})=\{0\}$. Then:\
[(i)]{} $\ker \Phi^X_{{\omega}{\tilde{\sigma}}}$ is an essential Banach $\ker \Phi^{L^1_{\sigma}(G)}_{{\omega}{\tilde{\sigma}}}$-module;\
[(ii)]{} $\ker \Pi^X_{\omega}$ is an essential Banach $\ker \Pi^{L^1_{\sigma}(G)}_{\omega}$-module.
It follows from Proposition \[P: pi\_phi\](iii) that $\ker \Phi^X_{{\omega}{\tilde{\sigma}}}$ is a Banach $\ker \Phi^{L^1_{\sigma}(G)}_{{\omega}{\tilde{\sigma}}}$-module. Thus it remains to show the essentiality.
Let $I_0$ be the augmentation ideal of $L^1_{{\omega}{\tilde{\sigma}}}(G)$. Since $L^1_{{\omega}{\tilde{\sigma}}}(G)$ has no non-zero continuous point derivations at $\varphi^{{\omega}{\tilde{\sigma}}}_0$, $I_0^2:=I_0I_0$ is dense in $I_0$ [@D Proposition 1.8.8]. Hence, from the essentiality of $X$, $$I_0\widehat{\otimes} X=[(I_0 \widehat{\otimes} L^1_{{\sigma}}(G)) ( I_0 \widehat{\otimes} X )]^-.$$ Therefore, by Lemma \[L:augmentation-Kernel-module\], if we apply $\iota\otimes id_X$ to the both sides of the above equality, we have $$\ker \phi ^X_{{\omega}{\tilde{\sigma}}}=[\ker \phi ^{L^1_{{\sigma}}(G)}_{{\omega}{\tilde{\sigma}}} \ker \phi ^X_{{\omega}{\tilde{\sigma}}}]^-.$$ The final result follows from applying $\alpha_X$ to the both sides of the preceding equality and using Theorem \[T: Alpha-algebraic\](i) and Corollary \[C: classification of kernel\](ii).\
(ii) It follows from part (i), Theorem \[T: Transform TH\](iv), and Theorem \[T: Kernel-Agumentation\].
Let $A$ be a Banach algebra, and $X$ be a Banach $A$-bimodule. There is a Banach $A$-bimodule actions on $A \widehat{\otimes} X$ specified by $$a \cdot(b \otimes x)=ab \otimes x \ \ , \ \ (b \otimes x)\cdot a=b \otimes
xa \ \ (a,b \in A, x\in X).$$
\[C: Characterization-kernel\] Let $\omega$ and ${\sigma}$ be weights on $G$ such that ${\omega}\geqslant 1$, ${\omega}\geqslant {\sigma}$, ${\omega}{\tilde{\sigma}}\geqslant 1$, and let $X$ be an essential Banach $L^1_{\sigma}(G)$-module. Suppose that $\mathcal{Z}^1(L^1_{{\omega}{\tilde{\sigma}}}(G), {\mathbb{C}}_{\varphi^{{\omega}{\tilde{\sigma}}}_0})=\{0\}$. Then $$\ker \pi^X_{\omega}=\overline{{\mathrm{span}}} \{ f\cdot u-u\cdot f \mid f \in L^1_{\omega}(G), u\in \ker \pi^X_{\omega}\}.$$
From Theorem \[T: Kernel-module\](ii), Theoem \[T: Alpha-algebraic\](i), and Proposition \[P: pi\_phi\](i), we have $$\ker \pi ^X_{\omega}=[\ker \pi ^{L^1_{\sigma}(G)}_{\omega}\ker \pi ^X_{\omega}]^-.$$ Hence it suffices to show that $$[\ker \pi ^{L^1_{\sigma}(G)}_{\omega}\ker \pi ^X_{\omega}]^- =\overline{{\mathrm{span}}} \{ f\cdot u-u\cdot f \mid f \in L^1_{\omega}(G),
u\in \ker \pi^X_{\omega}\}. \eqno{(1)}$$ Let $\{ e_i \}_{i\in I}$ be a bounded approximate identity in $L^1_{\omega}(G)$ for $L^1_{\omega}(G)$, $L^1_{\sigma}(G)$, and $X$. For every $f \in L^1_{\omega}(G)$ and $u\in \ker \pi^X_{\omega}$, $$f\cdot u-u\cdot f={\displaystyle}\lim_{i\to \infty} (f\otimes e_i-e_i\otimes f)u \in [\ker \pi ^{L^1_{\sigma}(G)}_{\omega}\ker \pi ^X_{\omega}]^-.$$ Hence “$\supseteq$" follows in (1).
Conversly, let $m\in \ker \pi ^X_{\omega}$ and $v={\displaystyle}\sum_{n=1}^\infty f_n \otimes g_n \in \ker \pi ^{L^1_{\sigma}(G)}_{\omega}$. We have $$\begin{aligned}
(e_i \otimes e_i) v
&=& {\displaystyle}\sum_{n=1}^\infty e_i f_n \otimes e_i g_n \\
&=& {\displaystyle}\sum_{n=1}^\infty [(e_i \otimes g_n) (f_n \otimes e_i-e_i \otimes f_n + e_i \otimes f_n )]\\
&=& {\displaystyle}\sum_{n=1}^\infty [(e_i \otimes g_n) (f_n \otimes e_i-e_i \otimes f_n)] + e_i^2 \otimes {\displaystyle}\sum_{n=1}^\infty f_n g_n\\
&=& {\displaystyle}\sum_{n=1}^\infty [(e_i \otimes g_n) (f_n \otimes e_i-e_i \otimes f_n)] + e_i \otimes \pi ^{L^1_{\sigma}(G)}_{\omega}(v) \\
&=& {\displaystyle}\sum_{n=1}^\infty [(e_i \otimes g_n) (f_n \otimes e_i-e_i \otimes f_n)].\end{aligned}$$ However, it is straightforward to check that, for each $n\in {\mathbb{N}}$, $m\cdot g_n \in \ker \pi ^X_{\omega}$ and $$[(e_i \otimes g_n) (f_n \otimes e_i-e_i \otimes f_n)]m \rightarrow f_n\cdot [m\cdot g_n]-[m\cdot g_n] \cdot f_n$$ as $i \rightarrow \infty$. Hence $$[(e_i \otimes g_n) (f_n \otimes e_i-e_i \otimes f_n)]m \in \overline{{\mathrm{span}}} \{ f\cdot u-u\cdot f \mid f \in L^1_{\omega}(G),
u\in \ker \pi^X_{\omega}\}.$$ The final result follows since $(e_i \otimes e_i) vm \rightarrow vm$ as $i \rightarrow \infty$.
\[T:Beurling alg-der\] Let $\omega$ and ${\sigma}$ be weights on $G$ such that ${\omega}\geqslant 1$, ${\omega}\geqslant {\sigma}$, ${\omega}{\tilde{\sigma}}\geqslant 1$. Suppose that $\mathcal{Z}^1(L^1_{{\omega}{\tilde{\sigma}}}(G), {\mathbb{C}}_{\varphi^{{\omega}{\tilde{\sigma}}}_0})=\{0\}$. Then, for any Banach $L^1_{\sigma}(G)$-module ${\mathfrak{X}}$, every continuous derivation from $L^1_{\omega}(G)$ into ${\mathfrak{X}}$ is zero.
First consider the case ${\mathfrak{X}}=X^*$, where $X$ is an essential Banach $L^1_{\sigma}(G)$-module.
Let $D {\colon}L^1_{\omega}(G) {\longrightarrow}X^*$ be a bounded derivation. Define the bounded operator $\tilde{D} {\colon}L^1_{\omega}(G) \widehat{\otimes} X \to {\mathbb{C}}$ specified by $$\tilde{D}(f \otimes x)={\langle}D(f) \ , \ x {\rangle}\ \ \ \ (f\in L^1_{\omega}(G), x\in X).$$ We first claim that $$\tilde{D}=0 \ \ \ \text{on} \ \ \ \ker \pi^X_{\omega}. \eqno{(1)}$$ A straightforward calculation shows that, for all $f,g,h\in L^1_{\omega}(G)$ and $x\in X$, $${\tilde{D}}[f\cdot (gh\otimes x-g\otimes hx)-(gh\otimes x-g\otimes hx)\cdot f]=0.$$ On the other hand, from Corollary \[C: Characterization-kernel\], $$\ker \pi^X_{\omega}=\overline{{\mathrm{span}}} \{ f\cdot u-u\cdot f \mid f \in L^1_{\omega}(G), u\in \ker \pi^X_{\omega}\}.$$ Hence (1) follows if we show that $$\ker \pi^X_{\omega}=\overline{{\mathrm{span}}} \{ gh\otimes x-g\otimes hx \mid g,h \in L^1_{\omega}(G), x\in X\}. \eqno{(2)}$$
Let $u={\displaystyle}\sum_{n=1}^\infty f_n \otimes x_n \in \ker \pi ^X_{\omega}$. Let $\{ e_i \}_{i\in I}$ be a bounded approximate identity in $L^1_{\omega}(G)$ for both $L^1_{\omega}(G)$ and $X$. Then $$\begin{aligned}
e_i\cdot u
&=& {\displaystyle}\sum_{n=1}^\infty (e_i f_n \otimes x_n-e_i \otimes f_n x_n +e_i \otimes f_n x_n) \\
&=& {\displaystyle}\sum_{n=1}^\infty (e_i f_n \otimes x_n-e_i \otimes f_n x_n) + e_i \otimes {\displaystyle}\sum_{n=1}^\infty f_n x_n\\
&=& {\displaystyle}\sum_{n=1}^\infty (e_i f_n \otimes x_n-e_i \otimes f_n x_n) + e_i \otimes \pi ^X_{\omega}(u) \\
&=& {\displaystyle}\sum_{n=1}^\infty (e_i f_n \otimes x_n-e_i \otimes f_n x_n).\end{aligned}$$ Thus (2) follows since $e_i \cdot u \rightarrow u$ as $i \rightarrow \infty$. Hence, since $D$ is a derivation, for all $g\in L^1_{\omega}(G)$ and $x\in X$, $$\begin{aligned}
{\langle}D(g) \ , \ x {\rangle}&=& {\tilde{D}}(g \otimes x) \\
&=& {\displaystyle}\lim_{i\to\infty} {\tilde{D}}(g \otimes xe_i) \\
&=& {\displaystyle}\lim_{i\to\infty} {\langle}e_iD(g) \ , \ x {\rangle}\\
&=& {\displaystyle}\lim_{i\to\infty} [ {\langle}D(e_ig) \ , \ x {\rangle}- {\langle}D(e_i)g \ , \ x {\rangle}] \\
&=& {\displaystyle}\lim_{i\to\infty} {\tilde{D}}(e_ig \otimes x- e_i \otimes gx)\\
&=& 0,\end{aligned}$$ where the final equality follows from (1). Hence $D=0$.
The general case follows by adapting an argument similar to the one made in the proof of [@D Theorem 2.8.63(iii)].
weak amenability {#S:weak amen}
================
In this section, we present our main results on weak amenibility. We first recall that if $\omega$ is a weight on $G$ such that ${\omega}\geqslant 1$, then the [*strong operator topology*]{} on $M_{\omega}(G)$ is defined as follows: a net $\{\mu_{\alpha}\}$ converges to $\mu$ ($\mu_{\alpha} \overset{s.o.}{{\longrightarrow}} \mu$) if and only if $\mu_{\alpha}*f \to \mu*f$ in norm for every $f\in L^1_{\omega}(G)$. From [@DL Lemma 13.5], both $L^1_{\omega}(G)$ and $l^1_{\omega}(G)$ are s.o. dense in $M_{\omega}(G)$.
The following proposition gives a sufficient condition on ${\omega}$ so that $L^1_{\omega}(G)$ has no non-zero continuous point derivation. This has been indicated in various references for particular cases but we have not seen it in its general form, so it seems worthwhile to provide a complete proof of it.
\[P:Point Der-Beurling alg\] Let $\omega$ be a weight on $G$ such that ${\omega}\geqslant 1$. Then\
$($i$)$ $\mathcal{Z}^1(L^1_{\omega}(G),{\mathbb{C}}_{\varphi^{\omega}_0})=\{0\}$ whenever $\mathcal{Z}^1(l^1_{\omega}(G),{\mathbb{C}}_{\varphi^{\omega}_0})=\{0\}$.\
$($ii$)$ Suppose that, for every $t\in G$, $\inf \{ \omega (nt)/n \mid n\in {\mathbb{N}}\}=0$. Then $\mathcal{Z}^1(L^1_{\omega}(G),{\mathbb{C}}_{\varphi^{\omega}_0})=\{0\}$.
\(i) Let $d {\colon}L^1_{\omega}(G) \to {\mathbb{C}}$ be a continuous point derivations at $\varphi^{\omega}_0$. By [@D Theorem 2.9.53] and [@DL Theorem 7.14], there is a unique extension of $d$ to a continuous derivation $\tilde{d} {\colon}M_{\omega}(G) \to {\mathbb{C}}$; the action of $M_{\omega}(G)$ on ${\mathbb{C}}$ is defined by $$\mu \cdot z=z\cdot \mu=\mu(G)z \ \ \ (\mu\in M_{\omega}(G), z\in {\mathbb{C}}).$$ Moreover, $\tilde{d}$ is continuous with respect to the strong operator topology on $M_{\omega}(G)$. Clearly the restriction of $\tilde{d}$ to $l^1_{\omega}(G)$ belongs to $\mathcal{Z}^1(l^1_{\omega}(G),{\mathbb{C}}_{\varphi^{\omega}_0})$. Therefore $\tilde{d}=0$ on $l^1_{\omega}(G)$. However, $l^1_{\omega}(G)$ is dense in $M_{\omega}(G)$ with respect to the strong operator topology. Hence $\tilde{d}=0$, and so $d=0$.
\(ii) Take $d\in \mathcal{Z}^1(l^1_{\omega}(G),{\mathbb{C}}_{\varphi^{\omega}_0})$ and $t\in G$. For every $n\in {\mathbb{N}}$, $d(\delta_{nt})=n[\delta_{(n-1)t} \cdot d(\delta_t)]=nd(\delta_t)$. Hence $$\Vert d(\delta_t) \Vert \leqslant \Vert d \Vert \Vert \delta_{nt} \Vert / n =
\Vert d \Vert {\omega}(nt) /n.$$ Thus, from the hypothesis, $d(\delta_t)=0$, i.e. $d=0$.
The final results follows from part (i).
We note that for any weight ${\omega}$ on $G$, its [*symmetrization*]{} is the weight defined by $\Omega(t):=\omega (t)\omega (-t)$ ($t\in G$). We can now use Theorem \[T:Beurling alg-der\] to present a class of weakly amenable Beurling algebras. This has been already established, by N. Groenbaek, for Beurling algebras on discrete abelian groups [@G1].
\[T:weak amenable\] Let $\omega$ be a weight on $G$ such that, for every $t\in G$, $\inf \{ \Omega (nt)/n \mid n\in {\mathbb{N}}\}=0$. Then $L^1_{\omega}(G)$ is weakly amenable.
Let ${\mathbb{R}}^{+\bullet}:=(0, \infty)$ be the group of positive real numbers with respect to multiplication. By [@DL Theorem 7.44], there is a continuous character (i.e. a non-zero group homomorphism) $\chi {\colon}G \to {\mathbb{R}}^{+\bullet}$ such that ${\omega}_1:={\omega}/\chi$ is a weight on $G$, ${\omega}_1 \geqslant 1$, and $L^1_{\omega}(G)$ is isometrically isomorphic to $L^1_{{\omega}_1}(G)$. Therefore it suffices to show that $L^1_{{\omega}_1}(G)$ is weakly amenable. Since $\chi$ is a group homomorphism, for every $t\in G$, $$\omega_1(t) \tilde{\omega_1}(t)=\omega (t){\tilde{\omega}}(t)=\Omega(t).$$ On the other hand, by Proposition \[P:Point Der-Beurling alg\], there is no non-zero continuous point derivation on $L^1_\Omega(G)$ at $\varphi^\Omega_0$. Hence, from Theorem \[T:Beurling alg-der\], every continuous derivation from $L^1_{{\omega}_1}(G)$ into any Banach $L^1_{{\omega}_1}(G)$-module is zero, i.e. $L^1_{{\omega}_1}(G)$ is weakly amenable.
We would like to point out that the condition in Proposition \[P:Point Der-Beurling alg\] is not necessary for vanishing of $\mathcal{Z}^1(L^1_{\omega}(G), {\mathbb{C}}_{\varphi^{\omega}_0})$. Indeed, in Theorem \[T:2-weak amenable-sharp weights\], we will present examples of Beurling algebras with sharp growing weights which have no non-zero continuous point derivations.
2-weak amenability {#S:2-weak amen}
==================
Let ${\omega}\geq 1$ be a weight on $G$. Define the function ${\omega}_1$ on $G$ by
$$\begin{aligned}
{\omega}_1(s) &=& {\displaystyle}\limsup_{t\to\infty} \frac{{\omega}(t+s)}{{\omega}(t)} \\
&=& \inf \left\{\sup \left \{\frac{{\omega}(t+s)}{{\omega}(t)} : t\notin K \right \} \mid K \ \text{is a compact subset of G} \right \}.\end{aligned}$$
It is clear that ${\omega}_1$ is a sub-additive function on $G$ such that ${\omega}_1 \leq {\omega}$ and ${\omega}\tilde{{\omega}}_1 \geq 1$. However, we do not know whether ${\omega}_1$ is continuous. Nevertheless we have the following lemma:
\[L:om-sg euivalent\] Let ${\omega}\geq 1$ be a weight on $G$, and let ${\omega}_1$ be as above. Then:\
[(i)]{} ${\omega}_1$ is measurable;\
[(ii)]{} there is a weight on $G$, denoted by ${\sigma}_{\omega}$, and positive real numbers $M$ and $N$ such that $$M{\omega}_1(t) \leq {\sigma}_{\omega}(t) \leq N{\omega}_1(t) \ \ (t\in G).$$ In particular, ${\sigma}_{\omega}$ is bounded if and only if ${\omega}_1$ is bounded.
\(i) Let $r>0$, $G_r:=\{x\in G \mid {\omega}_1(x)<r \}$ and $s\in G_r$. There is a compact subset $K$ of $G$ and $0<r_1<r$ such that $$\sup \{\frac{{\omega}(t+s)}{{\omega}(t)} : t\notin K \}< r_1. \eqno{(1)}$$ Let $U$ a compact neightborhood of identity in $G$ satisfying $$\sup \{{\omega}(s_1) : s_1\in U \}r_1<r. \eqno{(2)}$$ This is possible because ${\omega}$ is continuous and ${\omega}(e)=1$. Now take $s_1\in U$. From (1) and the fact that ${\omega}$ is sub-additive, we have $$\sup \{\frac{{\omega}(t+s+s_1)}{{\omega}(t)} : t\notin K \}\leq \sup \{\frac{{\omega}(t+s)}{{\omega}(t)} : t\notin K \}
{\omega}(s_1) < r_1{\omega}(s_1).$$ Hence it follows from (2) that $$\sup \{\frac{{\omega}(t+s+s_1)}{{\omega}(t)} : t\notin K, s_1\in U \}\leq \sup \{{\omega}(s_1) : s_1\in U \}r_1<r.$$ This implies that ${\omega}_1(s+s_1)<r$ for all $s_1\in U$ i.e. $s+U\subseteq G_r$. Thus $G_r$ is open. Hence ${\omega}_1$ is measurable.
\(ii) It follows from (i) and [@RS Definition 3.7.1 and Theorem 3.7.5].
The importance of ${\sigma}_{\omega}$ is presented in the following lemma in which we show that a certain Banach $L^1_{\omega}(G)$-module can be regarded as an $L^1_{{\sigma}_{\omega}}(G)$-module. This interesting phonomenon helps us to connect the space of derivations into the second dual of $L^1_{\omega}(G)$ to the behavior of the augmentation ideals of $L^1_{{\omega}{\sigma}_{\omega}}(G)$. Indeed, we will see in Theorem \[T:2-weak amenable\] that, for the case when ${\sigma}_{\omega}$ is bounded, this gives us a precise correspondence between continuous point derivations at $\varphi^{{\omega}}_0$ and 2-weak amenability of $L^1_{\omega}(G)$.
We recall from [@DL P. 77] that $L^\infty_{1/{\omega}}(G)$, as the dual of $L^1_{\omega}(G)$, is a Banach $M_{\omega}(G)$-module. In particular, for each $f\in L^\infty_{1/{\omega}}(G)$, we have $$f\cdot \delta_t=\delta_t \cdot f=\delta_{-t}* f \ \ \ (t\in G).$$ Moreover, $LUC_{1/{\omega}}(G)=L^1_{\omega}(G)L^\infty_{1/{\omega}}(G)$.
\[L:om-sg module\] Let ${\omega}\geq 1$ be a weight on $G$, and let $X_{\omega}=LUC_{1/{\omega}}(G)/C_{0,1/{\omega}}(G)$. Then the standard action of $G$ on $X_{\omega}$ extends continuously to an action of $M_{{\sigma}_{\omega}}(G)$ on $X_{\omega}$. In particular, $X_{\omega}$ is a unital Banach $M_{{\sigma}_{\omega}}(G)$-module and an essential Banach $L^1_{{\sigma}_{\omega}}(G)$-module.
For simplicity, put ${\sigma}={{\sigma}_{\omega}}$.
By Lemma \[L:om-sg euivalent\], there is $M>0$ such that ${\omega}_1\leq M^{-1}{\sigma}_{\omega}$. We first show that for every $t\in G$ and $\textbf{x}\in X_{\omega}$, $$\left\| \delta_t\cdot \textbf{x} \right\|\leq M^{-1}\left\| \textbf{x} \right\|{\sigma}(t). \eqno{(1)}$$ Let $\epsilon > 0$. There is a compact set $F_t$ in $G$ such that, for every $s \notin F_t$, ${\omega}(s+t)/{\omega}(s) \leq M^{-1}{\sigma}(t)+\epsilon$. Pick a continuous function $f_t$ on $G$ with compact support such that $$0 \leq f_t \leq 1 \ , \ f_t=1 \ \text{on} \ F_t. \eqno{(2)}$$ Let $\textbf{x}=\tilde{g}$ where $g\in LUC_{1/{\omega}}(G)$. Since $\delta_t\cdot g-(1-f_t)(\delta_t\cdot g)= f_t(\delta_t\cdot g) \in C_{0,1/{\omega}}(G)$, $$\left\|\delta_t\cdot \textbf{x} \right\| = \left\| \widetilde{(\delta_t\cdot g)} \right\| \leq \left\| (1-f_t)(\delta_t\cdot g) \right\|_{1/{\omega}}.$$ On the other hand, from (2),
$$\begin{aligned}
\left\| (1-f_t)(\delta_t\cdot g) \right\|_{1/{\omega}} & \leq & \sup
\{ \frac{| (\delta_t\cdot g)(s)|}{{\omega}(s)} \mid s\notin F_t \} \\
&=& \sup
\{ \frac{| g(s+t)|}{{\omega}(s)} \mid s\notin F_t \} \\
&=& \sup
\{ \frac{| g(s+t)|}{{\omega}(s+t)}\frac{ {\omega}(s+t)}{{\omega}(s)} \mid s\notin F_t \}\\
&\leq& \left\|g \right\|_{1/{\omega}} (M^{-1}{\sigma}(t)+\epsilon).\end{aligned}$$
Therefore $$\left\| \delta_t\cdot \textbf{x} \right\| \leq \left\|g \right\|_{1/{\omega}} (M^{-1}{\sigma}(t)+\epsilon),$$ and so, (1) follows since $ \left\| \textbf{x} \right\|=\inf \{ \left\|g \right\|_{1/{\omega}} \mid \tilde{g}=\textbf{x}
\}$ and $\epsilon$ was arbitrary.
Now suppose that $\mu \in M_{\omega}(G)$. By [@DL Proposition 7.15], the map $$s\mapsto \delta_s \cdot \textbf{x}, \ \ \ G \rightarrow X_{\omega}$$ is continuous. Hence it is $|\mu|$-measurable. Moreover, from (1), $$\int_G \left\| \delta_s \cdot \textbf{x} \right\| d|\mu| \leq M^{-1}\int_G \left\| \textbf{x} \right\| {\sigma}(s) d|\mu|
=M^{-1} \left\| \textbf{x} \right\|\left\| \mu \right\|_{\sigma}.$$ Therefore the Bohner integral $\mu \cdot \textbf{x}= \int_G \delta_s \cdot \textbf{x} d\mu$ is well-defined and $\left\| \mu \cdot \textbf{x} \right\| \leq M^{-1}\left\| \textbf{x} \right\|\left\| \mu \right\|_{\sigma}$ [@DF Appendix B.12].
The essentiality of $X_{\omega}$ follows simply because $X_{\omega}$ is the closure of $C_{00}(G)\cdot X_{\omega}$.
\[T:2-weak amenable-weaker version\] Let $\omega\geqslant 1$ be a weight on $G$. Suppose that $\mathcal{Z}^1(L^1_{{\omega}{\sigma}_{\omega}}(G),{\mathbb{C}}_{\varphi^{{\omega}{\sigma}_{\omega}}_0})=\{0\}$. Then $L^1_{\omega}(G)$ is $2$-weakly amenable.
Let $D {\colon}L^1_{\omega}(G) \to L^1_{\omega}(G)^{**}$ be a bounded derivation, and let $$X_{\omega}=LUC_{1/{\omega}}(G)/C_{0,1/{\omega}}(G).$$ By the first paragragh in the proof of [@DL Theorem 13.1], ${\text{im}}D \subseteq (C_{0,1/{\omega}}(G))^\bot$, where $$C_{0,1/{\omega}}(G)^\bot=\{ M\in L^1_{\omega}(G)^{**} \mid M=0 \ \text{on} \ C_{0,1/{\omega}}(G) \}.$$ Hence $D$ can be regarded as a bounded derivation from $L^1_{\omega}(G)$ into $C_{0,1/{\omega}}(G)^\bot=[L^\infty_{1/{\omega}}(G)/C_{0,1/{\omega}}(G)]^*$.
On the other hand, since $L^1_{\omega}(G)$ has a bounded approximate identity, by a result of Johnson [@D Corollary 2.9.27],
$$\begin{aligned}
\mathcal{Z}^1[L^1_{\omega}(G),(\frac{L^\infty_{1/{\omega}}(G)}{C_{0,1/{\omega}}(G)})^*]
&=& \mathcal{Z}^1[L^1_{\omega}(G),(\frac{L^1_{\omega}(G) L^\infty_{1/{\omega}}(G)}{C_{0,1/{\omega}}(G)})^*] \\
&=& \mathcal{Z}^1[L^1_{\omega}(G),X_{\omega}^*] \\
&=& 0,\end{aligned}$$
where the last equality follows from Lemma \[L:om-sg module\] and Theorem \[T:Beurling alg-der\]. We note that, by Lemma \[L:om-sg euivalent\], there are $M,N>0$ such that $N^{-1}{\sigma}_{\omega}\leq {\omega}$ and ${\omega}\tilde{{\sigma}}_{\omega}\geq M$. Hence one can easily verify that Theorem \[T:Beurling alg-der\] is valid for ${\omega}$ and ${\sigma}_{\omega}$.
\[T:2-weak amenable\] Let $\omega\geqslant 1$ be a weight on $G$. Suppose that ${\sigma}_{\omega}$ is bounded. Then $\mathcal{Z}^1(L^1_{\omega}(G),{\mathbb{C}}_{\varphi^{\omega}_0})=\{0\}$ if and only if $L^1_{\omega}(G)$ is $2$-weakly amenable.
“$\Rightarrow$" Since ${\sigma}_{\omega}$ is bounded, $L^1_{\omega}(G)$ is a dense subalgebra of $L^1_{{\omega}{\tilde{\sigma}}_{\omega}}(G)$. Hence $L^1_{{\omega}{\tilde{\sigma}}_{\omega}}(G)$ has no non-zero continuious point derivation at $\varphi^{{\omega}{\tilde{\sigma}}}_0$, and so, the result follows from Theorem \[T:2-weak amenable-weaker version\].
“$\Leftarrow$" Let $X:=X_{\omega}$ be as in Theorem \[T:2-weak amenable-weaker version\]. By [@GZ Proposition 2.2], $X^*$ is an $L^1_{\omega}(G)$-submodule of $L^1_{\omega}(G)^{**}$. Thus, by hypothesis, $$\mathcal{Z}^1[L^1_{\omega}(G),X^*]=0. \eqno{(1)}$$ On the other hand, by Lemma \[L:om-sg module\] and the fact that ${\sigma}_{\omega}$ is bounded, $X$ is an essential Banach $L^1(G)$-module. Therefore from the argument presented in the proof of Corollarly \[C: Characterization-kernel\] (equation 1) and Theorem \[T:Beurling alg-der\] (equation 2) we have
$$\begin{aligned}
[\ker \pi ^{L^1(G)}_{\omega}\ker \pi ^X_{\omega}]^- &=& \overline{{\mathrm{span}}} \{ f\cdot u-u\cdot f \mid f \in L^1_{\omega}(G), u\in \ker \pi^X_{\omega}\} \\ & \subseteq& \ker \pi^X_{\omega}\\ &=& \overline{{\mathrm{span}}} \{ gh\otimes x-g\otimes hx \mid g,h \in L^1_{\omega}(G), x\in X\}.\end{aligned}$$
We claim that $$\ker \pi ^X_{\omega}=[\ker \pi ^{L^1(G)}_{\omega}\ker \pi ^X_{\omega}]^-. \eqno{(\star)}$$ Let $T\in (L^1_{\omega}(G)\widehat{\otimes} X)^*$ such that $T=0$ on $\ker \pi ^{L^1(G)}_{\omega}\ker \pi ^X_{\omega}$. Hence, for every $f,g,h \in L^1_{\omega}(G)$ and $x\in X$, $$T[f\cdot (gh\otimes x-g\otimes hx)-(gh\otimes x-g\otimes hx)\cdot f]=0.$$ Thus if we let ${\widehat{T}}{\colon}L^1_{\omega}(G) \to X^*$ be the bounded operator defined by $${\langle}{\widehat{T}}(f) \ , \ x {\rangle}=T(f \otimes x) \ \ \ (f \in L^1_{\omega}(G), x\in X),$$ then a simple calculation shows that $${\widehat{T}}(fgh)-f{\widehat{T}}(gh)-{\widehat{T}}(fg)h+f{\widehat{T}}(g)h=0 \ \ \ (f,g,h \in L^1_{\omega}(G)). \eqno{(2)}$$ Define the bounded operator $D {\colon}L^1_{\omega}(G) \to
\mathcal{B}_{L^1_{\omega}(G)}(L^1_{\omega}(G),X^*)$ by $$D(f)(g)={\widehat{T}}(fg)-f{\widehat{T}}(g) \ \ (f,g\in L^1_{\omega}(G)). \eqno{(3)}$$ From (2), it is easy to verify that $D$ is well-defined. Moreover, upon setting $${\langle}f\cdot S \ , \ x {\rangle}={\langle}S\cdot f \ , \ x{\rangle}={\langle}S \ , \ fx {\rangle},$$ the space $\mathcal{B}_{L^1_{\omega}(G)}(L^1_{\omega}(G),X^*)$ becomes a Banach $A$-module and $\mathcal{D}$ becomes a bounded derivation from $L^1_{\omega}(G)$ into $\mathcal{B}_{L^1_{\omega}(G)}(L^1_{\omega}(G),X^*)$. However, since $L^1_{\omega}(G)$ has a bounded approximate identity, $\mathcal{B}_{L^1_{\omega}(G)}(L^1_{\omega}(G),X^*)$ is isometric with $X^*$ as Banach $L^1_{\omega}(G)$-module. Thus, from (1), $D=0$. Therefore, from (3), ${\widehat{T}}(fg)=f{\widehat{T}}(g)$. So $T$ vanishes on $$\overline{{\mathrm{span}}} \{ gh\otimes x-g\otimes hx \mid g,h \in L^1_{\omega}(G), x\in X\}=\ker \pi^X_{\omega}.$$ Thus ($\star$) holds. Hence, by Theorem \[T: Alpha-algebraic\](i) and Corollary \[C: classification of kernel\](i), $$\ker \Pi ^X_{\omega}=[\ker \Pi ^{L^1(G)}_{\omega}\ker \Pi ^X_{\omega}]^-.$$ However, from Theorem \[T: Transform TH\](iii), $\Lambda_X$ is invertible, and so, by Theorem \[T: Kernel-Agumentation\], $$\begin{aligned}
\ker \Phi ^X_{\omega}&=& \Lambda_X^{-1}(\ker \Pi ^X_{\omega}) \\
&=& [\Lambda_{L^1(G)}^{-1}(\ker \Pi ^{L^1(G)}_{\omega})\Lambda_X^{-1}(\ker \pi ^X_{\omega})]^- \\
&=& [\ker \Phi ^{L^1(G)}_{\omega}\ker \Phi ^X_{\omega}]^-.\end{aligned}$$ Therefore, by Corollary \[C: classification of kernel\](ii), $$\ker \phi ^X_{\omega}=[\ker \phi ^{L^1(G)}_{\omega}\ker \phi ^X_{\omega}]^-.$$ It follows from Lemma \[L:augmentation-Kernel-module\] that $$I_0\widehat{\otimes} X=[(I_0 \widehat{\otimes} L^1(G)) ( I_0 \widehat{\otimes} X )]^-.$$ Hence $I_0=\overline{I_0^2}$.
This completes the proof.
\[T:2-weak amenable-tensor\] Let $\{ L^1_{{\omega}_i}(G_i) \}_{i=1}^n$ be a finite set of Beurling algebras, and let ${\omega}:={\omega}_1\times \cdots \times {\omega}_n$ be the function on $G:=G_1\times \cdots \times G_n$ defined by $${\omega}(t_1, \cdots , t_n)=\prod_{i=1}^n {\omega}_i(t_i) \ \ (0\leq i \leq n, t_i \in G_i).$$ It is well-known that ${\omega}$ is a weight on $G$ so that $L^1_{\omega}(G)$ is algebraiclly isomorphic with $\hat{\otimes}_{i=1}^n L^1_{{\omega}_i}(G_i)$ [@G2 Proposition 1.2]. Now suppose that, for each $i$, ${\omega}_i \geq 1$, ${\sigma}_{{\omega}_i}$ is bounded, and $L^1_{{\omega}_i}(G_i)$ is 2-weakly amenable. Then $L^1_{\omega}(G)$ is 2-weakly amenable. Indeed, since ${\sigma}_{\omega}$ is bounded, by the preceding theorem, it suffices to show that $\mathcal{Z}^1(L^1_{{\omega}}(G),{\mathbb{C}}_{\varphi^{{\omega}}_0})=\{0\}.$ However from our assumption and Theorem \[T:2-weak amenable\], $\mathcal{Z}^1(L^1_{{\omega}_i}(G_i),{\mathbb{C}}_{\varphi^{{\omega}}_0})$ vanishes for each $i$. Therefore it can be easily shown that $\mathcal{Z}^1(L^1_{{\omega}}(G),{\mathbb{C}}_{\varphi^{{\omega}}_0})=\{0\}$: this follows from the fact that $\varphi^{{\omega}}_0=\otimes_{i=1}^n \varphi^{{\omega}_i}_0$.
We finish this section with the following corollary, which was obtained in [@GZ] by a different method.
\[C:2-weak amenable-inf\] Let $\omega\geqslant 1$ be a weight on $G$. Suppose that ${\sigma}_{\omega}$ is bounded and $\inf \{ \omega (nt)/n \mid n\in {\mathbb{N}}\}=0$. Then $L^1_{\omega}(G)$ is $2$-weakly amenable.
By Proposition \[P:Point Der-Beurling alg\], there is no non-zero continuous point derivation on $L^1_{\omega}(G)$ at $\varphi^{\omega}_0$. Hence the result follows from the preceding theorem.
weights on compactly generated abelian groups {#S:compactly generated groups}
=============================================
Let $G$ be compactly generated (abelian) group. Then, by the Structure Theorem, $$G\cong {\mathbb{R}}^k \times {\mathbb{Z}}^m \times T \eqno{(\star)}$$ where $k,m \in {\mathbb{N}}\cup \{0\}$ and $T$ is a compact (abelian) group. Therefore we can define a continuous sub-additive function on $G$ by $$\left| t \right|=\left\| t_1 \right\|,$$ where $t=(t_1,t_2)\in G$, $t_1 \in {\mathbb{R}}^k \times {\mathbb{Z}}^m$, and $||\cdot||$ is the Euclidean norm on ${\mathbb{R}}^{k+m}$. We can use this function to construct different weights on $G$.
For $\alpha \geq 0$ and $C>0$, let $${\omega}(t)=C(1+\left| t \right|)^\alpha \ \ (t\in G).$$ It is easy to see that ${\omega}$ is a weight on $G$; it is called [*polynomial of degree $\alpha$*]{}. The following theorem is a generalization of [@DL Theorem 13.2 and 13.9].
\[T:2-weak amenable-poly\] Let $G$ be a non-compact, compactly generated group, and let $\omega$ and ${\sigma}$ be polynomial weights of degree $\alpha$ and $\beta$, respectively. Then:\
$($i$)$ if $\alpha\geq \beta$ and $\alpha+\beta <1$, then, for any Banach $L^1_{\sigma}(G)$-module ${\mathfrak{X}}$, every continuous derivation from $L^1_{\omega}(G)$ into ${\mathfrak{X}}$ is zero:\
$($ii$)$ $L^1_{\omega}(G)$ is weakly amenable if and only if $\alpha< 1/2$;\
$($iii$)$ $L^1_{\omega}(G)$ is $2$-weakly amenable if and only if $\alpha < 1$.
\(i) follows from Theorem \[T:Beurling alg-der\] and Proposition \[P:Point Der-Beurling alg\](ii).
For (ii), if $\alpha <1/2$, then, from part (i), $L^1_{\omega}(G)$ is weakly amenable. Conversely, suppose that $\alpha \geq 1/2$. Since $G$ is not compact, it has a copy of ${\mathbb{R}}$ or ${\mathbb{Z}}$ as a direct sum. Hence there is a continuous algebraic homomorphism from $L^1_{\omega}(G)$ onto either $L^1_{{\omega}_{|{\mathbb{R}}}}({\mathbb{R}})$ or $l^1_{{\omega}_{|{\mathbb{Z}}}}({\mathbb{Z}})$. However, neither of these algebras are weakly amenable [@DL Theorem 7.43] and [@D Corollary 5.6.19]. Hence $L^1_{\omega}(G)$ is not weakly amenable.
Finally, for (iii), it is easy to see that ${\displaystyle}\limsup_{t\to\infty} \frac{{\omega}(t+s)}{{\omega}(t)}=1$. Hence, by Theorem \[T:2-weak amenable\], it suffices to show that there is no non-zero continuous point derivation on $L^1_{{\omega}}(G)$ at $\varphi^{{\omega}}_0$ if and only if $\alpha < 1$. The “if" part follows from Proposition \[P:Point Der-Beurling alg\](ii) and the “only if" part follows from the fact that, for $\alpha \geq 1$, the Fourier transform of the elements of $L^1_{{\omega}_{|{\mathbb{R}}}}({\mathbb{R}})$ and $l^1_{{\omega}_{|{\mathbb{Z}}}}({\mathbb{Z}})$ are continuously differentiable [@DL Theorem 13.2 and 13.9]. This gives us a non-zero continuous point derivation on $L^1_{{\omega}}(G)$.
Another family of weights that are considered on compactly generated groups are the exponential weights. A weight ${\omega}$ is said to be [*exponential of degree $\alpha$*]{}, $0\leq \alpha \leq 1$, if there exists $C>0$ such that $${\omega}(t)=e^{C|t|^\alpha}, \ \ (t\in G).$$ By our method, we can investigate the question of 2-weak amenability for these families of Beurling algebras.
\[T:2-weak amenable-exp\] Let $G$ be a non-compact, compactly generated group, and let ${\omega}$ be an exponential weight of degree $\alpha$. Then $L^1_{{\omega}}(G)$ is not $2$-weakly amenable if $0< \alpha < 1$.
It is easy to see that, for $0< \alpha < 1$, ${\displaystyle}\limsup_{t\to\infty} \frac{{\omega}(t+s)}{{\omega}(t)}$ is bounded by 1. Also a similar argument to that presented in Theorem \[T:2-weak amenable-poly\](iii) gives us a non-zero continuous point derivation on $L^1_{\omega}(G)$. Hence the result follows from Theorem \[T:2-weak amenable\].
\(i) We note that the result of the preceeding theorem holds, with the same argument, for weights of the form ${\omega}(t)=e^{\rho(|t|)}$, where $\rho$ is a positive increasing sub-additive function which belongs to the Lipschitz algebra on ${\mathbb{R}}^+=(0, \infty)$ with the degree $0< \alpha < 1$.\
(ii) We would like to point out that the result of Theorem \[T:2-weak amenable-exp\] is not true in general when $\alpha=1$. Indeed, it is demonstrated in [@DL Theorem 13.3] that for ${\omega}(n)=e^{|n|}$, $l^1_{\omega}({\mathbb{Z}})$ is 2-weakly amenable.
Theorem \[T:2-weak amenable-exp\] can be generalized to a larger class of weights. Let $q {\colon}{\mathbb{R}}^+ \to {\mathbb{R}}^+$ be a decreasing continuous function such that $${\displaystyle}\lim_{r\to+\infty} q(r)=0 \ \ \text{and} \ \ {\displaystyle}\lim_{r\to+\infty} rq(r)=\infty.$$ Then the function ${\omega}{\colon}G \to [1, \infty)$ given by $${\omega}(t)= e^{|t|q(|t|)}, \ \ (t\in G)$$ is a weight on $G$. All weights constructed as above belong to a family of weights that satisfy a so-called [*condition (S)*]{}. This condition is defined in order to get the symmetry of the certain weighted group algebras on non-abelian groups. We refer the reader to [@DLC] and [@FGLLC] for more details.
\[T:2-weak amenable-cond S\] Let $G$ be a non-compact, compactly generated group, and let $q$ and $\omega$ be as above. Suppose that $rq(r) \geq \ln (1+r)$ for sufficiently large $r$. Then $L^1_{\omega}(G)$ is not $2$-weakly amenable.
Let $t,s \in G$ with $|t|>|s|$, and put $r=|t|-|s|$. Then $|t-s|\geq r$, and so, from the fact that $q$ is decreasing, $$\begin{aligned}
|t|q(|t|)-|t-s|q(|t-s|) &\leq & [r+|s|]q(r+|s|)-rq(r) \\
&=& r[q(r+|s|)-q(r)]+|s|q(r+|s|) \\
&\leq & |s|q(r+|s|) \\
&=& |s|q(|t|).\end{aligned}$$ Hence $${\displaystyle}\limsup_{t\to\infty} \frac{{\omega}(t+s)}{{\omega}(t)}={\displaystyle}\limsup_{t\to\infty} \frac{{\omega}(t)}{{\omega}(t-s)} \leq {\displaystyle}\limsup_{t\to\infty} e^{|s|q(|t|)}=1,$$ since ${\displaystyle}\lim_{t\to\infty} q(|t|)=0$. Therefore ${\sigma}_{\omega}$ is bounded. On the other hand, by hypothesis, ${\omega}(x) \geq 1+|x|$ outside a compact set. Thus the result follows in a similar way to Theorem \[T:2-weak amenable-exp\].
Some examples of such weights are presented in [@DLC Example 1.7]. They are, for instance, given by\
(i) ${\omega}(t)=e^{C|t|^\alpha}=e^{|t|\frac{C}{|t|^{1-\alpha}}}, C>0, 0<\alpha<1,$
\(ii) ${\omega}(t)=e^{|t|{\displaystyle}\Sigma_{n=1}^\infty \frac{c_n}{1+|t|^{\alpha_n}}}, 0<\alpha_n<1, \{ \alpha_n \} \ \text{decreasing to 0}, \ {\displaystyle}\Sigma_{n=1}^\infty c_n < \infty,$
\(iii) ${\omega}(t)=e^{C\frac{|t|}{\ln(e+|t|)}}$,
\(iv) ${\omega}(t)=e^{C\frac{|t|}{(\ln(e+|t|))^k}}, k>0$\
In Theorem \[T:2-weak amenable-poly\], we gave examples of 2-weakly amenable Beurling algebras over (symmetric) polynomial weights. Now we will present a family of non-symmetric weights on ${\mathbb{R}}$ and ${\mathbb{Z}}$ for which the Beurling algebras are 2-weakly amenable and, at one side, they have a much faster growth. Let $0\leq \alpha < 1/2$ and define the functions ${\omega}_{\mathbb{R}}$ and ${\omega}_{\mathbb{Z}}$ on ${\mathbb{R}}$ and ${\mathbb{Z}}$, respectively, as follows: $${\omega}_{\mathbb{R}}(t)=1 \ \text{if} \ t\geq 0 \ \text{and} \ {\omega}_{\mathbb{R}}(t)=e^{|t|^\alpha} \ \text{if} \ t<0 ;$$ $${\omega}_{\mathbb{Z}}(n)=1 \ \text{if} \ n\geq 0 \ \text{and} \ {\omega}_{\mathbb{Z}}(n)=e^{|n|^\alpha} \ \text{if} \ n<0 .$$
It is straightforward to verify that ${\omega}_{\mathbb{R}}$ and ${\omega}_{\mathbb{Z}}$ are weights.
\[P:2-weak amenable-sharp weights\] Let $\alpha$, ${\omega}_{\mathbb{R}}$ and ${\omega}_{\mathbb{Z}}$ be as above. Then:\
$($i$)$ $l^1_{{\omega}_{\mathbb{Z}}}({\mathbb{Z}})$ is $2$-weakly amenable;\
$($ii$)$ $L^1_{{\omega}_{\mathbb{R}}}({\mathbb{R}})$ is $2$-weakly amenable.
\(i) It is easy to see that, ${\displaystyle}\limsup_{n\to\infty} \frac{{\omega}_{\mathbb{Z}}(n+m)}{{\omega}_{\mathbb{Z}}(n)}=1$. On the other hand, $l^1_{{\omega}_{\mathbb{Z}}}({\mathbb{Z}})$ is a commutative regular semisimple Banach algebra on $\mathbb{T}=\{ z\in {\mathbb{C}}\mid |z|=1 \}$. It is shown in [@Z] that, since $${\displaystyle}\limsup_{n\to+\infty} \ \ln{\omega}(-n)/\sqrt{n}=0,$$ singeltons in $\mathbb{T}$ are sets of spectral synthesis for $l^1_{{\omega}_{\mathbb{Z}}}({\mathbb{Z}})$. Thus there are no non-zero continuous point derivations on $l^1_{{\omega}_{\mathbb{Z}}}({\mathbb{Z}})$. Hence the result follows from Theorem \[T:2-weak amenable\].\
(ii) Since ${\sigma}_{{\omega}_{\mathbb{R}}}$ is bounded, by Theorem \[T:2-weak amenable\] and Proposition \[P:Point Der-Beurling alg\](i), it suffices to show that $$\mathcal{Z}^1(l^1_{{\omega}_{\mathbb{R}}}({\mathbb{R}}),{\mathbb{C}}_{\varphi^{{\omega}_{\mathbb{R}}}_0})=\{0\}.$$ Let $d\in \mathcal{Z}^1(l^1_{{\omega}_{\mathbb{R}}}({\mathbb{R}}),{\mathbb{C}}_{\varphi^{{\omega}_{\mathbb{R}}}_0})$. For every $r\in {\mathbb{R}}^+$, let ${\langle}r {\rangle}$ be the discrete additive subgroup of ${\mathbb{R}}$ generated by $r$. Clearly the closed subalgebra $A_r$ of $l^1_{{\omega}_{\mathbb{R}}}({\mathbb{R}})$ generated by the restriction to ${\langle}r {\rangle}$ is algebraically isomorphic to $l^1_{{\omega}_{\mathbb{Z}}}({\mathbb{Z}})$. Thus, from (i), $d=0$ on $A_r$, and so, $d(\delta_r)=d(\delta_{-r})=0$. Hence $d=0$.
The preceding theorem and Remarks \[T:2-weak amenable-tensor\] can be routinely employed to construct fast growing weights on compactly generated groups for which the Beurling algebras are 2-weakly amenable. For each $1\leq i \leq k$ and $1\leq j \leq m$, let $0\leq \alpha_i < 1/2$ and $0\leq \beta_j < 1/2$, and let ${\omega}_{\alpha_i}$ and ${\omega}_{\beta_j}$ be the weights on ${\mathbb{R}}$ and ${\mathbb{Z}}$ associated, as in Proposition \[P:2-weak amenable-sharp weights\], with $\alpha_i$ and $\beta_j$, respectively. Put $${\omega}=\prod_{i=1}^k {\omega}_{\alpha_i} \times \prod_{j=1}^m {\omega}_{\beta_j}.$$ By the identification $(\star)$, ${\omega}$ defines a weight on $G$.
\[T:2-weak amenable-sharp weights\] Let $G$ be a compactly generated group, and let ${\omega}$ be as above. Then $L^1_{{\omega}}(G)$ is $2$-weakly amenable.
[99]{} F. F. Bonsall and J. Duncan, *Complete Normed Algebras*, New York, Springer-Verlag 1973. W. G. Bade, P. C. Curtis, H. G. Dales, [*Amenability and weak amenability for Beurling and Lipschitz algebras*]{}, Proc. London Math. Soc. (3) [**55**]{} (1987), no. 2, 359-377. H. G. Dales, [*Banach algebras and automatic continuity*]{}, New York, Oxford University Press, 2000. H. G. Dales, F. Ghahramani, N. Gr[ø]{}nb[æ]{}k, [*Derivations into iterated duals of Banach algebras*]{}, Studia Math, [**128**]{} (1998), no.1, 19-54. H. G. Dales and A. T.-M. Lau, [*The Second Duals of Beurling Algebras*]{}, Mem. Amer. Math. Soc, [**177**]{} n. 836 (2005). A. Defant and K. Floret, [*Tensor Norms and Operator Ideals*]{}, North-Holland, 1993. J. Diestel and J.J. Uhl, Jr, [*Vector Measures*]{}, American Mathematical Society, 1977. J. Dziubanski, J. Ludwig and C. Molitor-Braun, [*Functional calculus in Weighted Group Algebras*]{}, Rev. Mat. Complut, [**17**]{}, n. 2 (2004) 321-357. G. Fendler, K. Gröchenig, M. Leinert, J. Ludwig and C. Molitor-Braun, [*Weighted Group Algebras on Group of polynomial Growth*]{}, Math. Z. [**245**]{}, n.4 (2003) 791-821. J. M. G. Fell and R. S. Doran, *Representaion of $^*$-algebras, Locally Compact Groups, and Banach $^*$-Algebraic Bundles*, Volume 2, Academic Press, 1988. B. Forrest, [*Amenability and derivations of the Fourier algebra*]{}, Proc. Amer. Math. Soc. [**104**]{} (1988), no. 2, 437-442. B. E. Forrest and L. W. Marcoux, [*Weak amenability of triangular Banach algebras*]{}, Trans. Amer. Math. Soc. [**354**]{} (2002), 1435-1452. F. Ghahramani and G. Zabandan, [*2-weak amenability of a Beurling algebra and amenability of its second dual*]{}, Int. J. Pure Appl. Math. [**16**]{} (2004), no.1, 75-86. N. Gr[ø]{}nb[æ]{}k, [*A characterization of weakly amenable Banach algebras*]{}, Studia Math. [**94**]{} (1989), no. 2, 149-162. N. Gr[ø]{}nb[æ]{}k, [*Amenability of Weighted convolution Algebras on Locally Compact Groups*]{}, Trans. Amer. Math. Soc. [**319**]{} (1990), no. 2, 765-775. N. Gr[ø]{}nb[æ]{}k, [*Bounded Hochschild cohomology of Banach algebras with a matrix-like structure*]{}, Trans. Amer. Math. Soc. [**358**]{} (2006), no.6, 2651-2662. B. E. Johnson, [*Derivations from $L\sp 1(G)$ into $L\sp 1(G)$ and $L\sp \infty(G)$*]{}, Harmonic analysis (Luxembourg, 1987), 191-198 B. E. Johnson, [*Non-amenability of the Fourier algebra of a compact group*]{}, J. London Math. Soc. (2) [**50**]{} (1994) 361-374. B. E. Johnson, [*Permanent weak amenability of group algebras of free groups*]{}, Bull. London Math. Soc. [**31**]{} (1999), no. 5, 569-573. H. Reiter and J. D. Stegeman,*Classical Harmonic Analysis and Locally compact Groups*, Oxford Universsity Press, 2000. R. Stokke, [*Approximate diagonals and F[ø]{}lner conditions for amenable group and semigroup algebras*]{}, Studia Math. [**164**]{} (2004), no. 2, 139-159. Y. Zhang. [*Weak amenability of module extensions of Banach algebras*]{}, Trans. Amer. Math. Soc. [**354**]{} (2002) 4131-4151. M. Zarrabi, [*Synthèse spectrale dans certaines algèbres de Beurling sur le cercle unité.*]{} (French) \[Spectral synthesis in certain Beurling algebras on the unit circle\] Bull. Soc. Math. France, [**118**]{} (1990), no.2, 241-249.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider numerical functions weakly computable by grammar-controlled vector addition systems (GVASes, a variant of pushdown Petri nets). GVASes can weakly compute all fast growing functions $F_\alpha$ for $\alpha<\omega^\omega$, hence they are computationally more powerful than standard vector addition systems. On the other hand they cannot weakly compute the inverses $F_\alpha^{-1}$ or indeed any sublinear function. The proof relies on a pumping lemma for runs of GVASes that is of independent interest.'
address:
- 'LaBRI, Univ. Bordeaux & CNRS, France'
- 'Chennai Mathematical Institute, India'
- 'LSV, ENS Paris-Saclay & CNRS, France'
- 'LaBRI, Univ. Bordeaux & CNRS, France'
author:
- Jérôme Leroux
- M Praveen
- Philippe Schnoebelen
- Grégoire Sutre
bibliography:
- 'wcca.bib'
title: On Functions Weakly Computable by Pushdown Petri Nets and Related Systems
---
[^1]
Introduction {#sec-intro}
============
Pushdown Petri nets are Petri nets extended with a pushdown stack. They have been used to model asynchronous programs [@SenV06] and, more generally, recursive programs with integer variables [@AtigG11]. They sometimes appear under a different but essentially equivalent guise: stack/pushdown/context-free vector addition systems [@lazic2013b; @leroux2014; @leroux2015d], partially blind multi-counter machines [@Greibach78a] with a pushdown stack, etc. It is not yet known whether reachability is decidable for pushdown Petri nets and this is one of the major open problems in computer science. However, a series of recent results improved our understanding of the computational power of these models: coverability, reachability and boundedness are <span style="font-variant:small-caps;">Tower</span>-hard [@lazic2013b; @lazic2017], and boundedness is solvable in hyper-Ackermannian time [@leroux2014].
With the present article, we contribute to this line of work. We recall *Grammar-Controlled Vector Addition Systems* [@leroux2015d], or GVAS, a variant model, close to Pushdown Petri nets, where the pushdown stack is replaced by a context-free restriction on the firing of rules. The runs are now naturally organized in a derivation tree, and the stack is not actually present in the configurations: this leads to a simplified mathematical treatment, where the usual monotonicity properties of VASes can be put to use.
As a step towards understanding the expressive power of these GVASes, we consider the number-theoretical functions that are *weakly computable* in this model. Restricting to weakly computing a numerical function is a natural idea when dealing with models like VASes and GVASes that lack zero-tests, or, more precisely, that cannot initiate a given action on the condition that a counter’s value is zero, only on the condition that it is not zero.
This notion has been used since the early days of Petri nets and has proved very useful in hardness or impossibility proofs: For Petri nets and VASSes, the undecidability of equivalence problems, and the Ackermann-hardness of the same problems for bounded systems, have been proved using the fact that multivariate polynomials with positive integer coefficients —aka positive Diophantine polynomials— and, respectively, the fast-growing functions $(F_i)_{i\in{{\mathbb{N}}}}$ in the Grzegorczyk hierarchy, are all weakly computable [@hack76b; @Mayr-Meyer81; @jancar2001]. More recently, the nonelementary complexity lower bound for VASS reachability is obtained thanks to a uniform (polynomial size) family of systems computing (exactly) $n$-EXP(2) from $n$ [@leroux-stoc19].
The above results rely on showing how *some* useful functions are weakly computable by Petri nets and VASSes. But not much is known about exactly which functions are weakly computable or not. It is known that all such functions are monotonic. They are all primitive-recursive. The class of weakly computable functions is closed under composition.
In this article, we show that functions weakly computable by GVASes go beyond those weakly computable by VASSes, in particular we show how to weakly compute the Fast Growing $(F_\alpha)$ for all $\alpha<\omega^\omega$.
A folklore conjecture states that the inverses of the fast-growing functions are not weakly computable by Petri nets. It is stated as fact in [@phs-IPL2002 p.252] but no reference is given. In this article, we settle the issue by proving that any unbounded function weakly computable by Petri nets and more generally by GVASes is in $\Omega(x)$, i.e., it eventually dominates $c\cdot x$ for some constant $c>0$. Thus any function that is sublinear, like $x\mapsto \lfloor \sqrt{x} \rfloor$, or $x\mapsto
\lfloor \log x \rfloor$, is not weakly computable by GVASes. The proof technique is interesting in its own right: it relies on a pumping lemma on runs of GVASes that could have wider applications. This pumping lemma follows from a well-quasi-ordering on the set of runs that further directs it.
Beyond Petri nets and VASSes. {#beyond-petri-nets-and-vasses. .unnumbered}
-----------------------------
Petri nets and VASSes are a classic example of well-structured systems [@abdulla2000c; @finkel2001]. In recent years, weakly computing numerical functions has proved to be a fundamental tool for understanding the expressive power and the complexity of some families of well-structured systems that are more powerful than Petri nets and VASSes [@phs-mfcs2010; @HSS-lics2012; @HSS-lmcs]. For such systems, the hardness proofs rely on weakly computing fast-growing functions $(F_\alpha)_{\alpha\in\textit{Ord}}$ that extend Grzegorczyk’s hierarchy. These hardness proofs also crucially rely on weakly computing the inverses of the $F_\alpha$’s.
There are several extensions of Petri nets for which reachability (or coverability or boundedness) remains decidable: nets with nested zero-tests [@Reinhardt08], recursive VASSes [@bouajjani2012] and Branching VASSes [@demri2013], VASSes with pointers to counters [@demri2013c], unordered data Petri nets [@datanetsfull08], etc., and of course pushdown VASes and GVASes. For the latter, while coverability and reachability are still open in general, partial decidability results have been obtained by looking at sub-classes, namely GVASes with finite-index grammars [@AtigG11] and GVASes of dimension one [@leroux2015d]. In many cases, it is not known how these extensions compare in expressive power and in complexity. We believe that weakly computable functions can be a useful tool when addressing these questions.
Related models. {#related-models. .unnumbered}
---------------
The GVAS model can simulate counter machines extended with nested zero-tests (from [@Reinhardt08]), and the vector addition systems extended with a pushdown stack (from [@leroux2014]). The first simulation was shown in [@AtigG11] and holds even for GVASes with finite-index grammars. The second one comes from the classical transformation of a pushdown automaton into a context-free grammar that recognizes the same language. There exists still other models that extend vector addition systems with stack-related mechanisms, e.g., Mayr’s Process Rewrite Systems [@mayr2000] or Haddad and Poitrenaud’s Recursive Petri Nets [@haddad2007]. Pending some further, more formal, comparison, it seems that these models are less expressive than Pushdown VASes since they only allow limited interactions between stack and counters.
Outline of the paper. {#outline-of-the-paper. .unnumbered}
---------------------
Section \[sec-gvas\] introduces GVASes and fixes some notation. In Section \[sec-pumping\], we introduce flow trees, a tree-shaped version of runs of GVASes for which we develop our two main tools: a well-quasi-ordering between flow trees and an Amalgamation Theorem. The following two sections explore applications of the Amalgamation Theorem in understanding the computing power of GVASes: via GVAS-definable sets in Section \[sec-gvas-sets\], via weakly computable function in Section \[sec-weakcomp\]. Finally, we show in Section \[sec-hypack\] that GVASes can weakly compute all Fast-Growing functions $F_\alpha$ for $\alpha<\omega^\omega$.
Grammar-Controlled Vector Addition Systems {#sec-gvas}
==========================================
This section recalls the model of grammar-controlled vector addition systems, originally from [@leroux2015d]. In a nutshell, these are intersections of classical VAS with context-free grammars. Remark \[rem:pvas\] relates them with the equivalent model of pushdown vector addition systems.
Vector Addition Systems {#vector-addition-systems .unnumbered}
-----------------------
For a *dimension* $d\in{{\mathbb{N}}}$, we consider *configurations* that are vectors ${{{\mathbf{c}}}},{{{\mathbf{d}}}},{{{\mathbf{x}}}},{{{\mathbf{y}}}},\ldots$ in ${{\mathbb{N}}}^d$, and *actions* that are vectors ${\mathbf{a}}\in{{{\mathbb{Z}}}}^d$. We write ${{{\mathbf{x}}}}\xrightarrow{{{{\mathbf{a}}}}}{{{\mathbf{y}}}}$ for two configurations ${{{\mathbf{x}}}},{{{\mathbf{y}}}}$ in ${{\mathbb{N}}}^d$ if ${{{\mathbf{y}}}}={{{\mathbf{x}}}}+{{{\mathbf{a}}}}$. A *vector addition system* (a *VAS*) is a transition system of the form $({{\mathbb{N}}}^d,\{\xrightarrow{{{{\mathbf{a}}}}}\}_{{{{\mathbf{a}}}}\in{\mathbf{A}}})$ generated by a finite set ${\mathbf{A}}\subseteq{{{\mathbb{Z}}}}^d$ of actions.
In a VAS, the one-step transition relations $\{\xrightarrow{{{{\mathbf{a}}}}}\}_{{{{\mathbf{a}}}}\in {\mathbf{A}}}$ are composed in a natural way: with any word $w={\mathbf{a}}_1\cdots {{{\mathbf{a}}}}_k\in {\mathbf{A}}^*$ of actions, we associate the binary relation $\xrightarrow{w}$ defined over configurations by ${\mathbf{x}}\xrightarrow{w}{\mathbf{y}}$ iff there exists a sequence ${\mathbf{c}}_0,\ldots,{\mathbf{c}}_k$ of configurations such that ${\mathbf{c}}_0={\mathbf{x}}$, ${\mathbf{c}}_k={\mathbf{y}}$ and such that ${\mathbf{c}}_{j-1}\xrightarrow{{{{\mathbf{a}}}}_j}{\mathbf{c}}_{j}$ for every $1\leq j\leq k$. Those relations are monotonic: $$\label{eq-vas-monotonic}
{\mathbf{x}}\xrightarrow{w}{\mathbf{y}} \text{ and }{\mathbf{v}}\in{{\mathbb{N}}}^d\text{
implies }{\mathbf{x}}+{\mathbf{v}}\xrightarrow{w}{\mathbf{y}}+{\mathbf{v}} \:.$$
Notation {#notation .unnumbered}
--------
When writing configurations ${{{\mathbf{c}}}}\in{{\mathbb{N}}}^d$, we sometimes split the vector in parts, writing e.g., ${{{\mathbf{c}}}}=({{{\mathbf{x}}}},{{{\mathbf{y}}}})$ for some ${{{\mathbf{x}}}}\in{{\mathbb{N}}}^{d_1}$ and ${{{\mathbf{y}}}}\in{{\mathbb{N}}}^{d_2}$ with $d=d_1+d_2$. We also write ${{{\mathbf{0}}}}_d$ for the null vector in ${{\mathbb{N}}}^d$, often leaving the dimension implicit.
Grammar-controlled Vector Addition Systems {#grammar-controlled-vector-addition-systems .unnumbered}
------------------------------------------
A $d$-dimensional *grammar-controlled vector addition system* (a *GVAS*) can be seen as a context-free grammar using terminals from ${{{\mathbb{Z}}}}^d$, or equivalently as a VAS where the valid sequences of actions are generated by a context-free grammar. Formally, a $d$-dimensional GVAS is some $G=(V,{\mathbf{A}},R,S)$ where $V$ is a finite set of *nonterminals*, where ${\mathbf{A}}\subseteq
{{{\mathbb{Z}}}}^d$ is a finite set of terminals called *actions*, where $R\subseteq
V\times (V\cup {\mathbf{A}})^*$ is a finite set of production *rules*, and $S\in V$ is the *start symbol*. Following the usual convention, a rule $(T,u)$ is also written $T \rightarrow u$. We denote nonterminals from $V$ with capital letters like $S,T,\ldots$ while symbols from the larger set $V\cup {\mathbf{A}}$ are denoted with $X,Y,\ldots$. Words in $(V\cup {\mathbf{A}})^*$ are denoted with $w,u,v,\ldots$. As usual, $\varepsilon$ denotes the empty word.
For all words $w,w'\in (V\cup {\mathbf{A}})^*$, we say that $w{\xLongrightarrow{}}w'$ is a *derivation step* of $G$ if there exist two words $v,v'$ in $(V\cup {\mathbf{A}})^*$ and a rule $(T,u)$ in $R$ such that $w=vTv'$ and $w'=vuv'$. Let ${\xLongrightarrow{*}}$ denote the reflexive and transitive closure of ${\xLongrightarrow{}}$. The language $L_G\subseteq {\mathbf{A}}^*$ generated by $G$ seen as a grammar is defined as usual with $w\in L_G{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{\Leftrightarrow}}S{\xLongrightarrow{*}}w\in
{\mathbf{A}}^*$. More generally, for any $u\in (V\cup {\mathbf{A}})^*$, the language $L_G(u)$ is $\{w\in
{\mathbf{A}}^*~|~u{\xLongrightarrow{*}}w\}$.
When $G$ is a GVAS, we are interested in what sequences of actions may occur between configurations in ${{\mathbb{N}}}^d$. For this, we extend the definition of the $\xrightarrow{w}$ relation and consider $\xrightarrow{u}$ for any $u\in (V\cup
{\mathbf{A}})^*$. Formally, we let $$\begin{gathered}
{\mathbf{x}}\xrightarrow{u}{\mathbf{y}}
{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{\Leftrightarrow}}\exists w\in L_G(u):{\mathbf{x}}\xrightarrow{w}{\mathbf{y}}\:.\end{gathered}$$ A labeled pair ${{{\mathbf{x}}}}\xrightarrow{u}{{{\mathbf{y}}}}$ is called a *run* of the GVAS, and should not be confused with the derivations $w{\xLongrightarrow{*}}w'$ that only involve the grammar part.
Like VASes, GVASes are monotonic: $$\label{eq-gvas-monotonic}
{\mathbf{x}}\xrightarrow{u}{\mathbf{y}} \text{ and }{\mathbf{v}}\in{{\mathbb{N}}}^d\text{
implies }{\mathbf{x}}+{\mathbf{v}}\xrightarrow{u}{\mathbf{y}}+{\mathbf{v}}\:.$$ The underlying grammar $G$ is left implicit in the above notation. We sometimes write ${{{\mathbf{x}}}}\xrightarrow{G}{{{\mathbf{y}}}}$ instead of ${{{\mathbf{x}}}}\xrightarrow{S}{{{\mathbf{y}}}}$, where $S$ is the start symbol of $G$, when several grammars are considered simultaneously.
\[ex-power-2\] Let $d=1$, $V=\{S,T\}$, and consider the 1-dimensional GVAS given by the following four rules in Backus-Naur form:
[2]{} S & | S T , & T & | T .
Since we shall claim in Section \[sec-weakcomp\] that this GVAS weakly computes the $2^n$ function, let us state and prove the main properties of its runs. Formally, for every $k,k',n,n'\in{{\mathbb{N}}}$, one has: $$\begin{aligned}
\label{eq-spec-power-2}
k\xrightarrow{T}k' & \text{ iff } k\leq k'\leq 2k \:,
&
n\xrightarrow{S}n' & \text{ iff } 1\leq n'\leq 2^n \:.\end{aligned}$$ Thus, even in dimension 1, the reachability relation may not be semilinear.
To prove assume first that $k\xrightarrow{T}k'$ for some natural numbers $k,k'$. There exists $m\in{{{\mathbb{N}}}}$ such that $k\xrightarrow{{\mathbf{-1}}^m{\mathbf{0}}{\mathbf{2}}^m}k'$. In particular $m\leq k$ and $k'=k+m$. We deduce that $k\leq k'\leq 2k$. Conversely, let $k,k'$ be two natural numbers such that $k\leq k'\leq
2k$. Observe that $T{\xLongrightarrow{*}}{\mathbf{-1}}^n{\mathbf{0}}{\mathbf{2}}^n$ where $n$ is defined as $k'-k$. The following relations show that $k\xrightarrow{T}k'$: $$k ~\xrightarrow{{\mathbf{-1}}^n}~ k-n ~\xrightarrow{{\mathbf{0}}}~
k-n ~\xrightarrow{{\mathbf{2}}^n}~ k'\:.$$
Now, assume that $n\xrightarrow{S}n'$ for some natural numbers $n,n'$, There exists $m\in{{{\mathbb{N}}}}$ such that $n\xrightarrow{{\mathbf{-1}}^m{\mathbf{1}}T^m}n'$. It follows that $m\leq n$, and from the previous paragraph, we deduce that $n'\leq (n-m+1)2^m\leq
2^n$ by observing that $x+1\leq 2^x$ for every $x\in{{\mathbb{N}}}$ and by replacing $x$ by $n-m$. Conversely, let $n,n'$ be two natural numbers such that $1\leq
n'\leq 2^n$. Observe that $S{\xLongrightarrow{*}}{\mathbf{-1}}^n{\mathbf{1}}T^n$. Let us introduce a natural number $m$ in $\{0,\ldots,n-1\}$ such that $2^m\leq
n'\leq 2^{m+1}$. The following relations show that $n\xrightarrow{S}n'$: $$\tag*{$\mathqed$}
n\xrightarrow{{\mathbf{-1}}^n}0\xrightarrow{{\mathbf{1}}}2^0\xrightarrow{T}2^1\cdots
\xrightarrow{T}2^{m}\xrightarrow{T}n'\xrightarrow{T^{n-1-m}}n'\:.$$
\[ex-gvas\] Let $G$ be the 2-dimensional GVAS with a single nonterminal symbol and the following three rules: $$S \rightarrow S\,S ~\big|~ {\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)} ~\big|~ {\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle -1\end{array}\bigr)}
\:.$$ Let $w={\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}{\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle -1\end{array}\bigr)}{\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}$ and observe that $S {\xLongrightarrow{*}}w$. We have ${\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 2\end{array}\bigr)} \xrightarrow{{\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}} {\bigl(\begin{array}{@{}c@{}}\scriptstyle 1\\[-.5em] \scriptstyle 4\end{array}\bigr)}
\xrightarrow{{\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle -1\end{array}\bigr)}} {\bigl(\begin{array}{@{}c@{}}\scriptstyle 3\\[-.5em] \scriptstyle 3\end{array}\bigr)} \xrightarrow{{\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}}
{\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 5\end{array}\bigr)}$ and hence ${\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 2\end{array}\bigr)} \xrightarrow{S} {\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 5\end{array}\bigr)}$.
\[rem:pvas\] It is well known that, from a formal language viewpoint, context-free grammars are equivalent to pushdown automata. Similarly, GVAS can be equivalently presented as VAS extended with a pushdown stack. Formally, a $d$-dimensional *Pushdown Vector Addition System* (a *PVAS*) is a transition system of the form $({{\mathbb{N}}}^d \times \Gamma^*, \{\xrightarrow{{{{\mathbf{p}}}}}\}_{{{{\mathbf{p}}}}\in{\mathbf{P}}})$ generated by a pair $(\Gamma, {\mathbf{P}})$ where $\Gamma$ is a finite *stack alphabet* and ${\mathbf{P}} \subseteq \Gamma^* \times \Gamma^* \times {{\mathbb{Z}}}^{d}$ is a finite set of *actions*. So configurations are now pairs $({\mathbf{x}}, u)$ where ${\mathbf{x}} \in {{\mathbb{N}}}^{d}$ is as for VAS and $u \in \Gamma^*$ is a word denoting the contents of the stack. Intuitively, an action ${{{\mathbf{p}}}}= (\alpha, \beta,{\mathbf{a}})$ pops the string $\alpha$ from the top of the stack, then pushes the string $\beta$ onto the top of the stack, and adds ${\mathbf{a}}$ to the vector of natural numbers. Formally, each action $(\alpha, \beta, {\mathbf{a}}) \in {\mathbf{P}}$ induces a binary relation $\xrightarrow{(\alpha,\beta,{\mathbf{a}})}$ on configurations defined by $({\mathbf{x}}, u)
\xrightarrow{(\alpha,\beta,{\mathbf{a}})}
({\mathbf{y}}, v)$ if ${\mathbf{y}} = {\mathbf{x}} + {\mathbf{a}}$ and there exists $w$ such that $u = \alpha w$ and $v = \beta w$. GVASes can be translated into equivalent PVASes and vice-versa.
For instance, the PVAS corresponding to Example \[ex-gvas\] is generated by the pair $(\Gamma, {\mathbf{P}})$ where $\Gamma = \{S\}$ and ${\mathbf{P}}$ is the set of actions $\{(S, SS, {\bigl(\begin{array}{@{}c@{}}\scriptstyle 0\\[-.5em] \scriptstyle 0\end{array}\bigr)}), (S, \varepsilon, {\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}),
(S, \varepsilon, {\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle -1\end{array}\bigr)})\}$. Corresponding to ${\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 2\end{array}\bigr)} \xrightarrow{w} {\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 5\end{array}\bigr)}$ with $w={\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}{\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle -1\end{array}\bigr)}{\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}$ there, we have $$(S,{\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 2\end{array}\bigr)})
\xrightarrow{\! S,SS,{\bigl(\begin{array}{@{}c@{}}\scriptstyle 0\\[-.5em] \scriptstyle 0\end{array}\bigr)}\!} (SS,{\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 2\end{array}\bigr)})
\xrightarrow{\! S, \varepsilon, {\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}\!} (S, {\bigl(\begin{array}{@{}c@{}}\scriptstyle 1\\[-.5em] \scriptstyle 4\end{array}\bigr)})
\xrightarrow{\! S,SS,{\bigl(\begin{array}{@{}c@{}}\scriptstyle 0\\[-.5em] \scriptstyle 0\end{array}\bigr)}\!} (SS, {\bigl(\begin{array}{@{}c@{}}\scriptstyle 1\\[-.5em] \scriptstyle 4\end{array}\bigr)})
\xrightarrow{\! S, \varepsilon, {\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle -1\end{array}\bigr)}\!} (S,{\bigl(\begin{array}{@{}c@{}}\scriptstyle 3\\[-.5em] \scriptstyle 3\end{array}\bigr)})
\xrightarrow{\! S, \varepsilon, {\bigl(\begin{array}{@{}c@{}}\scriptstyle -1\\[-.5em] \scriptstyle 2\end{array}\bigr)}\!} (\varepsilon,{\bigl(\begin{array}{@{}c@{}}\scriptstyle 2\\[-.5em] \scriptstyle 5\end{array}\bigr)})$$ in the PVAS.
Well-Quasi-Ordering Runs in GVASes {#sec-pumping}
==================================
In this section we define flow trees of GVASes and show that they satisfy an amalgamation property. This property is used in the next section to provide a geometrical decomposition of GVAS sets, and in the following section to show that unbounded weakly computable functions are in $\Omega(n)$.
Let $G=(V,{\mathbf{A}},R,S)$ be a $d$-dimensional GVAS. *Flow trees* of $G$ are trees that combine a transition ${{{\mathbf{x}}}}\xrightarrow{w}{{{\mathbf{y}}}}$ in the VAS part of $G$ with a derivation tree for the corresponding $S{\xLongrightarrow{*}}w$ in the grammar part of $G$.
Flow trees are finite rooted ordered trees labeled with transitions of $G$: we write $t = \sigma[t_1,\ldots,t_\ell]$ to denote a flow tree $t$ made of a root with $\ell$ subtrees $t_1,\ldots,t_\ell$. The root is labeled by a transition $\sigma$ of $G$, of the form ${{{\mathbf{c}}}}\xrightarrow{X}{{{\mathbf{d}}}}$ with $X\in V\cup {\mathbf{A}}$. We write ${\operatorname{root}}(t) = \sigma$. Formally, $F(G)$ is the least set of trees that contains all $({{{\mathbf{c}}}}\xrightarrow{{{{\mathbf{a}}}}}{{{\mathbf{d}}}})[]$ with ${{{\mathbf{a}}}}\in {\mathbf{A}}$ and ${{{\mathbf{c}}}}+{{{\mathbf{a}}}}={{{\mathbf{d}}}}$, and all $({{{\mathbf{c}}}}\xrightarrow{T}{{{\mathbf{d}}}})[t_1,\ldots,t_\ell]$ with $T\in V$ and $t_1,\ldots,t_\ell\in F(G)$ such that there is a rule $T{\xLongrightarrow{}}X_1\cdots X_\ell$ in $R$ and configurations ${{{\mathbf{c}}}}_0,{{{\mathbf{c}}}}_1,\ldots,{{{\mathbf{c}}}}_\ell$ with ${{{\mathbf{c}}}}_0={{{\mathbf{c}}}}$, ${{{\mathbf{c}}}}_\ell={{{\mathbf{d}}}}$ and such that, for $i=1,\ldots,\ell$, the root of $t_i$ is labeled with ${{{\mathbf{c}}}}_{i-1}\xrightarrow{X_i}{{{\mathbf{c}}}}_i$. A *subtree* of $t=\sigma[t_1,\ldots,t_\ell]$ is either $t$ itself or a subtree of some $t_i$ for $i=1,\ldots,\ell$. A (sub)tree $({{{\mathbf{c}}}}\xrightarrow{X}{{{\mathbf{d}}}})[t_1,\ldots,t_\ell]$ is a *leaf* when $\ell=0$: this requires that $X={{{\mathbf{a}}}}\in {\mathbf{A}}$ is an action (and then ${{{\mathbf{d}}}}={{{\mathbf{c}}}}+{{{\mathbf{a}}}}$) or that $X=T\in V$ is a non-terminal and $T{\xLongrightarrow{}}\varepsilon$ is a rule in $R$ (and then ${{{\mathbf{d}}}}={{{\mathbf{c}}}}$).
As is standard, we use *positions* to identify occurrences of subtrees inside $t$. Formally, a position is a finite sequence of natural numbers, and the positions of the subtrees of $t$, denoted ${{\mathit{Pos}}}(t)$ are given inductively by $${{\mathit{Pos}}}(\sigma[t_1,\ldots,t_l]) {\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}\{\varepsilon\}\cup\{i.q~|~1\leq i\leq l \land
q\in{{\mathit{Pos}}}(t_i)\} \:.$$ For $p\in{{\mathit{Pos}}}(t)$, the subtree of $t$ at position $p$ is denoted $t/p$.
\[ex-flow-tree\] Recall the 1-dimensional GVAS $G$ from Example \[ex-power-2\]. The grammar admits, among others, a derivation $S{\xLongrightarrow{*}}w$ for $w={\mathbf{-1}}\,{\mathbf{-1}}\,{\mathbf{1}}\,{\mathbf{0}}\,{\mathbf{0}}$. Thus $3\xrightarrow{w} 2$ is a transition in $G$.
In Fig. \[fig-ex-flowtree2\] we display a derivation tree witnessing $S{\xLongrightarrow{*}}w$ and a flow tree witnessing ${{{\mathbf{c}}}}\xrightarrow{w}{{{\mathbf{d}}}}$ for ${{{\mathbf{c}}}}=3$ and ${{{\mathbf{d}}}}=2$.
\(a) [$S$]{}; (a1) \[below=of a,xshift=-3em\] [${\mathbf{-1}}$]{}; (a2) \[below=of a\] [$S$]{}; (a3) \[below=of a,xshift=4em\] [$T$]{}; (a21) \[below=of a2,xshift=-2em\] [${\mathbf{-1}}$]{}; (a22) \[below=of a2\] [$S$]{}; (a23) \[below=of a2,xshift=2em\] [$T$]{}; (a31) \[below=of a3\] [${\mathbf{0}}$]{}; (l23) \[below=1em of a23\] [${\mathbf{0}}$]{}; (l22) \[below=1em of a22\] [${\mathbf{1}}$]{};
\(a) to (a1); (a) to (a2); (a) to (a3); (a2) to (a21); (a2) to (a22); (a2) to (a23); (a3) to (a31); (a22) to (l22); (a23) to (l23);
\(A) \[right=15em of a\] [$3\xrightarrow{S}2$]{}; (A1) \[below=of A,xshift=-6em\] [$3\xrightarrow{{\mathbf{-1}}}2$]{}; (A2) \[below=of A, xshift=-1em\] [$2\xrightarrow{S}2$]{}; (A3) \[below=of A,xshift=7em\] [$2\xrightarrow{T}2$]{}; (A21) \[below=of A2,xshift=-4em\] [$2\xrightarrow{{\mathbf{-1}}}1$]{}; (A22) \[below=of A2\] [$1\xrightarrow{S}2$]{}; (A23) \[below=of A2,xshift=4em\] [$2\xrightarrow{T}2$]{}; (A31) \[below=of A3\] [$2\xrightarrow{{\mathbf{0}}}2$]{}; (L23) \[below=1em of A23\] [$2\xrightarrow{{\mathbf{0}}}2$]{}; (L22) \[below=1em of A22\] [$1\xrightarrow{{\mathbf{1}}}2$]{};
\(A) to (A1); (A) to (A2); (A) to (A3); (A2) to (A21); (A2) to (A22); (A2) to (A23); (A3) to (A31); (A22) to (L22); (A23) to (L23);
We now extend to flow trees of GVASes an ordering initially introduced by Jančar for runs of VASes [@jancar90 Def. 6.4].
\[def-leq\_G\] For two transitions $\sigma={{{\mathbf{c}}}}\xrightarrow{X}{{{\mathbf{d}}}}$ and $\theta={{{\mathbf{c}}}}'\xrightarrow{X'}{{{\mathbf{d}}}}'$ with $X,X'\in V\cup {\mathbf{A}}$, we let $$\sigma\leq\theta \:{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{\Leftrightarrow}}\: {{{\mathbf{c}}}}\leq {{{\mathbf{c}}}}'\land {{{\mathbf{d}}}}\leq{{{\mathbf{d}}}}'\land X=X' \:.$$
The ordering $\leq_G$ between flow trees $s,t\in F(G)$ is defined by induction on the structure of trees: $s=\sigma[s_1,\ldots,s_k]\leq_G t=\theta[t_1,\ldots,t_\ell]$ if, and only if, $\sigma\leq \theta$ and there exists a subtree $t'$ of $t$ of the form $t'=\theta'[t'_1,\ldots,t'_{\ell'}]$ with $\sigma\leq
\theta'$, $\ell'=k$ and $s_j\leq_G t'_j$ for every $1\leq j\leq k$.
This definition is well-founded and, since the subtree relation is transitive, $\leq_G$ is clearly reflexive and transitive, i.e., is a quasi-ordering. In the appendix, we prove the following key property:
\[lem-FG-wqo\] $(F(G),\leq_G)$ is a well-quasi-ordering.
In other words, any infinite sequence $s_0,s_1,s_2,\ldots$ of flow trees contains an infinite increasing subsequence $s_{i_0}\leq_G s_{i_1}\leq_G s_{i_2}\leq_G \cdots$.
\[ex:flow-tree-ordering\] The flow trees shown in Fig. \[fig:ordering\] illustrate the ordering $\le_G$, on a $1$-dimensional GVAS with non-terminals $S,T,U,V$ and the following rules: $S \rightarrow {\mathbf{3}} \, T ~|~
{\mathbf{3}} \, U$, $T \rightarrow {\mathbf{-2}} ~|~ V \, T$, $U \rightarrow T$ and $V \rightarrow \epsilon$. We can see that $t_1
\leq_G t_2$ from the following orderings on the subtrees of $t_1$ and $t_2$: $t_1/1 \leq_G t_2/1$, $t_1/2 \leq_G t_2/2$ and $t_1/21 \leq_G
t_2/221$. It can be verified that $t_1 \not\leq_G t_0$. However, $t_1 \sqsubseteq t_0$, where $\sqsubseteq$ is the standard homeomorphic embedding [^2] of labeled trees. This can be seen by the following orderings: $t_1/1 \sqsubseteq
t_0/1$, $t_1/2 \sqsubseteq t_0/21$ and $t_1/21 \sqsubseteq t_0/211$.
When $\sigma\leq\theta$ for some $\sigma={{{\mathbf{c}}}}\xrightarrow{X}{{{\mathbf{d}}}}$ and $\theta={{{\mathbf{c}}}}'\xrightarrow{X}{{{\mathbf{d}}}}'$, we also write $\sigma\leq^\Delta\theta$ with $\Delta=({{{\mathbf{c}}}}'-{{{\mathbf{c}}}},{{{\mathbf{d}}}}'-{{{\mathbf{d}}}})$. Similarly, we write $s\leq_G^\Delta t$ for two flow trees $s$ and $t$ when $s\leq_G t$ and ${\operatorname{root}}(s)\leq^\Delta{\operatorname{root}}(t)$.
The pair $\Delta$ is called a *lifting*. Note that necessarily $\Delta$ belongs to $({{\mathbb{N}}}^d)^2$ and that $\sigma\leq^\Delta\theta$ and $\theta\leq^{\Delta'}\rho$ entail $\sigma\leq^{\Delta+\Delta'}\rho$. We write $\rho=\sigma+\Delta$ when $\sigma\leq^\Delta\rho$. Two liftings $\Delta=({{{\mathbf{a}}}},{{{\mathbf{b}}}})$ and $\Delta'=({{{\mathbf{a}}}}',{{{\mathbf{b}}}}')$ can be chained if ${{{\mathbf{b}}}}={{{\mathbf{a}}}}'$. In this case we let $\Delta\cdot\Delta'{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}({{{\mathbf{a}}}},{{{\mathbf{b}}}}')$. Note this partial operation is associative.
When $t/p=t'\leq_G u$ we can replace $t'$ by $u$ inside $t$ but this requires a bit of surgery to ensure the result is well-formed. First, for a flow tree $t$ and a displacement ${{{\mathbf{a}}}}\in{{\mathbb{N}}}^d$, we let $t+{{{\mathbf{a}}}}$ be the tree defined via $$\sigma[t_1,\ldots,t_\ell]+{{{\mathbf{a}}}}{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}(\sigma+({{{\mathbf{a}}}},{{{\mathbf{a}}}}))[t_1+{{{\mathbf{a}}}},\ldots,t_\ell+{{{\mathbf{a}}}}]
\:.$$ Obviously, $t+{{{\mathbf{a}}}}$ is a valid flow tree, with $t\leq_G^{({{{\mathbf{a}}}},{{{\mathbf{a}}}})} (t+{{{\mathbf{a}}}})$. Now, when $t/p=t'\leq_G^\Delta u$ for $\Delta=({{{\mathbf{a}}}},{{{\mathbf{b}}}})$, we define $t[u]_p$ by induction on $p$ in the following way:
[2]{} t\[u\]\_ &u, & t\[u\]\_[i.q]{} &(+)\[t\_1+,…,t\_[i-1]{}+,t\_i\[u\]\_q, t\_[i+1]{}+,…,t\_k+\].
\[cl-replace-one-subtree\] If $t/p\leq^\Delta_G u$ then $t[u]_p$ is a valid flow tree satisfying $t\leq_G^{\Delta}t[u]_p$.
By induction on $p$. If $p=\varepsilon$ the claim holds trivially. Assume $p=i.q$ with $1\leq i \leq k$ and let $u'=t_i[u]q$. By induction hypothesis, $t_i\leq_G^\Delta u'$. This implies that $t[u]_p$ is a well-defined flow tree. Since furthermore $t_j\leq_G
t_j+{{{\mathbf{a}}}}$ when $1\leq j<i$, and symmetrically, $t_j\leq_G t_j+{{{\mathbf{b}}}}$ when $i<j\leq k$, we see that $t\leq_G t[u]_p$. Finally, we observe that ${\operatorname{root}}(t)\leq^\Delta{\operatorname{root}}(t[u]_p)$.
\[lem-replace-children\] Let $t=\sigma[t_1,\ldots,t_k]$ and assume $t_i\leq_G^{\Delta_i} u_i$ for $i=1,\ldots,k$. If $\Delta_1\cdot \Delta_2 \cdots \Delta_k=\Delta$ is defined, then $u=(\sigma+\Delta)[u_1,\ldots,u_k]$ is a valid flow tree satisfying $t\leq_G^\Delta u$.
Since $\Delta=\Delta_1\cdots \Delta_k$ is defined, we can write $\Delta_i=({{{\mathbf{a}}}}_{i-1},{{{\mathbf{a}}}}_i)$ and $\Delta=({{{\mathbf{a}}}}_0,{{{\mathbf{a}}}}_k)$. Assume $\sigma={{{\mathbf{c}}}}_0\xrightarrow{X}{{{\mathbf{c}}}}_k$, with ${\operatorname{root}}(t_i)={{{\mathbf{c}}}}_{i-1}\xrightarrow{Y_i}{{{\mathbf{c}}}}_i$ for $i=1,\ldots,k$. Then ${\operatorname{root}}(u_i)=({{{\mathbf{c}}}}_{i-1}+{{{\mathbf{a}}}}_{i-1})\xrightarrow{Y_i}({{{\mathbf{c}}}}_i+{{{\mathbf{a}}}}_i)$ and $u$ is a valid transition. That $t\leq_G u$ is immediate.
\[thm:amalgamation\] If $s\leq_G^{\Delta_1} t_1$ and $s\leq_G^{\Delta_2} t_2$ then there exists $s'$ s.t. $t_1\leq_G^{\Delta_2}s'$ and $t_2\leq_G^{\Delta_1}s'$ (further entailing $s\leq_G^{\Delta_1+\Delta_2}s'$).
By induction on $s$. Assume $s=\sigma[s_1,\ldots,s_k]$. Since $s\leq_G^{\Delta_1} t_1$, there is a subtree $t_1/p=t^1=\rho_1[t_1^1,\ldots,t_k^1]$ of $t_1$ such that $\sigma\leq\rho_1$ and $s_j\leq_G t_j^1$ for all $j=1,\ldots,k$. Assume that $\sigma\leq^{\Delta'_1}\rho_1$ and that $s_j\leq_G^{\Gamma_j}t_j^1$ for $j=1,\ldots,k$. Since $s$ and $t^1$ are valid flow trees, we deduce that $\Delta'_1=\Gamma_1 \cdots
\Gamma_k$. Symmetrically, from $s\leq_G^{\Delta_2}t_2$, we know that there is a subtree $t^2=t_2/q$ of $t_2$, of the form $t^2=\rho_2[t^2_1,\ldots,t^2_k]$ with $\sigma\leq^{\Delta'_2}\rho_2$, $s_j\leq_G^{\Gamma'_j}t^2_j$ for $j=1,\ldots,k$, and $\Delta'_2=\Gamma'_1\cdots\Gamma'_k$.
By the induction hypothesis, there exists flow trees $s'_1,\ldots,s'_k$ such that $t^1_j\leq_G^{\Gamma'_j}s'_j$ and $t^2_j\leq_G^{\Gamma_j}s'_j$ for all $j=1,\ldots,k$. We now define
[3]{} u&(\_2+’\_1)\[s’\_1,…,s’\_k\], & u’&t\_2\[u\]\_q, & s’&t\_1\[u’\]\_p,
and claim that these are valid flow trees, $s'$ being the flow tree witnessing the Lemma.
To begin with, and since $\Delta'_1=\Gamma_1\cdots\Gamma_k$, $u$ is well-formed by Lemma \[lem-replace-children\] and satisfies $t^2\leq_G^{\Delta'_1}u$. Since $\rho_2+\Delta'_1=\rho_1+\Delta'_2$, and since $\Delta'_2=\Gamma'_1\cdots\Gamma'_k$, one also has $t^1\leq_G^{\Delta'_2}u$.
Then, and since $t^2\leq_G^{\Delta'_1} u$, we have $t_2=t_2[t^2]_q\leq_G^{\Delta'_1}t_2[u]_q=u'$ as in Claim \[cl-replace-one-subtree\]. Thus the root of $u'$ is $\sigma+\Delta_2+\Delta'_1=\rho_1+\Delta_2$. We deduce $t^1\leq_G^{\Delta_2}u'$, relying on $t^1\leq_G u$. As in Claim \[cl-replace-one-subtree\], we obtain $t_1\leq_G^{\Delta_2}t_1[u']_p=s'$, proving the first half of the Lemma.
On the other hand, from $t_2\leq_G u'$ we get $t_2\leq_G^{\Delta_1}
s'$ by just checking that the root of $t_2$, i.e., $\sigma+\Delta_2$, is smaller than the root of $s'$, i.e., $\sigma+\Delta_1+\Delta_2$. This provides the other half and completes the proof.
GVAS-Definable Predicates {#sec-gvas-sets}
=========================
We explore in this section a natural notion of computable sets and relations for the GVAS model, defined as projections of reachability sets. The context-free grammar ingredient of GVASes is essential in the proof that the class of computable sets is closed under intersection, while the Amalgamation Theorem proves that computable sets are finite union of shifted periodic sets.
A $n$-dimensional *GVAS-definable predicate* is a subset ${\mathbf{X}}$ of ${{\mathbb{N}}}^n$ such that there exists a $d$-dimensional GVAS $G$ with $d=n+\ell$ for some $\ell\in{{\mathbb{N}}}$ such that: $$\tag{$\dagger$}\label{eq-gvas-set}
{\mathbf{X}}=\{ {{{\mathbf{x}}}}\in{{\mathbb{N}}}^n \mid \exists {{{\mathbf{e}}}}\in{{\mathbb{N}}}^\ell : {{{\mathbf{0}}}}_d\xrightarrow{G}({{{\mathbf{x}}}},{{{\mathbf{e}}}}) \} \:.$$ When holds, we say that $G$ defines ${\mathbf{X}}$ using $\ell$ auxiliary counters.
The class of GVAS-definable predicates is clearly closed under union, cartesian product [^3], and by projecting away some components. We will provide additional closure properties in the remainder of this section. GVAS-definable predicates form a rich class that strictly contains all Presburger sets, i.e., subsets of ${{\mathbb{N}}}^n$ that are definable in ${\mathsf{FO}}({{\mathbb{N}}};+)$, the first-order theory of natural numbers with addition.
Presburger sets are GVAS-definable predicates. The proof is obtained by introducing the class of semilinear sets as follows. A *linear set* of ${{\mathbb{N}}}^n$ is a set of the form $\{{\mathbf{b}}+\lambda_1{\mathbf{p}}_1+\cdots+\lambda_k{\mathbf{p}}_k \mid \lambda_j\in{{\mathbb{N}}}\}$ where ${\mathbf{b}}$ and ${\mathbf{p}}_1,\ldots,{\mathbf{p}}_k$ are vectors in ${{\mathbb{N}}}^n$. A *semilinear set* of ${{\mathbb{N}}}^n$ is a finite union of linear sets of ${{\mathbb{N}}}^n$. Let us recall that a subset of ${{\mathbb{N}}}^n$ is Presburger if, and only if, it is semilinear [@gs66]. Since the class of GVAS-definable predicates is closed under union, it is sufficient to show that every linear set is GVAS-definable. We associate with a linear set ${\mathbf{X}}=\{{\mathbf{b}}+\lambda_1{\mathbf{p}}_1+\cdots+\lambda_k{\mathbf{p}}_k \mid
\lambda_j\in{{\mathbb{N}}}\}$ the $n$-dimensional GVAS $G$ that generates the regular language ${\mathbf{b}}{\mathbf{p}}_1^*\ldots{\mathbf{p}}_k^*$. Notice that $G$ defines the linear set ${\mathbf{X}}$ (using no auxiliary counter).
\[ex-power-2-as-gvas-set\] Let $G$ be the 2-dimensional GVAS given by the following four rules:
[2]{} S &
(
[@c@]{}0\
1
)
|
(
[@c@]{}1\
0
)
S T , & T &
(
[@c@]{}0\
0
)
|
(
[@c@]{}0\
-1
)
T
(
[@c@]{}0\
2
)
.
This GVAS is a variant of the 1-dimensional GVAS given in Example \[ex-power-2\]. Analogously to that example, it can be shown that $G$ defines the set ${\mathbf{X}} = \{ (x, y) \in {{\mathbb{N}}}^2 \mid 1 \leq y \leq 2^x \}$, which is not semilinear.
A geometrical decomposition of the GVAS-definable predicates can be shown thanks to the *periodic sets*. A subset ${\mathbf{P}}$ of ${{\mathbb{N}}}^d$ is said to be *periodic* [@DBLP:conf/popl/Leroux11] if it contains the zero vector, and if ${\mathbf{x}}+{\mathbf{y}}\in{\mathbf{P}}$ for every ${\mathbf{x}},{\mathbf{y}}\in{\mathbf{P}}$. Note that a periodic ${\mathbf{P}}$ is not necessarily finitely generated, hence not necessarily semilinear (see Example \[ex-power-2-as-gvas-set-cont\]). The following Proposition extends the known decomposition of Presburger sets into linear sets.
\[prop-gvas-set-periodic\] Every GVAS-definable predicate ${\mathbf{X}}\subseteq {{\mathbb{N}}}^n$ can be decomposed into a finite union of sets of the form ${\mathbf{b}}+{\mathbf{P}}$ where ${\mathbf{b}}\in{{\mathbb{N}}}^n$ and ${\mathbf{P}}$ is a periodic subset of ${{\mathbb{N}}}^n$.
There exists a $d$-dimensional GVAS $G$ that defines the set ${\mathbf{X}}$ with $\ell$ auxiliary counters. Let us consider the set $T$ of flow trees $t$ such that ${\operatorname{root}}({t})=({\mathbf{0}}_{d},S,({\mathbf{x}},{\mathbf{e}}))$ for some ${\mathbf{x}}\in{{\mathbb{N}}}^n$, ${\mathbf{e}}\in{{\mathbb{N}}}^\ell$ and where $S$ is the start symbol of $G$. For such a flow tree $t$ in $T$, we denote by $\mu(t)$ the vector ${\mathbf{x}}$. With each $s\in T$ we associate the set ${\uparrow} s=\{t\in T \mid s \leq_G t\}$. Since $(T,\leq_G)$ is a wqo, there exists a finite subset $T_0$ of $T$ such that $$T = \bigcup_{s\in T_0} {\uparrow} s \:.$$ Given $s\in T$, we introduce the set ${\mathbf{P}}_s=\{\mu(t)-\mu(s)
\mid t\in {\uparrow} s\}$. Theorem \[thm:amalgamation\] shows that ${\mathbf{P}}_s$ is a periodic set. Now, just observe that the following equality holds: $${\mathbf{X}}=\bigcup_{s\in T_0}\mu(s)+{\mathbf{P}}_{s} \:.$$ The proposition is proved.
\[ex-power-2-as-gvas-set-cont\] Continuing Example \[ex-power-2-as-gvas-set\], the set ${\mathbf{X}} = \{ (x, y) \in {{\mathbb{N}}}^2 \mid 1 \leq y \leq 2^x \}$ may be decomposed into ${\mathbf{X}} = (0, 1) + {\mathbf{P}}$ where ${\mathbf{P}}$ is the periodic set ${\mathbf{P}} = \{ (x, y) \in {{\mathbb{N}}}^2 \mid 0 \leq y < 2^x \}$.
The rest of this section discusses various closure properties of GVAS-definable predicates. We start with boolean operations. As mentioned previously, GVAS-definable predicates are closed under union. In order to prove closure under intersection, we first provide a technical lemma that shows how auxiliary counters of a GVAS can be assumed to be zero at the end of the computation.
\[lem:zero\] For every $d$-dimensional GVAS $G$ and for every subset $I$ of $\{1,\ldots,d\}$, there is a $(d+1)$-dimensional GVAS $G_I$ such that for every ${{{\mathbf{x}}}}\in{{\mathbb{N}}}^d$ and for every $c\in {{\mathbb{N}}}$, we have: $$\label{eq-lem:zero}
{{{\mathbf{0}}}}_{d+1}\xrightarrow{G_I}({{{\mathbf{x}}}},c)
\text{ iff } {{{\mathbf{0}}}}_d\xrightarrow{G}{{{\mathbf{x}}}}\land
c=0 \land \bigwedge_{i\in I}{{{\mathbf{x}}}}[i]=0
\:.$$
The idea of the proof is to put the counters in $I$ “on a budget” (see, e.g., [@phs-mfcs2010] for details on the budgeting construction) and to harness the expressive power given by context-free grammars to non-deterministically initialize the total budget, simulate $G$ with the given budget, and finally check that the budget is fully restored at the end of the computation, which guarantees that the counters in $I$ are zero.
Formally, let us introduce the function $\Delta_I$ that maps vectors ${{{\mathbf{x}}}}$ of ${{{\mathbb{Z}}}}^d$ to the number $\Delta_I({{{\mathbf{x}}}})=\sum_{i\in I}{\mathbf{x}}[i]$. We also introduce the mapping $\mu_I:{{{\mathbb{Z}}}}^d\rightarrow{{{\mathbb{Z}}}}^{d+1}$ defined by $\mu_I({{{\mathbf{x}}}})=({{{\mathbf{x}}}},-\Delta_I({{{\mathbf{x}}}}))$. This mapping is extended over words of actions as a word morphism, and over languages by $\mu_I(L)=\{\mu_I(w) \mid w\in L\}$. Let us introduce the actions ${\mathbf{a}}_+=({{{\mathbf{0}}}}_d,1)$, and ${\mathbf{a}}_-=({{{\mathbf{0}}}}_d,-1)$. In linear time, from $G$ we can define a $(d+1)$-dimensional GVAS $G_I$ that generates the following language: $$L_{G_I}=\bigcup_{k\in{{\mathbb{N}}}}{\mathbf{a}}_+^k\mu_I(L_G) {\mathbf{a}}_-^k \:.$$ Let us prove that this GVAS satisfies the lemma. We consider ${{{\mathbf{x}}}}\in{{\mathbb{N}}}^d$ and $c\in{{\mathbb{N}}}$.
Assume first that ${{{\mathbf{0}}}}_{d+1}\xrightarrow{G_I}({{{\mathbf{x}}}},c)$. In that case, there exists $k\in{{\mathbb{N}}}$ and $w\in L_G$ such that ${{{\mathbf{0}}}}_{d+1}\xrightarrow{{\mathbf{a}}_+^k\mu_I(w) {\mathbf{a}}_-^k}({{{\mathbf{x}}}},c)$. Observe that we have $${{{\mathbf{0}}}}_{d+1}\xrightarrow{{\mathbf{a}}_+^k}({{{\mathbf{0}}}}_d,k)\xrightarrow{\mu_I(w)}({{{\mathbf{x}}}},c+k)\xrightarrow{{\mathbf{a}}_-^k}({{{\mathbf{x}}}},c) \:.$$ Since $\mu_I(w)$ preserves the sum of the counters in $I$ and of the last counter, it follows that $\Delta_I({{{\mathbf{0}}}}_d)+k=\Delta_I({{{\mathbf{x}}}})+c+k$. Thus $c+\Delta_I({{{\mathbf{x}}}})=0$. It follows that $c=0$ and ${{{\mathbf{x}}}}[i]=0$ for every $i\in I$. Moreover, from ${{{\mathbf{0}}}}_d\xrightarrow{w}{{{\mathbf{x}}}}$ we derive ${{{\mathbf{0}}}}_d\xrightarrow{G}{{{\mathbf{x}}}}$.
Conversely, let us assume that ${{{\mathbf{0}}}}_d\xrightarrow{G}{{{\mathbf{x}}}}$, $c=0$ and ${{{\mathbf{x}}}}[i]=0$ for every $i\in I$. There exists $w\in L_G$ such that ${{{\mathbf{0}}}}_d\xrightarrow{w}{{{\mathbf{x}}}}$. There exists $k\in{{\mathbb{N}}}$ large enough such that $({{{\mathbf{0}}}}_d,k)\xrightarrow{\mu_I(w)}({{{\mathbf{x}}}},k-\Delta_I({{{\mathbf{x}}}}))=({{{\mathbf{x}}}},k)$. It follows that ${{{\mathbf{0}}}}_{d+1}\xrightarrow{{\mathbf{a}}_+^k\mu_I(w){\mathbf{a}}_-^k}({{{\mathbf{x}}}},0)=({{{\mathbf{x}}}},c)$. Thus ${{{\mathbf{0}}}}_{d+1}\xrightarrow{G_I}({{{\mathbf{x}}}},c)$.
We are now ready to prove that GVAS-definable predicates are closed under intersection.[^4]
\[lemma:closure-intersection\] The class of GVAS-definable predicates is closed under intersection.
Let ${\mathbf{X}},{\mathbf{Y}}\subseteq {{{\mathbb{N}}}}^n$ be GVAS-definable. Since the class of GVAS-definable predicates is closed under cartesian product, it follows that ${\mathbf{X}}\times{\mathbf{Y}}$ is also GVAS-definable. Hence, there exists a $d$-dimensional GVAS $H$ with $\ell$ auxiliary counters that defines that set. Let us consider the mapping $\mu:{{{\mathbb{Z}}}}^d\rightarrow {{{\mathbb{Z}}}}^n\times{{{\mathbb{Z}}}}^d$ defined by $\mu({\mathbf{x}},{\mathbf{y}},{\mathbf{e}})=({\mathbf{0}}_n,{\mathbf{x}},{\mathbf{y}},{\mathbf{e}})$ for every ${\mathbf{x}},{\mathbf{y}}\in{{{\mathbb{Z}}}}^n$ and ${\mathbf{e}}\in{{{\mathbb{Z}}}}^\ell$. The mapping $\mu$ is extended as a word morphism. We introduce the action ${\mathbf{a}}_i$ in ${{{\mathbb{Z}}}}^{n+d}$ defined as follows: $${\mathbf{a}}_i=({{{\mathbf{i}}}}_{i,n},-{{{\mathbf{i}}}}_{i,n},-{{{\mathbf{i}}}}_{i,n},{{{\mathbf{0}}}}_\ell) \:.$$ Obviously, we can build a $(d+n)$-dimensional GVAS $G$ such that $L_{G}=\mu(L_H){\mathbf{a}}_1^*\cdots{\mathbf{a}}_d^*$. Now, let $I=\{n+1,\ldots,3n\}$ and let us apply Lemma \[lem:zero\] on $G$ and $I$. We obtain a $(d+n+1)$-GVAS that defines ${\mathbf{X}}\cap{\mathbf{Y}}$.
The class of GVAS-definable predicates is not closed under taking complements, see Proposition \[prop:stabcomplement\].
We now investigate closure under sum [^5] and under the associated Kleene star, which we call periodic hull. Formally, the *periodic hull* of a subset ${\mathbf{X}} \subseteq {{{\mathbb{N}}}}^n$ is the set of finite sums of vectors in ${\mathbf{X}}$. It turns out that the class of GVAS-definable predicates is closed under sum and periodic hull. Closure under sum can be proved along the same lines as closure under intersection (see Lemma \[lemma:closure-intersection\]). The detailed proof is left as an exercise. Closure under periodic hull is more involved and requires well-behaved GVASes.
In the definition of a GVAS-definable predicate ${\mathbf{X}}$ given in (\[eq-gvas-set\]), the vector ${\mathbf{e}}$ can be seen as auxiliary counters that can have arbitrary values at the end of the computation. We say that a $d$-dimensional GVAS $G$ defining a predicate ${\mathbf{X}}\subseteq {{{\mathbb{N}}}}^n$ using $\ell$ auxiliary counters is *auxiliary-resetting* if for every $({\mathbf{x}},{\mathbf{e}})\in{{{\mathbb{N}}}}^n\times{{{\mathbb{N}}}}^\ell$ such that ${\mathbf{0}}_d\xrightarrow{G}({{{\mathbf{x}}}},{{{\mathbf{e}}}})$ we have ${{{\mathbf{e}}}}={\mathbf{0}}_\ell$. We also say that $G$ is *output-increasing* if every action of $G$ is an element of ${{{\mathbb{N}}}}^n\times{{{\mathbb{Z}}}}^\ell$ meaning that the output counters ${{{\mathbf{x}}}}$ cannot be decremented during a computation.
Let us first prove that GVAS-definable predicates can be defined by auxiliary-resetting output-increasing GVAS. To do so, we introduce, for every $k\in{{{\mathbb{N}}}}$, the unit vector ${\mathbf{i}}_{i,k}$ of ${{{\mathbb{N}}}}^k$ defined by ${\mathbf{i}}_{i,k}[j]=0$ if $j\not=i$ and ${\mathbf{i}}_{i,k}[i]=1$.
\[lem:strongdef\] For every $d$-dimensional GVAS $G$ defining a set ${\mathbf{X}}\subseteq{{{\mathbb{N}}}}^n$, there is a $(d+n+1)$-dimensional auxiliary-resetting output-increasing GVAS defining ${\mathbf{X}}$.
We introduce the mapping $\mu:{{{\mathbb{Z}}}}^d\rightarrow{{{\mathbb{Z}}}}^{n+d}$ defined by $\mu({{{\mathbf{a}}}})=({\mathbf{0}}_n,{{{\mathbf{a}}}})$. This mapping is extended as a word morphism. Observe that there exists a $(d+n)$-dimensional GVAS $G'$ satisfying the following equality: $$L_{G'}=\mu(L_G)({\mathbf{i}}_{1,n},-{\mathbf{i}}_{1,n},{\mathbf{0}}_\ell)^*\cdots
({\mathbf{i}}_{n,n},-{\mathbf{i}}_{n,n},{\mathbf{0}}_\ell)^*({\mathbf{0}}_n,{\mathbf{0}}_n,-{\mathbf{i}}_{1,\ell})^*\cdots
({\mathbf{0}}_n,{\mathbf{0}}_n,-{\mathbf{i}}_{\ell,\ell})^* \:.$$ This GVAS $G'$ satisfies ${\mathbf{0}}_{d+n}\xrightarrow{G'}({{{\mathbf{x}}}},{\mathbf{0}}_n,{\mathbf{0}}_\ell)$ with ${{{\mathbf{x}}}}\in
{{{\mathbb{N}}}}^n$ if, and only if, there exists ${{{\mathbf{e}}}}\in{{{\mathbb{N}}}}^\ell$ such that ${\mathbf{0}}_d\xrightarrow{G}({{{\mathbf{x}}}},{{{\mathbf{e}}}})$. Moreover, actions of that GVAS are in ${{{\mathbb{N}}}}^n\times{{{\mathbb{Z}}}}^d$. By applying the construction given in the proof of Lemma \[lem:zero\] on $G'$ with $I=\{n,\ldots,d+n\}$ observe that we get a $(d+n+1)$-dimensional auxiliary-resetting output-increasing GVAS defining ${\mathbf{X}}$.
The class of GVAS-definable predicates is closed under periodic hull.
Assume that a predicate ${\mathbf{X}}\subseteq {{{\mathbb{N}}}}^n$ is GVAS-definable. Lemma \[lem:strongdef\] shows that ${\mathbf{X}}$ is defined by a $d$-dimensional auxiliary-resetting output-increasing GVAS $G$ using $\ell$ auxiliary counters. As context-free languages are closed under Kleene star, there exists a GVAS $G'$ satisfying $L_{G'} = L_G^*$. We show that $G'$ defines the periodic hull of ${\mathbf{X}}$. Let ${\mathbf{y}}={\mathbf{x}}_1+\cdots+{\mathbf{x}}_k$ with $k\in{{{\mathbb{N}}}}$ and ${\mathbf{x}}_1,\ldots,{\mathbf{x}}_k\in{\mathbf{X}}$ and let us prove that ${\mathbf{0}}_d\xrightarrow{G'}({\mathbf{y}},{\mathbf{0}}_\ell)$. We introduce ${\mathbf{y}}_j={\mathbf{x}}_1+\cdots+{\mathbf{x}}_j$ for every $j\in\{0,\ldots,k\}$. Let $i\in\{1,\ldots,k\}$. Since ${{{\mathbf{x}}}}_i\in
{\mathbf{X}}$ and $G$ is auxiliary-resetting, we get ${\mathbf{0}}_d\xrightarrow{G}({\mathbf{x}}_i,{\mathbf{0}}_\ell)$. By monotony, we can add on both sides the vector $({\mathbf{y}}_{i-1},{\mathbf{0}}_\ell)$ and derive $({\mathbf{y}}_{i-1},{\mathbf{0}}_{\ell})\xrightarrow{G}({\mathbf{y}}_{i},{\mathbf{0}}_\ell)$. We get $({\mathbf{y}}_0,{\mathbf{0}}_{\ell}) \xrightarrow{G} ({\mathbf{y}}_{1},{\mathbf{0}}_\ell) \cdots \xrightarrow{G} ({\mathbf{y}}_k,{\mathbf{0}}_\ell)$. Since ${\mathbf{y}}_0={\mathbf{0}}_n$ and ${\mathbf{y}}_k={\mathbf{y}}$, we have proved that ${\mathbf{0}}_d\xrightarrow{G'}({\mathbf{y}},{\mathbf{0}}_\ell)$. Conversely, let ${\mathbf{y}}\in{{{\mathbb{N}}}}^n$ and ${\mathbf{e}}\in{{{\mathbb{N}}}}^\ell$ such that ${\mathbf{0}}_d\xrightarrow{G'}({\mathbf{y}},{\mathbf{e}})$ and let us prove that ${\mathbf{y}}$ is in the periodic hull of ${\mathbf{X}}$. Since $L_{G'} = L_G^*$, we have $({\mathbf{y}}_0,{\mathbf{e}}_0) \xrightarrow{G} ({\mathbf{y}}_{1},{\mathbf{e}}_1) \cdots \xrightarrow{G} ({\mathbf{y}}_k,{\mathbf{e}}_k)$ for some sequence $({\mathbf{y}}_0,{\mathbf{e}}_0), \ldots, ({\mathbf{y}}_k,{\mathbf{e}}_k)$ such that $({\mathbf{y}}_0,{\mathbf{e}}_0)={\mathbf{0}}_d$ and $({\mathbf{y}}_k,{\mathbf{e}}_k)=({\mathbf{y}},{\mathbf{e}})$. Since $G$ is output-increasing we deduce that ${\mathbf{x}}_i$, defined as ${\mathbf{x}}_i = {\mathbf{y}}_i-{\mathbf{y}}_{i-1}$, is in ${{{\mathbb{N}}}}^n$ and satisfies $({\mathbf{0}}_n,{\mathbf{e}}_{i-1})\xrightarrow{G}({\mathbf{x}}_i,{\mathbf{e}}_i)$. As $G$ is auxiliary-resetting, by induction we deduce that ${\mathbf{e}}_i={\mathbf{0}}$ for every $i$. It follows that ${\mathbf{x}}_i\in{\mathbf{X}}$. As ${\mathbf{y}}={\mathbf{x}}_1+\cdots+{\mathbf{x}}_k$, we conclude that ${\mathbf{y}}$ is in the periodic hull of ${\mathbf{X}}$.
To conclude this section, we discuss closure under relational composition and under the associated Kleene star (namely, the reflexive-transitive closure). For the purpose of GVAS-definability, we view binary relations on ${{\mathbb{N}}}^n$ as subsets of ${{\mathbb{N}}}^{2n}$. The class of GVAS-definable binary relations on ${{\mathbb{N}}}^n$ is closed under relational composition. This claim follows from closure of GVAS-definable predicates under cartesian product, intersection and projection. However, GVAS-definable binary relations are not closed under reflexive-transitive closure in general, as the following example shows.
Consider the binary relation ${\mathbf{R}}$ on ${{\mathbb{N}}}$ defined by ${\mathbf{R}} = \{(x, 2x) \mid x \in {{\mathbb{N}}}\}$. The binary relation ${\mathbf{R}}$ is clearly GVAS-definable. However, its reflexive-transitive closure ${\mathbf{R}}^* = \{(x, 2^k x) \mid x, k \in {{\mathbb{N}}}\}$ is not GVAS-definable. Indeed, if ${\mathbf{R}}^*$ were GVAS-definable then, by Proposition \[prop-gvas-set-periodic\], there would exist some ${\mathbf{b}} \in {{\mathbb{N}}}^2$, some periodic ${\mathbf{P}} \subseteq {{\mathbb{N}}}^2$ and two distinct powers $2\leq 2^k<2^\ell$ such that $$(1, 2^k), (1, 2^\ell) \in {\mathbf{b}} + {\mathbf{P}} \subseteq {\mathbf{R}}^* \:.$$ We now use the assumption that ${\mathbf{P}}$ is periodic and derive a contradiction. Let us write ${\mathbf{b}} = (b_1, b_2)$. Note that $b_1 \leq 1$ and $b_2 \leq 2^k$ since $(1, 2^k) \in {\mathbf{b}} + {\mathbf{P}}$. There are three cases.
- If ${\mathbf{b}} = {\mathbf{0}}$ then $(1, 2^k)$ and $(1, 2^\ell)$ are both in ${\mathbf{P}}$, hence, $(2, 2^k + 2^\ell)$ is also in ${\mathbf{P}}$ by periodicity, and so $(2, 2^k + 2^\ell) \in {\mathbf{R}}^*$. This is impossible since $2^k + 2^\ell$ is not a power of two (as $k \neq \ell$).
- If $b_1 > 0$ then $b_1 = 1$, hence, $(1, 2^\ell) = (1, b_2) + (0, p_2)$ for some $(0, p_2) \in {\mathbf{P}}$. Note that $p_2 > 0$ since $b_2 \leq 2^k < 2^\ell$. Since ${\mathbf{b}} + {\mathbf{P}} \subseteq {\mathbf{R}}^*$, we get by periodicity that $(1, b_2 + n p_2) \in {\mathbf{R}}^*$ for every $n \in {{\mathbb{N}}}$. This means that $b_2 + n p_2$ is a power of two for every $n \in {{\mathbb{N}}}$, which is impossible as $p_2 > 0$.
- If $b_1 = 0$ and $b_2 > 0$ then $(1, 2^\ell) = (0, b_2) + (1, p_2)$ for some $(1, p_2) \in {\mathbf{P}}$. Since ${\mathbf{b}} + {\mathbf{P}} \subseteq {\mathbf{R}}^*$, we get by periodicity that $(n, b_2 + n p_2) \in {\mathbf{R}}^*$ for every $n \in {{\mathbb{N}}}$. Taking $n = b_2 + 1$, we derive that $(b_2 + 1, b_2 + (b_2 + 1) p_2) \in {\mathbf{R}}^*$. This is impossible since $b_2 > 0$ and $y$ is a multiple of $x$ for every $(x, y) \in {\mathbf{R}}^*$.
Weakly Computable Functions {#sec-weakcomp}
===========================
There is a classical notion of number-theoretical functions weakly computable by Petri nets [@hack76b]. In this section, we extend the idea to GVASes.
As we argued in the introduction, the notion of weakly computable functions has recently gained new relevance with the development of well-structured systems that go beyond Petri nets and VASSes in expressive power, while sharing some of their characteristics.
The expected way for a GVAS to compute a numerical function $f:{{\mathbb{N}}}\to{{\mathbb{N}}}$ is to start with some input number $n$ stored in a designated input counter and, from that configuration, eventually reach a configurations with $f(n)$ in a designated output counter. In order for that GVAS to be correct (as a computer for $f$), it should be impossible that it reaches a value differing from $f(n)$ in the output counter. In that case, we say that the GVAS *strongly computes* $f$. This notion of correctness is fine with other models like Minsky machines but it is too strong for GVASes and does not lead to an interesting family of computable functions. In fact, GVAS are essentially nondeterministic devices, and the above notion of strongly computing some function does not accommodate nondeterminism nicely.
With this in mind, and in the setting of VASes, Rabin defined a notion of “weakly computing $f$” that combines the following two principles:
Completeness
: For any $n\in{{\mathbb{N}}}$, there is a computation with input $n$ and output $f(n)$;
Safety
: Any computation from input $n$ to some output $r$ satisfies $r\leq f(n)$.
This leads to our definition of weak GVAS computers, where the input and output counters are the first two components.
\[def-wpn\] Let $f:{{\mathbb{N}}}\to{{\mathbb{N}}}$ be a total function. A *weak GVAS computer* (with $\ell$ auxiliary counters) for $f$ is a $d$-dimensional GVAS $G$ with $d=2+\ell$ that satisfies the following two properties: $$\begin{aligned}
\tag{CO}\label{eq-co}
\forall n: \exists n',{\mathbf{e}}:\: &
(n,0,{{{\mathbf{0}}}}_\ell)
\xrightarrow{G}
(n',f(n),{\mathbf{e}})
\:,
\\
\tag{SA}\label{eq-sa}
\forall n, n',r,{\mathbf{e}}:\: &
(n,0,{{{\mathbf{0}}}}_\ell)\xrightarrow{G}(n',r,{\mathbf{e}})
\text{ implies } r\leq f(n)
\:.\end{aligned}$$ We say that $f$ is *weakly computable*, or WC, if there is a weak GVAS computer for it.
For convenience, Definition \[def-wpn\] assumes that the input is given in the first counter of $G$, and that the result is found in the second counter. Note that $G$ may use its $\ell$ last counters for auxiliary calculations. We focus on total functions over the natural numbers rather than total functions over the vectors of natural numbers to simplify the presentation. However, results given in this section can be easily extended to this more general setting.
\[ex:power-WC\] Example \[ex-power-2\] shows that the function $f:{{\mathbb{N}}}\rightarrow{{\mathbb{N}}}$ defined by $f(n)=2^{n}$ is WC.
Only monotonic functions can be weakly computed in the above sense. This is an immediate consequence of the monotonicity of GVASes (see ). Recall that a total function $f:{{\mathbb{N}}}\to{{\mathbb{N}}}$ is *non-decreasing* if $n\leq m$ implies $f(n)\leq f(m)$.
\[prop-WCPN-mono\] If $f$ is WC then $f$ is non-decreasing.
Assume that $n\leq m$ and pick any weak GVAS computer $G$ for $f$. By , we have $(n,0,{{{\mathbf{0}}}}_\ell) \xrightarrow{G}
(n',f(n),{\mathbf{e}})$ for some $n'\in{{\mathbb{N}}}$ and ${\mathbf{e}}\in{{\mathbb{N}}}^\ell$. By monotonicity, it follows that $(n+(m-n),0,{{{\mathbf{0}}}}_\ell) \xrightarrow{G}
(n'+(m-n),f(n),{\mathbf{e}})$. We get $f(n)\leq f(m)$ by .
We may now relate WC computability with GVAS-definability.
\[lem:WCGVAS\] A total function $f:{{\mathbb{N}}}\to{{\mathbb{N}}}$ is WC if, and only if, $f$ is non-decreasing and the following set is GVAS-definable. $$\{(x,y)\in{{{\mathbb{N}}}}\times{{{\mathbb{N}}}}\mid y\leq f(x)\} \:.$$
Assume first that $f$ is WC. There exists a weak GVAS computer (with $\ell$ auxiliary counters) for $f$ given as a $d$-dimensional GVAS $G$ with $d=2+\ell$ that satisfies (\[eq-co\]) and (\[eq-sa\]). Let us consider the mapping $\mu:{{{\mathbb{Z}}}}^d\rightarrow{{{\mathbb{Z}}}}^{d+1}$ defined by $\mu(a,b,{\mathbf{e}})=(0,b, a,{\mathbf{e}})$ for every $a,b\in{{{\mathbb{Z}}}}$, and ${\mathbf{e}}\in{{{\mathbb{Z}}}}^\ell$. The mapping $\mu$ is extended as a word morphism. Let us show that a GVAS $G'$ such that $L_{G'}=(1,0,1,{\mathbf{0}}_\ell)^*\mu(L_G)(0,-1,0,{\mathbf{0}}_\ell)^*$ is defining the set $\{(x,y)\in{{{\mathbb{N}}}}\times{{{\mathbb{N}}}}\mid y\leq f(x)\}$. Let $(x,y)$ in that set. From (\[eq-co\]), there exists a word $\sigma\in L_G$, $x'\in{{{\mathbb{N}}}}$ and ${\mathbf{e}}\in{{{\mathbb{N}}}}^\ell$ such that $(x,0,{\mathbf{0}}_\ell)\xrightarrow{\sigma}(x',f(x),{\mathbf{e}})$. The word $\sigma'=(1,0,1,{\mathbf{0}}_\ell)^x\mu(\sigma)(0,-1,0,{\mathbf{0}})^{f(x)-y}$ shows that $(0,0,0,{\mathbf{0}}_\ell)\xrightarrow{\sigma'}(x,y,x',{\mathbf{e}})$. As $\sigma'\in L_{G'}$, we get $(0,0,0,{\mathbf{0}}_\ell)\xrightarrow{G'}(x,y,x',{\mathbf{e}})$. Conversely, assume that $(0,0,0,{\mathbf{0}}_\ell)\xrightarrow{\sigma'}(x,y,x',{\mathbf{e}})$ for some $x,y,x'\in{{{\mathbb{N}}}}$, ${\mathbf{e}}\in{{{\mathbb{N}}}}^\ell$ and $\sigma'\in
L_{G'}$, and let us prove that $y\leq f(x)$. By definition of $G'$, there exists $n,m\in{{{\mathbb{N}}}}$ and a word $\sigma\in L_G$ such that $\sigma'=(1,0,1,{\mathbf{0}}_\ell)^n\mu(\sigma)(0,-1,0,{\mathbf{0}}_\ell)^m$. It follows that $(n,0,n,{\mathbf{0}}_\ell)\xrightarrow{\mu(\sigma)}(x,y+m,x',{\mathbf{e}})$. Since actions occurring in $\mu(\sigma)$ cannot modify the first counter, we get $n=x$. Moreover, $(x,0,{\mathbf{0}}_\ell)\xrightarrow{\sigma}(x',y+m,{\mathbf{e}})$. From (\[eq-sa\]), we derive $y+m\leq f(x)$. Hence $y\leq f(x)$. We have proved that $\{(x,y)\in{{{\mathbb{N}}}}\times{{{\mathbb{N}}}}\mid y\leq f(x)\}$ is GVAS-definable.
Conversely, let us assume that $f$ is non-decreasing and that $\{(x,y)\in{{{\mathbb{N}}}}\times{{{\mathbb{N}}}}\mid
y\leq f(x)\}$ is GVAS-definable. There exists a $d$-dimensional GVAS $G$ with $d=2+\ell$ such that: $$\{(x,y) \mid y\leq f(x)\}=\{(x,y)\mid \exists
{\mathbf{e}}\in{{{\mathbb{N}}}}^\ell : (0,0,{\mathbf{0}}_\ell)\xrightarrow{G}(x,y,{\mathbf{e}})\} \:.$$ Let us consider the mapping $\mu:{{{\mathbb{Z}}}}^d\rightarrow{{{\mathbb{Z}}}}^{d+1}$ defined by $\mu(a,b,{\mathbf{e}})=(-a,b, a,{\mathbf{e}})$ for every $a,b\in{{{\mathbb{Z}}}}$, and ${\mathbf{e}}\in{{{\mathbb{Z}}}}^\ell$. The mapping $\mu$ is extended as a word morphism. Let us show that a GVAS $G'$ such that $L_{G'}=\bigcup_{k\in{{{\mathbb{N}}}}}(1,0,0,{\mathbf{0}}_\ell)^k\mu(L_G)(-1,0,0,{\mathbf{0}}_\ell)^k$ is a weak GVAS computer for $f$. Let us first consider $x\in{{{\mathbb{N}}}}$. By definition of $G$, there exists a word $\sigma\in
L_G$ and ${\mathbf{e}}\in{{{\mathbb{N}}}}^\ell$ such that $(0,0,{\mathbf{0}}_\ell)\xrightarrow{\sigma}(x,f(x),{\mathbf{e}})$. Notice that for $k$ large enough, we have $(k+x,0,0,{\mathbf{0}}_\ell)
\xrightarrow{\mu(\sigma)}(k,f(x),x,{\mathbf{e}})$. The word $\sigma'= (1,0,0,{\mathbf{0}}_\ell)^k\mu(\sigma)(-1,0,0,{\mathbf{0}}_\ell)^k$ is such that $(x,0,0,{\mathbf{0}}_\ell)\xrightarrow{\sigma'}(0,f(x),x,{\mathbf{e}})$. Hence (\[eq-co\]) is satisfied by $G'$. Finally, let us assume that $(x,0,0,{\mathbf{0}}_\ell)\xrightarrow{\sigma'}(z,y,x',{\mathbf{e}})$ for a word $\sigma'\in L_{G'}$ and $x',y,z\in{{{\mathbb{N}}}}$ and ${\mathbf{e}}\in{{{\mathbb{N}}}}^\ell$. There exists $k\in{{{\mathbb{N}}}}$ and $\sigma\in L_G$ such that $\sigma'=(1,0,0,{\mathbf{0}}_\ell)^k\mu(\sigma)(-1,0,0,{\mathbf{0}}_\ell)^k$. It follows that $(x+k,0,0,{\mathbf{0}}_\ell)\xrightarrow{\mu(\sigma)}(z+k,y,x',{\mathbf{e}})$. By definition of $\mu$, since the effect of the sum of the first and third counters is zero, we get $x+k+0=z+k+x'$. Hence $x'\leq x$ and in particular $f(x')\leq f(x)$. Moreover, we have $(0,0,{\mathbf{0}}_\ell)\xrightarrow{\sigma}(x',y,{\mathbf{e}})$. By definition of $G$, we get $y\leq f(x')$. We have proved that $y\leq
f(x)$. Hence (\[eq-sa\]) is satisfied by $G'$. We have proved that $G'$ is a weak GVAS computer for $f$
By combining Lemma \[lem:WCGVAS\] and the decomposition of GVAS-definable sets given by Proposition \[prop-gvas-set-periodic\], we obtain two interesting, albeit negative, results on WC functions and GVAS-definable sets.
\[prop:sublin\] Let $f$ be an unbounded WC function. Then there exists a rational number $c>0$ and some $z\in{{{\mathbb{Z}}}}$ such that $f(n)\geq c n+z$ for every $n\in{{\mathbb{N}}}$.
Lemma \[lem:WCGVAS\] shows that the set ${\mathbf{X}}$ defined as $\{(n,m) \mid m\leq f(n)\}$ is GVAS-definable. Proposition \[prop-gvas-set-periodic\] shows that ${\mathbf{X}}$ can be decomposed into a finite union of sets of the form $(a,b)+{\mathbf{P}}$ where $(a,b)\in{{\mathbb{N}}}^2$ and ${\mathbf{P}}$ is a periodic subset of ${{\mathbb{N}}}^2$. Since $f$ is unbounded, there exists $(p,q)\in {\mathbf{P}}$ such that $q>0$. It follows $(a,b)+k(p,q)\in{\mathbf{X}}$ for every $k\in{{\mathbb{N}}}$. In particular $f(a+kp)\geq b+kq$ for every $k\in{{\mathbb{N}}}$. As $f(a)\in{{\mathbb{N}}}$ and $q>0$, we deduce that $p>0$. Let us consider $n\in{{{\mathbb{N}}}}$ such that $n\geq a$ and observe that there exists $k\in{{{\mathbb{N}}}}$ such that: $$k\leq \frac{n-a}{p}<k+1 \:.$$ It follows that $a+kp\leq n$ and in particular $f(a+kp)\leq f(n)$. Hence $f(n)\geq b+kq\geq b+(\frac{n-a}{p}-1)q$. Introducing $c=\frac{q}{p}$, we deduce that $f(n)- c n\geq b-c(a+p)$ for every $n\geq a$. We have proved the lemma with any $z\in{{{\mathbb{Z}}}}$ satisfying $z\leq f(n)-cn$ for every $0\leq n<a$ and $z\leq b-c(a+p)$.
\[prop:stabcomplement\] The complement of a GVAS-definable set is not always GVAS-definable.
Recall from Example \[ex:power-WC\] that the function $f:{{\mathbb{N}}}\to{{\mathbb{N}}}$ defined by $f(n)=2^{n}$ is WC. We derive from Lemma \[lem:WCGVAS\] that ${\mathbf{X}}{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}\{(n,m) \mid m\leq 2^n\}$ is GVAS-definable. Assume, by way of contradiction, that the complement ${\mathbf{Y}}{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}\{(n,m) \mid 2^n< m\}$ is GVAS-definable. From a GVAS defining ${\mathbf{Y}}$, we easily derive a GVAS defining ${\mathbf{Z}}{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}\{(n,m) \mid 2^m \leq n+1\}$, by swapping the first two counters and then decrementing the first counter by two at the end. It follows from Lemma \[lem:WCGVAS\] that the mapping $g:{{\mathbb{N}}}\to{{\mathbb{N}}}$ defined by $g(n)=\lfloor \log_2 (n+1)\rfloor$ is WC, contradicting Proposition \[prop:sublin\] since $g$ is unbounded and sublinear. Hence ${\mathbf{Y}}$, i.e., ${{{\mathbb{N}}}}^2\setminus{\mathbf{X}}$, cannot be GVAS-definable.
Hyper-Ackermannian GVAS {#sec-hypack}
=======================
In this section we construct GVASes that weakly compute functions from the Fast Growing Hierarchy. Our main result is the following.
\[thm-Fa-WC\] The Fast Growing functions $(F_\alpha)_{\alpha<\omega^\omega}$ are weakly computable (by GVASes).
Note that these are exactly the multiply-recursive functions $F_\alpha$. They include functions that are not primitive-recursive (the $F_\alpha$ for $\omega\leq\alpha<\omega^\omega$) and that are thus not weakly computable by VASSes (see [@jantzen80 section 2]). We do not know whether $F_{\omega^\omega}$ is weakly computable by a GVAS, or whether there exist WC functions that are not multiply-recursive.\
The rest of this section proves Theorem \[thm-Fa-WC\]. The detailed proof illustrates how the GVAS model makes it manageable to define complex constructions precisely, and to formally prove their correctness. By contrast, observe how in less abstract models e.g., the Timed-Arc Petri Nets of [@HSS-lics2012], only schematic constructions are given for weakly computing functions, and only an outline for a correctness proof can be provided.
We follow notation and definitions from [@schmitz-toct2016] and consider functions $F_\alpha:{{\mathbb{N}}}\to{{\mathbb{N}}}$ indexed by an ordinal $\alpha<\epsilon_0$ (though we shall only build GVASes for functions with $\alpha<\omega^\omega$). Any such ordinal can be written in Cantor normal form (CNF) $\alpha =
\omega^{\alpha_1}+\cdots+\omega^{\alpha_m}$ with $\alpha>\alpha_1\geq
\cdots\geq\alpha_m$. When $m=0$, $\alpha$ is $0$. When $\alpha_m=0$, $\alpha$ is a successor of the form $\beta+\omega^0$, i.e., $\beta+1$, and when $\alpha_m>0$, $\alpha$ is a limit ordinal. When $\alpha\neq
0$, we often decompose $\alpha$ under the form $\alpha =
\gamma+\omega^{\alpha_m}$ so that the smallest summand in $\alpha$’s CNF is exposed. CNFs are often written more concisely using coefficients, as in $\alpha=\omega^{\alpha_1}\cdot c_1 + \cdots +
\omega^{\alpha_m}\cdot c_m$, with now $\alpha>\alpha_1>
\cdots>\alpha_m$ and $\omega>c_1,\ldots,c_m>0$.
With each limit ordinal $\lambda<\epsilon_0$, one associates a fundamental sequence $(\lambda(n))_{n<\omega}$ such that $\lambda=\sup_n\lambda(n)$. These are defined inductively as follows. $$\begin{aligned}
\label{eq-L1}\tag{L1}
(\gamma + \omega^{\beta + 1})(n) &= \gamma +
\omega^{\beta}\cdot(n+1)
\:,
\\
\label{eq-LL}\tag{LL}
(\gamma + \omega^{\lambda})(n) &= \gamma
+ \omega^{\lambda(n)}
\:.\end{aligned}$$
For instance, Eq. gives $\omega(n)$, i.e., $\omega^1(n)
= \omega^0\cdot(n + 1)=n+1$ and $(\omega^{3}\cdot 6 + \omega^{2}\cdot
3)(n) = \omega^{3}\cdot 6 + \omega^{2}\cdot 2 + \omega\cdot(n+1)$. Similarly, Eq. gives $\omega^{\omega}(n) = \omega^{\omega(n)}
= \omega^{n+1}$. Note that the fundamental sequences satisfy $\lambda(0) < \cdots <
\lambda(n) < \lambda(n+1) < \cdots < \lambda$ for any limit ordinal $\lambda$ and index $n$.
We may now define our fast growing functions $F_\alpha:{{\mathbb{N}}}\to{{\mathbb{N}}}$ for $\alpha<\epsilon_0$ by induction on the $\alpha$ index. $$\begin{aligned}
\label{eq-F0}\tag{F0}
F_0(x) &= x+1
\:,\\[-1.5em]
\label{eq-F1}\tag{F1}
F_{\alpha + 1}(x) &= F_{\alpha}^{\omega(x)}(x) =
\overbrace{F_{\alpha}(\cdots (F_{\alpha}}^{x+1\text{
times}}(x) )\cdots )
\:,\\
\label{eq-FL}\tag{FL}
F_{\lambda}(x) &= F_{\lambda(x)}(x)
\:.\end{aligned}$$ As shown —e.g., in [@schmitz-toct2016]— these functions are *strictly expansive* and *monotonic*, i.e., for all ordinals $\alpha<\epsilon_0$ and all $n,n'\in{{\mathbb{N}}}$: $$\begin{gathered}
\label{eq-F-exp}\tag{FX}
n < F_\alpha(n)
\:,
\\
\label{eq-F-mono}\tag{FM}
n\leq n' \implies F_\alpha (n)\leq F_\alpha (n')
\:.\end{gathered}$$
Given two ordinals in Cantor normal form $\alpha=\omega^{\beta_1}+ \cdots
+ \omega^{\beta_m}$ and $\alpha'=\omega^{\beta'_1} + \cdots + \omega^{\beta'_n}$, we denote by $\alpha \oplus \alpha'$ their *natural sum* $\sum_{k=1}^{m+n} \omega^{\gamma_k}$, where $\gamma_1 \geq \ldots \geq \gamma_{m+n}$ is a reordering of $\beta_1, \ldots, \beta_m, \beta'_1, \ldots, \beta'_n$. The $F_\alpha$ functions are not monotonic in the ordinal index, i.e., $\alpha\leq\alpha'$ does not always entail $F_\alpha(n)\leq F_{\alpha'}(n)$, see [@schmitz-toct2016 section A.2]. However, our construction relies on similar monotonicity properties, albeit for special cases of $\alpha$ and $\alpha'$, that we now state.
\[lem-Fmono-alpha\] For any ordinals $\alpha,\alpha'<\epsilon_0$ and any $n\in{{\mathbb{N}}}$, $F_{\alpha}(n) \leq F_{\alpha\oplus\alpha'}(n)$.
\[lem-Fmono-alphaSeq\] For any ordinal $\alpha<\epsilon_0$ and limit ordinal $\lambda<\omega^\omega$, for any $m, n\in{{\mathbb{N}}}$, if $m \leq n$ then $F_{\alpha\oplus\lambda(m)}(n)
\leq
F_{\alpha\oplus\lambda}(n)$.
For these two results, detailed proofs are given in the appendix. We note that Lemma \[lem-Fmono-alpha\] is a rewording of Lemma 2.2a from [@CS-lics08], however that paper uses a different definition for the fundamental sequences $(\lambda(n))_{n\in{{\mathbb{N}}}}$, resulting in slightly different $F_\alpha$ functions, hence the need of an independent proof. Similarly, Lemma \[lem-Fmono-alphaSeq\] is a generalization of Lemma VI.5 from [@leroux2014], using different notation and allowing a simpler proof.\
We now define weak GVAS computers for the $F_{\alpha}$ functions such that $\alpha<\omega^\omega$. Our construction is in two steps: we first pick an arbitrary exponent $d\in{{\mathbb{N}}}$ and define $G_d$, a GVAS with a structure suitable for correctness proofs. We then obtain a weak GVAS computer for $F_\alpha$ by slightly modifying $G_d$, provided $\alpha<\omega^d$. The whole construction is an adaptation into the GVAS framework of the pushdown VAS from [@leroux2014].
The dimension of $G_d$ is $d+2$ and we use $d+2$ counters named $r, \overline{r}, \kappa_0, \ldots, \kappa_{d-1}$, in this order. The set of actions ${\mathbf{A}}\subseteq {{{\mathbb{Z}}}}^{d+2}$ consists of all vectors ${\mathbf{d}}_{x}$ and ${{{\mathbf{i}}}}_{x}$ where $x$ is one of the $d+2$ counters: formally ${\mathbf{d}}_x$ is the vector that decrements $x$, while ${{{\mathbf{i}}}}_x{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}-{\mathbf{d}}_x$ increments it. For instance, ${\mathbf{d}}_{\kappa_0} = (0, 0, -1, {{{\mathbf{0}}}}_{d-1})$ and ${{{\mathbf{i}}}}_{r} = (1, 0, {{{\mathbf{0}}}}_{d})$. The set of non-terminals of $G_d$ is $V=\{{{{\mathtt{F}}}}, {{{\mathtt{Rec}}}}, {{{\mathtt{Pop}}}}, {{{\mathtt{Lim}}}}_1, \ldots, {{{\mathtt{Lim}}}}_{d-1}\}$. The start symbol is ${{{\mathtt{F}}}}$. The other non-terminals are used for intermediate steps (see the rules below).
The first two counters, $r$ and $\overline{r}$, are used to manipulate the arguments of the functions being computed. The other $d$ counters are used as a data structure representing an ordinal $\alpha<\omega^d$. Formally, with any $d$-tuple ${\langle {c_0,\ldots,c_{d-1}} \rangle}$ of natural numbers, we associate the ordinal $\alpha = \omega^{d-1} \cdot c_{d-1} + \cdots + \omega^{0}
\cdot c_0$. We will follow the convention of writing the contents of the counters of our GVAS in the form ${\langle {n, m, \alpha} \rangle}$, where $n$ and $m$ are the value of $r$ and $\overline{r}$, respectively, and where $\alpha$ is the ordinal associated with the values in $\kappa_0,\ldots,\kappa_{d-1}$.
The rules of $G_d$ are given below. The rules involving the ${{{\mathtt{Lim}}}}_{i}$ non-terminals are present for every $i \in \{1,
\ldots, d-1\}$.
[2]{} & \_[r]{} , \[ruleF0\]\
& \_[\_0]{} \_[\_0]{} , \[ruleF+1\]\
& \_[\_i]{} \_[\_[i-1]{}]{} \_i \_[\_[i-1]{}]{}\_[\_[i]{}]{} , \[ruleFlim\]\
&, \[ruleRec0\]\
& \_r\_ , \[ruleRec1\]\
& , \[ruleRest0\]\
& \_r \_ , \[ruleRest1\]\
\_i & , \[ruleLim0\]\
\_i & \_r\_ \_[\_[i-1]{}]{} \_[i]{} \_[\_[i-1]{}]{} . \[ruleLim1\]
Our first goal is to prove that $G_d$ has computations of the form ${\langle {n,0,\alpha} \rangle} \xrightarrow{{{{\mathtt{F}}}}} {\langle {F_{\alpha}(n),0,\alpha} \rangle}$, for any $n\in{{\mathbb{N}}}$ and $\alpha<\omega^d$. We start with a lemma exposing some specific sentential forms that can be derived from ${{{\mathtt{F}}}}$ and ${{{\mathtt{Pop}}}}$. As will be clear from the proof of Lemma \[lem:completenessF\], these derivations (namely , and ) correspond to our inductive definition of the fast growing functions $F_\alpha$ (namely , and ).
\[lem:completeness-derivations\] For every $n \in {{\mathbb{N}}}$ and $0<i<d$, $G_d$ admits the following derivations: $$\begin{aligned}
\label{derivF0}\tag{D0}
{{{\mathtt{F}}}}& {\xLongrightarrow{*}}{{{\mathbf{i}}}}_{r}
\:,
\\
\label{derivF+1}\tag{D1}
{{{\mathtt{F}}}}& {\xLongrightarrow{*}}{\mathbf{d}}_{\kappa_0} \ ({\mathbf{d}}_r {{{\mathbf{i}}}}_{\overline{r}})^n \ {{{\mathtt{Pop}}}}\ ({{{\mathtt{F}}}})^{n+1} \ {{{\mathbf{i}}}}_{\kappa_0}
\:,
\\
\label{derivFlim}\tag{DL}
{{{\mathtt{F}}}}& {\xLongrightarrow{*}}{\mathbf{d}}_{\kappa_i} {{{\mathbf{i}}}}_{\kappa_{i-1}} \ ({\mathbf{d}}_r {{{\mathbf{i}}}}_{\overline{r}} {{{\mathbf{i}}}}_{\kappa_{i-1}})^n \ {{{\mathtt{Pop}}}}\ {{{\mathtt{F}}}}\ ({\mathbf{d}}_{\kappa_{i-1}})^{n+1} \ {{{\mathbf{i}}}}_{\kappa_{i}}
\:,
\\
\label{derivPop}\tag{DP}
{{{\mathtt{Pop}}}}& {\xLongrightarrow{*}}({{{\mathbf{i}}}}_r {\mathbf{d}}_{\overline{r}})^n
\:.
\end{aligned}$$
Derivation is an immediate consequence of rule and derivation similarly follows from rules and . To prove derivation , we use $$\begin{aligned}
{{{\mathtt{F}}}}\xLongrightarrow{\eqref{ruleF+1}}
{\mathbf{d}}_{\kappa_0} \ {{{\mathtt{Rec}}}}\ {{{\mathtt{F}}}}\ {{{\mathbf{i}}}}_{\kappa_0}
\xLongrightarrow{\eqref{ruleRec1}} \cdots
& \xLongrightarrow{\eqref{ruleRec1}}
{\mathbf{d}}_{\kappa_0} \ ({\mathbf{d}}_r{{{\mathbf{i}}}}_{\overline{r}})^n \ {{{\mathtt{Rec}}}}\ ({{{\mathtt{F}}}})^n \ {{{\mathtt{F}}}}\ {{{\mathbf{i}}}}_{\kappa_0}
\\
& \xLongrightarrow{\eqref{ruleRec0}}
{\mathbf{d}}_{\kappa_0} \ ({\mathbf{d}}_r{{{\mathbf{i}}}}_{\overline{r}})^n \ {{{\mathtt{Pop}}}}\ ({{{\mathtt{F}}}})^{n+1} \ {{{\mathbf{i}}}}_{\kappa_0}
\:.\end{aligned}$$ Finally, derivation is obtained with $$\begin{aligned}
{{{\mathtt{F}}}}&
\xLongrightarrow{\eqref{ruleFlim}}
{\mathbf{d}}_{\kappa_i} {{{\mathbf{i}}}}_{\kappa_{i-1}} \ {{{\mathtt{Lim}}}}_i
\ {\mathbf{d}}_{\kappa_{i-1}}{{{\mathbf{i}}}}_{\kappa_{i}}
\\
&
\xLongrightarrow{\eqref{ruleLim1}}\cdots\xLongrightarrow{\eqref{ruleLim1}}
{\mathbf{d}}_{\kappa_i} {{{\mathbf{i}}}}_{\kappa_{i-1}} \
({\mathbf{d}}_r{{{\mathbf{i}}}}_{\overline{r}} {{{\mathbf{i}}}}_{\kappa_{i-1}})^n \ {{{\mathtt{Lim}}}}_i \
({\mathbf{d}}_{\kappa_{i-1}})^n \
{\mathbf{d}}_{\kappa_{i-1}}{{{\mathbf{i}}}}_{\kappa_{i}}
\\
&
\xLongrightarrow{\eqref{ruleLim0}}
{\mathbf{d}}_{\kappa_i} {{{\mathbf{i}}}}_{\kappa_{i-1}} \
({\mathbf{d}}_r{{{\mathbf{i}}}}_{\overline{r}} {{{\mathbf{i}}}}_{\kappa_{i-1}})^n \
{{{\mathtt{Pop}}}}\ {{{\mathtt{F}}}}\
({\mathbf{d}}_{\kappa_{i-1}})^{n+1} \
{{{\mathbf{i}}}}_{\kappa_{i}}
\:.
\qedhere\end{aligned}$$
\[lem:completenessF\] For every ordinal $\alpha < \omega^{d}$ and every $n \in {{\mathbb{N}}}$, $G_d$ has a computation ${\langle {n,0,\alpha} \rangle} \xrightarrow{{{{\mathtt{F}}}}}
{\langle {F_{\alpha}(n),0,\alpha} \rangle}$.
We first observe that $G_d$ has a computation ${\langle {0,n,\alpha} \rangle} \xrightarrow{{{{\mathtt{Pop}}}}} {\langle {n,0,\alpha} \rangle}$ for every ordinal $\alpha < \omega^{d}$ and every $n \in {{\mathbb{N}}}$. This computation exists because $G_d$ admits derivation , i.e., ${{{\mathtt{Pop}}}}{\xLongrightarrow{*}}({{{\mathbf{i}}}}_r {\mathbf{d}}_{\overline{r}})^n$. We now prove the lemma by induction on $\alpha$.
For the base case $\alpha = 0$, we use derivation , i.e., ${{{\mathtt{F}}}}{\xLongrightarrow{*}}{{{\mathbf{i}}}}_{r}$, yielding the following computation: $${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle 0\\[-.5em] \scriptstyle 0\end{array}\Bigr)} \xrightarrow{{{{\mathtt{F}}}}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle n+1\\[-.5em] \scriptstyle 0\\[-.5em] \scriptstyle 0\end{array}\Bigr)} = {\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_0(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle 0\end{array}\Bigr)}
\:.$$
In the case of a successor ordinal $\alpha=\beta + 1$, we use derivation , i.e., ${{{\mathtt{F}}}}{\xLongrightarrow{*}}{\mathbf{d}}_{\kappa_0} \ ({\mathbf{d}}_r {{{\mathbf{i}}}}_{\overline{r}})^n \ {{{\mathtt{Pop}}}}\ ({{{\mathtt{F}}}})^{n+1} \ {{{\mathbf{i}}}}_{\kappa_0}$, leading to the following computation (recall that $F_{\alpha}(n) = F_{\beta}^{n+1}(n)$ by Eq. ): $${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle 0\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)}
\xrightarrow{{\mathbf{d}}_{\kappa_0}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle 0\\[-.5em] \scriptstyle \beta\end{array}\Bigr)}
\xrightarrow{({\mathbf{d}}_r {{{\mathbf{i}}}}_{\overline{r}})^n}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle 0\\[-.5em] \scriptstyle n\\[-.5em] \scriptstyle \beta\end{array}\Bigr)}
\xrightarrow{{{{\mathtt{Pop}}}}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle 0\\[-.5em] \scriptstyle \beta\end{array}\Bigr)}
\xrightarrow[\text{ind.\ hyp.}]{{{{\mathtt{F}}}}^{n+1}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_{\beta}^{n+1}(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle \beta\end{array}\Bigr)}
\xrightarrow{{{{\mathbf{i}}}}_{\kappa_0}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_{\beta}^{n+1}(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle \beta+1\end{array}\Bigr)}
={\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_{\alpha}(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)}
\:.$$
Finally, in the case of a limit ordinal $\alpha = \lambda$, say of the form $\lambda = \gamma + \omega^{i}$ where $0<i<d$, we use derivation , i.e., ${{{\mathtt{F}}}}{\xLongrightarrow{*}}{\mathbf{d}}_{\kappa_i} {{{\mathbf{i}}}}_{\kappa_{i-1}} \ ({\mathbf{d}}_r {{{\mathbf{i}}}}_{\overline{r}} {{{\mathbf{i}}}}_{\kappa_{i-1}})^n \ {{{\mathtt{Pop}}}}\ {{{\mathtt{F}}}}\ ({\mathbf{d}}_{\kappa_{i-1}})^{n+1} \ {{{\mathbf{i}}}}_{\kappa_{i}}$. Before inspecting the computation below, note that if $\lambda=\gamma+\omega^i$ is represented by the values in $\kappa_0, \ldots, \kappa_{d-1}$, then one obtains a representation for $\gamma$ by decrementing $\kappa_i$. Then, by incrementing $(n+1)$ times $\kappa_{i-1}$, one obtains $\gamma+\omega^{i-1}\cdot(n+1)$ which is $\lambda(n)$ by . This leads to the following computation (recall that $F_{\lambda}(n) = F_{\lambda(n)}(n)$ by Eq. ): $$\begin{aligned}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle 0\\[-.5em] \scriptstyle \lambda\end{array}\Bigr)}
\xrightarrow{{\mathbf{d}}_{\kappa_{i}}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle 0\\[-.5em] \scriptstyle \gamma\end{array}\Bigr)}
\xrightarrow{{{{\mathbf{i}}}}_{\kappa_{i-1}} ({\mathbf{d}}_r {{{\mathbf{i}}}}_{\overline{r}} {{{\mathbf{i}}}}_{\kappa_{i-1}})^n}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle 0\\[-.5em] \scriptstyle n\\[-.5em] \scriptstyle \lambda(n)\end{array}\Bigr)}
\xrightarrow{{{{\mathtt{Pop}}}}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle 0\\[-.5em] \scriptstyle \lambda(n)\end{array}\Bigr)}
\xrightarrow[\text{ind.\ hyp.}]{{{{\mathtt{F}}}}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_{\lambda(n)}(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle \lambda(n)\end{array}\Bigr)}
\:,\text{ and}
\\
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_{\lambda(n)}(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle \lambda(n)\end{array}\Bigr)}
=
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_{\lambda}(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle \lambda(n)\end{array}\Bigr)}
\xrightarrow{({\mathbf{d}}_{\kappa_{i-1}})^{n+1}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_{\lambda}(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle \gamma\end{array}\Bigr)}
\xrightarrow{{{{\mathbf{i}}}}_{\kappa_{i}}}
{\Bigl(\begin{array}{@{}c@{}}\scriptstyle F_{\lambda}(n)\\[-.4em] \scriptstyle 0\\[-.5em] \scriptstyle \lambda\end{array}\Bigr)}
\:.
\end{aligned}$$
In all three cases, $G_d$ has a computation ${\langle {n,0,\alpha} \rangle} \xrightarrow{{{{\mathtt{F}}}}}
{\langle {F_{\alpha}(n),0,\alpha} \rangle}$.
We showed in Lemma \[lem:completenessF\] that there are computations of $G_d$ that end in $F_{\alpha}(n)$. This corresponds to the completeness of weak computers. We will now show the safety part, i.e., that no successful computation of $G_d$ may reach a value greater than $F_{\alpha}(n)$.
\[lem:safetyGd\] For all $n,n',m,m'\in{{\mathbb{N}}}$, $\alpha,\alpha'<\omega^d$, and $0<i<d$, the following hold: $$\begin{aligned}
\label{safe-F}
{\langle {n,m,\alpha} \rangle}\xrightarrow{{{{\mathtt{F}}}}}{\langle {n',m',\alpha'} \rangle}
&\implies \alpha'=\alpha \land n'+m'\leq F_{\alpha}(n+m)
\:,
\\
\label{safe-Rec}
{\langle {n,m,\alpha} \rangle}\xrightarrow{{{{\mathtt{Rec}}}}}{\langle {n',m',\alpha'} \rangle}
&\implies \alpha'=\alpha \land n'+m'\leq F^n_{\alpha}(n+m)
\:,
\\
\label{safe-Pop}
{\langle {n,m,\alpha} \rangle} \xrightarrow{{{{\mathtt{Pop}}}}} {\langle {n',m',\alpha'} \rangle}
&\implies \alpha' = \alpha \land n'+m' = n+m
\:,
\\
\label{safe-Lim}
{\langle {n,m,\alpha} \rangle}\xrightarrow{{{{\mathtt{Lim}}}}_i}{\langle {n',m',\alpha'} \rangle}
&\implies \alpha'=\alpha \land n'+m'\leq F_{\alpha\oplus(\omega^{i-1}\cdot n)}(n+m)
\:.\end{aligned}$$
By structural induction on the flow trees witnessing the transitions.
Top rule is ${{{\mathtt{F}}}}\rightarrow {{{\mathbf{i}}}}_{r}$
: Then the flow tree has the following shape (in this and following illustrations, *we only display the top node of each immediate subtree* of the flow tree under consideration):
\(a) [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{F}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{}; (a1) \[below=of a\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathbf{i}}}}_{r}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{};
\(a) to (a1);
Using action ${{{\mathbf{i}}}}_{r}$ in the subtree implies $n'=n+1$, and also $\alpha'=\alpha$ and $m'=m$. With , we deduce $n'+m'=n+m+1\leq F_\alpha(n+m)$ as required by .
Top rule is ${{{\mathtt{F}}}}\rightarrow {\mathbf{d}}_{\kappa_0}
\ {{{\mathtt{Rec}}}}\ {{{\mathtt{F}}}}\ {{{\mathbf{i}}}}_{\kappa_0}$
: We note that the first action, ${\mathbf{d}}_{\kappa_0}$, can only be fired if $\alpha$ is a successor ordinal $\beta+1$. Then decrementing $\kappa_0$ transforms $\alpha$ into $\beta$, and the flow tree has the following form.
\(b) [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{F}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \beta'+1\end{array}\Bigr)}$]{};
(b1) \[below=of b, xshift=-11em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{\mathbf{d}}_{\kappa_0}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \beta\end{array}\Bigr)}$]{}; (b21) \[below=of b, xshift=-3.5em, yshift=0em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \beta\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Rec}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_{1}\\[-.5em] \scriptstyle m_{1}\\[-.5em] \scriptstyle \beta_{1}\end{array}\Bigr)}$]{}; (b22) \[below=of b, xshift=3.5em, yshift=0em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_{1}\\[-.5em] \scriptstyle m_{1}\\[-.5em] \scriptstyle \beta_{1}\end{array}\Bigr)} \xrightarrow{{{{\mathtt{F}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \beta'\end{array}\Bigr)}$]{}; (b3) \[below=of b, xshift=11em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \beta'\end{array}\Bigr)} \xrightarrow{{{{\mathbf{i}}}}_{\kappa_0}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \beta'+1\end{array}\Bigr)}$]{};
\(b) to (b1); (b) to (b3); (b) to (b21); (b) to (b22);
Invoking the induction hypothesis on the second and third subtrees yields
[2]{} \_1&=, & n\_1+m\_1&F\^n\_(n+m),\
’&=\_1, & n’+m’&F\_[\_1]{}(n\_1+m\_1).
Combining these results, we obtain $\beta'+1=\alpha$ as needed, and $$\begin{aligned}
n'+m' \leq F_{\beta_1}(n_1+m_1)
&\leq F_{\beta_1}(F^n_{\beta}(n+m))
\tag*{by \eqref{eq-F-mono}}
\\
&= F^{n+1}_{\beta}(n+m)
\\
&\leq F^{n+m+1}_{\beta}(n+m)
\tag*{by \eqref{eq-F-exp}}
\\
&= F_{\beta+1}(n+m)
= F_\alpha(n+m) \:.
\tag*{by \eqref{eq-F1}}\end{aligned}$$ Finally, $n'+m'\leq F_\alpha(n+m)$ as required by .
Top rule is ${{{\mathtt{F}}}}\rightarrow
{\mathbf{d}}_{\kappa_{i}} {{{\mathbf{i}}}}_{\kappa_{i-1}} \ {{{\mathtt{Lim}}}}_{i}
\ {\mathbf{d}}_{\kappa_{i-1}} {{{\mathbf{i}}}}_{\kappa_{i}}$
: The flow tree has the following form.
\(a) [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{F}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{}; (a1) \[below=of a, xshift=-10em, yshift=1em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{\mathbf{d}}_{\kappa_{i}}
{{{\mathbf{i}}}}_{\kappa_{i-1}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_1\\[-.5em] \scriptstyle m_1\\[-.5em] \scriptstyle \alpha_1\end{array}\Bigr)}$]{};
(a2) \[below=of a, xshift=0em, yshift=1em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_1\\[-.5em] \scriptstyle m_1\\[-.5em] \scriptstyle \alpha_1\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Lim}}}}_i} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_2\\[-.5em] \scriptstyle m_2\\[-.5em] \scriptstyle \alpha_2\end{array}\Bigr)}$]{};
(a3) \[below=of a, xshift=10em, yshift=1em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_2\\[-.5em] \scriptstyle m_2\\[-.5em] \scriptstyle \alpha_2\end{array}\Bigr)} \xrightarrow{{\mathbf{d}}_{\kappa_{i-1}}{{{\mathbf{i}}}}_{\kappa_{i}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{};
\(a) to (a1); (a) to (a2); (a) to (a3);
Firing the first two actions requires decrementing $\kappa_i$, hence $\alpha$ is some $\alpha_0\oplus\omega^i$. After these actions, one has $n_1=n$, $m_1=m$ and $\alpha_1=\alpha_0\oplus\omega^{i-1}$. Similarly, the last two actions require decrementing $\kappa_{i-1}$, hence $\alpha_2$ is some $\alpha_3\oplus\omega^{i-1}$ and one has $n'=n_2$, $m'=m_2$ and $\alpha'=\alpha_3\oplus\omega^i$. One obtains $\alpha_1=\alpha_2$, hence $\alpha'=\alpha$, with the induction hypothesis, as well as $$\begin{aligned}
n'+m'=n_2+m_2
&\leq
F_{\alpha_1\oplus(\omega^{i-1}\cdot n_1)}(n_1+m_1)
\tag*{by ind.\ hyp.}
\\
&=
F_{(\alpha_0\oplus\omega^{i-1})\oplus(\omega^{i-1}\cdot n)}(n+m)
\\
&=
F_{\alpha_0\oplus(\omega^{i}(n))}(n+m)
\\
&\leq F_{\alpha_{0}\oplus\omega^{i}}(n+m)
\tag*{by Lemma~\ref{lem-Fmono-alphaSeq}}
\\
&= F_{\alpha}(n+m)
\:,\end{aligned}$$ as required by .
Top rule is ${{{\mathtt{Rec}}}}\rightarrow {{{\mathtt{Pop}}}}$
: Then the flow tree has the following form.
\(b) [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Rec}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{}; (b1) \[below=of b, yshift=0em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Pop}}}}} {{\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}}$]{};
\(b) to (b1);
The induction hypothesis gives $\alpha'=\alpha$ and $n'+m'=n+m$. We deduce $n'+m'\leq F_\alpha^n(n+m)$, as required by , by invoking .
Top rule is ${{{\mathtt{Rec}}}}\rightarrow
{\mathbf{d}}_r{{{\mathbf{i}}}}_{\overline{r}}\ {{{\mathtt{Rec}}}}\ {{{\mathtt{F}}}}$
: Then $n>0$ and the flow tree has the following form.
\(b) [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Rec}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \beta\end{array}\Bigr)}$]{};
(b1) \[below=of b, xshift=-9em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{\mathbf{d}}_{r} {{{\mathbf{i}}}}_{\overline{r}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n-1\\[-.5em] \scriptstyle m+1\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)}$]{}; (b2) \[below=of b, yshift=-0em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n-1\\[-.5em] \scriptstyle m+1\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Rec}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_{1}\\[-.5em] \scriptstyle m_{1}\\[-.5em] \scriptstyle \beta_{1}\end{array}\Bigr)}$]{}; (b3) \[below=of b, xshift=9em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_{1}\\[-.5em] \scriptstyle m_{1}\\[-.5em] \scriptstyle \beta_{1}\end{array}\Bigr)} \xrightarrow{{{{\mathtt{F}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \beta\end{array}\Bigr)}$]{};
\(b) to (b1); (b) to (b2); (b) to (b3);
Here we can use the induction hypothesis on the second subtree, yielding
[2]{} \_1&=, & n\_1+m\_1&F\_\^[n-1]{}(n-1+m+1)=F\_\^[n-1]{}(n+m) ,
and on the third subtree, yielding
[2]{} &=\_1, & n’+m’&F\_[\_1]{}(n\_1+m\_1) = F\_(n\_1+m\_1) .
Combining these and invoking , yields $\beta=\alpha$ and $n'+m'\leq F_\alpha(n_1+m_1)\leq F_\alpha(F_\alpha^{n-1}(n+m))=
F_\alpha^n(n+m)$ as required by .
Top rule is ${{{\mathtt{Pop}}}}\rightarrow\varepsilon$
: Then the flow tree $(n,m,\alpha)\xrightarrow{{{{\mathtt{Pop}}}}}(n',m',\alpha')$ is a leaf, entailing $\alpha'=\alpha$ and $n'+m'=n+m$ as required by .
Top rule is ${{{\mathtt{Pop}}}}\rightarrow
{{{\mathbf{i}}}}_r {\mathbf{d}}_{\overline{r}} \ {{{\mathtt{Pop}}}}$
: Then the flow tree has the form.
\(b) [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Pop}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{}; (b1) \[below=of b, xshift=-5em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathbf{i}}}}_{r} {\mathbf{d}}_{\overline{r}} }{\Bigl(\begin{array}{@{}c@{}}\scriptstyle n+1\\[-.5em] \scriptstyle m-1\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)}$]{}; (b2) \[below=of b, xshift=5em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n+1\\[-.5em] \scriptstyle m-1\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Pop}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{};
\(b) to (b1); (b) to (b2);
The induction hypothesis on the second subtree gives $\alpha'=\alpha$ and $n'+m'=n+1+m-1=n+m$ as required by .
Top rule is ${{{\mathtt{Lim}}}}_{i}\rightarrow {{{\mathtt{Pop}}}}\ {{{\mathtt{F}}}}$
: The flow tree has the following form.
\(a) [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Lim}}}}_{i}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{};
(a1) \[below=of a, xshift=-7em, yshift=1em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Pop}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_{1}\\[-.5em] \scriptstyle m_{1}\\[-.5em] \scriptstyle \alpha_{1}\end{array}\Bigr)}$]{}; (a2) \[below=of a, xshift=7em, yshift=1em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_{1}\\[-.5em] \scriptstyle m_{1}\\[-.5em] \scriptstyle \alpha_{1}\end{array}\Bigr)} \xrightarrow{{{{\mathtt{F}}}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{};
\(a) to (a1); (a) to (a2);
On these subtrees, the induction hypothesis yields $\alpha'=\alpha_{1}
= \alpha$ and $n_{1} + m_{1} = n + m$. Furthermore we have $$\begin{aligned}
n' + m' &\leq F_{\alpha_1}(n_{1} + m_{1})
\tag*{by ind.\ hyp.}
\\
& \leq F_{\alpha\oplus(\omega^{i-1}\cdot n)}(n+m)
\:,
\tag*{by Lemma~\ref{lem-Fmono-alpha}}\end{aligned}$$ as required by .
Top rule is ${{{\mathtt{Lim}}}}_{i} \rightarrow {\mathbf{d}}_{r}
{{{\mathbf{i}}}}_{\overline{r}} {{{\mathbf{i}}}}_{\kappa_{i-1}} \ {{{\mathtt{Lim}}}}_{i} \ {\mathbf{d}}_{\kappa_{i-1}}$
: The flow tree has the following form.
\(a) [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Lim}}}}_{i}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{}; (a1) \[below=of a, xshift=-10em, yshift=1em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n\\[-.5em] \scriptstyle m\\[-.5em] \scriptstyle \alpha\end{array}\Bigr)} \xrightarrow{{\mathbf{d}}_{r} {{{\mathbf{i}}}}_{\overline{r}} {{{\mathbf{i}}}}_{\kappa_{i-1}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_1\\[-.5em] \scriptstyle m_1\\[-.5em] \scriptstyle \alpha_1\end{array}\Bigr)}$]{}; (a2) \[below=of a, xshift=0em, yshift=1em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_1\\[-.5em] \scriptstyle m_1\\[-.5em] \scriptstyle \alpha_1\end{array}\Bigr)} \xrightarrow{{{{\mathtt{Lim}}}}_{i}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_2\\[-.5em] \scriptstyle m_2\\[-.5em] \scriptstyle \alpha_2\end{array}\Bigr)}$]{}; (a3) \[below=of a, xshift=10em, yshift=1em\] [${\Bigl(\begin{array}{@{}c@{}}\scriptstyle n_2\\[-.5em] \scriptstyle m_2\\[-.5em] \scriptstyle \alpha_2\end{array}\Bigr)} \xrightarrow{{\mathbf{d}}_{\kappa_{i-1}}} {\Bigl(\begin{array}{@{}c@{}}\scriptstyle n'\\[-.5em] \scriptstyle m'\\[-.5em] \scriptstyle \alpha'\end{array}\Bigr)}$]{};
\(a) to (a1); (a) to (a2); (a) to (a3);
With its three actions, the first subtree implies
[3]{} n\_1&=n-1 , & m\_1&=m+1 , &\_1&=\^[i-1]{} .
Similarly, the last subtree yields
[3]{} n’&=n\_2 , & m’&=m\_2 , &\_2&=’\^[i-1]{}.
With the second subtree, the induction hypothesis yields $\alpha_2=\alpha_1$ and $n_2+m_2\leq
F_{\alpha_1\oplus(\omega^{i-1}\cdot n_1)}(n_1+m_1)$. Combining these gives $\alpha'=\alpha$ and $n'+m'=n_2+m_2\leq
F_{\alpha_1\oplus(\omega^{i-1}\cdot n_1)}(n_1+m_1) =
F_{(\alpha\oplus\omega^{i-1})\oplus(\omega^{i-1}\cdot (n-1))}(n+m) =
F_{\alpha\oplus(\omega^{i-1}\cdot n)}(n+m)$ as required by .
We have proved that the counters in the GVAS $G_{d}$ will never have counter values more than $F_{\alpha}(n)$, where $\alpha$ and $n$ are represented by the initial counter values. We will now use this GVAS to design a Weak GVAS computer for the function $F_{\alpha}$, where $\alpha < \omega^{\omega}$. Given $\alpha =
\omega^{d}$, the weak GVAS computer will have $d+4$ counters. The first two counters are for the input and output, as given in Definition \[def-wpn\]. The remaining $d+2$ counters will work exactly as in $G_{d}$. The weak computer has to transfer the input value $n$ to the appropriate counters of $G_{d}$ and initialize the other counters of $G_{d}$ to correctly represent $\alpha$. After $G_{d}$ finishes, the weak computer has to transfer the result to the output counter. The details are routine and skipped. Now, for any ordinal $\alpha<\omega^d$, we may extend $G_d$ and obtain a GVAS $G_{F_\alpha}$ that weakly computes $F_\alpha$. This new GVAS inherits the counters, actions, non-terminals and rules of $G_d$. It furthermore includes two additional non-terminals, ${{{\mathtt{S}}}}$ and ${{{\mathtt{Pop}}}}'$, and associated rules. The start symbol will be ${{{\mathtt{S}}}}$ and, if $\alpha$’s CNF is $\sum_{i=d-1}^{0}\omega^i\cdot c_i$, the extra rules are:
[1]{} & \_[\_0]{}\^[c\_0]{} \_[\_1]{}\^[c\_1]{} \_[\_[d-1]{}]{}\^[c\_[d-1]{}]{} ’ , \[ruleS\]\
’ & , \[ruleTrans0\]\
’ & \_[r]{}\_ ’ . \[ruleTrans1\]
It is clear that, since there are no new rules for the non-terminals inherited from $G_d$, the properties stated in Lemmas \[lem:completeness-derivations\] to \[lem:safetyGd\] hold for $G_{F_\alpha}$ as they hold for $G_d$. Note also that ${{{\mathtt{Pop}}}}'$ behaves as ${{{\mathtt{Pop}}}}$ but exchanging the roles of $r$ and $\overline{r}$.
\[lem-GFa-WC\] $G_{F_\alpha}$ weakly computes $F_{\alpha}$.
We start with the completeness part of Definition \[def-wpn\]. For this it is needed to show that, for any $n\in{{\mathbb{N}}}$, $G_{F_\alpha}$ has a computation of the form $(n,0,{{{\mathbf{0}}}}_d)\xrightarrow{{{{\mathtt{S}}}}}
(n',F_{\alpha}(n),{\mathbf{e}})$. For this we use the rule and exhibit the following computation: $$(n,0,{{{\mathbf{0}}}}_d)
\xrightarrow{{{{\mathbf{i}}}}_{\kappa_0}^{c_0}
{{{\mathbf{i}}}}_{\kappa_1}^{c_1}
\cdots
{{{\mathbf{i}}}}_{\kappa_{d-1}}^{c_{d-1}}}
(n,0,\alpha)
\xrightarrow{{{{\mathtt{F}}}}}
(F_{\alpha}(n),0,\alpha)
\xrightarrow{{{{\mathtt{Pop}}}}'}
(0,F_{\alpha}(n),\alpha)
\:.$$ The first part of that computation just relies on our convention for reading $\kappa_{d-1},\ldots,\kappa_0$ as the encoding of an ordinal, the second (crucial) part is given by Lemma \[lem:completenessF\], and the last part is by an analog of derivation for ${{{\mathtt{Pop}}}}'$.
For the safety part, we consider an arbitrary computation of the form $(n,0,{{{\mathbf{0}}}}_d)\xrightarrow{{{{\mathtt{S}}}}}(n',r,{\mathbf{e}})$. The only rule for ${{{\mathtt{S}}}}$ is , so there must exist some steps of the form $$(n,0,{{{\mathbf{0}}}}_d)
\xrightarrow{{{{\mathbf{i}}}}_{\kappa_0}^{c_0}
{{{\mathbf{i}}}}_{\kappa_1}^{c_1}
\cdots
{{{\mathbf{i}}}}_{\kappa_{d-1}}^{c_{d-1}}}
(n,0,\alpha)
\xrightarrow{{{{\mathtt{F}}}}}
(n',m',\alpha')
\xrightarrow{{{{\mathtt{Pop}}}}'}
(n'',r,{\mathbf{e}})
\:.$$ Necessarily, this satisfies $n'+m'\leq F_{\alpha}(n)$ by . And since ${{{\mathtt{Pop}}}}'$ behaves like ${{{\mathtt{Pop}}}}$, we have $n''+r=n'+m'$, as in . All this entails $r\leq
F_{\alpha}(n)$ as required by .
Concluding Remarks {#sec-concl}
==================
We proved that Grammar-controlled VASes or Pushdown VASes cannot weakly compute numerical functions that are sublinear. This was recently shown for plain VASes [@LerSch-rp2014]. We also proved that GVASes can weakly compute the fast-growing functions $F_\alpha$ for all $\alpha<\omega^\omega$ while VASes can only weakly compute $F_\alpha$ for $\alpha<\omega$.
This research is motivated by verification questions for well-structured systems, in particular VASes and their extensions. In this area, weakly computable functions have traditionally been used to prove hardness results. Recent hardness proofs for well-structured systems crucially rely on the ability to weakly compute both fast-growing and slow-growing functions.
This work raises some new questions that are left for future work, including whether GVASes can weakly compute $F_{\omega^\omega}$ and whether slow-growing functions can be weakly computed in other VAS extensions like the VASes with nested zero-tests of [@Reinhardt08].
Another open question is the decidability of the boundedness problem for GVASes. Boundedness is decidable for PVASes [@leroux2014] but the two problems do not coincide: on the one hand, in GVASes we only consider sequences of actions that are the yields of complete derivation trees of a grammar, corresponding to configurations in PVAS that have empty stack content; on the other hand unboundedness in PVASes can come from unbounded counters or unbounded stack, while in GVASes only counters are measured. Indeed, the counter-boundedness problem for PVASes reduces to the boundedness problem for GVASes and is still open, while the stack-boundedness problem was shown decidable in [@leroux2015c].
The reachability problem for GVASes is also a source of open problems. Recently, the complexity of the reachability problem for plain VASes was proved to be between $\mathbf{F}_3$ and $\mathbf{F}_\omega$ [@leroux-stoc19; @leroux-lics19] in the complexity hierarchy set up by Schmitz [@schmitz-toct2016]. Improving the $\mathbf{F}_3$ lower bound in the case of GVASes is an open question, as is the decidability status of the reachability problem.
Well-quasi-ordering flow trees {#app-wqo}
==============================
We now prove Lemma \[lem-FG-wqo\], stating that $\leq_G$ is a well-quasi-ordering of $F(G)$, this for any GVASS $G$. A simple way to prove this is to reformulate $\leq_G$ as an homeomorphic embedding on flow trees labeled with enriched information equipped with further labels.
With a flow tree $t=\sigma[t_1,\ldots,t_k]$ we associate a tree ${\operatorname{adorn}}(t)$ made of a root labeled with $\langle{\operatorname{root}}(t),{\operatorname{root}}(t_1),\ldots,{\operatorname{root}}(t_k)\rangle$ and having ${\operatorname{adorn}}(t_1),\ldots,{\operatorname{adorn}}(t_k)$ as immediate subtrees. The nodes of ${\operatorname{adorn}}(t)$ are labeled by tuples of the form $\langle({{{\mathbf{c}}}}_0,X,{{{\mathbf{c}}}}_k),({{{\mathbf{c}}}}_0,X_1,{{{\mathbf{c}}}}_1),\ldots,
({{{\mathbf{c}}}}_{k-1},X_k,{{{\mathbf{c}}}}_k)\rangle$ where ${{{\mathbf{c}}}}_0,\ldots,{{{\mathbf{c}}}}_k\in {{\mathbb{N}}}^d$ are configurations and where $X{\xLongrightarrow{}}X_1\cdots X_k$ is a rule in $R$, or $X\in {\mathbf{A}}$ is a terminal action and $k=0$. Such a tuple is called an *instance* of the rule $X{\xLongrightarrow{}}X_1\cdots X_k$ (or of the action $X\in {\mathbf{A}}$). Given two instances $\lambda=\langle({{{\mathbf{c}}}}_0,X,{{{\mathbf{c}}}}_k),({{{\mathbf{c}}}}_0,X_1,{{{\mathbf{c}}}}_1),\ldots,
({{{\mathbf{c}}}}_{k-1},X_k,{{{\mathbf{c}}}}_k)\rangle$ and $\lambda'=\langle({{{\mathbf{d}}}}_0,Y,{{{\mathbf{d}}}}_\ell),({{{\mathbf{d}}}}_0,Y_1,{{{\mathbf{d}}}}_1),\ldots,
({{{\mathbf{d}}}}_{\ell-1},Y_\ell,{{{\mathbf{d}}}}_\ell)\rangle$, we write $\lambda\leq\lambda'$ when $(X{\xLongrightarrow{}}X_1\cdots X_k)$ and $(Y{\xLongrightarrow{}}Y_1\cdots Y_\ell)$ are the same rule or action —entailing $k=\ell$— and when ${{{\mathbf{c}}}}_j\leq {{{\mathbf{d}}}}_j$ for all $1\leq j\leq k$. Suppose $s$ is a flow tree with immediate subtrees $s_1, \ldots,
s_k$. We write ${\operatorname{adorn}}(s) \sqsubseteq {\operatorname{adorn}}(t)$ if there is a subtree $t'$ of $t$ with immediate subtrees $t_1, \ldots, t_l$ such that ${\operatorname{root}}({\operatorname{adorn}}(s)) \le {\operatorname{root}}({\operatorname{adorn}}(t'))$ (entailing $k = l$) and inductively ${\operatorname{adorn}}(s_j) \sqsubseteq {\operatorname{adorn}}(t_j)$ for all $1 \le j \le k$. It is easy to see that on derived trees of flow trees, $\sqsubseteq$ coincides with the standard homeomorphic embedding of labeled trees.
\[lem-FG-kruskal\] $s\leq_G t$ if, and only if, ${\operatorname{root}}(s)\leq{\operatorname{root}}(t) \land
{\operatorname{adorn}}(s)\sqsubseteq{\operatorname{adorn}}(t)$.
We assume $s=\sigma[s_1,\ldots,s_k]$, $t=\theta[t_1,\ldots,t_\ell]$ and prove the claim by structural induction.\
${\implies}$: Assume $s\leq_G t$. Thus ${\operatorname{root}}(s)\leq{\operatorname{root}}(t)$ and, by definition of $\leq_G$, $t$ contains a subtree $t'=\theta'[t'_1,\ldots,t'_k]$ with $$\sigma\leq\theta' \land s_1\leq_G t'_1 \land \cdots \land s_k\leq_G
t'_k \:.$$ Now this entails ${\operatorname{root}}(s)\leq{\operatorname{root}}(t')$ and ${\operatorname{root}}(s_i)\leq{\operatorname{root}}(t'_i)$ for all $i$, as well as (by ind. hyp.) ${\operatorname{adorn}}(s_i)\sqsubseteq{\operatorname{adorn}}(t'_i)$ for all $i$. Hence ${\operatorname{adorn}}(s)\sqsubseteq{\operatorname{adorn}}(t')$, entailing ${\operatorname{adorn}}(s)\sqsubseteq{\operatorname{adorn}}(t)$.
${\impliedby}$: We assume ${\operatorname{root}}(s)\leq{\operatorname{root}}(t)$ and ${\operatorname{adorn}}(s)\sqsubseteq{\operatorname{adorn}}(t)$. Hence there is a subtree $t'$ of $t$ with immediate subtrees $t'_1, \ldots, t'_k$ such that ${\operatorname{root}}({\operatorname{adorn}}(s))\leq{\operatorname{root}}({\operatorname{adorn}}(t'))$ and ${\operatorname{adorn}}(s_i)\sqsubseteq {\operatorname{adorn}}(t'_i)$ for all $i=1,\ldots,k$. From ${\operatorname{root}}({\operatorname{adorn}}(s))\leq{\operatorname{root}}({\operatorname{adorn}}(t'))$, we infer that ${\operatorname{root}}(s_i) \leq
{\operatorname{root}}(t'_i)$ for all $i=1, \ldots,k$. Now one witnesses $s\leq_G t$ by observing that $s_i\leq_G t_i$ by ind. hyp.
Since the instances of rules are well-quasi-ordered by $\leq$ (there are only finitely many rules), $\sqsubseteq$ is a well-quasi-ordering by Kruskal’s Tree Theorem [@kruskal60]. With Lemma \[lem-FG-kruskal\] we immediately infer that $\leq_G$ is a well-quasi-ordering.
The flow trees equipped with a the wqo $\leq_G$ satisfies an amalgamation property. This property can be expressed by associating to a flow tree $t$ rooted by a triple $({\mathbf{c}},X,{\mathbf{d}})$ the pair $({\mathbf{c}},{\mathbf{d}})$ called the *direction* of $t$ and denoted by ${\operatorname{dir}(t)}$.
\[thm:amalgamation\] For every flow trees $s\leq_G t_1,t_2$ there exists a flow tree $s'$ such that $t_1,t_2\leq_G s'$ and such that: $${\operatorname{dir}(s')}-{\operatorname{dir}(s)}={\operatorname{dir}(t_1)}-{\operatorname{dir}(s)}+{\operatorname{dir}(t_2)}-{\operatorname{dir}(s)}\:.$$
The proof is based on the notion of iterable context. An iterable context for a triple $\sigma=({\mathbf{c}},X,{\mathbf{d}})$ parameterized by the $4$-tuple ${\mathbf{p}}=({\mathbf{x}},{\mathbf{y}},{\mathbf{u}},{\mathbf{v}})$ is a tree defined like the flow trees except that that the root is labeled by $({\mathbf{c}}+{\mathbf{x}},X,{\mathbf{d}}+{\mathbf{y}})$ and exactly one leaf is labeled by $({\mathbf{c}}+{\mathbf{u}},X,{\mathbf{d}}+{\mathbf{v}})$. All the other leaves are labeled by triples involving terminal symbols in ${\mathbf{A}}$. Observe that the underlying derivation tree and ${\mathbf{p}}$ uniquely define the iterable context. The composition of two iterable contexts $C$ and $C'$ for $\sigma$ parameterized respectively by the $4$-tuples ${\mathbf{p}}$ and ${\mathbf{p}}'$ is the unique iterable context for $\sigma$ parameterized by ${\mathbf{p}}+{\mathbf{p'}}$ with an underlying derivation tree that is equal to the one of $C'$ plugged in the special node of the one of $C$.
The proof of the theorem is performed by induction on the structure of $s$. Assume that $s$ is the flow tree $\sigma[s_1,\ldots,s_k]$ where $\sigma=({\mathbf{c}},X,{\mathbf{d}})$ and $t_1,t_2$ are two flow trees with roots $({\mathbf{c}}+{\mathbf{x}}_1,X,{\mathbf{d}}+{\mathbf{y}}_1)$ and $({\mathbf{c}}+{\mathbf{x}}_2,X,{\mathbf{d}}+{\mathbf{y}}_2)$ for ${\mathbf{x}}_1,{\mathbf{y}}_1,{\mathbf{x}}_2,{\mathbf{y}}_2\in{{\mathbb{N}}}^d$ such that there exist two sub-flow trees $t^1$ and $t^2$ of $t_1$ and $t_2$ such that $t^1=({\mathbf{c}}+{\mathbf{x}}^1,X,{\mathbf{d}}+{\mathbf{y}}^1)[t^1_{1},\ldots,t^1_{k}]$ and $t^2=({\mathbf{c}}+{\mathbf{x}}^2,X,{\mathbf{d}}+{\mathbf{y}}^2)[t^2_{1},\ldots,t^2_{k}]$ where ${\mathbf{x}}^1,{\mathbf{y}}^1,{\mathbf{x}}^2,{\mathbf{y}}^2\in{{\mathbb{N}}}^d$.
Assume first that $s$ is a leaf, i.e. $k=0$. In that case $X$ is a terminal, i.e. a vector ${\mathbf{a}}\in{{{\mathbb{Z}}}}^d$. It follows that $t_1$ and $t_2$ are leaves as well. Notice that ${\mathbf{c}}+{\mathbf{a}}={\mathbf{d}}$, ${\mathbf{d}}+{\mathbf{y}}_1={\mathbf{c}}+{\mathbf{x}}_1+{\mathbf{a}}$, and ${\mathbf{d}}+{\mathbf{y}}_2={\mathbf{c}}+{\mathbf{x}}_2 + {\mathbf{a}}$. Thus ${\mathbf{y}}_1={\mathbf{x}}_1$ and ${\mathbf{y}}_2={\mathbf{x}}_2$. We introduce the flow tree $s'$ defined as the root $({\mathbf{c}}+{\mathbf{x}}_1+{\mathbf{x}}_2,{\mathbf{a}},{\mathbf{d}}+{\mathbf{y}}_1+{\mathbf{y}}_2)$. Notice that $s'$ satisfies $t_1,t_2\leq_G s'$.
Next, let us assume that $s$ is not a leaf, i.e. $k\geq 1$. In that case, since $t^1$ and $t^2$ are two subtrees of $t_1$ and $t_2$, by removing $t^1$ from $t_1$ and $t^2$ from $t_2$ we get two iterable contexts $C_1$ and $C_2$ for $\sigma$ parameterized by ${\mathbf{p}}_1=({\mathbf{x}}_1,{\mathbf{y}}_1,{\mathbf{x}}^1,{\mathbf{y}}^1)$ and ${\mathbf{p}}_2=({\mathbf{x}}_2,{\mathbf{y}}_2,{\mathbf{x}}^2,{\mathbf{y}}^2)$. We introduce the sequence ${\mathbf{c}}_0,\ldots,{\mathbf{c}}_k$ of configurations and the sequence $X_1,\ldots,X_k$ in $V\cup{\mathbf{A}}$ such that ${\operatorname{root}}(s_j)=({\mathbf{c}}_{j-1},X_j,{\mathbf{c}}_j)$ for every $j\in\{1,\ldots,k\}$. There also exists two sequences ${\mathbf{c}}_0^1,\ldots,{\mathbf{c}}_k^1$ and ${\mathbf{c}}_0^2,\ldots,{\mathbf{c}}_k^2$ of configurations such that ${\operatorname{root}}(t_j^1)=({\mathbf{c}}_{j-1}+{\mathbf{c}}_{j-1}^1,X_j,{\mathbf{c}}_{j}+{\mathbf{c}}_j^1)$ and ${\operatorname{root}}(t_j^2)=({\mathbf{c}}_{j-1}+{\mathbf{c}}_{j-1}^2,X_j,{\mathbf{c}}_{j}+{\mathbf{c}}_j^2)$. For any $j=1,\ldots,k$, and from $s_j\leq_G t^1_{j},t^2_{j}$ we deduce from the induction hypothesis that there exists a flow tree $s'_j$ such that $t_{j}^1,t^2_{j}\leq_G
s'_j$ and such that the following equality holds: $${\operatorname{dir}(s_j')}-{\operatorname{dir}(s_j)}={\operatorname{dir}(t_{j}^1)}-{\operatorname{dir}(s_j)}+{\operatorname{dir}(t^2_{j})}-{\operatorname{dir}(s_j)}\:.$$ It follows that: $${\operatorname{root}}(s'_j)=({\mathbf{c}}_{j-1}+{\mathbf{c}}_{j-1}^1+{\mathbf{c}}_{j-1}^2,X_j,{\mathbf{c}}_j+{\mathbf{c}}_j^1+{\mathbf{c}}_j^2)\:.$$ Notice that ${\mathbf{c}}_0+{\mathbf{c}}_0^1+{\mathbf{c}}_0^2={\mathbf{c}}+{\mathbf{x}}^1+{\mathbf{x}}^2$ and ${\mathbf{c}}_k+{\mathbf{c}}_k^1+{\mathbf{c}}_k^2={\mathbf{d}}+{\mathbf{y}}^1+{\mathbf{y}}^2$. In particular, by plugging-in $[s_1',\ldots,s_k']$ into the iterable context $C_1\circ C_2$ we get a flow trees $s'$ that satisfies the lemma. We have proved the induction.
Monotonicity for Fast-Growing functions {#app-mon-fg}
=======================================
We give detailed proofs for the two monotonicity lemmas stated after the definitions of the $F_\alpha$ functions in section \[sec-hypack\].
For any ordinals $\alpha,\alpha'<\epsilon_0$ and any $n\in{{\mathbb{N}}}$, $F_{\alpha}(n) \leq F_{\alpha\oplus\alpha'}(n)$.
By induction on $\alpha'$, then on $\alpha$. We first observe that, if the claim holds for some given $\alpha$ and $\alpha'$, then it entails $F^m_{\alpha}(n) \leq F^m_{\alpha\oplus\alpha'}(n)$ for any $m>0$ as a consequence of monotonicity, i.e., .
We now consider several cases for $\alpha$ and $\alpha'$:
- If $\alpha'=0$, then $\alpha\oplus\alpha'=\alpha$ and the claim holds trivially.
- If $\alpha=0$ then the claim becomes $F_0(n)\leq F_{\alpha'}(n)$, which holds since $F_0(n)=n+1$ by and $n+1\leq F_{\alpha'}(n)$ by .
- If $\alpha'=\beta'+1$ is a successor then $\alpha\oplus\alpha'$ is $(\alpha\oplus\beta')+1$ and we have $F_{\alpha}(n)\leq
F_{\alpha\oplus\beta'}(n)$ by ind. hyp., $\leq
F_{\alpha\oplus\beta'}^{n+1}(n)$ by , $=
F_{\alpha\oplus\beta'+1}(n)=F_{\alpha\oplus\alpha'}(n)$, and we are done.
- If $\alpha=\beta+1$ is a successor then $\alpha\oplus\alpha'$ is $(\beta\oplus\alpha')+1$ and we have $F_\alpha(n)
=F_{\beta}^{n+1}(n) \leq F_{\beta\oplus\alpha'}^{n+1}(n)$ by ind. hyp., $=F_{\alpha\oplus\alpha'}(n)$.
- The only remaining possibility is that both $\alpha$ and $\alpha'$ are limit ordinals. Then $(\alpha\oplus\alpha')(n)$ is $\alpha\oplus\alpha'(n)$ or $\alpha(n)\oplus\alpha'$, depending on which limit has the CNF with smallest last summand. In the first case we have $F_\alpha(n)\leq F_{\alpha\oplus\alpha'(n)}(n)$ by ind. hyp. since $\alpha'(n)<\alpha'$, $=F_{(\alpha\oplus\alpha')(n)}(n)=F_{\alpha\oplus\alpha'}(n)$. In the second case we have $F_\alpha(n)=F_{\alpha(n)}(n)\leq
F_{\alpha(n)\oplus\alpha'}(n)$ by ind. hyp. since $\alpha(n)<\alpha$, $=F_{(\alpha\oplus\alpha')(n)}(n)=F_{\alpha\oplus\alpha'}(n)$.
For any ordinal $\alpha<\epsilon_0$ and limit ordinal $\lambda<\omega^\omega$, for any $m, n\in{{\mathbb{N}}}$, if $m \leq n$ then $F_{\alpha\oplus\lambda(m)}(n)
\leq
F_{\alpha\oplus\lambda}(n)$.
Let us decompose $\lambda$ under the form $\lambda=\delta+\omega^{k+1}$ so that $\lambda(m)=\delta
+\omega^{k}\cdot (m+1)$. We prove the lemma by induction on $\alpha$. We first observe that, if the claim holds for some given $\alpha$, then $F^p_{\alpha\oplus\lambda(m)}(n)
\leq
F^p_{\alpha\oplus\lambda}(n)$ for every $p > 0$ and every $m, n\in{{\mathbb{N}}}$ such that $m \leq n$. This observation, which is easily proved by induction on $p$, is a consequence of strict expansivity and monotonicity, i.e., and , respectively.
- If $\alpha=0$ we have $F_{\alpha\oplus\lambda(m)}(n)
= F_{\lambda(m)}(n)
\leq F_{\lambda(n)}(n)$ by Lemma \[lem-Fmono-alpha\] since $\lambda(n)=\lambda(m)\oplus
\omega^k\cdot (n-m)$, $= F_{\lambda}(n)
= F_{\alpha\oplus\lambda}(n)$ by .
- If $\alpha=\beta+1$ is a successor, we have $F_{\alpha\oplus\lambda(m)}(n)
= F_{(\beta\oplus\lambda(m))+1}(n)
= F^{n+1}_{\beta\oplus\lambda(m)}(n)$ by , $\leq F^{n+1}_{\beta\oplus\lambda}(n)$ by ind. hyp., $= F_{(\beta\oplus\lambda)+1}(n) = F_{\alpha\oplus\lambda}(n)$ again by .
- If $\alpha=\gamma+\omega^\beta$ is a limit, then $\alpha\oplus\lambda$ is a limit too and we can compare $\omega^\beta$ and $\omega^{k+1}$, the last summands of $\alpha$ and $\lambda$. There are two subcases:
- If $0 < \beta \leq k$ then $(\alpha\oplus \lambda)(n) = \alpha(n)\oplus \lambda$. Moreover, $\alpha\oplus\lambda(m)$ is a limit since $0 < k$, and $(\alpha\oplus\lambda(m))(n)=\alpha(n)\oplus\lambda(m)$. We deduce $F_{\alpha\oplus\lambda(m)}(n) = F_{\alpha(n)\oplus\lambda(m)}(n)$ by , $\leq F_{\alpha(n)\oplus\lambda}(n)$ by ind. hyp., $= F_{(\alpha\oplus\lambda)(n)}(n) = F_{\alpha\oplus\lambda}(n)$ again by .
- If $k+1 \leq \beta$ then $(\alpha\oplus \lambda)(n) = \alpha\oplus
\lambda(n)$. We deduce $F_{\alpha\oplus\lambda(m)}(n)
\leq F_{\alpha\oplus\lambda(n)}(n)$ by Lemma \[lem-Fmono-alpha\] since $m\leq n$, $= F_{(\alpha\oplus\lambda)(n)}(n)
= F_{\alpha\oplus\lambda}(n)$ by and we are done.
For any ordinal $\alpha<\epsilon_0$ and limit ordinal $\lambda<\omega^\omega$, for any $x,y,n,n',m\in{{\mathbb{N}}}$: $$x\leq y\land m\leq n\leq n'
\implies
F^x_{\alpha\oplus\lambda(m)}(n)
\leq
F^y_{\alpha\oplus\lambda}(n')
\:.$$
Let us decompose $\lambda$ under the form $\lambda=\delta+\omega^{k+1}$ so that $\lambda(m)=\delta
+\omega^{k}\cdot (m+1)$. We prove the lemma by induction on $\alpha$, then $y$. There are several cases:
- If $x<y$, we have $F^x_{\alpha\oplus\lambda(m)}(n)\leq
F^{y-1}_{\alpha\oplus\lambda}(n')$ by ind. hyp., and $F^{y-1}_{\alpha\oplus\lambda}(n')\leq
F^{y}_{\alpha\oplus\lambda}(n')$ by .
- If $x=y=0$ the claims trivially holds since $n\leq n'$.
- If $x>1$, the induction hypothesis yields both $(N{\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}})F^{x-1}_{\alpha\oplus\lambda(m)}(n) \leq
F^{y-1}_{\alpha\oplus\lambda}(n')({\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}N')$ and then $F_{\alpha\oplus\lambda(m)}(N) \leq F_{\alpha\oplus\lambda}(N')$, so we are done.
For now on we may assume $x=y=1$.
- If $\alpha=0$ we have $F_{\alpha\oplus\lambda(m)}(n)
=F_{\lambda(m)}(n)\leq F_{\lambda(n)}(n)$ by Lemma \[lem-Fmono-alpha\] since $\lambda(n)=\lambda(m)\oplus
\omega^k\cdot (n-m)$, $= F_{\alpha\oplus\lambda}(n) \leq
F_{\alpha\oplus\lambda}(n')$ by .
- If $\alpha=\beta+1$ is a successor, we have $F_{\alpha\oplus\lambda(m)}(n) =F_{(\beta\oplus\lambda(m))+1}(n)
=F^{n+1}_{\beta\oplus\lambda(m)}(n)$ by , $\leq
F^{n'+1}_{\beta\oplus\lambda}(n')$ by ind. hyp., $=F_{(\beta\oplus\lambda)+1}(n')=F_{\alpha\oplus\lambda}(n')$ again by .
- If $\alpha=\gamma+\omega^\beta$ is a limit, then $\alpha\oplus\lambda$ is a limit too and we can compare $\omega^\beta$ and $\omega^{k+1}$, the last summands of $\alpha$ and $\lambda$. There are two subcases:
- If $0<\beta<k+1$ then $(\alpha\oplus \lambda)(n') =\alpha(n')\oplus \lambda$ and $(\alpha\oplus\lambda(m))(n)=\alpha(n)\oplus\lambda(m)$. We deduce $F_{\alpha\oplus(\lambda(m))}(n) =F_{\alpha(n)\oplus\lambda(m)}(n) \leq
F_{\alpha(n)\oplus\lambda}(n')$ by ind. hyp., $\leq
F_{\alpha(n')\oplus\lambda}(n') =F_{\alpha\oplus\lambda}(n')$ by Lemma \[lem-Fmono-alpha\].
- If $k+1\leq \beta$ then $(\alpha\oplus \lambda)(n') = \alpha\oplus
\lambda(n')$. We deduce $F_{\alpha\oplus\lambda(m)}(n) \leq
F_{\alpha\oplus\lambda(n')}(n)$ by Lemma \[lem-Fmono-alpha\] since $m\leq n'$, $\leq F_{\alpha\oplus\lambda(n')}(n')$ by , $= F_{\alpha\oplus\lambda}(n')$ by and we are done.
[^1]: This work was partly supported by grant ANR-17-CE40-0028 of the French National Research Agency ANR (project BRAVAS), and by the Indo-French CNRS UMI 2000 ReLaX.
[^2]: defined by $s = \sigma[s_1,\ldots,s_k] \sqsubseteq t$ if, and only if, there exists a subtree $t' = \theta'[t'_1,\ldots,t'_{\ell'}]$ of $t$ such that $\sigma \leq \theta'$, $k \leq \ell'$ and $s_1 \sqsubseteq t'_{j_1}, \ldots, s_k \sqsubseteq t'_{j_k}$ for some subsequence $1 \leq j_1 < \cdots < j_k \leq \ell'$.
[^3]: defined via ${\mathbf{X}}\times{\mathbf{Y}} {\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}\{ ({{{\mathbf{x}}}},{{{\mathbf{y}}}}) \mid {{{\mathbf{x}}}}\in{\mathbf{X}},\ {{{\mathbf{y}}}}\in{\mathbf{Y}}\}$.
[^4]: By contrast, we believe that “VAS-definable” predicates are not closed under intersection (unless one requires auxiliary counters to be zero at the end of the computation). This conjecture remains to be proved.
[^5]: defined via ${\mathbf{X}} + {\mathbf{Y}} {\stackrel{\mbox{\begin{tiny}def\end{tiny}}}{=}}\{ {{{\mathbf{x}}}}+ {{{\mathbf{y}}}}\mid {{{\mathbf{x}}}}\in{\mathbf{X}},\ {{{\mathbf{y}}}}\in{\mathbf{Y}}\}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In a previous paper we proposed a factorization principle for the correlation functions of BMN (Berenstein-Maldacena-Nastase) operators in free $\mathcal{N}=4$ super-Yang-Mills theory. These correlators are conjectured to described physical string amplitudes in an infinitely curved Ramond-Ramond pp-wave background. There was a puzzle that the factorization seems to break down for S-channel in the $2\rightarrow 2$ scattering process. Here we resolve this puzzle by including some diagrams missed in the previous paper. We also observe some interesting relations which further support the interpretation of higher genus correlators as physical string loop amplitudes.'
---
3 cm
**A note on S-channel factorization**
**in multi-trace BMN correlators**
0.5 cm
[Min-xin Huang [^1] ]{}
.5cm
[Interdisciplinary Center for Theoretical Study,\
0.2 cm University of Science and Technology of China, Hefei, Anhui 230026, China]{}
Introduction and Summary
========================
In string theory realizations of particle physics models, we usually assume the string length scale is very small, or the string energy scale is very high. The stringy excitation modes are not observable in low energy, and the ground states of string mode give rise to the various fundamental particles that we know.
However, if string theory is consistent at arbitrary energy scale, it is interesting to also consider the opposite limit where the string length is very long. The motivation is probably not for the purpose of constructing realistic particle physics model, but for purely theoretical reasons, for example a deeper understanding of the still mysterious AdS/CFT correspondence [@Maldacena; @Gubser; @Witten], which states that type IIB string theory on the $AdS_5\times S^5$ background is equivalent to the maximally supersymmetric $SU(N)$ gauge theory in 4 dimensions. A particularly interesting limit is the pp-wave geometry, a Penrose limit zooming in the null geometry of the $AdS_5\times S^5$ spacetime. The string excitation modes are described by the BMN (Berenstein-Maldacena-Nastase) operators on the field theory side [@BMN].
In our previous papers [@Huang:2002; @Huang:2010] we studied the (higher genus) correlation functions of BMN operators in free gauge theory. The field theory side of the correspondence is easy. However, on the string theory side, this corresponds to an infinitely negatively curved Ramond-Ramond pp-wave background, where strings are effectively infinitely long and tensionless, and all stringy excitation modes have degenerate mass. Usually the effective field theory approach breaks down in this scenario, and we can not say much about the underlying physics. However, it seems somehow luckily the stringy physics also becomes extremely simplified. We proposed that the string amplitudes can be computed simply by cubic diagrams, and there is a so called “factorization" principle relating the string diagram calculations and field theory calculations, in the spirit of AdS/CFT correspondence. Here we do not have a Lagrangian description on the string side because of the infinite string length. The nice situation arises, probably due to the fact that the spacetime is highly compressed by the infinite curvature. Its structure becomes effectively that of a single point and is thus extremely simple. The string diagrams have only cubic vertices, but no propagator between the vertices, signaling string interactions occurring instantaneously without mediation in an ambient spacetime. The general form of the factorization rule is $$\begin{aligned}
S_i = \sum_j m_{ij} F_j,\end{aligned}$$ where $S_i$ and $F_j$ denote string and field theory diagram contributions, and $m_{ij}$ are non-negative integers denoting the multiplicity of expanding the “short process" of field theory diagrams into “long process" of string diagrams.
If our claim is valid, the higher genus correlation functions represent string loop amplitudes in this simple background. We can straightforwardly compute higher string loop amplitudes which are notoriously difficult. In flat space for critical string theories most calculations have been restricted to less than two loops, see, e.g. [@DHoker:1988pdl]. We can compare with other special situations where higher genus string amplitudes are computable. One case is the non-critical string theories dual to matrix models or matrix quantum mechanics, intensely studied in early 1990’s, see, e.g. [@Ginsparg:1993is]. Another case is topological string theory, where all genus partition function are computed for certain class of non-compact Calabi-Yau manifolds, by e.g. the topological vertex method [@Aganagic:2003db]. For compact Calabi-Yau manifolds, one can compute the amplitudes to very high but not all genera, e.g. in [@Huang:2006hq; @Huang:2015sta]. In these special situations the physical string degrees of freedom are rather simple, mostly “minimal" or topological in nature, lacking the infinite stringy oscillator modes. Here although the spacetime structure of the infinitely curved pp-wave background is extremely simple, we still have the rich physical degrees of freedom of infinite stringy oscillator modes of conventional critical string theory. In this case there is no apparent technical obstruction to compute string amplitudes to any higher genus.
We think it is worthwhile to revisit the proposal in our previous papers [@Huang:2002; @Huang:2010] from time to time with fresh perspective, and to make incremental improvements and corrections. Many interesting developments over the years seem to support the physical significance of infinite curvature limit in AdS/CFT correspondence, and are potentially relevant for further explorations of our work. Gopakumar have long promoted the studies of free field theory in the context of AdS/CFT correspondence in a series of papers [@Gopakumar:2003ns]. Many useful techniques have been developed to compute correlation function of large charge operators, see e.g. [@Pasukonis:2010rv; @Berenstein:2019esh]. More recently, some progress are made to understand string perturbation theory in pp-wave background in the honest Neveu-Schwarz-Ramond formalism [@Cho:2018nfn]. Also, Berkovits use pure spinor formalism to understand free super-Yang-Mills Feynman diagrams in the the small AdS radius limit [@Berkovits:2019ulm].
Since this is a short note, we should not provide lengthy review of the topic here. Basic properties of BMN operators are studied in the early pioneering papers [@Constable1; @KPSS]. We use the notations in our previous paper [@Huang:2010]. In the current work, in Section \[sec2\], we clarify a puzzle about S-channel factorization in our previous paper. in Section \[sec3\], we comment on some properties of the higher genus BMN correlators.
The $2\rightarrow 2$ scattering process {#sec2}
=======================================
The $2\rightarrow 2$ scattering process is particularly familiar from collider physics. Since our spacetime is collapsed to a point, the process here should be thought of as instantaneous, instead of scattering to and from space infinity as in usual quantum field theory in flat space. To be consistent we use the same symbols for various diagrams in our previous paper [@Huang:2010]. We also consider 3 cases, and point out the missing diagrams in each case. For convenience we list the formulas for BMN operators $$\begin{aligned}
\label{BMNoperators}
&& O^{J} = \frac{1}{\sqrt{JN^J}}TrZ^J, ~~~~~~~~~
O^{J}_{0} = \frac{1}{\sqrt{N^{J+1}}} Tr(\phi^{I} Z^{J}), \nonumber \\
&& O^J_{-m,m} = \frac1{\sqrt{JN^{J+2}}} \sum_{l=0}^{J}e^{\frac{2\pi iml}{J}}
Tr(\phi^{I_1} Z^l\phi^{I_2} Z^{J-l}).\end{aligned}$$ Here $Z$ is a complex scalar in the $\mathcal{N}=4$ super-Yang-Mills theory, and $\phi^{I_1}$ and $\phi^{I_2}$ are the two different real scalar fields out of the 4 remaining ones besides those in $Z$. We take the BMN limit $J, N \rightarrow \infty$ with finite $g:=\frac{J^2}{N}$ as the effective string coupling constant. The planar three point functions can be derived from the Green-Schwarz string field theory vertex [@SV; @Huang1], and serve as the cubic vertex in the string diagrams. As in the previous paper, we omit the universal spacetime factor in the correlators.
$$\begin{aligned}
\label{planar1}
&& \langle\bar{O}^JO^{J_1}O^{J_2}\rangle=\frac{g}{\sqrt{J}}\sqrt{x(1-x)}, ~~~
\langle\bar{O}^J_{0}O^{J_1}O^{J_2}_{0}\rangle = \frac{g}{\sqrt{J}}x^{\frac{1}{2}}(1-x), \nonumber \\
&& \langle\bar{O}^J_{00}O^{J_1}_{0}O^{J_2}_{0}\rangle = \frac{g}{\sqrt{J}}x(1-x), ~~~
\langle\bar{O}^J_{00}O^{J_1}_{00}O^{J_2}\rangle = \frac{g}{\sqrt{J}}x^{\frac{3}{2}}(1-x)^{\frac{1}{2}}, \nonumber \\
&& \langle\bar{O}^J_{-m,m
}O^{J_1}_{0}O^{J_2}_{0}\rangle = - \frac{g}{\sqrt{J}}\frac{\sin^2(\pi mx)}{\pi^2m^2}, \nonumber \\
&& \langle\bar{O}^J_{-m,m}O^{J_1}_{-n,n}O^{J_2}\rangle = \frac{g}{\sqrt{J}}x^{\frac{3}{2}}(1-x)^{\frac{1}{2}}\frac{\sin^2(\pi mx)}{\pi^2 (mx-n)^2},\end{aligned}$$
where $x=\frac{J_1}{J}$, and it is implicit that $J=J_1+J_2$. If the denominator is zero in the last two formulas, then one uses special case formulas. The following integral form is valid for all integers $m,n$, and are sometimes quite useful for checking the factorization rule $$\begin{aligned}
\label{integralform}
\langle\bar{O}^J_{-m,m
}O^{J_1}_{0}O^{J_2}_{0}\rangle&=& \frac{g}{\sqrt{J}} (\int_0^x dy_1 e^{2\pi i m y_1}) (\int_x^1 dy_2 e^{-2\pi i m y_2}) \\ \nonumber
\langle\bar{O}^J_{-m,m
}O^{J_1}_{-n,n}O^{J_2}\rangle&=& \frac{g}{\sqrt{J}} ( \frac{1-x}{x} )^{\frac{1}{2}} (\int_0^x dy_1 e^{2\pi i (m-\frac{n}{x})y_1}) (\int_0^x dy_2 e^{-2\pi i (m-\frac{n}{x})y_2}). \end{aligned}$$
### Case one: $\bra \bar{O}^{J_1} \bar{O}^{J_4}O^{J_2}O^{J_3}\ket$
![Some field theory diagrams for $\bra \bar{O}^{J_1} \bar{O}^{J_4}O^{J_2}O^{J_3}\ket$.[]{data-label="F3"}](F3.pdf){width="6.5in"}
This is the correlator of the vacuum operators, and it is implicit that $J_1+J_4=J_2+J_3$, and we denote $x_i=\frac{J_i}{J}$.. Without loss of generality we assume $J_1>J_2>J_3>J_4$. In the previous paper we calculated the field theory amplitudes for diagrams depicted in Fig. \[F3\] and found $$\begin{aligned}
F^{(3)}_1 &=& \frac{J_1J_2J_3J_4(J_1-J_2)}{N^2\sqrt{J_1J_2J_3J_4}}= \frac{g^2}{J}(x_1x_2x_3x_4)^{\frac{1}{2}}(x_1-x_2), \nonumber \\
F^{(3)}_2 &=& \frac{J_1J_2J_3J_4(J_1-J_3)}{N^2\sqrt{J_1J_2J_3J_4}}= \frac{g^2}{J}(x_1x_2x_3x_4)^{\frac{1}{2}}(x_1-x_3). \label{F3eq} \end{aligned}$$
![The new diagram missed in our previous paper for $\bra \bar{O}^{J_1} \bar{O}^{J_4}O^{J_2}O^{J_3}\ket$.[]{data-label="F3a"}](F3a.pdf){width="3.3in"}
The missing diagram in the previous paper [@Huang:2010] is depicted in Fig. \[F3a\]. Here the double trace is also further dissected into 4 segments, and one can check it has the the same large $N$ factor as the diagrams in Fig. \[F3\]. To compute its contribution, we note that the diagram is determined by the dissecting the shortest trace $O^{J_4}$ into 2 parts, and then the positions of the other dissecting points uniquely follow. This contributes a factor of $J_4$ and the other factors are the same as in the (\[F3eq\]). So the contribution is $$\begin{aligned}
F^{(3)}_3 &=& \frac{J_1J_2J_3J_4}{N^2\sqrt{J_1J_2J_3J_4}} J_4= \frac{g^2}{J}(x_1x_2x_3x_4)^{\frac{1}{2}}x_4.\end{aligned}$$
![The string diagrams for $\bra \bar{O}^{J_1} \bar{O}^{J_4}O^{J_2}O^{J_3}\ket$. We denote the contributions of the 3 diagrams $S^{(3)}_1$, $S^{(3)}_2$, $S^{(3)}_3$ respectively. The 3 diagrams represent the $T$, $U$, $S$ channels in $2\rightarrow 2$ scattering. []{data-label="S3"}](S3.pdf){width="6.5in"}
The string diagrams are depicted in Fig. \[S3\], and their contributions are $$\begin{aligned}
S^{(3)}_1 &=& \frac{g^2}{J}(x_1x_2x_3x_4)^{\frac{1}{2}}(x_1-x_2) \\
S^{(3)}_2 &=& \frac{g^2}{J}(x_1x_2x_3x_4)^{\frac{1}{2}}(x_1-x_3) \\
S^{(3)}_3 &=& \frac{g^2}{J}(x_1x_2x_3x_4)^{\frac{1}{2}}\end{aligned}$$
We follow the proposal in our previous paper [@Huang:2010] to count the multiplicity, by expanding the short process of the field theory diagrams into long processes $$\begin{aligned}
F^{(3)}_1 : && (123)_1(4)_4 \rightarrow (1)_2(23)(4)_4 \rightarrow (1)_2(324)_3 \nonumber \\
&& (123)_1(4)_4 \rightarrow (3124) \rightarrow (1)_2(243)_3 \\
F^{(3)}_2 : && (123)_1(4)_4 \rightarrow (1)_3(23)(4)_4 \rightarrow (1)_3(324)_2 \nonumber \\
&& (123)_1(4)_4 \rightarrow (3124) \rightarrow (1)_3(243)_2
\\
F^{(3)}_3 : && (12)_1(34)_4 \rightarrow (2134) \rightarrow (13)_2(24)_3 \nonumber \\
&& (12)_1(34)_4 \rightarrow (1243) \rightarrow (13)_2(24)_3\end{aligned}$$ We find the string diagram $S^{(3)}_1$ has a multiplicity of 1 with respect to $F^{(3)}_1$, the string diagram $S^{(3)}_2$ has a multiplicity of 1 with respect to $F^{(3)}_2$. The string diagram $S^{(3)}_3$ has the multiplicities of 1 with respect to both $F^{(3)}_1$ and $F^{(3)}_2$, and a multiplicity of 2 with $F^{(3)}_3$ from the previously missing diagram. We find that for $S^{(3)}_1$ and $S^{(3)}_2$, which represent the $T$ and $U$ channels of the $2\rightarrow 2$ scattering, the factorization relation holds, namely $$\begin{aligned}
S^{(3)}_1 &=& F^{(3)}_1 \nonumber \\
S^{(3)}_2 &=& F^{(3)}_2 \end{aligned}$$ Now by including the new contribution $F^{(3)}_3$, we find that the factorization also work for the S-channel process $$\begin{aligned}
S^{(3)}_3 &=& F^{(3)}_1+F^{(3)}_2 + 2F^{(3)}_3\end{aligned}$$ The total contribution to the correlator is now $$\begin{aligned}
\bra \bar{O}^{J_1} \bar{O}^{J_4}O^{J_2}O^{J_3}\ket = F^{(3)}_1+F^{(3)}_2+ F^{(3)}_3
= \frac{1}{2}( S^{(3)}_1+S^{(3)}_2 +S^{(3)}_3) \end{aligned}$$
### Case two: $\bra \bar{O}^{J_1} \bar{O}^{J_4}_{-m,m} O^{J_2}_0O^{J_3}_0 \ket$ ($J_1>J_2>J_3>J_4$)
![Some field theory diagrams for $\bra \bar{O}^{J_1} \bar{O}^{J_4}_{-m,m} O^{J_2}_0O^{J_3}_0 \ket$ ($J_1>J_2>J_3>J_4$ ). These diagrams turn out to give vanishing contributions. []{data-label="F4"}](F4.pdf){width="6.5in"}
The field theory diagrams in the previous paper [@Huang:2010] are depicted in Fig. \[F4\]. In both cases it is impossible to put in the scalar insertions without violating planarity, so the contributions of these diagrams vanish $$\begin{aligned}
\label{F4eq}
F^{(4)}_1=0, ~~~~F^{(4)}_2=0\end{aligned}$$
![The new diagram missed in our previous paper for $\bra \bar{O}^{J_1} \bar{O}^{J_4}_{-m,m} O^{J_2}_0O^{J_3}_0 \ket$.[]{data-label="F4a"}](F4a.pdf){width="3.3in"}
The missing diagram is depicted in Fig. \[F4a\], which is structurally the same as Fig. \[F3a\]. But now it is possible to contract the scalar insertion in $O^{J_2}_0$ and $O^{J_3}_0$ without violating planarity. After carefully tracking the combinatorial factors, and separating the $Z$ fields in $O^{J_4}_{-m,m}$ into two parts, we can write the contributions as integral $$\begin{aligned}
\label{F43eq}
F^{(4)}_3 &=& \frac{g^2}{J} \sqrt{\frac{J_1}{J_4}} \int_{0}^{x_4} dy \int_0^yd y_1 e^{2\pi i m \frac{y_1}{x_4}} \int_y^{x_4} d y_2 e^{- 2\pi i m \frac{y_2}{x_4}} \nonumber \\
&=& - \frac{g^2}{J} \frac{(x_1)^{\frac{1}{2}} (x_4)^{\frac{5}{2}} }{2\pi^2 m^2} \end{aligned}$$
![The string diagram for $\bra \bar{O}^{J_1} \bar{O}^{J_4}_{-m,m} O^{J_2}_0O^{J_3}_0 \ket$ ($J_1>J_2>J_3>J_4$ ). This is the only non-vanishing S-channel diagram, which we denote $S^{(4)}$. []{data-label="S4"}](S4.pdf){width="6.5in"}
We look at the string diagrams. The longest operator is $O^{J_1}$, but it has no scalar insertion so it can not decay to $O^{J_2}_0$ or $O^{J_3}_0$. So the $T$, $U$ channels are impossible and we are left only with the S-channel contribution $S^{(4)}$ depicted in Fig. \[S4\]. The vanishing of the $T$, $U$ channels is consistent with the factorization rules since the field theory diagram contributions vanish (\[F4eq\]). We can calculate the S-channel contribution $$\begin{aligned}
S^{(4)} &=& \sum_{k=-\infty}^{+\infty} \bra \bar{O}^{J_1} \bar{O}^{J_4}_{-m,m} O^{J_1+J_4}_{-k,k}\ket \bra \bar{O}^{J_1+J_4}_{-k,k} O^{J_2}_0O^{J_3}_0 \ket \nonumber \\
&=& -\frac{g^2}{J} (x_1)^{\frac{1}{2}} (x_4)^{\frac{3}{2}} \sum_{k=-\infty}^{+\infty} \frac{\sin^2(\pi k x_4)}{\pi^2 (kx_4-m)^2} \frac{\sin^2(\pi k x_2)}{\pi^2 k^2}. \label{S4eq} \end{aligned}$$ The infinite sum can be performed analytically using the useful summation formulas in the Appendix in the previous paper [@Huang:2010]. To apply the summation formulas we note that it is convenient to use a trigonometry identity $4\sin^2(\beta_1) \sin^2(\beta_2) = 2\sin^2(\beta_1) +2\sin^2(\beta_2) - \sin^2(\beta_1+\beta_2) -\sin^2(\beta_1-\beta_2) $. Quite nicely, it turns out that the infinite sum in (\[S4eq\]) is actually independent of $x_2$ as long as $x_2>x_4$ as assumed here. With the new contributions, we check indeed the S-channel factorization is satisfied $$\begin{aligned}
\bra \bar{O}^{J_1} \bar{O}^{J_4}_{-m,m} O^{J_2}_0O^{J_3}_0 \ket = F^{(4)}_3 = \frac{1}{2} S^{(4)}\end{aligned}$$
Here it is probably more illuminating to check the factorization rule with the integral form of the vertex (\[integralform\]), as it was done in the previous paper for torus two-point function [@Huang:2010]. The sum over intermediate states can be done by the Poisson resummation formula $\sum_{k=-\infty}^{\infty} e^{2\pi i kx}= \sum_{p=-\infty}^{+\infty}\delta(x-p) $. We can write the S-channel contribution as $$\begin{aligned}
S^{(4)} &=& \frac{g^2}{J} (\frac{x_1}{x_4})^{\frac{1}{2}} \int_0^{x_4} d y_1 e^{-2\pi i \frac{m}{x_4} y_1} \int_0^{x_4} d y_2 e^{ 2\pi i \frac{m}{x_4} y_2}\nonumber \\ && \times \sum_{p=-\infty}^{\infty}\int_0^{x_2} d y_3 \int_{x_2}^1 d y_4 \delta(y_1-y_2+y_3-y_4-p).
\end{aligned}$$ We can analyze the integral. Since the final result is real, the contributions from $y_1>y_2$ and $y_2>y_1$ are the same. We consider for example $y_1>y_2$, then the sum over $p$ has non-vanishing contribution only for $p=0$. Since $x_4$ is the smallest, we have the integral $\int_{0}^{x_2} d y_3 \int_{x_2}^{1} d y_4 \delta(y_1-y_2+y_3-y_4) = y_1 - y_2 =\int_{y_2}^{y_1} d y $. Finally, we can re-arrange the integral domain by $\int_0^{x_4} d y_1 \int_0^{y_1} d y_2 \int_{y_2}^{y_1} d y (\cdots)= \int_{0}^{x_4} d y \int_0^{y} d y_2 \int_y^{x_4} d y_1 (\cdots)$. This is exactly the the integral formula in (\[F43eq\]). Counting the double contributions from $y_1>y_2$ and $y_2>y_1$ we have the factorization rule $S^{(4)} = 2 F^{(4)}_3 $.
### Case three: $\bra \bar{O}^{J_1}_{-m,m} \bar{O}^{J_4} O^{J_2}_{-n,n}O^{J_3} \ket$ ($J_1>J_2>J_3>J_4$)
![Some field theory diagrams for $\bra \bar{O}^{J_1}_{-m,m} \bar{O}^{J_4} O^{J_2}_{-n,n}O^{J_3} \ket$. []{data-label="F5"}](F5.pdf){width="5.5in"}
![The new diagram missed in our previous paper for $\bra \bar{O}^{J_1}_{-m,m} \bar{O}^{J_4} O^{J_2}_{-n,n}O^{J_3} \ket$. []{data-label="F5a"}](F5a.pdf){width="3.3in"}
The field theory diagrams are depicted in Fig \[F5\] and Fig \[F5a\]. The two diagrams in Fig \[F5\] have been computed in the previous paper, and the results are $$\begin{aligned}
F^{(5)}_1 &=& \frac{g^2}{J} (x_1)^{-\frac{1}{2}}x_2^{\frac{3}{2}} (x_3x_4)^{\frac{1}{2}} (x_1-x_2)\int_0^1 dy_1e^{-2\pi i (\frac{mx_2}{x_1}-n)y_1} \int_0^1 dy_2e^{2\pi i (\frac{mx_2}{x_1}-n)y_2} \nonumber \\
&=& \frac{g^2}{J} (x_1x_2)^{\frac{3}{2}} (x_3x_4)^{\frac{1}{2}} (x_1-x_2)\frac{1-\cos(2m\pi\frac{x_2}{x_1})}{2\pi^2(mx_2-nx_1)^2},\end{aligned}$$ $$\begin{aligned}
F^{(5)}_2 &=& \frac{g^2}{J} (x_1x_2)^{-\frac{1}{2}} (x_3x_4)^{\frac{1}{2}} (x_1-x_3)^3 \int_0^1 dy \times \nonumber \\ &&
| (e^{- 2\pi i n\frac{x_1-x_3}{x_2}} \int_0^y dy_1+\int_y^1 dy_1) e ^{2\pi i (x_1-x_3)(\frac{m}{x_1}-\frac{n}{x_2})y_1} |^2 \nonumber \\
&=& \frac{g^2}{J} \frac{(x_1x_2)^{\frac{3}{2}} (x_3x_4)^{\frac{1}{2}}}{2\pi^3(nx_1-mx_2)^3}\{ \pi(nx_1-mx_2)(x_1-x_3)
[2-\cos(2m\pi\frac{x_3}{x_1})-\cos(2n\pi\frac{x_4}{x_2})] \nonumber \\ &&
+x_1x_2[\sin(2m\pi \frac{x_3}{x_1})-\sin(2n\pi\frac{x_4}{x_2})-\sin(2\pi\frac{(nx_1-mx_2)(x_1-x_3)}{x_1x_2})] \}
\label{F52}\end{aligned}$$
![The string diagrams for $\bra \bar{O}^{J_1}_{-m,m} \bar{O}^{J_4} O^{J_2}_{-n,n}O^{J_3} \ket$ ($J_1>J_2>J_3>J_4$ ). We denote the contributions of the 2 diagrams $S^{(5)}_1$, $S^{(5)}_2$, $S^{(5)}_3$ respectively. []{data-label="S5"}](S5.pdf){width="6.5in"}
The string diagrams are depicted in Fig. \[S5\]. In the previous paper we have computed the $T,U$ channels contributions denoted by $S^{(5)}_1$, $S^{(5)}_2$, and verified the factorization relation. $$\begin{aligned}
S^{(5)}_1= \bra \bar{O}^{J_1}_{-m,m} O^{J_2}_{-n,n}O^{J_1-J_2}\ket \bra \bar{O}^{J_1-J_2}\bar{O}^{J_4} O^{J_3} \ket
=F^{(5)}_1,\end{aligned}$$ $$\begin{aligned}
S^{(5)}_2 = \sum_{k=-\infty}^{+\infty} \bra \bar{O}^{J_1}_{-m,m} O^{J_3} O^{J_1-J_3}_{-k,k}\ket \bra \bar{O}^{J_1-J_3}_{-k,k}\bar{O}^{J_4} O^{J_2}_{-n,n} \ket = F^{(5)}_2\end{aligned}$$
We calculate the new diagram in Fig. \[F5a\]. The two oscillator modes are located at the left side of the diagram, $$\begin{aligned}
F^{(5)}_3 &=& \frac{g^2}{J} (\frac{x_3x_4}{x_1x_2})^{\frac{1}{2}} \int_0^{x_4} dy~ |\int_0^{x_2-y} e^{-2\pi i (\frac{m}{x_1} - \frac{n}{x_2}) y_1 } d y_1|^2
\nonumber \\
&=& \frac{g^2}{J} \frac{(x_1x_2)^{\frac{3}{2}} (x_3x_4)^{\frac{1}{2}}}{4\pi^3(nx_1-mx_2)^3} [2\pi (nx_1-mx_2)x_4
+x_1x_2\sin(2m\pi\frac{x_2}{x_1}) \nonumber \\ && + x_1x_2 \sin(2\pi\frac{(nx_1-mx_2)(x_2-x_4)}{x_1x_2}) ]. \end{aligned}$$
It is quite involved to use the integral form and Poisson resummation formula to check the S-channel factorization. It is more straightforward to perform the infinite sum using the formulas in the Appendix of our previous paper [@Huang:2010]. We check indeed $$\begin{aligned}
S^{(5)}_2 = \sum_{k=-\infty}^{+\infty} \bra \bar{O}^{J_1}_{-m,m} \bar{O}^{J_4}O^{J}_{-k,k}\ket \bra \bar{O}^{J}_{-k,k} O^{J_2}_{-n,n} O^{J_3} \ket = F^{(5)}_1+F^{(5)}_2 + 2F^{(5)}_3 .\end{aligned}$$
The total contributions to the correlator is $$\begin{aligned}
\bra \bar{O}^{J_1}_{-m,m} \bar{O}^{J_4} O^{J_2}_{-n,n}O^{J_3} \ket =F^{(5)}_1+F^{(5)}_2+F^{(5)}_3=\frac{1}{2} (S^{(5)}_1+S^{(5)}_2+S^{(5)}_3) .\end{aligned}$$
.
Some observations about higher genus correlators {#sec3}
================================================
In quantum mechanics, for a initial state $|\psi\ket$ with the proper normalization ${\bra \psi|\psi\ket=1}$, we may make a measurement about a physical observable such as the energy. It is well known that the eigenstates of the physical observable form a complete orthonormal basis. Summing the probability of all possible final states is simply the unity probability $\sum_{n} |\bra \psi|n\ket|^2=1$.
Analogously, we would like to interpret the two single-string correlator $\bra \bar{O}^{J}_{-m,m} O^{J}_{-n,n} \ket_{h}$ of genus $h$ as the physical $h$-loop probability amplitude of preparing the initial as $O^{J}_{-m,m}$ and observing the final state $O^{J}_{-n,n}$. We note that since the spacetime is infinitely curved, a hypothetical observer are performing the gedanken experiment of measurement without an ambient spacetime. First we consider the torus two-point function $$\begin{aligned}
\label{F6torus}
&& \bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket_{\textrm{torus}} \\
&=& \left\{
\begin{array}{cl}
\frac{g^2}{24}, & m=n=0; \\
0, & m=0, n\neq0, \\
& \textrm{or}~n=0, m\neq0; \\
g^2(\frac{1}{60} - \frac{1}{24 \pi^2 m^2} + \frac{7}{16 \pi^4 m^4}), & m=n\neq0; \\
\frac{g^2}{16\pi^2m^2} ( \frac{1}{3}+\frac{35}{8\pi^2m^2}), & m=-n\neq0; \nonumber \\
\frac{g^2}{4\pi ^{2}(m-n)^2} ( \frac{1}{3}+\frac{1}{\pi
^2n^2}+\frac{1}{\pi ^2m^2}-\frac{3}{2\pi ^2mn}-\frac{1}{2\pi
^2(m-n)^2}) & \textrm{all~other~cases}
\end{array}
\right.\end{aligned}$$ We find the sum of the probability amplitude over final states is actually quite simple and independent of the initial state mode $$\begin{aligned}
\label{sumover}
\sum_{n=-\infty}^{\infty} \bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket_{\textrm{torus}} = \frac{g^2}{24}. \end{aligned}$$ It is easy to check the infinite sum for some fixed small integers $m$. The result can be better derived using the integral formula $$\begin{aligned}
\label{torusintegral}
&& \bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket_{\textrm{torus}} \\
&=& g^2 \int_{0}^1 d x_1 dx_2 dx_3 dx_4 \delta(x_1+x_2+x_3+x_4-1)
[\int_0^{x_1} dy_1 e^{-2\pi i (m-n) y_1}] \nonumber \\
&& \cdot [\int_0^{x_1} dy_2 e^{2\pi i (m-n) y_2} + e^{2\pi im(x_3+x_4) } \int_{x_1}^{x_1+x_2} dy_2 e^{2\pi i (m-n) y_2}
\nonumber \\ &&
+e^{2\pi i m(x_4-x_2) } \int_{x_1+x_2}^{1-x_4} dy_2 e^{2\pi i (m-n) y_2}
+e^{-2\pi i m(x_2+x_3) } \int_{1-x_4}^{1} dy_2 e^{2\pi i (m-n) y_2} ]. \nonumber\end{aligned}$$ Summing over $n$ with Poisson resummation formula gives a delta function, which only has non-vanishing contribution for the first of the 4 terms in the last factor of the integrand. The delta function also ensure the dependence on initial state mode $m$ cancels out. The final integral is simply $\int_{0}^1 d x_1 dx_2 dx_3 dx_4 \delta(x_1+x_2+x_3+x_4-1) x_1=\frac{1}{24}$, proving the formula (\[sumover\]).
An alternative derivation is to use factorization formula $$\begin{aligned}
\bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket_{\textrm{torus}} &=& \frac{1}{2} (\sum_{J_1=1}^{J-1}\sum_{k=-\infty}^{+\infty} \langle
\bar{O}^J_{-m,m} O^{J_1}_{-k,k} O^{J-J_1} \rangle \langle
\bar{O}^{J_1}_{-k,k} \bar{O}^{J-J_1} O^J_{-n,n}
\rangle \nonumber \\ &&+
\sum_{J_1=1}^{J-1}\langle \bar{O}^J_{-m,m} O^{J_1}_0 O^{J-J_1}_0
\rangle \langle \bar{O}^{J_1}_0 \bar{O}^{J-J_1}_0
O^J_{-n,n} \rangle ) \label{factorizationtorus}\end{aligned}$$ We can perform the calculations using the integral form of the vertex (\[integralform\]) and sum over final states. Again using Poisson resummation formula, we derive the formula (\[sumover\]). Here the second term in (\[factorizationtorus\]) vanishes with summing over $n$, and the contribution comes entirely from only the first term.
The formula (\[sumover\]) supports the interpretation of $\bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket _{\textrm{torus}} $ as a kind of observable probability amplitude, as the proper normalization to unity is independent of initial string mode $m$. It might seem a little strange that unlike usual quantum mechanics, we do not need to take the norm square of the amplitude. The underlying reason will become clear later. For now this is not a problem since one can easily show $\bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket _{\textrm{torus}} $ is already real and non-negative. Furthermore, summing instead the square of torus two point function over final modes $n$ would be dependent on initial mode $m$, and would not give a nice formula like (\[sumover\]).
A subtlety here is that this interpretation only works for single string states. For example, the planar three point amplitude $\langle\bar{O}^J_{-m,m}O^{J_1}_{0}O^{J_2}_{0}\rangle$ for $m\neq 0$ seems to be negative. However, this is not really a problem because they actually vanish in the BMN limit where $J, N\rightarrow \infty$, noticing the factor of $\sqrt{J}$ in all three-point functions in (\[planar1\]). Only the (higher genus) two-point functions remains finite in this strict BMN limit. So we argue that the single string states form a complete Hilbert space by themselves, and we can regard the multi-strings as virtual states only appearing in the intermediate steps of a physical process. In this sense an amplitude with external multi-string states such as the $2\rightarrow 2$ amplitudes discussed in section \[sec2\] are virtual amplitudes. The virtual amplitudes are still useful because they can make a finite contribution to the physical two-point amplitudes in a string loop diagram calculation due to also a large number $J$ ways to split a single string into two strings, as e.g. in the factorization formula (\[factorizationtorus\]). In this way, the contributions of general multi-strings have been already accounted for when we study two-point functions, due to formulas like (\[factorizationtorus\]). It would be redundant to include them as part of the physical Hilbert space.
In quantum mechanics we can multiply the quantum state by a complex phase factor and do not change the underlying physical quantum state. This seems to pose a problem for our probability interpretation. For example, if we change the initial state BMN operator by a minus sign, then the two point function would become negative, and inconsistent with a probability interpretation. Our strategy here is to first work with the BMN states, and fix a uniform phase factor for the operators, avoiding the problem for the moment. Indeed the BMN operators form a preferred natural basis for physical states. If we deform away from the infinite curvature limit in the pp-wave background, corresponding to turning on gauge interactions on the field theory side, the mass degeneracy of string excited states is broken. The BMN operators were originally proposed to correspond to the string mass eigenstates, and their conformal dimensions in the planar limit compute the free string mass spectrum [@BMN]. In this sense the BMN strings are the “on-shell" states which appear as external states in usual quantum field theory calculations of scattering amplitudes. We argue this basis is preferred even in the infinite curvature limit where the string mass spectrum is degenerate. Of course, the fundamental principle of quantum mechanics implies there should be complex linear superpositions of BMN states. We will explain later how to deal with them.
In quantum mechanics, an observer performs measurement with respect to an orthogonal basis of states. Here since there appears to be non-trivial overlaps between BMN strings, one may wonder whether they form an orthogonal basis for measurement. To understand this point, we can compare again with scattering amplitudes in quantum field theory, where the external states are mass eigenstates of free theory. The interactions are turned off with the external states propagating asymptotically away to spacetime infinity. Likewise, we argue here that when a hypothetical observer measures the external BMN states, the string interactions are turned off. The BMN operators are certainly orthogonal for zero string coupling $g=0$ where we have only planar contributions, so are consistent for quantum measurement. While in quantum field theory we put the interaction vertices in the middle of initial and final states to compute the scattering amplitudes, in our case the BMN strings self interact by splitting and rejoining when we turn on the string coupling $g>0$.
For general higher genus $h$, the field theory has the $ \frac{(4h-1)!!}{2h+1}$ cyclically different diagrams [@HZ]. Similar to the genus one case, using the field theory integrals, one can show the contribution of each diagram summing over all final states is $ \frac{g^{2h}}{(4h)!}$. So the formula is $$\begin{aligned}
\sum_{n=-\infty}^{\infty} \bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket_{h} = \frac{(4h-1)!!}{(2h+1)(4h)!} g^{2h}.\end{aligned}$$ This is the coefficient of the series expansion of the all-genus formula for the vacuum correlator $\bra \bar{O}^J O^J \ket_{\textrm{all genera}} = \frac{2\sinh(g/2)}{g}$.
An immediate consequence of the probability interpretation is that $\bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket_{h}$ must always be non-negative. This fact can be understood from two perspectives. From the field theory perspective, we are doing integrals of two oscillator modes along the string. For example, in the formula for torus integral (\[torusintegral\]), for one oscillator mode we only need to do the integral for one of the 4 segments, due to cyclicity. We can also write it in a more complicated but symmetric form that treats both oscillator modes equally. Then the integrand is positively proportional to the product of two complex conjugate parts, so is always non-negative. On the other hand, from string theory perspective, up to 2 scalar modes, the only negative cubic vertex is $\langle\bar{O}^J_{-m,m}O^{J_1}_{0}O^{J_2}_{0}\rangle$ for $m\neq 0$. This type of vertex separates two string oscillator modes into two strings, and must always appears in even numbers in a string diagram of $\bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket_{h}$. Here we need to be a little careful to discuss some special cases since $\langle\bar{O}^J_{0,0}O^{J_1}_{0}O^{J_2}_{0}\rangle$ is positive. If both $m,n$ are not zero, or $m=n=0$, the negative vertex always appear in pairs and the string diagram contribution is positive. In the special case one of $m,n$ is zero and the other is not, some string diagram contributions can be negative, but the overall contribution actually vanishes e.g. $\bra \bar{O}_{-m,m}^J O_{0,0}^J \ket_{h} = 0$ for $m\neq 0$, and is still of course consistent with a probability interpretation.
In the field theory computations, we see that the BMN correlators are secretly accounting for the norm square in the usual quantum probability in the integrand. The underlying reason is the close string level matching condition which requires the two string modes in a BMN operator have opposite sign.
Putting together the results, we can write the total probability $p_{m,n}\geq 0$ of preparing an initial single string state $O_{-m,m}$, then observing a final single string state $O_{-n,n}$, including all string loop contributions, as $$\begin{aligned}
\label{pmatrix}
p_{m,n} = \frac{g}{2\sinh(g/2) } \sum_{h=0}^{\infty} \bra \bar{O}_{-m,m}^J O_{-n,n}^J \ket_{h} .\end{aligned}$$ This is properly normalized by the vacuum correlator so that $\sum_{n=-\infty}^{\infty} p_{m,n} =1$ for any initial mode $m$.
We can now discuss our proposal more precisely in terms of the usual formulation of quantum mechanics. We denote the orthonormal BMN states of free string theory by $|n\ket$. Let us assume the transition amplitude between BMN states can be described by a unitary operator $e^{i\hat{H}(g)}$, where $\hat{H}(g)$ is a Hermitian operator corresponding to the time integral of Hamiltonian in a usual quantum mechanics system. Here $\hat{H}(0)=0$ for free string theory and the operator $\hat{H}(g)$ models string interactions at finite coupling $g$. Our probability interpretation implies that the matrix element $p_{m,n}$ in (\[pmatrix\]) does not correspond naively to the usual transition amplitude $\bra m | e^{i\hat{H}(g)} |n\ket$, but rather to its norm square $|\bra m | e^{i\hat{H} (g)} |n\ket|^2$. This is already strongly supported by the same normalization relation $\sum_{n=-\infty}^{\infty} p_{m,n} =\sum_{n=-\infty}^{\infty} |\bra m | e^{i\hat{H} (g)} |n\ket|^2 = 1$.
In general, a purely real perturbative series is not expected to be unitary. Here in particular we show that naively taking $p_{m,n}$ in (\[pmatrix\]) as the transition amplitude $\bra m | e^{i\hat{H}(g)} |n\ket$ would be indeed inconsistent with unitarity. We allow the generous possibility that the BMN operator $O_{-m,m}^J$ may be normalized by a function $f_m(g)$, depending on both mode number and string coupling. The unitarity condition is then $$\begin{aligned}
\label{naiveunitarity}
\sum_{n=-\infty}^{\infty} f_m( g)^* f_{m^{\prime}}( g) |f_n( g)|^2 p_{m,n} p^*_{m^{\prime},n} = \delta_{m,m^{\prime}} . \end{aligned}$$ Consider the simple case $m=m^{\prime}$ and $g=0$ we deduce $|f_m(0)|=1$ for any mode $m$. Next we consider a non-trivial case $m\neq m^{\prime}$ and both non-zero. Expanding for small $g$, the leading term of left hand side of (\[naiveunitarity\]) is of order $g^2$ from two contributions $n=m, m^{\prime}$ in the summation $$\begin{aligned}
&& \sum_{n=-\infty}^{\infty} f_m( g)^* f_{m^{\prime}}( g) |f_n( g)|^2 p_{m,n} p^*_{m^{\prime},n} \nonumber \\
&=&2 f_m( 0)^* f_{m^{\prime}}( 0) \bra \bar{O}_{-m,m}^J O_{-m^{\prime}, m^{\prime}}^J \ket_{\textrm{torus}} +\mathcal{O}(g^3) . \end{aligned}$$ This is clearly non-zero, violating unitarity. So we conclude if we believe the free single-string states form a complete Hilbert space $\sum_n |n\ket \bra n|=1$, the naive proposal is basically ruled out.
Coming back to the correct proposal, by introducing some real phase angles, we can write the transition amplitude as $$\begin{aligned}
\label{transition}
\bra m | e^{i\hat{H}(g)} |n\ket = e^{i\theta_{m,n}(g)} \sqrt{p_{m,n}}. \end{aligned}$$ For $g=0$ this is the identity matrix $\delta_{m,n}$, so the phase angle $\theta_{m,m}(0)=0$. For $m\neq n$, the phase angle $\theta_{m,n}(0)$ is not determined this way since in this case $p_{m,n}$ already vanishes for $g=0$. Another special case is that for $m\neq 0$, the matrix elements $p_{m,0}$ and $p_{0,m}$ vanish for any coupling $g$, so the phase angles $\theta_{m,0}(g), \theta_{0, m}(g)$ are actually redundant. The unitarity condition is $$\begin{aligned}
\label{unitarity}
\sum_{n=-\infty}^{+\infty} e^{i [ \theta_{m,n}(g) - \theta_{m^{\prime},n}(g) ]} \sqrt{p_{m,n} p_{m^{\prime},n}} = \delta_{m,m^{\prime}}. \end{aligned}$$ For $m=m^{\prime}$ this is already satisfied, while the cases of $m\neq m^{\prime}$ may provide some constrains for the phase angles. For the special case where one of $m,m^{\prime}$ is zero and the other is not, the unitary condition (\[unitarity\]) is automatically satisfied without any constrain for the phase angles. Again we consider the non-trivial case of $m\neq m^{\prime}$ and both non-zero. Expanding the left hand side of (\[unitarity\]) for small $g$, we see the leading order term is now of order $g$ with two contributions also from $n=m, m^{\prime}$ in the summation $$\begin{aligned}
&& \sum_{n=-\infty}^{+\infty} e^{i [ \theta_{m,n}(g) - \theta_{m^{\prime},n}(g) ]} \sqrt{p_{m,n} p_{m^{\prime},n}} \nonumber \\ &=& [e^{i \theta_{m,m^{\prime} }(0)} + e^{- i \theta_{m^{\prime},m }(0) }] \sqrt{ \bra \bar{O}_{-m,m}^J O_{-m^{\prime}, m^{\prime}}^J \ket_{\textrm{torus}} } +\mathcal{O}(g^2) . \end{aligned}$$ Since the torus two-point function is non-zero, we arrive at an interesting relation imposed by unitarity that $ \theta_{m,m^{\prime} }(0) + \theta_{m^{\prime} ,m}(0) =\pi$, up to an integer multiple of $2\pi$, for non-zero $m\neq m^{\prime}$. So the phase angles can not be trivially all set to zero even for free string theory, due to the consistency of string interactions. With these extra phase angles, we see that we are able to preserve the unitarity condition violated by the naive proposal.
To illustrate the power of unitarity, let us consider an analogous simpler situation of a Hilbert space of finite dimension $D$, and count the degrees of freedom. In this case there are $D^2$ real phase angles, and the unitarity conditions provide $\frac{D(D-1)}{2}$ complex equations. So generically, there are $D$ remaining free real parameters. There are also some freedom of gauge choices. Firstly there is a freedom to rotate the base states by a phase factor $|n\ket \rightarrow e^{i \theta_n} |n\ket$, so the phase angles in transition amplitude (\[transition\]) are shifted $\theta_{m,n} \rightarrow \theta_{m,n} + \theta_n - \theta_m$. There are $D-1$ free parameters for such choices since $\theta_{m,n}$ is unchanged if all $ \theta_n$’s are the same. Secondly, the overall phase corresponds to a shift of the operator $\hat{H}$ by a real number, which is a zero-point energy, usually considered physically unobservable without coupling the quantum mechanics to gravity. So the number of remaining free parameters $D$ exactly equals the total number of gauge choices. Generically, this feature should remain valid in our case of infinite dimensional Hilbert space. So we expect that the unitarity conditions (\[unitarity\]) can completely determine the phase angles $\theta_{m,n}(g)$ up to the gauge choices discussed above, with a discrete number of inequivalent solutions. We note that in our case since the zero mode decouples from the non-zero modes, there are 2 free choices of the “zero-point energy", the phase angle $\theta_{0,0}(g)$ and the overall phase for the non-zero modes [^2].
It would be interesting to fully explicitly compute these phase angles $\theta_{m,n}(g)$ for general coupling $g$. In any case, now we can in principle follow the usual rule of quantum mechanics to compute the transition amplitudes between complex linear superpositions of BMN states with the equation (\[transition\]). We note that we can not compute by naively putting linear combinations of BMN operators in the correlators in (\[pmatrix\]). Of course, this does not violate the fundamental principle of linear superposition in quantum mechanics, as we now understand that the two point functions in (\[pmatrix\]) do not directly correspond to the inner product of the underlying quantum system. This simply means we need to work a little more carefully instead with the right formula (\[transition\]). Our earlier strategy of fixing a uniform phase for the BMN operators is also now justified.
In usual quantum mechanics, the diagonalization of the transition amplitude gives rise to the energy eigenstates. Since here we restrict ourself to free gauge theory with only string interactions, the BMN states already have completely degenerate mass. So the eigenstates that diagonalize the transition amplitude (\[transition\]), i.e. eigenstates of the Hermitian operator $\hat{H}(g)$, are probably not the conventional energy eigenstates. It is interesting to explore the physical interpretations of such eigenstates in the future.
Similar results are also true for BMN operators with more oscillator modes. For example, the BMN operators with 3 scalar excitation modes and orthonormal at planar level are $$\begin{aligned}
O^{J}_{(m_1,m_2,m_3)} = \frac{1}{\sqrt{N^{J+2}}J} \sum_{l_1, l_2=0}^{J} e^{\frac{2\pi im_2l_1}{J}} e^{\frac{2\pi im_3l_2}{J}} {\textrm{Tr}}(\phi^1 Z^{l_1} \phi^2 Z^{l_2-l_1} \phi^3 Z^{J-l_2}), \label{operator3}\end{aligned}$$ where the 3 scalar modes satisfy the close string level matching condition $m_1+m_2+m_3=0$. The factorization formula was studied in our previous paper [@Huang:2010]. We can show similarly the sum over final states is independent of the initial state modes $$\begin{aligned}
\label{sumover3}
\sum_{n_1+n_2+n_3=0}^{\infty} \bra \bar{O}^{J}_{(m_1,m_2,m_3)} O^{J}_{(n_1,n_2,n_3)} \ket_{\textrm{torus}} = \frac{g^2}{24}. \end{aligned}$$ We also check $ \bra \bar{O}^{J}_{(m_1,m_2,m_3)} O^{J}_{(n_1,n_2,n_3)} \ket_{\textrm{torus}} \geq 0$ is always true including various degenerate cases, consistent with the physical probability interpretation.
Recently Erbin, Maldacena and Skliros consider the two-point string amplitudes, and show that they are not as trivial as previously thought [@Erbin:2019uiz]. These authors consider the conventional flat space, while we consider the infinitely curved pp-wave background. Despite the different settings, we see the development as another supporting evidence that the higher genus two point amplitudes discussed here should have much relevant physical interpretations.
This work was supported by the National Natural Science Foundation of China (Grant No. 11675167) and the national “Young Thousand People" program.
[10]{}
M. Aganagic, A. Klemm, M. Marino and C. Vafa, “The Topological vertex,” Commun. Math. Phys. [**254**]{}, 425 (2005) \[hep-th/0305132\].
D. Berenstein and R. de Mello Koch, “Gauged fermionic matrix quantum mechanics,” JHEP [**1903**]{}, 185 (2019) \[arXiv:1903.01628 \[hep-th\]\].
D. E. Berenstein, J. M. Maldacena and H. S. Nastase, “Strings in flat space and pp waves from N = 4 super Yang Mills,” JHEP [**0204**]{}, 013 (2002) \[arXiv:hep-th/0202021\].
N. Berkovits, “Sketching a Proof of the Maldacena Conjecture at Small Radius,” JHEP [**1906**]{}, 111 (2019) \[arXiv:1903.08264 \[hep-th\]\].
N. R. Constable, D. Z. Freedman, M. Headrick, S. Minwalla, L. Motl, A. Postnikov and W. Skiba, “PP-wave string interactions from perturbative Yang-Mills theory,” JHEP [**0207**]{}, 017 (2002) \[arXiv:hep-th/0205089\].
M. Cho, S. Collier and X. Yin, “Strings in Ramond-Ramond Backgrounds from the Neveu-Schwarz-Ramond Formalism,” arXiv:1811.00032 \[hep-th\].
H. Erbin, J. Maldacena and D. Skliros, “Two-Point String Amplitudes,” JHEP [**1907**]{}, 139 (2019) \[arXiv:1906.06051 \[hep-th\]\].
P. H. Ginsparg and G. W. Moore, “Lectures on 2-D gravity and 2-D string theory,” Yale Univ. New Haven - YCTP-P23-92 (92,rec.Apr.93) 197 p. Los Alamos Nat. Lab. - LA-UR-92-3479 (92,rec.Apr.93) 197 p. e: LANL hep-th/9304011 \[hep-th/9304011\].
S. S. Gubser, I. R. Klebanov and A. M. Polyakov, “Gauge theory correlators from non-critical string theory,” Phys. Lett. B [**428**]{}, 105 (1998) \[arXiv:hep-th/9802109\].
E. D’Hoker and D. H. Phong, “The Geometry of String Perturbation Theory,” Rev. Mod. Phys. [**60**]{}, 917 (1988).
M. x. Huang, “Three point functions of N = 4 super Yang Mills from light cone string field theory in pp-wave,” Phys. Lett. B [**542**]{}, 255 (2002) \[arXiv:hep-th/0205311\].
M. x. Huang, “String interactions in PP wave from N=4 superYang-Mills,” Phys. Rev. D [**66**]{}, 105002 (2002) \[hep-th/0206248\]. M. x. Huang, “Higher Genus BMN Correlators: Factorization and Recursion Relations,” Adv. Theor. Math. Phys. [**16**]{}, no. 2, 421 (2012) \[arXiv:1009.5447 \[hep-th\]\].
M. x. Huang, S. Katz and A. Klemm, “Topological String on elliptic CY 3-folds and the ring of Jacobi forms,” JHEP [**1510**]{}, 125 (2015) \[arXiv:1501.04891 \[hep-th\]\].
M. x. Huang, A. Klemm and S. Quackenbush, “Topological string theory on compact Calabi-Yau: Modularity and boundary conditions,” Lect. Notes Phys. [**757**]{}, 45 (2009) \[hep-th/0612125\].
R. Gopakumar, “From free fields to AdS,1, 2, 3” Phys. Rev. D [**70**]{}, 025009 (2004) \[hep-th/0308184\]; Phys. Rev. D [**70**]{}, 025010 (2004) \[hep-th/0402063\]; Phys. Rev. D [**72**]{}, 066008 (2005) \[hep-th/0504229\].
C. Kristjansen, J. Plefka, G. W. Semenoff and M. Staudacher, “A new double-scaling limit of N = 4 super Yang-Mills theory and PP-wave strings,” Nucl. Phys. B [**643**]{}, 3 (2002) \[arXiv:hep-th/0205033\].
J. M. Maldacena, “The large N limit of superconformal field theories and supergravity,” Adv. Theor. Math. Phys. [**2**]{}, 231 (1998) \[Int. J. Theor. Phys. [**38**]{}, 1113 (1999)\] \[arXiv:hep-th/9711200\].
J. Pasukonis and S. Ramgoolam, “From counting to construction of BPS states in N=4 SYM,” JHEP [**1102**]{}, 078 (2011) \[arXiv:1010.1683 \[hep-th\]\].
M. Spradlin and A. Volovich, “Superstring interactions in a pp-wave background,” Phys. Rev. D [**66**]{}, 086004 (2002) \[arXiv:hep-th/0204146\].
E. Witten, “Anti-de Sitter space and holography,” Adv. Theor.Math. Phys. [**2**]{}, 253 (1998) \[arXiv:hep-th/9802150\].
[^1]: [email protected]
[^2]: For free string theory we have already made such choices by setting $\hat{H}(0)=0$. We can now further simply choose $\theta_{0,0}(g) = 0$ for any string coupling $g$, and focus on the non-trivial non-zero modes. A simple example of fixing all gauge choices for the non-zero mode block is to choose $\theta_{1,m}(g) = 0$ for all $m$’s and any coupling $g$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss nuclear structure functions in lepton scattering including neutrino reactions. First, the determination of nuclear parton distribution functions is explained by using the data of electron and muon deep inelastic scattering and those of Drell-Yan processes. Second, NuTeV $sin^2 \theta_W$ anomaly is discussed by focusing on nuclear corrections in the iron target. Third, we show that the HERMES effect, which indicates nuclear modification of the longitudinal-transverse structure function ratio, should exist at large $x$ with small $Q^2$ in spite of recent experimental denials at small $x$.'
address: 'Department of Physics, Saga University, Saga, 840-8502, Japan'
author:
- 'S. Kumano [^1]'
title: Nuclear modification of structure functions in lepton scattering
---
\
\
\
\
------------------------------------------------------------------------
\
[\* Email: [email protected]. URL: http://hs.phys.saga-u.ac.jp.]{}\
Introduction
============
Modification of nuclear structure functions or nuclear parton distribution functions (NPDFs) is known especially in electron and muon scattering. In neutrino scattering, such nuclear effects have not been seriously investigated due to the lack of accurate deuteron data. Nuclear effects in the PDFs have been investigated mainly among hadron structure physicists. However, demands for accurate NPDFs have been growing from other fields in the recent years. In fact, one of the major purposes of this workshop [@nuint02] is to describe neutrino-nucleus cross sections for long baseline neutrino experiments, so that neutrino oscillations could be understood accurately [@sakuda; @py].
In the near future, neutrino cross sections should be understood within a few percent level for the oscillation studies [@sakuda; @py]. Because typical nuclear corrections in the oxygen nucleus are larger than this level, they should be precisely calculated. In low-energy scattering, the nuclear medium effects are discussed in connection with nuclear binding, Fermi motion, short-range correlations, Pauli exclusion effects, and other nuclear phenomena. In this paper, the nuclear corrections are discussed in the structure functions and the PDFs by focusing on the high-energy region. These studies are important not only for the neutrino studies but also for other applications. For example, they are used in heavy-ion physics [@heavy] for understanding accurate initial conditions of heavy nuclei, so that one could make a definitive statement, for example on quark-gluon plasma formation, in the final state. They could be also used in understanding nuclear shadowing mechanisms [@recent-shadow].
In this paper, recent studies are explained on the nuclear effects which are relevant to high-energy neutrino scattering. First, a recent NPDF $\chi^2$ analysis is reported. Although the unpolarized PDFs in the nucleon have been investigated extensively [@pdf], the NPDFs are not well studied. However, there are some studies to obtain optimum NPDFs by using a simple parametrization form and nuclear scattering data [@ekrs; @hkm]. We explain the current situation. Second, NuTeV $sin^2 \theta_W$ anomaly [@nutev02] is investigated in a conservative way, namely in terms of nuclear corrections [@nutevmod; @nucl-sinth; @sk02; @kulagin]. The NuTeV collaboration obtained anomalously large $sin^2 \theta_W$. Before discussing any new physics mechanisms [@new], we should exclude possible nuclear physics explanations. In particular, the used target is the iron and it may cause complicated nuclear medium effects. Third, the HERMES effect [@hermes00], which is nuclear modification of the longitudinal-transverse structure function ratio, is investigated in a simple convolution model. It is intended to show that such an effect should exist in the medium and large $x$ regions [@ek03] in spite of recent experimental denials at small $x$ [@ccfr01; @hermes02]. In particular, the nucleon Fermi motion in a nucleus could play an important role for the nuclear modification.
This paper consists of the following. In section \[npdf\], global NPDF analysis results are shown The $sin^2 \theta_W$ anomaly topic is discussed in section \[sin2th\]. The HERMES effect is explained in section \[hermes\]. The results are summarized in section \[sum\].
Nuclear parton distribution functions {#npdf}
=====================================
The determination of the NPDFs is not still satisfactory in comparison with the one for the nucleon. It is partly because enough data are not obtained for fixing each distribution from small $x$ to large $x$. For example, various scaling violation data are not available, unlike the HERA data for the proton, at very small $x$ for fixing gluon distributions. However, the determination of nuclear PDFs has been awaited for describing high-energy nuclear scattering phenomena, including neutrino-nucleus and heavy-ion reactions. Some efforts have been made to provide practical parametrizations for the NPDFs, such as the ones by Eskola, Kolhinen, Ruuskanen, Salgado [@ekrs] and the ones by the HKM analysis [@hkm]. In the following, the NPDFs are discussed based on the latter study in Ref. [@hkm].
First, the parametrization form should be selected. From the studies of nuclear $F_2$ structure function ratios $F_2^A/F_2^D$, one knows the existence of shadowing phenomena at small $x$, anti-shadowing at $x\approx 0.2$, depletion at medium $x$, and then a positive nuclear modification at large $x$. In order to express such $x$ dependence, the following functions are used for the initial NPDFs at $Q_0^2$=1 GeV$^2$: $$\begin{aligned}
&
f_i^A (x, Q_0^2) = w_i(x,A,Z) \, f_i (x, Q_0^2),
\nonumber \\
&
w_i(x,A,Z) = 1 + \left( 1 - \frac{1}{A^{1/3}} \right)
\nonumber \\
& \ \ \ \ \ \ \ \ \ \ \
\times
\frac{a_i(A,Z) +b_i x+c_i x^2 +d_i x^3}{(1-x)^{\beta_i}} .
\label{eqn:w}\end{aligned}$$ Here, $Z$ is the atomic number, $A$ is the mass number, and the subscript $i$ indicates a distribution type: $i$=$u_v$, $d_v$, $\bar q$, or $g$. The functions $f_i^A$ and $f_i$ are the PDFs in a nucleus and the nucleon, respectively, so that the weight function $w_i$ indicates nuclear medium effects. The nuclear modification $w_i -1$ is assumed to be proportional to $1-1/A^{1/3}$, and its $x$ dependence is taken to be a cubic functional form with the $1/(1-x)^{\beta_i}$ factor for describing the Fermi-motion part. The parameters $a_i$, $b_i$, $c_i$, and $d_i$ are determined by a $\chi^2$ analysis of experimental data.
Although the flavor dependence of the antiquark distributions is known in the nucleon [@flavor], the details of nuclear antiquark distributions cannot be investigated at this stage. Therefore, flavor symmetric antiquark distributions are assumed in the parametrization.
The electron and muon deep inelastic experimental data and Drell-Yan data are fitted by the NPDFs in Eq. (\[eqn:w\]). The initial NPDFs are, of course, evolved to various experimental $Q^2$ points, and $\chi^2$ values are calculated in comparison with the data for electron and muon deep inelastic scattering and Drell-Yan processes: $$\chi^2 = \sum_j \frac{(R_j^{data}-R_j^{theo})^2}
{(\sigma_j^{data})^2}.
\label{eqn:chi2}$$ Here, $R$ is the ratio $F_2^A/F_2^{A'}$ or $\sigma_{DY}^A/\sigma_{DY}^{A'}$. These structure functions and the DY cross sections are calculated in the leading order. The experimental error is given by systematic and statistical errors as $(\sigma_j^{data})^2 = (\sigma_j^{sys})^2 + (\sigma_j^{stat})^2$. The first version was published in 2001, and then the research is in progress by including the Drell-Yan data. We discuss the obtained NPDFs by these analyses.
Obtained optimum distributions are shown for the calcium nucleus at $Q^2$=1 GeV$^2$ in Fig. \[fig:wxca1\]. The solid, dashed, and dotted curves indicate the weight functions for the valence-quark, antiquark, and gluon distributions. The valence distribution is well determined in the medium $x$ region, but it is difficult to determine it at small $x$ although it is constrained by the baryon-number and charge conservations. In fact, it will be one of the NuMI projects [@numi] to determine the valence-quark ($F_3$) shadowing in comparison with the antiquark ($F_2$) shadowing by neutrino-nucleus scattering. On the other hand, the antiquark distribution is well determined at small $x$; however, it cannot be fixed at medium $x$ ($x>0.2$) in spite of the momentum-conservation constraint. Because this is the leading order analysis, the gluon distribution is not fixed in the whole $x$ region.
{width="45.00000%"}
\[fig:wxca1\]
The obtained NPDFs are available at the web site http://hs.phys.saga-u.ac.jp/nuclp.html, where computer codes are available for calculating the distributions at given $x$ and $Q^2$ for a requested nucleus. The nuclear type should be in the range, $2 \le A \le 208$, in principle, because the analyzed nuclei are in this range. However, the distributions could be also calculated for larger nuclei ($A>208$) because variations of the NPDFs are rather small in such a large-$A$ region. If one wishes to use an analytical form, the distributions at $Q^2$=1 GeV$^2$ are provided in the appendix of Ref. [@hkm]. After the first version was published, a new analysis has been investigated. The second version will become available within the year of 2003.
A nuclear physicist’s view of $sin^2 \theta_W$ anomaly {#sin2th}
======================================================
The NuTeV collaboration announced that their measurement of the weak mixing angle $sin^2 \theta_W$ is significantly different from collider measurements. If the neutrino-nucleus scattering data are excluded, a global analysis indicates $sin^2 \theta_W^{on-shell}= 0.2227 \pm 0.0004$ [@lep01]. On the other hand, the NuTeV reported [@nutev02] $$sin^2 \theta_W = 0.2277 \pm 0.0013 \, \text{(stat)}
\pm 0.0009 \, \text{(syst)}
\, ,$$ by using their neutrino and antineutrino scattering data.
Because it is one of the important constants in the standard model, we should find a reason for the discrepancy. Although it may indicate the existence of a new mechanism [@new], we should seek a conservative explanation first. In particular, the NuTeV target is the iron nucleus so that nuclear medium effects might have altered the $sin^2 \theta_W$ value [@nutevmod; @nucl-sinth; @sk02; @kulagin]. In the following, we explain nuclear effects on the $sin^2 \theta_W$ determination.
The neutrino and antineutrino cross section data are analyzed by a special Monte Carlo code, so that it is not theoretically straightforward to investigate a possible explanation. In order to simplify the investigation, we study nuclear effects on the Paschos-Wolfenstein (PW) relation, which is considered to be “implicitly" used in the NuTeV analysis. The PW relation [@pw] is given by the ratio of charged current (CC) and neutral current (NC) cross sections: $$R^- = \frac{ \sigma_{NC}^{\nu N} - \sigma_{NC}^{\bar\nu N} }
{ \sigma_{CC}^{\nu N} - \sigma_{CC}^{\bar\nu N} }
= \frac{1}{2} - sin^2 \theta_W
\, .
\label{eqn:pw}$$ This relation is valid for the isoscalar nucleon; however, corrections should be carefully investigated for the non-isoscalar iron target.
If the relation is calculated for a nucleus in the leading order of $\alpha_s$, we obtain [@sk02] $$\begin{aligned}
R_A^- & = \frac{ \sigma_{NC}^{\nu A} - \sigma_{NC}^{\bar\nu A} }
{ \sigma_{CC}^{\nu A} - \sigma_{CC}^{\bar\nu A} }
\nonumber \\
& \! \! \! \! \! \! \! \! \! \!
= \{ 1-(1-y)^2 \} \, [ \, (u_L^2 -u_R^2 ) \{ u_v^A(x) + c_v^A (x) \}
\nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
+ (d_L^2 -d_R^2 ) \{ d_v^A (x) + s_v^A (x) \} \, ]
\nonumber \\
& \! \! \! \! \! \! \! \!
/ \, [ \, d_v^A (x) + s_v^A (x)
- (1-y)^2 \, \{ u_v^A (x) + c_v^A (x) \} \, ]
\, ,
\label{eqn:apw1}\end{aligned}$$ where the valence quark distributions are defined by $q_v^A \equiv q^A -\bar q^A$. The couplings are expressed by the weak mixing angle as $u_L = 1/2- (2/3) \, sin^2 \theta_W$, $u_R = -(2/3) \, sin^2 \theta_W$, $d_L = -1/2 +(1/3) \, sin^2 \theta_W$, and $d_R = (1/3) \, sin^2 \theta_W$. It is known that the nuclear distributions are modified from those for the nucleon. The modification for $u_v^A$ and $d_v^A$ could be expressed by the weight functions $w_{u_v}$ and $w_{d_v}$ at any $Q^2$: $$\begin{aligned}
u_v^A (x) & = w_{u_v} (x,A,Z) \, \frac{Z \, u_v (x) + N \, d_v (x)}{A},
\nonumber \\
d_v^A (x) & = w_{d_v} (x,A,Z) \, \frac{Z \, d_v (x) + N \, u_v (x)}{A},
\label{eqn:wpart}\end{aligned}$$ although $w_i$ in section 2 is defined at fixed $Q^2$ (=$Q_0^2$). Here, $u_v$ and $d_v$ are the distributions in the proton, and $N$ is the neutron number.
In order to find possible deviation from the PW relation, we first define a function which is related to the neutron excess in a nucleus: $\varepsilon_n (x) = [(N-Z)/A] (u_v-d_v)/(u_v+d_v) $, and then a difference between the weight functions is defined by $$\varepsilon_v (x) = \frac{w_{d_v}(x,A,Z)-w_{u_v}(x,A,Z)}
{w_{d_v}(x,A,Z)+w_{u_v}(x,A,Z)}
\, .
\label{eqn:en}$$ Furthermore, there are correction factors associated with the strange and charm quark distributions, so that we define $\varepsilon_s$ and $\varepsilon_c$ by $\varepsilon_s = s_v^A /[w_v \, (u_v+d_v)]$ and $\varepsilon_c = c_v^A /[w_v \, (u_v+d_v)]$ with $w_v = (w_{d_v}+w_{u_v})/2$.
Neutron-excess effects are taken into account in the NuTeV analysis as explained by McFarland [*et al.*]{} [@nutevmod], and they are also investigated by Kulagin [@kulagin]. The strange quark ($\varepsilon_s$) contribution is small according to Zeller [*et al.*]{} [@sv], and it increases the deviation. Here, we investigate a different contribution from the $\varepsilon_v (x)$ term [@sk02]. Writing Eq. (\[eqn:apw1\]) in terms of the factors, $\varepsilon_n$, $\varepsilon_v$, $\varepsilon_s$, and $\varepsilon_c$, and then expanding the expressions by these small factors, we obtain $$\begin{aligned}
&
R_A^- = \frac{1}{2} - sin^2 \theta_W
\nonumber \\
&
- \varepsilon_v (x) \bigg \{ \bigg ( \frac{1}{2} - sin^2 \theta_W \bigg )
\frac{1+(1-y)^2}{1-(1-y)^2} - \frac{1}{3} sin^2 \theta_W
\bigg \}
\nonumber \\
&
+O(\varepsilon_v^2)+O(\varepsilon_n)+O(\varepsilon_s)+O(\varepsilon_c)
\, .
\label{eqn:apw3}\end{aligned}$$ Because only the $\varepsilon_v$ contribution is discussed in the following, other terms are not explicitly written in the above equation. This equation indicates that the observed $sin^2 \theta_W$ in neutrino-nucleus scattering is effectively larger if the ratio is calculated without the $\varepsilon_v$ correction.
The nuclear modification difference $\varepsilon_v(x)$ is not known at all at this stage. We try to estimate it theoretically by using charge and baryon-number conservations: $Z = \int dx \, A \sum_q e_q (q^A - \bar q^A)$ and $A = \int dx \, A \sum_q (1/3) \, (q^A - \bar q^A)$. These equations are expressed by the valence-quark distributions, then they becomes $$\begin{aligned}
&
\int dx \, (u_v+d_v) \, [ \, \Delta w_v
+ w_v \, \varepsilon_v (x) \, \varepsilon_n (x) \, ] = 0
\, ,
\label{eqn:b} \\
&
\int dx \, (u_v+d_v) \, [ \, \Delta w_v \,
\{ 1-3 \, \varepsilon_n(x) \} \,
\nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \
- w_v \, \varepsilon_v (x) \,
\{ 3 - \varepsilon_n (x) \} \, ] =0
\, ,
\label{eqn:c}\end{aligned}$$ where $\Delta w_v$ is defined by $\Delta w_v=w_v -1$. These equations suggest that there should exist a finite distribution for $\varepsilon_v(x)$ due to the charge and baryon-number conservations. However, there is no unique solution for these integral equations, so that the following discussions become inevitably model dependent.
We provide two examples for estimating the order of magnitude of the effect on $sin^2 \theta_W$. First, the integrands of Eqs. (\[eqn:b\]) and (\[eqn:c\]) are assumed to vanish by neglecting the higher-order terms $O(\varepsilon_v \varepsilon_n)$: $$\text{case 1:}\ \
\varepsilon_v (x)
= - \varepsilon_n (x) \, \frac{\Delta w_v(x)}{w_v(x)}
\, .
\label{eqn:evx2}$$ Second, the $\chi^2$ analysis result [@hkm], which is explained in section \[npdf\], could be used for the estimation: $$\text{case 2:}\ \
\varepsilon_v (x) = \left [ \frac{w_{d_v}(x)-w_{u_v}(x)}
{w_{d_v}(x)+w_{u_v}(x)}
\right ]_{\text{$\chi^2$ analysis}}
\, .
\label{eqn:ev}$$ These two descriptions are numerically estimated and the results are shown at $Q^2$=20 GeV$^2$ in Fig. \[fig:epsv12\]. The solid and dashes curves indicate the case 1 and 2, respectively.
{width="45.00000%"}
\[fig:epsv12\]
In the first case, the function $\varepsilon_v$ is directly proportional to the nuclear modification $\Delta w_v(x)$, so that it changes the sign at $x \sim 0.2$. In comparison with the NuTeV deviation 0.005, which is shown by the dotted line, $\varepsilon_v^{(1)}$ is of the same order of magnitude. On the other hand, the second one $\varepsilon_v^{(2)}$ is rather small. This is partly because of the assumed functional form in the $\chi^2$ analysis [@hkm], which was not intended especially to obtain the nuclear modification difference $\varepsilon_v$. Because the distributions are much different depending on the model, numerical estimates are merely considered to be an order of magnitude estimate.
![Contributions to $sin^2 \theta_W$ are calculated by taking the $x$ average and they are shown by the solid curves. The dashed curves are calculated by taking the NuTeV kinematics into account [@sk02].](q2dep.eps){width="45.00000%"}
\[fig:q2dep\]
As far as we see Fig. \[fig:epsv12\], our mechanism seems to a promising explanation for the NuTeV anomaly. If a simple $x$ average is taken for the $\varepsilon_v$ contribution to the $sin^2 \theta_W$ determination, we obtain the solid curves in Fig. \[fig:q2dep\], and they are of the order of the NuTeV deviation. However, the situation is not so simple. Although the function $\varepsilon_v^{(1)}$ is large in the large $x$ region in Fig. \[fig:epsv12\], few NuTeV data exist in such a region. It means that the $\varepsilon_v^{(1)}$ contribution to $sin^2 \theta_W$ could be significantly reduced if the NuTeV kinematics is taken into account. The guideline of incorporating such experimental kinematics is supplied in Fig. 1 of Ref. [@sv]. The distribution $\varepsilon_v$ could be effectively simulated by the NuTeV functions, $u_v^p-d_v^p$ and $d_v^p-u_v^p$, although physics motivation is completely different. Using the NuTeV functionals [@sv; @mz], we obtain the dashed curves in Fig. \[fig:q2dep\]. Because of the lack of large $x$ data, the contributions are significantly reduced.
The mechanism due to the nuclear modification difference between $u_v$ and $d_v$ could partially explain the NuTeV deviation, but it is not a major mechanism for the deviation according to Fig. \[fig:q2dep\]. However, the distribution $\varepsilon_v$ itself is not known at all, so that it would be too early to exclude the mechanism. On the other hand, it should be an interesting topic to investigate $\varepsilon_v$ experimentally by the NuMI [@numi] and neutrino-factory [@nufact] projects.
HERMES effect {#hermes}
=============
The HERMES effect indicates nuclear modification of the longitudinal-transverse structure function ratio $R(x,Q^2)$. It was originally reported at small $x$ ($0.01<x<0.03$) with small $Q^2$ ($0.5<Q^2<1$ GeV$^2$) in the HERMES paper in 2000 [@hermes00]. There are theoretical investigations on this topic in terms of shadowing [@shadow] and an isoscalar meson [@mbk00].
This interesting nuclear effect is, however, not observed in the CCFR/NuTeV experiments [@ccfr01]. Although the CCFR/NuTeV target is the iron nucleus, observed R values agree well with theoretical calculations for the nucleon in the same kinematical region with the HERMES. Furthermore, a more careful HERMES analysis of radiative corrections showed such modification does not exist anymore [@hermes02].
Considering these experimental results, one may think that such a nuclear effect does not exist at all. However, we point out that the effect should exist in a different kinematical region, namely at large $x$ with small $Q^2$ [@ek03]. The existence of a nuclear effect in $R(x,Q^2)$ is important not only for investigating nuclear structure in the parton model but also for many analyses of lepton scattering data. For example, the SLAC parametrization in 1990 [@r1990] has been used as a popular one. However, the data contain nuclear ones, so that one cannot use it for nucleon scattering studies if large nuclear effects exist in the data. Because of the importance of $R(x,Q^2)$ in lepton scattering analyses, we investigate the possibility of nuclear modification theoretically.
The structure function for the photon polarization $\lambda$ is $W^{A,N}_\lambda = \varepsilon_\lambda^{\mu *} \,
\varepsilon_\lambda^\nu \, W^{A,N}_{\mu\nu}$, so that longitudinal and transverse ones are defined by $ W^{A,N}_T = ( W^{A,N}_{+1} + W^{A,N}_{-1} ) /2$ and $ W^{A,N}_L = W^{A,N}_0 $. Here, $N$ and $A$ denote the nucleon and a nucleus, respectively. Lepton-hadron scattering cross section is described by a lepton tensor multiplied by a hadron tensor $W_{\mu \nu}$. In the electron scattering, the tensors for the nucleon and a nucleus are given by $$\begin{aligned}
W^{A,N}_{\mu\nu} (p_{_{A,N}}, & q) =
- W^{A,N}_1 (p_{_{A,N}}, q)
\left ( g_{\mu\nu} - \frac{q_\mu q_\nu}{q^2} \right )
\nonumber \\
& + W^{A,N}_2 (p_{_{A,N}}, q) \, \frac{\pt_{_{A,N} \mu}
\, \pt_{_{A,N} \nu}}{p_{_{A,N}}^2}
\ ,
\label{eqn:hadron}\end{aligned}$$ where $\pt_{\mu} = p_{\mu} -(p \cdot q) \, q_\mu /q^2$. In terms of these structure functions, the longitudinal and transverse structure functions are given $ W^{A,N}_T = W^{A,N}_1 $ and $ W^{A,N}_L = (1+\nu_{_{A,N}}^2/Q^2) W^{A,N}_2 - W^{A,N}_1 $ by taking the nucleus or nucleon rest frame. Here, $\nu_A \equiv \nu$, and the photon momentum in the nucleon rest frame is denoted $(\nu_N,\vec q_N)$ with $\nu_N^2 = (p_N \cdot q)^2 /p_N^2$.
We use a conventional convolution description for nuclear structure functions: $$W^A_{\mu\nu} (\pa, q) = \int d^4 \pn \, S(\pn) \, W^N_{\mu\nu} (\pn, q)
\ ,
\label{eqn:conv}$$ where $p_N$ is the nucleon momentum and $S(p_N)$ is the spectral function which indicates the nucleon momentum distribution in a nucleus. In order to investigate the longitudinal and transverse components, we introduce projection operators which satisfy $\widehat P_1^{\, \mu\nu} W^A_{\mu\nu} = W_1^A$ and $\widehat P_2^{\, \mu\nu} W^A_{\mu\nu} = W_2^A$. They are explicitly written as $ \widehat P_1^{\, \mu\nu} = - (1/2)
\left ( g^{\mu\nu} - \pt_A^{\, \mu} \, \pt_A^{\, \nu}
/ \pt_A^{\, 2} \right ) $ and $ \widehat P_2^{\, \mu\nu} = - p_A^2 / (2\, \pt_A^{\, 2})
\left ( g^{\mu\nu} - 3 \, \pt_A^{\, \mu} \, \pt_A^{\, \nu}
/ \pt_A^{\, 2} \right ) $. Instead of $W_1$ and $W_2$ structure functions, the functions $F_1$ and $F_2$ are usually used: $F_1^{A,N} = \sqrt{p_{_{A,N}}^2} \, W_1^{A,N}$ and $F_2^{A,N} = ( p_{_{A,N}} \cdot q / \sqrt{p_{_{A,N}}^2}) \, W_2^{A,N} $. Then, the longitudinal structure function is given by $$\begin{aligned}
F_L^{A,N} (x_{_{A,N}}, Q^2) & = \bigg ( 1 + \frac{Q^2}{\nu_{_{A,N}}^2} \bigg )
F_2^{A,N} (x_{_{A,N}}, Q^2)
\nonumber \\
& - 2 x_{_{A,N}} F_1^{A,N} (x_{_{A,N}}, Q^2)
\, ,
\label{eqn:flla}\end{aligned}$$ where $x_A = Q^2 /(2 \, p_A \cdot q)$ and $x_N = Q^2 /(2 \, p_N \cdot q)$. The ratio $R_A$ of the longitudinal cross section to the transverse one is expressed by the function $R_A(x_A,Q^2)$: $$R_A (x_A, Q^2) = \frac{F_L^A (x_A, Q^2)}{2 \, x_A F_1^A (x_A, Q^2)}
\ .$$
Applying the projection operators $\widehat P_1^{\, \mu\nu}$ and $\widehat P_2^{\, \mu\nu}$ to Eq. (\[eqn:conv\]), we have $$\begin{aligned}
& \! \! \!
2 \, x_A F_1^A (x_A, Q^2) = \int d^4 \, p_N \, S(p_N) \, z \,
\frac{M_N}{\sqrt{p_N^2}}
\nonumber \\
& \! \! \! \!
\times
\bigg [ \bigg ( 1
+ \frac{\vec p_{N\perp}^{\ 2}}{2 \, \pt_N^{\, 2}} \bigg )
2 x_N F_1^N (x_N, Q^2)
+ \frac{\vec p_{N\perp}^{\ 2}}{2 \pt_N^{\, 2}}
F_L^N (x_N, Q^2) \bigg ]
\, ,
\label{eqn:trans}
\\
& \! \! \!
F_L^A (x_A, Q^2) = \int d^4 \, p_N \, S(p_N) \, z \,
\frac{M_N}{\sqrt{p_N^2}}
\nonumber \\
& \! \! \! \!
\times
\bigg [ \bigg ( 1
+ \frac{\vec p_{N\perp}^{\ 2}}{\pt_N^{\, 2}} \bigg )
F_L^N (x_N, Q^2)
+ \frac{\vec p_{N\perp}^{\ 2}}{\pt_N^{\, 2}}
2 x_N F_1^N (x_N, Q^2) \bigg ]
\, .
\label{eqn:longi}\end{aligned}$$ These results are interesting. The transverse structure function for a nucleus is described not only by the transverse one for the nucleon but also by the longitudinal one with the admixture coefficient $\vec p_{N\perp}^{\ 2}/(2 \pt_N^{\, 2})$. The $\vec p_{N\perp}$ is the nucleon momentum component perpendicular to the photon one $\vec q$. Equations (\[eqn:trans\]) and (\[eqn:longi\]) indicate that the transverse-longitudinal admixture exists because the nucleon momentum direction is not necessary along the virtual photon direction.
These expressions are numerically estimated for the nitrogen nucleus by taking a simple shell model for the spectral function with density dependent Hartree-Fock wave functions. Parton distribution functions are taken from the MRST-1998 version and the nucleonic $R(x,Q^2)$ is taken from the SLAC analysis in 1990 [@r1990]. The nitrogen-nucleon ratios $R_{^{14}N}/R_N$ are shown at $Q^2$=1, 10, 100 GeV$^2$ by the solid curves in Fig. \[fig:rratio03\]. In order to clarify the admixture effects, the ratios are also calculated by suppressing the $\vec p_{N\perp}^{\ 2}$ terms, and the results are shown by the dashed curves. In addition, the nuclear modification is calculated at $Q^2$=0.5 GeV$^2$ by using the GRV94 parametrization for the PDFs. It is intended to find the modification magnitude at smaller $Q^2$, where JLab experiments could possibly probe [@jlab]. Because the admixture is proportional to $\vec p_{N\perp}^{\ 2}/(2 \pt_N^{\, 2}) \sim \vec p_{N\perp}^{\ 2}/Q^2$, the modification effects are large at small $Q^2$ (=0.5 $-$ 1 GeV$^2$) and they become small at large $Q^2$. However, the modification does not vanish even at $Q^2$=100 GeV$^2$ due to the Fermi-motion and binding effects which are contained implicitly in the spectral function.
{width="45.00000%"}
\[fig:rratio03\]
In this way, we found that the nucleon Fermi motion, especially the perpendicular motion to the virtual photon direction, and the nuclear binding give rise to the nuclear modification of the longitudinal-transverse ratio $R(x,Q^2)$. However, nuclear modification of $R$ in the large $x$ region with small $Q^2$ has not been investigated experimentally. The situation is clearly illustrated in Fig. 3 of Ref. [@ccfr01], where the data does not exist at $x=0.5$ with $Q^2 \approx$1 GeV$^2$. We hope that future measurements, for example those of JLab experiments [@jlab], are able to provide clear information on the nuclear modification in this region.
Summary {#sum}
=======
Current neutrino scattering experiments are done with nuclear targets, so that precise nuclear corrections should be taken into account in order to investigate underlying elementary processes, for example neutrino oscillation phenomena. In this paper, the discussions are focused on high-energy reactions.
First, the optimum nuclear parton distribution functions were determined by the $\chi^2$ analysis of DIS and Drell-Yan data. They could be used for calculating high-energy nuclear cross sections.
Second, a possibility of explaining the NuTeV $sin^2 \theta_W$ was investigated by the nuclear correction difference between $u_v$ and $d_v$ in the iron nucleus. Although the contribution to the $sin^2 \theta_W$ deviation may not be large at this stage, the distribution $\varepsilon_v (x)$ should be investigated by future experiments.
Third, a possible HERMES-type effect was proposed in the medium and large $x$ regions due to the nucleon Fermi motion and binding. Especially, we found that the perpendicular nucleon motion to the virtual photon direction gives rise to the admixture of longitudinal and transverse structure functions in the nucleon. Such an effect should be tested by electron and neutrino scattering experiments at large $x$ with small $Q^2$.
Acknowledgments {#acknowledgments .unnumbered}
===============
S.K. was supported by the Grant-in-Aid for Scientific Research from the Japanese Ministry of Education, Culture, Sports, Science, and Technology. He thanks M. Sakuda for his financial support for participating in this workshop.
[9]{} http://nuint.ps.uci.edu. M. Sakuda, http://nuint.ps.uci.edu/slides/ Sakuda.pdf; in proceedings of this workshop. E. A. Paschos and J. Y. Yu, Phys. Rev. D65 (2002) 033002. Shi-yuan Li and Xin-Nian Wang, Phys. Lett. B527 (2002) 85; X. Zhang and G. Fai, Phys. Rev. C65 (2002) 064901; A. Chamblin and G. C. Nayak, Phys. Rev. D66 (2002) 091901. B. Z. Kopeliovich and A. V. Tarasov, Nucl. Phys. A710 (2002) 180; L. Frankfurt, V. Guzey, and M. Strikman, hep-ph/0303022; N. Armesto [*et. al.*]{}, hep-ph/0304119. http://durpdg.dur.ac.uk/hepdata/pdf.html. K. J. Eskola, V. J. Kolhinen, and P. V. Ruuskanen, Nucl. Phys. B535 (1998) 351; K. J. Eskola, V. J. Kolhinen, and C. A. Salgado, Eur. Phys. J. C9 (1999) 61. M. Hirai, S. Kumano, and M. Miyama, Phys. Rev. D64 (2001) 034003; research in progress. See http://hs.phys.saga-u.ac.jp/nuclp.html. G. P. Zeller [*et. al.*]{}, Phys. Rev. Lett. 88 (2002) 091802. K. S. McFarland [*et. al.*]{}, Nucl. Phys. B 112 (2002) 226. G. A. Miller and A. W. Thomas, hep-ex/0204007; G. P. Zeller [*et. al.*]{}, hep-ex/0207052. W. Melnitchouk and A. W. Thomas, Phys. Rev. C67 (2003) 038201; S. Kovalenko, I. Schmidt, and J.-J. Yang, Phys. Lett. B546 (2002) 68. S. Kumano, Phys. Rev. D66 (2002) 111301. S. A. Kulagin, Phys. Rev. D67 (2003) 091301. S. Davidson [*et. al.*]{}, J. High Energy Phys. 0202, 037 (2002); E. Ma and D. P. Roy, Phys. Rev. D65 (2002) 075021; C. Giunti and M. Laveder, hep-ph/0202152; W. Loinaz, N. Okamura, T. Takeuchi, and L. C. R. Wijewardhana, Phys. Rev. D67 (2003) 073012. K. Ackerstaff [*et al.*]{}, Phys. Lett. B475 (2000) 386. M. Ericson and S. Kumano, Phys. Rev. C 67 (2003) 022201. U. K. Yang [*et al.*]{}, Phys. Rev. Lett. 87 (2001) 251802. A. Airapetian [*et al.*]{}, hep-ex/0210067 & 0210068. S. Kumano, Phys. Rep. 303 (1998) 183; G. T. Garvey and J.-C. Peng, Prog. Part. Nucl. Phys. 47 (2001) 203. J. G. Morfin, Nucl. Phys. B112 (2002) 251. D. Abbaneo [*et. al.*]{}, hep-ex/0112021. See also the reference \[21\] in Ref. [@nutev02]. E. A. Paschos and L. Wolfenstein, Phys. Rev. D7 (1973) 91. G. P. Zeller [*et. al.*]{}, Phys. Rev. D65 (2002) 111103 . K. S. McFarland and G. P. Zeller, personal communications. http://www.cap.bnl.gov/nufact03/. V. Barone and M. Genovese, hep-ph/9610206; B. Kopeliovich, J. Raufeisen, and A. Tarasov, Phys. Rev. C62 (2000) 035204. G. A. Miller, S. J. Brodsky, and M. Karliner, Phys. Lett. B481 (2000) 245; G. A. Miller, Phys. Rev. C64 (2001) 022201. L. W. Whitlow, S. Rock, A. Bodek, S. Dasu, and E. M. Riordan, Phys. Lett. B250 (1990) 193; L. W. Whitlow, report SLAC-357 (1990). H. P. Blok, personal communications. A. Brüll [*et al.*]{}, http://www.jlab.org/exp\_prog /proposals/99/PR99-118.pdf.
[^1]: [email protected], http://hs.phys.saga-u.ac.jp
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Parallel algorithms and simulators with good scalabilities are particularly important for large-scale reservoir simulations on modern supercomputers with a large number of processors. In this paper, we introduce and study a family of highly scalable multilevel restricted additive Schwarz (RAS) methods for the fully implicit solution of subsurface flows with Peng-Robinson equation of state in two and three dimensions. With the use of a second-order fully implicit scheme, the proposed simulator is unconditionally stable with the relaxation of the time step size by the stability condition. The investigation then focuses on the development of several types of multilevel overlapping additive Schwarz methods for the preconditioning of the resultant linear system arising from the inexact Newton iteration, and some fast solver technologies are presented for the assurance of the multilevel approach efficiency and scalability. We numerically show that the proposed fully implicit framework is highly efficient for solving both standard benchmarks as well as realistic problems with several hundreds of millions of unknowns and scalable to 8192 processors on the Tianhe-2 supercomputer.'
address:
- 'School of Mathematics, Hunan University, Changsha, Hunan 410082, PR China'
- 'School of Mathematical Sciences, Peking University, Beijing 100871, PR China'
author:
- Rui Li
- Haijian Yang
- Chao Yang
title: 'Parallel multilevel restricted Schwarz preconditioners for implicit simulation of subsurface flows with Peng-Robinson equation of state'
---
Reservoir simulation ,Fully implicit method ,Multilevel method ,Restricted Schwarz preconditioners ,Parallel computing
Introduction {#sc-int}
============
Simulation of subsurface fluid flows in porous media is currently an important topic of interest in many applications, such as hydrology and groundwater flow, oil and gas reservoir simulation, CO$_2$ sequestration, and waste management [@Chen; @Firoozabadi]. The extensive growing demand on accurate modeling subsurface systems has produced persistent requirements on the understanding of many different physical processes, including multiphase fluid flow and transport, fluid phase behavior with sophisticated equation of state, and geomechanical deformations. To simulate these processes, one needs to solve a set of coupled, nonlinear, time-dependent partial differential equations (PDEs) with complicated nonlinear behaviors arising from the complexity of the geological media and reservoir properties. Additional computational challenge comes from the involving of Peng-Robinson equation of state, which governs the equilibrium distribution of fluid and has a remarkable influence on the high variation of the Darcy velocity. Due to their high computational complexity, numerically approximating subsurface phenomena with high resolution is critical to the reservoir engineering for accurate predictions of costly projects. Hence, parallel reservoir simulators equipped with robust and scalable solvers on high performance computing platforms are crucial to achieve efficient simulations of this intricate problem [@Chen07].
In terms of the numerical solution of partial differential equations, the fully implicit method [@bui17; @bui18; @Lacroix; @Liu2018; @qiao; @Sko13; @wang; @haijian-cmame] is a widely preferred approach to solve various problems arising [from the discretization of subsurface fluid flows]{}. The fully implicit scheme is unconditionally stable and can relax the stability requirement of the Courant-Friedrichs-Lewy (CFL) condition. It therefore provides a consistent and robust way to simulate the subsurface fluid flow problem in long-term and extreme-scale computations. In particular, a great advantage of the fully implicit algorithm is that the corresponding nonlinear equations are implicitly solved in a monolithic way. This characteristic feature strengthens its potential to allow the addition of more physics and the introduction of more equations without changing much of the simulator framework, which greatly expands the scope of application of the fully implicit approach. In spite of being stable for arbitrary large time steps, when a fully implicit scheme is applied, one must solve a large, sparse, and ill-conditioned linear system at each nonlinear iteration. It remains challenging and important to design robust and scalable solvers for the large scale simulations on high performance computing platforms. In this study, we employ the framework of Newton–Krylov algorithms [@Cai98; @cs02a; @fim5; @Knoll04; @liu; @haijian-sisc1] to guarantee the nonlinear consistency, and mainly focus our efforts on designing an efficient preconditioning strategy to substantially reduce the condition number of the linear system.
In reservoir simulations, some effective preconditioning techniques, such as block preconditioners [@Haga; @Lee; @White], the Constrained Residual Pressure (CPR) preconditioning approach [@Wallis; @Wallis85; @wang], the domain decomposition methods [@Liu2018; @Sko13; @haijian-cmame; @haijian-jcp19], and the algebraic or geometric multigrid algorithms [@bui17; @bui18; @qiao; @wang], are flexible and capable of addressing the inherent ill-conditioning of a complicated physics system, and hence have received increasingly more attentions in recent years. The focus of this paper is on the domain decomposition approach by which we propose a family of one-level or multilevel restricted additive Schwarz preconditioners. The original overlapping additive Schwarz (AS) method was introduced for the solution of symmetric positive definite elliptic finite element problems, and was later extended to many other nonlinear or linear systems [@cs99; @Toselli]. In the AS method, it follows the divide-and-conquer technique by recursively breaking down a problem into more sub-problems of the same or related type, and its communication only occurs between neighboring subdomains during the restriction and extension processes. Moreover, the additive Schwarz preconditioner does not require any splitting of the nonlinear system. Hence, it can serve as a basis of efficient algorithms that are naturally suitable for massively parallel computing. In particular, the approach can be combined with certain variants of restriction operators in a robust and efficient way, such as the so-called restricted additive Schwarz (RAS) method proposed by Xiao-Chuan Cai et al. in [@cs03; @cs99]. Hence, it can significantly improve the physical feasibility of the solver and substantial reduce the total computing time, and has proven to be very efficient in a variety of applications [@kong; @Prudencio; @Sko13; @wang; @haijian-cmame; @haijian-jcp]. Recently, the overlapping restricted additive Schwarz method in conjunction with a variant of inexact Newton methods has been applied successfully to the two-phase flow problem [@haijian-cmame] and the black oil model [@wang; @haijian-jcp19]. It was demonstrated that the one-level Schwarz preconditioner is scalable in terms of the numbers of nonlinear and linear iterations, but not in terms of the total compute time [when the number of processors becomes large]{}, which means the requirement of the family of multilevel methods [@kong; @Prudencio; @sbg96; @Toselli].
In this work, to take advantage of modern supercomputers for large-scale reservoir simulations, we propose and develop the overlapping restricted additive Schwarz preconditioner into a general multilevel framework for solving discrete systems coming from fully implicit discretizations. We show that, with this new feature, our proposed approach is more robust and efficient for problems with highly heterogeneous media, and can scale optimally to a number of processors on the Tianhe-2 supercomputer. We would like to pointing out that designing a good strategy for the multilevel approach is both time-consuming and challenging, as it requires extensive knowledge of the general-purpose framework of interests, such as the selection of coarse-to-fine mesh ratios, the restriction and interpolation operators, and the solvers for the smoothing and coarse-grid correction. To the best of our knowledge, very limited research has been conducted to apply the multilevel restricted additive Schwarz preconditioning technique for the fully implicit solution of petroleum reservoirs.
The outline of the paper is as follows. In Section \[Model\], we introduce the governing equations of subsurface flows with Peng-Robinson equation of state, followed by the corresponding fully implicit discretization. In Section \[two-RAS\], a family of multilevel restricted additive Schwarz preconditioners, as the most important part of the fully implicit solver, is presented in detail. We show numerical results for some 2D and 3D realistic benchmark problems in Section \[Numerical-scheme\] to demonstrate the robustness and scalability of the proposed preconditioner. The paper is concluded in Section \[conclusion\].
Mathematical model and discretizations {#Model}
======================================
The problem of interest in this study is the compressible Darcy flow in a porous medium for the description of gas reservoir simulations [@Chen; @Chen07; @Firoozabadi]. Let $\Omega\in \mathbb{R}^d$ $(d =1,2,3)$ be the computational domain with the boundary $\Gamma=\partial \Omega$, and $t_f$ denote the final simulated time. The mass balance equation with the Darcy law for the real gas fluid is defined by $$\label{mass conservation}
\left\{
\begin{array}{ll}
\displaystyle \frac{\partial(\phi\rho)}{\partial{t}}+\nabla \cdot (\rho \mathbf{v})= q, ~\mbox{in~} \Omega_{T}=\Omega\times (0, t_f], \\
\displaystyle \mathbf{v}=-\frac{\mathbf{K}}{\mu}(\nabla p -\rho g\nabla h), \\
\end{array}
\right.$$ where $\phi$ is the porosity of the porous medium, $\mathbf{K}$ is the permeability tensor, $\mathbf{v}$ is the Darcy’s velocity, $\rho$ denotes the density, $\mu$ is the viscosity of the flow, $p$ is the pressure, and $q$ is the external mass flow rate. The Peng-Robinson (PR) equation of state (EOS) [@Firoozabadi] is used to describe the density as a function of the composition, temperature and pressure: $$\label{equ:density}
\left\{
\begin{array}{ll}
\rho=cW,\\
c=\displaystyle\frac{p}{ZRT},\\
\end{array}
\right.$$ where $W$ is the molecular weight, $T$ is the temperature, $R$ is the gas constant, and $Z$ is the compressibility factor of gas defined by $$\label{equ:PR}
Z^3-(1-B)Z^2+(A-3B^2-2B)Z-(AB-B^2-B^3)=0.$$ Moreover, $A$ and $B$ are the PR-EOS parameters defined as follows: $$\displaystyle A=\frac{ap}{R^2T^2}, ~ B=\frac{bp}{RT},$$ where the parameters $a$ and $b$ are modeled by imposing the criticality conditions $$\label{attrfactor}
\left\{
\begin{array}{ll}
\displaystyle a=a(T)=0.45724\cdot\frac{R^2T_c^2}{p_c}\left(1+m\left(1-\sqrt{T_r}\right)\right)^2,\\
\displaystyle b=0.0778\cdot\frac{RT_c}{p_c},
\end{array}
\right.$$ with $T_c$ ($p_c$) being the specific pressure (temperature) in the critical state of the gas, and $T_r$ being the reduced temperature defined as ${T}/{T_c}$. In addition, the parameter $m$ in is a fitting formula of the acentric factor $w$ of the substance: $$m= \left\{
\begin{aligned}
0.37464+1.54226w-0.26992w^2 \quad 0\leq w\leq 0.491,\\
0.3796+1.485w-0.1644w^2+0.01667w^3 \quad 0.491\leq w\leq 2.0.\\
\end{aligned}
\right.$$ Here, the acentric factor $w$ is calculated by the following formula: $$w= \frac{-ln(\frac{p_c}{1\;atm})-f^{(0)}(T_{br})}{f^{(1)}(T_{br})},$$ where $f^0(T_{br})$ and $f^1(T_{br})$ are given by $$\left\{
\begin{aligned}
f^0(T_{br})&=\frac{-5.97616(1-T_{br})+1.29874(1-T_{br})^{\frac{3}{2}}-0.60394(1-T_{br})^{\frac{5}{2}}-1.06841(1-T_{br})^5}{T_{br}},\\
f^1(T_{br})&=\frac{-5.03365(1-T_{br})+1.11505(1-T_{br})^{\frac{ 3}{2}}-5.41217(1-T_{br})^{\frac{5}{2}}- 7.46628(1-T_{br})^5}{T_{br}},
\end{aligned}
\right.$$ with the normal boiling point temperature $T_b$ and the reduced normal boiling point temperature defined as $T_{br}={T_b}/{T_c}$.
In this study, we take the pressure $p$ as the primary variable. Suppose that the porosity $\phi$ is not changed with the time, then the equations of the mathematical model can be rewritten by substituting and into as follows: $$\label{gover_sys}
\frac{\phi p}{Z}\left(\frac{1}{p}-\frac{1}{Z}\frac{\partial Z}{\partial p}\right)\frac{\partial p}{\partial t}- \nabla \cdot \left(\frac{\mathbf{K}p}{\mu Z} \left(\nabla p -\frac{pW}{ZRT} g\nabla h\right)\right)- \frac{qRT}{W}=0,$$ where the relations are given by $$\left\{
\begin{aligned}
&\frac{\partial Z}{\partial p}=-\frac{b}{RTC}Z^2-
\left(\frac{ a}{R^2T^2C}-\frac{2b+6Bb}{RTC}\right)Z+\left(\frac{2 abp}{R^3T^3C}-\frac{2Bb+3B^2b}{RTC}\right),\\
&C=3Z^2-2(1-B)Z +(A-2B-3B^2),\\
&Z^3-(1-B)Z^2+(A-2B-3B^2)Z-(AB-B^2-B^3)=0,
\end{aligned}
\right.$$ with the initial condition $p=p_0$. Suppose the boundary of the computational domain $\Omega$ is composed of two parts $\partial\Omega=\Gamma^{D}+\Gamma^{N}$ with $\Gamma^{D}\cap\Gamma^{N}=\emptyset$. The boundary conditions associated to the model problem are $$\label{boundary}
\left\{\begin{array}{ll}
p= p^{D} &\mbox{on}\, \Gamma^D,\\
\mathbf{v}\cdot\mathbf{n}=q^{N} &\mbox{on}\, \Gamma^N,\\
\end{array}\right.$$ where $\textbf{n}$ is the outward normal of the boundary $\partial\Omega$. We remark that the compressibility factor $Z$ as an intermediate variable is obtained by solving the algebraic cubic equation with the primary variable $p$, see the references [@Chen; @Chen07] for the computation of $Z$ in details.
We employ a cell-centered finite difference (CCFD) method for the spatial discretization, for which the details can be found in [@Chen; @Chen07; @Monteagudo07; @haijian-siam], and then a fully implicit scheme is applied for the time integration [@bui17; @kong; @wang; @haijian-cmame]. For a given time-stepping sequence $0=t^{(1)}<t^{(2)}<...$, define the time step size $\Delta t^{(l)}=t^{(l+1)}-t^{(l)}$ and use superscript $(l)$ to denote the discretized evaluation at time level $t=t^{(l)}$. After spatially discretizing by the CCFD scheme, we have a semi-discrete system $$\label{semi-discretized-system}
\frac{\phi p}{Z}\left(\frac{1}{p}-\frac{1}{Z}\frac{\partial Z}{\partial p}\right)\frac{\partial p}{\partial t}+\mathcal{F}\left(p^{(l+1)}\right)=0,$$ where $\mathcal{F}\left(p^{(l+1)}\right)$ is the operator of the spatial discretization at time level $t=t^{(l+1)}$ by using the CCFD method. For the purpose of comparison, we implement both the first-order backward Euler scheme (BDF-1) and second-order backward differentiation formula (BDF-2) for the temporal integration of . For the BDF-1 scheme, the fully discretized system reads $$\label{implicit-method-bdf1}
\left[\frac{\phi p}{Z}\left(\frac{1}{p}-\frac{1}{Z}\frac{\partial Z}{\partial p}\right)\right]^{(l+1)} \frac{p^{(l+1)}-p^{(l)}}{\triangle t^{(l)}}+\mathcal{F}\left(p^{(l+1)}\right)=0,$$ where $p^{(l)}$ is the evaluation of $p$ at the $l^{th}$ time step with a time step size $\triangle t^{(l)}$. And the fully discretized system for the BDF-2 method is $$\label{implicit-method-bdf2}
\left[\frac{\phi p}{Z}\left(\frac{1}{p}-\frac{1}{Z}\frac{\partial Z}{\partial p}\right)\right]^{(l+1)} \frac{3p^{(l+1)}-4p^{(l)}+p^{(l-1)}}{2\triangle t^{(l)}}+\mathcal{F}\left(p^{(l+1)}\right)=0.$$ Note that, for the BDF-2 scheme, the BDF-1 method is used at the initial time step.
The employed CCFD discretization for on rectangular meshes can be viewed as the mixed finite element method with Raviart-Thomas basis functions of lowest order equipped with the trapezoidal quadrature rule [@Chen; @Chen07; @Monteagudo07]. We also would like to pointing out that the proposed solver technologies introduced in the following section can be extended to solve the systems arising from the other spatial discretization schemes.
Multilevel restricted Schwarz preconditioners {#two-RAS}
=============================================
When the fully implicit method is applied, a system of nonlinear algebraic equations after the temporal and spatial discretizations, $$\label{equ:nonlinear-system}
F(X)=0,$$ is constructed and solved at each time step. In , the vector $F(X): \mathbb{R}^n\rightarrow \mathbb{R}^n$ is a given nonlinear vector-valued function arising from the residuals function in or given by $F(X)=(F_1, F_2, \cdots, F_n)^{T}$ with $F_i=F_i(X_1, X_2,\cdots, X_n)^{T}$, and $X=(X_1, X_2,\cdots, X_n)^{T}$. In this study, the nonlinear system is solved by a parallel Newton-Krylov method with the family of domain decomposition type preconditioners [@Cai98; @cs02a; @fim5; @Knoll04].
Let the initial guess for Newton iterations $X^{0}\in\mathbb{R}^n$ at current time step be the solution of the previous time step, then the new solution $X^{k+1}$ is updated by the current solution $X^{k}$ and a Newton correction vector $s^{k}$ as follows: $$X^{k+1}=X^{k}+\lambda^{k}s^{k}, \quad k=0,1,2,\cdots,
\label{newton_iterate}$$ [ where $\lambda^{k}$ is the step length calculated by a cubic line search method [@ds] to satisfy $$\label{eq:line-search}
\begin{array}{ll}
\left\| F(X^k+\lambda^{k} s^{k})\right\|\leq (1-\alpha\lambda^{k})\left\|F(X^k)\right\|,
\end{array}$$ with the parameter $\alpha$ being fixed to $10^{-4}$ in practice]{}. And the Newton correction $s^{k}$ is obtained by solving the following Jacobian linear system $$\label{jacobian1}
J_k s^k =-F(X^k),$$ with a Krylov subspace iteration method [@Sa03]. Here, $J_k=\nabla F(X^k)$ is the Jacobian matrix obtained from the current solution $X^{k}$. Since the corresponding linear system from the Newton iteration is large sparse and ill-conditioned, the family of one-level or multilevel restricted Schwarz preconditioners is taken into account to accelerate the convergence of the linear system, i.e., we solve the following right-preconditioned Jacobian system $$\label{jacobian}
J_k M_k^{-1} M_k s^k =-F(X^k),$$ with an overlapping Schwarz preconditioned Krylov subspace method. In practice, the Generalized Minimal RESidual (GMRES) algorithm as a Krylov subspace method is used for the linear solver.
The accuracy of approximation is controlled by the relative and absolute tolerances $\eta^r_k$ and $\eta^a_k$ until the following convergence criterion is satisfied $$\label{equ:linear_stop}
\| J_k s^k +F(X^k)\| \leq \min\{\eta^r_k \| F(X^k)\|,\eta^a_k\}.$$ We set the stopping conditions for the Newton iteration as $$\label{equ:nonlinear_stop}
||{F}(X^{k+1})|| \leq \max\{ \varepsilon_{r}
||{F}(X^{0})||, \varepsilon_{a} \},$$ with the relative $\varepsilon_{r}$ and absolute $\varepsilon_{a}$ solver tolerances, respectively.
The most important component of a robust and scalable solver for solving the nonlinear or linear system is the choice of suitable preconditioners. In large-scale parallel computing, the additive Schwarz (AS) preconditioner, as a family of domain decomposition methods, can help to improve the convergence and meanwhile is beneficial to the scalability of the linear solver, which is the main focus of this paper.
One-level Schwarz preconditioners
---------------------------------
Let $\Omega\subset \mathbb{R}^d$, $d=1,2,3$, be the computational domain on which the PDE system are defined. Then a spatial discretization is performed with a mesh $\Omega_h$ of the characteristic size $h>0$. To define the one-level Schwarz preconditioner, we first divide $\Omega_h$ into nonoverlapping subdomains $\Omega_i$, $i=1,\cdots,N_p$, and then expand each $\Omega_i$ to $\Omega_i^{\delta}$, i.e., $\Omega_i\subset\Omega_i^{\delta}$, to obtain the overlapping partition. Here, the parameter $\delta>0$ is the size of the overlap defined as the minimum distance between $\partial \Omega_i^{\delta}$ and $\partial \Omega_i$, in the interior of $\Omega$. For boundary subdomains we simply cut off the part outside $\Omega$. We also denote $H>0$ as the characteristic diameter of $\Omega_i$, as shown in the left panel of Figure \[fig:partition-domain\].
(0,0) rectangle (4,4); (0,0) rectangle (4,4); (0.02,0.02) grid (3.98,3.98); (1.2,0)node\[right\][$\mathsf{~ }$]{} to (1.2,4); (2.8,0)node\[right\][$\mathsf{~ }$]{} to (2.8,4); (0,1.2)node\[right\][$\mathsf{~ }$]{} to (4,1.2); (0,2.8)node\[right\][$\mathsf{~ }$]{} to (4,2.8);
(1.2,1.2) rectangle (2.8,2.8); (0.8,0.8) rectangle (3.2,1.2); (0.8,2.8) rectangle (3.2,3.2); (0.8,1.2) rectangle (1.2,2.8); (2.8,1.2) rectangle (3.2,2.8);
(0.8,0.8)node\[right\][$\mathsf{~ }$]{} to (0.8,3.2); (0.8,0.8)node\[right\][$\mathsf{~ }$]{} to (3.2,0.8); (3.2,0.8)node\[right\][$\mathsf{~ }$]{} to (3.2,3.2); (0.8,3.2)node\[right\][$\mathsf{~ }$]{} to (3.2,3.2);
(1.2,0.8) – (2.8,0.8); (0,-46)[(1.0,1.0)[$H_x$]{}]{} (0.8,0.8) – (0.8,1.2); (-45,-32)[(1.0,1.0)[$\delta$]{}]{}
(0,0) rectangle (4,4); (0,0) rectangle (4,4); (0.02,0.02) grid (3.98,3.98);
(1.2,0)node\[right\][$\mathsf{~ }$]{} to (1.2,4); (2.8,0)node\[right\][$\mathsf{~ }$]{} to (2.8,4); (0,1.2)node\[right\][$\mathsf{~ }$]{} to (4,1.2); (0,2.8)node\[right\][$\mathsf{~ }$]{} to (4,2.8);
(1.2,1.2) rectangle (2.8,2.8); (0.8,0.8) rectangle (3.2,1.2); (0.8,2.8) rectangle (3.2,3.2); (0.8,1.2) rectangle (1.2,2.8); (2.8,1.2) rectangle (3.2,2.8);
(0.8,0.8)node\[right\][$\mathsf{~ }$]{} to (0.8,3.2); (0.8,0.8)node\[right\][$\mathsf{~ }$]{} to (3.2,0.8); (3.2,0.8)node\[right\][$\mathsf{~ }$]{} to (3.2,3.2); (0.8,3.2)node\[right\][$\mathsf{~ }$]{} to (3.2,3.2);
(0,-70)[(1.0,1.0)[$x$]{}]{} (-70,0)[(1.0,1.0)[$z$]{}]{} (100,-30)[(1.0,1.0)[$y$]{}]{}
(0,0) rectangle (4,4); (0,0) rectangle (4,4); (0,1.2)node\[right\][$\mathsf{~ }$]{} to (4,1.2); (0,2.8)node\[right\][$\mathsf{~ }$]{} to (4,2.8); (1.2,0)node\[right\][$\mathsf{~ }$]{} to (1.2,4); (2.8,0)node\[right\][$\mathsf{~ }$]{} to (2.8,4);
(0,0) rectangle (4,4); (0,0) rectangle (4,4); (0.02,0.02) grid (3.98,3.98); (1.2,0)node\[right\][$\mathsf{~ }$]{} to (1.2,4); (2.8,0)node\[right\][$\mathsf{~ }$]{} to (2.8,4); (0,1.2)node\[right\][$\mathsf{~ }$]{} to (4,1.2); (0,2.8)node\[right\][$\mathsf{~ }$]{} to (4,2.8);
(1.2,1.2) rectangle (2.8,2.8); (0.4,0.4) rectangle (3.6,1.2); (0.4,2.8) rectangle (3.6,3.6); (0.4,1.2) rectangle (1.2,2.8); (2.8,1.2) rectangle (3.6,2.8);
(0.4,0.4)node\[right\][$\mathsf{~ }$]{} to (0.4,3.6); (0.4,0.4)node\[right\][$\mathsf{~ }$]{} to (3.6,0.4); (3.6,0.4)node\[right\][$\mathsf{~ }$]{} to (3.6,3.6); (0.4,3.6)node\[right\][$\mathsf{~ }$]{} to (3.6,3.6);
(0,0) rectangle (4,4); (0,0) rectangle (4,4); (0.02,0.02) grid (3.98,3.98);
(1.2,0)node\[right\][$\mathsf{~ }$]{} to (1.2,4); (2.8,0)node\[right\][$\mathsf{~ }$]{} to (2.8,4); (0,1.2)node\[right\][$\mathsf{~ }$]{} to (4,1.2); (0,2.8)node\[right\][$\mathsf{~ }$]{} to (4,2.8);
(1.2,1.2) rectangle (2.8,2.8); (0.4,0.4) rectangle (3.6,1.2); (0.4,2.8) rectangle (3.6,3.6); (0.4,1.2) rectangle (1.2,2.8); (2.8,1.2) rectangle (3.6,2.8);
(0.4,0.4)node\[right\][$\mathsf{~ }$]{} to (0.4,3.6); (0.4,0.4)node\[right\][$\mathsf{~ }$]{} to (3.6,0.4); (3.6,0.4)node\[right\][$\mathsf{~ }$]{} to (3.6,3.6); (0.4,3.6)node\[right\][$\mathsf{~ }$]{} to (3.6,3.6);
(0,0) rectangle (4,4); (0,0) rectangle (4,4); (0,1.2)node\[right\][$\mathsf{~ }$]{} to (4,1.2); (0,2.8)node\[right\][$\mathsf{~ }$]{} to (4,2.8); (1.2,0)node\[right\][$\mathsf{~ }$]{} to (1.2,4); (2.8,0)node\[right\][$\mathsf{~ }$]{} to (2.8,4);
(0,-70)[(1.0,1.0)[$x$]{}]{} (-70,0)[(1.0,1.0)[$z$]{}]{} (100,-30)[(1.0,1.0)[$y$]{}]{}
Let $N$ and $N_i$ denote the number of degrees of freedom associated to $\Omega$ and $\Omega_j^{\delta}$, respectively. Let $A\in \mathbb{R}^{N \times N}$ be the Jacobian matrix of the linear system defined on a mesh $\Omega_h$ $$\label{lin_system1}
AX=b.$$ Then we can define the matrices ${R}^{\delta}_i$ and ${R}^0_i$ as the restriction operator from $\Omega_h$ to its overlapping and non-overlapping subdomains as follows: Its element $({R}^{\delta}_i)_{l_1,l_2}$ is either (a) an identity block, if the integer indices $1\leq l_1\leq N_i$ and $1\leq l_2\leq N$ are related to the same mesh point and this mesh point belongs to $\Omega_j^{\delta}$, or (b) a zero block, otherwise. The multiplication of ${R}_i^{\delta}$ with a $N \times 1$ vector generates a smaller $N_i\times1$ vector by discarding all elements corresponding to mesh points outside $\Omega_i^{\delta}$. The matrix ${R}_i^0\in \mathbb{R}^{N_i\times N}$ is defined in a similar way, the only difference to the operator ${R}^{\delta}_i$ is that its application to a $N\times 1$ vector also zeros out all of those elements corresponding to mesh points on $\Omega_i^{\delta}\backslash \Omega_i$. Then the classical one-level additive Schwarz (AS, [@Toselli]) preconditioner is defined as $$\begin{array}{ll}
{M}^{-1}_{\delta,\delta} =\sum\limits_{i=1}^{N_p}({R}^{\delta}_i)^T
{A}_i^{-1}{R}^\delta_i.
\end{array}
\label{equ:as-class}$$ with ${A}_i= {R}^\delta_i {A}({R}^\delta_i)^T$ and $N_p$ is the number of subdomains, which is the same as the number of processors. In addition to that, there are two modified approaches of the one-level additive Schwarz preconditioner that may have some potential advantages for parallel computing. The first version is the left restricted additive Schwarz (left-RAS, [@cs99]) method defined by $$\begin{array}{ll}
{M}^{-1}_{0,\delta} =\sum\limits_{i=1}^{N_p}({R}^0_i)^T
{A}_i^{-1}{R}^\delta_i.
\end{array}
\label{equ:as-left}$$ and the other modification to the original method is the right restricted additive Schwarz (right-RAS, [@cs03]) preconditioner as follow: $$\begin{array}{ll}
{M}^{-1}_{\delta,0} =\sum\limits_{i=1}^{N_p}({R}^\delta_i)^T
{A}_i^{-1}{R}^0_i.
\end{array}
\label{equ:as-right}$$ In the above preconditioners, we use a sparse LU factorization or incomplete LU (ILU) factorization method to solve the subdomain linear system corresponding to the matrix ${A}_i^{-1}$. In the following, we will denote a Schwarz preconditioner simply by $M^{-1}_{h}$ defined on the mesh $\Omega_h$ with the characteristic size $h$, when the distinction is not important.
Multilevel additive Schwarz preconditioners
-------------------------------------------
To improve the scalability and robustness of the one-level additive Schwarz preconditioners, especially when a large number of processors is used, we employ the family of multilevel Schwarz preconditioners. Multilevel Schwarz preconditioners are obtained by combining the single level preconditioner $M^{-1}_{h}$ assigned to each level, as shown in Figure \[fig:partition-domain\]. For the description of multilevel Schwarz preconditioners [@kong; @Prudencio; @sbg96], we use the index $j=0,1,...,L-1$ to designate any of the $L\geq 2$ levels. The meshes from coarse to fine are denoted by $\Omega_{h_j}$, and the corresponding matrices and vectors are denoted $A_{h_j}$ and $X_{h_j}$. Let us denote $I_{h_j}$ as the identity operator defined on the level $j$, and the restriction operator from the level $j$ to the level $j-1$ be defined by $$\mathcal{I}_j^{j-1}: \mathbb{R}^{N_j} \rightarrow \mathbb{R}^{N_{j-1}},
\label{equ:restriction}$$ where $N_j$ and $N_{j-1}$ denote the number of degrees of freedom associated to $\Omega_{h_j}$ and $\Omega_{h_{j-1}}$. Moreover, $$\label{equ:interpolation}
\mathcal{I}_{j-1}^{j}: \mathbb{R}^{N_{j-1}} \rightarrow \mathbb{R}^{N_{j}}$$ is the interpolation operator from the level $j-1$ to the level $j$.
For the convenience of introduction, we present the construction of the proposed multilevel Schwarz preconditioners from the view of the multigrid (MG) V-cycle algorithm [@kong; @Prudencio; @sbg96; @Toselli]. More precisely speaking, in this sense, at each level $j>0$, the Schwarz preconditioned Richardson method works as the pre-smoother and post-smoother, i.e., $M^{-1}_{h_j}$ preconditioning the $\mu_i\geq0$ presmoother iterations and $M^{-1}_{h_j}$ preconditioning the $\nu_i\geq0$ postsmoother iterations. In the general multigrid V-cycle framework, let $X^k$ be the current solution for the linear system , the new solution is computed by the iteration $X^{k+1}=MG(b,L,X^{k})$, as described in Algorithm \[vcyclealgorithm\].
Solve $A_{h_0}X_{h_0}=b_{h_0}$ Smooth $\mu_i$ times $A_{h_j}X_{h_j}=b_{h_j}$: $(b_{h_j}-A_{h_j}X_{h_j})=(I_{h_j}-A_{h_j}M^{-1}_{h_j})^{\mu_i}(b_{h_j}-A_{h_j}X_{h_j})$; $b_{h_{j-1}}=\mathcal{I}_j^{j-1}(b_{h_j}-A_{h_j}X_{h_j})$; $X_{h_{j-1}}=MG(b_{h_{j-1}},j-1,0)$; $X_{h_j}=X_{h_j}+\mathcal{I}_{j-1}^{j} X_{h_{j-1}}$; Smooth $\nu_i$ times $A_{h_j}X_{h_j}=b_{h_j}$: $(b_{h_j}-A_{h_j}X_{h_j})=(I_{h_j}-A_{h_j}M^{-1}_{h_j})^{\nu_i}(b_{h_j}-A_{h_j}X_{h_j})$; **return** $X_{h_j}$
### With application to the two-level case
In the following, we restrict ourselves to the two-level case, and use the geometry preserving coarse mesh that shares the same boundary geometry with the fine mesh, as shown in Figure \[fig:partition-domain\]. The V-cycle two-level Schwarz preconditioner $M_{V}^{-1}$, i.e, $L=2$ in Algorithm \[vcyclealgorithm\], is constructed by combining the fine level $M^{-1}_{h_1}$ and the coarse level $M^{-1}_{h_0}$ preconditioners as follows: $$\label{two_level_preconditioner}
M_{V}^{-1}=A_{h_1}^{-1}\left[I_{h_1}-(I_{h_1}-A_{h_1}M^{-1}_{h_1})^{\nu_1}(I_{h_1}-A_{h_1}\mathcal{I}_0^{1}M^{-1}_{h_0}\mathcal{I}_1^0)(I_{h_1}-A_{h_1}M^{-1}_{h_1})^{\mu_1}\right].$$ More precisely, if the pre- and post-smoothing parameters are fixed to $\mu_i=1$ and $\nu_i=1$, then the matrix-vector product of the two-level Schwarz preconditioner in with any given vector $e_{h_1}=M_{V}^{-1}r_{h_1}$ is obtained by the following three steps: $$\label{equ:v-cycle}
\left\{
\begin{array}{ll}
\displaystyle{ e_{h_1}^{1/3}=M^{-1}_{h_1}r_{h_1} },\\
\\
\displaystyle{ e_{h_1}^{2/3}=e_{h_1}^{1/3}+\mathcal{I}^{1}_{0}M^{-1}_{h_0}\mathcal{I}^{0}_{1} (r_{h_1}-A_{h_1}e_{h_1}^{1/3}),}\\
\\
\displaystyle{ e_{h_1}=e_{h_1}^{2/3}+M^{-1}_{h_1} (r_{h_1}-A_{h_1}e_{h_1}^{2/3}).}
\end{array}
\right.$$
There are some modified versions of the V-cycle two-level Schwarz preconditioner that can be used for parallel computing. If we set $\mu_1=0$ and $\nu_1=1$ in , then we obtain the left-Kaskade Schwarz method defined by $$\label{equ:left-Kas}
M_{left-Kas}^{-1}=\mathcal{I}_0^{1}M^{-1}_{h_0}\mathcal{I}_1^0+M^{-1}_{h_1}-M^{-1}_{h_1}A_{h_1}\mathcal{I}_0^{1}M^{-1}_{h_0}\mathcal{I}_1^0.$$ And the other modification to the original method is the right-Kaskade Schwarz preconditioner with $\mu_1=1$ and $\nu_1=0$ in as follow: $$\label{equ:right-Kas}
M_{right-Kas}^{-1}=M^{-1}_{h_1}+\mathcal{I}_0^{1}M^{-1}_{h_0}\mathcal{I}_1^0-\mathcal{I}_0^{1}M^{-1}_{h_0}\mathcal{I}_1^0A_{h_1}M^{-1}_{h_1}.$$ Moreover, we can define other two hybrid versions of two-level Schwarz algorithms. The first two-level method is the pure additive two-level Schwarz preconditioner as $$\label{equ:two-additive}
M_{additve}^{-1}=\mathcal{I}_0^{1}M^{-1}_{h_0}\mathcal{I}_1^0+M^{-1}_{h_1},$$ and the coarse-level only type two-level Schwarz preconditioner $$\label{equ:coarse-level}
M_{coarse}^{-1}=\mathcal{I}_0^{1}M^{-1}_{h_0}\mathcal{I}_1^0.$$ The motivation of the above two-level preconditioners, including, additively or multiplicatively, coarse preconditioners to an existing fine mesh preconditioner, is to make the overall solver scalable with respect to the number of processors or the number of subdomains. Hence, some numerical results will be shown later to compare the performance of these two-level preconditioners.
\[remark-2\] In the multilevel Schwarz preconditioners, the use of a coarse level helps the exchange of information, and the following linear problem defined on the coarsest mesh needs to be solved $$A_{h_0}X_{h_0}=b_{h_0},
\label{equ:coarse-system}$$ with the use of the one-level additive Schwarz preconditioned GMRES, owing to the face that it is too large and expensive for some direct methods in the applications of large-scale simulations. When an iterative method is used for solving the linear system on the coarsest level , the overall preconditioner is an iterative procedure, and the preconditioner changes from iteration to iteration. Hence, when the multilevel Schwarz method is applied, a flexible version of GMRES (fGMRES, [@Sa03]) is used for the solution of the outer linear system .
\[remark-3\] When the classical additive Schwarz preconditioner is applied to symmetric positive definite systems arising from the discretization of elliptical problems, then the condition number $\kappa$ of the preconditioned linear system satisfies $\kappa\leq C(1+H/\delta)/{H^2}$ for the one-level method and $\kappa\leq C(1+H/\delta)$ for the two-level method, where the parameter $C$ is independent of $\delta$, $H$, and $h$, see the references [@sbg96; @Toselli]. However, these condition number estimates can neither be applied to the restricted additive Schwarz preconditioners, nor adapted to the family of non-elliptic systems like our model problem, and thereby there are very little theoretical literatures on the convergence of multilevel restricted Schwarz preconditioners for this system. The scalability tests in this paper will provide more understanding of restricted type domain decomposition methods for the hyperbolic problem.
### Selection of interpolation operators
Algorithm \[vcyclealgorithm\] provides a general framework for choosing the coarse-to-fine mesh ratio, the restriction and interpolation operators, and the solvers for the smoothing and coarse-grid correction. And in this among them, choosing the right restriction and interpolation operators at each level is very important for the overall performance of the preconditioner in terms of the trade-off between rate of convergence and the cost of each iteration. Generally speaking, in most cases the overall solver makes a profit of the addition of coarse preconditioners with the decrease of the total number of linear iterations. However, for some classes of important problems such as reservoir simulation, we found that the number of iterations does not decrease as expected, if the operators are not chosen properly. After some attempts, we figure out the source of this phenomenon from the cell-centered spatial discretization. When the cell-centered scheme is involved, it happens that a coarse mesh point does not coincide with any fine mesh point, as shown in Figure \[two-level-domain\]. Hence, in the rest of this paper, we focus on the some strategies for the selection of interpolation operators, i.e, the first-order, second-order and third-order interpolation schemes.
(0,0) rectangle (5,5); (0,0) rectangle (5,5); (0,0) grid (5,5); (1.00,1.00) rectangle (1.5,1.5); (1.5,1.5) rectangle (2.0,2.0); (1.5,1.0) rectangle (2.0,1.5); (1.0,1.5) rectangle (1.5,2.0); (0.5,0.5) rectangle (1.0,2.5); (1.0,0.5) rectangle (2.5,1.0); (2.0,1.0) rectangle (2.5,2.5); (1.0,2.0) rectangle (2.0,2.5);
(0,0) grid (4,4); in [(0,0), (1,0), (2,0),(3,0),(0,1),(1,1),(2,1),(3,1),(0,1),(1,1),(2,1),(3,1),(0,2),(1,2),(2,2),(3,2),(0,3),(1,3),(2,3),(3,3)]{} + (0.5,0.5) circle (0.1);
(0,0) rectangle (5,5); (0,0) rectangle (5,5); (0,0) grid (5,5); (1.00,1.00) rectangle (1.5,1.5); (1.5,1.5) rectangle (2.0,2.0); (1.5,1.0) rectangle (2.0,1.5); (1.0,1.5) rectangle (1.5,2.0);
(0,0) grid (2,2); in [(0,0), (1,0), (0,1),(1,1)]{} + (0.5,0.5) circle (0.1);
(0,0) rectangle (5,5); (0,0) rectangle (5,5); (0,0) grid (1,1); (0,0) grid (5,5); (1.00,1.00) rectangle (2,2);
(0,0) grid (1,1); in [(0,0)]{} + (1.0,1.0) circle (0.1);
(7.0,1.3)node\[right\][$\mathsf{~ }$]{} to\[out=180,in=90\] (0,1.5); (7.0,-1.4)node\[right\][$\mathsf{~ }$]{} to\[out=180,in=90\] (0,-1.4); (6.3,-4.5)node\[right\][$\mathsf{~ }$]{} to\[out=180,in=90\] (0,-4);
Let $\Omega$ be the computational domain covered with $N_x\times N_y$ mesh cells. Then we consider $(x, y)$ is a point at the position of the rectangular subdivision $[x_i,x_{i+1}]\times[y_j,y_{j+1}]$ with $i=1, \dots, N_x$, $j=1, \dots, N_y$, and $f(x,y)$ is the interpolation point value on this point. The nearest neighbor interpolation operator, which approximate the interpolation point data of the nearest node according to the shortest distance between the interpolation point and the sample point in space, is a first-order method defined by $$\label{equ:inter-frist}
f(x,y)= \sum_{l=0}^1\sum_{m=0}^1f\left(x_{i+l},y_{j+m}\right)W_{i+l,j+m},$$ where $f(x_{j+l},y_{k+m})$ is the interpolation point value located at the point $(x_{j+l},y_{k+m})$, and $W_{i+l,j+m}$ is a weight defined as $$W_{i+l,j+m}=\left\{\begin{aligned}
1, \quad &\mbox{if} ~\|(x,y)-(x_{i+l},y_{j+m})\|~\mbox{is minimum}, \\
0, \quad &\mbox{others}.
\end{aligned}
\right.$$ The bilinear interpolation operator, which utilizes the weighted average of the nearest neighboring values $f(x_{i+l},y_{j+m})$ to approximately generate a interpolation point value. Due to the point $(x, y)$ being in the subdivision $[x_i,x_{i+1}]\times[y_j,y_{j+1}]$, is a second-order interpolation technique defined as follows: $$\label{equ:inter-second}
f(x,y)=\sum_{l=0}^1\sum_{m=0}^1f\left(x_{i+l},y_{j+m}\right)W_{i+l}W_{j+m},$$ where the weight for the bilinear interpolation scheme is defined as: $$\left\{
\begin{aligned}
W_{i+l}&=\left(\frac{2x-x_i-x_{i+1}}{x_{i+1}-x_{i}}\right)l+\frac{x_{i+1}-x}{x_{i+1}-x_{i}},\\
W_{j+m}&=\left(\frac{2y-y_j-y_{j+1}}{y_{j+1}-y_{j}}\right)m+\frac{y_{j+1}-y}{y_{j+1}-y_{j}}.
\end{aligned}
\right.$$ In contrast to the bilinear interpolation, which only takes 4 ponits $(2\times2)$ into account, the bicubic interpolation use 16 points $(4\times4)$ and is a third-order interpolation scheme defined as follows: $$\label{equ:inter-third}
f(x,y)=\sum_{l=-1}^2\sum_{m=-1}^2f\left(x_{i+l},y_{j+m}\right)W\left(\frac{x-x_{i+l}}{x_{i+1}-x_{i}}\right)W\left(\frac{y-y_{j+m}}{y_{j+1}-y_{j}}\right),$$ where the weight for the cubic convolution interpolation is defined as $$W(s)=\left\{\begin{aligned}
&\frac{3}{2}|s|^3-\frac{5}{2}|s|^2+1,\quad &0< |s|< 1, \\
&-\frac{1}{2}|s|^3+\frac{5}{2}|s|^2-4|s|+2, \quad&1<|s|< 2, \\
&\;0, \quad &2<|s|.
\end{aligned}
\right.$$ We remark that the two dimensional computational domain is used in the above description only for the ease of demonstration.
Numerical experiments {#Numerical-scheme}
=====================
In this section, we implement the proposed algorithm described in the previous sections using the open-source Portable, Extensible Toolkit for Scientific computation (PETSc) [@petsc], which is built on the top of Message Passing Interface (MPI), and investigate the numerical behavior and parallel performance of the newly proposed fully implicit solver with a variety of test cases.
[[Robustness and efficiency of the solver]{}]{}
-----------------------------------------------
[[The focus of the subsection is on the robustness and efficiency of the proposed algorithm for both standard benchmarks and realistic problems in highly heterogeneous media.]{}]{} Unless otherwise specified, the values of physical parameters used in the test cases are set as follows: $\phi=0.2$, $\mu=11.067\times10^{-3} \; \mbox{cp}$, $R= 8.3147295\; \mbox{J}/(\mbox{mol}\cdot \mbox{K})$, $W=16\times10^{-3} \; \mbox{kg/mol}$, $T=298 \; \mbox{K}$, $p_c= 4.604\times10^6 \; \mbox{pa}$, $T_c=190.58 \; \mbox{K}$, and $T_b=111.67 \; \mbox{K}$.
[[We first present results from a 2D test case (denoted as Case-1) in a horizontal layer with the permeability tensor $\mathbf{K}$ in as follows: $$\mathbf{K}=\left[\begin{array}{cc}
k_{xx} & 0 \\
0 & k_{yy}\\
\end{array}
\right].
\label{2.4}$$ With the media being horizontal, we neglect the effect of the gravity. In the configuration, the distribution of permeabilities includes two domains with different isotropy, as shown in Figure \[case2\_perm1\]. The computational domain is 100 meters long and 100 meters wide.]{}]{} In the simulation, we assume that the left and right boundaries of the domain are impermeable. Then we flood the system by gas from the top to the bottom, i.e., we set $p_{in}=10$ atm at the top boundary and set $p_{out}=1$ atm at the right boundary. And there is no injection/extraction inside the domain. Compared with the previous example, in this test case there is an $H$-shape zone and the value of permeabilities has a huge jump inside and outside the zone, which brings about even greater challenges to the fully implicit solver. In the test, the simulation is performed on a $512\times512$ mesh, the time step size is fixed to $\Delta t=0.1$ day, and the simulation is stopped at $10$ day. [Figure \[case2\_perm1\] also illustrates the contour plots of the pressure.]{} Table \[Heterogeneous\_isotropic\] shows the performance of the proposed fully implicit solver with respect to different permeability configurations. It is clearly seen that the simulation spends more computing time when the isotropic medium is used, which attributes to the increase in nonlinearity of the problem that affects the number of linear iterations.
Medium type [Isotropic case]{} [Anisotropic case ]{}
------------------------------ -------------------- -----------------------
Average nonlinear iterations 2.2 2.2
Average linear iterations 54.7 41.5
Execution time (second) 172.0 151.1
: [[[A comparison of the fully implicit solver with different permeability configurations for Case-1.]{}]{}]{}
\[Heterogeneous\_isotropic\]
[[In the following, the experiment is conducted to simulate some 3D test cases. The focus is on the flow model with the medium being highly heterogeneous and isotropic, which significantly increases the nonlinearity of the system and imposes a severe challenge on the fully implicit solver. We first consider a 3D domain with dimension $\Omega=(0,100\; \mbox{m})^3$, in which the permeabilities of the porous medium are random, denoted as Case-2.]{}]{} The random distribution of permeability with the range \[3.1,14426.2\] is generated by a geostatistical model using the open source toolbox MRST [@mrst], as shown in Figure \[case3\_perm\]. [In the test, two different flow configurations are used for the injection boundary, i.e., the fixed injection pressure and the fixed injection rate. In the fixed pressure configuration, the flow with the pressure of 10 atm is injected at the part of the left-hand side $x=0\; \mbox{m},y\in[0,10\; \mbox{m}],z\in[0,10\; \mbox{m}]$; while, for the case of the fixed injection rate, the flow is injected from the same sub-boundary with the Darcy’s velocity $\mathbf{v}=5\times10^{-4}~\mbox{m/s}$. The fluid flows from the right-hand side of the domain with a fixed pressure $p_{out}=1$ atm, and no-flow boundary conditions are imposed on the other boundaries of the domain.]{} The initial pressure of $p_0=1$ atm is specified in the whole domain and the parameters required for this example are consistent with the previous examples.
![Permeability field for Case-2. In the figure, we use the logarithmic scale for the contour plot of the permeability. []{data-label="case3_perm"}](./fig/case3_perm1.png){width="0.4\linewidth"}
In this case, the domain is highly heterogeneous, leading to the increase in nonlinearity of the problem that is challenging for the numerical techniques. [Figure \[3D\_pressure\] shows the plots of the pressure profiles for Case-2 at different times under the two flow configurations. It is demonstrated from the figures that the proposed approach successfully resolves the different stages of the simulation, and the flow is close to reach the breakthrough at the time $t=5$ days.]{} In Table \[random\_permeaiblity\], we present the history of the values of nonlinear and linear iterations and the total execution time for the proposed solver at different times. The simulation is carried out on a $128\times128\times128$ mesh and the time step size is $\Delta t=0.1$ day. It can be observed from Table \[random\_permeaiblity\] that the number of nonlinear and linear iterations has barely changed, while the total execution time increases constantly with the advance of time as expected, which displays the robustness and effectiveness of the implicit solver to handle the highly variety of physical parameters in the model problem.
\
Boundary Type End of the simulation (day) [$t=1$ ]{} [$t=2$]{} [$t=3$ ]{} [ t=5 ]{}
--------------- ------------------------------ ------------ ----------- ------------ -----------
Average nonlinear iterations 3.0 2.9 2.9 2.9
Average linear iterations 22.6 23.8 24.3 24.5
Execution time (second) 645.3 1095.9 1545.2 2455.6
Average nonlinear iterations 2.8 2.7 2.7 2.6
Average linear iterations 24.8 26.3 27.1 27.9
Execution time (second) 519.5 965.4 1325.1 2142.9
: [A comparison of the fully implicit solver at different times for Case-2 with the fixed pressure or velocity boundary.]{}
\[random\_permeaiblity\]
In the all of the above test cases, the porosity $\phi$ in the model problem is assumed to a constant. In the next 3D test case (Case-3) of this subsection, we import the porosity and the permeability from the Tenth SPE Comparative Project (SPE10) [@spe10] as an example of a realistic realization with geological and petrophysical properties, in which the porosity and the permeability are capable of variation with the change of the position. It is a classic and challenging benchmark problem for reservoir due to highly heterogeneous permeabilities and porosities. As shown in Figure \[spe10permporo\], the permeability is characterized by variations of more than six orders of magnitude and is ranged from $6.65\times10^{-4}$ to $2\times10^4$, and the porosity scale ranges from 0 to 0.5. The 3D domain dimensions are 1200 ft long $\times$ 2200 ft wide $\times$ 170 ft thick. In the test, the boundary condition for pressure on the left-hand side of the domain is uniformly imposed to $p_{in}=10$ atm and on the opposite side is imposed to $p_{out}=1$ atm. No-flow boundary condition is set to other boundary of the domain. Other used parameters are the same as the previous case.
\
[In Figure \[spe10\_pressure\], we display the distributions for the pressure when the simulation is finished at different time points $t=1,3,6,10$ days.]{} In the simulation, the mesh is $60\times220\times85$ and the time step size is fixed to $0.1$ day. The results shown in Figure \[spe10\_pressure\] demonstrate that the proposed approach successfully resolves the rapid and abrupt evolution of the simulation at different stages, while keeping the solver in a robust and efficient way. Finally, we analyze the behavior of the proposed fully-implicit method when the time step size $\Delta t$ is changed. In the test, we again run the SPE-10 model on a fixed $60\times220\times85$ mesh. The simulation is stopped at $t=3$ year. The results on the average numbers of nonlinear and linear iterations as well as the total compute time are summarized in Table \[spe10\_timesteps\]. The results in the table clearly indicate that the combination of nonlinear and linear iterations works well for even very large value of time steps. The implicit approach converges for all time steps and is unconditionally stable. In addition to that, we also notice that, as the time step size $\Delta t$ decreases, the average number of nonlinear and linear iterations become smaller, whereas the total computing time increases. This behavior is somehow expected for the fully implicit approach in a variety of applications [@bui17; @kong; @wang; @haijian-cmame20; @haijian-cmame].
Time step size $\Delta t$ [0.06]{} [0.075]{} [0.1]{} [0.2]{} [0.3]{}
------------------------------ ---------- ----------- --------- --------- ---------
Number of time steps 50 40 30 15 10
Average nonlinear iterations 3.5 3.6 3.8 4.4 4.5
Average linear iterations 45.4 49.3 52.1 67.5 80.0
Execution time (second) 1424.2 1191.7 951.4 579.8 408.8
: The effect of different time step sizes in the fully implicit solver for SPE10.
\[spe10\_timesteps\]
\
Performance of Schwarz preconditioners {#sec:Schwraz}
--------------------------------------
In this subsection, we focus on the parallel performance of the proposed fully implicit solver with respect to the one-level or multilevel Schwarz preconditioner by Case-1, [Case-2]{}, and a new 3D test case. In the new 3D case, [denoted as Case-4]{}, the computational domain is $\Omega=(0,100 \; \mbox{m})^3$ with the medium being heterogeneous and isotropic, and the distribution of permeability in each layer $z\in[25\; \mbox{m},75\; \mbox{m}]$ is the same as the distribution of Case-1. We flood the system by gas from the behind face ($x\in[0,100\; \mbox{m}]$, $y=100\; \mbox{m}$, $z\in[0,100\; \mbox{m}]$) to the front face, i.e., we set $p_{in}=10$ atm at the behind face and set $p_{out}=1$ atm at the front face. And there is no injection/extraction inside the domain.
In the test, we use the following stopping parameters and notations, unless specified otherwise. The relative and absolute tolerances in for Newton iterations are set to $10^{-6}$ and $10^{-10}$, respectively. The linear systems are solved by the one-level or two-level Schwarz preconditioned GMRES method with absolute and relative tolerances of $10^{-8}$ and $10^{-5}$ in , except for the coarse solve of the two-level preconditioner, for which we use $2\times 10^{-1}$ as the relative stopping tolerance. The restart value of the GMRES method is fixed as 30. In the tables of the following tests, the symbol “$N_p$" denotes the number of processor cores that is the same as the number of subdomains, “N. It" stands for the average number of Newton iterations per time step, “L. It" the average number of the one-level or two-level Schwarz preconditioned GMRES iterations per Newton iteration, and “Time" the total computing time in seconds.
### One-level Schwarz preconditioners
Under the framework of one-level Schwarz preconditioners, we study the performance of the fully implicit solver, when different types of the additive Schwarz preconditioners are employed and several important parameters such as the subdomain solver and the overlapping size are taken into consideration.
We first look at the influence of subdomain solvers. In this test, we focus on the classical AS preconditioner and fix the overlapping factor to $\delta=1$. The ILU factorization with different levels of fill-in and the full LU factorization are considered for the subdomain solvers. The simulation of Case-1 is applied to a fixed $512\times512$ mesh using $N_p=16$ processors. The time step size is 0.1 day, and we run the 2D case for the first 20 time steps. The simulation of Case-4 is applied to a fixed $128\times128\times128$ mesh by using $N_p=64$ processors. The time step size is 0.1 day, and we run this case for the first 10 time steps. The numerical results are summarized in Table \[different-subpreconditioner\]. We observe from the table that the number of nonlinear iterations is not sensitive to the choice of subdomain solvers. The linear system converges with less iterations when the LU factorization method is used as the subdomain solver; while the ILU approach is beneficial to the AS preconditioner in terms of the computing time, especially for the 3D test case.
-------- ------- ------- ------ ------- ------- -------
Case
Solver N. It L. It Time N. It L. It Time
3.0 52.9 37.1 4.2 32.9 954.8
3.0 99.1 34.3 4.3 161.7 282.4
3.0 75.7 29.9 4.3 98.0 253.0
3.0 73.9 32.1 4.3 76.1 239.6
3.0 66.2 33.8 4.3 51.0 245.8
-------- ------- ------- ------ ------- ------- -------
: The impact of different subdomain solvers in the one-level Schwarz preconditioner.
\[different-subpreconditioner\]
[Moreover, in Table \[tab:ilu-case4\], we also compare the impact of different subdomain solvers for Case-2, when the fixed pressure and velocity boundaries are imposed for the flow configurations. In the test, the simulation is carried out on a $128\times128\times128$ mesh and the time step size is fixed to $\Delta t=0.1$ day by using $N_p=64$ processors. It is clearly illustrated from the results that the choice of ILU(2) is still optimal for the 3D flow problem under the random permeability case with different boundary conditions, when the flow undergoes different stages in the complex simulation as shown in Figure \[3D\_pressure\].]{}
--------- ------- ------- ------- ------- ------- -------- ------- ------- -------- ------- ------- --------
Solver
Model-A N. It L. It Time N. It L. It Time N. It L. It Time N. It L. It Time
3.0 33.0 685.2 2.9 36.6 1167.0 2.9 38.7 1652.7 2.9 41.2 2625.9
3.0 161.5 435.3 2.9 183.0 760.1 2.9 195.3 1086.8 2.9 208.3 1742.7
3.0 100.6 415.5 2.9 110.3 715.9 2.9 116.4 1015.4 2.9 123.6 1616.7
3.0 80.0 403.4 2.9 88.9 695.0 2.9 93.3 991.1 2.9 98.4 1572.5
3.0 52.4 417.8 2.9 58.8 725.1 2.9 62.1 1030.5 2.9 65.7 1638.9
Model-B N. It L. It Time N. It L. It Time N. It L. It Time N. It L. It Time
2.8 39.3 563.7 2.7 42.9 1046.9 2.7 45.2 1438.8 2.6 47.7 2335.0
2.8 178.5 366.0 2.7 190.4 679.2 2.7 201.8 934.4 2.6 214.9 1518.6
2.8 113.9 349.6 2.7 124.0 641.3 2.7 131.6 873.7 2.6 138.4 1408.2
2.8 87.8 340.2 2.7 95.7 626.8 2.7 101.2 859.0 2.6 105.0 1363.8
2.8 67.3 351.8 2.7 74.1 650.5 2.7 78.5 895.2 2.6 81.6 1455.1
--------- ------- ------- ------- ------- ------- -------- ------- ------- -------- ------- ------- --------
: [The impact of different subdomain solvers in the one-level Schwarz preconditioner for Case-2 with the fixed pressure boundary (denoted as “Model-A") and the fixed velocity boundary (denoted as “Model-B").]{}
\[tab:ilu-case4\]
We then perform test with the 2D and 3D test cases by varying the combination of the overlapping factor $\delta$ and the Schwarz preconditioner types, i.e., the classical-AS , the left-RAS , and the right-RAS preconditioners. Based on the above observations, we take the choice of the subdomain solver with ILU(1) for the 2D test case and ILU(2) for the 3D test case. The numbers of nonlinear and linear iterations together with the execution time are illustrated in Table \[different-overlap\]. The results in the table suggest that the more robust combination is the left-RAS preconditioner with the overlapping size $\delta=1$ for the compromise between the linear iteration and the total computing time. We remark that, when the overlapping size $\delta=0$, these preconditioners degenerates into the block-Jacobi preconditioner.
---------------- ---------- ------- ------- ------ ----------- ----------- ----------- ------- ------- -------
Case
Preconditioner $\delta$ N. It L. It Time [N. It]{} [L. It]{} [Time]{} N. It L. It Time
0 3.0 80.4 31.5 [2.8]{} [73.9]{} [321.6]{} 4.2 53.0 211.9
1 3.0 75.7 29.9 [2.8]{} [87.8]{} [340.2]{} 4.3 76.1 239.6
2 3.0 76.9 31.8 [2.8]{} [99.1]{} [352.7]{} 4.3 92.6 286.7
3 3.0 76.8 32.5 [2.8]{} [103.4]{} [368.2]{} 4.3 93.8 326.6
1 3.0 55.8 28.3 [2.8]{} [43.7]{} [306.2]{} 4.2 37.5 205.0
2 3.0 51.9 28.8 [2.8]{} [41.6]{} [312.8]{} 4.3 34.0 218.0
3 3.0 49.4 28.9 [2.8]{} [40.8]{} [322.5]{} 4.3 33.2 232.3
1 3.0 55.6 28.2 [2.8]{} [44.5]{} [307.5]{} 4.3 37.1 207.3
2 3.0 51.4 28.0 [2.8]{} [42.0]{} [311.4]{} 4.3 34.3 215.8
3 3.0 48.7 28.0 [2.8]{} [41.2]{} [326.2]{} 4.3 33.4 223.4
---------------- ---------- ------- ------- ------ ----------- ----------- ----------- ------- ------- -------
: The impact of different one-level Schwarz preconditioners with several overlapping sizes.
\[different-overlap\]
### Two-level Schwarz preconditioners
For the reservoir simulation with high accuracy, the supercomputer with a large number of processors is a must, and therefore the scalabilities of the algorithm with respect to the number of processors are critically important. As introduced in Section \[two-RAS\], in the one-level Schwarz preconditioner, the average number of linear iterations per Newton iteration grows with the number of processors $N_p$, resulting in the deterioration of the implicit solver. It is clear that some stabilization is needed, which is achieved by the two-level method with the use of a coarse mesh and the interpolation operators explained in Section \[two-RAS\]. The performance of the two-level method depends heavy on the two linear solvers defined on the coarse and fine meshes. Here, we restrict ourselves within the framework of Schwarz preconditioned GMRES methods, i.e., we refer to these iterative methods as smoothers on the each levels.
There are several assembly techniques available to construct a hybrid two-level Schwarz preconditioner by composing the one-level additive Schwarz preconditioner with a coarse-level preconditioner in a multiplicative or additive manner. Choosing the right type of two-level Schwarz preconditioners is very important for the overall performance of the preconditioner. A large number of numerical experiments is often necessary to identify the right selection. As introduced in Section \[two-RAS\], in the study we investigate the pure-coarse , the additive , the left-Kaskade , the right-Kaskade , and the V-cycle type two-level Schwarz preconditioners. For each numerical case, the the overlapping sizes on the fine and coarse levels are fixed to [$\delta_f=1$ and $\delta_c=1$]{}, respectively. The subdomain solvers on the fine and coarse levels is solved by the LU factorization. In the test, we again run the 2D model with a fixed time step size $0.1$ day [on a $512 \times 512$ mesh]{} and run Case-4 with a fixed time step size $0.2$ day [on a $128 \times 128 \times 128$ mesh]{}, and the simulation is finished at $t= 2$ and $1$ days, respectively. [For the simulation of Case-4, the flow problem with the fixed velocity model is solved on a $128\times128\times128$ mesh, and the computation is ended at $t=1$ day with $\Delta t=0.1$ day.]{} The coarse-to-fine mesh ratio is used to $2$ in each direction. In Table \[different-level\], we report the performance of the proposed two-level Schwarz preconditioners with respect to different interpolation operators that includes the first-order scheme , the second-order scheme , and the third-order scheme . Below we list the observations made from the results.
- We know that the effectiveness of the Schwarz preconditioner relies on its ability to mimic the spectrum of the linear operator and at the same time is relatively cheap to apply. We see that using the high order schemes in the construction of the Schwarz preconditioner provides less linear iteration counts. Moreover, when compared with the Schwarz method with high order methods, the low order approach is more attractive in the terms of the execution time, owing to its lower bandwidth and a less number of nonzeros in the sparse matrix. The results in the table suggest that the second order scheme is a suitable choice for compromise between the iterations and the total computing time.
- The best choices for some of the options in the multilevel Schwarz preconditioner are problem dependent. For the implicit solution of subsurface flows problems, from Table \[different-level\], we can see that the additive or Kaskade type two-level Schwarz preconditioners is exacerbated by a larger number of outer linear iterations, when compared with the V-cycle approach. The pure coarse version of the two-level approach performs considerably worse than the hybrid methods. Hence, our experiments suggest that there is a benefit to include both pre- and post-swipes of the one-level preconditioning for the simulation of model problems, especially for the 3D test cases.
-- ------------------- --------- ---------- ----------- --------- ---------- ----------- --------- ---------- ------------
N. It L. It Time N. It L. It Time N. It L. It Time
Coarse – – – – – – – – –
Additive 3.0 114.6 233.9 3.0 71.3 207.3 3.0 45.9 734.7
Left-Kaskade 3.0 71.1 204.2 3.0 34.1 163.9 3.0 8.05 565.2
Right-Kaskade 3.0 75.0 194.8 3.0 37.7 157.8 3.0 9.03 565.5
V-Cycle 3.0 10.0 129.6 3.0 9.4 129.1 3.0 5.4 579.3
[Coarse]{} [–]{} [–]{} [–]{} [–]{} [–]{} [–]{} [–]{} [–]{} [–]{}
[Additive]{} [2.8]{} [87.4]{} [371.0]{} [2.8]{} [58.3]{} [349.4]{} [2.8]{} [42.0]{} [5186.7]{}
[Left-Kaskade]{} [2.8]{} [31.3]{} [325.7]{} [2.8]{} [20.6]{} [312.4]{} [2.8]{} [14.7]{} [4921.4]{}
[Right-Kaskade]{} [2.8]{} [34.5]{} [326.6]{} [2.8]{} [21.8]{} [315.9]{} [2.8]{} [15.1]{} [4940.3]{}
[V-Cycle]{} [2.8]{} [14.9]{} [314.2]{} [2.8]{} [12.6]{} [308.8]{} [2.8]{} [5.8]{} [4770.9]{}
Coarse – – – – – – – – –
Additive 4.0 79.9 453.2 4.0 44.5 402.1 4.0 36.6 6114.6
Left-Kaskade – – – – – – 4.0 13.4 5763.8
Right-Kaskade – – – – – – 4.0 14.8 5308.8
V-Cycle 4.0 11.4 374.2 4.0 10.3 372.8 4.0 5.1 5156.5
-- ------------------- --------- ---------- ----------- --------- ---------- ----------- --------- ---------- ------------
: The impact of different two-level Schwarz preconditioners with three interpolation operators. In the table, the symbol “–" denotes the divergence of the solver caused by the failure of linear iterations.
\[different-level\]
For the two-level preconditioner, the size of the coarse mesh has a strong impact on the efficiency and robustness of the method. It is clear that using a relatively fine coarse mesh gains a stronger two-level Schwarz preconditioner, and therefore it can help reduce the total number of linear iterations. On the other hand, finer coarse meshes generates plenty of amount of memory and cache, leading to the increase of the total compute time. An important implementation detail to consider in designing the two-level method is to balance the effects of preconditioning and the computing time of the coarse solve. To understand the impact of different coarse meshes on the convergence of the algorithm, we show the results with different coarse mesh sizes with respect to different interpolation operators for the 2D and 3D test cases. In the test, the fine meshes for the 2D and 3D problems are $512 \times 512$ and $128 \times 128 \times 128$, respectively. As shown in Table \[different-interpolation\], we observe from the table that: (a) for the first-order scheme, the bad quality of the second coarse mesh does lead to a large increase of the number of iterations; (b) with the help of higher order schemes, the linear iterations increase slowly with growth of coarse-to-fine mesh ratios. This implies that a higher order scheme does a good job on preconditioning the fine mesh problem, since the second coarse mesh even with worse quality is able to keep the whole algorithm efficient in terms of the number of iterations.
-- -------- --------- ---------- ----------- --------- ---------- ----------- --------- ---------- ------------
N. It L. It Time N. It L. It Time N. It L. It Time
2 3.0 10.0 129.6 3.0 9.4 129.1 3.0 5.4 579.3
4 3.0 18.8 123.0 3.0 16.9 120.9 3.0 13.3 553.1
8 3.0 26.8 130.1 3.0 19.2 120.0 3.0 14.1 535.5
16 3.0 36.0 140.5 3.0 19.5 118.5 3.0 14.2 532.8
[2]{} [2.8]{} [14.9]{} [314.2]{} [2.8]{} [12.6]{} [308.8]{} [2.8]{} [5.8]{} [4770.9]{}
[4]{} [2.8]{} [16.2]{} [310.7]{} [2.8]{} [13.1]{} [303.1]{} [2.8]{} [8.9]{} [4741.6]{}
[8]{} [2.8]{} [18.9]{} [308.5]{} [2.8]{} [13.7]{} [299.6]{} [2.8]{} [9.4]{} [4725.3]{}
[16]{} [2.8]{} [30.4]{} [326.2]{} [2.8]{} [15.8]{} [302.4]{} [2.8]{} [11.3]{} [4709.2]{}
2 4.0 11.4 374.2 4.0 10.3 372.8 4.0 5.1 5156.5
4 4.0 17.7 372.1 4.0 10.6 368.4 4.0 9.5 5106.9
8 4.0 18.8 370.2 4.0 11.0 354.4 4.0 10.2 5091.3
16 4.0 39.8 415.7 4.0 12.7 356.3 4.0 11.6 5051.4
-- -------- --------- ---------- ----------- --------- ---------- ----------- --------- ---------- ------------
: The impact of coarse-to-fine mesh ratios for the V-Cycle two-level Schwarz preconditioner.
\[different-interpolation\]
Parallel scalability
--------------------
Achieving good parallel scalability is important in parallel computing, especially when solving large-scale problems with many processors. Hence, in the following we focus on the parallel performance of the proposed fully implicit method with one-level and two-level restricted additive Schwarz preconditioners. Again, the 2D and 3D test problems descried in subsectionn \[sec:Schwraz\] are used for the scalability simulations. The numerical tests are carried out on the Tianhe-2 supercomputer. The computing nodes of Tianhe-2 are interconnected via a proprietary high performance network, and there are two 12-core Intel Ivy Bridge Xeon CPUs and 24 GB local memory in each node. In the numerical experiments, we use all 24 CPU cores in each node and assign one subdomain to each core.
-- -- ------ ------- ------- ------- ------- ------- -------
N. It L. It Time N. It L. It Time
256 7.2 65.4 564.3 7.2 18.3 460.7
512 7.2 95.0 291.5 7.2 18.9 187.6
1024 7.2 96.3 136.6 7.2 19.6 90.0
2048 7.2 142.1 82.1 7.2 20.4 46.1
4096 7.2 144.4 44.9 7.2 21.1 26.3
8192 7.2 214.2 34.6 7.2 24.0 18.8
512 7.8 82.7 775.7 7.8 31.0 735.1
1024 7.8 97.7 250.1 7.8 35.2 231.5
2048 7.8 109.8 97.6 7.8 37.9 76.2
4096 7.8 122.9 68.7 7.8 43.2 37.6
8192 7.8 147.0 29.9 7.8 35.0 21.8
-- -- ------ ------- ------- ------- ------- ------- -------
: Strong scalability with different number of processors $N_p$.
\[impact-strong-scale\]
The strong scalability (denoted by $Speedup$) and the parallel efficiency (denoted by $E_{f}$) are respectively defined as follows: $$Speedup=\displaystyle\frac{T(N_{min})}{T(N_p)},~E_{f}=\displaystyle\frac{T(N_{min})\times N_{min}}{T(N_p)\times N_p},$$ where $N_{min}$ denotes the smallest processors number of the compassion, $T(N_p)$ denotes computational time [with]{} $N_p$ processors. In the strong scalability test of the 2D problem, we use a fixed mesh $8192\times 8192$, and also fixed time step sizes $\triangle t=0.05$ and 0.01 days, in which the largest simulation consists of $8192\times 8192=67,108,864$ degrees of freedom. Moreover, we investigate the scalability of the the proposed fully implicit solver by using the fixed mesh $256\times 256\times 256$, also a fixed time step size $\triangle t=0.5$ day for Case-4. The number of processors is changed from $256$ to $8192$. Table \[impact-strong-scale\] shows the number of nonlinear and linear iterations as well as the computing time with respect to the number of processors, and Figure \[fig:strong-scale\] provides the compute time and $Speedup$ curve for the strong scalability. The table clearly indicates that the number of Newton iterations remains to be independent of the number of processors, and the number of linear iterations depends on the preconditioner employed in the solver. For the one-level solver, the number of linear iterations suffers as the number of processors increases. While, with the help of the coarse mesh, the number of linear iterations for the two-level preconditioner is kept to a low level as the number of processors increases. [Moreover, to investigate the strong scalability of the proposed method with respect to the heterogeneity property in the reservoir, we again use the Case-2 benchmark, and compare the number of nonlinear and linear iterations and the total computing time at $t=1,2,3,5$ days, respectively. As shown in Table \[tab:strong-scale-case-4\], a good strong scalability is also achieved for the random case test under longer simulations.]{}
-- ------ ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- --------
N. It L. It Time N. It L. It Time N. It L. It Time N. It L. It Time
64 2.8 44.5 307.5 2.7 52.1 588.6 2.7 58.7 801.4 2.6 63.5 1275.7
128 2.8 53.7 125.2 2.7 64.8 245.2 2.7 69.3 353.2 2.6 75.2 583.1
256 2.8 63.1 57.8 2.7 71.6 110.9 2.7 80.1 169.6 2.6 84.9 252.7
512 2.8 72.6 28.3 2.7 81.2 52.6 2.7 87.7 80.2 2.6 92.5 134.9
1024 2.8 84.9 18.9 2.7 94.5 37.2 2.7 99.4 58.5 2.6 104.7 96.3
2048 2.8 92.5 12.4 2.7 106.7 28.5 2.7 112.8 39.1 2.6 118.5 62.5
64 2.8 18.8 283.6 2.7 20.2 554.7 2.7 22.6 772.0 2.6 24.3 1236.8
128 2.8 21.5 104.5 2.7 24.5 218.1 2.7 27.3 312.4 2.6 29.7 545.9
256 2.8 25.8 42.6 2.7 28.9 91.5 2.7 30.6 137.9 2.6 33.5 227.5
512 2.8 29.4 22.8 2.7 31.7 46.5 2.7 34.3 74.1 2.6 36.4 118.4
1024 2.8 32.3 15.7 2.7 35.4 32.9 2.7 36.8 51.8 2.6 38.2 89.2
2048 2.8 37.6 11.0 2.7 39.2 25.3 2.7 40.2 35.7 2.6 42.5 58.3
-- ------ ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- --------
: [Strong scalability with different number of processors $N_p$ for Case-2.]{}
\[tab:strong-scale-case-4\]
The weak scalability is used to examine how the execution time varies with the number of processors when the problem size per processor is fixed. In the weak scaling test, we start with a small $2048\times2048$ mesh with the number of processors $N_p=64$ and end up with a large $16384\times16384$ mesh ($268,435,456$ degrees of freedom) using up to $4096$ processor cores for the 2D test case. Also, we further test our algorithms in terms of the weak scalability starting with a $96\times96\times96$ mesh and $N_p=216$ for Case-4. We refine the mesh and increase the number of processors simultaneously to keep the number of unknowns per processor as a constant. Table \[tab:weak-scale\] presents the results of weak scaling tests, which are run with fixed time step sizes $\triangle t=0.01$ and 0.5 days respectively, and are stopped after $5$ implicit time steps, i.e., the simulation are terminated at $t=0.05$ and $2.5$ days. We observe that, with the increase of the number of processors form $64$ to $4096$ and $216$ to $6859$ for the 2D and 3D test problems, a reasonably good weak scaling performance is obtained, especially for the two-level restricted Schwarz approach, which indicates that the proposed solver has a good weak scalability for this range of processor counts.
-- ------------------------- ------ ------- ------- ------- ------- ------- -------
N. It L. It Time N. It L. It Time
$2048\times2048$ 64 3.8 26.4 37.4 3.8 12.7 35.2
$4096\times4096$ 256 4.8 46.4 61.1 4.8 18.0 54.3
$6144\times6144$ 576 6.2 69.5 95.1 6.2 14.4 67.2
$8192\times8192$ 1024 7.2 96.3 136.6 7.2 19.6 90.0
$12288\times12288$ 2304 9.8 173.9 316.8 9.8 17.8 131.8
$16384\times16384$ 4096 12.4 294.7 607.8 12.4 17.0 197.1
$96\times96\times96$ 216 3.8 40.6 12.5 3.8 16.9 10.7
$128\times128\times128$ 512 4.2 56.3 15.0 4.2 15.9 12.2
$192\times192\times192$ 1728 5.8 90.3 27.6 5.8 35.9 22.5
$224\times224\times224$ 2744 6.8 107.7 35.2 6.8 38.2 28.1
$256\times256\times256$ 4096 7.8 122.9 68.7 7.8 43.2 37.6
$304\times304\times304$ 6859 9.4 147.4 85.9 9.4 44.7 54.6
-- ------------------------- ------ ------- ------- ------- ------- ------- -------
: Weak scalability with different number of processors.
\[tab:weak-scale\]
In summary, observing from the above tables and the figures, we highlight that, compared with the one-level method, the two-level restricted additive Schwarz method results in a very sharp reduction in the number of linear iterations and therefore brings about a good reduction in compute time. Hence, the two-level method is much more effective and scalable than the one-level approach in terms of the strong and weak scalabilities.
Conclusions {#conclusion}
===========
The simulation of subsurface flows with high resolution solutions is of paramount importance in reservoir simulation. In this work, we have presented a parallel fully implicit framework based on multilevel restricted Schwarz preconditioners for subsurface flow simulations with Peng-Robinson equation of state. The proposed framework can get rid of the restriction of the time step size, and is flexible and allows us to construct different type of multilevel preconditioners, based on plenitudinous choices for additively or multiplicatively strategies, interpolation and restriction operators. After experimenting with many different overlapping Schwarz type preconditioners, we found that the class of V-cycle-type restricted Schwarz methods based on the second order scheme is extremely beneficial for the problems under investigation. Numerical experiments also showed that the proposed algorithms and simulators are robust and scalable for the large-scale solution of some benchmarks as well as realistic problems with highly heterogeneous permeabilities in petroleum reservoirs.
Acknowledgment {#acknowledgment .unnumbered}
==============
[[The authors would like to express their appreciations to the anonymous reviewers for the invaluable comments that have greatly improved the quality of the manuscript.]{}]{}
[00]{}
S. Balay, J. Brown, K. Buschelman, V. Eijkhout, W. Gropp, D. Kaushik, M. Knepley, L.C. McInnes, B.F. Smith, H. Zhang, PETSc Users Manual, Argonne National Laboratory, 2019.
Q.M. Bui, H.C. Elman, J.D. Moulton, Algebraic multigrid preconditioners for multiphase flow in porous media, SIAM J. Sci. Comput. 39 (2017) S662–S680.
Q.M. Bui, L. Wang, D. Osei-Kuffuor, Algebraic multigrid preconditioners for two-phase flow in porous media with phase transitions, Adv. Water Resour. 114 (2018) 19–28.
X.-C. Cai, M. Dryja, M. Sarkis, Restricted additive Schwarz preconditioners with harmonic overlap for symmetric positive definite linear systems, SIAM J. Numer. Anal. 41 (2003) 1209–1231.
X.-C. Cai, W.D. Gropp, D. E. Keyes, R. G. Melvin, D. P. Young, Parallel Newton–Krylov–Schwarz algorithms for the transonic full potential equation, SIAM J. Sci. Comput. 19 (1998) 246–265.
X.-C. Cai, M. Sarkis, A restricted additive Schwarz preconditioner for general sparse linear systems, SIAM J. Sci. Comput. 21 (1999) 792–797.
X.-C. Cai, D.E. Keyes, Nonlinearly preconditioned inexact Newton algorithms, SIAM J. Sci. Comput. 24 (2002)183–200.
Z. Chen, G. Huan, Y. Ma, Computational Methods for Multiphase Flows in Porous Media, SIAM, Philadelphia, PA, 2006.
Z. Chen, Reservoir Simulation: Mathematical Techniques in Oil Recovery, SIAM, Philadelphia, PA, 2007.
M. Christie, M. Blunt, Tenth SPE comparative solution project: A comparison of upscaling techniques. In SPE Reservoir Simulation Symposium. Society of Petroleum Engineers. (2001) 308–317.
C.N. Dawson, H. Klíe, M.F. Wheeler, C.S. Woodward, A parallel, implicit, cell-centered method for two-phase flow with a preconditioned Newton–Krylov solver, Computat. Geosci. 1 (1997) 215–249.
J.E. Dennis, R.B. Schnabel, Numerical Methods for Unconstrained Optimization and Nonlinear Equations, SIAM, Philadelphia, 1996.
A. Firoozabadi, Thermodynamics of Hyrocarbon Reservoirs, McGraw-Hill, New York, 1999.
J.B. Haga, H. Osnes, H.P. Langtangen, A parallel block preconditioner for large-scale poroelasticity with highly heterogeneous material parameters, Comput. Geosci. 16 (2012) 723–734.
D.A. Knoll, D.E. Keyes, Jacobian-free Newton–Krylov methods: a survey of approaches and applications, J. Comput. Phys. 193 (2004) 357–397.
F. Kong, X.-C. Cai, A highly scalable multilevel Schwarz method with boundary geometry preserving coarse spaces for 3D elasticity problems on domains with complex geometry, SIAM J. Sci. Comput. 38 (2016) C73–C95.
S. Lacroix, Y.V. Vassilevski, J.A. Wheeler, M.F. Wheeler, Iterative solution methods for modeling multiphase flow in porous media fully implicitly, SIAM J. Sci. Comput. 25 (2003) 905–926.
J.J. Lee, K.-A. Mardal, R. Winther, Parameter-robust discretization and preconditioning of biot’s consolidation model, SIAM J. Sci. Comput. 39 (2017) A1–A24.
K. Lie, S. Krogstad, I. Ligaarden, J. Natvig, H. Nilsen, B. Skaflestad, Open-source MATLAB implementation of consistent discretisations on complex grids, Comput. Geosci. 16 (2012) 297–322.
L. Liu, D.E. Keyes, R. Krause, A note on adaptive nonlinear preconditioning techniques, SIAM J. Sci. Comput. 40 (2018) A1171–A1186.
Y. Liu, H. Yang, Z. Xie, P. Qin, R. Li, Parallel simulation of variably saturated soil water flows by fully implicit domain decomposition methods, J. Hydrol. 582 (2020) 124481.
J.E.P. Monteagudo, A. Firoozabadi, Comparison of fully implicit and IMPES formulations for simulation of water injection in fractured and unfractured media, Int. J. Numer. Meth. Engng. 69 (2007) 698–728.
E.E. Prudencio, X.-C. Cai, Parallel multilevel restricted Schwarz preconditioners with pollution removing for PDE-constrained optimization, SIAM J. Sci. Comput. 29 (2007) 964–985.
C. Qiao, S. Wu, J. Xu, C.S. Zhang, Analytical decoupling techniques for fully implicit reservoir simulation. J. Comput. Phys. 336 (2017) 664–681.
Y. Saad, Iterative Methods for Sparse Linear Systems. SIAM, [S]{}econd ed., 2003.
J.O. Skogestad, E. Keilegavlen, J.M. Nordbotten, Domain decomposition strategies for nonlinear flow problems in porous media, J. Comput. Phys. 234 (2013) 439–451.
G. Singh , M.F. Wheeler, A space-time domain decomposition approach using enhanced velocity mixed finite element method, J. Comput. Phys. 374 (2018) 893–911.
B. Smith, P. Bj[ø]{}rstad, W. Gropp, Domain [D]{}ecomposition: [P]{}arallel [M]{}ultilevel [M]{}ethods for [E]{}lliptic [P]{}artial [D]{}ifferential [E]{}quations, Cambridge University Press, 1996.
A. Toselli, O. Widlund, Domain Decomposition Methods-Algorithms and Theory, Springer, Berlin, 2005.
J.R. Wallis, Incomplete gaussian elimination as a preconditioning for generalized conjugate gradient acceleration, in SPE Reservoir Simulation Symposium, Society of Petroleum Engineers, 1983.
J.R. Wallis, R.P. Kendall, L.E. Little, Constrained residual acceleration of conjugate residual methods, in SPE Reservoir Simulation Symposium, Society of Petroleum Engineers, 1985.
K. Wang, H. Liu, Z. Chen, A scalable parallel black oil simulator on distributed memory parallel computers, J. Comput. Phys. 301 (2015) 19–34.
J.A. White, N. Castelletto, H.A. Tchelepi, Block-partitioned solvers for coupled poromechanics: A unified framework, Comput. Methods Appl. Mech. Engrg. 303 (2016) 55–74.
H. Yang, F.-N. Hwang, X.-C. Cai, Nonlinear preconditioning techniques for full-space Lagrange-Newton solution of PDE-constrained optimization problems, SIAM J. Sci. Comput. 38 (2016) A2756–A2778.
H. Yang, Y. Li, S. Sun, Nonlinearly preconditioned constraint-preserving algorithms for subsurface three-phase flow with capillarity, Comput. Methods Appl. Mech. Engrg. 367 (2020) 113140. H. Yang, S. Sun, Y. Li, C. Yang, A scalable fully implicit framework for reservoir simulation on parallel computers, Comput. Methods Appl. Mech. Engrg. 330 (2018) 334–350.
H. Yang, S. Sun, Y. Li, C. Yang, A fully implicit constraint-preserving simulator for the black oil model of petroleum reservoirs, J. Comput. Phys. 396 (2019) 347–363. H. Yang, S. Sun, C. Yang, Nonlinearly preconditioned semismooth Newton methods for variational inequality solution of two-phase flow in porous media, J. Comput. Phys. 332 (2017) 1–20.
H. Yang, C. Yang, S. Sun, Active-set reduced-space methods with nonlinear elimination for two-phase flow problems in porous media, SIAM J. Sci. Comput. 38 (2016) B593–B618.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this article we propose some Maple procedures, for teaching purposes, to study the basics of General Relativity (GR) and Cosmology. After presenting some features of GRTensorII, a package specially built to deal with GR, we give two examples of how one can use these procedures. In the first example we build the Schwarzschild solution of Einstein equations, while in the second one we study some simple cosmological models. \
[**Keywords:**]{} general relativity $\bullet$ GRTensorII $\bullet$ cosmology
[**PACS (2010):**]{} 98.80.-k $\bullet$ 04.30.-w $\bullet$ 01.40.Ha\
author:
- |
[**Ciprian A. Sporea[^1] , Dumitru N. Vulcanov**]{}[^2]\
\
[West University of Timişoara, Faculty of Physics,]{}\
[V. Parvan Ave. no. 4, 300223, Timişoara, Romania]{}\
title: '**Using Maple + GRTensorII in teaching basics of General Relativity and Cosmology**'
---
Introduction
============
Cosmology (i.e. the modern theory of Universe dynamics) has became in the last two decades an attractive field of human knowledge including also an intense media campaign. This was possible, among other reasons, also because in this time period several space missions performed cosmological measurements -like COBE, WMAP, Plank, BICEP2- thus transforming cosmology from a pure theoretical field also into an experimental one. Thus teaching cosmology, even at undergraduate level, in physics faculties comes to be a compulsory topic. Cosmology is based on two major pillars [@8],[@13]: astrophysics as a phenomenological tool and general relativity (GR) as the main theory, thus making it, unfortunately, difficult to teach to undergrad students. The mathematical structure or GR is based on differential geometry [@9], [@10] and learning it means that the student must first get familiar with the main instruments of differential geometry (such as tensor calculus on curved manifolds, Riemannian curvature, metric connection, covariant derivative, etc), which often means cumbersome and lengthy hand calculations. To illustrate these facts let us remind that GR is based on field equations known as Einstein equations, namely $$\label{EE}
R_{\mu\nu}-\frac{1}{2}g_{\mu\nu}R+ \Lambda g_{\mu\nu}=-\kappa T_{\mu\nu}$$ where $R_{\mu \nu}$ is the Ricci tensor, $g_{\mu\nu}$ the metric tensor and $T_{\mu\nu}$ represents the stress-energy tensor. We denoted by $\Lambda$ the cosmological constant and $\kappa=8\pi G/c^4$. The components of the Ricci tensor are given by [@8],[@9] $$\label{Ricci}
R_{\mu\nu}=\partial_\lambda\Gamma^\lambda_{\ \mu\nu}-\partial_\nu\Gamma^\lambda_{\ \mu\lambda}+\Gamma^\lambda_{\ \mu\nu}\Gamma^\sigma_{\ \lambda\sigma}-\Gamma^\sigma_{\ \mu\lambda}\Gamma^\lambda_{\ \nu\sigma}$$ where $\Gamma^\lambda_{\ \mu\nu}$ are the so called Chrisstoffell symbols wich in Riemannian geometry [@10] describe the structure of the curved space-time underlying GR and the associated metric tensor is compatible with the connection described by the above Chrisstoffel symbols, namely $$\label{Chris}
\Gamma^\lambda_{\ \mu\nu} = g^{\lambda\sigma}\left( \partial_\mu g_{\nu\sigma} + \partial_\nu g_{\mu\sigma} - \partial_\sigma g_{\mu\nu}\right)$$ All these make even the most determined students to lose interest in studying cosmology (and GR too). Dozens of pages with hundreds of terms containing partial differentials to be hand processed could scare any student (and not only!).
As today students are by passing day more skilled in computer manipulation and with a more and more advanced practice in programming, the modern teaching of GR and cosmology should use intensively computer facilities for algebraic programming, tensor manipulation and of course numerical and graphical facilities [@11], [@14], [@15]. The use of computer algebra was in the view of physicists even since the beginning of computer science both for teaching and research purposes. Computer algebra (or algebraic programming codes) evolved from early days of REDUCE package (see for example [@19],[@20], [@11]) till recent developments using integrated platforms as Maple and Mathematica in different fields of physics, not only in general relativity (see for example [@11], [@17], [@18] and more recently [@22], [@21]).
Some years ago [@4] we published our experience in this using the REDUCE platform. Unfortunately, in the last years REDUCE lost the market in favour of more intergraded and visual platforms such as Maple [@1] and Mathematica, thus we adapted our experience and program packages to Maple and we’ve made use of the free package GTTensorII [@2] adapted for doing GR.
The aim of the present article is to report our new experience in this direction. The article is organised as follows. The next section introduces the main features of GRTensorII in doing tensorial symbolic computation in GR and Riemannian geometry. Section no. 3 describes the way we can obtain an exact solution of Einstein equations. We used again, as in the main classical text on GR, the Schwarzschild solution. This is the most famous solution used today intensively in describing the motion in the solar system and for studying black-holes physics [@12]. The last section is dedicated to describe how one can use Maple and GRTensorII for cosmology (and teaching it). In both these two above sections we gave the main Maple commands which can be put together to have short programs to be used during the computer lab hours and even during the lectures. The article ends with a short section where we present the main conclusions and some ideas for future developments.
Short presentation of the GRTensorII package
=============================================
GRTensorII is a computer algebra package built within the Maple platform as a special set of libraries [@1]. It is a free distributed package (see [@2]) and it is adapted for dealing with computer algebra manipulation in general relativity. Thus it is designed for dealing with tensors and other geometric objects specific to Riemannian geometry (a metric that is compatible with the connection, symmetric connection - torsion free manifolds). In what follows, we will present some of the main features offered by GRTensorII. The library is based on a series of special commands all starting with $"gr"$ (for example $grcalc, grdisplay, gralter, grdefine$, etc) for dealing with a series of (pre)defined geometric objects such as the metric tensor, Ricci tensor and scalar, Einstein tensor, Chrisstoffell symboles, etc.
To start the GRTensorII package one must type in a (new) Maple session the following commands
> restart; > grtw();
The restart command causes the Maple kernel to clear its internal memory so that Maple acts (almost) as if just started. The second command initializes the GRTensorII package and gives some information about the current version.
The GRTensorII library allows us to build our own space-time manifold. The most easiest way to do this is by creating a metric tensor $ g_{\mu\nu} $ with the help of $>makeg( )$ command. The result of operating this command will be a special ASCII file containing the main information about the metric and stored in a folder called “metrics” within the GRTensorII library. The folder “metrics” contains also a collection of predefined metric files distributed with the package (for egz. schw.mpl, schmidt.mpl, vdust.mpl, etc).
Another way in which one can specify a space-time manifold is by loading a predefined metric from the “metrics” folder. This can be done with the help of two commands
> qload(metricName);
> grload(metricName, meticFile);
The geometry built using $>makeg()$, $>qload() $ and $>grload()$ commands will fix the background on which all the later operations and calculations will be performed.
One of the main advantages of GRTensorII library is that it allows us to do complicated operations on tensorial objects, regardless of how many indices those objects poses. The main command that permit us to do those calculations is
$>grcalc( objectSeq )$ that calculates the components of tensors;
For example $>grcalc ( R(dn,dn), R(up,dn,dn,dn) )$ ask the program to calculate the covariant components of the Ricci tensor $ R_{\mu\nu} $ and the components of the standard curvature Riemann tensor $ R^{\sigma}_{\mu\nu\lambda} $.
The command $ >grdisplay()$ can be used to display the components of GRTensorII objects which have been previously calculated for a particular space-time. Before displaying the calculated components of an object it is indicated to use the command $>gralter()$ in order to simplify them. The $>grcalc()$ calculates all the components of a given object, so if one wants to calculate only a specific component then it can use the command $>grcalc1 ( object, indexList )$. For example: $>gecalc1 ( R(dn,dn,dn,dn), [t, r, \theta, \phi])$.
Besides the predefined objects that exist in GRTensorII we can also define new objects (scalars, vectors, tensors) with the help of the command $>grdef()$. For example
$>grdef ( `G2\{a\ \ b\} := R\{a\ \ b\} - (1/2)*Ricciscalar*g\{a\ \ b\} + Lambda*g\{a\ \ b\}`)$
defines a contravariant two index tensor, $G2_{a b}$, which it is explicitly assigned to an expression involving a number of previously defined (or predefined) tensors. The syntax in $> grdef()$ command follows naturally the usual tensorial operations which defines the new objects.
Another important command of GRTensorII is $> grcoponent()$ which allows us to extract a certain component of a tensorial object. The extracted component can be used as a standard Maple object for later processing (symbolically, graphically and numerically).
Although GRTensorII was designed initially for Riemannian differential geometry it can be easily extended to other types of geometries, such as ones with torsion or higher order alternative theories of gravity [@5], [@6].
We end this section by making the important observation that using GRTensorII does not impose any restriction in using all the numerical, graphical and symbolic computation facilities of Maple (as it happens with other packages even for Maple). Thus we can combine al these facilities for an efficient use of the Maple platform.
Example 1: Schwarzschild type solutions
=======================================
General relativity and its applications (such as cosmology) are based on Einstein equations (\[EE\]) as main field equations. They have many exact solutions, although these second order nonlinear differential equations have no unique analytical general solution. The most famous exact solution of Einstein equations is the Schwarzschild solution [@8], [@9] describing the gravitational field around a pointlike mass M (or outside a sphere of mass M). This solution is used today for describing the black-hole dynamics and was used in the first attempts in applying GR to the motion of planets and planetoids in our solar system.
It is obvious that from a pedagogical point of view finding an exact solution of Einstein equations could be a good introductory lesson in applications of GR. Next we will derive this solution following the natural steps :
- identifying the symmetries of the system for which we build the solution;
- building a metric tensor compatible with the above symmetry;
- building the shape of the stress-energy tensor components (if any exists);
- calculating the Ricci tensor and the components of Einstein equations;
- solving the above equations after a close inspection of them.
These above steps could be done manually and usually it takes several hours of hard calculations (even straightforward and even for an experienced person). Our advise for anyone who wants to teach GR and/or cosmology is to do this traditional step with the students. It will be a good lesson and a motivation to proceed in using algebraic computing facilities (here Maple+GRTensorII).
Thus the above steps are clearly transposable in computer commands in Maple+ GRTensorII. For the Schwarzschild solution the symmetry is clearly spherical and static (no time dependence including the time inversion, namely $t \rightarrow -t$ ). Thus we will use a spherical symmetric metric tensor as [@8]: $$\label{sferic}
ds^2=e^{2\lambda(r)}dt^2 + e^{2\mu(r)}dr^2 + r^2\left( d\theta^2 + sin^2(\theta)d\phi^2\right)$$ in spherical coordinates $(t,r,\theta,\phi)$, where $\lambda(r)$ and $\mu(r)$ are the two unknown functions of the radial coordinate $r$ to be found at the end. Thus the student already in front of a computer or station having started a Maple session will be guided to compose the next sequence of commands
> restart; grtw();
> makeg(sferic);
>......
> grdisplay(metric); grdisplay(ds);
where after the two commands for starting the GRTensortII the command $> makeg$ will create the ASCII file $sferic.mpl$ containing the information on the metric we will built. The series of dots above represent those steps where the user has to answer with the type of the metric, symmetry and of course its components one by one. The last two lines given above are calculating the metric and displays its shape in the form of a matrix. After this we can continue to do some calculations or to close the session. The metric we produce can be loaded anytime later in another sessions.
The next step of our demonstrative program will be to point out and calculate the Einstein equations, but not before introducing the stress-energy components. It is obvious that in this case the strass-energy components are cancelled as we calculate the gravity field outside the source (the pointlike mass or a sphere). Thus in this case we will solve the so called vacuum Einstein equations i.e. $R_{\mu \nu}=0$ [@9].
In this view we will built a sequence of Maple commands (for a new session):
> restart; grtw();
> qload(sferic); grcalc(R(dn,dn));
> gralter(R(dn,dn),simplify); grdisplay(R(dn,dn));
where after loading the metric tensor with the command $>qload$ we calculate the Ricci tensor components, simplify and display them as a 4x4 matrix (using the last $>grdisplay$ command).
It is a good and interesting experience, before proceeding with the solving of Einstein eqs. to insert in the above lines the next ones
> grcalc(Chr(up,dn,dn)); grdisplay(Chr(up,dn,dn));
immediately after loading the metric tensor with $> qload$ which very fast calculates and displays the 40 components of the Chrisstoffel symbols using relation (\[Chris\]).
Of course if we want maximum effect of these, before calculating with Maple and GRTensorII we advise the teacher to calculate, by hand, together with the students at least 3 of 4 of these components. This will take some time of hard calculations and many mistakes when done for the first time.
To continue it is now much more simple to extract the components of the Ricci tensor one by one as Maple objects in order to process them to solve the obtained equations. It will be a sequence of $>grcomponent$ commands, namely :
> ec0:=grcomponent(R(dn,dn),[t,t]);
> ec1:=grcomponent(R(dn,dn),[r,r]);
> ec2:=grcomponent(R(dn,dn),[theta,theta]);
> ec3:=grcomponent(R(dn,dn),[phi,phi]);
obtaining the four Einstein equations of the problem. The rest of the Ricci tensor components are zero.
A simple inspection of the above obtained four equations reveals that only two of them are independent. Also we can eliminate the second order derivative of the $\lambda(r)$ function between $ec0$ and $ec1$. These can be checked by using the next command lines
> expand(simplify(ecu2-ecu3/sin(theta)^2));
> ecu0;ecu1;ecu2;
> l2r:=solve(subs(diff(lambda(r),r,r)=l2r,ecu0),l2r);
> expand(simplify(subs(diff(lambda(r),r,r)=l2r,ecu0)));
> ecu11:=expand(simplify(subs(diff(lambda(r),r,r)=l2r,ecu1)));
The first of the above commands checks the equality between $ecu2$ and $ecu3$, which simply gives a “0” (zero) and the next ones simply display the remaining three equations. The command that follows extracts the second order derivative of $\lambda(r)$ from $ecu0$ (substituting it with an intermediate constant $lr2$) and the next one substitutes the result in $ecu1$ and we obtained a new equation $ecu11$. Thus we have now only two equations, $ecu11$ and $ecu2$, namely $$\label{ecu11}
\frac{2}{r}\partial_r\lambda(r)+\frac{2}{r}\partial_r\mu(r)=0$$ $$\label{ecu2}
1-e^{-2\mu(r)}+ \left[ r\,\partial_r\mu(r)-r\,\partial_r\lambda(r) \right]e^{-2\mu(r)}=0$$ These two differential equations need to be solved next, in order to obtain the solution. Of course we can now follow a classical strategy solving them manually as is done in any textbook (see [@8] for example). But it is also possible to continue with Maple, using the $>dsolve$ command for solving differential equations (including systems of differential equations). Thus we write the next command
> dsolve({ecu11,ecu2},{mu(r),lambda(r)});
which gives us the function $\mu(r)$ as $$\begin{aligned}
\mu(r)=\frac{1}{2} ln \left( \frac{r}{r e^{C1}
-1} \right) +\frac{C1}{2}
\nonumber\end{aligned}$$ In particular a close inspection of both equations $(ecu11, ecu2)$ reveals that $ecu11$ is simply a relation between the derivatives of the two functions, namely $$\begin{aligned}
\frac{d \lambda (r)}{d r} +\frac{d \mu (r)}{d r} =0 \nonumber\end{aligned}$$ This shows that we will need only one integration constant and we can solve the equations only for one function as done above. One can write the constant $C1$ as $$\begin{aligned}
C1= ln \left( \frac{1}{r_s} \right) \nonumber\end{aligned}$$ where the new introduced constant $r_s$ will be determined later. With these we can rewrite the $ecu2$ and applying again the $>dsolve$ command on it we obtain in the Schwarzschild (type) solution as $$\label{FSS}
ds^2 = \left(1-\frac{r_s}{r} \right)c^2dt^2 + \left(1-\frac{r_s}{r}\right)^{-1}dr^2 + r^2\left[d\theta^2 + sin^2(\theta)d\phi^2 \right]$$ The constant $r_s$ is known under the name of Schwarzschild radius and can be determined using the newtonian limit of the field equations as it is done in any textbook (see [@8] for example). The precise value of it is $r_s= 2MG/c^2 $ but this has nothing to do with algebraic computing.
Example 2: Simple cosmological models
=====================================
Modern cosmology is based on general relativity and Einstein equations, from which we can derive the so called Friedman equations. The latter ones form the core of all cosmological models. The most used metric for describing the dynamics of the universe in a cosmological model is the Friedman-Robertson-Walker metric (FRW), which in spherical coordinates has the following line element [@9] $$\label{cm1}
ds^2=c^2dt^2-a^2(t)\left[\frac{dr^2}{1-kr^2}+r^2(d\theta^2+\sin^2\theta d\phi^2)\right]$$ where $k$ is the curvature constant and we are using the $(+,-,-,-)$ signature for the metric. Usually, this $k$ constant is taken to be $1$ (for closed universes), $-1$ (for open universe) and $0$ for a flat one. In (\[cm1\]) we denoted by $a(t)$ the scale factor, that in the end will be the only unknown function of a cosmological model. This scale factor is directly related to the evolution of the universe. For the FRW metric (obtained form the cosmological principle i.e. the universe is spatially homogenous and isotropic) the scale factor is a function only on time. By introducing FRW metric (\[cm1\]) into the Einstein equations (\[EE\]) and assuming that the energy-momentum tensor is of a prefect fluid form [@8] $$\label{rm3}
T^{\mu\nu}=\left( \rho+\frac{p}{c^2} \right) u^\mu u^\nu-p g^{\mu\nu}$$ one arrives to the Friedman-Lemaitre equations $$\label{rm4}
\begin{split}
&\ddot a=-\frac{4\pi G}{3}\left(\rho+\frac{p}{c^2} \right)a + \frac{1}{3}\Lambda c^2a\\
&\dot a^2=\frac{8\pi G}{3}\rho a^2+\frac{1}{3}\Lambda c^2a^2-c^2k
\end{split}$$ If the cosmological constant $\Lambda$ is set to zero in eqs.(\[rm4\]) then the equations are called simply the Friedman equations. In eq. (\[rm3\]) $u^\mu$ represnets the 4-velocity of the cosmological fluid, while $p$ and $\rho$ stands for the pressure, respectively the mass density of the fluid.
In the same manner as done in [@3] we can compose a sequence of GRTensorII commands for obtaining the Friedman equations (\[rm4\]). A student can write the program on a computer in less than an hour to the a job that if it is done by hand calculations it will take several good hours to a very good student. The basic lines of the GRTensorII program are as follows:
> restart; > grtw(); qload(metrica_FRW);
> grdef(`u{^a}:=[1,0,0,0]`);
> grdef(`T{^a ^b}:=(rho(t)+p(t)/c^2)*u{^a}*u{^b}-p(t)*g{^a ^b}`);
> ...
> ec0:=R1_0-T1_0; > ec1:=R1_1-T1_1;
> ec1:=subs(diff(a(t),t,t)=-(1/6)*K*c^4*rho(t)*a(t)-
(1/2)*K*c^2*p(t)*a(t)+(1/3)*Lambda*c^2*a(t),ec1): ec0; ec1;
The first line of commands starts the GRTensorII and loads the FRW metric (\[cm1\]). In the next lines we define the 4-velocity and the stress-energy tensor (\[rm3\]) of the cosmological fluid. Now follows a series of more technical commands which can be found in the supplementary web material [@16], commands that allows us to calculate and write the final form of Friedman equations (\[rm4\]).
Friedman equations (\[rm4\]) are in fact a system of two differential equations with three unknowns: $a(t)$, $p$ and $\rho$. Thus one needs to find a third equation in order to completely solve the problem. In standard cosmology we use as a third relation the equation of state $$\label{rm5}
p(t)=w\rho(t)c^2$$ where $w$ is a constant ( $w=0$ for pressureless ’dust’, $w=1/3$ for radiation and $w=-1$ for vacuum).
Let us further introduce the dimensionless quantities (see for example [@9]), usually called density parameters, which are defined by $$\label{rm5}
\Omega_i(t)\equiv \frac{8\pi G}{3H^2(t)}\rho_i(t)$$ where $ H(t)=\dot a(t)/a(t)$ is known as the Hubble parameter and $i$ stands for matter, radiation and the cosmological constant $\Lambda$. Besides these three quantities one can also define a curvature density parameter $$\label{rm5}
\Omega_k(t)= -\frac{c^2k}{H^2(t)a^2(t)}$$ Rewriting the second equation of (\[rm4\]) in terms of the new defined density parameters we arrive at a very simple expression
$$\label{rm6}
\Omega_m + \Omega_r + \Omega_{\Lambda} + \Omega_k = 1$$
Taking all the above into account and introducing a normalised scale factor $ A(t)=a(t)/a_0(t_{now})$ (with $a_0(t_{now})$ representing the value of the scale factor at the present epoch) we can finally write an equation for the evolution of the scale factor, namely $$\label{rm7}
H^2(t)=H_0^2(\Omega_{r,0}a^{-4}+ \Omega_{m,0}a^{-3}+ \Omega_{k,0}a^{-2} +\Omega_{\Lambda,0})$$ where $\Omega_{i,0}$ are the values of the densities measured at the present epoch.
Below we give the main GRTensorII command lines with the help of which we arrive at equation (\[rm7\]). The complete sequence of commands can be found in the supplementary web material [@16].
> ec1:=subs(diff(a(t),t)=H(t)*a(t),ec1);
> rho(t):=rho_m0*(a_0/a(t))^3+rho_r0*(a_0/a(t))^4;
> ec1:=subs(Omega_k0=1-Omega_m0-Omega_r0-Omega_Lambda0,ec1);
We can now use the other numerical and computational facilities of Maple in order to numerically solve equation (\[rm7\]) and express the results as plots (see Fig. 1) of the scale factor as a function of cosmic time. For that we use the following code (see also the supplementary web material [@16]):
> ecu:=diff(A(t), t)-sqrt(Omega_m0/A(t)+Omega_r0/A(t)^2+
Omega_Lambda0*A(t)^2+1-Omega_m0-Omega_r0-Omega_Lambda0);
> ecu_a:=subs(Omega_m0 =0.3,Omega_Lambda0=0.7,Omega_r0=0,ecu);
> sys1:={ecu_a,A(0)=1}: > f1:=dsolve(sys1,numeric):
> odeplot(f1,t=-2..2,axes=boxed,numpoints=1000,color=black);
[0.47]{} ![Time evolution of the scale factor[]{data-label="fig1ab"}](plot3 "fig:"){width="\textwidth"}
[0.47]{} ![Time evolution of the scale factor[]{data-label="fig1ab"}](plot2 "fig:"){width="\textwidth"}
Conclusions and further developments
====================================
The article describes a way in which some simple computer programs in Maple and GrTensorII can be used in teaching GR and cosmology. It is obvious that the speed of learning main concepts in GR (and subsequently differential geometry) can be successfully enhanced, avoiding large hand computation steps and a lot of natural mistakes. On the other hand it is clear that the lectures will need to take place in a computer lab, which can be another way to increase the attractiveness of GR and cosmology. We illustrated our experience with short and simple commands without sophisticated tricks normally a professional in the field is using (for example building procedures and libraries via the symbolic computation facilities of Maple).
The small and short programs we described here can be also used as a strong basis for further developments in view of more sophisticated and advanced examples. For instance one can develop the above procedures for cosmology in generalised theories of gravity, like those with higher order Lagrangians (as we done in [@6]).
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by a grant of the Romanian National Authority for Scientific Research, Programme for research-Space Technology and Advanced Research-STAR, project nr. 72/29.11.2013 between Romanian Space Agency and West University of Timisoara.
C.A.Sporea was supported by the strategic grant POSDRU/159/1.5/S/137750, Project “Doctoral and Postdoctoral programs support for increased competitiveness in Exact Sciences research” cofinanced by the European Social Found within the Sectorial Operational Program Human Resources Development 2007 – 2013.
[99]{}
S. Weinberg, “Cosmology”, Oxford University Press,2008.
B. Schutz, “A first curse in General Relativity”, Cambridge University Press, 2009.
M. P. Hobson, G. P. Efstathiou, A. N. Lasenby, “General Relativity, An Introduction for Physicists”, Cambridge University Press, 2006.
L.I. Nicolaescu, “Lectures on the Geometry of Manifolds”, World Scientific, 1996.
J. Grabmeier, E. Kaltofen, V. Weispfenning, “Computer Algebra Handbook: Foundations · Applications · Systems”, Springer, 2003.
F.W. Hehl; R.A. Puntigam; H. Ruder, “Relativity and Scientific Computing ”, Springer, 1996.
D. Stauffer; F.W. Hehl; N. Ito; V. Winkelmann; J.G. Zabolitzky, “Computer Simulations and Computer Algebra”, Springer-Verlag, 1993.
R.A. d’Inverno, “Algebraic computing in general relativity”, GRG 6(6), 576-593 (1975).
C. Heinicke, F.W. Hehlhel, “Computer algebra in gravity”, http://arxiv.org/abs/gr-qc/0105094
F.Y. Wang, “Physics with MAPLE: The Computer Algebra Resource for Mathematical Methods in Physics”, WILEY-VCH Verlag GmbH Co. KGaA,Weinheim, 2006.
F.Y.-H. Wang, “Relativistic orbits with computer algebra”, Am. J. Phys. 72,1040 (2004).
Maple for Physics Students: Complete Set of Lectures.
http://www.maplesoft.com/applications/view.aspx?SID=4743
Trandafir et al., “Elementary tight-binding method for simple electronic structure calculations - An educational approach to modeling conjugated dyes for dye-sensitized solar cells”, Rom. Rep. Phys. 66, 574 (2014).
F.A. Ghergu, D.N. Vulcanov, “The Use of the algebraic programming in teaching general relativity”, Comput.Sci.Eng. 3 (2001) 65-70.
Maple User Manual, Waterloo Maple Inc.2005.
http://grtensor.phy.queensu.ca/
I.I. Cotaescu, C. Crucean, C.A. Sporea, “Elastic scattering of Dirac fermions on Schwarzschild black holes”, arXiv:1409.7201.
D.N. Vulcanov, “Calculation of the Dirac equation in curved spacetimes with possible torsion using MAPLE and REDUCE”, Comp. Phys. Comm., 154 (3), pp. 205-218.
D.N. Vulcanov, G.S. Djordjevic, C.A. Sporea, “REM - the Shape of Potentials for f(R) Theories in Cosmology and Tachyons”, in Cosmology and Particle Physics beyond Standard Models. Ten Years of the SEENET-MTP Network, edited by Luis Alvarez-Gaume, Goran S. Djordjevic, Dejan Stojkovc, CERN-Proceedings-2014-001, pp. 165-169.
D.N. Vulcanov, V. D. Vulcanov, “Maple+GrTensorII libraries for cosmology” - http://arxiv.org/abs/cs/0409006v1
D.N. Vulcanov, G.S. Djordjevic, “On cosmologies with non-minimally coupled scalar fields, the ’reverse engineering method’ and the Einstein frame”, Rom.J.Phys. 57 (2012) 1011-1016.
The Supplementary Materials consist of a Maple worksheet illustrating the calculation of Einstein and Friedman equations for some simple cosmological models.
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Fully Developed Turbulence (FDT) occurs at the infinite extreme of the Reynolds spectrum. It is a theoretical phenomenon which can only be approximated experimentally or computationally, and thus its precise properties are only hypothetical, though widely accepted. It is considered to be a chaotic yet steady flow field, with self-similar fractalline features. A number of approximate models exist, often exploiting this self-similarity. We hereby present the exact mathematical model of Fractal Potential Flows, and link it philosophically to the phenomenon of FDT, building on its experimental characteristics. The model hinges on the recursive iteration of a fluid dynamical transfer operator. We show the existence of its unique attractor in an appropriate function space - called the invariant flow - which will serve as our model for the FDT flow field. Its sink singularities are shown to form an IFS fractal, resolving Mandelbrot’s Conjecture. Meanwhile we present an isometric isomorphism between flows and probability measures, hinting at a wealth of future research.[^1]\
[: 76F20 (primary); 76F25, 37A30 (secondary)]{}
[: turbulence, chaotic advection, attractors, fractals, invariant measure]{}
author:
- |
József Vass\
Department of Applied Mathematics\
University of Waterloo\
[email protected]
bibliography:
- 'mybib.bib'
date: 'December 20, 2012'
title: |
Fractal Potential Flows as an Exact Model for\
Fully Developed Turbulence
---
Introduction {#s01}
============
We begin our discussion in Section \[s0201\] with an extensive overview of Fully Developed Turbulence (FDT), in an attempt to identify the definitive characteristics of this hypothetical phenomenon at the infinite end of the Reynolds spectrum, which can only be approximated and observed via experiments or simulation. We then proceed to give an account of the beautiful idea of Potential Flows and their superposition, heretofore seemingly unrelated to FDT, yet intriguingly sharing its inviscid and steady characteristics. We proceed to detailing the influential direction of non-deterministic experimental Fluid Dynamics, called Chaotic Advection, which we shall reformulate in the language of Optimal Transport Theory. Since the well-known IFS invariant measure is the fixed point of the iterative evolution under the measure transfer operator, it will serve as our inspiration for showing the unique existence of an invariant flow. Since the transfer operator distributes singularities over an IFS fractal, this adds impetus to pursuing a transfer model, considering Mandelbrot’s Conjecture. In our remarks on this overview, we make a decisive step to discard Navier-Stokes evolution on mathematical grounds, observing the non-smooth nature of intermittency, and turn our attention to discrete inertial transfer evolution on piecewise steady flows.
The fusion of the above theories will come naturally and effortlessly in the form of Fractal Potential Flows, introduced in Section \[s03\]. The random intermittent interaction of a finite system of eddies, is directly modelled via the recursive iteration of a weighted transfer operator in Section \[s0302\]. The contractive action of each eddy sink on the flow field, is modelled with a corresponding pushforward transfer map, induced by a contraction map of an IFS. Since the collective action of the eddies is probabilistically weighted, it can be interpreted as an “expected value” expression, while physically speaking as the weighted effect each eddy has on the flow field. The iteration of the transfer operator, is reasoned to correspond to the non-smooth discrete energy bursts of intermittency.
The evolution equation towards FDT having been reasoned to be the discrete transfer recursion, our goal consequently becomes to show that it converges to a unique attractor flow - the FDT flow field itself - in some appropriate function space. In finding the correct space, eddies remain our guiding inspiration, considering that a system of eddies becomes a superposed eddy when zoomed out, possessing a complex character at complex infinity as discussed in Section \[s0303\]. In their local universe, the eddies interact to produce an increasingly fractalline flow field, approaching the desired invariant flow shown in Section \[s0305\]. The required iteration number to reach FDT is infinite, while the total time can be potentially finite, if the time-spacing is for instance in geometric regression, as discussed in Section \[s0306\].
Lastly, the sinks of the invariant flow field are shown to form an IFS fractal in Section \[s030701\], implying the ultimate conclusion that the geometrical study of such sets, is of fundamental relevance to analyzing an FDT flow field.
Preliminary Concepts {#s02}
====================
Fully Developed Turbulence {#s0201}
--------------------------
Turbulence is a challenging phenomenon to model, primarily because it is difficult to define rigorously. One must isolate from experimental observation its definitive characteristics, before attempting to model it. However, even its main characteristics are difficult to identify, and one may never feel secure that some crucial characteristic may have been excluded from the assumptions of a model, or that some assumption has been made which is not in agreement with reality. Furthermore, one must carefully differentiate between observed and intuitive characteristics as described by other researchers, as we hereby will. One must also restrict the problem of modelling sufficiently, in order for a specific solution to exist, which is why we shall focus on Fully Developed Turbulence (FDT). Although other approximate models may exist, one only needs to consider the phenomenon itself as defined from experiments, when attempting to build a new model. Currently FDT has no known exact model, which is a major issue considering its widespread occurrence in Nature and technology.
A rich literature exists on the topic of FDT, and we recommend the overview of Ecke [@ba00032] for an introduction. Our overview assumes that the reader is familiar with elementary Fluid Dynamics.
### Definition {#s020101}
Turbulence has been the interest of classical physics for centuries. Its name comes from the Italian word “turbolenza” given by Leonardo da Vinci, one of the first to study and depict the phenomenon. In terms of the Reynolds number, flows above an $Re$ of $5000$ or higher are considered to be turbulent.
The phenomenon of FDT is an extremal case of turbulence, based on experimental observations at high Reynolds numbers, typically above $10,000$. The words “fully developed” in FDT refer to its main theoretical characteristics: an infinite Reynolds number and a steady unchanging yet chaotic velocity field, independent of viscosity. This final unchanging nature of the field is meant to characterize an equilibrium state or invariance with respect to its evolution equation, whatever it may be. Furthermore, the velocity field, its set of singularities, as well as its evolution is hypothesized to be self-similar or fractalline in some sense.
The precise ideal conditions for FDT to occur are currently unknown, and its exact causes, evolution, and characteristics are hypothetical. With some contemplation, one may realize that a mathematical theory is the only possible resolution. It is currently believed that the Navier-Stokes Equations can potentially describe turbulence at high Reynolds numbers.
### Relevance {#s020102}
The modelling of turbulence - whether approximate or exact - is of tremendous importance, considering its various and widespread applications.
The study of atmospheric and oceanic turbulence occurs in the context of Geophysical Fluid Dynamics and Climate Theory - primarily in the form of 2D turbulence. Quantum turbulence in quantum fluids, such as superfluids, occurs at close to zero viscosity. Magnetohydrodynamic (MHD) turbulence is relevant to astro- and geophysics, ranging from the study of planetary magnetic fields to large-scale industrial facilities, and plasma flow in stars and fusion reactors. In biology, plankton distribution is influenced by turbulence, and is important since plankton comprise the bottom of the oceanic food chain.
### Characteristics, Causes, and Evolution {#s020103}
Turbulence literature describes a set of observed and intuitive characteristics, and one must carefully distinguish between them. The causes and evolution towards Fully Developed Turbulence are also hypothetical, yet reasonable and experimentally inspired. Observed characteristics include high and erratic velocity fluctuation over a large range of coupled spatial and temporal scales. Consequently, a turbulent flow field exhibits structure at many length scales, often described as self-similar or scale-independent in some sense.
The onset and evolution of turbulence can be fuelled by various circumstances, such as the interaction of eddies in the flow, by injected kinetic energy into the flow such as via stirring, or even by the shape or surface of the container. Examples include flow from a faucet, flow past rocks or a curved wall, or the rotation of the Earth. A periodicity in the circumstances often plays a role in inducing intermittency. Mathematically speaking, development can occur due to periodic changes in the external forcing, or variations in the boundary conditions. In intermittency, transitions are observed to occur abruptly to successively more complex states. The evolution of turbulence is suspected to be purely inertial, meaning due to the actual motion of the fluid, even if fuelled externally. Flows where inertial effects are small, tend to remain laminar.
The induced eddies and their intermittent inertial interaction are considered to be the intrinsic elements of turbulent dynamics. Since the injected energy at the top scale is periodic, the eddies in their evolutionary hierarchy receive periodic bursts of energy as well, causing their interaction to be intermittent. In between each energy burst, a brief quiescent period occurs, where the fluid flows along its natural course i.e. its steady streamlines.
A hypothetical picture of the eddy hierarchy is the idea of vortex stretching and the Richardson energy cascade. The eddies on the free surface of the fluid, will have axes roughly perpendicular to the surface. A thinning of the eddies will occur due to the conservation of fluid elements. This results in the breakdown of larger structures into hierarchically smaller ones, causing a cascade of energy. This is a purely inertial and inviscid process, which continues until the local structures are small enough that viscosity causes their kinetic energy to dissipate as heat.
\[s02010301\] (Mandelbrot [@bc00008]) At infinite Reynolds numbers (FDT), the energy dissipation of the fluid concentrates in a set of non-integer Hausdorff dimension.
The conjecture has been resolved by V. Scheffer in 1980, though the proof is non-constructive. If one wishes to construct an FDT flow field, this conjecture may serve as an inspiration.
The Richardson cascade is primarily a 2D phenomenological hypothesis, taking place on the surface of the fluid. Turbulent fluids typically occupy 3D space, but motion in the vertical direction can become less relevant, and we are left with a quasi-2D hydrodynamics. The degree to which the third dimension is suppressed, primarily depends on the ratio of the vertical to the lateral scales, and the effect of external forces causing dominant lateral velocities over vertical ones.
### Existing Models {#s020104}
Previous efforts at modelling turbulence, and FDT in particular, have been approximate or statistical in nature, mostly in the form of scaling laws derived from the equations of motion. This direction was inspired by Kolmogorov’s influential model based on the Richardson cascade, culminating in the current Beta Model developed by Novikov, Stewart, Mandelbrot, Kraichan, Frisch, Sulem and Nelkin et al. For a detailed overview, we recommend Jou [@ba00033]. There is also the computational direction of simulating large Reynolds number flows on high-performance computers, originally suggested by John von Neumann.
The theory of Chaotic Advection is an experimental and mathematical direction, meant to investigate the intermittent interaction of eddies directly. Its language is Potential Flow theory and Dynamical Systems. We will dedicate an entire section to its review.
Detailed understanding and prediction from first principles still elude turbulence theory, which mostly bases itself on the Navier-Stokes Equations. Currently, it is considered to be highly unlikely that an exact rigorous mathematical model should be possible for either turbulence and its evolution, or its end result, Fully Developed Turbulence itself. If such a model exists, it is conjectured to be geometrical. Nevertheless, an exact model must only be in agreement with the observed phenomenon itself, and not with other derived models.
Potential Flows {#s0202}
---------------
For laminar flows outside the boundary layer, potential flows are considered to provide a sufficient model. Yet the turbulent boundary layer itself is often modelled approximately with a superposition of vortices (Vortex Dynamics). These flows are assumed to be steady (time-independent), ideal (zero viscosity / inviscid, uniform density, and incompressible), and irrotational (Lebesgue almost everywhere, denoted as a.e.).
By these conditions, conservation of mass for a velocity field $v=(v_1,v_2):{\ensuremath{\mathbb{R}}}^2 {\ensuremath{\shortrightarrow}}{\ensuremath{\mathbb{R}}}^2$ is $$0 = {\mathrm{div}}\ v = \partial_1 v_1 + \partial_2 v_2\ \ {\mathrm{a.e.}}$$ and the equations of motion become $$\frac12{\ensuremath{\nabla}}|v|^2 - v\times{\mathrm{curl}}\ v = -\frac{1}{{\ensuremath{\varrho}}}{\ensuremath{\nabla}}p$$ where ${\ensuremath{\varrho}}$ is the density and $p$ is pressure, further reducing to $$p(z) = \left(p(z_0)+\frac{{\ensuremath{\varrho}}}{2} |v(z_0)|^2\right) - \frac{{\ensuremath{\varrho}}}{2} |v(z)|^2$$ under the irrotationality requirement $$0 = {\mathrm{curl}}\ v = \partial_2 v_1 - \partial_1 v_2\ \ {\mathrm{a.e.}}$$
Supposing that $\psi,\phi\in C^2_{ae}({\ensuremath{\mathbb{R}}}^2,{\ensuremath{\mathbb{R}}})$ are harmonic conjugates, meaning they satisfy the Cauchy-Riemann Equations $${\ensuremath{\Delta}}\psi={\ensuremath{\Delta}}\phi=0\ {\mathrm{a.e.}}\ \ \ \mathrm{and}\ \ \ \partial_1\phi = \partial_2\psi,\ \partial_2\phi = -\partial_1\psi\ {\mathrm{a.e.}}$$ then $v:= -{\ensuremath{\nabla}}\phi = {\mathrm{i}}{\ensuremath{\nabla}}\psi$ satisfies mass conservation and irrotationality ${\mathrm{a.e.}}$ The pressure field can be calculated from $v$ as above. We call such ${\mathrm{a.e.}}$ harmonic $\psi$ for which a conjugate exists, the stream function of a potential flow, and $\phi$ the potential function. We note that the potential function corresponding to a stream function is only unique up to a gradient. Nevertheless we denote it as $\tilde{\psi}:=\phi$, and denote equivalence in the gradient as $$\phi_1\equiv\phi_2\ {\ensuremath{\Leftrightarrow}}\ {\ensuremath{\nabla}}\phi_1={\ensuremath{\nabla}}\phi_2\ {\mathrm{a.e.}}$$ A harmonic function is known to admit a conjugate if its domain is simply connected. Furthermore $\tilde{\tilde{\psi}}=-\psi$.
By their relation to the velocity field, the curves of constant $\psi$ represent the streamlines of the flow, and the curves of constant $\phi$ the equipotential lines. Having two pairs of harmonic conjugates $\psi_1, \phi_1$ and $\psi_2, \phi_2$ their linear combinations $a\psi_1 + b\psi_2$ and $a\phi_1 + b\phi_2$ for any $a,b\in{\ensuremath{\mathbb{R}}}$ are also harmonic conjugates. So the principle of superposition holds, as long as the boundary conditions are also correspondingly combined. Since the equations of motion have a unique solution for a set of boundary conditions, and they translate to Laplace’s equation (the solution of which for appropriate boundary conditions also exists and is unique up to a gradient), we may conclude that potential flows fully characterize the set of all flows which are steady, ideal, and irrotational ${\mathrm{a.e.}}$
Elementary potential flows, which are often superposed to create more complex ones, include sinks, sources and the (circular) vortex. The stream and potential functions of a source ($q>0$) or a sink ($q<0$) with strength $q\in{\ensuremath{\mathbb{R}}}$ are defined at $z\in{\ensuremath{\mathbb{C}}}$ as $$\psi(z) = -\frac{q}{2\pi}{\mathrm{Arg}}\ z\ {\mathrm{mod}}\ |q|\ \ \ {\mathrm{and}}\ \ \ \phi(z) = -\frac{q}{2\pi}\ln |z|$$ while for the orthogonal circular vortex $$\psi(z) = \frac{\Gamma}{2\pi}\ln |z|\ \ \ {\mathrm{and}}\ \ \ \phi(z) = -\frac{\Gamma}{2\pi}{\mathrm{Arg}}\ z\ {\mathrm{mod}}\ |\Gamma|$$ where $\Gamma$ is the circulation around any closed path containing the vortex (strength of the vortex).
Superposing a sink and a circular vortex with a general center $p\in{\ensuremath{\mathbb{C}}}$ results in a logarithmic vortex (eddy) as follows $$\psi(z) = -\frac{q}{2\pi}{\mathrm{Arg}}(z-p) + \frac{\Gamma}{2\pi}\ln |z-p|\ \ {\mathrm{mod}}\ |q|$$ $$\phi(z) = -\frac{q}{2\pi}\ln |z-p| - \frac{\Gamma}{2\pi}{\mathrm{Arg}}(z-p)\ \ {\mathrm{mod}}\ |\Gamma|$$ Since $\psi$ is differentiable in the ${\ensuremath{\mathbb{R}}}^2$ sense almost everywhere (except at $p$ and along the branch $\psi(z)=q$) we have that the corresponding velocity field (extended continuously to the entire plane) is the following $$v(z) = \frac{q+\Gamma{\mathrm{i}}}{2\pi}\frac{1}{|z-p|}\frac{z-p}{|z-p|}\ \ (z\in{\ensuremath{\mathbb{C}}})$$ Here we consider the gradient vectors to be on the complex plane and differentiation in the bivariate sense, meaning ${\ensuremath{\nabla}}\psi = \partial_1\psi +{\mathrm{i}}\partial_2\psi$. In general, we will sloppily identify $z\mapsto\psi(z),\ z\in{\ensuremath{\mathbb{C}}}$ with $(x,y)\mapsto\psi(x+y{\mathrm{i}}),\ (x,y)\in{\ensuremath{\mathbb{R}}}^2$ when it is more convenient to do so.
Chaotic Advection {#s0203}
-----------------
### Aref’s Blinking Vortex-Sink System {#s020301}
Advection is the idea of a fluid transporting light matter (typically tracer particles) on its free surface, so that the velocity of the particles is given by the velocity field of the flow. If $\psi$ is the stream function of the velocity field, then this corresponds to the Hamiltonian equations $v_1 = -\partial_2\psi,\ v_2 = \partial_1\psi$. Hence the configuration space of an advected particle is the phase space of this Hamiltonian system. Often such a system exhibits chaos, or sensitive dependence on initial conditions. The notion that laminar flows can produce chaotic particle trajectories, is considered today to be a cornerstone of Fluid Dynamics. Hassan Aref is the originator of this idea, and his influence has propagated far.
The flows mostly studied in this context, are piecewise steady flows. Meaning that on the evolution timeline of the flow, the velocity field is steady for disjoint consecutive time intervals, and a non-smooth abrupt transition (jump) to a new velocity field occurs at each partitioning moment in time. The jumps act along the streamlines of two circular (or logarithmic) vortices, which take alternating (non-random) turns in perturbing the flow field. This is called the Blinking Vortex(-Sink) System, and has been examined extensively by a number of theorists and experimentalists. It is meant to directly simulate the intermittent interaction of (circular or) logarithmic vortices, which is considered to be the elementary driving mechanism of the evolution towards Fully Developed Turbulence.
Aref in his highly influential paper [@ba00034] originally defined the setup with two circular vortices, which has been modified in [@ba00035] to two vortex-sinks (eddies), and was further examined by Károlyi and Tél [@ba00036] as well as a number of other enlightened researchers.
In particular, Wiggins [@ba00043] discusses the method and implications of modelling this blinking as the alternating (non-random) infinite iterative composition of two maps (called linked twist maps, the study of which goes back to Devaney [@bp00001]) in order to show the existence of a chaotic invariant set via the Conley-Moser conditions. These ideas are further examined with Ottino in [@ba00044], and bear an inspirational resemblance to our own, to be presented here. For an overview of Chaotic Advection (or Mixing), we recommend the book by Ottino [@bb00008].
### Ott’s Chaos Game {#s020302}
Ott et al. [@ba00038] has made a significant improvement on Aref’s model. The regular iterative alternation between two sinks has been modified to a random iteration over a finite family of mappings that transform the flow at the intermittent time partitions, in between steady flow periods. So at step $l\in{\ensuremath{\mathbb{N}}}$ (time partition $t_l$) the state of the flow $\sigma_l$ (usually the position of a tracer particle on the fluid surface) is transformed as $\sigma_{l+1} = T_l(\sigma_l)$ where $T_l\in\{T_1,\ldots,T_n\}$ is chosen at random, and where the precise form of $T_l$ can vary and is often left implicit. This random discrete evolution is reminiscent of the Chaos Game of Barnsley [@bb00002]. Though as we will see, this formulation can be closely aligned with experiments, it is difficult however to treat it mathematically. Ott et al. further assume that $t_l$ are spaced evenly and their stream functions are time-dependent, as typical in the theory of Chaotic Advection. However as we will see, both of these are unnecessary constraints, in our explicit and more natural formulation of this chaos game, in Section \[s0302\]. An overview of the work of Ott and Sommerer et al. in collaboration with Tél has been presented in [@ba00037], building upon [@ba00038; @ba00039; @ba00040].
### Sommerer’s Experiment {#s020303}
Chaotic Advection is a de facto experimental theory, and its predictions have been verified by both computational and real-world experiments. Sommerer’s apparatus [@ba00039; @ba00037] carried out at the JHU Applied Physics Laboratory, consists of a tank of water with a set of sources at the bottom, which transmit pulsing injections from an external pulsatile flow modulator that randomly regulates which source receives the next injection at equal time intervals, circulated by an external pump. Each pulse corresponds to an iterative step $\sigma_{l+1} = T_l(\sigma_l)$ where the fixed set of $T_l\in\{T_1,\ldots,T_n\}$ represents the actions of each source on the particles. These fluorescent particles on the fluid surface are then traced over their chaotic paths. In a sense, this apparatus is an inversion of the Aref setup, since it utilizes sources instead of sinks. Nevertheless, its relevance lies in attempting to simulate the intermittent evolution of turbulence, in the spirit of Ott’s chaos game recursion.
IFS Fractals {#s0204}
------------
The attractors of Iterated Function Systems - IFS Fractals - were pioneered by Hutchinson [@ba00007], further discussed by Barnsley and Demko [@ba00017], and may be the most elementary fractals possible. They are the attractors of a finite set of affine linear contraction mappings on the plane - the “function system” - which when combined and iterated to infinity, converges to an attracting limit set, the IFS fractal itself. Our formulation follows [@bu00006].
### Definition and Existence {#s020401}
\[s02040101\] Let a 2-dimensional affine contractive mapping (briefly contraction or contraction map) $T:{\ensuremath{\mathbb{C}}}\shortrightarrow{\ensuremath{\mathbb{C}}}$ be defined for all $z\in{\ensuremath{\mathbb{C}}}$ as $T(z):=p+\varphi(z-p)$ where $p\in{\ensuremath{\mathbb{C}}}$ is the fixed point of $T$, and ${\ensuremath{\varphi}}={\ensuremath{\lambda}}e^{{\ensuremath{\vartheta}}i}\in{\ensuremath{\mathbb{C}}}$ is the factor of $T$, with ${\ensuremath{\lambda}}\in(0,1)$ the contraction factor of $T$, and ${\ensuremath{\vartheta}}\in(-\pi,\pi]$ the rotation angle of $T$.
Note that an equivalent definition may be given using unitary rotation matrices $R\in{\ensuremath{\mathbb{R}}}^{2\times 2},\\ R^{T}\!R=I$ corresponding to $e^{{\ensuremath{\vartheta}}i}$. Then the contraction maps take the form $$T(z)=p+{\ensuremath{\lambda}}R(z-p)\ \ \ (z\in{\ensuremath{\mathbb{R}}}^2,\ p\in{\ensuremath{\mathbb{R}}}^2,\ {\ensuremath{\lambda}}\in (0,1))$$ This version shall be useful for certain proofs in later sections.
\[s02040102\] Let a 2-dimensional affine contractive $n$-map iterated function system (briefly IFS or $n$-map IFS, $n\in{\ensuremath{\mathbb{N}}}$) be defined as a finite set of contractions, and denoted as $\mathcal{T}:=\{T_1,\ldots,T_n\}$. We will denote the set of indices as ${\ensuremath{\mathcal{N}}}:=\{1,\ldots,n\}$, the set of fixed points as ${\ensuremath{\mathcal{P}}}:=\{p_1,\ldots,p_n\}$, and the set of factors as $\Phi:=\{{\ensuremath{\varphi}}_1,\ldots,{\ensuremath{\varphi}}_n\}$.
\[s02040103\] Let $\mathcal{T}=\{T_1,\ldots,T_n\}, n\in{\ensuremath{\mathbb{N}}}$ be an IFS. We define the Hutchinson operator $H$ belonging to $\mathcal{T}$ as $$H(S):=\bigcup_{k=1}^n T_k(S)\ \ \mathrm{where}\ \ T_k(S):=\{T_k(z): z\in S\}\ \ \mathrm{for\ any}\ \ S\subset{\ensuremath{\mathbb{C}}}$$ and call $H(S)$ the Hutchinson of the set $S$.
\[s02040104\] (Hutchinson [@ba00007]) For any IFS with Hutchinson operator H, there exists a unique compact set $F\subset{\ensuremath{\mathbb{C}}}$ such that $H(F)=F$. Furthermore, for any compact $S_0\subset{\ensuremath{\mathbb{C}}}$, the recursive iteration $S_{n+1}:=H(S_n)$ converges to $F$ in the Hausdorff metric.
[**Proof** ]{}The proof follows from the Banach Fixed Point Theorem, once we show that $H$ is contractive in the Hausdorff metric over compact sets, which form a complete space. [$\square$]{}
\[s02040105\] Let the set $F$ in the above theorem be called a fractal generated by an IFS with Hutchinson operator $H$ (briefly IFS fractal). Denote $\langle T_1,\ldots,T_n\rangle = \langle\mathcal{T}\rangle :=F$.
![Generation of a fractal by iterating a square (created by Scott Draves).[]{data-label="s02040106"}](./p00006b.jpg){width="285pt"}
### The Address Set {#s020402}
The address set arises from the iteration of the Hutchinson operator, and it is a way to label each fractal point. Since we can start the iteration towards $F$ with any compact set, we will choose the primary fixed point for simplicity, which is any point in ${\ensuremath{\mathcal{P}}}$ of our preference.
\[s02040201\] Let ${\ensuremath{\mathcal{N}}}^L:={\ensuremath{\mathcal{N}}}\times\ldots\times{\ensuremath{\mathcal{N}}}$ be the index set to the $L$-th Cartesian power, and call this $L\in{\ensuremath{\mathbb{N}}}$ the iteration level. Then define the address set as ${\ensuremath{\mathcal{A}}}:=\{0\}\cup\bigcup_{L=1}^\infty{\ensuremath{\mathcal{N}}}^L$. For any $a\in{\ensuremath{\mathcal{A}}}$ denote its $k$-th coordinate as $a(k),\ k\in{\ensuremath{\mathbb{N}}}$. Let its dimension or length be denoted as $|a|\in{\ensuremath{\mathbb{N}}}$ so that $a\in{\ensuremath{\mathcal{N}}}^{|a|}$ and let $|0|:=0$. Define the map with address $a\in{\ensuremath{\mathcal{A}}}$ acting on any $z\in{\ensuremath{\mathbb{C}}}$ as the function composition $T_a(z):=T_{a(1)}\circ\ldots\circ T_{a(|a|)}(z)$. Let the identity map be $T_0:=Id$. For a set of weights $\{w_1,\ldots, w_n\}\subset (0,1)$ define $w_a:= w_{a(1)}\cdot\ldots\cdot w_{a(|a|)}$, for a set of factors $\{{\ensuremath{\varphi}}_1,\ldots,{\ensuremath{\varphi}}_n\}$ define ${\ensuremath{\varphi}}_a:= {\ensuremath{\varphi}}_{a(1)}\cdot\ldots\cdot{\ensuremath{\varphi}}_{a(|a|)}$, and for a set of angles $\{{\ensuremath{\vartheta}}_1,\ldots,{\ensuremath{\vartheta}}_n\}$ define ${\ensuremath{\vartheta}}_a:={\ensuremath{\vartheta}}_{a(1)}+\ldots +{\ensuremath{\vartheta}}_{a(|a|)}$.
\[s02040202\] For any primary fixed point $p\in{\ensuremath{\mathcal{P}}}$ we have $$F=\langle T_1,\ldots,{\mathrm{T}}_n\rangle = \lim_{L\rightarrow\infty} H^L(\{p\}) = \mathrm{Cl}\{T_a(p):a\in{\ensuremath{\mathcal{A}}}\} = \mathrm{Cl}\{T_a(p_k):a\in{\ensuremath{\mathcal{A}}},\ p_k\in{\ensuremath{\mathcal{P}}}\}$$ We call this the address generation of $F$.
[**Proof** ]{}The proof follows from Theorem \[s02040104\] with either of the initial sets $\{p\}$ or ${\ensuremath{\mathcal{P}}}$. [$\square$]{}
\[s02040203\] Inspired by the above theorem, we will call $F_L:=\{T_a(p): a\in{\ensuremath{\mathcal{A}}}, |a|=L\}$ for some $L\in{\ensuremath{\mathbb{N}}}$ and primary fixed point $p\in{\ensuremath{\mathcal{P}}}$ the $L$-th level iterate towards $(T_1,\ldots,T_n):=\{T_a(p): a\in{\ensuremath{\mathcal{A}}}\}$ the countably infinite IFS fractal.
Optimal Transport over Measures {#s0205}
-------------------------------
Transportation Theory, often referred to as the theory of Optimal Transport, is the study of the optimal transportation and allocation of resources. Given two finite sets of producers and consumers equal in number, with each producer supplying a specific consumer according to a bivariate cost function for transportation, the goal is to find the minimal arrangement between producers and consumers, called the optimal transport plan or map.
The discrete problem can be generalized to probability measures - the Monge-Kantorovich formulation, named after its two most influential creators. Like many optimization problems, this minimization problem has a dual maximization equivalent. When the cost function over a compact metric space $(X, d_X)$ is the metric itself, then the dual becomes the first Wasserstein distance (also called Hutchinson metric) between the elements of the set of Borel probability measures $\nu_{1,2}\in{\ensuremath{\mathcal{M}_X}}$ over $X$, defined as $$d(\nu_1, \nu_2) = \sup \left\{\int_X f\ {\mathrm{d}}(\nu_1-\nu_2): f\in{\ensuremath{\mathcal{L}_1(X)}}\right\}$$ where ${\ensuremath{\mathcal{L}_1(X)}}$ are the Lipschitz-1 functions over $X$. That this is indeed the dual to the minimization problem, was shown by Kantorovich and Rubinstein [@ba00041]. Furthermore $({\ensuremath{\mathcal{M}_X}}, d)$ is known to be a complete metric space.
Generally the transport plan is a pushforward map (transfer operator) of the form $T^*\nu = \nu{\ensuremath{\circ}}{\mathrm{T}}^{-1}\ (\nu\in{\ensuremath{\mathcal{M}_X}})$. If $T^*$ is contractive in $d$, then by the Banach Fixed Point Theorem it has a unique invariant measure ${\ensuremath{\overline{\nu}}}= T^*{\ensuremath{\overline{\nu}}}$. Taking a finite weighted average of such pushforward operators, the resulting transfer operator $${\mathrm{T}}= w_1 T_1^* +\ldots +w_n T_n^*\ \ \ \mathrm{where}\ \ \ w_k\in (0,1),\ w_1+\ldots+w_n =1$$ will also be contractive in $d$ over ${\ensuremath{\mathcal{M}_X}}$, and possesses a fixed point. Thus the following theorems hold for an IFS $\{T_1,\ldots,{\mathrm{T}}_n\}$ in the form $T_k(z)=p_k+{\ensuremath{\lambda}}_k R_k(z-p_k),\ z\in{\ensuremath{\mathbb{R}}}^2$ supposing that $F\subset X\subset {\ensuremath{\mathbb{R}}}^2$.
\[s02050001\] (Hutchinson [@ba00007]) For any IFS $\{T_1,\ldots,T_n\}$ and weights $\{w_1,\ldots,w_n\}$, the transfer operator ${\mathrm{T}}= w_1 T_1^* +\ldots +w_n T_n^*$ is contractive on $({\ensuremath{\mathcal{M}_X}},d)$ and attains a unique invariant probability measure ${\ensuremath{\overline{\nu}}}= {\mathrm{T}}{\ensuremath{\overline{\nu}}}$ over the compact set $X\subset{\ensuremath{\mathbb{R}}}^2$, with support $\langle T_1,\ldots,T_n\rangle$. We call this the invariant measure with respect to the IFS $\{T_1,\ldots,T_n\}$. For any $\nu_0\in{\ensuremath{\mathcal{M}_X}}$ the recursion $\nu_{L+1}=w_1 T_1^*\nu_L +\ldots +w_n T_n^*\nu_L\in{\ensuremath{\mathcal{M}_X}},\ L\in{\ensuremath{\mathbb{N}}}$ converges to the invariant measure.
We remark that with the primary fixed point $p$ and the notations of Section \[s020402\], the measures $$\nu_0 = \delta_p,\ \ \nu_L = \sum_{|a|=L} w_a T_a^*\delta_p\in{\ensuremath{\mathcal{M}_X}},\ L\in{\ensuremath{\mathbb{N}}}$$ satisfy the above recursion, and so ${\ensuremath{\overline{\nu}}}(S)=\lim (\nu_L)$ as discussed in [@bu00006]. We also observe that the support of $\nu_L$ is concentrated on the $L$-th iterate $F_L:= H^L(\{p\})=\{T_a(p): a\in{\ensuremath{\mathcal{A}}}, |a|=L\}$ approaching the compact IFS fractal $F=\langle T_1,\ldots,T_n\rangle$.
We further remark that defining $${\ensuremath{\mathcal{D}_X}}:= \{{\ensuremath{\varrho}}\in C_{ae}({\ensuremath{\mathbb{R}}}^2,{\ensuremath{\mathbb{R}}}): {\ensuremath{\varrho}}\geq 0,\ {\mathrm{Supp}}\ {\ensuremath{\varrho}}\subset X,\ \int_X{\ensuremath{\varrho}}=1\}$$ $${\ensuremath{\mathcal{P}_X}}:= \left\{\nu\in{\ensuremath{\mathcal{M}_X}}:\ \exists{\ensuremath{\varrho}}\in{\ensuremath{\mathcal{D}_X}}\ \forall S\subset X\ \mathrm{Borel\ set}: \nu(S)=\int_S{\ensuremath{\varrho}}\ \right\}$$ and keeping the above metric $d$, then ${\mathrm{Conv}}\{\delta_p: p\in{\ensuremath{\mathbb{R}}}^2\}\subset{\ensuremath{\mathcal{P}_X}}\subset{\ensuremath{\mathcal{M}_X}}$ and ${\ensuremath{\overline{\nu}}}\in{\ensuremath{\mathcal{P}_X}}$, so it remains true that there exists a unique invariant measure in $({\ensuremath{\mathcal{P}_X}}, d)$, and it is the same as the earlier. Furthermore ${\mathrm{T}}:{\ensuremath{\mathcal{P}_X}}{\ensuremath{\shortrightarrow}}{\ensuremath{\mathcal{P}_X}}$ since ${\mathrm{Supp}}({\mathrm{T}}\nu)=H({\mathrm{Supp}}\ \nu)\subset X$ if $\nu\in{\ensuremath{\mathcal{P}_X}}$.
The study of transfer operators - also called Ruelle(-Perron-Frobenius-Markov) operators - is a rich field, and their largest eigenvalue is typically one, while their eigenfunctions are usually fractalline or self-similar in some sense. This has profound implications for classical mechanics, such as the increase of entropy or the irreversibility of time. An eigenvalue of one, corresponds to a state of equilibrium.
For a more complete discussion of Optimal Transport, we refer the reader to Villani [@bb00007] and McCann [@bc00009]. For an overview of IFS, we recommend articles [@bc00002; @ba00007; @ba00017].
Remarks {#s0206}
-------
We hereby wish to make some remarks on the previous sections, which will shape our approach to be presented in Section \[s03\].
First of all, it seems reasonable to raise the doubt whether Fully Developed Turbulence in full three dimensions even exists. Certainly, flows can approach above $10,000$ Reynolds numbers, but it is doubtable whether the other defining characteristics of FDT, including the equilibrium state of an unchanging velocity field at $Re = \infty$ that is independent of viscosity, can evolve in a fully three-dimensional flow field. Furthermore, we recall that the Richardson cascade that inspired the notion of self-similarity, is primarily a 2D picture on the free surface of the fluid. It seems that 3D space has one too many degrees of freedom, while a limiting 2D free surface forces the fluid to self-organize via self-similarity. Thus we restrict our attention to the 2D free surface in our effort to model FDT, in the spirit of Chaotic Advection.
Secondly, we point out the mathematically unreasonable effort of various other approaches, which attempt to model FDT via the Navier-Stokes Equations. The discussed energy bursts fuelling the erratic characteristic of intermittency, are clear signs of the non-smooth nature of the evolution towards FDT. The Navier-Stokes Equations however, require temporal differentiability of the velocity field, which does not hold by the observed phenomenon of intermittent evolution, summarized above. Therefore one has no choice mathematically, but to discard the Navier-Stokes Equations, in hope of a more fitting evolution equation, in line with the experimental picture of intermittent inertial evolution.
Thirdly, we point out that prior to the advent of Chaotic Advection, potential flows had been considered to be useless for the study of turbulence. Yet ironically, potential flows share with FDT the characteristics of independence with respect to viscosity and time, both being steady flow fields. In fact for increasing Reynolds numbers, the Navier-Stokes Equations approach the inviscid Euler Equations, implying that the flow would converge to the inviscid solutions of Potential Flow Theory. Thus the Reynolds spectrum appears to be an ouroboros, similarly to the Riemann circle.
Lastly, we point out that in the discussion of Chaotic Advection, we have not introduced the standard mathematics in full, generally employed in the theory. This is partly because we find the existing methods to be suboptimal for actualizing the potential in its concepts, and partly because the theory in its current form is often intuitive and in need of a rigorous mathematical language. So we feel that the theory needs a complete reformulation unifying the Aref and Ott directions in the spirit of Wiggins, which is one of our goals in this paper.
So to project our intentions, the model to be presented shall employ the methods of Optimal Transport Theory to reformulate in a new language the theory of Chaotic Advection, in the setting of Potential Flows, with the purpose of creating an exact geometrical model for the evolution towards the final steady equilibrium state of Fully Developed Turbulence.
Fractal Potential Flows {#s03}
=======================
We have introduced above the theoretical elements that we intend to weave together in this section and the next, in our effort to create an exact model for Fully Developed Turbulence and the evolution towards it. We will introduce Fractal Potential Flows as our exact model, which arise due to the intermittent transfer interaction of eddies. We define a weighted transfer operator over a certain class of stream functions, and show the existence of its unique fixed point - the invariant flow - characterized by a stream function whose rotated gradient field will serve as our model. The set of sink singularities of this flow field is shown to be an IFS fractal, explicitly resolving Mandelbrot’s Conjecture.
Eddy Invariance {#s0301}
---------------
In Section \[s0202\] we have defined the stream function of an eddy with sink strength $q = C\ln{\ensuremath{\lambda}}$ and vortex strength or circulation $\Gamma=C{\ensuremath{\vartheta}}$ as $$\psi(z) = -\frac{C\ln{\ensuremath{\lambda}}}{2\pi}{\mathrm{Arg}}(z-p) + \frac{C{\ensuremath{\vartheta}}}{2\pi}\ln |z-p|\ \ {\mathrm{mod}}\ -C\ln{\ensuremath{\lambda}}$$ where ${\ensuremath{\varphi}}\in{\ensuremath{\mathbb{C}}},\ {\ensuremath{\lambda}}=|{\ensuremath{\varphi}}|<1,\ {\ensuremath{\vartheta}}={\mathrm{Arg}}\ {\ensuremath{\varphi}}$ and $C>0$ is an arbitrary parameter. The associated velocity field (continuously extended to the entire plane) becomes $$v[\psi](z):=\frac{C{\mathrm{Log}}\ {\ensuremath{\varphi}}}{2\pi}\frac{1}{|z-p|}\frac{z-p}{|z-p|}$$ We observe that the eddy stream function is invariant under the contraction $T(z)=p+{\ensuremath{\varphi}}(z-p)$, meaning $T^*\psi =\psi$. This is expected geometrically, since the orbit $T^t(z_0)=p+{\ensuremath{\varphi}}^t(z_0-p),\ t\in{\ensuremath{\mathbb{R}}}$ traces out a logarithmic spiral centered at $p\in{\ensuremath{\mathbb{C}}}$, and the streamlines of constant $\psi$ are spirals of the same pitch. So in other words, $\psi$ is the fixed point of the pushforward transfer operator $T^*$. This raises the question whether $\psi$ is unique, and if not, then what space would guarantee its uniqueness. Defining $\psi_0(z):={\mathrm{Arg}}(z-p)\ {\mathrm{mod}}\ {\ensuremath{\vartheta}}$ we see that $T^*\psi_0 = \psi_0$ also holds, therefore $\psi$ cannot be unique in such a general setting. So our ultimate goal becomes finding the proper function space where uniqueness can be guaranteed, with a reasonable physical interpretation. We begin our search by deriving some fundamental properties of transfer, which will have profound implications.
\[s03010001\] For a similarity contraction of the form $T(z)=p+{\ensuremath{\varphi}}(z-p),\ |{\ensuremath{\varphi}}|<1$ the following properties hold (with differentiation in the ${\ensuremath{\mathbb{R}}}^2$ sense, and $v[\psi]$ being the continuous extension of ${\mathrm{i}}{\ensuremath{\nabla}}\psi$). $${\ensuremath{\Delta}}T^* = \frac{1}{{\ensuremath{\lambda}}^2}\ T^*\!{\ensuremath{\Delta}}\ \ \ \mathrm{and}\ \ \ \widetilde{T^*\psi}\equiv T^*\tilde{\psi}\ \ \ \mathrm{and}\ \ \ v[T^*\psi] = \frac{{\ensuremath{\varphi}}}{|{\ensuremath{\varphi}}|^2}\ T^{*}v[\psi]$$
[**Proof** ]{}We show the above in the ${\ensuremath{\mathbb{R}}}^2$ sense first, which translates to the complex sense, using $T(z)=p+{\ensuremath{\lambda}}R(z-p),\ p,z\in{\ensuremath{\mathbb{R}}}^2$ where $R$ is the rotation matrix corresponding to ${\mathrm{e}}^{{\ensuremath{\vartheta}}{\mathrm{i}}}$. We first show the third property, keeping in mind that $v[\psi](z)=i{\ensuremath{\nabla}}\psi(z)\ {\mathrm{a.e.}}$ Gradient is a column vector in ${\ensuremath{\mathbb{R}}}^2$, which corresponds to the complex vector $\partial_1+{\mathrm{i}}\partial_2$. Differentiation ${\mathrm{D}}$ results in a row vector however, so we must take a transpose, meaning ${\ensuremath{\nabla}}={\mathrm{D}}^T$. Applying the generalized chain rule, we get $${\ensuremath{\nabla}}(\psi{\ensuremath{\circ}}T^{-1}) = {\mathrm{D}}(\psi{\ensuremath{\circ}}T^{-1})^T = \left(({\mathrm{D}}\psi{\ensuremath{\circ}}T^{-1})\cdot\frac{1}{{\ensuremath{\lambda}}}R^{-1}\right)^T = \frac{1}{{\ensuremath{\lambda}}}R\cdot{\ensuremath{\nabla}}\psi{\ensuremath{\circ}}T^{-1}$$ Multiplying by a $90^\circ$ rotation (or ${\mathrm{i}}$), we get the third property. To get the first property, we observe that ${\ensuremath{\Delta}}= {\mathrm{tr}}\ D{\ensuremath{\nabla}}$, so by the chain rule, the properties of trace, and the above $${\ensuremath{\Delta}}T^*\psi = {\mathrm{tr}}\ {\mathrm{D}}{\ensuremath{\nabla}}T^*\psi = {\mathrm{tr}}\ {\mathrm{D}}\left(\frac{1}{{\ensuremath{\lambda}}}R\cdot T^*{\ensuremath{\nabla}}\psi\right) = {\mathrm{tr}}\left(\frac{1}{{\ensuremath{\lambda}}}R\cdot T^*{\mathrm{D}}{\ensuremath{\nabla}}\psi\cdot\frac{1}{{\ensuremath{\lambda}}}R^{-1}\right) = \frac{1}{{\ensuremath{\lambda}}^2}T^*{\ensuremath{\Delta}}\psi$$ The second property also follows from the third, observing that in the above complex sense $\psi$ and $\tilde{\psi}$ are conjugates iff ${\ensuremath{\nabla}}\psi={\mathrm{i}}{\ensuremath{\nabla}}\tilde{\psi}$ by Section \[s0202\]. So taking $T^*$ of both sides, multiplying by ${\mathrm{i}}$ and dividing by ${\ensuremath{\lambda}}$, we get that ${\ensuremath{\nabla}}(T^*\psi) = {\mathrm{i}}{\ensuremath{\nabla}}(T^*\tilde{\psi})$ implying that ${\ensuremath{\nabla}}(T^*\tilde{\psi})={\ensuremath{\nabla}}(\widetilde{T^*\psi})$ and thus $\widetilde{T^*\psi}\equiv T^*\tilde{\psi}$. [$\square$]{}
Among many things, the first property also implies that if $\psi$ is ${\mathrm{a.e.}}$ harmonic then so is $T^*\psi$. Together with the second property, this implies that the space of potential flows is closed under pushforward transfer by a similarity contraction.
We end our discussion of a single eddy by emphasizing the underlying idea that one application of $T^*$ to a flow field represented by the stream function, corresponds to one intermittent energy burst fuelling an eddy to transform the flow field along the logarithmic spiral orbits of $T$. If $T^*$ can be shown to be contractive over stream functions in some complete space, then the final equilibrium fixed point stream can also be shown.
The Transfer Operator {#s0302}
---------------------
Considering that potential flows form a vector space as discussed in Section \[s0202\] and by Theorem \[s03010001\], the weighted transfer operator $${\mathrm{T}}\psi = w_1 T_1^*\psi +\ldots + w_n T_n^*\psi$$ for any IFS $\{T_1,\ldots,{\mathrm{T}}_n\},\ T_k(z)=p_k+{\ensuremath{\varphi}}_k(z-p_k)$ and weights $w_k\in(0,1), \sum_k w_k=1$, maps the stream function of a potential flow to another such flow. One application of ${\mathrm{T}}$ to a flow field corresponds to the random application of some $T_k^*\in\{T_1^*,\ldots, T_n^*\}$ with probability ${\mathrm{P}}({\mathrm{T}}\psi(z_0) = T_k^*\psi(z_0))=w_k$ at some point $z_0$ in the plane. Indeed, the iteration of the transfer operator ${\mathrm{T}}$ corresponds to the intermittent randomly alternating contractive action of a system of $n$ eddies along the orbits of $T_k$. Here the flow field is piecewise steady in between the energy bursts fuelling some particular eddy at a certain intermittent moment. During the steady periods, the flow field is of the form ${\mathrm{T}}^L\psi_0,\ L\in{\ensuremath{\mathbb{N}}}$ for some initial flow $\psi_0$. The partitioning intermittent moments in time $t_L$ can be spaced in any way over the timeline. This constitutes our mathematical reformulation of Chaotic Advection.
We see that after the $L$-th iteration we have $$\psi_L = {\mathrm{T}}^L\psi_0 = \sum_{|a|=L} w_a T_a^*\psi_0$$ where $T_a$ and $w_a$ were defined in Definition \[s02040201\]. Probabilistically speaking, this implies that $\psi_L$ is the stochastic superposition of $n^L$ eddies with transfers $T_a^*$, transforming the flow field at the $L$-th intermittent moment $t_L$ with probabilities ${\mathrm{P}}(\psi_L(z_0) = T_a^*\psi_0(z_0))=w_a$. We also note that by Theorem \[s03010001\] the corresponding velocity transfer operator is $$v[{\mathrm{T}}\psi]= w_1\frac{{\ensuremath{\varphi}}_1}{|{\ensuremath{\varphi}}_1|^2}\ T_1^*v[\psi] +\ldots + w_n\frac{{\ensuremath{\varphi}}_n}{|{\ensuremath{\varphi}}_n|^2}\ T_n^*v[\psi]$$ There are two possible interpretations of the iteration of the above transfer operator - probabilistic and physical. Probabilistically it means that considering the next iterate $\psi_{L+1}$ to be a random variable, its expected value is ${\mathrm{T}}\psi_L$. A related financial interpretation is that the collective action of the eddies, is a portfolio. Physically, weighted transfer can be interpreted as weighting the effect each eddy can have on the fluid at an intermittent moment.
Succinctly, we observe that iteration corresponds to intermittency, and hereby theorize that the intermittent evolution of turbulence towards the equilibrium state of Fully Developed Turbulence, corresponds to the iteration of ${\mathrm{T}}$ ad infinitum to an invariant flow ${\mathrm{T}}{\ensuremath{\overline{\psi}}}= {\ensuremath{\overline{\psi}}}$. Therefore ${\ensuremath{\overline{\psi}}}$ will represent our model for the fully developed turbulent flow field, and showing its unique existence now becomes our primary goal, along with finding an appropriate function space where this is possible and physically most reasonable.
\[s03020001\] For the weighted transfer operator ${\mathrm{T}}$ the following properties hold. $${\ensuremath{\Delta}}{\mathrm{T}}= \sum_{k=1}^n \frac{w_k}{{\ensuremath{\lambda}}_k^2}\ T_k^*{\ensuremath{\Delta}}\ \ \ \mathrm{and}\ \ \ \widetilde{{\mathrm{T}}\psi}\equiv {\mathrm{T}}\tilde{\psi}\ \ \ \mathrm{and}\ \ \ v[{\mathrm{T}}\psi] = \sum_{k=1}^n w_k \frac{{\ensuremath{\varphi}}_k}{|{\ensuremath{\varphi}}_k|^2}\ T_k^* v[\psi]$$
[**Proof** ]{}The proof follows trivially from Theorem \[s03010001\]. Note that the first property implies the preservation of harmonicity ${\mathrm{a.e.}}$ or physicality upon transfer, as discussed in Section \[s0202\]. [$\square$]{}
Flow Character {#s0303}
--------------
In constructing the proper flow space, our inspiring objective is to ensure the unique existence of an eddy as the attractor of its generating pushforward transfer map. In resolving this question, we show a correspondence between eddies and the Dirac delta function, which hints at a possible general correspondence between stream functions and the density functions of probability measures, via Poisson’s Equation. Finding the ideal kind of boundary condition will prove crucial to our quest, and meanwhile the proper space of flows shall slowly reveal itself to us.
First of all, we observe that the velocity field of the eddy discussed in Section \[s0301\] is characterized by the complex parameter $c=\frac{C}{2\pi}{\mathrm{Log}}\ {\ensuremath{\varphi}}$ which can be extracted from the stream function $\psi$ by the operation $$v[\psi](z)\ \overline{z-p} = c\ \frac{z-p}{|z-p|^2}\ \overline{z-p} =c$$ We will denote the stream function of an eddy with character $c\in{\ensuremath{\mathbb{C}}},\ {\mathrm{Re}}\ c<0$ and centered at $p\in{\ensuremath{\mathbb{C}}}$, with ${\ensuremath{q_c}}:=-2\pi{\mathrm{Re}}\ c,\ {\ensuremath{\Gamma_c}}:=2\pi{\mathrm{Im}}\ c$ as $$\psi_{c,p}(z) := \frac{{\ensuremath{q_c}}}{2\pi}{\mathrm{Arg}}(z-p) + \frac{{\ensuremath{\Gamma_c}}}{2\pi}\ln |z-p|\ \ {\mathrm{mod}}\ {\ensuremath{q_c}}\ \ \ \mathrm{and}\ \ \ \psi_c:=\psi_{c,0}$$ Certain flows may have a similar “eddy character” when their flow field is zoomed out, even if they exhibit varying streamline behaviour locally around the origin, as defined below.
\[s03030001\] We say that a stream function $\psi:{\ensuremath{\mathbb{C}}}{\ensuremath{\shortrightarrow}}{\ensuremath{\mathbb{R}}}$ satisfies the boundary condition at infinity with character $c\in{\ensuremath{\mathbb{C}}}$, if for any $p\in{\ensuremath{\mathbb{C}}}$ and any sequence $(z_j)\subset{\ensuremath{\mathbb{C}}}, |z_j-p|{\ensuremath{\shortrightarrow}}\infty$ we have $\exists\lim_{j{\ensuremath{\shortrightarrow}}\infty} v[\psi](z_j)\ \overline{z_j-p} = c$ where the limit is taken in the complex sense. We denote this property as $\psi\in{\ensuremath{BC_c^\infty}}$ or ${\mathrm{char}}(\psi)=c$.
We remark that character is independent of the choice of $p\in{\ensuremath{\mathbb{C}}}$. To see this, take $$\lim_{j{\ensuremath{\shortrightarrow}}\infty} |v[\psi](z_j)| = \lim_{j{\ensuremath{\shortrightarrow}}\infty} \left|\frac{v[\psi](z_j)\ \overline{z_j-p}}{\overline{z_j-p}}\right| = \lim_{j{\ensuremath{\shortrightarrow}}\infty} \frac{|c|}{|z_j-p|} = 0$$ so $\lim (v[\psi](z_j))=0$. Taking some other $p'\in{\ensuremath{\mathbb{C}}}$ we have for the above $(z_j)$ that $|z_j-p'|{\ensuremath{\shortrightarrow}}\infty$ since $|z_j-p|\leq |z_j-p'|+|p'-p|$ and that $$v[\psi](z_j)\ \overline{z_j-p'} = v[\psi](z_j)\ \overline{z_j-p} + v[\psi](z_j)\ \overline{p-p'}\ {\ensuremath{\shortrightarrow}}\ c+0= c\ \ \mathrm{as}\ \ j{\ensuremath{\shortrightarrow}}\infty$$
\[s03030002\] Character is a linear map that is invariant under weighted transfer, or convolution with a density function over the plane.
[**Proof** ]{}The linearity of ${\mathrm{char}}$ is trivial by definition. Let us suppose that ${\mathrm{char}}(\psi)=c$. Since for some $p$ and $(z_j),\ |z_j-p|{\ensuremath{\shortrightarrow}}\infty$ sequence $z_j-p = {\ensuremath{\varphi}}_k(T_k^{-1}(z_j)-p_k)+(p_k-p)$ and applying Theorem \[s03010001\] we have that $$v[{\mathrm{T}}\psi](z_j)\ \overline{z_j-p} = \sum_{k=1}^n w_k\frac{{\ensuremath{\varphi}}_k}{|{\ensuremath{\varphi}}_k|^2}\ v[\psi](T_k^{-1}(z_j))\ \overline{{\ensuremath{\varphi}}_k(T_k^{-1}(z_j)-p_k)+(p_k-p)} =$$ $$= \sum_{k=1}^n w_k\frac{{\ensuremath{\varphi}}_k\bar{{\ensuremath{\varphi}}}_k}{|{\ensuremath{\varphi}}_k|^2}\ v[\psi](T_k^{-1}(z_j))\ \overline{T_k^{-1}(z_j)-p_k} + \sum_{k=1}^n w_k\frac{{\ensuremath{\varphi}}_k}{|{\ensuremath{\varphi}}_k|^2}\ v[\psi](T_k^{-1}(z_j))\ \overline{p_k-p}$$ which approaches $\sum_k w_k c + \sum_k 0 = c$ as $j{\ensuremath{\shortrightarrow}}\infty$ meaning ${\mathrm{char}}({\mathrm{T}}\psi)=c$.\
Now let ${\ensuremath{\varrho}}$ be a density function over ${\ensuremath{\mathbb{R}}}^2$ and $(z_j)$ be the above sequence. We employ scalar and complex vector Riemann integrals as follows.
$${\ensuremath{\nabla}}(\psi\ast{\ensuremath{\varrho}})(z) = \int_{{\ensuremath{\mathbb{R}}}^2} {\ensuremath{\nabla}}\psi(z-w){\ensuremath{\varrho}}(w)\ {\mathrm{d}}w$$ $$v[\psi\ast{\ensuremath{\varrho}}](z_j)\ \overline{z_j-p} = \int_{{\ensuremath{\mathbb{R}}}^2} {\mathrm{i}}{\ensuremath{\nabla}}\psi(z_j-w)\ \overline{(z_j-w)-p}\ {\ensuremath{\varrho}}(w)\ {\mathrm{d}}w\ +$$ $$+ \int_{{\ensuremath{\mathbb{R}}}^2} {\mathrm{i}}{\ensuremath{\nabla}}\psi(z_j-w)\ \bar{w}\ {\ensuremath{\varrho}}(w)\ {\mathrm{d}}w\ \ {\ensuremath{\shortrightarrow}}\ \ \int_{{\ensuremath{\mathbb{R}}}^2} c{\ensuremath{\varrho}}(w)\ {\mathrm{d}}w + 0 = c\ \ \mathrm{as}\ \ j{\ensuremath{\shortrightarrow}}\infty$$ [$\square$]{}
In what follows, we will denote with $L^\infty({\ensuremath{\mathbb{R}}}^2)$ the set of bounded functions over the plane, and with $C_{ae}^2({\ensuremath{\mathbb{R}}}^2)$ the ${\mathrm{a.e.}}$ defined and twice ${\mathrm{a.e.}}$ continuously differentiable functions $\psi$, for which ${\mathrm{i}}{\ensuremath{\nabla}}\psi$ can be continuously extended to the entire plane, denoted as $v[\psi]$, furthermore the set of singularities $S_{\psi}$ of $v[\psi]$ is compact and nonempty.
\[s03030003\] For any density function ${\ensuremath{\varrho}}$ over the plane and any $c\in{\ensuremath{\mathbb{C}}},\ {\mathrm{Im}}\ c\neq 0$, there exists a $\psi\in L^\infty({\ensuremath{\mathbb{R}}}^2)\cap C_{ae}^2({\ensuremath{\mathbb{R}}}^2)\cap BC_c^\infty$ for which ${\ensuremath{\Delta}}\psi = {\ensuremath{\Gamma_c}}{\ensuremath{\varrho}}\ {\mathrm{a.e.}}$ Furthermore, this $\psi$ is unique up to a gradient (${\mathrm{a.e.}}$) and $\psi = \psi_c\ast{\ensuremath{\varrho}}$.
[**Proof** ]{}The existence of such a $\psi$ is guaranteed using the Green function method of convolving the fundamental solution $\psi_c$ (${\ensuremath{\Delta}}\psi_c = {\ensuremath{\Gamma_c}}\delta_0\ {\mathrm{a.e.}}$) with ${\ensuremath{\varrho}}$ which will also have a character of $c$ by the previous theorem. So we turn to the question of uniqueness (up to a gradient). Let us suppose indirectly that $\exists\psi_{1,2}\in{\ensuremath{BC_c^\infty}}, \psi_1\not\equiv\psi_2$ such that ${\ensuremath{\Delta}}\psi_1 = {\ensuremath{\Gamma_c}}{\ensuremath{\varrho}}= {\ensuremath{\Delta}}\psi_2\ {\mathrm{a.e.}}$ Then ${\ensuremath{\Delta}}\psi = 0\ {\mathrm{a.e.}}$ with $\psi:=\psi_1-\psi_2$. By the Divergence Theorem, for some compact $S\subset{\ensuremath{\mathbb{R}}}^2$ with piecewise smooth boundary $$\int_S |{\ensuremath{\nabla}}\psi|^2 = \int_S |{\ensuremath{\nabla}}\psi|^2+\psi{\ensuremath{\Delta}}\psi = \int_S {\mathrm{div}}(\psi{\ensuremath{\nabla}}\psi) = \int_{\partial S} \psi\langle{\ensuremath{\nabla}}\psi, u\rangle$$ Since $\psi_{1,2}$ are bounded, then so is $\psi$ with say $|\psi|\leq M$. So for any sequence of disks $S_j:=B(p,r_j),\ 0<r_j<r_{j+1},\ j\in{\ensuremath{\mathbb{N}}},\ \lim (r_j)=\infty$ for which $S_{\psi}\subset S_1,\ \partial S_{\psi}\cap\partial S_1={\ensuremath{\varnothing}}$ $$\int_{S_j} |{\ensuremath{\nabla}}\psi|^2 = \left|\int_{\partial S_j} \psi \langle{\ensuremath{\nabla}}\psi, u\rangle\right|\leq M\!\int_{\partial S_j} \left|\left\langle{\ensuremath{\nabla}}\psi(z), \frac{z-p}{|z-p|}\right\rangle\right|{\mathrm{d}}z \leq M\!\int_{\partial S_j} |v[\psi](z)|\ {\mathrm{d}}z$$ where the integrals are in the usual ${\ensuremath{\mathbb{R}}}^2$ sense. If we can show that this last integral vanishes as $j{\ensuremath{\shortrightarrow}}\infty$, then $\int_{{\ensuremath{\mathbb{R}}}^2} |{\ensuremath{\nabla}}\psi|^2 =0$ and so ${\ensuremath{\nabla}}\psi = 0\ {\mathrm{a.e.}}$ which would contradict $\psi_1\not\equiv\psi_2$.\
Let us define $z_j:={\mathrm{argmax}}_{z\in\partial S_j} |v[\psi](z)|,\ j\in{\ensuremath{\mathbb{N}}}$ which exist by the Extreme Value Theorem, considering the compactness of $\partial S_j$ and the continuity of $v[\psi]$ on $\partial S_j\subset{\ensuremath{\mathbb{R}}}^2\setminus S_{\psi}$. Since $z_j\in\partial S_j,\ j\in{\ensuremath{\mathbb{N}}}$ we have $|z_j-p|=r_j{\ensuremath{\shortrightarrow}}\infty,\ j{\ensuremath{\shortrightarrow}}\infty$ which by ${\mathrm{char}}(\psi_{1,2})=c$ implies that $v[\psi](z_j)\ \overline{z_j-p}{\ensuremath{\shortrightarrow}}0$ so $$\int_{\partial S_j} |v[\psi](z)|\ {\mathrm{d}}z \leq \int_{\partial S_j} |v[\psi](z_j)|\ {\mathrm{d}}z = |v[\psi](z_j)\ \overline{z_j-p}|\ \frac{1}{|z_j-p|}\ 2\pi r_j\ {\ensuremath{\shortrightarrow}}\ 0\ \ \mathrm{as}\ \ j{\ensuremath{\shortrightarrow}}\infty\ \ {\ensuremath{\square}}$$
By the above results, we may conclude among many things that the eddy $\psi_{c,p}$ uniquely corresponds via Poisson’s Equation to the Dirac delta function $\delta_p$ among all bounded functions of character $c$, where the uniqueness is guaranteed up to a gradient.
The Flow Space and the Measure Map {#s0304}
----------------------------------
\[s03040001\] Let $X\subset{\ensuremath{\mathbb{R}}}^2$ be a compact set such that $F=\langle T_1,\ldots,T_n\rangle\subset X$ and $c\in{\ensuremath{\mathbb{C}}}$ some number for which ${\mathrm{Re}}\ c <0,\ {\mathrm{Im}}\ c\neq 0$. For a function $\psi\in C_{ae}^2({\ensuremath{\mathbb{R}}}^2)$ define $$\mu[\psi](S)=\mu_c[\psi](S):=\frac{1}{{\ensuremath{\Gamma_c}}}\int_S{\ensuremath{\Delta}}\psi\ \ \ (S\subset{\ensuremath{\mathbb{R}}}^2)$$ Let the set of flows with character $c$ be defined as $$\Psi_{X,c}:= \{\psi\in C_{ae}^2({\ensuremath{\mathbb{R}}}^2)\cap{\ensuremath{BC_c^\infty}}:\ {\mathrm{Ran}}\ \psi\subset [0,{\ensuremath{q_c}}),\ \mu[\psi]\in{\ensuremath{\mathcal{P}_X}}\}$$ We will call $\mu:\Psi_{X,c}{\ensuremath{\shortrightarrow}}{\ensuremath{\mathcal{P}_X}}$ the measure map, and $\mu[\psi]$ the measure induced by $\psi$.\
Recalling the metric $d$ from Section \[s0205\], for any $\psi_{1,2}\in{\ensuremath{\Psi_{X,c}}}$ denote $$D(\psi_1,\psi_2)=D_{X,c}(\psi_1,\psi_2):=d(\mu[\psi_1],\mu[\psi_2])=\sup\left\{\frac{1}{{\ensuremath{\Gamma_c}}}\int_X f\ {\ensuremath{\Delta}}(\psi_1-\psi_2): f\in{\ensuremath{\mathcal{L}_1(X)}}\right\}$$ As in Section \[s0202\], we will consider two functions to be congruent $\psi_1\equiv\psi_2$ iff ${\ensuremath{\nabla}}\psi_1={\ensuremath{\nabla}}\psi_2\ {\mathrm{a.e.}}$ over $X$. Let the equivalence class of $\psi$ be denoted as $[\psi]:=\{\psi_0\in\Psi_{X,c}: \psi\equiv\psi_0\}$ and let $$\left[\Psi_{X,c}\right]:=\{[\psi]: \psi\in\Psi_{X,c}\} = \Psi_{X,c}/{\mathrm{Ker}}(\mu_c)$$ Further denote $D([\psi_1], [\psi_2])=D_{X,c}([\psi_1], [\psi_2]):= D_{X,c}(\psi_1,\psi_2)$.\
Then we call $([\Psi_{X,c}], D)$ the flow space over $X$ with character $c$.
\[s03040002\] $([\Psi_{X,c}], D)$ is a well-defined metric space. The measure map $\mu:[\Psi_{X,c}]{\ensuremath{\shortrightarrow}}{\ensuremath{\mathcal{P}_X}}$ is an isometric isomorphism, and the weighted transfer maps ${\mathrm{T}}:\Psi_{X,c}{\ensuremath{\shortrightarrow}}\Psi_{X,c}$. Furthermore, the two maps commute.
[**Proof** ]{}First of all, we observe that $\mu:{\ensuremath{[\Psi_{X,c}]}}{\ensuremath{\shortrightarrow}}{\ensuremath{\mathcal{P}_X}}$ is clearly an isometry, since by definition $d(\mu[\psi_1],\mu[\psi_2])=D([\psi_1],[\psi_2])$. $\mu$ is bijective iff $\forall\nu\in{\ensuremath{\mathcal{P}_X}}\ \exists! [\psi]\in{\ensuremath{[\Psi_{X,c}]}}:\ \mu[\psi]=\nu$. The equality $\mu[\psi]=\nu$ here means that with the density function ${\ensuremath{\varrho}}$ belonging to $\nu\in{\ensuremath{\mathcal{P}_X}}$ $$\mu[\psi](S)=\frac{1}{{\ensuremath{\Gamma_c}}}\int_S {\ensuremath{\Delta}}\psi = \nu(S) = \int_S{\ensuremath{\varrho}}\ \ \ \forall S\subset X$$ which is equivalent to ${\ensuremath{\Delta}}\psi = {\ensuremath{\Gamma_c}}{\ensuremath{\varrho}}\ {\mathrm{a.e.}}$ on $X$. The unique existence of such a $\psi\in{\ensuremath{\Psi_{X,c}}}$ for any ${\ensuremath{\varrho}}$ was shown in Theorem \[s03030003\] up to a gradient, meaning up to a congruence or equivalence class. So we have that $\mu$ is a bijective isometry, also called an isometric isomorphism.\
For the well-definition of the space, we must show that $D$ is a metric over the set [$[\Psi_{X,c}]$]{}. $D$ is positive, symmetric, and inherits the triangle inequality from $d$. The only question that remains is whether $D([\psi_1],[\psi_2])=0$ implies $[\psi_1]=[\psi_2]$. Since $0=D([\psi_1],[\psi_2])=d(\mu[\psi_1],\mu[\psi_2])$ we have that $\mu[\psi_1]=\mu[\psi_2]$, and since $\mu$ has been shown to be bijective, this implies that $[\psi_1]=[\psi_2]$ or $\psi_1\equiv\psi_2$.\
We go on to showing that ${\mathrm{T}}$ and $\mu$ commute, meaning ${\mathrm{T}}\mu = \mu{\mathrm{T}}$. We recall from Theorem \[s03020001\] that $${\ensuremath{\Delta}}({\mathrm{T}}\psi)(z) = \sum_{k=1}^n \frac{w_k}{{\ensuremath{\lambda}}_k^2}\ ({\ensuremath{\Delta}}\psi){\ensuremath{\circ}}T_k^{-1}(z)$$ and also see that for any $S\subset X\subset {\ensuremath{\mathbb{R}}}^2$ and $k\in\{1,\ldots,n\}$, by taking Riemann integrals over ${\ensuremath{\mathbb{R}}}^2$ we have that $$\mu[T_k^*\psi](S) = \frac{1}{{\ensuremath{\Gamma_c}}}\int_S{\ensuremath{\Delta}}(\psi{\ensuremath{\circ}}T_k^{-1}) = \frac{1}{{\ensuremath{\Gamma_c}}}\int_S \frac{1}{{\ensuremath{\lambda}}_k^2}({\ensuremath{\Delta}}\psi)(T_k^{-1}(z))\ {\mathrm{d}}z =$$ $$= \frac{1}{{\ensuremath{\Gamma_c}}}\int_{T_k^{-1}(S)}{\ensuremath{\Delta}}\psi(w)\ {\mathrm{d}}w = \mu[\psi](T_k^{-1}(S)) = (T_k^*\mu[\psi])(S)$$ so taking a convex combination and using the linearity of $\mu$ we have the desired property $\mu[{\mathrm{T}}\psi]={\mathrm{T}}\mu[\psi],\ \psi\in{\ensuremath{\Psi_{X,c}}}$ or succinctly $\mu{\mathrm{T}}= {\mathrm{T}}\mu$.\
Lastly, we show that ${\mathrm{T}}$ maps from ${\ensuremath{\Psi_{X,c}}}$ to itself. ${\mathrm{T}}$ preserves $C_{ae}^2({\ensuremath{\mathbb{R}}}^2)$ since $S_{{\mathrm{T}}\psi}=H(S_{\psi})$, and it also preserves character according to Theorem \[s03030002\]. Since ${\mathrm{T}}$ is a convex combination, it preserves that ${\mathrm{Ran}}({\mathrm{T}}\psi)\subset[0,{\ensuremath{q_c}})$. Furthermore, we have that $\mu[{\mathrm{T}}\psi]={\mathrm{T}}\mu[\psi]\in{\ensuremath{\mathcal{P}_X}}$ since $\mu[\psi]\in{\ensuremath{\mathcal{P}_X}}$ and the measure transfer is known to map ${\mathrm{T}}:{\ensuremath{\mathcal{P}_X}}{\ensuremath{\shortrightarrow}}{\ensuremath{\mathcal{P}_X}}$ by Section \[s0205\]. [$\square$]{}
Physically speaking, the above space $([\Psi_{X,c}], D)$ represents the fact that the intermittent interaction of a system of $n$ eddies, can be considered in a locally isolated manner over a compact $X\subset{\ensuremath{\mathbb{R}}}^2$ of reasonable scale. ${\ensuremath{[\Psi_{X,c}]}}$ versus ${\ensuremath{\Psi_{X,c}}}$ signifies that it is only the velocity field induced by a stream function that matters. Prior to the start of their interaction, the eddies $\psi_{c_k,p_k}$ (notation of Section \[s0303\]) superpose initially in a flow with a certain character $c$ as follows. Letting $\psi_0:=\sum_k w_k T_k^*\psi_{c_k,p_k}$ then by Theorem \[s03010001\] $$v[\psi_0]= \sum_{k=1}^n w_k\frac{{\ensuremath{\varphi}}_k}{|{\ensuremath{\varphi}}_k|^2}\ T_k^*v[\psi_{c_k,p_k}] = \sum_{k=1}^n w_k\ v[\psi_{c_k,p_k}]$$ so clearly $c={\mathrm{char}}(\psi_0)=\sum_k w_k c_k$. As the intermittent interaction of the eddies progresses according to the transfer operator ${\mathrm{T}}$, their overall character will be preserved to be this $c$ by Theorem \[s03030002\].
The Invariant Flow {#s0305}
------------------
\[s03050001\] The transfer operator ${\mathrm{T}}$ is contractive over $({\ensuremath{[\Psi_{X,c}]}}, D)$. For any $\psi_0\in{\ensuremath{\Psi_{X,c}}}$ the iteration $\psi_{L+1}={\mathrm{T}}\psi_L\in{\ensuremath{\Psi_{X,c}}}$ converges to $${\ensuremath{\overline{\psi}}}= \psi_c\ast{\ensuremath{\overline{\nu}}}= \lim_{L{\ensuremath{\shortrightarrow}}\infty} \sum_{|a|=L} w_a\psi_c(\cdot-T_a(p))\in{\ensuremath{\Psi_{X,c}}}$$ for which ${\mathrm{T}}{\ensuremath{\overline{\psi}}}\equiv{\ensuremath{\overline{\psi}}}$, and here ${\ensuremath{\overline{\nu}}}$ is the invariant measure of Theorem \[s02050001\] and $p$ is the primary fixed point. We will call this ${\ensuremath{\overline{\psi}}}$ the invariant flow (or fractal potential flow) with respect to the IFS $\{T_1,\ldots,T_n\}$ and character $c\in{\ensuremath{\mathbb{C}}}$.
[**Proof** ]{}The contractivity of transfer over stream functions, follows directly from that of the measure transfer in Theorem \[s02050001\], considering that the measure map is isometric by definition.\
At this point one would expect to show the completeness of $({\ensuremath{[\Psi_{X,c}]}},D)$ in order to conclude the unique existence of a fixed flow of ${\mathrm{T}}$ by the Banach Fixed Point Theorem. This however becomes unnecessary, considering that we have an isometric isomorphism $\mu:[\Psi_{X,c}]{\ensuremath{\shortrightarrow}}{\ensuremath{\mathcal{P}_X}}$ between flows and measures, which commutes with ${\mathrm{T}}$ by Theorem \[s03040002\].\
By Theorem \[s02050001\] and the remarks we have made on it, there exists a unique ${\ensuremath{\overline{\nu}}}\in{\ensuremath{\mathcal{P}_X}}:{\mathrm{T}}{\ensuremath{\overline{\nu}}}={\ensuremath{\overline{\nu}}}$. By Theorem \[s03040002\] there exists a unique $[{\ensuremath{\overline{\psi}}}]\in{\ensuremath{[\Psi_{X,c}]}}:\mu[{\ensuremath{\overline{\psi}}}]={\ensuremath{\overline{\nu}}}$ and $$\mu[{\mathrm{T}}{\ensuremath{\overline{\psi}}}] = {\mathrm{T}}\mu[{\ensuremath{\overline{\psi}}}] = {\mathrm{T}}{\ensuremath{\overline{\nu}}}= {\ensuremath{\overline{\nu}}}= \mu[{\ensuremath{\overline{\psi}}}]\ {\ensuremath{\Rightarrow}}\ {\mathrm{T}}{\ensuremath{\overline{\psi}}}\equiv{\ensuremath{\overline{\psi}}}$$ since $\mu$ is bijective. Now let us suppose indirectly that $[{\ensuremath{\overline{\psi}}}]$ is not a unique fixed point, meaning there is another $\psi\not\equiv{\ensuremath{\overline{\psi}}}$ for which ${\mathrm{T}}\psi\equiv\psi$. Then $$\mu[\psi] = \mu[{\mathrm{T}}\psi] = {\mathrm{T}}\mu[\psi]\ {\ensuremath{\Rightarrow}}\ \mu[\psi]={\ensuremath{\overline{\nu}}}=\mu[{\ensuremath{\overline{\psi}}}]\ {\ensuremath{\Rightarrow}}\ \psi\equiv{\ensuremath{\overline{\psi}}}$$ which is a contradiction.\
Now let $\psi_0\in{\ensuremath{\Psi_{X,c}}}$ be any flow, and consider the recursion $\psi_{L+1}={\mathrm{T}}\psi_L\in{\ensuremath{\Psi_{X,c}}}$ and define $\nu_L:=\mu[\psi_L]$. Then $\nu_{L+1}=\mu[{\mathrm{T}}\psi_L]={\mathrm{T}}\nu_L$ so by Theorem \[s02050001\] we have that $\exists\lim (\nu_L)={\ensuremath{\overline{\nu}}}$ in $({\ensuremath{\mathcal{P}_X}},d)$. Since $D([\ \cdot\ ],[\ \cdot\ ])=d(\mu[\ \cdot\ ],\mu[\ \cdot\ ])$ and $\mu$ is injective, we have that $\exists\lim ([\psi_L])=[{\ensuremath{\overline{\psi}}}]$.\
We go on to characterizing the exact form of ${\ensuremath{\overline{\psi}}}$. From Section \[s0205\] we recall that $${\ensuremath{\overline{\nu}}}= \lim_{L{\ensuremath{\shortrightarrow}}\infty} \sum_{|a|=L} w_a T_a^*\delta_p$$ From the proof of Theorem \[s03030003\] we see that $\psi_c\in{\ensuremath{\Psi_{X,c}}}: {\ensuremath{\Delta}}\psi_c= {\ensuremath{\Gamma_c}}\delta_0$ so $${\ensuremath{\overline{\psi}}}= \psi_c\ast{\ensuremath{\overline{\nu}}}= \lim_{L{\ensuremath{\shortrightarrow}}\infty} \sum_{|a|=L} w_a\psi_c\ast(\delta_p{\ensuremath{\circ}}T_a^{-1}) = \lim_{L{\ensuremath{\shortrightarrow}}\infty} \sum_{|a|=L} w_a\psi_c(\cdot-T_a(p))\ \ \ {\ensuremath{\square}}$$ We remark that ${\ensuremath{\Delta}}{\ensuremath{\overline{\psi}}}=0\ {\mathrm{a.e.}}$ and if ${\ensuremath{\Delta}}\psi_0=0\ {\mathrm{a.e.}}$ then ${\ensuremath{\Delta}}\psi_L=0\ {\mathrm{a.e.}}\ (L\in{\ensuremath{\mathbb{N}}})$.\
Furthermore, taking $\psi_0:=\psi_{c,p}$ $${\mathrm{T}}^L\psi_{c,p} = \sum_{|a|=L} w_a T_a^*\psi_{c,p} \equiv \sum_{|a|=L} w_a\psi_c(\cdot-T_a(p))\ \ ({\mathrm{mod}}\ {\ensuremath{q_c}})$$ The first equality is clear, while the second equivalence requires some consideration. It can be shown by induction that $z-T_a(p)={\ensuremath{\varphi}}_a(T_a^{-1}(z)-p)$. So since $\psi_{c,p}=\psi_c(\cdot-p)$ we have $$\psi_{c,p}(T_a^{-1}(z)) = \psi_c(T_a^{-1}(z)-p) \equiv \psi_c(z-T_a(p))-\psi_c({\ensuremath{\varphi}}_a)\ \ ({\mathrm{mod}}\ {\ensuremath{q_c}})$$ Summing over all addresses $|a|=L$ with factors $w_a$ we get the earlier congruence.
Examining the intriguing quantities $\psi_c({\ensuremath{\varphi}}_a)$ further, using the notations of Definition \[s02040201\], we see that $$\sum_{|a|=L} w_a\psi_c({\ensuremath{\varphi}}_a) \equiv \frac{{\ensuremath{q_c}}}{2\pi}\sum_{|a|=L} w_a{\ensuremath{\vartheta}}_a + \frac{{\ensuremath{\Gamma_c}}}{2\pi}\sum_{|a|=L} w_a\ln{\ensuremath{\lambda}}_a\ \ ({\mathrm{mod}}\ {\ensuremath{q_c}})$$ Choosing the weights $w_k={\ensuremath{\lambda}}_k^s$ (where $s>0,\ \sum_k{\ensuremath{\lambda}}_k^s=1$ is the so-called pseudodimension of the IFS fractal $\langle T_1,\ldots,T_n\rangle$ which corresponds to its Hausdorff dimension if the open set condition holds) we have as the factor of $\frac{{\ensuremath{\Gamma_c}}}{2\pi}$ above the following quantity $$-\frac{1}{s}\sum_{|a|=L} {\ensuremath{\lambda}}_a^s\ln{\ensuremath{\lambda}}_a^s = -\frac{1}{s}\sum_{|a|=L} w_a\ln w_a$$ Considering the weights $w_k={\ensuremath{\lambda}}_k^s$ to be probabilities, this is the Gibbs Entropy Formula for a collection of classical particles. This can be interpreted as the entropy of the set of points $F_L=\{T_a(p): a\in{\ensuremath{\mathcal{A}}}, |a|=L\}$ at each level $L\in{\ensuremath{\mathbb{N}}}$ during the Chaos Game evolution towards the IFS fractal attractor $F=\langle T_1,\ldots,T_n\rangle = {\mathrm{Cl}}\{T_a(p):a\in{\ensuremath{\mathcal{A}}}\}$. Remarkably $\frac{1}{s}$ plays the role of Boltzmann’s constant, hinting at some potentially deeper interpretations.
The Evolution towards Invariance {#s0306}
--------------------------------
As we have reasoned in Section \[s0302\], the intermittent evolution towards a fully developed turbulent flow field, corresponds to the iteration of the weighted / probabilistic transfer operator towards the invariant flow ${\ensuremath{\overline{\psi}}}$ characterized in Theorem \[s03050001\].
One of the advantages of our model over classical Chaotic Advection formalism, is that the intermittent / iterative steps need not occur at equal time intervals, and in fact the spacing in time can be arbitrary. An interesting question is whether the fully developed state can be achieved in finite time. The answer is, it certainly can. If the time spacing of each iteration decreases say geometrically, then the total time to reach the $L=\infty$ level iteration (corresponding to ${\ensuremath{\overline{\psi}}}$) will be finite. This is clearly true for any convergent sequence of time spacing.
Another question which may arise when considering the previous sections, is how the global superposition of $n$ pushforwarded stream functions according to the transfer operator ${\mathrm{T}}$ relates to the local experimental picture of the splitting of eddies. Upon some contemplation, we may realize that the global and local viewpoints are the direct consequence of the associativity of transfer ${\mathrm{T}}({\mathrm{T}}^L\psi_{c,p}) = {\mathrm{T}}^L({\mathrm{T}}\psi_{c,p})$. Executing an iteration of ${\mathrm{T}}$ over the entire flow field ${\mathrm{T}}^L\psi_{c,p}$ corresponds to the local intermittent splitting of the initial eddy $\psi_{c,p}$ into $n$ eddies ${\mathrm{T}}\psi_{c,p}$ and therefore the splitting of all of its level $L$ iterates.
Equilibrium Points of the Invariant Flow {#s0307}
----------------------------------------
### Sink Singularities {#s030701}
From the representation shown for the invariant flow $${\ensuremath{\overline{\psi}}}(z) = \lim_{L{\ensuremath{\shortrightarrow}}\infty} \sum_{|a|=L} w_a\psi_c(z-T_a(p))$$ where the limit is taken in the sense of $({\ensuremath{[\Psi_{X,c}]}}, D)$, we see that ${\ensuremath{\overline{\psi}}}$ is an infinite shifted superposition of the eddy $\psi_c$ of character $c$, preserved by ${\ensuremath{\overline{\psi}}}$. Since $\psi_c$ is centered at the origin, each weighted eddy $w_a\psi_c(z-T_a(p))$ is centered at $T_a(p)$.
Therefore the set of sink singularities of ${\ensuremath{\overline{\psi}}}$ is precisely the countably infinite IFS fractal $(T_1,\ldots,T_n)=\{T_a(p):a\in{\ensuremath{\mathcal{A}}}\}$ where $p$ is the primary fixed point. Since as we have reasoned ${\ensuremath{\overline{\psi}}}$ represents a fully developed turbulent flow field, this resolves Mandelbrot’s Conjecture \[s02010301\]. Note that IFS fractals mostly have a non-integer Hausdorff dimension, but in certain special cases, such as Dragon Curves, the dimension can equal two.
Since the flow concentrates on an IFS fractal, we must focus our efforts on the study of such sets, in order to uncover the characteristics of fractal potential flows. We also emphasize that by Theorems \[s03010001\] and \[s03020001\], having similarity contractions is necessary for preserving the physicality of flows under transfer.
### Saddle Points {#s030702}
When visualizing ${\mathrm{T}}^L\psi_{c,p}$ for large enough $L\in{\ensuremath{\mathbb{N}}}$, one observes a thinning of the basins of attraction of each eddy center $F_L = \{T_a(p): a\in{\ensuremath{\mathcal{A}}}, |a|=L\}$ which raises the question whether the thinning continues on to a width of zero. The basins are partitioned by directrices (manifolds) of the saddle points (hyperbolic equilibria) of the flow field. In the neighbourhood of each partitioning infinite separatrix (stable manifold) belonging to these saddles (the unstable one ending in sinks), the flow behaviour becomes chaotic - meaning a tracer particle starting down the streamline near one side of the separatrix, may end up at a distant sink singularity, relative to if it had started near the other side. When the flow field is considered as the phase portrait of a Hamiltonian system, then this signifies sensitivity to initial conditions near the particular separatrix. For the entire flow field to be considered chaotic, we must show that such partitioning separatrices “cover” the flow field, in some sense. We also intuit that this covering would be implied if the saddle points could be shown to be dense in the countably infinite IFS fractal $(T_1,\ldots,T_n)$.
\[s03070201\] The infinite separatrices of the saddle points of ${\ensuremath{\overline{\psi}}}$ are dense in the plane.
\[s03070202\] The saddle points of ${\ensuremath{\overline{\psi}}}$ are dense in $(T_1,\ldots,{\mathrm{T}}_n)$.
We further conjecture that the second conjecture implies the first, which in turn guarantees the chaotic nature of the invariant flow field.
Visualization {#s04}
=============
When considering the transfer operator over stream functions, one encounters a proliferation of branch cuts, as the transfer iteration progresses. This can make the visualization of streamlines - the level curves of the stream function - quite difficult. Note that stream functions in ${\ensuremath{[\Psi_{X,c}]}}$ are only required to be almost everywhere smooth, and the transfer iteration results in a countably infinite Hutchinson union of branch cuts in ${\ensuremath{\overline{\psi}}}$, which pose no issue theoretically, despite the visual and algorithmic mess they create in practice. This is illustrated by Figure \[s04000001\] with transfer parameters $$p_1=0,\ p_2=1,\ {\ensuremath{\varphi}}_1=0.65 {\mathrm{e}}^{-\frac{2\pi}{6}{\mathrm{i}}},\ {\ensuremath{\varphi}}_2=0.65{\mathrm{e}}^{\frac{2\pi}{4}{\mathrm{i}}},\ w_{1,2}=0.5$$ Therefore in order to iteratively visualize the flow field, one seems to have no option but to resort to an iteration over the velocity field. $\psi_0$ can be chosen arbitrarily, as apparent from Theorem \[s03050001\], so we choose $\psi_0:=\sum_k w_k T_k^*\psi_{c_k,p_k}$ as discussed at the end of Section \[s0304\], where $c_k={\mathrm{Log}}\ {\ensuremath{\varphi}}_k$ (choosing $C=2\pi$ in Section \[s0301\]) for an IFS of Section \[s0204\]. The evolution in the velocity field progresses according to Theorem \[s03020001\]. So we have the following recursive iteration for the velocity field, over the flow space $({\ensuremath{[\Psi_{X,c}]}}, D)$ with preserved character $c=\sum_k w_k{\mathrm{Log}}\ {\ensuremath{\varphi}}_k$ and $X\supset\langle T_1,\ldots,T_n\rangle$. $$v_0(z)= v[\psi_0] = \sum_{k=1}^n w_k({\mathrm{Log}}\ {\ensuremath{\varphi}}_k)\frac{z-p_k}{|z-p_k|^2}$$ $$v_{L+1}(z)= v[{\mathrm{T}}\psi_L](z)= \sum_{k=1}^n w_k \frac{{\ensuremath{\varphi}}_k}{|{\ensuremath{\varphi}}_k|^2}\ v_L(T_k^{-1}(z))$$ Once we have arrived at a large enough iteration level $L\in{\ensuremath{\mathbb{N}}}$, we may execute a streamline solver algorithm, resulting in a preferable image as in Figure \[s04000002\]. The transfer iteration is visualized in Figure \[s04000003\]. Note the apparent convergence to an attractor flow field, as predicted by Section \[s03\].
In order to visualize the level curves of the corresponding potential function - the equipotential lines - we first note the correspondence in character ${\mathrm{char}}(\tilde{\psi})= -{\mathrm{i}}\ {\mathrm{char}}(\psi)$ by Definition \[s03030001\]. If one does not wish to deal with the arising sources in the conjugate to the above $\psi_0$ (as we did in Figure \[s04000004\]), we may choose simply instead $\psi_0:=\psi_{c,p}$ and thus $\tilde{\psi}_0=\psi_{-{\mathrm{i}}c,p}$ with the primary fixed point $p$ and the above $c$, as discussed in Section \[s0305\]. Then in order to arrive at the potential field at level $L$, we execute the same velocity recursion on $\tilde{v}_0(z):=v[\tilde{\psi}_0]$ as above on $v_0(z)= v[\psi_0]$. The iterative formula remains the same because of Theorems \[s03010001\] and \[s03020001\] implying the following, with $\tilde{v}_L:=v[\tilde{\psi}_L],\ L\in\{0\}\cup{\ensuremath{\mathbb{N}}}$. $$\tilde{v}_0(z)= v[\tilde{\psi}_{0}](z)= -{\mathrm{i}}c\ \frac{z-p}{|z-p|^2}$$ $$\tilde{v}_{L+1}= v[\tilde{\psi}_{L+1}]= v[\widetilde{{\mathrm{T}}\psi_L}]= v[{\mathrm{T}}\tilde{\psi}_L]= \sum_{k=1}^n w_k\frac{{\ensuremath{\varphi}}_k}{|{\ensuremath{\varphi}}_k|^2}\ T_k^*v[\tilde{\psi}_L]= \sum_{k=1}^n w_k\frac{{\ensuremath{\varphi}}_k}{|{\ensuremath{\varphi}}_k|^2}\ T_k^*\tilde{v}_L$$ Depending on the sign of ${\mathrm{Im}}\ c$, it may be more convenient from the viewpoint of a streamline solver, to take the negative of $\tilde{v}_0$ so that ${\mathrm{Re}}\ {\mathrm{char}}(\tilde{v}_0)<0$.
![Level curves of the stream function ${\mathrm{T}}^4\psi_0$ with logarithmic spiral branch cuts.[]{data-label="s04000001"}](./p00054_1d.jpg){width="405pt"}
![The same streamlines solved from the velocity field.[]{data-label="s04000002"}](./p00054_2b.jpg){width="405pt"}
![Converging transfer iteration from $\psi_0$ to ${\mathrm{T}}^7\psi_0$.[]{data-label="s04000003"}](./p00055_all3.jpg){width="470pt"}
![The equipotential lines at iteration level $L=6$, and a detail image.[]{data-label="s04000004"}](./p00053_Potentials_s.jpg "fig:"){width="405pt"}\
![The equipotential lines at iteration level $L=6$, and a detail image.[]{data-label="s04000004"}](./p00053_Potentials_s2.jpg "fig:"){width="405pt"}\
Conclusive Remarks {#s05}
==================
After attempting an extensive and analytical overview of Fully Developed Turbulence (FDT) in Section \[s0201\], we proceeded to introduce some theoretical concepts which we fused together in Section \[s03\] in the form of Fractal Potential Flows. Our goal had been two-fold: (1) to formulate the random intermittent evolution towards FDT as the iteration of a weighted transfer operator; (2) to show that this operator has a unique attracting fixed point, modelling the FDT flow field.
As far as suggestions for future research, we mention the conjectures stated in Section \[s030702\], as well as the inspirational remarks on entropy in Section \[s0305\]. Furthermore, the isometric isomorphism introduced in Section \[s0304\] between flows and probability measures, may be exploited to translate former results on measures to the language of flows, while examining the physical implications. Considering Section \[s030701\], the study of Fractal Potential Flows (and thus FDT) reduces essentially to the geometry of IFS Fractals with similarity contractions, so their investigation appears to be of fundamental importance.
We wish to express our appreciation to Dr. Sommerer and Dr. Tél for their clarifications of their work with Dr. Ott and Dr. Károlyi et al. and its relation to Aref’s theory of Chaotic Advection. We express gratitude for the support of Prof. Edward R. Vrscay whose unbiased skepticism also motivated the author to chisel this attractor and its literature foundations to an arguably sufficient level of iteration, with his never sharp enough IFS maps.
We dedicate our paper to the memory of Profs. Benoit B. Mandelbrot and Hassan Aref.
\[s05000001\] {width="310pt"}
We gratefully acknowledge the support of the Faculty of Mathematics at the University of Waterloo, and primarily the gracious support of Prof. Vrscay, supported by NSERC.
[^1]: Presented at the 2012 Winter Meeting of the Canadian Mathematical Society on December 10. To be submitted for publication. Comments are appreciated.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Quantum algorithms have been developed for efficiently solving linear algebra tasks. However they generally require deep circuits and therefore universal fault-tolerant quantum computers. In this work, we propose variational algorithms for linear algebra tasks that are compatible with Noisy Intermediate Scaled Quantum devices. We show that the solutions of linear systems of equations and matrix-vector multiplications can be translated as the ground states of the constructed Hamiltonians. Based on recently proposed variational quantum algorithms, we introduce Hamiltonian morphing together with an adaptive ansatz for efficiently finding the ground state. As the ground state always has zero energy, it also enables us to verify the solution. Our algorithms have wide applications in machine learning and optimisation problems, and the algorithm for matrix multiplications can be also used for variational Hamiltonian simulation. Finally, we conduct numerical simulations of the algorithm for solving linear systems of equations.'
author:
- Xiaosi Xu
- Jinzhao Sun
- Suguru Endo
- Ying Li
- 'Simon C. Benjamin'
- Xiao Yuan
bibliography:
- 'bib.bib'
title: Variational algorithms for linear algebra
---
*Introduction.—*Quantum computing has wide applications in various linear algebra tasks. A prominent example is the quantum algorithm for solving linear systems of equations by Harrow, Hassidim, and Lloyd (HHL) in Ref. [@PhysRevLett.103.150502]. Given an $N$ by $N$ sparse matrix ${{M}}$ and a state vector $\ket{v_0}$, the HHL quantum algorithm can prepare a state that is proportional to $
\ket{v_{{{M}}}} = {{{M}}}^{-1} \ket{v_0}$, with a complexity polynomial in $\log_2N$ and condition number $c$, which is the ratio of the largest eigenvalue to the smallest eigenvalue of a given matrix. As classical algorithms generally have polynomial complexity in $N$, the HHL algorithm suggests exponential speed-ups of quantum computers. Recent developments of quantum algorithms for linear systems of equations can be found in Ref. [@doi:10.1137/16M1087072; @ambainis2012variable; @PhysRevLett.110.250504; @PhysRevLett.120.050502; @chakraborty2018power; @gilyen2018quantum; @PhysRevLett.122.060504]. Another common linear algebra task is matrix-vector multiplication by applying a sparse matrix ${{M}}$ to a vector $\ket{v_0}$ as $
\ket{v_{{{M}}^{-1}}} = {{{M}}} \ket{v_0}$. The quantum algorithms for linear equations can be similarly applied for matrix-vector multiplications. Furthermore, they can be applied for quantum optimisation and machine learning [@PhysRevLett.113.130503; @lloyd2014quantum; @biamonte2017quantum].
The conventional quantum algorithms generally require a long-depth circuit for a fault tolerant quantum computer. This is challenging for current technology. Recently, there has been great interest in quantum computing in the Noisy Intermediate Scaled Quantum (NISQ) regime, for finding energy spectra of a many-body Hamiltonian [@peruzzo2014variational; @wang2015quantum; @PRXH2; @PhysRevA.95.020501; @VQETheoryNJP; @PhysRevLett.118.100503; @PhysRevX.8.011021; @Santagatieaap9646; @kandala2017hardware; @kandala2018extending; @PhysRevX.8.031022; @kokail2018self], simulating real and imaginary time dynamics of many-body systems [@Li2017; @mcardle2018variational; @2018arXiv181208767Y; @endo2018variational; @Fujirealtime19; @chen2019demonstration], applications with machine learning [@romero2017quantum; @farhi2018classification; @perdomo2018opportunities; @benedetti2018generative; @havlivcek2019supervised; @lamata2018quantum; @khoshaman2018quantum; @lloyd2018quantum], circuit learning [@mitarai2018quantum; @cincio2018learning; @jones2018quantum; @2019arXiv190304500B; @sharma2019noise], and others [@2019arXiv190709032L; @lee2019hybrid]. These algorithms are generally hybrid in a sense that they only solve the core problem with a shallow quantum circuit and leave the higher level calculation to be performed with a classical computer. Furthermore, even without error correction, noise in the shallow circuit implementation can be suppressed via error mitigation [@subspace1; @Li2017; @PhysRevLett.119.180509; @endo2017practical; @subspace2; @recoveringnoisefree; @PhysRevA.99.012334; @samerrormitigation; @bonet2018low], indicating the feasibility of quantum computing with NISQ hardware.
In this work, we propose variational algorithms for linear algebra problems, including linear systems of equations and matrix-vector multiplications, implemented with NISQ hardware. The main idea is to construct a many-body Hamiltonian so that its ground state corresponds to the solution of the linear algebra problem. Then we apply the recently proposed variational methods, such as variational quantum eigensolver (VQE) [@peruzzo2014variational] or imaginary time evolution (ITE) [@mcardle2018variational; @2018arXiv181208767Y], to find the ground state and hence the solution. We also introduce Hamiltonian morhping with an adaptive ansatz to address the barren plateau problem [@mcclean2018barren]. Different from the conventional scenario where the ground state energy is generally unknown, the ground state energy of the linear algebra Hamiltonians is designed to be zero in our algorithms, and this further enables us to verify the correctness of the solution. Meanwhile, we show that the variational matrix-vector multiplication algorithm can be applied for Hamiltonian simulation as an alternative to the one proposed by @Li2017. Finally, we present numerical simulation results for solving linear systems of equations with random matrices that have sizes up to 64\*64.
*Variational algorithms for linear algebra.—*We first introduce our variational algorithms for matrix-vector multiplication. Given a sparse $N$ by $N$ matrix ${{M}}$ and an initial state vector $\ket{v_0}$, our task is to calculate the normalised state $$\begin{aligned}
\label{Eq:MatrixMulti}
\ket{v_{{{M}}}} = \frac{{{{M}}} \ket{v_0}}{\|{{{M}}} \ket{v_0}\|},\end{aligned}$$ with $\|{{{M}}} \ket{v_0}\| = \bra{v_0}{{{M}}}^\dag {{{M}}} \ket{v_0}$. Here, we consider the case that ${{M}}$ can be a general (non-Hermitian) matrix. We find that $\ket{v_{{{M}}}}$ is the ground state of the Hamiltonian $$\begin{aligned}
H_{{{{M}}}}=I-\frac{{{{M}}}\ket{v_0}\bra{v_0}{{{M}}^\dag}}{\|{{{M}}} \ket{v_0}\|^2},\end{aligned}$$ with energy 0. Using universal quantum computers, one may apply the conventional techniques [@Abrams97; @aspuru2005simulated], such as adiabatic state preparation and phase estimation to find the ground state of $H_{{{{M}}}}$. With NISQ devices, we can instead consider the variational method with a parametrised state. The main idea is to first consider a state or ansatz, $\ket{\phi(\vec{\theta})}=U(\vec{\theta})\ket{0}$, with $U(\vec{\theta}) = U_L(\theta_L)\dots U_2(\theta_2)U_1(\theta_1)$, $\vec{\theta}=(\theta_1,\theta_2,...,\theta_L)$. Suppose the ground state of $H_{{{{M}}}}$ can be represented by the ansatz with certain parameters, the ground state finding problem is then converted to the minimisation, $$\begin{aligned}
\label{overlap}
\vec{\theta}_{\min}=\arg \min_{\vec{\theta}}\bra{\phi(\vec{\theta})} H_{{{{M}}}} \ket{\phi(\vec{\theta})},\end{aligned}$$ with the solution given by $\ket{v_{{{M}}}} = \ket{\phi(\vec{\theta}_{\min})}$. The minimisation can be accomplished with various methods, such as gradient based classical optimisation VQE [@peruzzo2014variational], global search algorithms [@kokail2018self], ITE [@mcardle2018variational; @2018arXiv181208767Y], etc. In the main text, we focus on the VQE method and leave the discussion of other method in Supplementary Materials. With VQE via gradient descent [@peruzzo2014variational], we start with a guess of the solution, $\vec{\theta}_0$ and update the parameters along the negative gradient of the energy $E_{{{M}}}(\vec\theta) = \bra{\phi(\vec{\theta})}H_{{{{M}}}}\ket{\phi(\vec{\theta})}$, $$\vec{\theta}_{i+1} = \vec{\theta}_{i} - a \nabla E_{{{M}}}(\vec\theta_i), \,\forall i = 0, 1, \dots, T$$ where $a$ is the time step, and $T$ is the total number of steps. As the energy $E_{{{M}}}$ monotonically decreases with sufficiently small $a$, the solution via VQE always at least corresponds to a local minimum. Starting from several random initial positions, one may find the true solution, i.e., the global minimum. In conventional VQE, it is generally hard to verify whether the true ground state is achieved. As we will discuss shortly, we show that we can verify the correctness of the final state found.
Next, we consider the problem of solving linear equations. Given a sparse $N$ by $N$ matrix ${{M}}$ and an initial state vector $\ket{v_0}$, we want to calculate $$\begin{aligned}
\label{Eq:MatrixLinear}
\ket{v_{{{M}}^{-1}}} = \frac{{{M}}^{-1} \ket{v_0}}{\|{{{M}}}^{-1} \ket{v_0} \|},\end{aligned}$$ where $\|{{{M}}}^{-1} \ket{v_0} \| = \sqrt{\bra{v_0}{{{M}}}^{-1} {{{M}}}^{-1} \ket{v_0}}$. When the matrix ${{M}}$ is not Hermitian, it can always be converted into an equivalent problem with a Hermitian matrix by doubling the number of qubits [@PhysRevLett.103.150502]. We therefore focus on the case where ${{M}}$ is Hermitian and invertible. It is not hard to see that the solution $\ket{v_{{{M}}^{-1}}}$ is the ground state of the Hamiltonian $$\begin{aligned}
H_{{{{M}}}^{-1}}={{{M}}}^\dag (I-\ket{v_0}\bra{v_0}){{{M}}}\end{aligned}$$ with energy $0$. Again, one can apply adiabatic algorithms to find the ground state with a universal quantum computer [@PhysRevLett.122.060504]. Here, we focus on the variational algorithm and consider the following optimisation problem $$\begin{aligned}
\label{overlap}
\vec{\theta}_{\min}=\arg \min_{\vec{\theta}}\bra{\phi(\vec{\theta})} H_{{{{M}}}^{-1}} \ket{\phi(\vec{\theta})},\end{aligned}$$ with $\ket{v_{{{M}}^{-1}}} = \ket{\phi(\vec{\theta}_{\min})}$. With VQE, we start with a guess $\vec{\theta}_0$ and update the parameters by $$\vec{\theta}_{i+1} = \vec{\theta}_{i} - a \nabla E_{{{M}}^{-1}}(\vec\theta_i), \,\forall i = 0, 1, \dots, T$$ with energy $E_{{{M}}^{-1}}(\vec\theta) = \bra{\phi(\vec{\theta})}H_{{{{M}}}^{-1}}\ket{\phi(\vec{\theta})}$, time step $a$, and total number of steps $T$.
*Implementation with quantum circuits—*Now, we consider the implementation of our algorithms with quantum circuits. To realise the VQE optimisation, we need to obtain $\nabla E_{{{M}}^{-1}}$ or $\nabla E_{{{M}}}$, which can be estimated either with the finite difference formulae or quantum gradient finding methods recently considered in Ref. [@Li2017; @dallaire2018low; @romero2018strategies]. For simplicity, we only focus on $E = E_{{{M}}}$ in the main text and leave the discussion of $E = E_{{{M}}^{-1}}$ in Supplementary Materials.
With the finite difference formulae, each component ${\partial E_{{{M}}}}/{\partial \theta_i}$ of the gradient $\nabla E_{{{M}}}=(\partial E_{{{M}}}/\partial \theta_i)$ around $\vec\theta$ can be calculated as $$\frac{\partial E_{{{M}}}}{\partial \theta_i} \approx \frac{E_{{{M}}}(\vec{\theta}+\delta \theta_i) - E_{{{M}}}(\vec{\theta}-\delta \theta_i)}{2\delta \theta_i},$$ with $\delta\theta_i \ll 1$. Note that $E_{{{M}}}(\vec\theta)=1-|\bra{\phi(\vec{\theta})}{{{M}}}\ket{v_0}|^2$ with an arbitrary angle $\vec\theta$. Suppose ${{M}}$ can be decomposed as ${{{M}}}=\sum_j \lambda_j \sigma_j$, we can then obtain $E_{{{M}}}(\vec\theta)$ as $$\begin{aligned}
E_{{{M}}}(\vec\theta) = 1 - \big|\sum_i \lambda_i \bra{\phi(\vec{\theta})}\sigma_i \ket{v_0} \big|^2.
\label{overlap2}\end{aligned}$$ For each term, we denote $\bra{\phi(\vec{\theta})}\sigma_i \ket{v_0}=\braket{0|U|0}$ with $U=U(\vec\theta)^\dag\sigma_iU_{v_0}$ and $\ket{v_0} = U_{v_0}\ket{0}$. The real and imaginary part of $\bra{\phi(\vec{\theta})}\sigma_i \ket{v_0}$ can be evaluated with the Hardamard test or swap test quantum circuits. Suppose ${{M}}$ consists of a polynomial number (with respect to the number of qubits) of tensor products of local operators $\sigma_j$, we can then measure each term and efficiently calculate $E_{{{M}}}(\vec\theta)$. Meanwhile, even if ${{{M}}}=\sum_j \lambda_j \sigma_j$ has exponential terms with respect to the number of qubits, the algorithm is still efficient if we can efficiently sample the probability distribution $|\lambda_i|/\sum_i|\lambda_i|$ with a polynomial dependence of $\sum_i|\lambda_i|$ on the number of qubits. This includes cases where the coefficients have analytical expressions, which enables efficient sampling, or the matrix is an oracle-based sparse matrix. With the quantum gradient finding method, we first express the gradient as $$\frac{\partial E_{{{M}}}}{\partial \theta_i} = -2\Re\left(\frac{\partial \bra{\phi(\vec{\theta})}}{\partial \theta_i}{{M}}\ket{v_0}\bra{v_0}{{M}}^\dag\ket{\phi(\vec{\theta})}\right).
\label{Equ::energyequation}$$ The term $\bra{v_0}{{M}}^\dag\ket{\phi(\vec{\theta})}$ can be efficiently measured with the aforementioned method. To measure $\frac{\partial \bra{\phi(\vec{\theta})}}{\partial \theta_i}{{M}}\ket{v_0}$, we have to first decompose the derivative of the state as $$\frac{\partial \ket{\phi(\vec{\theta})}}{\partial \theta_i} = \sum_s f_i^s \ket{\phi_{i}^s(\vec\theta)}.$$ Here $\ket{\phi_{i}^s(\vec\theta)} = U_L(\theta_L)\dots \sigma_i^s U_i(\theta_i)\dots U_2(\theta_2)U_1(\theta_1)\ket{0}$, and the derivative of the unitary is decomposed as $\partial U_i(\theta_i)/\partial \theta_i = \sum_sf_i^s\sigma_i^sU_i(\theta_i)$ with tensor products of local operators $\sigma_i^s$ and coefficients $f_i^s$. Then we have $\frac{\partial \bra{\phi(\vec{\theta})}}{\partial \theta_i}{{M}}\ket{v_0}=\sum_s f_i^{s*} \bra{\phi_{i}^s(\vec\theta)}{{M}}\ket{v_0}$ where each term can be efficiently measured with a quantum circuit.
*Solution verification.—*After finding the solution, it is also practically important to verify whether it is the correct one. This is in general impossible for conventional VQE as neither the ground state nor the ground state energy is known. However, we can verify the solution of the linear algebra problems as the exact solution should always have zero energy. We first consider the matrix-vector multiplication problem. With a solution $\ket{\phi(\vec\theta_{\min})}$, we can estimate $|\bra{\phi(\vec\theta_{\min})}{v_{{{M}}}}\rangle|^2$ via $$|\bra{\phi(\vec\theta_{\min})}{v_{{{M}}}}\rangle|^2 = 1 - E_{{{M}}}(\vec\theta_{\min}).$$ Therefore, we can also have the fidelity of $\ket{\phi(\vec\theta_{\min})}$ to the exact solution $\ket{v}$ from the energy $E_{{{M}}}(\vec\theta_{\min})$ and the solution can be verified as correct whenever $|\bra{\phi(\vec\theta_{\min})}{v}\rangle|^2$ is close to 1. Similarly, the solution of the linear equation can be verified by measuring the energy $E_{{{M}}^{-1}}(\vec\theta_{\min})$ to be close to $0$. Note that the fidelity $|\bra{\phi(\vec\theta_{\min})}{v_{{{M}}^{-1}}}\rangle|^2$ cannot be determined as we cannot directly measure $|\bra{\phi(\vec\theta_{\min})}{{M}}^{-1}\ket{v_0}|^2$ or $\bra{v_0}{{{M}}}^{-1} {{{M}}}^{-1} \ket{v_0}$. Alternatively, we can verify the solution by checking whether $\ket{v_0} \propto {{M}}\ket{\phi(\vec\theta_{\min})}$ is satisfied. Therefore, we can measure ${|\bra{v_0}{{M}}\ket{\phi(\vec\theta_{\min})}|^2}/{\bra{\phi(\vec\theta_{\min})}{{{M}}}^\dag {{{M}}} \ket{\phi(\vec\theta_{\min})}}$ and the solution is verified as correct when the value is close to 1. This can be an alternative cost function for VQE and we leave the discussion to future works.
*Optimisation via Hamiltonian morphing.—*The VQE method tries to find the global minimum of a large multi-parameter space. For each trial, one starts from a randomly initialised parameter set and obtains a local minimum after several optimisation steps. One can then verify whether the solution is correct by measuring the energy and restart from another random parameter set until the energy is close to 0. This may need to be repeated many times until a satisfactory solution can be found. There are several potential issues that can affect the practical performance. First, as the parameter space is large, it may require a large number of repetitions starting from different initial random parameters. This is a common issue in optimisation and machine learning, where a combination of different algorithms may help to speed up the search time. A more serious problem is that the gradient may vanish for randomly initialised parameters in some random quantum circuits of large systems [@mcclean2018barren]. In quantum computational chemistry, physically motivated ansatz and a good guess of initial parameters are considered to avoid vanishing gradient, see Refs. [@ourReview; @cao2018quantum] for recent reviews. However, this is not straightforward for the linear algebra tasks.
Here we propose a Hamiltonian morphing optimisation method to avoid these problems. Similar methods have been introduced in Refs. [@PhysRevA.92.042303; @saez2018adiabatic] for variational quantum simulation. The morphing method is an analog to adiabatic state preparation, where one slowly varies the Hamiltonian from $H_i$ to $H_f$, as to gradually evolve the corresponding ground state from $\ket{\psi_i}$ to $\ket{\psi_f}$. For our morphing method, we consider the linear equation problem as an example and it works similarly for the matrix-vector multiplication problem. Denote a time-dependent Hamiltonian $H_{{{M}}^{-1}}(t)={{M}}(t)(I-\ket{v_0}\bra{v_0}){{M}}(t)$ with ${{M}}(t)$ satisfying ${{M}}(t=0)=I$ and ${{M}}(t=T)={{M}}$ respectively. Our algorithm starts from the ground state $\ket{v_0}$ of $H_{{{M}}^{-1}}(0)$ and reaches the ground state of the target Hamiltonian $H_{{{M}}^{-1}}(T)$ as follows. Consider discretised time step $\delta t$ and ${{M}}(t)=(1-t/T)I+ t/T\cdot {{M}}$.
1\. At the $n=0$ step with time $t=0$, we denote the parameters that represent $\ket{v_0}$ as $\vec \theta(0)$, i.e., $\ket{v_0} = \ket{\phi(\vec \theta(0))}$.
2\. At the $n^{\textrm{th}}\in [1, T/\delta t)$ step, we use parameters $\vec \theta((n-1)\delta t)$ found in the last step for $H_{{{M}}^{-1}}((n-1)\delta t)$ as the initial position to find parameters $\vec \theta(n\delta t)$ that correspond to the ground state of $H_{{{M}}^{-1}}(n\delta t)$.
For any positive matrix $M$, when we use a sufficiently small $\delta t$ and a powerful enough ansatz, our method is guaranteed to find the exact solution according to the adiabatic theorem [@RevModPhys.90.015002]. For a general ${{M}}$, one can introduce an extra qubit so that the gap $H_{{{M}}^{-1}}(t)$ is non-vanishing. We refer to Ref. [@PhysRevLett.122.060504] for details. In practice, the initially specified ansatz may not be powerful enough to represent the states at all time $t$. Fortunately, when the ansatz is insufficient at any time $t$, one can detect it by measuring a high averaged energy. Then one can either simply choose a more powerful ansatz or adaptively vary the ansatz by changing its structure and adding more gates [@jones2018quantum]. Note that this is in general not possible for conventional many-body ground state problems with VQE, owing to the unknown ground state energy.
*Hamiltonian simulation.—*Our algorithms can be useful for tasks in quantum machine learning and optimisation. The matrix-vector multiplication algorithm can also be applied for variational Hamiltonian simulation. Starting from $\ket{\psi_0}$, the Hamiltonian simulation task is to prepare the state $\ket{\psi_t} = e^{-iHt}\ket{\psi_0}$ at time $t$ with sparse Hamiltonian $H$. For instance, when the Hamiltonian can be decomposed as a linear sum of polynomial terms $H=\sum_j \lambda_j \sigma_j$, we can make use of the Trotterization method to have $$\ket{\psi_t} \approx \left(\prod_ie^{-i\lambda_j \sigma_jt/K}\right)^K\ket{\psi_0} + O(t^2/K),$$ with $K$ Trotter steps. As the circuit depth increases with the evolution time, the circuit depth can be large for long evolution time. The variational quantum algorithm for simulating real-time dynamics with a constant circuit depth has been proposed by @Li2017. Here, we present an alternative algorithm based on matrix-vector that can be more robust to the one in Ref. [@Li2017]. The key idea of Ref. [@Li2017] is to prepare the state at any time $t$ via a parametrised state $\ket{\phi(\vec\theta(t))}$ and update the parameters from $\vec\theta(t))$ to $\vec\theta(t+\delta t))$ that effectively realise $e^{-iH\delta t}\ket{\phi(\vec\theta(t))}$ via a deterministic process $\vec\theta(t+\delta t)) = \vec\theta(t) + \dot{\vec{\theta}}(t) \delta t$ with $\dot{\vec{\theta}}(t) = A^{-1}C$, where $A_{i,j}(t)=-\textrm{Im}\left(\frac{\partial\bra{\phi(\vec{\theta})}}{\partial\theta_i} \frac{\partial\ket{\phi(\vec{\theta})}}{\partial \theta_j}\right)$ and $C_i(t)=\textrm{Re}\left(\frac{\partial\bra{\phi(\vec{\theta})}}{\partial\theta_i} H \ket{\phi(\vec{\theta})}\right)$. Each term of $A$ and $C$ can be efficiently measured via a quantum circuit.
Instead of deterministically obtaining $\vec\theta(t+\delta t))$ via $\dot{\vec{\theta}}(t)$, we make use of the algorithm introduced in this paper. That is, to realise $e^{-iH\delta t}\ket{\phi(\vec\theta(t))}$, we can consider ${{M}}= e^{-iH\delta t}\approx 1 - iH\delta t$ and update the parameters as $\ket{\phi(\vec\theta(t+\delta t))} = {{M}}\ket{\phi(\vec\theta(t))}$. Therefore, for each time step, we can update the parameters according to the matrix-vector algorithm.
*Numerical simulation.—*We numerically test the variational algorithm for solving linear systems of equations. The simulation is based on the Quantum Exact Simulation Toolkit (QuEST) package [@jones2019quest], which is a high performance classical simulator of general quantum circuits. In our simulation, we consider $2^n$ by $2^n$ positive complex Hermitian matrices $M$ with $n$ qubits. Note that even though the matrix size is assumed to be an exponential of 2, an $m$ by $m$ matrix with $2^{n-1}<m\le2^n$ can be also handled with an $n$-qubit circuit. The matrix is randomly generated with a given condition number $c$ and [the input state is also randomly generated.]{} As we use no prior information of the matrix, we consider the ansatz as shown in Fig. \[fig:/ansatz\], where each green blocks $V_i$ has a fixed structure. We also vary the number of blocks, called the circuit depth $m$, to endow the ansatz with different powers. In a more practical scenario, we can assume that we have a sparse decomposition of the matrix and the input state is simple to prepare or it is obtained from a previous calculation. We can also try different ansätze such as the Hamiltonian anstaz, defined based on the Pauli decomposition of the matrix. As this work only aims to verify the basic function of the algorithm, we leave the simulation of the practical cases to a future work.
We also implement the optimisation based on Hamiltonian morphing with an adaptive ansatz. [We divide the total time $T$ into 10 intervals with an equal duration, where $T$ ranges from 20 to 100 depending on the matrix size.]{} At $t=0$, we set $H_{M^{-1}}=I-|v_0\rangle\langle v_0|$ with the ground state being the input state $|v_0\rangle$ and start with all single-qubit gates parameterised with a small random rotation angle. In the $i^{\textrm{th}}$ variational cycle, the gradient descent method is used to search for the ground state energy of a time-dependent Hamiltonian $H_{{{M}}^{-1}}(t)={{M}}(t)(I-\ket{v_0}\bra{v_0}){{M}}(t)$ with ${{M}}(t)=(1-i/10)I+i/10\cdot M$, [and a step size $\delta t=0.1$.]{} We choose initial parameters to be the ones obtained from the previous cycle. A small circuit depth $m$ is tried at the beginning and is increased gradually until the fidelity of the qubit state to the target state is higher than 99%, which is regarded as a success.
![The circuit ansatz. In total $n$ qubits are required to solve a problem with a $2^n*2^n$ matrix. Here, $R_y$ and $R_z$ represent single qubit rotations around the $Y$ and $Z$ axes, respectively. The rotation angle (parameter) for each single-qubit gate is initialised from a small random value, which is updated in each of the variational cycle. At the beginning two single-qubit gates are applied to each of the data qubit, followed with sets of CNOT gates, as depicted as the green block $V$. Within each set, a CNOT gate is followed with two single-qubit gates. The number of sets represents the depth of the circuit, which is gradually increased until a desired target state is found. In the end, a set of CNOT gates is applied, which is in the reversed order as in $V$. The aim for this set is to ensure that the eigenstate can be found at $t=0$, when the Hamiltonian morphing technique is applied.[]{data-label="fig:/ansatz"}](la.pdf "fig:"){width="1\columnwidth"}\
![The probability to find the target state with increasing condition number and circuit depths. This experiment is based on a 6-qubit circuit, thus the matrix size is fixed to be 64 \*64. For a given condition number, the success (having a solution with a fidelity higher than 99%) probability is shown for the circuit with increasing depths, until it reaches 100% (i.e., all the 4940 runs). The minimum depth required to find the target state with 100% success probability is plotted in the inserted graph.[]{data-label="fig:/prob1"}](prob1.pdf "fig:"){width="1\columnwidth"}\
![The probability to find the target state with increasing number of qubits and circuit depths. For a given number of qubits, circuit depth is increased gradually until the probability reaches 1. The inserted graph shows the minimum depth found in the simulation when the target state is found with 100% probability.[]{data-label="fig:/prob2"}](prob2.pdf "fig:"){width="1.04\columnwidth"}\
We study the complexity of our algorithm with respect to the matrix size and the condition number of the matrix. We first fix the matrix size to be $2^6*2^6$ with $6$ qubits and consider random matrices with different condition numbers. For different condition numbers from 10 to 100, we test our algorithm with different circuit depth. For each case, we run 4940 random trials to calculate the averaged success probability with results shown in Fig. \[fig:/prob1\]. Not surprisingly, we found that a larger circuit depth ansatz generally leads to a higher success probability. We also obtain the minimum depth that is required to guarantee the success of finding $|v_{M^{-1}}\rangle$, plotted as an insert. Overall, we find that the minimal required depth varies linearly with the condition number. Next, we vary the size $2^n*2^n$ of the matrix $M$, by changing the number of data qubits $n$ in the circuit ansatz. For an $n$-qubit system, the condition number is adopted as $c=10n$. As shown in Fig. \[fig:/prob2\], the circuit depth leading to a non-zero success probability also increases with respect to the number of qubits. The inserted plot shows the minimum circuit depth for achieving a 100% success probability. Overall, we find a super-linear increase of the circuit depth with respect to the number of qubits.
This is not surprising as we consider random matrices with random input states. Even though we consider fixed condition numbers, it only determines properties of eigenvalues of the matrix. For any matrix $M$, all other matrices $UMU^\dag$ with unitary $U$ have the same eigenvalues and hence the same condition number. As the unitary $U$ is an arbitrary unitary, it can lead to a general solution state, which may not be able to be prepared with shallow circuits. This explains the super-linear dependence of the circuit depth with respect to the number of qubits. One possible solution to this problem is to consider sparse matrices as in Ref. [@PhysRevLett.103.150502]. Meanwhile, a better ansatz whose design is informed by the matrix can be designed. For example, we note that the chosen ansatz is not optimal for the problem, and a much more compact one can be obtained from circuit compiling [@jones2018quantum] as shown in the Supplementary Materials. As our algorithm has the capability of verifying the solution, it also enables us to dynamically varying the ansatz until the solution is found. In our simulation, we dynamically increase the circuit depth and we expect more profound circuit morphing techniques can be designed and exploited.
*Discussion.—*In this work we present variational algorithms for solving linear algebra problems including linear system of equations and matrix-vector multiplications with NISQ devices. We design a Hamiltonian with the ground state as the solution of the target problem, and solve it with the recently-proposed variational algorithms. As the ground state has zero energy, it also enables us to verify the solution and adjust the ansatz. We also introduce the Hamiltonian morphing technique to avoid local minima and to accelerate the search progress, and numerically verify our algorithm for solving the linear systems of equations. Owing to computation limitations, we consider randomly chosen matrices with a small range of matrix size. Within this range and with the non-optimal ansatz, the result suggests that the circuit depth based on our algorithm scales linearly with the condition number and super-linearly with the number of qubits. We expect further improvements of the performance of the algorithm for sparse matrices together with more sophisticated ansätze. We also expect that our results may shed light on quantum machine learning and quantum optimisations in the NISQ era.
Acknowledgements. {#acknowledgements. .unnumbered}
=================
This work is supported by the EPSRC National Quantum Technology Hub in Networked Quantum Information Technology (EP/M013243/1). SE is supported by Japan Student Services Organization (JASSO) Student Exchange Support Program (Graduate Scholarship for Degree Seeking Students). YL is supported by NSAF (Grant No. U1730449).
The lowest eigenstate of $H_{{{{M}}}^{-1}}$
===========================================
Here, we show the lowest eigenstate of $H_{{{{M}}}^{-1}}$ is ${{{M}}}^{-1} \ket{v_0}$. For an arbitrary vector $\ket{\varphi}$, $$\begin{aligned}
\bra{\varphi}H_{{{{M}}}^{-1}}\ket{\varphi}
&= \bra{\varphi}{{{M}}} (I-\ket{v_0}\bra{v_0}){{{M}}}\ket{\varphi} \\
&=\sum_i |\bra{\varphi}{{{M}}}\ket{\varphi_i} |^2 \geq 0, \end{aligned}$$ where we denoted $I-\ket{v_0}\bra{v_0}=\sum_i \ket{\varphi_i}\bra{\varphi_i}$, $\braket{\varphi_i|\varphi_j}=\delta_{i,j}$. Therefore $H_{{{{M}}}^{-1}}$ is a positive semidefinite matrix. On the other hand, $H_{{{{M}}}^{-1}} {{{M}}}^{-1} \ket{v_0} =0$, thus ${{{M}}}^{-1} $ is the lowest eiegnvector of $H_{{{{M}}}^{-1}}$.
Optimisation methods searching for $\vec\theta_{min}$
=====================================================
Search for the optimal set of parameters based on the cost function can be generalised as an optimisation problem, which can be handled with many methods. We describe the approach of imaginary time evolution approach [@2018arXiv181208767Y] here. Based on the McLachlan’s variational principle, the minimisation starts with a initial guess of parameters $\vec\theta_0$, which is to be updated in each cycle with its derivative given by: $$\sum_j M_{i,j}\dot{\theta_j}=-V_i,$$ where $$\begin{aligned} &M_{i,j}=\Re\left(\frac{\partial \langle\phi(\vec\theta)|}{\partial\theta_i}\frac{\partial |\phi(\vec\theta)\rangle}{\partial\theta_j}\right),
&V_i=\Re\left(\frac{\partial \langle\phi(\vec\theta)|}{\partial\theta_i}H|\phi(\vec\theta)\rangle\right).
\end{aligned}$$ The Hamiltonian $H$ for the two cases described in the main text would be $H_M$ and $H_{M^{-1}}$ respectively. By updating $\vec\theta$ via: $$\vec\theta_{i+1}=\vec\theta_i+\Delta T\dot{\vec{\theta_i}}$$ with a sufficiently small time step $\Delta T$, the corresponding energy $E_M$ or $E_{M^{-1}}$ would decrease along the optimisation trajectory.
Circuit implementation of $\nabla E_{M^{-1}}$
=============================================
Here we show that the circuits to implement $\nabla E_{M^{-1}}$ can be found similarly as we do for $\nabla E_M$. We have $$\begin{aligned}
H_{{{{M}}}^{-1}}={{{M}}}^\dag (I-\ket{v_0}\bra{v_0}){{{M}}},\end{aligned}$$ thus $$\begin{aligned}
\frac{\partial E_{M^{-1}}}{\partial \theta_i}&=\frac{\partial}{\partial \theta_i}(\bra{\phi(\vec\theta)}H_{M^{-1}}\ket{\phi(\vec\theta)})=2\Re\left(\frac{\partial\bra{\phi(\vec\theta)}}{\partial \theta_i}H\ket{\phi(\vec\theta)}\right)\\
&=2\Re\left(\frac{\partial\bra{\phi(\vec\theta)}}{\partial \theta_i}M^\dag M\ket{\phi(\vec\theta)}\right)-2\Re\left(\frac{\partial\bra{\phi(\vec\theta)}}{\partial\theta_i}M^{\dagger}\ket{v_0}\bra{v_0}M\ket{\phi(\vec\theta)}\right).
\end{aligned}
\label{Equ::energyinverse}$$
Suppose $M$ can be decomposed into combinations of Pauli terms $M=\sum_j\alpha_j\sigma_j$, $M^{\dag}M$ can then be expressed as $M^{\dag}M=\sum_j\beta_j\sigma_j$, where the coefficients $\alpha$ and $\beta$ for each term are different. We decompose the derivative of the state as $$\frac{\partial \ket{\phi(\vec{\theta})}}{\partial \theta_i} = \sum_s f_i^s\ket{\phi_{i}^s(\vec\theta)},$$ the same way as described in the main text. Then the first term in Equation \[Equ::energyinverse\] is expressed as
$$\begin{aligned}
\frac{\partial \bra{\phi(\vec{\theta})}}{\partial \theta_i}M^{\dag}M\ket{\phi(\vec{\theta})}&=\sum_s f_i^{s*} \bra{\phi_{i}^s(\vec\theta)}M^{\dag}M\ket{\phi(\vec\theta)}\\
&=\sum_j\sum_s c_i^{s,j}\bra{0}U_1^{\dag}(\theta_1)U_2^{\dag}(\theta_2)\dots U_i^{\dag}(\theta_i)\sigma_i^s\dots U_L^{\dag}(\theta_L)\sigma_jU_L(\theta_L)\dots U_2(\theta_2)U_1(\theta_1)\ket{0}.
\end{aligned}$$
The real and imaginary value of each term $\bra{0}U_1^{\dag}(\theta_1)U_2^{\dag}(\theta_2)\dots U_i^{\dag}(\theta_i)\sigma_i^s\dots U_L^{\dag}(\theta_L)\sigma_jU_L(\theta_L)\dots U_2(\theta_2)U_1(\theta_1)\ket{0}$ can be evaluated with the circuit shown in Figure \[fig:circuitimp\](a), by initialising the ancilla qubit to be at $\frac{\sqrt{2}}{2}(\ket{0}+\ket{1})$ and $\frac{\sqrt{2}}{2}(\ket{0}+i\ket{1})$ respectively.
The second term can be separated as two parts, $\frac{\partial\bra{\phi(\vec\theta)}}{\partial\theta_i}M^{\dagger}\ket{v_0}$ and $\bra{v_0}M\ket{\phi(\vec\theta)}$. The first part can be expressed as $$\begin{aligned}
\frac{\partial\bra{\phi(\vec\theta)}}{\partial\theta_i}M^{\dagger}\ket{v_0}&=\sum_s f_i^{s*}\bra{\phi_{i}^s(\vec\theta)}M^{\dag}U_{v_0}\ket{0}\\
&=\sum_j\sum_s d_i^{s,j}\bra{0}U_1^{\dag}(\theta_1)U_2^{\dag}(\theta_2)\dots U_i^{\dag}(\theta_i)\sigma_i^s\dots U_L^{\dag}(\theta_L)\sigma_jU_{v_0}\ket{0},
\end{aligned}$$ where each term can be evaluated with the circuit shown in Figure \[fig:circuitimp\](b). The second part is written as $$\bra{v_0}M\ket{\phi(\vec\theta)}=\bra{0}U_{v_0}^\dag M U(\vec\theta)\ket{0}=\sum_jk_j\bra{0}U_{v_0}^\dag\sigma_jU(\vec\theta)\ket{0}$$ where each term can be evaluated with the circuit shown in Figure \[fig:circuitimp\](c).
![The circuits to implement $\nabla E_{M^{-1}}$. The first qubit is the ancilla, and the second refers to the data qubit register. The real part can be obtained from measuring $\braket{X}$ when the ancilla initialised at $\frac{\sqrt{2}}{2}(\ket{0}+\ket{1})$, while the imaginary part is evaluated if the ancilla is initialised at $\frac{\sqrt{2}}{2}(\ket{0}+i\ket{1})$.[]{data-label="fig:circuitimp"}](circuitImp.pdf "fig:"){width="0.8\columnwidth"}\
Example with a more compact circuit
===================================
In this paper we apply the hardware-efficient ansatz for practical soundness, however, this is not the optimal ansatz for most cases. Here we take one case as an example, where the solution is found difficult to reach in our numerical simulations.
![A more compact circuit found with quantum compilation to realise the target state. Each single- and multi-qubit gate is controlled with one parameter. In total 126 parameters are included in the circuit, which realises the state with a fidelity higher than 99.9%. []{data-label="fig:circuitrec"}](localCircuit.pdf){width="1\columnwidth"}
The matrix in this case is a complex matrix with size of 64\*64, and condition number $c=60$. $\ket{v_0}$ is a complex vector, and the target state is $\ket{v_{{{M}}^{-1}}} = \frac{{{M}}^{-1} \ket{v_0}}{\|{{{M}}}^{-1} \ket{v_0} \|}$. It is found in the numerical simulation that the target state can not be represented by a circuit with depth 7 but can by a circuit with depth 8. On the other hand, applying quantum compiling, it is found that the target state can be realised with a circuit that requires only half the number of parameters. Figure \[fig:circuitrec\] is one example of the circuits found to prepare a state with fidelity higher than 99.9% compared with the target state. A limited topology with only nearest-neighbor interactions is considered here as it is more hardware efficient. The detailed description and the parameters can be found at https://questlink.qtechtheory.org/.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A peer-to-peer system is a distributed system in which equal nodes (in terms of role and usage) exchange information and services directly. This paper describes a distributed peer-to-peer protocol that allows wifi-enabled smart devices (especially Android smartphones) to exchange data using only wifi. The protocol is designed to allow the automatic establishment of a distributed peer-to-peer network of any size without geography constraint. It is applied to file sharing between devices, but can easily be adapted to support any data sharing. The protocol defines two layers: (1) a *kernel layer* responsible for creating, routing, establishing and maintaining links between nodes (peer), addressing a node and adding and removing nodes; and an (2) application layer to support data sharing (file sharing in this case). The structure of the peer-to-peer network is hybrid. Regarding the file sharing use case, the application allows a node to (i) search for a file in the catalog of its subnet that is held by the *root* node, and (ii) download a file: if the file is not in the subnet, the root node delegates a node to make vouchers in neighboring subnetworks to get the file and make it available. A proof of concept was made on Android.'
address: 'National Advanced School of Engineering, Yaoundé, Cameroon'
author:
- 'Steve Tueno, Romeo Tabue, Forentin Jiechieu, Yacynth Ndonna, Billy Zafack, Audric Feuyan, Jonas Atibita, Alex Djouontse and Rodrigue Mbinkeu'
bibliography:
- 'references.bib'
title: |
Pear2Pear (On Wifi):\
A Data Sharing Protocol Over Wifi through a Peer to Peer Network
---
Distributed System ,Peer-to-Peer ,Network ,File Sharing ,Wifi ,Routing
Introduction {#introduction .unnumbered}
============
Since the beginning of the Internet, the client-server model was the reference model for the provision of resources. In this model, the system relies on a dedicated server that centralizes and maintains all resources and services. Therefore, the increase in the number of users requires a greater investment of service providers. It is indeed required to guarantee the availability of resources and services, despite the large number of simultaneous requests. This requires significant resources and imposes software and hardware constraints, which make such systems very expensive and difficult to secure and scale because of the single point of failure. However, the increase in the number of users also implies an increase in cumulative resources (computing and storage), hence the peer-to-peer paradigm.
The peer-to-peer paradigm consists of connecting users in order to pool resources and distribute tasks [@DBLP:conf/vldb/YangG01]. A peer-to-peer system is thus a distributed system of connected peers, consumers and service providers. Systems based on this paradigm are implemented as logical networks, connecting participants over physical networks. The work presented in this paper reports on the design of an hybrid peer-to-peer protocol in which peers are wifi-enabled mobiles not connected to Internet.
Background
==========
The Peer-to-Peer Paradigm
-------------------------
The peer-to-peer paradigm (P2P) has been focused from the outset on the management of large amounts of data spread over a very large scale. Although it has been used so far mainly for file sharing purposes, many research projects are interested in other types of applications including collaboration, distributed computing, sharing and distribution of content.
### Definition
A peer-to-peer system is a distributed system in which equal nodes (in terms of role and usage) directly exchange information and services [@DBLP:conf/vldb/YangG01]. Behind this academic definition is the reality of the context of use of these systems. Indeed, it is the role of peers to put and maintain each peer-to-peer system in place. There is no central entity that controls the system. This entails a number of advantages, but also constraints. For benefits, increasing the number of participants also increases the resources available while distributing the load among participants. It is therefore theoretically possible to scale to a relatively low cost. In addition, the absence of a central control entity gives the system greater robustness, since the availability of resources is not directly related to a particular peer [@dalle2008analysis]. For constraints, the absence of a central entity requires the establishment of a self-organizing mechanism. Peers must connect to or disconnect from the network without affecting the availability of resources. In addition, peer-to-peer systems must manage a network of highly heterogeneous nodes (hardware architecture: processor, memory and connection type: network interface, connection rate, ...).
### Different centralization levels
P2P systems can be classified into three major families according to the level of centralization that results in a more or less important distribution of tasks to be accomplished. These are centralized, distributed and decentralized (hybrid) P2P models (Fig. \[p2p\_systems\_classification\]).
![\[p2p\_systems\_classification\] P2P systems classification](images/p2p_systems_classification.png){width=".52\textwidth"}
The *centralized model* is at the boder of the P2P model, because it is based on a dedicated server that centralizes and maintains the body of knowledge of peers, the resources always being hosted on peers. In this model, the server is a location entity and contains no resources. Only the retrieval of resources is decentralized. The advantage of this model is that it allows a comprehensive search, fast resource localization and simple management thanks to the central entity. The major disadvantage is that the system operation is based solely on the central entity. The latter must be able to support a large number of queries and perform a lot of research. A well-known example falling into this category is that of the file-sharing system *MP3 Napster* [@DBLP:journals/mms/SaroiuGG03].
*Distributed P2P systems* try to distribute all the system functions among the peers, namely: searching, routing and retrieving objects in the logical network. Peers need to organize to form an effective dynamic architecture connecting them to each other. There are two approaches to this: build a random graph and rely on random explorations/searches of the network, or build a structured graph in which an efficient routing is gradually supported [@viennot2005autour]. A well-known example falling into this category is that of *Gnutella v0.4* [@ripeanu2001peer]. Gnutella in its first version v0.4, proposes to connect the nodes in a random way and uses the flood for its searches in the logical network thus formed. For a better use of resources, the following version, *Gnutella v0.6*, proposes to organize the logical network according to a decentralized architecture on two levels.
The *decentralized model*, also called hybrid model, adds an hierarchy to the centralized model. This model relies on super node interconnections at the top of the hierarchy, according to the distributed model. Each leaf node is attached to one or more super nodes. A super node manages a set of leaf nodes. Objects shared by a leaf node are saved to the super node responsible for that node. When a leaf node looks for an object, it sends its request to its super node. This one then carries out the search among the objects of the nodes which are attached to it, or even the neighboring super nodes if necessary. A well-known example falling into this category is that of *FastTrack* [@liang2006fasttrack]: 1 super node for 100 leaf nodes.
Existing peer-to-peer systems rely on Internet for data sharing. Our goal is to design a peer-to-peer system able to circumvent the need for an active Internet connexion by relying solely on device wifi network interfaces.
The Pear2Pear On Wifi Protocol
==============================
The protocol is designed according to the decentralized or hybrid model.
![\[Pear2Pear\_architectural\_view\] Architectural View of Pear2Pear On Wifi Subnetworks](images/Pear2Pear_architectural_view.pdf){width="100.00000%"}
![\[pear2pear\_protocol\] Building Blocks of a Pear2Pear On Wifi Subnetwork](images/pear2pear_protocol.pdf){width=".7\textwidth"}
A proof of concept was made on Android (Fig. \[apercu\]).
Problems / Contributions
------------------------
Several problems were identified as soon as this project started:
- Establishing and maintaining links between nodes
- Adding and removing a node from the network (does a node exit the network or is it removed ?)
- Routing (how to choose the best route for a given packet ?)
- Node addressing
- Management of available resources (battery, memory,...)
- File sharing protocol
- How to identify a file on the network ?
- Uniqueness, completeness, error handling for a given file
- Managing duplicate files (if a peer wants to download a file it already has)
- Storing files (file block, complete file)
- Parallelism management in file sharing
- Management of identical files but with different names
Algorithms
----------
To solve the identified problems, several algorithms have been set up. This section presents some among the most relevant ones.
### Subnet Creation
A wifi-enabled device wants to connect:
1. It scans surrounding wifi networks for available networks.
- If it identifies a network, it connects to it; otherwise, it creates a network (hotspot) and becomes its *root*. The ssid of each network is defined so as to make it unique and quickly identifiable by other peers. Each network is protected. To access it, a complex computation function allows each device equipped with *Pear2Pear* to compute the passphrase from the ssid.
- If it identifies a network and cannot connect to it after a while, it creates a network (hotspot) and becomes its *root*. Indeed, some operating systems, including Android, limit the number of devices likely to join the hotspot. This number can also be limited by the protocol in order to optimize the use of ressources of the network root (battery, memory, ...).
Regarding the management of peer departures, if a peer notifies its root of its departure, it is removed from catalogs after the expiration of a certain countdown. Moreover, the root must make pings to scan for available peers and deals with silent leaves.
### The Network File Catalog
Each subnet is associated with a file catalog. The catalog identifies for each file the proprietary device and its subnet. Each file is represented by its original name and an unique identifier, related to its content, obtained through the computation of the hash of this content. This makes it possible to manage the duplication of files and to optimize the acquisition of data.
1. The root device initializes the catalog of its subnet, when it is created, with the list of its shared files.
2. Each peer that connects to it, communicates the list of its available files.
Each non-root device in a subnet periodically scans other accessible subnets and informs its root to allow it to build a subnet catalog. The subnet catalog identifies other accessible subnets and for each, the list of peers that can access it.
### File Search and File Management
**Note:** when a file search fails, a special frame must (or can) be emitted (once or periodically) in order to invite the peers to make this file available as far as possible.
To allow inter-network searches, each root must periodically designate, for each neighboring subnet, a peer that will switch to that subnet and retrieve the file catalog to bring it back to its original subnet and make it available to the root. This will enrich the network file catalog of the subnet which will support all file requests. A subnet that is no longer accessible is automatically removed from the catalog after the expiration of a certain countdown. Each time a file is added or deleted, a notification is sent to the root to update the catalog.
The choice of the peer which will retrieve the catalog of another subnet, among all peers having access to this subnet, must be done by an algorithm ensuring optimal use of peer resources. We choosed *round robin* for our proof of concept.
A catalog makes it possible to reference for each subnet whose file is referenced the number of jumps required to reach it and the nearest network that serves as gateway to it. This supports the computation of the shortest path to retrieve a file (also taking into account file duplicates that enable distribute file retrieval).
### File Retrieval
A peer that searches for an available file gets the unique identifier of the file.
#### A Peer to a Peer of the Same Subnet
: A peer *P1* of the subnet *x* wants to retrieve the file *f* present in the shared folder of peer *P2* of the subnet *x*:
1. *P1* sends a request to download the file *f* specifying the unique identifier of *f* in the network file catalog.
2. The root of subnet *x* identifies the peer *P2* as the closest (same subnet) peer that has a copy of the requested file and sends to *P1* the network address of *P2*.
3. *P1* proceeds directly to download the file from *P2*.
If multiple peers of *x* have a copy of the file, the root makes all their addresses available to *P1* so that it can proceeds with a distributed download of the blocks of *f*.
#### A Peer to a Peer of a Different Subnet
: The peer *P1* of the subnet *S1* wants to retrieve the file *f* of a peer of the subnet *Sn*, where *n* is the number of hops separating the 2 subnets:
1. *P1* sends a request to download the file *f* specifying the unique identifier of *f* in the network file catalog.
2. The root of *S1* identifies the closest subnet that owns the file, based on available file copies and number of hops between networks, as *Sn*. It also identifies the peer *P1’* of *S1* as the most likely to go to *S2*, the gateway from *S1* to *Sn*. Therefore, it sends *P1’* to *S2* with the file request.
3. *P1’*, once in *S2*, checks if $S2 == Sn$.
4. If $S2 == Sn$, it sends the file request to the root of *S2* to receive the network address of the peer *P2* of *S2* hosting the file.
5. *P1’* proceeds directly to download the file from *P2*.
6. If $S2 /= Sn$, the protocol resumes in step 1 with $P1=P1'$ and $S1 = S2$.
7. *P1’* jumps back to *S1* and directly delivers the file to *P1* since it has its network address.
If multiple peers of *S2* have a copy of the file, the root makes all their addresses available to *P1’* so that it can proceeds with a distributed download of the blocks of *f*. In addition, if the file is very large, several peers of *S1* having access to *S2* can jump to *S2* to request specific blocks of *f* and return to distributively deliver it to *P1*.
References {#references .unnumbered}
==========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the construction of the minimum cost spanning geometric graph of a given rooted point set $P$ where each point of $P$ is connected to the root by a path that satisfies a given property. We focus on two properties, namely the monotonicity w.r.t. a single direction (*$y$-monotonicity*) and the monotonicity w.r.t. a single pair of orthogonal directions (*$xy$-monotonicity*). We propose algorithms that compute the rooted $y$-monotone ($xy$-monotone) minimum spanning tree of $P$ in $O(|P|\log^2 |P|)$ (resp. $O(|P|\log^3 |P|)$) time when the direction (resp. pair of orthogonal directions) of monotonicity is given, and in $O(|P|^2\log|P|)$ time when the optimum direction (resp. pair of orthogonal directions) has to be determined. We also give simple algorithms which, given a rooted connected geometric graph, decide if the root is connected to every other vertex by paths that are all monotone w.r.t. the same direction (pair of orthogonal directions).'
author:
- Konstantinos Mastakas
- Antonios Symvonis
bibliography:
- 'ic-sa.bib'
- 'templib.bib'
- 'monotone.bib'
- 'upward.bib'
- 'pointLoc.bib'
- 'NN.bib'
title: Rooted Uniform Monotone Minimum Spanning Trees
---
Introduction
============
A geometric path $W = (w_0$, $w_1$, …, $w_l)$ is *monotone in the direction of* $y$, also called $y-$*monotone*, if it is $y$-*decreasing*, i.e. $y(w_0) \geq y(w_1) \geq$ …$\geq y(w_l)$ or if it is $y-$*increasing*, i.e. $y(w_0) \leq y(w_1) \leq$ …$\leq y(w_l)$, where $y(p)$ denotes the $y$ coordinate of a point $p$. $W$ is *monotone* if there exists an axis $y'$ s.t. $W$ is $y'-$monotone. Arkin et al. [@ArkCM89] proposed a polynomial time algorithm which connects two given points by a geometric path that is monotone in a given (an arbitrary) direction and does not cross a set of obstacles, if such a path exists. Furthermore, the problem of drawing a directed graph as an *upward graph*, i.e. a directed geometric graph such that each directed path is $y-$increasing, has been studied in the field of graph drawing, e.g. see [@BatT88; @GarT01].
A geometric graph $G = (P,E)$ is *monotone* if every pair of points of $P$ is connected by a monotone geometric path, where the direction of monotonicity does not need to be the same for each pair. If there exists a single direction of monotonicity, we denote the graph as *uniform monotone*. Uniform monotone graphs were also denoted as *$1-$monotone graphs* by Angelini [@Ang15]. When the direction of monotonicity is known, say $y$, the graph is called *$y-$monotone*. Monotone graphs were introduced by Angelini et al. [@AngCBFP12]. The problem of drawing a graph as a monotone graph has been studied in the field of graph drawing; e.g. see [@Ang15; @AngCBFP12; @AngDKMRSW15; @HeH16]. The reverse problem, namely, given a point set $P$ we are asked to construct a monotone spanning geometric graph on the points of $P$, has trivial solutions, i.e. the complete graph $K_{|P|}$ on the points of $P$ as well as the path graph $W_{|P|}$ which visits all points of $P$ in increasing order of their $y$ coordinates are both $y-$monotone spanning geometric graphs of $P$.
The *Euclidean minimum spanning tree problem*, i.e. the problem of constructing the minimum cost spanning geometric tree of a plane point set $P$ (where the cost of the tree is taken to be the sum of the Euclidean lengths of its edges), has also received attention [@preparata1988computational]. Shamos and Hoey [@ShaH75] showed that it can be solved in $\Theta(|P|\log|P|)$ time.
Combining the Euclidean minimum spanning tree problem with the notion of monotonicity leads to a large number of problems that, to the best of our knowledge, have not been previously investigated. The most general problem can be stated as follows: *“Given a point set $P$ find the minimum cost monotone spanning geometric graph of $P$, i.e. the geometric graph such that every pair of points of $P$ is connected by a monotone path”.* Since in a monotone graph the direction of monotonicity need not be the same for all pairs of vertices, it is not clear whether the minimum cost monotone spanning graph is a tree. We call this problem the *Monotone Minimum Spanning Graph problem.* We note that there exist point sets for which the Euclidean minimum spanning tree is not monotone and hence does not coincide with the monotone minimum spanning graph. Consider for example the point set with four points depicted in Figure \[fig:EMSTvsMonotone\] for which the Euclidean minimum spanning tree is a geometric path that is not monotone.
![The Euclidean minimum spanning tree which is depicted is not monotone[]{data-label="fig:EMSTvsMonotone"}](EMSTvsM.pdf)
We focus on a simple variant of the general monotone minimum spanning graph problem. Let $P$ be a rooted point set, i.e. a point set having a designated point, say $r$, as its root. We do not insist on having monotone paths between every pair of points of $P$ but rather only between the root $r$ with all other points of $P$. Moreover, we insist that all paths are *uniform* in the sense that they are all monotone with respect to the same direction, i.e. we build *rooted uniform monotone graphs*. Actually, as it turns out (Corollary \[cor:y-tree\]), in this problem the sought graphs are trees and, thus, we refer to it as the *rooted Uniform Monotone Minimum Spanning Tree* (for short, *rooted UMMST*) *problem*. In the rooted UMMST problem we have the freedom to select the direction of monotonicity. When we are restricted to have monotone paths in a specific direction, say $y$, we have the *rooted $y$-Monotone Minimum Spanning Tree* (for short, *rooted $y$-MMST*) *problem*. Figure \[fig:illYrSpanning\](a) illustrates a rooted $y-$monotone spanning graph of a rooted point set $P$, while the rooted $y-$MMST of $P$ is given in Figure \[fig:illYrSpanning\](b).
![Illustration of rooted $y-$monotone spanning graphs []{data-label="fig:illYrSpanning"}](rootedMonotoneEx.pdf "fig:"){width="45.00000%"}\
(a)
![Illustration of rooted $y-$monotone spanning graphs []{data-label="fig:illYrSpanning"}](rootedMonotoneMinimum.pdf "fig:"){width="45.00000%"}\
(b)
Rooted point sets have been previously studied in the context of minimum spanning trees. The *capacitated minimum spanning tree* is a tree that has a designated vertex $r$ (its root) and each of the subtrees attached to $r$ contains no more than $c$ vertices. $c$ is called the *tree capacity*. Solving the capacitated minimum spanning tree problem optimally has been shown by Jothi and Raghavachari to be NP-hard [@Jothi:2005:AAC:1103963.1103967]. In the same paper, they have also presented approximation algorithms for the case where the vertices correspond to points on the Euclidean plane.
If the geometric path $W$ is both $x-$monotone and $y-$monotone then it is denoted as $xy-$*monotone*. Furthermore, if there exists a Cartesian System $x'y'$ s.t. $W$ is $x'y'-$monotone then $W$ is *2D-monotone*. Based on $xy-$monotone geometric paths and in analogy to the (rooted) monotone, uniform monotone and $y-$monotone graphs, we define the (*rooted*) *2D-monotone*, *uniform 2D-monotone* and *$xy-$monotone graphs*. 2D-monotone paths/graphs were also recently denoted by Bonichon et al. [@BonBCKLV16] as *angle-monotone paths/graphs*. Bonichon et al. [@BonBCKLV16] gave a $O(|P|\cdot|E|^2)$ time algorithm that decides if a geometric graph $G = (P,E)$ is 2D-monotone. In order to do so, Bonichon et al. [@BonBCKLV16] gave a $O(|E|^2)$ time algorithm which is used as a subroutine and decides if the graph is rooted 2D-monotone, where the root is a specified vertex. Bonichon et al. [@BonBCKLV16] also noted that it is not always feasible to construct a planar 2D-monotone spanning geometric graph of a given point set. Similarly to the rooted UMMST and rooted $y$-MMST problems we define the corresponding *rooted Uniform 2D-Monotone Minimum Spanning Tree* (for short, *rooted 2D-UMMST*) and *rooted $xy$-Monotone Minimum Spanning Tree* (for short, *rooted $xy$-MMST*) problems, which ask for the minimum cost rooted Uniform 2D-Monotone spanning tree and rooted $xy$-monotone spanning tree of a given rooted point set, respectively.
A path/curve $W$ is *increasing-chord* (see [@LarM72; @Rot94]) if for every four points $p_1, p_2, p_3, p_4$ traversed in this order along it, it holds that $d(p_2, p_3) \leq d(p_1, p_4)$ where $d(p,q)$ denotes the Euclidean distance between the points $p$ and $q$. A geometric graph $G = (P,E)$ is *increasing-chord* if each two points of $P$ are connected by an increasing-chord path. Increasing-chord graphs were introduced by Alamadari et al. [@AlaCGLP13]. Alamdari et al. [@AlaCGLP13] noted that any 2D-monotone path/graph is also increasing-chord. Drawing a graph as an increasing-chord graph is studied in [@AlaCGLP13; @NolPR16]. On the other hand, constructing increasing-chord graphs that span a given point set is studied in [@AlaCGLP13; @DehFG15; @MasS15]. In all the papers that construct increasing-chord spanning graphs of a point set, i.e. in [@AlaCGLP13; @DehFG15; @MasS15], the constructed increasing-chord paths connecting the vertices are additionally 2D-monotone.
**Our contribution:**
Let $P$ be a rooted point set. We give algorithms that produce the rooted $y-$MMST of $P$ and the rooted $xy-$MMST of $P$ in $O(|P|\log^2 |P|)$ time and $O(|P|\log^3 |P|)$ time, respectively. We also propose algorithms that build the rooted UMMST of $P$ and the rooted 2D-UMMST of $P$ in $O(|P|^2 \log |P|)$ time when the optimum direction and the optimum pair of directions has to be determined, respectively. For all these four problems, we provide a $\Omega(|P| \log |P|)$ time lower bound which is easily derived.
We also propose simple algorithms that decide whether a given connected geometric graph on a rooted point set is (i) rooted $y-$monotone, (ii) rooted uniform monotone, (iii) rooted $xy-$monotone and (iv) rooted uniform 2D-monotone.
Definitions and Preliminaries
=============================
In this article we deal with the Euclidean plane, i.e. every point set is a subset of $\mathbb{R}^2$, and we consider only rooted point sets. Let $x,y$ be the axes of a Cartesian System. The $x$ and $y$ coordinates of a point $p$ are denoted by $x(p)$ and $y(p)$, respectively. W.l.o.g., we assume that the root $r$ of the point sets coincides with the origin of the Cartesian System, i.e., $x(r) = y(r) = 0$. We also assume that the point sets are *in general position*, i.e. no three points are collinear.
Let $P$ be a point set. $P$ is called *positive* (*negative*) w.r.t. the direction of $y$ or $y-$positive ($y-$negative) if for each $p \in P$, $y(p) \geq 0$ (resp. $y(p) \leq 0$). Let $a$ be a real number. By $P_{y \leq a}$ we denote the set of points of $P$ that have $y$ coordinate less than or equal to $a$. Subsets $P_{y \geq a}$, $P_{x \leq a}$, $P_{x \geq a}$, $P_{|y| \leq a}$, $P_{|y| \geq a}$, $P_{|x| \leq a}$ and $P_{|x| \geq a}$ are similarly defined. $P_{x \leq a,y \leq a}$ denotes the set $P_{x \leq a}\cap P_{y \leq a}$. Subsets $P_{x \geq a,y \geq a}$, $P_{x \leq a, y \geq a}$ and $P_{x \geq a, y \leq a}$ are defined similarly. The Euclidean plane is divided into four quadrants, i.e. the quadrants $\mathbb{R}^2_{x \geq 0, y \geq 0}$, $\mathbb{R}^2_{x \leq 0, y \geq 0}$, $\mathbb{R}^2_{x \leq 0, y \leq 0}$ and $\mathbb{R}^2_{x \geq 0, y \leq 0}$.
Let $P$ be a point set and $p$ be a point of the plane, then $d(p,P)$ denotes the Euclidean distance from $p$ to the point set $P$, i.e. $d(p,P) = \min\limits_{q \in P} d(p,q)$.
The line segment with endpoints $p$ and $q$ is denoted as $\overline{pq}$. The *slope* of a line $L$ is the angle that we need to rotate the $x$ axis counterclockwise s.t. the $x$ axis becomes parallel to $L$. Each slope belongs to the range $[0,\pi)$.
A *geometric graph* $G = (P,E)$ consists of a set $P$ of points which are denoted as its vertices and a set $E$ of line segments with endpoints in $P$ which are denoted as its edges. If $P$ is rooted then $G$ is a *rooted geometric graph*. The cost of a geometric graph $G = (P,E)$, denoted as $\text{cost}(G)$, is the sum of the Euclidean lengths of its edges. Let $G_1 = (P_1, E_1), G_2 = (P_2, E_2), \ldots, G_n = (P_n ,E_n)$ be $n$ geometric graphs then the union $G_1\cup G_2\cup \ldots \cup G_n$ is the geometric graph $G = (P, E)$ s.t. $P = P_1 \cup P_2 \cup \ldots \cup P_n$ and $E = E_1 \cup E_2 \cup \ldots \cup E_n$.
The *closest point* (or *nearest neighbor*) *problem* is an important problem in computational geometry. It was initially termed as the *post-office problem* by Knuth [@Knu73]. In this problem there exists a set $S$ of points that is static (it cannot be changed by inserting points to it or deleting points from it) and the goal is to find the closest point (or nearest neighbor) from $S$ to a given query point. This problem is usually reduced to the problem of locating in which region of a planar subdivision the query point is located [@Sha75; @DobL76]. Efficient static data structures have been constructed to answer these queries in logarithmic time by performing fast preprocessing algorithms, e.g. see [@LipT80; @Kir83]. Concerning the semi-dynamic version of the closest point problem, in which insertions of points to $S$ are allowed, Bentley [@Ben79] gave a very useful semi-dynamic data structure.
\[fact:BentleyStructure\] There exists a semi-dynamic data structure that allows only two operations, the insertion of a point and a closest point query. Where, a closest point query takes $O(\log ^2 n)$ time (with $n$ denoting the size of the structure) and inserting $n$ elements in the structure takes $O(n \log ^2 n)$ total time.
A variant of the closest point problem that was studied in the fourth section of the article of Bentley [@Ben79], comes in handy when we study the rooted $xy-$MMST problem. More specifically, in this variant, each point in $S$ is associated with a numerical attribute value. Given a query point $q$, the goal is to find the closest point to $q$ from the subset of points of $S$ for which their attribute value belongs to a specified range. For the static version, i.e. when $S$ cannot be mutated, Bentley [@Ben79] gave a static data structure that adds a multiplicative logarithmic factor in the query and preprocessing time of the static data structure for the closest point problem. Concerning the semi-dynamic version of this problem, Bentley [@Ben79] implicitly produced, i.e from his corresponding static data structure (section 4 of [@Ben79]) and his results about the decomposable problems (section 3 of [@Ben79]), a very useful semi-dynamic data structure.
\[fact:rangeClosest\] There exists a semi-dynamic data structure that supports the operations (i) insertion of a point, and (ii) the closest point to a query point from the subset of points in the data structure whose attribute value is within a given range. Where, a closest point query takes $O(\log ^3 n)$ time (with $n$ denoting the number of points in the data structure) and inserting $n$ points in the data structure takes $O(n \log ^3 n)$ total time.
The Rooted $y$-Monotone Minimum Spanning Tree (rooted $y-$MMST) Problem {#sec:y-uMMST}
=======================================================================
In this section we study the construction of the rooted $y-$MMST of a rooted point set $P$. We initially show that we can deal with $P_{y \leq 0}$ and $P_{y\geq 0}$ separately. Then, we provide a characterization of the rooted $y-$MMST of rooted $y-$positive (or $y-$negative) point sets. Using the previous two, we develop an algorithm that constructs the rooted $y-$MMST of $P$. We also provide a lower bound for the time complexity of any algorithm that solves the rooted $y-$MMST problem as well as a simple recognition algorithm for rooted $y-$monotone graphs.
We recall that we assume that the root $r$ of a rooted point set is located at the point $(0,0)$.
\[obs:unnesEdge\] Let $P$ be a rooted point set and $G = (P,E)$ be a rooted $y-$monotone spanning graph of $P$ and let $\overline{p_dp_u} \in E$ with $y(p_d) < 0 < y(p_u)$. Then, every path from the root $r$ to a point $p \in P\setminus\{r\}$ that contains $\overline{p_dp_u}$ is not $y-$monotone since it moves “south” to $p_d$ and then “north” to $p_u$, or vice versa.
\[cor:unEdges\] Let $P$ be a rooted point set, $G^{\text{opt}} = (P,E)$ be the rooted $y-$monotone minimum spanning graph of $P$ and $p_d,p_u \in P$ such that $y(p_d) < 0 < y(p_u)$. Then, $\overline{p_dp_u} \notin E$.
Corollary \[cor:unEdges\] implies that the root $r$ of a rooted point set $P$ splits the problem of finding the rooted $y-$monotone minimum spanning graph of $P$ into two independent problems. Namely, producing the rooted $y-$monotone minimum spanning graph of (i) $P_{y\leq 0}$, and (ii) $P_{y\geq 0}$. Hence, we obtain the following Lemma.
\[lem:splitUpDown\] Let $P$ be a rooted point set and $G^{\text{opt}}$ be the rooted $y-$monotone minimum spanning graph of $P$. Furthermore, let $G_{y \leq 0}^{\text{opt}}$ and $G_{y \geq 0}^{\text{opt}}$ be the rooted $y-$monotone minimum spanning graphs of $P_{y \leq 0}$ and $P_{y \geq 0}$, respectively. Then, $G^{\text{opt}}$ is the union of $G_{y \leq 0}^{\text{opt}}$ and $G_{y \geq 0}^{\text{opt}}$.
We now study the construction of the rooted $y-$monotone minimum spanning graph of a rooted $y-$positive (or $y-$negative) point set $P$ with root $r$. The main idea is to connect each $p \in P \setminus\{r\}$ to its closest point with absolute $y$ coordinate less than the absolute $y$ coordinate of $p$. However, since we have assumed that there might be (at most) two points with the same $y$ coordinate the analysis becomes a little more complicated. We define $S[P,y]$ to be the sequence of points of $P$ ordered by the following rule: *“The points of $S[P,y]$ are ordered w.r.t. their absolute $y$ coordinates and, if two points have the same $y$ coordinate, then they are ordered w.r.t. their distance from the preceding points in $S[P,y]$.”*. More formally, $S[P,y] = (r = p_0,p_1,p_2, \ldots, p_n)$ s.t. $|y(p_0)| \leq |y(p_1)| \leq |y(p_2)| \leq \ldots \leq |y(p_n)|$ and $|y(p_i)| = |y(p_{i+1})|$ implies that $d(p_i, \{p_0, p_1, \ldots, p_{i-1}\}) \leq d(p_{i+1}, \{p_0, p_1, \ldots, p_{i-1}\})$ and $P = \{p_0,p_1,p_2, \ldots, p_n\}$. We now give a characterization of the rooted $y-$monotone minimum spanning graph of $P$.
\[lem:yMonotoneChar\] Let $G=(P,E)$ be a rooted geometric graph where $P$ is a rooted $y-$positive (or $y-$negative) point set with $S[P,y]$ $= (r$ $= p_0,p_1,p_2$ , …, $p_n)$. Then, $G$ is the rooted $y-$monotone minimum spanning graph of $P$ if and only if (i) $p_n$ is connected in $G$ only with its closest point (or nearest neighbor) from $\{p_0$, $p_1$, …, $p_{n-1}\}$, i.e. the point $p_j$ such that $d(p_n,p_j) = d(p_n$, $\{p_0$, $p_1$, …, $p_{n-1}\})$, and (ii) $G\setminus\{p_n\}$ is the rooted $y-$monotone minimum spanning graph of $P\setminus\{p_n\}$.
The ($\Rightarrow$) direction can be easily proved by contradiction. We now prove the ($\Leftarrow$) direction. $G$ is a rooted $y-$monotone graph that spans $P$ since $G\setminus\{p_n\}$ is a rooted $y-$monotone graph that spans $P\setminus\{p_n\}$ and $p_n$ is connected with another point in the graph. Let $G^{\text{opt}}$ be the rooted $y-$monotone minimum spanning graph of $P$. Then, $p_n$ is connected in $G^{\text{opt}}$ with some other point in $P$, and thus $G^{\text{opt}}$ contains an edge of cost at least $d(p_n, \{p_0, p_1, \ldots, p_{n-1}\})$. Furthermore, the graph $G^{\text{opt}} \setminus\{p_n\}$ is also rooted $y-$monotone, hence its cost is at least the cost of the rooted $y-$monotone minimum spanning graph of $P\setminus\{p_n\}$. Thus, $G^{\text{opt}}$ has at least the same cost as $G$.
Lemma \[lem:splitUpDown\] and Lemma \[lem:yMonotoneChar\] lead to the next Corollary.
\[cor:y-tree\] The rooted $y-$monotone minimum spanning graph of a rooted point set $P$ is a geometric tree.
Let $P$ be a rooted $y-$positive (or $y-$negative) point set and $S[P,y] =$ $(r =$ $p_0$, $p_1$, …, $p_n)$. We call the closest point to $p_i$ from $\{p_0, p_1, \ldots, p_{i-1}\}$ the *parent* of $p_i$ and we denote it as *par$(p_i)$*. More formally, par$(p_i) = p_j$ if and only if $p_j \in \{p_0, p_1, \ldots, p_{i-1}\}$ and $d(p_i, p_j) = d(p_i, \{p_0, p_1, \ldots, p_{i-1}\})$. Then, Lemma \[lem:yMonotoneChar\] implies the following Corollary.
\[cor:parentCharact\] The edges of the rooted $y-$MMST of $P$ are exactly the line segments $\overline{\text{par}(p_i)p_i},$ for $i =1, 2, \ldots, n$.
Corollary \[cor:parentCharact\] implies a $O(|P|^2)$ time algorithm for producing the rooted $y-$MMST of $P$. However, using the semi-dynamic data structure for closest point queries given by Bentley [@Ben79], the time complexity of our rooted $y-$MMST algorithm becomes $O(|P|\log^2|P|)$. Our rooted $y-$MMST algorithm is described in Algorithm \[alg:GIyMonotone\].
**Input:** A rooted $y-$positive (or $y-$negative) point set $P$ with root $r$.
**Output:** The rooted $y$-Monotone Minimum Spanning Tree of $P$.
$T \gets$ the geometric graph with $P$ as its vertex set and $\emptyset$ as its edge set. Sort the points of $P$ w.r.t. their absolute $y$ coordinates, constructing the sequence $S$. If some points tie, use an arbitrary order (their relative order may be changed later on, during the execution of the algorithm, since $S$ is used in order to calculate the $S[P,y]$, i.e. at the end of the algorithm $S$ equals to $S[P,y]$). More formally, $S$ = $(r$ = $S[0]$, $S[1]$, …, $S[n])$ with $|y(S[i])| \leq |y(S[i+1])|$. ProximityDS is an (initially empty) semi-dynamic data structure for closest point queries. Insert $r$ in ProximityDS. $i \gets 1$ par$(S[i]) \gets $ closest point to $S[i]$ from ProximityDS, obtained by performing a closest point query in ProximityDS. insert $S[i]$ in ProximityDS and insert the edge $\overline{\text{par}(S[i])S[i]}$ in $T$. $i \gets i + 1$ par$(S[i+1]) \gets $ closest point to $S[i+1]$ from ProximityDS, obtained by performing a closest point query in ProximityDS. Swap the points with indices $i$ and $i+1$ in $S$ par$(S[i+1]) \gets S[i]$. insert $S[i]$ and $S[i+1]$ in ProximityDS Insert the edges $\overline{\text{par}(S[i])S[i]}$ and $\overline{\text{par}(S[i+1])S[i+1]}$ in $T$. $i \gets i + 2$ **return** $T$.
\[thm:yRootedMonotoneMain\] The rooted $y-$MMST of a rooted point set $P$ can be computed in $O(|P|\log^2 |P|)$ time.
We first construct $P_{y\leq 0}$ and $P_{y\geq 0}$. Then, we apply our rooted $y-$MMST algorithm (Algorithm \[alg:GIyMonotone\]) on $P_{y\leq 0}$ and $P_{y\geq 0}$ constructing $T_{y \leq 0}$ and $T_{y \geq 0}$, respectively. By Corollary \[cor:parentCharact\], $T_{y \leq 0}$ and $T_{y \geq 0}$ are the rooted $y-$MMSTs of $P_{y\leq 0}$ and $P_{y\geq 0}$, respectively. Using Fact \[fact:BentleyStructure\] and since $O(|P|)$ insertions and $O(|P|)$ closest point queries are performed, computing $T_{y \leq 0}$ and $T_{y \geq 0}$ takes $O(|P|\log^2 |P|)$ time. By Lemma \[lem:splitUpDown\], $T_{y \leq 0} \cup T_{y \geq 0}$ is the rooted $y-$MMST of $P$.
In the next Theorem, we give a lower bound for the time complexity of any algorithm which given a rooted point set $P$ produces the rooted $y-$MMST of $P$.
\[thm:lowerBoundyMonotone\] Any algorithm which given a rooted point set $P$, produces the rooted $y-$MMST of $P$ requires $\Omega (|P|\log |P|)$ time.
We use the reduction from sorting that was given by Shamos [@Sha78]. Let $(a_1$, $a_2$, …, $a_n)$ be a sequence of nonnegative integers. We reduce this sequence to the rooted point set $P = \{r= (0,0)$, $(a_1, a_1^2)$, $(a_2, a_2^2)$, …, $(a_n, a_n^2)\}$. Then, the rooted $y-$MMST of $P$ contains exactly the edges $\overline{rp_1}$, $\overline{p_1p_2}$, …,$\overline{p_{n-1}p_n}$ s.t. $a_i' = x(p_i), i = 1,2, \ldots, n$, where $(a_1', a_2', \ldots, a_n')$ is the sorted permutation of $(a_1, a_2, \dots, a_n)$. The lower bound follows since sorting $n$ numbers requires $\Omega (n\log n)$ time.
We note that using the same reduction, i.e. the reduction from sorting that was given by Shamos [@Sha78], the same lower bound can be easily obtained for the rooted UMMST and (rooted) Monotone Minimum Spanning Graph Problem.
We conclude this section, by showing that rooted $y-$monotone graphs can be efficiently recognized. Our approach is similar to the approach employed in the third section of the article of Arkin et al. [@ArkCM89].
\[thm:recDirected\] Let $G = (P,E)$ be a rooted connected geometric graph. Then, we can decide in $O(|E|)$ time if $G$ is rooted $y-$monotone.
We first transform $G$ into a directed geometric graph $\overrightarrow{G}$ in $O(|E|)$ time, by assigning direction to the edges and removing some of them. Let $\overline{pq}$ be an edge of $G$. If $p$ and $q$ belong to opposite half planes w.r.t. the $x$ axis then $\overline{pq}$ cannot be used in a $y-$monotone path from the root $r$ to a point of $P\setminus \{r\}$ (see Observation \[obs:unnesEdge\]). Hence, we remove the edge $\overline{pq}$ from the graph. If $y(p) = y(q)$ then we insert both $\overrightarrow{pq}$ and $\overrightarrow{qp}$ in $\overrightarrow{G}$. Otherwise, assuming w.l.o.g., that $|y(p)| < |y(q)|$, we insert $\overrightarrow{pq}$ in $\overrightarrow{G}$. $G$ is rooted $y-$monotone if and only if $r$ is connected with all other points of $P$ in $\overrightarrow{G}$. The latter can be easily decided in $O(|E|)$ time by a breadth first search or a depth first search traversal.
Rooted Uniform Monotone Graphs: Minimum Spanning Tree Construction, and Recognition {#sec:generalRootMon}
===================================================================================
In this section we study the rooted uniform monotone graphs. We initially study the computation of the rooted UMMST of a rooted point set (Subsection \[subsec:UMMST\]) and then we deal with the recognition of rooted uniform monotone graphs (Subsection \[subsec:recUM\]).
Building the Rooted UMMST {#subsec:UMMST}
-------------------------
In this subsection, we focus on the rooted UMMST problem. In contrast with the rooted $y-$MMST problem where the direction $y$ of monotonicity was given, here we are asked to determine the optimum direction of monotonicity, say $y'$, and the corresponding rooted $y'-$MMST. We tackle the problem by giving a rotational sweep algorithm. Rotational sweep is a well known technique in computational geometry in which a (directed) line is rotated counterclockwise (or clockwise) and during this rotation important information about the solution of the problem is updated. We note that crucial to the development of our algorithm for the rooted UMMST problem is the fact that it is sufficient to take into account only a quadratic number of directions of monotonicity (to be proved in Lemma \[lem:MonotoneNecessaryDirections\]). This fact is based on the following observations.
\[obs:changeInRotation\] Let $y'$ be an axis. If we rotate the $y'$ axis counterclockwise then the sequence $S[P_{y' \geq 0}, y']$ or the sequence $S[P_{y' \leq 0}, y']$ changes only when the $y'$ axis reaches (moves away from) a line perpendicular to a line passing through two points of $P$. Then, by Lemma \[lem:yMonotoneChar\] and Corollary \[cor:parentCharact\], the rooted $y'-$MMST of $P$ may only change at the same time.
\[obs:slopeLessThanPi\] Let $y'$ and $y''$ be axes of opposite directions. Then, the rooted $y'-$MMST of $P$ is the same as the rooted $y''-$MMST of $P$. Hence, when computing the rooted UMMST of $P$ we only need to take into account the $y'$ axes such that the angle that we need to rotate the $x$ axis counterclockwise to become codirected with the $y'$ is less than $\pi$.
Based on Observations \[obs:changeInRotation\] and \[obs:slopeLessThanPi\], we define the set $\Theta = \{\theta \in [0,\pi): \theta$ is the slope of a line perpendicular to a line passing through two points of $P\}$. We also define $S[\Theta]$ to be the sorted sequence that contains the slopes of $\Theta$ in increasing order, i.e. $S[\Theta] = (\theta_0, \theta_1, \ldots, \theta_{m-1})$, $\theta_i < \theta_{i+1}$, $i =0, 1, \ldots, m-2$ and $m \leq \binom{|P|}{2}$. We further define the set $\Theta_{\text{critical}} = \{\theta_0$, $\theta_1$, …, $\theta_{m-1}\} \cup \{$ $\frac{\theta_0 + \theta_1}{2}$, $\frac{\theta_1 + \theta_2}{2}$, …, $\frac{\theta_{m-2} + \theta_{m-1}}{2}$, $\frac{\theta_{m-1} + \pi}{2}\}$ which we call the *critical set of slopes* since, as we show in Lemma \[lem:MonotoneNecessaryDirections\], examining the axes with slope in $\Theta_{\text{critical}}$ is sufficient for computing the rooted UMMST of $P$. $|\Theta_{\text{critical}}| = O(|P|^2)$. We now assign “names” to the axes with slopes in $\Theta_{\text{critical}}$. Let $y_{2i}$ be the axis with slope $\theta_i$, $i = 0, 1, \ldots, m-1$ and let $y_{2i + 1}$ be the axis with slope $\frac{\theta_i + \theta_{i+1}}{2}$, $i = 0, 1, \ldots, m-2$, and $y_{2m-1}$ be the axis with slope $\frac{\theta_{m-1} + \pi}{2}$. Note that the subscript of each axis gives its order when the axes are sorted w.r.t. their slope. Note also that axes with even subscripts correspond to lines perpendicular to lines passing through two points of $P$.
\[lem:MonotoneNecessaryDirections\] The rooted UMMST of $P$ is one of the rooted $y'-$MMST of $P$ over all axes $y'$ with slope in $\Theta_{\text{critical}}$ and, more specifically, the one of minimum cost.
Let $y'$ and $y''$ be axes of slope $\theta'$ and $\theta''$, respectively, such that $\theta_i < \theta', \theta'' < \theta_{i+1}$ for some $0 \leq i \leq m - 2$ and let $T'$, $T''$ be the rooted $y'-$MMST and the rooted $y''-$MMST of $P$, respectively. By Observation \[obs:changeInRotation\], cost$(T') = $cost$(T'')$. As a result, we need to take into account only one of the $y'$ axes of slope $\theta '$ with $\theta_i < \theta' < \theta_{i+1}$. We take into account the axis $y_{2i+1}$. Additionally, we take into account the axis $y_{2m-1}$ for which the rooted $y_{2m-1}-$MMST of $P$ is the same as the rooted $y'-$MMST of $P$ for any $y'$ axis with slope in the range $(\theta_{m-1},\pi) \cup [0,\theta_0)$.
We also need to take into account all the axes of slope $\theta_{i}$, i.e. the axes $y_{2i}$, $ 0 \leq i \leq m-1$, since, by Observation \[obs:changeInRotation\], the rooted $y_{2i}-$MMST of $P$ might differ from all the other rooted $y'-$MMST of $P$, $y' \neq y_{2i}$ (the sorted sequence $S[P_{y_{2i} \geq 0},y_{2i}]$ (or $S[P_{y_{2i} \leq 0},y_{2i}]$) might be different from every other $S[P_{y' \geq 0},y']$ (resp., $S[P_{y' \leq 0},y']$) ). See, for example, Figure \[fig:consAxes\].
\
(a)
\
(b)
\
(c)
We now describe our algorithm which produces the rooted UMMST of a rooted point set $P$. Our rooted UMMST algorithm is a rotational sweep algorithm. It considers an axis $y'$, which initially coincides with $y_0$, and then it rotates it counterclockwise until $y'$ becomes opposite to the $x$ axis. Throughout this procedure, it updates the rooted $y'-$MMST of $P$. By Lemma \[lem:MonotoneNecessaryDirections\], it only needs to obtain each rooted $y_i-$MMST of $P$, where $y_i$ is an axis with slope in $\Theta_{\text{critical}}$, $0 \leq i \leq 2m-1$. Let $T^{\text{opt}}_i$ be the rooted $y_i-$MMST of $P$, $0 \leq i \leq 2m-1$. Our rooted UMMST algorithm can now be stated as follows: It initially constructs $T^{\text{opt}}_0$ using Theorem \[thm:yRootedMonotoneMain\]. Then, it iterates for $i = 1, 2, \ldots, 2m - 1$ obtaining at the end of each iteration $T^{\text{opt}}_{i}$ by modifying $T^{\text{opt}}_{i-1}$. In order to do this efficiently, it maintains a tree $T$ which is initially equal to $T^{\text{opt}}_0$ and throughout its operation it evolves to $T^{\text{opt}}_{1}$, $T^{\text{opt}}_{2}$, …, $T^{\text{opt}}_{2m-1}$. Similarly, it maintains the sequences $S^-$ and $S^+$ which are initially equal to $S[P_{y_{0} \leq 0}, y_{0}]$ and $S[P_{y_{0} \geq 0}, y_{0}]$, respectively, and evolve to $S[P_{y_{i} \leq 0}, y_{i}]$ and $S[P_{y_{i} \geq 0}, y_{i}]$, respectively, $i =1, 2, \ldots, 2m-1$. Our algorithm stores the axis which corresponds to the produced rooted UMMST of $P$, so far, in the variable “minAxis”. In its final step, it recomputes the rooted [“minAxis”]{}-MMST of $P$ using Theorem \[thm:yRootedMonotoneMain\] and returns this tree. The pseudocode of our rooted UMMST algorithm is presented in Algorithm \[alg:rootedMonotoneAlgo\].
**Input:** A rooted point set $P$.
**Output:** The rooted Uniform Monotone Minimum Spanning Tree of $P$.
Compute the axes $y_0, y_1, \ldots, y_{2m-1}$ with slopes in $\Theta_{\text{critical}}$. Compute $T^{\text{opt}}_0$, $S[P_{y_0 \leq 0}, y_0]$ and $S[P_{y_0 \geq 0}, y_0]$ using Theorem \[thm:yRootedMonotoneMain\]. $T \gets T^{\text{opt}}_0$, $S^- \gets S[P_{y_0 \leq 0}, y_0]$, $S^+ \gets S[P_{y_0 \geq 0}, y_0]$, minCost$\gets$ cost$(T^{\text{opt}}_0)$ and minAxis$\gets y_0$. Update $T, S^-, S^+$ such that $T$ equals to $T^{\text{opt}}_{i}$ and $S^-$ (resp. $S^+$) equals to $S[P_{y_{i} \leq 0}, y_{i}]$ (resp. $S[P_{y_{i} \geq 0}, y_{i}]$). minCost $\gets$ cost$(T)$ and minAxis$\gets y_i$. **return** the minAxis-MMST of $P$, computed using Theorem \[thm:yRootedMonotoneMain\].
\[thm:spanningTreeOneDir\] The rooted UMMST of a rooted point set $P$ can be computed in $O(|P|^2 \log |P|)$ time.
By Lemma \[lem:MonotoneNecessaryDirections\], our rooted UMMST algorithm (Algorithm \[alg:rootedMonotoneAlgo\]) produces the rooted UMMST of $P$. We now show that its time complexity is $O(|P|^2 \log |P|)$. The axes $y_0$, $y_1$, …, $y_{2m-1}$ with slopes in $\Theta_{\text{critical}}$ can be computed in $O(|P|^2 \log |P|)$ time. Let $k_i$ be the number of pairs of points of $P$ that have the same projection onto the $y_{2i}$ axis, $0 \leq i \leq m-1$. Then $\sum\limits_{i =0}^{m-1} k_i = \binom{|P|}{2}$. For each $i = 0$, $1$, …, $m-1$, we compute a list $L_i$ which contains these $k_i$ pairs. All $L_i, 0\leq i \leq m-1$, can be computed in $O(|P|^2 \log |P|)$ total time.
For each point $p$ in $S^-\setminus \{r\}$ (resp. in $S^+\setminus \{r\}$) we maintain a data structure PD$(p)$ which is a self-balancing binary search tree that contains all the points that precede $p$ in $S^-$ (resp. $S^+$) accompanied with their distance from $p$. More formally, let $S^-$ be equal to $(r = p_0, p_1, \ldots, p_s)$. Then, for each $p_j, j = 1, 2, \ldots, s$, PD$(p_j)$ contains the pairs $(p_0, d(p_0, p_j))$, $(p_1, d(p_1, p_j))$, …, $(p_{j-1}, d(p_{j-1}, p_j))$. The key of each $(p_l, d(p_l, p_j)), l = 0, 1, \ldots, j-1$, is the distance $d(p_l, p_j)$. Similarly, we define PD$(p)$ for each $p \in S^+\setminus \{r\}$. We employ these PD$(p)$, $p \in P\setminus \{r\}$, data structures since using the information stored in them we can obtain the parent of each $p$ efficiently. In more detail, for each $p \in P\setminus \{r\}$ the par$(p)$ in $T$ can be obtained or updated in $O(\log |P|)$ time by taking into account the PD$(p)$, since the pair $(\text{par}(p), d(\text{par}(p),p))$ is the element with the minimum key in PD$(p)$.
Computing the initial values of $T, S^-,S^+$ and PD$(p), p \in P\setminus \{r\}$ can be done in $O(|P|^2 \log |P|)$ time. This is true since $T^{\text{opt}}_0$, $S[P_{y_0 \leq 0}, y_0]$ and $S[P_{y_0 \geq 0}, y_0]$ are computed in $O(|P|\cdot \log ^2 |P|)$ time (see Theorem \[thm:yRootedMonotoneMain\]). Furthermore, computing PD$(p_j)$ for some $p_j$ in $S^-\setminus \{r\}$ (resp. in $S^+\setminus \{r\}$), when $S^-$ (resp. $S^+$) equals to $S[P_{y_0 \leq 0}, y_0]$ (resp. $S[P_{y_0 \geq 0}, y_0]$) takes $O(|P| \log |P|)$ time since we have to insert each $(p_i,d(p_i,p_j)$, with $i < j$, to PD$(p_j)$ and each such insertion takes $O(\log |P|)$ time. Hence, the total running time for initially computing all PD$(p),p \in P\setminus \{r\}$, is $O(|P|^2 \log |P|)$.
Let $T$ be equal to $T^{\text{opt}}_{i-1}$ and let $S^-$ (resp. $S^+$) be equal to $S[P_{y_{i-1} \leq 0},y_{i-1}]$ (resp. $S[P_{y_{i-1} \geq 0},$ $ y_{i-1}]$) then $T$ and $S^-$ (resp. $S^+$) can be updated such that $T$ becomes equal to $T^{\text{opt}}_{i}$ and $S^-$ (resp. $S^+$) becomes equal to $S[P_{y_{i} \leq 0},y_{i}]$ (resp. $S[P_{y_{i} \geq 0},y_{i}]$) in:
1. \[en:first\] $O(k_{\lfloor\frac{i}{2}\rfloor}\log |P|)$ time if $i$ is even and $y_i$ is not perpendicular to a line passing through the root $r$ and another point in $P$.
2. \[en:second\] $O(k_{\lfloor\frac{i}{2}\rfloor}\log |P|)$ time if $i$ is odd and $y_{i-1}$ is not perpendicular to a line passing through the root $r$ and another point in $P$.
3. \[en:third\] $O(|P| \log |P|)$ time if $i$ is even and $y_i$ is perpendicular to a line passing through the root $r$ and another point $q \in P\setminus \{r\}$.
4. \[en:fourth\] $O(|P| \log |P|)$ time if $i$ is odd and $y_{i-1}$ is perpendicular to a line passing through $r$ and another point $q \in P\setminus \{r\}$.
We first explain how to maintain the data structures so that all points relevant to case \[en:first\] are treated in $O(k_{\lfloor\frac{i}{2}\rfloor}\log |P|)$ time. Since $y_i$ is not perpendicular to a line connecting the root $r$ with another point in $P$ then $P_{y_{i} \leq 0} = P_{y_{i-1} \leq 0}$ and $P_{y_{i} \geq 0} = P_{y_{i-1} \geq 0}$. Hence, no point is inserted into (or removed from) $S^-$ or $S^+$. However, some points which had different projections onto the $y_{i-1}$ axis, now have the same projection onto the $y_{i}$ axis.
We only explain how to deal with the points in $S^+$; the points in $S^-$ can be treated similarly. Recall that $L_{\lfloor\frac{i}{2}\rfloor}$ contains the $k_{\lfloor\frac{i}{2}\rfloor}$ pairs of points of $P$ that have the same projection onto the $y_i$ axis. Let $S^+ = (r = p_0, p_1, \ldots, p_s)$ and let $(p_{j_1}, p_{j_1+1})$, $(p_{j_2}, p_{j_2+1})$, …, $(p_{j_k}, p_{j_k+1})$ with $j_1 < j_2 < \ldots < j_k$ be the $k$ pairs of points in $S^+$ that belong to $L_{\lfloor\frac{i}{2}\rfloor}$, i.e. they are connected by line perpendicular to $y_i$. Then, each $p \in S^+\setminus \{p_{j_1}$, $p_{j_1+1}$, $p_{j_2}$, $p_{j_2+1}$, …, $p_{j_k}$, $p_{j_k+1}\}$ has the same relative order with the other points of $S^+$ w.r.t. both the $y_{i}$ axis and the $y_{i-1}$ axis. Hence, $p$ is placed at the correct position in $S^+$, the parent of $p$ in $T$ is correct (see Lemma \[lem:yMonotoneChar\]) and PD$(p)$ does not need to be updated.
As a result, the only changes that may be necessary regard the points in $\{p_{j_1}$, $p_{j_1+1}$, $p_{j_2}$, $p_{j_2+1}$, …, $p_{j_k}$, $p_{j_k+1}\}$. For these points we may need to recalculate their parent in $T$. We may also need to swap some consecutive $p_{j_l}$ and $p_{j_l+1}, l =1,2, \ldots, k$, in $S^+$. Finally, we may need to update some of the PD$(p_{j_l})$ and PD$(p_{j_l+1}), l =1,2, \ldots, k$.
For each $l = 1$ to $k$ we do the following:
We compute $d(p_{j_l+1}, \{p_0, p_1, \ldots, p_{j_l-1}\})$ in $O(\log |P|)$ time using PD$(p_{j_l+1})$.
If $d(p_{j_l+1},$ $\{p_0, p_1$, …, $p_{j_l-1}\}) \geq$ $d(p_{j_l},$ par$(p_{j_l}))$, then we do not update anything, since the points $p_{j_l}$ and $p_{j_l+1}$ are placed at the correct position in $S^+$.
If, on the other hand, $d(p_{j_l+1},$ $\{p_0, p_1$, …, $p_{j_l-1}\}) <$ $d(p_{j_l},$ par$(p_{j_l}))$, we remove the edges $\overline{\text{par}(p_{j_l})p_{j_l}}$ and $\overline{\text{par}(p_{j_l+1})p_{j_l+1}}$ from $T$. Then, we swap the order of the points $p_{j_l+1}$ and $p_{j_l}$ in $S^+$, i.e. if $S^+$ was previously equal to $(r = p_0,$ $p_1,$ …, $p_{j_l}, p_{j_l+1},$ …, $p_n)$, now $S^+$ becomes equal to $(r = p_0$, $p_1$, …, $p'_{j_l}, p'_{j_l+1}$, …, $p_{n})$ with $p'_{j_l}$ equal to $p_{j_l+1}$ and $p'_{j_l+1}$ equal to $p_{j_l}$. We then insert the pair $(p'_{j_l}, d(p'_{j_l}, p'_{j_l+1}))$ into PD$(p'_{j_l+1})$ and remove the pair $(p'_{j_l+1}, d(p'_{j_l+1}, p'_{j_l}))$ from PD$(p'_{j_l})$. Finally, in $T$ we connect the point $p'_{j_l}$ (resp. $p'_{j_l+1}$) with the point $p$ s.t. $(p,d(p,p'_{j_l}))$ (resp. $(p,d(p,p'_{j_l+1}))$) has the minimum key in PD$(p'_{j_l})$ (resp. PD$(p'_{j_l+1})$) and update its parent accordingly. Using PD$(p'_{j_l})$ and PD$(p'_{j_l+1})$ all this process which concerns a single pair of points is completed in $O(\log |P|)$ time. Thus, $k_{\lfloor\frac{i}{2}\rfloor}$ pairs of points are treated in $O(k_{\lfloor\frac{i}{2}\rfloor} \log |P|)$ time.
Case \[en:second\] is treated similarly to Case \[en:first\].
We now treat Case \[en:third\]. First, observe that this case occurs exactly $|P|-1$ times, i.e. one time for each pair $(r,p), p \in P\setminus\{r\}$. In this case, either $P_{y_{i-1} \leq 0}$ is a strict subset of (not equal to) $P_{y_{i} \leq 0}$ or $P_{y_{i-1} \geq 0}$ is a strict subset of (not equal to) $P_{y_{i} \geq 0}$. W.l.o.g. we assume that $P_{y_{i-1} \geq 0}$ is a strict subset of $P_{y_{i} \geq 0}$ and that $y_{i-1}(q) < 0$, i.e. $q$ belongs to $S^-$ w.r.t. $y_{i-1}$ while it belongs to both $S^+$ and $S^-$ w.r.t. the $y_i$ axis.
We first explain how to deal with the insertion of $q$ into the point set $P_{y_{i} \geq 0}$. We insert $q$ into the sequence $S^+$ right after $r$. We do not need to update the PD$(q)$ since it already contains only the pair $(r,d(r,q))$. We also do not need to update par$(q)$. Then, for each point $p \in S^+$ with $p \notin \{r,q\}$ we insert the pair $(q, d(q,p))$ in the PD$(p)$ and if $d(q,p)$ is the lowest key in PD$(p)$, then we remove $\overline{\text{par}(p)p}$ from $T$, we assign $q$ to the par$(p)$ and then we insert the edge $\overline{pq}$ to $T$. We are now done with $q$. All the previously described actions take $O(|P|\log |P|)$ time, for the single pair $(r,q)$.
Then, for the pairs of points that have the same projection onto the $y_{i}$ axis, except for $(r,q)$, we apply the procedure described in Case \[en:first\]. As shown in that Case, this is done in $O(k_{\lfloor\frac{i}{2}\rfloor} \log |P|) = O(|P| \log |P|)$ time, since $ k_{\lfloor\frac{i}{2}\rfloor} < |P|$.
Case \[en:fourth\] is treated similar to Case \[en:third\].
Since $\sum\limits_{i =0}^{m-1} k_i = O(|P|^2)$, the total running time of our algorithm is $O(|P|^2 \log |P|)$.
Recognizing Rooted Uniform Monotone Graphs {#subsec:recUM}
------------------------------------------
We now proceed to the problem of deciding if a given rooted connected geometric graph is rooted uniform monotone. Like we did for the rooted UMMST problem, we tackle this decision problem with a rotational sweep algorithm. Again, it is sufficient to take into account only polynomially many directions and more specifically linearly many (to be proved in Lemma \[lem:recArbNecDirections\]).
We first define some auxiliary sets. Let $G = (P, E)$ be a rooted connected geometric graph with root $r$ and $p$ be a point of $P\setminus \{r\}$. Let $A(p,y)$ be the set that contains all the adjacent points to $p$ that are on the same side with $p$ w.r.t. the $x$ axis and are strictly closer to the $x$ axis than $p$. More formally, $A(p,y) = \{q: q\in$ Adj$(p)$, $q$ lies on the same half plane with $p$ w.r.t. the $x$ axis and $|y(q)| < |y(p)|$ $\}$. Let $B(y)$ denote the set $\{p: p\in P\setminus\{r\}$ and $A(p,y) \neq \o \}$. Let $C(y)$ be the set that consists of the points $p\in P\setminus \{r\}$ that (i) do not belong to $B(y)$ and (ii) are connected with some other point $q$ with the same $y$ coordinate such that $A(q,y) \neq \o$. More formally, $C(y) = \{p: p\in P\setminus (B(y)\cup\{r\})$ such that there exists $q \in $Adj$(p)$ with $y(q) = y(p) $ and $A(q,y) \neq \o \}$. An example of a rooted geometric graph and the corresponding sets is given in Figure \[fig:illSets\].
We now give a characterization of rooted $y-$monotone graphs based on the previously defined sets.
\[lem:charYrootSetTerm\] Let $G = (P,E)$ be a rooted connected geometric graph such that for each $p \in P\setminus\{r\}$, $y(p) \neq 0$. Then, $G$ is rooted $y-$monotone if and only if $|B(y)| +$ $|C(y)| = |P| - 1$.
We first prove the ($\Rightarrow$) direction. We prove that each $p \in P\setminus\{r\}$ is included in exactly one of $B(y)$ and $C(y)$ (i.e. $P\setminus\{r\} = B(y) \cup C(y)$) which is equivalent to $|B(y)| +$ $|C(y)|$ = $|P| - 1$. Each $p \in P\setminus\{r\}$ is connected with $r$ by a $y-$monotone path, hence there exists a point $q \in P\setminus \{p\}$ on the same side with $p$ w.r.t. the $x$ axis that is adjacent to $p$ and $|y(q)| \leq |y(p)|$. If for such a point $q$ it holds that $|y(q)| < |y(p)|$, then $A(p,y) \neq \o$ and hence $p \in B(y)$. Otherwise, there exists exactly one point $q \in P\setminus \{r,p\}$ that is adjacent to $p$ and on the same side with $p$ w.r.t. the $x$ axis and $|y(q)| = |y(p)|$. Then, for $q$ it holds that $y(q) = y(p)$ and hence $p$ belongs to $C(y)$. The ($\Leftarrow$) direction can be easily proved by induction on the number of points.
\[rem:generCharYrootSetTerm\] If there exists a point $p \in P\setminus\{r\}$ with $y(p) = 0$ then $G$ is rooted $y-$monotone if and only if (i) $p$ is connected with $r$ and (ii) $|B(y)| +$ $|C(y)|$ equals to $|P| - 2$.
\[rem:extToZeroCoord\] If we know $B(y)$, $C(y)$ and whether there exists a point $p \in P\setminus \{r\}$ with $y(p) = 0$ connected to $r$, we can decide if $G$ is rooted $y-$monotone. This implies a $O(|E|)$ time algorithm, different from the one given in Theorem \[thm:recDirected\], which recognizes rooted $y-$monotone geometric graphs.
\[obs:changeInSets\] Let $G = (P,E)$ be a rooted connected geometric graph. If we rotate an axis $y'$ counterclockwise, then $B(y')$, $C(y')$ and the points $p$ in $P\setminus \{r\}$ with $y'(p) = 0$, change only when the $y'$ axis reaches (or moves away from) a line that is perpendicular to an edge of $G$ or that is perpendicular to a line passing through the root $r$ and a point of $P\setminus\{r\}$.
Using similar arguments to the ones employed for solving the rooted UMMST problem, we define a set of critical slopes and appropriate axes which we have to test in order to decide if the rooted connected geometric graph $G = (P,E)$ is rooted uniform monotone. Let $\Theta = \{\theta \in [0,\pi): \theta$ is the slope of a line perpendicular to either an edge of $G$ or to a line passing through the root $r$ and another point of $P\}$. $S[\Theta]$ is the sorted sequence that contains the slopes of $\Theta$ in increasing order, i.e. $S[\Theta] = (\theta_0, \theta_1, \ldots, \theta_{m-1})$, $\theta_i < \theta_{i+1}$, $i =0, 1, \ldots, m-2$ and $m < |E| +|P|$. We define the critical set of slopes, $\Theta_{\text{critical}} = \{\theta_0$, $\theta_1$, …, $\theta_{m-1}\} \cup \{$ $\frac{\theta_0 + \theta_1}{2}$, $\frac{\theta_1 + \theta_2}{2}$, …, $\frac{\theta_{m-2} + \theta_{m-1}}{2}$, $\frac{\theta_{m-1} + \pi}{2}\}$. We now assign “names” to the axes with slope in $\Theta_{\text{critical}}$. Let $y_{2i}$ be the axis with slope $\theta_i, 0\leq i \leq m-1$. Moreover, let $y_{2i+1}$ be the axis of slope $\frac{\theta_i + \theta_{i+1}}{2}$, $0\leq i \leq m-2$ and $y_{2m-1}$ be the axis of slope $\frac{\theta_{2m-1} + \pi}{2}$. In analogy to Lemma \[lem:MonotoneNecessaryDirections\] we obtain the next Lemma.
\[lem:recArbNecDirections\] $G$ is rooted uniform monotone if and only if it is rooted $y'-$monotone for some $y'$ axis of slope in $\Theta_{\text{critical}}$.
We now give a rotational sweep algorithm that tests whether a given rooted connected geometric graph $G = (P,E)$ is rooted uniform monotone. Our rooted uniform monotone recognition algorithm rotates an axis $y'$ which initially coincides with $y_0$ until it becomes opposite to the $x$ axis. Throughout this rotation, it checks if $G$ is rooted $y'-$monotone. Taking into account Lemma \[lem:recArbNecDirections\], our algorithm only needs to test if $G$ is rooted $y_i-$monotone for some $i = 0,1, \ldots, 2m-1$. The pseudocode of our rooted uniform monotone recognition algorithm is given in Algorithm \[alg:recRooted\].
**Input:** A rooted connected geometric graph $G = (P, E)$.
**Output:** The axis of monotonicity if it exists, otherwise, null.
axis $\gets $ null. compute the axes $y_0, y_1, \ldots, y_{2m-1}$ with slopes in $\Theta_{\text{critical}}$. axis $\gets y_0$. axis $\gets y_i$. **return** axis
\[thm:recRootedMonotone\] Let $G = (P,E)$ be a rooted connected geometric graph. Then, we can decide in $O(|E|\log |P|)$ time if $G$ is rooted uniform monotone.
By Lemma \[lem:recArbNecDirections\], it is immediate that our rooted uniform monotone recognition algorithm (Algorithm \[alg:recRooted\]) decides if $G$ is rooted uniform monotone. We now show that its time complexity is $O(|E|\log |P|)$. Computing the axes $y_0, y_1$, …, $y_{2m-1}$, with slope in $\Theta_{\text{critical}}$ can be done in $O(|E| \log |P|)$ time. Let $k_i$ be the number of pairs of points of $P$ connected by an edge perpendicular to $y_{2i}, 0 \leq i \leq m-1$. Then, $\sum\limits_{i = 0}^{m-1} k_i = |E|$. For each $i = 0,1$,…,$m-1$, we construct a list $L_i$ containing the $k_i$ pairs of points of $P$ that are connected by an edge perpendicular to $y_{2i}$. All $L_i$, $0 \leq i \leq m-1$, can be computed in $O(|E|\log |P|)$ total time.
Let $y_i$ be the last axis taken into account. Our algorithm maintains for each $p \in P\setminus\{r\}$ a data structure $A(p)$ which represents the set $A(p,y_i)$ (which is a subset of the Adj$(p)$). $A(p)$ contains the indices of the points of $P$ that belong to $A(p,y_i)$. $A(p)$ can be implemented by any data structure which supports insert, delete and retrieve operations in $O(\log |P|)$ time (e.g. a $2-3$ tree). Our algorithm also maintains the data structure $B$ that represents the $B(y_i)$. In order to performing the insert and delete operations in $O(1)$ time, $B$ is implemented as an array of boolean with size $O(|P|)$.
Let $p$ be a point in $P\setminus \{r\}$ then computing the $A(p)$ s.t. $A(p)$ equals to $A(p, y_0)$ takes $O(|\text{Adj}(p)|\log |P|)$ time since all the points $q$ adjacent to $p$ are checked and if $q$ belongs to the same half plane with $p$ w.r.t. the $x$ axis and $|y(q)| < |y(p)|$ then $q$ is inserted into $A(p)$ in $O(\log |P|)$ time. Hence, computing all $A(p), p \in P\setminus \{r\}$, takes $O(|E|\log |P|)$ total time. Given each $A(p), p \in P\setminus \{r\}$, computing $B$ s.t. $B$ equals to $B(y_0)$ takes $O(|P|)$ time, i.e. $O(1)$ time to check if $A(p) \neq \o$ for each $p \in P\setminus \{r\}$.
When we move from the $y_{i-1}$ to the $y_i$ axis, the necessary updates we need to make s.t. for each $p \in P\setminus\{r\}$, $A(p)$ represents the set $A(p,y_i)$ and $B$ represents the $B(y_i)$ take:
1. $O(k_{\lfloor \frac{i}{2}\rfloor} \log |P|)$ time if $i$ is even and there is no point $p \in P\setminus \{r\}$ with the same projection with $r$ onto $y_i$
2. $O(k_{\lfloor \frac{i}{2}\rfloor} \log |P|)$ time if $i$ is odd and there is no point $p \in P\setminus \{r\}$ with the same projection with $r$ onto $y_{i-1}$
3. $O((|\text{Adj}(q)| + k_{\lfloor \frac{i}{2}\rfloor}) \log |P|)$ time if $i$ is even and $y_i$ is perpendicular to the line passing through $r$ and the point $q \in P\setminus \{r\}$
4. $O( (|\text{Adj}(q)| + k_{\lfloor \frac{i}{2}\rfloor}) \log |P|)$ time if $i$ is odd and $y_{i-1}$ is perpendicular to the line passing through $r$ and the point $q \in P\setminus \{r\}$
Cases (1),(2),(3) and (4) are proved similarly to the Cases (1),(2),(3) and (4) in Theorem \[thm:spanningTreeOneDir\].
We also note that given $B(y_{2i})$ and each $A(p,y_{2i}), p\in P\setminus \{r\}$, then computing $C(y_{2i}), 0 \leq i \leq m-1$, takes $O(k_i)$ time using the list $L_i$. Furthermore, if we have both $B$ equal to $B(y_i)$ and $C$ equal to $C(y_i)$ and know if $y_i$ is perpendicular to some line passing through $r$ and another point $p$ of $P$ with $\overline{pr} \in E$, then we can test if $G$ is $y_i-$rooted-monotone in $O(1)$ time (see Lemma \[lem:charYrootSetTerm\] and Remark \[rem:extToZeroCoord\]).
Since, $\sum\limits_{i = 0}^{m-1} k_i = |E|$ and $\sum\limits_{q \in P\setminus\{r\}} |\textrm{Adj}(q)| = O(|E|)$, the time complexity of the algorithm is $O(|E| \log |P|)$.
We note that the approach we took for deciding if a given rooted connected geometric graph is rooted uniform monotone has some similarities with the approach employed in the third section of the article of Arkin et al. [@ArkCM89].
Rooted Uniform 2D-monotone Graphs: Minimum Spanning Tree Production, and Recognition {#sec:2D-UMMST}
====================================================================================
In this section we study monotonicity w.r.t. two perpendicular axes. Our treatment is analogous to that of Section \[sec:y-uMMST\] and Section \[sec:generalRootMon\].
The Rooted $xy-$MMST Problem
----------------------------
We first study the case where the perpendicular axes are given, i.e. they are the $x$ and $y$ axes.
In analogy with Observation \[obs:unnesEdge\] we obtain the following Observation.
\[obs:xyUnnesEdges\] Let $P$ be a rooted point set, $G = (P,E)$ be a rooted $xy-$monotone spanning graph of $P$ and let $\overline{pp'} \in E$ where either (i) $p$ and $p'$ lie on different quadrants of the plane or (ii) $p$ and $p'$ lie on the same quadrant of the plane and $(|x(p) | - |x(p')|)(|y(p)| - |y(p')|) < 0$. Then, every path from the root $r$ to a point $q \in P\setminus\{r\}$ that contains $\overline{pp'}$ is not $xy-$monotone.
\[cor:pointsOnDifQuadr\] Let $P$ be a rooted point set and $G^{\text{opt}}$ be the rooted $xy-$monotone minimum spanning graph of $P$. Let $p$ and $q$ be points of $P$ that do not lie on the same quadrant of the plane, then $G^{\text{opt}}$ does not contain the $\overline{pq}$.
The previous Corollary implies that producing the rooted $xy-$monotone minimum spanning graph of $P$ can be split into four independent problems. More specifically, obtaining the rooted $xy-$monotone minimum spanning graph of (i) $P_{x \leq 0,y \leq 0}$, (ii) $P_{x \geq 0,y \geq 0}$, (iii) $P_{x \leq 0, y \geq 0}$ and (iv) $P_{x \geq 0, y \leq 0}$. This is stated more formally in the following Lemma.
\[lem:splitQuadrants\] Let $P$ be a rooted point set and $G^{\text{opt}}$ be the rooted $xy-$monotone minimum spanning graph of $P$. Let $G_{x \leq 0,y \leq 0}^{\text{opt}}$, $G_{x \geq 0,y \geq 0}^{\text{opt}}$, $G_{x \leq 0, y \geq 0}^{\text{opt}}$ and $G_{x \geq 0, y \leq 0}^{\text{opt}}$ be the rooted $xy-$monotone minimum spanning graph of $P_{x \leq 0,y \leq 0}$, $P_{x \geq 0,y \geq 0}$, $P_{x \leq 0, y \geq 0}$ and $P_{x \geq 0, y \leq 0}$, respectively. Then, $G^{\text{opt}}$ is $G_{x \leq 0,y \leq 0}^{\text{opt}} \cup G_{x \geq 0,y \geq 0}^{\text{opt}}\cup G_{x \leq 0, y \geq 0}^{\text{opt}} \cup G_{x \geq 0, y \leq 0}^{\text{opt}}$.
Let $P$ be a rooted point set confined to one quadrant of the plane. Then, we define $S[P,y,x]$ to be the sequence that consists of the points of $P$, such that the points in $S[P,y,x]$ are ordered w.r.t. their absolute $y$ coordinates and if two points have the same absolute $y$ coordinate, then they are ordered w.r.t. their absolute $x$ coordinates. More formally, $S[P,y,x] = (r= p_0,p_1,p_2, \ldots, p_n)$ such that $|y(p_0)| \leq |y(p_1)| \leq |y(p_2)| \leq \ldots \leq |y(p_n)|$ and $|y(p_i)| = |y(p_{i+1})|$ implies that $ |x(p_i)| < |x(p_{i+1})|$ and $P = \{p_0,p_1,p_2, \ldots, p_n\}$. Using similar arguments with the proof of Lemma \[lem:yMonotoneChar\] we obtain a characterization of the rooted $xy-$monotone minimum spanning graph of $P$.
\[lem:minCostxy2DChar\] Let $P$ be a rooted point set confined to one quadrant of the plane, $S[P,y,x]$ $=$ $(r=$ $p_0$, $p_1$, $p_2$, …, $p_n)$ and $G$ be a geometric graph with vertex set $P$. Then, $G$ is the rooted $xy-$monotone minimum spanning graph of $P$ if and only if (i) $p_n$ is connected only with its closest point with absolute $x$ coordinate less than or equal to $|x(p_n)|$ from $\{p_0$, $p_1$, …, $p_{n-1}\}$, i.e. the point $p_j \in P\setminus \{p_n\}$ such that $|x(p_j)| \leq |x(p_n)|$ and $d(p_n,p_j) = d(p_n, P_{|x| \leq |x(p_n)|}\setminus \{p_n\})$, and (ii) $G\setminus\{p_n\}$ is the rooted $xy-$monotone minimum spanning graph of $P\setminus\{p_n\}$.
Lemma \[lem:splitQuadrants\] and Lemma \[lem:minCostxy2DChar\] lead to the next Corollary.
The rooted $xy-$monotone minimum spanning graph of a rooted point set $P$ is a geometric tree.
Let $P$ be a rooted point set confined to one quadrant of the plane and $S[P,y,x]$ $=$ $(r=$ $p_0$, $p_1$, $p_2$, …, $p_n)$. For each $i = 1, 2, \ldots, n$, we call the closest point with absolute $x$ coordinate in the range $[0, |x(p_i)|]$ to $p_i$ from $\{p_0$, $p_1$, …, $p_{i-1}\}$, the parent of $p_i$ and denote it as par$(p_i)$. Equivalently, par$(p_i) = p_j$ if and only if $j < i$, $|x(p_j)| \leq |x(p_i)|$ and $d(p_i, p_j)$ $=$ $d(p_i$, $\{p_0$, $p_1$,…, $p_{i-1}$$\}_{|x| \leq |x(p_i)|})$. Then, Lemma \[lem:minCostxy2DChar\] yields the following Corollary.
\[cor:xyEdgesChar\] The edges of the rooted $xy-$MMST of $P$ are exactly the line segments $\overline{\text{par}(p_i)p_i}$, $i = 1$, $2$, …, $n$.
The previous Corollary implies a $O(|P|^2)$ time algorithm that produces the rooted $xy-$MMST of $P$. However, using the semi-dynamic data structure for closest point with attribute value in specified range queries that was implicitly produced by Bentley [@Ben79], the time complexity of our rooted $xy-$MMST algorithm is reduced to $O(|P| \cdot \log ^3 |P|)$.
\[thm:minSpanningXYrooted\] The rooted $xy-$MMST of a rooted point set $P$ can be computed in $O(|P| \cdot \log ^3 |P|)$ time.
We initially construct $P_{x \leq 0,y \leq 0}$, $P_{x \geq 0,y \geq 0}$, $P_{x \leq 0, y \geq 0}$ and $P_{x \geq 0, y \leq 0}$. Then, we apply our rooted $xy-$MMST algorithm on $P_{x \leq 0, y \leq 0}$, $P_{x \geq 0, y \geq 0}$, $P_{x \leq 0, y \geq 0}$ and $P_{x \geq 0, y \leq 0}$ and we obtain the trees $T_{x \leq 0, y \leq 0}^{\text{opt}}$, $T_{x \geq 0,y \geq 0}^{\text{opt}}$, $T_{x \leq 0, y \geq 0}^{\text{opt}}$ and $T_{x \geq 0, y \leq 0}^{\text{opt}}$, respectively. Using Fact \[fact:rangeClosest\] and since $O(|P|)$ insertions and $O(|P|)$ closest point with attribute value in specified range queries are performed, computing $T_{x \leq 0, y \leq 0}^{\text{opt}}$, $T_{x \geq 0,y \geq 0}^{\text{opt}}$, $T_{x \leq 0, y \geq 0}^{\text{opt}}$ and $T_{x \geq 0, y \leq 0}^{\text{opt}}$ takes $O(|P| \cdot \log ^3 |P|)$ time. Finally, we return the union of $T_{x \leq 0, y \leq 0}^{\text{opt}}$, $T_{x \geq 0,y \geq 0}^{\text{opt}}$, $T_{x \leq 0, y \geq 0}^{\text{opt}}$ and $T_{x \geq 0, y \leq 0}^{\text{opt}}$ which by Lemma \[lem:splitQuadrants\] is the rooted $xy-$MMST of $P$.
Using the same reduction that we used in Theorem \[thm:lowerBoundyMonotone\], i.e. the reduction from sorting given by Shamos [@Sha78], we obtain a lower bound for the time complexity of every algorithm which solves the rooted $xy-$MMST (or the rooted 2D-UMMST or the rooted 2D-MMST) problem.
Producing the rooted $xy-$MMST (or the rooted 2D-UMMST or the rooted 2D-MMST) of a rooted point set $P$ requires $\Omega (|P|\log |P|)$ time.
We also give a linear time algorithm that recognizes rooted $xy-$monotone graphs.
\[thm:recRootedXY\] Let $G = (P, E)$ be a rooted connected geometric graph. Then, we can decide in $O(|E|)$ time if $G$ is rooted $xy-$monotone.
Our proof is similar to the proof of Theorem \[thm:recDirected\]. We transform $G$ to a directed graph $\overrightarrow{G}$ in $O(|E|)$ time as follows. Let $\overline{pq}$ be an edge of $G$. If $p$ and $q$ lie on the same quadrant of the plane and $|x(p)| \leq |x(q)|$ and $|y(p)| \leq |y(q)|$ (resp. $|x(q)| \leq |x(p)|$ and $|y(q)| \leq |y(p)|$) we direct $\overline{pq}$ from $p$ to $q$ (resp. from $q$ to $p$). By observation \[obs:xyUnnesEdges\], it follows that all the other edges cannot be traversed by a $xy-$monotone path connecting $r$ with another point of $P$. Hence, we remove them. Then, $G$ is rooted $xy-$monotone if and only if $r$ is connected with all other points of $P$ in $\overrightarrow{G}$. We can decide the latter in $O(|E|)$ time by applying a breadth first search or a depth first search traversal.
The Rooted 2D-UMMST Problem
---------------------------
We now study the problem of computing the rooted 2D-UMMST of a given rooted point set $P$. We give a $O(|P|^2 \log |P|)$ time rotational sweep algorithm that solves the problem, analogous to our rooted UMMST algorithm given in Subsection \[subsec:UMMST\].
\[obs:rotationEventsXYproduction\] Let $x'y'$ be a Cartesian System. If we rotate the Cartesian System $x'y'$ counterclockwise, then the rooted $x'y'-$MMST of $P$ changes only when the $y'$ axis reaches (or moves away from) a line that is perpendicular or parallel to a line passing through two points of $P$.
When the $y'$ axis reaches a line that is perpendicular or parallel to a line passing through two points of $P\setminus \{r\}$, say $p$ and $q$, then it might become feasible (while previously this was not feasible) to connect $p$ and $q$, with the line segment $\overline{pq}$, such that $p$ or $q$ traverses $\overline{pq}$ in a $x'y'-$monotone path from it to $r$. In this case, by Corollary \[cor:xyEdgesChar\], the rooted $x'y'-$MMST of $P$ may change. Furthermore, when the $y'$ becomes perpendicular or parallel to a line passing through $r$ and a point $p \in P\setminus \{r\}$ the point $p$ belongs to two quadrants (while previously belonged only to one of them). As a result, for some points $q_1$, $q_2$, …, $q_k \in P\setminus\{r,p\}$ belonging to the new quadrant, in which $p$ was inserted, may now be feasible to connect with $r$ via the $x'y'-$monotone path $(r,p,q_i), 1\leq i \leq k$, respectively. Hence, by Corollary \[cor:xyEdgesChar\], the rooted $x'y'-$MMST of $P$ may change.
On the other hand, when the $y'$ axis moves away from a line that is perpendicular or parallel to a line passing through $p$ and $q$, with $p,q \neq r$, then if the line segment $\overline{pq}$ was previously an edge of the rooted $x'y'-$MMST of $P$, hence it was previously traversed by the $x'y'-$monotone path from $p$ or $q$, say $p$, to $r$, now might not be feasible to be traversed in the $x'y'-$monotone path from $p$ to $r$ (see Observation \[obs:xyUnnesEdges\]). Hence, the rooted $x'y'-$MMST of $P$ may change. Furthermore, if one of $p$ or $q$ coincides with $r$, say $p$, then $q$ now does not belong to one of the quadrants that it belonged previously. Hence, some points of $P\setminus\{r,q\}$ which previously were adjacent to $q$ and previously belonged to the same quadrant with $q$ are now on different quadrants. Hence, from Corollary \[cor:pointsOnDifQuadr\] it follows that the edges that previously connected $q$ and these points cannot belong to the current rooted $x'y'-$MMST.
At no other moment during the rotation of the $x'y'$ Cartesian System, i.e. when the $y'$ axis does not reach (move away from) a line that is perpendicular or parallel to a line passing through two points of $P$, may become feasible/infeasible (while previously it was infeasible/feasible) to traverse a line segment connecting two points of $P$ in a $x'y'-$monotone path from $r$ to a point of $P\setminus\{r\}$. From the previous sentence and Corollary \[cor:xyEdgesChar\], it follows that at no other moment during the rotation of the $x'y'$ Cartesian System may the rooted $x'y'-$MMST of $P$ change.
\[obs:quarterPlaneRot\] Let $x' , y'$ and $x'', y''$ be axes of two different Cartesian Systems s.t. $x''$ (resp. $y''$) forms with $x'$ (resp. $y'$) a counterclockwise angle equal to $k\frac{\pi}{2}, k = 1, 2, 3$. Then, the rooted $x'y'-$MMST of $P$ coincides with the rooted $x''y''-$MMST of $P$.
Based on the previous Observations, we define the set $\Theta = \{\theta \in [0, \frac{\pi}{2}):$ a line of slope $\theta$ is either perpendicular or parallel to a line connecting two points of $P\}$. Let $S[\Theta]$ be the sorted sequence that contains the slopes in $\Theta$ in increasing order, i.e. $S[\Theta] = (\theta_0, \theta_1, \ldots, \theta_{m-1})$, $\theta_i < \theta_{i+1}, 0 \leq i \leq m-2$ and $ m \leq \binom{|P|}{2}$. Then, we define the critical set of slopes $\Theta_{\text{critical}} = \{\theta_0$, $\theta_1$, …, $\theta_{m-1}\} \cup \{\frac{\theta_0 + \theta_1}{2}$, $\frac{\theta_1+ \theta_2}{2}$, …, $\frac{\theta_{m-2}+ \theta_{m-1}}{2}$, $\frac{\theta_{m-1}+ \frac{\pi}{2}}{2}\}$. We now “name” the Cartesian Systems such that their vertical axis has slope in $\Theta_{\text{critical}}$. More formally, let $x_0y_0$, $x_1y_1$, …, $x_{2m-1}y_{2m-1}$ be the Cartesian Systems such that $y_{2i}$ has slope $\theta_i, 0 \leq i \leq m-1$, $y_{2i+1}$ has slope $\frac{\theta_i + \theta_{i+1}}{2}, 0 \leq i \leq m-2$ and $y_{2m-1}$ has slope $\frac{\theta_{m-1} + \frac{\pi}{2}}{2}$.
In analogy with Lemma \[lem:MonotoneNecessaryDirections\] we obtain the following Lemma.
\[lem:2Dmonotone\_Directions\] The rooted 2D-UMMST of $P$ is one of the rooted $x_iy_i-$MMST of $P$, $i =0, 1, 2, \ldots, 2m-1$, and more specifically the one of minimum cost.
We can produce the rooted 2D-UMMST of a rooted point set $P$ in $O(|P|^2 \log |P|)$ time.
We give our algorithm which produces the rooted 2D-UMMST of $P$. Our rooted 2D-UMMST algorithm is a rotational sweep algorithm analogous to the rooted UMMST algorithm given in Subsection \[subsec:UMMST\]. It rotates a Cartesian System $x'y'$ which initially coincides with $x_0y_0$ counterclockwise until it coincides with the Cartesian System $xy$, i.e. the given Cartesian System. Throughout this rotation, it updates the rooted $x'y'-$MMST of $P$. By Lemma \[lem:2Dmonotone\_Directions\], we only need to compute the rooted $x_iy_i-$MMST of $P$ for the Cartesian Systems $x_iy_i$, $i = 0, 1$, …, $2m-1$. Let $T_{i}^{\text{opt}}$ be the rooted $x_iy_i-$MMST of $P$, $i =$ $0$, $1$, …, $2m-1$. Then, our rooted 2D-UMMST algorithm is restated as follows. The algorithm initially computes the Cartesian Systems $x_0y_0$, $x_1y_1$, …, $x_{2m-1}y_{2m-1}$ in $O(|P|^2\log |P|)$ time. Then, it constructs $T_{0}^{\text{opt}}$ using Theorem \[thm:minSpanningXYrooted\]. Then, it iterates for $i =1$, $2$, …, $2m-1$, obtaining $T_{1}^{\text{opt}}$, $T_{2}^{\text{opt}}$, …, $T_{2m-1}^{\text{opt}}$ in this order. Throughout its execution it stores the Cartesian System min$X$min$Y$ in which it encountered the minimum cost solution found so far. Finally, it returns the rooted min$X$min$Y-$MMST of $P$, which is recomputed using Theorem \[thm:minSpanningXYrooted\].
From Lemma \[lem:2Dmonotone\_Directions\], our rooted 2D-UMMST algorithm produces the rooted 2D-UMMST of $P$. We only need to analyze its time complexity. We use similar data structures PD$(p), p \in P\setminus\{r\}$, to the ones employed in Theorem \[thm:spanningTreeOneDir\]. More specifically, for each point $p \in P \setminus \{r\}$ the data structure PD$(p)$ is a self-balancing binary search tree that contains the pairs $(q,d(p,q))$ for all the points $q \in P\setminus\{p\}$ such that $\overline{pq}$ can be traversed in a $x_iy_i-$monotone path from $p$ to $r$ (i.e. for the points $q \in P\setminus\{p\}$ that lie on the same quadrant with $p$ w.r.t. the Cartesian System $x_iy_i$ and $|x_i(q)| \leq |x_i(p)|$ and $|y_i(q)| \leq |y_i(p)|$), where $i$ is the index of the current iteration of our algorithm. Then, using similar arguments to the arguments employed in Theorem \[thm:spanningTreeOneDir\], the time complexity of the algorithm is $O(|P|^2 \log |P|)$.
Recognizing Rooted Uniform 2D-monotone Graphs
---------------------------------------------
We now study the problem of recognizing if a given rooted connected geometric graph $G = (P, E)$, with root $r$, is rooted uniform 2D-monotone. Our approach is analogous to the approach we took for recognizing rooted uniform monotone graphs in Subsection \[subsec:recUM\].
For each $p \in P\setminus \{r\}$ let $A(p,x,y)$ be the set $\{q: $ $q\in $Adj$(p)$ and $q$ lies on the same quadrant of the plane with $p$ and $|x(q)| \leq |x(p)|$ and $|y(q)| \leq |y(p)|$ $\}$. Let $B(x,y)$ be the set $\{p: p\in P\setminus \{r\}$ and $A(p,x,y) \neq \o$ $\}$. Then, similarly to Lemma \[lem:charYrootSetTerm\] we obtain the following Lemma.
\[lem:2DmonotneChar\] $G$ is rooted $xy-$monotone if and only if $|B(x,y)|$ equals to $|P|-1$.
The previous Lemma implies a $O(|E|)$ time recognition algorithm, that decides if $G$ is rooted $xy-$monotone, different from the algorithm given in Theorem \[thm:recRootedXY\].
Similarly to Observation \[obs:rotationEventsXYproduction\], we obtain the following Observation.
If we rotate a Cartesian System $x'y'$ counterclockwise then the sets $A(p,x',y'), p \in P\setminus \{r\}$ and $B(x',y')$ change only when the $y'$ axis reaches (or moves away from) a line perpendicular or parallel to an edge of $G$ or when the $y'$ axis reaches (or moves away from) a line perpendicular or parallel to a line connecting $r$ with another point of $P$.
From the previous Observation, it follows that we need to take into account only the Cartesian Systems $x_0y_0$, $x_1y_1$, …, $x_{2m-1}y_{2m-1}$, $m < |E| + |P|$ such that $y_{0}, y_2, \ldots, y_{2m-2}$, are all the axes that are either (i) perpendicular or parallel to some edge of $E$ or (ii) perpendicular or parallel to some line connecting $r$ with another point in $P$. The slope of each $y_{2i}$ is $\theta_i, 0 \leq i \leq m-1$ and it holds that $0 \leq \theta_0 < \theta_1 <$ …$< \theta_{m-1} < \frac{\pi}{2}$. Moreover, the slope of each $y_{2i+1}$ is equal to $\frac{\theta_{i} + \theta_{i+1}}{2}, i = 0, 1, \ldots, m-2$ and the slope of $y_{2m-1}$ is equal to $\frac{\theta_{m-1} + \frac{\pi}{2}}{2}$. Similarly to Lemma \[lem:recArbNecDirections\], we obtain the following Lemma.
\[lem:sufDirections2DUni\] $G$ is rooted uniform 2D-monotone if and only if it is rooted $x_iy_i-$monotone for some Cartesian System $x_iy_i$, $i = 0, 1, \ldots, 2m-1$.
Given a rooted connected geometric graph $G = (P, E)$, we can decide in $O(|E|\log |P|)$ time if $G$ is rooted uniform 2D-monotone.
The proof is similar to the proof of Theorem \[thm:recRootedMonotone\]. We employ a rotational sweep algorithm that decides if $G$ is rooted uniform 2D-monotone. From Lemma \[lem:sufDirections2DUni\], our rooted uniform 2D-monotone recognition algorithm decides if $G$ is rooted uniform 2D-monotone by testing if $G$ is rooted $x_iy_i-$monotone for some $i = 0,1, \ldots, 2m-1$. It tests that in this order, i.e. it first checks $x_0y_0$ then $x_1y_1$, …, and at the end it checks $x_{2m-1}y_{2m-1}$.
We now show that its complexity is $O(|E| \log |P|)$. We can compute the Cartesian Systems $x_0y_0$, $x_1y_1$, …, $x_{2m-1}y_{2m-1}$ in $O(|E| \log |P|)$ time. The algorithm maintains for each $p \in P\setminus\{r\}$ a data structure $A(p)$ which represents the $A(p,x_i,y_i)$ (when the algorithm checks the $x_iy_i$ Cartesian System) and can be implemented as a $2-3$ tree that stores the indices of the points that it contains. Moreover, the algorithm maintains a data structure $B$ that represents the $B(x_i,y_i)$ (when the algorithm checks the $x_iy_i$ Cartesian System) and is implemented as an array of boolean of $O(|P|)$ size. Using similar analysis to the one presented in Theorem \[thm:recRootedMonotone\], the initial construction of all $A(p), p \in P\setminus\{r\}$ s.t. $A(p)$ equals to $A(p,x_0,y_0)$ takes $O(|E|\log |P|)$ total time. Then, the construction of $B$ s.t. $B$ equals to $B(x_0,y_0)$ takes $O(|P|)$ time. Furthermore, using similar arguments to the ones presented in Theorem \[thm:recRootedMonotone\], the updates of all $A(p), p \in P\setminus\{r\}$, and $B$ throughout all the execution of the algorithm take $O(|E|\log |P|)$ total time. Additionally, from Lemma \[lem:2DmonotneChar\], given $B$ equal to $B(x_i,y_i)$, it can be decided in $O(1)$ time if $G$ is rooted $x_iy_i-$monotone. Hence, performing all the tests, i.e. if $G$ is rooted $x_iy_i-$monotone for $i = 0$, $1$, …, $2m-1$, using the data structure $B$ take $O(|E|)$ total time. From all the previous, it follows that the time complexity of the algorithm is $O(|E|\log |P|)$.
Conclusions and Future Work
===========================
In this article we studied the problem of constructing the minimum cost spanning geometric graph of a given rooted point set in which the root is connected to all other vertices by paths that are monotone w.r.t. a single direction, i.e. they are $y$-monotone (or w.r.t. a pair of orthogonal directions, i.e. they are $xy$-monotone). We showed that the minimum cost spanning geometric graph is actually a tree and we proposed polynomial time algorithms that construct it for the case where the direction (the pair of orthogonal directions) of monotonicity is given or remains to be determined.
Several directions for further research are open.
1. We studied rooted point sets and we built the minimum cost spanning tree that contains monotone paths w.r.t. a single direction from the root $r$ to any other point in the point set. What about the case where we are given a $k$-rooted point set, i.e. a set with $k$ designated points as its roots, and we are asked to find the minimum cost spanning geometric graph containing monotone paths w.r.t. a single direction from each root to every other point in the point set. In this case, is a wanted graph a tree and additionally, can we find a polynomial time algorithm for this problem?
In the extreme case where all points in the point set $P$ are designated as roots, the problem is trivial. Since a geometric graph $G=(P,E)$ is $y-$monotone only if it contains as subgraph the graph path $W_{|P|}$ visiting all points in increasing order of their $y$ coordinates [@Ang15], the $y-$monotone minimum spanning tree of $P$ is actually the graph path $W_{|P|}$. Furthermore, the uniform monotone minimum spanning tree of $P$ can be efficiently produced by a rotational sweep algorithm similar to the one employed for the rooted UMMST.
2. We showed that computing the rooted UMMST (or 2D-UMMST) of a rooted point set can be done in polynomial time. But is this also the case for the (rooted) monotone (or 2D-monotone) minimum spanning graph of a (rooted) point set or is the problem NP-hard?
3. We studied the problem of building rooted minimum cost spanning geometric graphs that possess a specific property, and we focused on the property of monotonicity (w.r.t. one or two orthogonal directions). What if we consider a different requirement/property? For example, we can ask for the minimum cost spanning geometric graph containing increasing-chord paths or self-approaching paths (see [@IckKL99; @AlaCGLP13]) from the root to any other point in the point set. In this case, is the sought graph a tree and does there exist an efficient algorithm that produces it?
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Finding the best way to leverage non-volatile memory (NVM) on modern database systems is still an open problem. The answer is far from trivial since the clear boundary between memory and storage present in most systems seems to be incompatible with the intrinsic memory-storage duality of NVM. Rather than treating NVM either solely as memory or solely as storage, in this work we propose how NVM can be simultaneously used as both in the context of modern database systems. We design a persistent buffer pool on NVM, enabling pages to be directly read/written by the CPU (like memory) while recovering corrupted pages after a failure (like storage). The main benefits of our approach are an easy integration in the existing database architectures, reduced costs (by replacing DRAM with NVM), and faster peak-performance recovery.'
author:
- Lucas Lersch
- Wolfgang Lehner
- Ismail Oukid
bibliography:
- 'sample-base.bib'
title: Persistent Buffer Management with Optimistic Consistency
---
Introduction
============
NVM is a persistent media promising higher bandwidth (2$\times$) and lower latency (10$\times$) than modern NAND-flash SSDs. Furthermore, NVM can be attached to the memory bus, thus allowing it to be directly accessed by the CPU through its caches in a much smaller granularity (cache-lines) than regular block devices. Therefore, NVM introduces not only a new layer within the storage hierarchy [@bonnet2017s], but it also enables more flexibility regarding data placement.
While a few factors slower, reading data directly from NVM can be done the same way as with DRAM. However, writing data directly to NVM imposes challenges in terms of data consistency. The root of these challenges is the lack of control the application has over data movement between CPU cache and NVM in comparison to the data movement between DRAM and SSD. In other words, the programmer cannot prevent cache lines from being evicted from the CPU cache and written-back to NVM at arbitrary points in time.
Related work address these challenges with solutions that fall in one of three categories (also identified by previous work [@van2018managing]). First, early proposals leverage NVM as a cheaper alternative to extend DRAM, enabling larger buffer pools [@wu2014app; @ou2014wear]. These approaches focus on reducing write amplification and improving wear leveling on NVM, but they do not enforce any consistency when writing data, and therefore do not leverage persistency. Second, persistent data structures [@oukid2016fptree; @yang2015nv; @chen2015persistent; @arulraj2018bztree; @lee2017wort] enable direct fine-grained writes to NVM by issuing out-of-place writes and relying on instruction ordering (*SFENCE*) and eagerly flushing cache lines (*CLFLUSH/CLWB*) to make the operation visible (usually by flipping a validity bit). Third, buffer managers were proposed to integrate NVM with the storage hierarchy below DRAM and above SSD [@pelley2013storage; @van2018managing; @eisenman2018reducing]. These approaches access data on DRAM, and therefore they have full control of when data is persisted. Optimizations can be made regarding block sizes, but they still impose movement of data between DRAM and NVM.
The first two categories treat NVM more like traditional memory, while the last one treats NVM more like traditional storage. However, NVM is actually both. We consider that a database system should allow NVM to be read and written directly, thus exploring its memory-storage duality to its full potential. Our system accesses NVM like memory, but it guarantees consistency of writes to NVM like storage. We achieve that by integrating NVM in the buffer pool of a database system to either extend DRAM or completely replace it, while still leveraging its persistency in an optimistic way. In other words, we never enforce ordering of writes or eagerly flush cache lines. The main motivation is that corruption occurs when a write operation is partially evicted from the CPU cache to NVM. With the capacity of NVM being significantly higher than CPU caches, the probability of corruption tends to be low and therefore pessimistically enforcing consistency of every write introduces a relatively high overhead by not leveraging the CPU cache.
Recent work [@DBLP:journals/corr/abs-1901-10938] has gone into the same direction of allowing data to be accessed directly on NVM. This approach complements our approach by focusing on cost models for optimizing data movement, while we focus on enabling direct writes to NVM to be consistent.
Background
==========
We give an overview of the techniques that we use to implement a persistent buffer management in a database system.
**Buffer Management:** We assume a traditional transactional storage manager having a B+Tree as its workhorse data structure. The B+tree is organized such that nodes are represented by pages, which are the unit of data movement and buffering. Optimizations like pointer swizzling [@graefe2014memory] and low-overhead replacement policies [@leis2018leanstore] may apply. The atomicity and durability of writes to pages buffered in DRAM is guaranteed by write-ahead logging (WAL).
**Single-page Recovery:** Similar to traditional ARIES [@mohan1992aries], we assume a page-level physiological logging. This implies that pages are not only a unit of data movement, but also of fault containment and repair [@DBLP:journals/pvldb/GraefeK12]. This enables techniques such as write-elision, on-demand instant restart and restore, and single-page repair [@graefe2014instant]. These techniques are the base to enable direct writes directly made to NVM to be consistent without eagerly flushing cache lines.
System Design
=============
The pages of our system are primarily located on SSD and only the warm pages are buffered in NVM. Hot pages might be buffered in a DRAM portion of the buffer pool, as seen in Figure \[fig:nvm\_buffer\].
**Normal Processing:** During normal transactional processing, a page to be updated will be either on SSD, NVM, or DRAM. In case the page is on SSD, it is loaded to the buffer pool (either to DRAM or NVM, to be decided by a placement policy). In case the page is on DRAM, we have a hit and the page is updated normally. If the page is on NVM, two actions might occur. First, the page might be identified as “heating up” by a placement policy such as 2Q [@lersch2017rethinking], in which case it will be copied to DRAM and updated there. Second, the page might be simply warm, in which case the update is done directly on NVM. In the last case, since atomicity and durability is guaranteed by WAL, issuing *CLFLUSH/CLWB* after updating the page is not necessary as the log serves as the single source of truth. However, when restarting after a system failure, pages that were on NVM might be corrupted because updates were not properly persisted. As a consequence, the current state of a page on NVM is unknown and therefore persistency of NVM cannot be leveraged.
We address this challenge by dividing it in two dimensions: corruption detection and page repair. Each page contains an 8 Byte checksum of the whole page. Whenever the content of a page is modified directly on NVM, the page checksum is re-calculated and updated. At this point, the modifications and the checksum of a page might be persistent or not, since we do not explicitly flush them from the CPU cache to NVM. In case the overhead of calculating the checksum for the whole page after every update is too high, it can be reduced either by updating the checksum only after a certain number of updates (at the cost of higher corruption ratio) or by introducing multiple checksums per page corresponding to fractions of the page (at the cost of higher space consumption).
![During normal processing reads and writes can access both DRAM and NVM directly (dashed lines). Whole pages can be moved across any devices (solid lines). The trade-offs between DRAM and NVM can be analysed by sliding an abstract persistency bar.[]{data-label="fig:nvm_buffer"}](NVM_Buffer_Pool.pdf)
**Corruption Detection** After a system failure, the restart process starts with log analysis, which identifies the state (not the content) of pages right before the crash. We assume state-of-the-art page-based on-demand restart [@graefe2014instant], therefore a page being requested might still be lingering in the NVM portion of the buffer pool since before the crash. Two steps are required to determine if the page can be used. First, its checksum is calculated and compared to the checksum stored within the page. If the checksums do not match, it is because either the checksum or the updated data were not evicted from the CPU cache to NVM. Second, if the checksums match, the log sequence number of the last update in the page (*pageLSN*) is compared to the LSN retrieved during log analysis (*expectedLSN*). To summarize, a page might be in one of the following states:
- Corrupted, if checksum does not match
- Behind, if checksum matches and $ pageLSN < expectedLSN $
- Current, if checksum matches and $ pageLSN = expectedLSN $
- Ahead, if checksum matches and $ pageLSN > expectedLSN $
The page is *behind*, if it was updated by a committed transaction (the log records were flushed) but neither the update nor the checksum were evicted from the CPU cache. In this case, the page is in a consistent but outdated state since it violated the write-ahead rule. However, this violation is tolerated on NVM as long as it is guaranteed not to happen on SSD. Analogously, the page is *ahead* if both update and checksum were persisted to NVM but the transactions that made these updates did not commit. Finally, if both checksum and LSNs match, the page is in its most *current* state and ready to be accessed. Except for the *current* state, all other states must be recovered by replaying log records.
**Page Repair** No assumption can be made about a *corrupted* page, and therefore the only alternative is to retrieve its more recent version from SSD (which acts as a backup) and replay the log records referring to this page to bring it up to its most recent state. A *behind* page is missing committed updates, therefore it can be used as the basis for replaying the log records, not requiring an older version of the page to be fetched from SSD. *Ahead* pages are consistent but contain updates made by uncommitted transactions. Since ARIES-style recovery only allows logical UNDO on a transaction level, the updates of a single page cannot be independently rolled back. Therefore, the same procedure of a *corrupted* page applies for rolling forward. Since it is required that an older version of the page is read from the database, it is convenient that frequently modified pages are flushed to the database often to bound the recovery time in such cases. Fortunately, this can be achieved by an asynchronous cleaning job that iterates through the buffer pool and picks dirty pages to be flushed to the database.
Conclusion
==========
In this work we propose extending database systems with a persistent buffer pool on NVM. We consider such extension to impose a low implementation effort, since NVM is treated very similarly to DRAM during runtime, while leveraging well-understood recovery algorithms to enforce its consistency. The consequence is that the persistency aspect can be leveraged in an optimistic way, without major changes in the code base to enforce ordering of writes to NVM. Furthermore, cache lines do not have to be eagerly flushed, which enables write operations to be cached in the CPU and amortize the higher latencies of NVM. We also consider that in the short and medium term, NVM will not replace neither DRAM or SSD, but should act in synergy with both. Our design also enables the user to explore different proportions of DRAM and NVM in the buffer pool: more DRAM will lead to higher performance, while more NVM will lead to reduced costs and faster recovery (higher availability). This adds more flexibility for analyzing trade-offs and navigating the performance continuum between modern expensive in-memory databases and more traditional low-cost disk-based systems.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Mohi Khansari$^{1}$, Daniel Kappler$^{1}$, Jianlan Luo$^{2}$, Jeff Bingham$^{1}$, Mrinal Kalakrishnan$^{1}$[^1][^2]'
bibliography:
- 'main.bib'
title: '**Action Image Representation: Learning Scalable Deep Grasping Policies with Zero Real World Data** '
---
[^1]: $^{1}$X, The Moonshot Factory, [{khansari, kappler, jeffbingham, kalakris}@google.com]{},
[^2]: $^{2}$University of California, Berkeley (the work was done as an intern at X.), [[email protected]]{}
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Luis Fuentes[^1]'
- Manuel Pedreira
title: 'Involutions of a canonical curve.'
---
0.2cm
\[section\] \[teo\][Definition]{} \[teo\][Proposition]{} \[teo\][Corollary]{} \[teo\][Lemma]{} \[teo\][Example]{} \[teo\][Remark]{}
\[subsection\] \[teo2\][Definition]{} \[teo2\][Proposition]{} \[teo2\][Corollary]{} \[teo2\][Lemma]{} \[teo2\][Example]{} \[teo\][Remark]{}
2 Ł[[L]{}]{} =eufm10 at 12pt c å ¶[[**P**]{}]{} Å[[**A**]{}]{}
[ Departamento de Algebra, Universidad de Santiago de Compostela. $15706$ Santiago de Compostela. Galicia. Spain. e-mail: [[email protected]]{};\
[[email protected]]{}\
[**Abstract:**]{} We give a geometrical characterization of the ideal of quadrics containing a canonical curve with an involution. This implies to study involutions of rational normal scrolls and Veronese surfaces.\
[**Mathematics Subject Classifications (1991):**]{} Primary, 14H37; secondary, 14H30, 14J26.\
[**Key Words:**]{} Canonical curve, involution, rational normal scrolls.]{}
[**Introduction:**]{} Let $C$ be a nonhyperelliptic smooth curve of genus $\pi$. An involution of $C$ is an automorphism $\vhi:C\lrw C$ such that $\vhi^2=id$. It induces a double cover $\gamma:C\lrw C/\vhi=X$, where $X$ is a smooth curve of genus $g$. We say that $C$ has an involution of genus $g$. By Hurwitz formula, we know that $\pi\geq 2g-1$. It is well know that a general smooth curve of genus $\pi\geq 3$ has not nontrivial automorphisms. In particular a smooth curve with an involution is not generic.
In this paper we give a geometric characterization of the ideal of quadrics containing the canonical model of a nonhyperelliptic curve $C$ with an involution. Let $C_{\K}\subset \P^{\pi-1}$ be the canonical model of $C$. We see that an involution of $C_{\K}$ is a harmonic involution; that is, it can be extended to $\P^{\pi-1}$.
An involution $\ov{\vhi}$ of $\P^n$ has two complementary spaces of base points $S\mas$ and $S\menos$. Moreover, $\ov{\vhi}$ induces an involution $\ov{\vhi}^*$ in the space of quadrics of $\P^n$. This involution has two spaces of base points: the [*base quadrics*]{}, that is, quadrics containing the spaces $S\mas$ and $S\menos$ and the [*harmonic quadrics*]{}, that is, quadrics such that $S\mas$ and $S\menos$ are polar respect to them. A subspace $\Sigma\subset
\P(H^0(\Te_{P^n}(2)))$ is called a [*base-harmonic system*]{} respect to $S\mas,S\menos$ when it is a fixed space of $\ov{\vhi}^*$. In this case $\Sigma_b$ and $\Sigma_h$ will denote the base quadrics and the harmonic quadrics of $\Sigma$ respectively.
We prove the following Theorem:\
\
[**Theorem \[fundamental\]**]{}
* Let $C_{\K}\subset \P^{\pi-1}$ be the canonical curve of genus $\pi$, with $\pi>4$. If $C_{\K}$ has an involution of genus $g$ then $\pi\geq 2g-1$ and the quadrics of $\P(H^0(I_{C_{\K}}(2)))$ are a base-harmonic system respect to the base spaces $\P^{g-1}$, $\P^{\pi-g-1}$ that contains $(g-1)(\pi-g-2)$ independent base quadrics. Conversely, these conditions are sufficient to grant the existence of an involution, except when:*
1. $\pi=6,g=2$ and $C_{\K}$ has a $g^2_5$; or
2. $\pi=2g$,$2g+1$ or $2g+2$ and $C_{\K}$ is trigonal.
First we prove that the conditions are sufficient when the curve $C_{\K}$ is complete intersection of quadrics. The Enriqües-Babbage Theorem says that $C_{\K}$ is the complete intersection of quadrics except when it is trigonal or $C$ is a quintic smooth curve. In these cases the quadrics containing the canonical curve intersect in a rational normal scroll and in the Veronese surface respectively.
In order to examine the special cases we make an study of the harmonic involutions of the rational normal scrolls and the Veronese surface. We compute the number of base quadrics on each case. From this calculus we obtain the Corollaries \[coro1\] and \[coro2\]:\
\
[**Corollary \[coro1\]**]{}[* The unique involutions on a trigonal canonical curve of genus $\pi$, $\pi>4$ are of genus $\frac{\pi}{2},\frac{\pi-1}{2}$ or $\frac{\pi-2}{2}$.* ]{}\
\
[**Corollary \[coro2\]**]{}[* The unique involutions on a smooth quintic plane curve are of genus $2$.*]{}
Furthermore, we make a particular geometrical study of the canonical curves of genus $4$ with an involution of genus $2$ and genus $1$.
Note that to compute the number of independent base quadrics containing the canonical curves with and involution we need the result about the projective normality of the canonical scrolls (see [@pedreira2],$\S5$). We will follow the notation of [@fuentes] and [@hartshorne] to work with scrolls and ruled surfaces.
We thanks Lawrence Ein by his interest on this work during his visit to our Department on November, 2001.
Harmonic involutions. {#involutions}
=====================
\[definvolucion\]
Let $X\subset \P^n$ be a projective variety. An isomorphism $\vhi:X\lrw X$ is called an involution if $\vhi^2=Id$. Moreover, if $\vhi$ is the restriction of an involution $\ov{\vhi}:\P^n\lrw \P^n$ then $\vhi$ is called a harmonic involution.
\[caracterizaarmonicas\]
Let $X\subset \P^n$ be a linearly normal projective variety. An involution $\vhi:X\lrw X$ is harmonic if and only if $\vhi^*(X\cap H)\sim X\cap H$ for all hyperplane $H$.
[**Proof:**]{} If $\vhi$ is harmonic we clearly have that $\vhi^*(X\cap H)\sim X\cap H$.
Conversely, if $\vhi^*(X\cap H)\sim X\cap H$ for all hyperplane, then we have an involution $\vhi^*:H^0(\Te_X(1))\lrw H^0(\Te_X(1))$ that makes the following diagram commutative:
$$\setlength{\unitlength}{5mm}
\begin{picture}(15,4)
\put(2.6,3){\makebox(0,0){$\P(H^0(\Te_X(1))^{\vee})$}}
\put(3.6,0){\makebox(0,0){$X$}}
\put(12.5,3){\makebox(0,0){$\P(H^0(\Te_X(1))^{\vee})$}}
\put(10.5,0){\makebox(0,0){$X$}}
\put(3.6,0.7){\vector(0,1){1.5}}
\put(10.5,0.7){\vector(0,1){1.5}}
\put(5.5,3){\vector(1,0){4}}
\put(5.5,0){\vector(1,0){4}}
\put(7.5,3.5){\makebox(0,0){$\P({\vhi^*}^{\vee})$}}
\put(7.5,0.6){\makebox(0,0){$\vhi$}}
\end{picture}$$ so $\vhi$ extends to $\P^n$.
[**Examples:**]{}
1. [*Any involution of a rational curve $D_n\subset \P^n$ is harmonic.*]{}
It is sufficient to note that $\vhi^*(D_n\cap H)$ has degree $n$; because $D_n$ is a rational curve, it follows that $\vhi^*(D_n\cap H)\sim D_n\cap H$.
2. [*Any involution of a canonical curve $C_{\K} \subset \P^{\pi-1}$ of genus $\pi$ is harmonic.*]{}
The linear system $|\vhi^*(C_{\K}\cap H)|$ has degree $2\pi-2$ and it has dimension $\pi$, so it is the canonical linear system and $\vhi^*(C_{\K}\cap H)\sim C_{\K}\cap H$.
3. [*Any involution of a normal rational scroll $R_{n-1}\subset \P^n$ with invariant $e>0$ is harmonic.*]{}
Let $S_e=\P(\Te_{P^1}\oplus \Te_{P^1}(-e))$ be the ruled surface associated to $R_{n-1}$. We know that $H\cap R_{n-1}\sim X_0+\b f$. Since $\vhi$ is an isomorphism, $\vhi^*(X_0)^2=X_0^2$ and $\vhi^*(f)^2=f^2=0$. But, in $S_e$ with $e>0$, $X_0$ is the unique curve with negative self-intersection and $f$ is the unique curve with self-intersection $0$. From this $\vhi^*(X_0)\sim X_0$ and $\vhi^*(f)\sim f$. Then: $$\vhi^*(H\cap R_{n-1})\sim \vhi^*(X_0+\b f)\sim \vhi^*(X_0)+\b \vhi^*(f)\sim X_0+\b f\sim H\cap
R_{n-1}$$
We recall some basic facts about involutions in a projective space.
\[espaciosfijos\]
Any involution $\ov{\vhi}$ of $\P^n$ has two complementary spaces $S\mas,S\menos \\ \subset
\P^n$ of base points. In this way, the image of a point $P$ is the fourth harmonic of $P$, $l\cap S\mas$ and $l\cap S\menos$, where $l$ is the unique line passing through $P$ verifying $l\cap S\mas\neq
\emptyset$ and $l\cap S\menos\neq \emptyset$.
Conversely, any pair of complementary spaces $S\mas,S\menos$ of $\P^n$ defines an involution of $\P^n$.
\[expresionmatricial\] [ If we take a base of $\P^n$, $W=\{ P_1,\ldots,P_{k+1},P'_1,\ldots,P'_{k'+1} \}$ where $\langle P_1,\ldots,P_{k+1} \rangle=S\mas$ and $\langle P'_1,\ldots,P'_{k'+1} \rangle=S\menos$ then the involution $\vhi$ is given by the matrix: $$M_{\vhi}=\left(
\begin{array}{cc}
{Id}&{0}\\
{0}&{-Id}\\
\end{array}\right)$$ ]{}
\[espaciosbase\]
Under the above assumptions, we say that $\vhi$ is harmonic respect to $S\mas$ and $S\menos$. Moreover, $S\mas$ and $S\menos$ are called the base spaces of $\vhi$.
\[cuadricas\]
*The linear isomorphism $\ov{\vhi}$ induces an isomorphism: $$\ov{\vhi}^*:\P(H^0(\Te_{P^n}(2)))\lrw \P(H^0(\Te_{P^n}(2)))$$ Because $\ov{\vhi}$ is an involution, $\ov{\vhi}^*$ is an involution too. Therefore, it has two complementary spaces of base points. Taking coordinates respect to the base $W$, a quadric $Q\subset \P^n$ has a matrix: $$M_{Q}=\left(
\begin{array}{cc}
{A}&{C}\\
{C^t}&{B}\\
\end{array}\right)$$*
We see that $\ov{\vhi}^*(Q)=Q$ if and only if $M_{\ov{\vhi}^*(Q)}=M_{\vhi}M_QM_{\vhi}=\lambda
M_Q$ for some $\lambda\neq 0$, if and only if $A=B=0$ or $C=0$. In the first case $Q$ is called a harmonic quadric and in the second case $Q$ is called a base quadric. The set of harmonic quadrics will be denoted by $\P(H^0(\Te_{P^n}(2))_h)$ and the set of base quadrics by $\P(H^0(\Te_{P^n}(2))_b)$. We have that:
1. $Q$ is a harmonic quadric if and only if $P'^tM_QP=0$ for each $P\in S\mas,P'\in
S\menos$; that is, if $S\mas$ and $S\menos$ are polar respect to $Q$.
2. $Q$ is a base quadric if and only if $S\mas,S\menos\subset Q$.
\[sistemaarmonico\]
Let $\Sigma\subset \P(H^0(\Te_{P^n}(2)))$ be a projective subspace $\Sigma=\P(V)$ and let $\ov{\vhi}:\P^n\lrw\P^n$ a harmonic involution respect to two spaces $S\mas,S\menos$. $\Sigma$ is called a base-harmonic system respect to $S\mas,S\menos$ when it is a fixed space of $\ov{\vhi}^*$.
\[notasistemaarmonico\]
[*If we denote $\Sigma_h=\Sigma\cap \P(H^0(\Te_{P^n}(2))_h)$ and $\Sigma_b=\Sigma\cap
\P(H^0(\Te_{P^n}( \\ 2))_b)$, we see that $\Sigma$ is a base-harmonic system if and only if $\Sigma=\Sigma_h+\Sigma_b$; that is, $V$ has a base composed by harmonic and base quadrics.*]{}
\[equivalencia\]
Let $X\subset \P^n$ be a projective variety.
1. If $X$ has a harmonic involution $\vhi$ respect to two spaces $S\mas,S\menos$ then the system $\P(H^0(I_X(2)))$ is base-harmonic system respect to $S\mas,S\menos$.
2. Suppose that $X$ is the complete intersection of quadrics. If $\P(H^0(I_X(2)))$ is a base-harmonic system respect to $S\mas,S\menos$, then $X$ has a harmonic involution respect to two spaces $S\mas,S\menos$.
[**Proof:**]{}
1. Let $Q\in \P(H^0(I_X(2)))$, that is, $X$ is contained on $Q$. It is sufficient to show that $\ov{\vhi}^*(Q)$ contains $X$. Since $\ov{\vhi}(X)=X$ and $X\subset Q$, $X=\ov{\vhi}(X)\subset
\ov{\vhi}(Q)$ and the conclusion follows.
2. Let $\ov{\vhi}:\P^n\lrw \P^n$ the harmonic involution defined by the spaces $S\mas,S\menos$. If $X$ is the complete intersection of quadrics, then $X=Q_1\cap\ldots\cap Q_k$ where $\{ Q_1,\ldots ,Q_k\}$ is a base of $\P(H^0(I_X(2)))$. If $\P(H^0(I_X(2)))$ is a base-harmonic system respect to $S\mas,S\menos$ we can choose a base of fixed quadrics, so $$\ov{\vhi}(X)=\ov{\vhi}(Q_1\cap\ldots\cap Q_k)=\ov{\vhi}(Q_1)\cap\ldots\cap\ov{\vhi}(Q_k)=
Q_1\cap\ldots\cap Q_k=X$$ and we can restrict $\ov{\vhi}$ to $X$.
\[cuadricasbase\]
Let $X\subset \P^n$ be a projective variety and let $\vhi:X\lrw X$ a harmonic involution of $X$ respect to spaces $S\mas,S\menos$.
Let $F=\overline{\{P\in \P^n/P\in \langle x,\vhi(x) \rangle,
x\in X\}}$ be the variety of lines joining points of $X$ related by the involution. Then: $$\begin{array}{rl}
{h^0(I_{X,P^n}(2))_b}&{=h^0(I_{F,P^n}(2))_b=}\\
{}&{=h^0(I_{F,P^n}(2))-h^0(I_{F\cap
S\mas,S\mas}(2))-h^0(I_{F\cap S\menos,S\menos}(2))}\\
\end{array}$$
[**Proof:**]{} Let us first prove that $H^0(I_{X,P^n}(2))_b=H^0(I_{F,P^n}(2))_b$.
Since $X\subset F$, a quadric containing $F$ contains $X$ too, so $H^0(I_{F,P^n}(2))_b\subset H^0(I_{X,P^n}(2))_b$.
Conversely, if $Q\in H^0(I_{X,P^n}(2))_b$, $X,S\mas,S\menos\subset Q$. Therefore, each line $l$ of $F$ meets $Q$ in four points: $(x,\vhi(x),l\cap S\mas,l\cap S\menos)$ and then it is contained on $Q$. Thus, $F\subset Q$ and $H^0(I_{X,P^n}(2))_b\subset H^0(I_{F,P^n}(2))_b$.
Now, let us consider the following exact sequence: $$0\lrw H^0(I_{F\cup S\mas,P^n}(2))\lrw H^0(I_{F,P^n}(2))\stackrel{\A}{\lrw} H^0(I_{F\cap
S\mas,S\mas}(2))$$ Let us see that $\A$ is a surjective map. Let $Q\mas\subset S\mas$ be a quadric which contains $F\cap S\mas$. Taking the cone of $Q\mas$ over $S\menos$, we obtain a quadric of $\P^n$ that contains the lines joining $Q\mas$ and $S\menos$ so it contains $F$. From this, we deduce that: $$h^0(I_{F\cup S\mas,P^n}(2))=h^0(I_{F,P^n}(2))-h^0(I_{F\cap S\mas,S\mas}(2))$$ Similarly, we have $$0\lrw H^0(I_{F\cup S\mas\cup S\menos,P^n}(2))\lrw H^0(I_{F\cup S\mas,P^n}(2))\stackrel{\beta}{\lrw}
H^0(I_{F\cap S\menos,S\menos}(2))$$ where $H^0(I_{F\cup S\mas\cup S\menos,P^n}(2))=H^0(I_{F,P^n}(2))_b$ and $\beta$ is a surjective map. Therefore: $$\begin{array}{rl}
{h^0(I_{F,P^n}(2))_b}&{=h^0(I_{F\cup S\mas,P^n}(2))-h^0(I_{F\cap S\menos,S\menos}(2))=}\\
{}&{=h^0(I_{F,P^n}(2))-h^0(I_{F\cap S\mas,S\mas}(2))-h^0(I_{F\cap S\menos,S\menos}(2))}\\
\end{array}$$
[*We finish this section by computing the dimension of the spaces of harmonic and base quadrics in an involution over a normal rational curve.*]{}
Let $D_n\subset \P^n$ be a rational normal curve of degree $n$ and let $\vhi:D_n\lrw D_n$ an involution of degree $2$. The lines joining points related by the involution generate a rational normal ruled surface $R_{n-1}\subset \P^n$. This ruled surface has two directrix curves $C\mas,C\menos$ on the base spaces $S\mas,S\menos$ of the involution.
We now ([@hartshorne], IV, $2.17$ and $2.19$) that $R_{n-1}\cong \P(\Te_{P^1}\oplus
\Te_{P^1}(-e))$ for some $e\geq 0$. In this way the divisor of hyperplane sections of $R_{n-1}$ is $H\sim X_0+(n-1+e)/2
f$. Moreover the curve $D_n$ corresponds to a $2$-secant curve on the surface, so $D_n\sim 2X_0+kf$. Since $deg(D_n)=D_n.H=n$, we obtain $k=e+1$. Because $D_n$ is irreducible, $D_n.X_0\geq 0$ and we obtain $e\leq 1$. We deduce that $e=0$ if $n$ is odd and $e=1$ if $n$ is even.
In this way we see that if $n$ is odd, the directrix curves $C\mas,C\menos$ have degree $(n-1)/2$ and if $n$ is even, the directrix curves have degree $(n-2)/2$ and $n/2$.
By applying Proposition \[cuadricasbase\] we have that: $$h^0(I_{D_n}(2))_b=h^0(I_{R_{n-1}}(2))-h^0(I_{C\mas}(2))-h^0(I_{C\menos}(2))$$ The dimension of theses spaces are well known. Thus, we obtain: $$h^0(I_{D_n}(2))_b=\left\{
\begin{array}{l}
{(\frac{n-1}{2})^2\mbox{ if $n$ is odd}}\\
{}\\
{\frac{n(n-2)}{4}\mbox{ if $n$ is even}}\\
\end{array}\right.$$
Note that when $n$ is odd, $D_n\sim 2X_0+f$, so $D_n.X_0=1$ with $C\mas,C\menos\sim X_0$. Thus, the base points of the involution of $D_n$ are $D_n\cap C\mas$ and $D_n\cap C\menos$. When $n$ is even, $D_n\sim 2X_0+2f$, so $D_n.X_0=0$ and $D_n.(X_0+f)=2$ where $C\mas\sim X_0$ and $C\menos\sim
X_0+f$. In this case the base points of the involutions of $D_n$ are the two points of $D_n\cap C\menos$.
Involutions of the canonical curve. {#involutionscanonica}
===================================
Let $C_{\K}$ be the canonical curve of genus $\pi$ and let $\vhi:C_{\K}\lrw C_{\K}$ be an involution. We saw that it is a harmonic involution. We will use the results obtained in [@pedreira2]. The scroll generated by the involution is a canonical scroll $R_{\bb}$. We call genus of the involution $\vhi$ to the genus of the ruled surface $R_{\bb}$. Thus we have a $2:1$-morphism $\gamma:C_{\K}\lrw X$. $R_{\bb}$ has a canonical directrix curve $\ov{X_0}$ of genus $g$, and a nonspecial curve $\ov{X_1}$ with degree $\pi-1$. They lie on disjoint spaces $\P^{g-1}$ and $P^{\pi-g-1}$. The involution of $C_{\K}$ has $2(\pi-1-2(g-1))$ base points that are the ramifications of $\gamma$. We denote them by $\R$ and we know that $\R\sim C_{\K}\cap \ov{X_1}$.
Let us compute the number of base quadrics containing $C_{\K}$, that is, the dimension of $H^0(I_{C_{\K}}(2))_b$. By Proposition \[cuadricasbase\] we know that:
$$h^0(I_{C_{\K}}(2))_b=h^0(I_{R_{\bb}}(2))-h^0(I_{\ov{X_0},P^{g-1}}(2))-h^0(I_{\ov{X_1},P^{\pi-g-1}}(2))$$
We can compute this dimension because we know that (see [@pedreira2]): $$\begin{array}{l}
{h^0(I_{R_{\bb}}(2))=h^0(\Te_{P^{\g2-1}}(2))-h^0(\Te_{S_{\bb}}(2H))+dim(s(H,
H))}\\
{h^0(I_{\ov{X_0}}(2))=h^0(\Te_{P^{g-1}}(2))-h^0(\Te_X(2\k))+dim(s(\k,\k))}\\
{h^0(I_{\ov{X_1}}(2))=h^0(\Te_{P^{\g2-g-1}}(2))-h^0(\Te_X(2\b))+dim(s(\b,\b))}\\
{}\\
{dim(s(H,H))=dim(s(\k,\k))+dim(s(\b,\b))}\\
\end{array}$$ From this, we obtain: $$h^0(I_{C_{\K}}(2))_b=h^0(\Te_{P^{\g2-1}}(2))_b-h^0(\Te_X(\b+\k))$$ Thus the number of base quadrics containing $C_{\K}$ is: $$h^0(C_{\K}(2))_b=(g-1)(\g2-g-2)$$
\[involucioncanonica\]
Let $C_{\K}\subset \P^{\pi-1}$ be a nonhyperelliptic canonical curve of genus $\pi$ that it is complete intersection of quadrics. $C_{\K}$ has an involution of degree $2$ if and only if $\P(H^0(I_{C_{\K}}(2)))$ is harmonic respect to two disjoint complementary subspaces of dimensions $k$ and $\pi-k-2$, with $k\leq \pi-k-1$. Moreover the involution has genus $g=k+1$ if and only if $h^0(I_{C_{\K}}(2))_b=(g-1)(\pi-g-2)$.
[**Proof:**]{} The first assertion follows from Proposition \[equivalencia\].
By the above discussion we know the number of base quadrics $h^0(I_{C_{\K}}(2))_b=(g'-1)(\pi-g'-2)$ where $g'$ is the genus of the involution. This genus is $k+1$ or $\pi-k-1$. Suppose that $(g'-1)(\pi-g'-2)=(g-1)(\pi-g-2)$ with $g=k+1$. If $g'=\pi-k-1=\pi-g$ then we obtain $\pi=2g$ so the genus $g'$ of the involution is $k+1$.
\[icompleta\]
[*It is well known that a nonhyperelliptic canonical curve of genus $\pi\geq 5$ is the complete intersection of quadrics, except when it is trigonal or when $\pi=6$ and it has a $g^2_5$. In the first case the intersection of the quadrics of $\P(H^0(I_{C_{\K}}(2)))$ is a rational normal scroll; in the second case it is the Veronese surface of $\P^5$.*]{}
Involutions of the canonical curve of genus $\pi=4$. {#involuciones4}
----------------------------------------------------
Let $C\cuatro\subset \P^3$ be a canonical curve of genus $4$. It is well known that this curve is the complete intersection of a quadric and a cubic surface (see [@hartshorne],IV,Example $5.2.2$). Suppose that $C\cuatro$ has an involution $C\cuatro\lrw X$ where $X$ is a smooth curve of genus $g$. We have that $\pi-1\geq 2g-2$, so then genus of $X$ can be $1$ or $2$ (if $g=0$ the curve $C\cuatro$ is hyperelliptic).
\[involucion41\]
A canonical curve of genus $4$ has an elliptic involution if and only if it is the complete intersection of an elliptic cubic cone $S$ and a quadric that doesn’t pass through the vertex of $S$.
[**Proof:**]{} Suppose that $C\cuatro$ has an elliptic involution. We know that the involution generates a scroll $R$. In this case the directrix curve $X_0$ has degree $g-1=0$, so it is a point. Then $R$ is an elliptic cone. If $Q$ is the unique quadric that contains $C\cuatro$, then necessary $C\cuatro=Q\cap R$. Moreover, we know that $C\cuatro\cap X_0=\emptyset$, so $Q$ does not pass through the vertex of $R$. Conversely, if $C\cuatro=Q\cap R$ and $Q$ does not pass through the vertex of $R$, the generators of $S$ provide an elliptic involution of $C\cuatro$.
\[involucion42\]
A canonical curve of genus $4$ has an involution of genus $2$ if and only if it is the complete intersection of a quadric and a cubic surface which has a harmonic involution respect to polar lines respect to the quadric.
[**Proof:**]{}
1. Suppose that $C\cuatro=Q_2\cap Q_3$ where $Q_2$ is a quadric and $Q_3$ is a cubic surface with a harmonic involution $\vhi:Q_3\lrw Q_3$. Let $l$ and $l'$ the base spaces. Suppose that they are polar respect to $Q_2$. From this, $\vhi(Q_2)=Q_2$ so $C\cuatro$ has a harmonic involution. Because the base spaces have dimension $1$, the involution has genus $2$.
2. Suppose that $C\cuatro$ has a harmonic involution $\vhi$ of genus $2$. From this, the base spaces are two lines $l$ and $l'$. Let $R\subset \P^3$ the ruled surface generated by the involution. We know that $R$ contains $l$ and $l'$ with multiplicities $3$ and $2$ respectively. Moreover, $l'\cap C\cuatro=\emptyset$ and $l\cap C\cuatro$ consist of two points. We know that $\P(H^0(I_{C\cuatro}(2)))=\{ Q_2
\}$ and $dim\P(H^0(I_{C\cuatro}(3)))=4$. In this way $C\cuatro=Q_2\cap Q_3$ for any cubic surface $Q_3\in
\P(H^0(I_{C\cuatro}(3)))$. Consider the involution $\ov{\vhi}:\P(H^0(I_{C\cuatro}(3)))\lrw
\P(H^0(I_{C\cuatro}(3)))$ induced by $\vhi$. Let us see that there is a fixed irreducible element. Let $V$ be the set of reducible elements of $\P(H^0(I_{C\cuatro}(3)))$. Because $Q_2$ is the unique quadric containing $C\cuatro$ then $V=\{Q_2+H; H\subset \P^3 \}$ and $dim(V)=3$. Since the base points of $\ov{\vhi}$ generate $\P(H^0(I_{C\cuatro}(3)))$, then there exist at least a fixed irreducible element.
Involutions of the Veronese surface. {#involucionesV}
------------------------------------
Let $v_{2,n}$ be the Veronese map of $\P^n$: $$\begin{array}{rcl}
{v_{2,n}:{\P^n}^*}&{\lrw}&{V_{2,n}\subset \P(H^0(\Te_{P^n}(2)))}\\
{[x_0:\ldots :x_n]}&{\lrw}&{[x_0^2:x_0x_1:\ldots :x_n^2]}\\
\end{array}$$ We will denote the image of this map by $V_{2,n}$. If $n=2$, it is the Veronese surface and we will denote it by $V_2$.
\[iveronesevariedad\]
The involutions of the Veronese variety $V_{2,n}$ are harmonic respect to two base spaces which are the harmonic and base quadrics of two subspaces $S\mas,S\menos$ of $\P^n$.
[**Proof:**]{} A harmonic involution $\vhi:\P^n\lrw \P^n$ respect to spaces $S\mas,S\menos$ induces a harmonic involution $\ov{\vhi}$ in $\P(H^0(\Te_{P^n}(2)))$ respect to $\P(H^0(\Te_{P^n}(2))_h)$ and $\P(H^0(\Te_{P^n}(2))_b)$.
Because $\ov{\vhi}\circ v_{2,n}=v_{2,n}\circ \ov{\vhi}$, we see that $\ov{\vhi}$ restricts to $V_{2,n}$. In this way we have a harmonic involution $\ov{\vhi}:V_{2,n}\lrw V_{2,n}$ respect to the spaces $\P(H^0(\Te_{P^n}(2))_h)$ and $\P(H^0(\Te_{P^n}(2))_b)$.
Conversely, given an involution $\eta$ of $V_2$. Applying the isomorphism $v_{2,n}$, we obtain an involution $\vhi^*$ of ${\P^n}^*$ with base spaces ${S\mas}^*,{S\menos}^*$. In this way, the dual map $\vhi=\vhi^{**}$ provides an involution $\ov{\vhi}$ of $\P(H^0(\Te_{P^n}(2)))$ such that $\ov{\vhi}|_{V_{2,n}}=\eta$.
We will denote the Veronese surface by $V_2$.
\[iveronesesuperficie\]
Any nontrivial involution of the Veronese surface $V_2$ is harmonic respect to a line which corresponds to the conics of $\P^2$ passing through a point $P$ and a line $r$, and a $3$-dimensional space $V$ corresponding to the polar conics of $\P^2$ respect to $P$ and $r$.
[**Proof:**]{} It is sufficient to note that an involution $\eta$ of $V_2$ is induced by and involution $\vhi$ of $P^2$. If $\vhi$ is nontrivial, then its base spaces are a point $P$ and a line $r$.
\[veroneseqbase\]
Let $\eta$ be a nontrivial harmonic involution of the Veronese surface. Then $h^0(\Te_{V_2}(2))_b=2$.
[**Proof:**]{} Consider the Veronese map of $\P^2$: $$\begin{array}{rcl}
{v_2:{\P^n}^*}&{\lrw}&{V_2\subset \P(H^0(\Te_{P^n}(2)))}\\
{[x_0:x_1:x_2]}&{\lrw}&{[x_0^2:x_0x_1:\ldots :x_2^2]=[y_0:y_1:\ldots:y_5]}\\
\end{array}$$
Let $Y$ be the matrix: $$\left(\begin{array}{ccc}
{y_0}&{y_1}&{y_2}\\
{y_1}&{y_3}&{y_4}\\
{y_2}&{y_4}&{y_5}\\
\end{array}\right)$$ We know that $H^0(\Te_{V_2}(2))$ is generated by the quadrics of $\P^5$ whose equations are defined by the minors of order $2$ of the matrix $Y$.
Moreover the base spaces of $\vhi$ are a line $l$ corresponding to the conics of $\P^2$ passing through a line $r$ and a point $P$ and a space $V$ corresponding to the polar conics respect to $P$ and $r$.
Taking an adequate system of coordinates we can consider $P$ generated by the equations $\{x_1=x_2=0\}$ and $r$ generated by the equation ${x_0=0}$.
A conic containing $P$ and $r$ has an equation $ax_0x_1+bx_0x_2=0$. Then the equations of $l$ are $\{y_0=y_3=y_4=y_5=0\}$.
A polar conic respect to $P$ and $r$ has an equation $ax_0^2+bx_1^2+cx_1x_2+dx_2^2=0$. Then the equations of $V$ are $\{y_1=y_2=0\}$.
Applying the conditions to contain $V$ and $l$ to the equations of $H^0(\Te_{V_2}(2))$, we obtain that the quadrics containing $V$ and $l$ are generated by $\{y_1y_4-y_2y_3=y_1y_5-y_2y_4=0\}$. From this, $h^0(\Te_{V_2}(2))_b=2$.
Harmonic involutions of the rational ruled surfaces. {#involucionesR}
----------------------------------------------------
Let $R_{n-1}\subset \P^n$ a rational normal ruled surface of degree $n-1$, with $n>3$. Let $\vhi:R_{n-1}\lrw R_{n-1}$ a harmonic involution of the surface. Then $\vhi$ conserves the degree of the curves. From this, it applies generators into generators if $n>3$. In this way, we have the following induced harmonic involutions: $$\begin{array}{rccl}
{\vhi_0:\P^1}&{\lrw}&{\P^1}&{}\\
{P}&{\lrw}&{Q}&{/\vhi(Pf)=Qf}\\
\end{array}$$ where $\P^1$ parameterizes the generators. $$\begin{array}{rcl}
{\vhi_l:|D_l|}&{\lrw}&{|D_l|}\\
{D}&{\lrw}&{\vhi(D)}\\
\end{array}$$ where $|D_l|$ is the linear system of curves of degree $l$.
Let $k$ be the degree of the curve of minimum degree of $R_{n-1}$:
1. If $k=\frac{n-1}{2}$, then there is a $1$-dimensional family of irreducible curves of degree $k$. The involution $\vhi_k$ has at least $2$ base points, so there are two disjoint curves $D_k$ that are invariant by $\vhi$.
2. If $k<\frac{n-1}{2}$, then there is a unique curve of minimum degree, so it is invariant by $\vhi$. Moreover, if $l=n-1-k$ the linear system $|D_l|$ has dimension $l-k$. Its generic curve is an irreducible curve disjoint from $D_k$. In particular, the set of reducible curves of $|D_l|$ are an hyperplane composed by curves of the form $D_k+\sum f_i$. Thus, we have a harmonic involution: $$\begin{array}{rcl}
{\vhi_l:\P^{l-k}\cong |D_l|}&{\lrw}&{|D_l|\cong \P^{l-k}}\\
\end{array}$$ We know that $\vhi_l$ has two disjoint spaces of base points. Both of them can not be contained on the hyperplane (because they generate $\P^{n-k}$), so necessary there exists an irreducible curve in $|D_l|$ that is fixed by the involution $\vhi_l$; that is, it is invariant by $\vhi$.
We conclude the following proposition:
\[conclusion1\]
Given a harmonic involution on a rational normal ruled surface of degree $n-1$, there exist two disjoint rational normal curves $D_k$,$D_l$ with degrees $k$ and $l$, $k+l=n-1$, that are invariant by the involution.
Let $D_k\subset \P^k$ and $D_l\subset \P^l$ be the two invariant curves. The involution $\vhi$ restricts to these spaces. Thus, we have a harmonic involution $\vhi_k:\P^k\lrw \P^k$. It has two invariant spaces $\P^{k\mas}$, $\P^{k\menos}$ with $k\mas+k\menos+1=k$. Similarly, the harmonic involution $\vhi_l:\P^l\lrw \P^l$ has two invariant spaces $P^{l\mas}$, $P^{l\menos}$ with $l\mas+l\menos+1=l$. From this, we have two possibilities for the base spaces $S\mas$,$S\menos$ of the involution $\vhi$: $$\begin{array}{ll}
{\begin{array}{l}
{S\mas=\langle \P^{k\mas},\P^{l\mas} \rangle=\P^{k\mas+l\mas+1}}\\
{S\menos=\langle \P^{k\menos},\P^{l\menos} \rangle=\P^{k\menos+l\menos+1}}\\
\end{array}}&
{\begin{array}{l}
{S\mas=\langle \P^{k\mas},\P^{l\menos} \rangle=\P^{k\mas+l\menos+1}}\\
{S\menos=\langle \P^{k\menos},\P^{l\mas} \rangle=\P^{k\menos+l\mas+1}}\\
\end{array}}\\
\end{array}$$
Conversely if we have two harmonic involutions in $\P^k$ and $\P^l$ we can recuperate an involution in $\P^n$. Note that this involution is not unique, because we have two ways to define it. Moreover, in order to restrict the involution to the ruled surface $q:R_{n-1}\lrw
\P^1$ we need that the involutions in $\P^k$ and $\P^l$ restrict to $D_k$ and $D_l$ and that they are compatible, that is, the images of the points on the same generator lay on the same generator: $q(\vhi_k(D_k\cap P f))=q(\vhi_l(D_l\cap P f)), \forall P\in \P^1$.
Thus, if $\vhi_k$ and $\vhi_l$ verify these conditions we have a harmonic involution $\vhi$ that restricts to $R_{n-1}$.
\[conclusion2\]
A harmonic involution on a normal rational ruled surface $R_{n-1}$ defines two harmonic involutions $\vhi_k$, $\vhi_l$ on two disjoint rational curves $D_k$, $D_l$ that generate the surface. Moreover, they make commutative the diagram $(1)$.
Conversely, if two harmonic involutions $\vhi_k$, $\vhi_l$ on two rational curves generating a rational ruled surface $R_{n-1}$ verifying $q(\vhi_k(D_k\cap P f))=q(\vhi_l(D_l\cap P f)),
\forall P\in \P^1$. , then they define two possible harmonic involutions on $R_{n-1}$, taking the space bases generate by the space bases of $\vhi_k$ and $\vhi_l$.
\[compatibilidad\]
*In order to obtain compatible involutions $\vhi_k$, $\vhi_l$ it is sufficient to define a involution $\eta$ on $\P^1$ and to translate it to $D_k$ and $D_l$: $$\begin{array}{c}
{\vhi_k(D_k\cap P f):=D_k\cap \eta(P) f}\\
{\vhi_l(D_l\cap P f):=D_l\cap \eta(P) f}\\
\end{array}$$ for all $P\in \P^1$.*
Moreover, if $\vhi_k$ and $\vhi_l$ are compatible involutions and one of them is the identity, then the other one is the identity too.
We saw how are the (nontrivial) harmonic involutions on a normal rational curve $D_m\sub \P^m$:
1. If $m=2\mu$ the involution is defined by two base spaces $\P^{\mu}$, $\P^{\mu-1}$ such that $\P^{\mu}\cap D_m=P\cup Q$ and $\P^{\mu-1}\cap D_m=\emptyset$ ($P,Q$ base points).
2. If $m=2\mu+1$ then involution is defined by two base spaces $\P^{\mu}_1$, $\P^{\mu}_2$ such that $\P^{\mu}_1\cap D_m=P$ and $\P^{\mu}_2\cap D_m=Q$ ($P,Q$ base points).
In both cases the involution generates a normal rational ruled surface $R_{m-1}$ of degree $m-1$, whose directrix curves lay on the space bases. We call them base curves.
Thus, let $\vhi$ be a harmonic involution on $R_{n-1}$. Let $\vhi_k$, $\vhi_l$ be the harmonic involutions induced on the directrix curves $D_k$, $D_l$. Let $\P^{k\mas}$, $\P^{k\menos}$, $\P^{l\mas}$, $\P^{l\menos}$ be the base spaces of $\vhi_k$ and $\vhi_l$. We know that the base spaces of $\vhi$ are $S\mas=\langle \P^{k\mas},\P^{l\mas} \rangle$, $S\menos=\langle
\P^{k\menos},\P^{l\menos} \rangle$. Let $C_{k\mas}$, $C_{k\menos}$, $C_{l\mas}$, $C_{l\menos}$ the corresponding base curves. Let $F$ be the variety of lines that join the points of the involution: $F=\overline{\{P\in \P^n/P\in \langle x,f(x)\rangle,x\in R_{n-1}\}}$. Let us identificate $F\cap S\mas$ and $F\cap S\menos$.
\[variedadderectas\]
The variety $F\cap S\mas$ $(F\cap S\menos)$ is a normal rational ruled surface of degree $k\mas+l\mas=n\mas$ $(k\menos+l\menos=n\menos)$ generated by the directrix curves $C_{k\mas}$ and $C_{l\mas}$ ($C_{k\menos}$ and $C_{l\menos}$). We call it base ruled surface $R_{n\mas-1}$ ($R_{n\menos-1}$).
[**Proof:**]{} Given a point $P\in R_{n-1}$, consider the line $r=\langle P,f(P) \rangle$ of $F$. $r$ meets $S\mas$ in a point $P\mas$ that corresponds to project $P$ from $S\menos$ onto $S\mas$. Thus, given a generator $f\in R_{n-1}$, the lines of $F$ defined by the points of $f$ meet $S\mas$ in a line $f\mas$; this line is the projection of $f$ from $S_b$. Moreover, since $f$ meets $D_k$ and $D_l$, its projection on $S\mas$ meets $C_{k\mas}$ and $C_{l\mas}$. In this way we see that the generator of $R_{n-1}$ project into lines joining $C_{k\mas}$ and $C_{l\menos}$, so $F\cap S\mas$ is the rational ruled surface defined by these directrix curves.
We saw that a harmonic involution on a normal rational ruled surface is defined by the involutions of the directrix curves $D_k$ and $D_l$. From this, we distinguish several types of involutions:
1. [*$\vhi_k$ and $\vhi_l$ are the identity.*]{}
Then the base spaces of $\vhi$ are the spaces $\P^k$ and $\P^l$ that contain the directrix curves. All the generators are invariants by $\vhi$ and the variety $F$ is the ruled surface $R_{n-1}$.
2. [*$\vhi_k$ and $\vhi_l$ are not trivial.*]{}
1. $n-1=2\lambda$ (even).
1. $k=2\mu, l=2(\lambda-\mu)$.
Then the involutions on $D_k$ and $D_l$ have the following base spaces and base curves:
$C_{\mu}\in \P^{\mu}$, $C_{\mu-1}\in \P^{\mu-1}$, with $P_k,Q_k\in C_{\mu}$ base points of $D_k$.
$C_{\lambda-\mu}\in \P^{\lambda-\mu}$, $C_{\lambda-\mu-1}\in \P^{\lambda-\mu-1}$, with $P_l,Q_l\in C_{\lambda-\mu}$ base points of $D_l$.
Then, the base spaces of $\vhi$ are:
1. Case A:
$S\mas=\langle \P^{\mu},\P^{\lambda-\mu}\rangle=\P^{\lambda+1}\ni P_k,Q_k,P_l,Q_l$.
$S\menos=\langle \P^{\mu-1},\P^{\lambda-\mu-1}\rangle=\P^{\lambda-1}$.
Where the generators $f_P,f_Q\in \P^{\lambda+1}$ are fixed.
2. Case B: \[B1\]
$S\mas=\langle \P^{\mu},\P^{\lambda-\mu-1}\rangle=\P^{\lambda}\ni P_k,Q_k$.
$S\menos=\langle \P^{\mu-1},\P^{\lambda-\mu}\rangle=\P^{\lambda}\ni P_l,Q_l$.
Where the generators $f_P,f_Q$ are invariant (not fixed).
2. $k=2\mu+1, l=2(\lambda-\mu)-1$.
Then the involutions on $D_k$ and $D_l$ have the following base spaces and base curves:
$P_k\in C_{\mu}\in \P^{\mu}$, $Q_k\in C_{\mu}\in \P^{\mu}$, with $P_k,Q_k$ base points of $D_k$.
$P_l\in C_{\lambda-\mu}\in \P^{\lambda-\mu}$, $Q_l\in C_{\lambda-\mu}\in \P^{\lambda-\mu}$, with $P_l,Q_l$ base points of $D_l$.
Then, the base spaces of $\vhi$ are:
1. Case C:
$S\mas=\langle \P^{\mu},\P^{\lambda-\mu-1}\rangle=\P^{\lambda}\ni P_k,P_l$.
$S\menos=\langle \P^{\mu},\P^{\lambda-\mu-1}\rangle=\P^{\lambda}\ni Q_k,Q_l$.
Where the generators $f_P\in \P^a,f_Q\in S\menos$ are fixed.
2. Case B: (similar to case \[B1\]).
$S\mas=\langle \P^{\mu},\P^{\lambda-\mu-1}\rangle=\P^{\lambda}\ni P_k,Q_l$.
$S\menos=\langle \P^{\mu},\P^{\lambda-\mu-1}\rangle=\P^{\lambda}\ni P_l,Q_k$.
Where the generators $f_P,f_Q$ are invariant (not fixed).
2. $n-1=2\lambda-1$ even.
Then the curves $D_k$ and $D_l$ have degrees $k=2\mu$ and $l=2(\lambda-\mu)-1$. The base spaces and base curves are:
$C_{\mu}\in \P^{\mu}$, $C_{\mu-1}\in \P^{\mu-1}$, with $P_k,Q_k\in C_{\mu}$ base points of $D_k$.
$P_l\in C_{\lambda-\mu-1}\in \P^{\lambda-\mu-1}$, $Q_l\in C_{\lambda-\mu-1}\in
\P^{\lambda-\mu-1}$, with $P_l,Q_l$ base points of $D_l$.
In any case, the base spaces of $\vhi$ are:
1. Case D:
$S\mas=\langle \P^{\mu},\P^{\lambda-\mu-1}\rangle=\P^{\lambda}\ni P_k,Q_k,P_l$.
$S\menos=\langle \P^{\mu},\P^{\lambda-\mu-1}\rangle=\P^{\lambda}\ni Q_l$.
Where $f_P$ is a fixed generator and $f_Q$ are an invariant generator.
Let $\vhi:R_{n-1}\lrw R_{n-1}$ a harmonic involution over the rational normal ruled surface $R_{n-1}\subset \P^n$. By the proposition \[equivalencia\] we know that $H^0(I_{R_{n-1}}(2))$ is a base-harmonic system; that is, $H^0(I_{R_{n-1}}(2))=H^0(I_{R_{n-1}}(2))_h\oplus H^0(I_{R_{n-1}}(2))_b$. Let us see the dimension of these spaces. We know that $h^0(I_{R_{n-1}}(2))=\left(^{n-1}_{\;\;\, 2}\right)$. We will treat each case separated:
1. All generators are invariant by the involution.
We use the proposition \[cuadricasbase\]. In this case $F=R_{n-1}$, $F\cap \P^a=D_k$ and $F\cap \P^b=D_l$. Thus, $$h^0(I_{R_{n-1}}(2))_b=h^0(I_{R_{n-1}}(2))-h^0(I_{D_k}(2))-h^0(I_{D_l}(2))=kl$$ and $$h^0(I_{R_{n-1}}(2))_h=h^0(I_{R_{n-1}}(2))-h^0(I_{R_{n-1}}(2))_b=\left(^{n-1}_{\;\;\,
2}\right)-kl$$
2. The generic generator is not invariant by the involution.
We use the proposition \[cuadricasbase\]. But in this case, $F\cap S\mas=R_{n\mas-1}$ and $F\cap
S\menos=R_{n\menos}$. Then we have: $$h^0(I_{R_{n-1}}(2))_b=h^0(I_{F}(2))-h^0(I_{R_{n\mas-1}}(2))-h^0(I_{R_{n\menos-1}}(2))
\eqno(2)$$
A quadric containing $R_{n-1}\cup R_{n\mas-1}$ meets each line of $F$ in three points. Then such quadric contains $F$, so $H^0(I_F(2))=H^0(I_{R_{n-1}\cup R_{n\mas-1}}(2))$. Consider the exact sequence: $$0\lrw H^0(I_{R_{n-1}\cup R_{n\mas-1}}(2))\lrw H^0(I_{R_{n-1}}(2))\stackrel{\alpha}{\lrw}
H^0(\Te_{R_{n\mas-1}}(2-Y))$$ where $Y=R_{n-1}\cap R_{n\mas-1}$. Then: $$h^0(I_F(2))\geq h^0(I_{R_{n-1}}(2))-h^0(\Te_{R_{n\mas-1}}(2-R_{n-1}\cap R_{n\mas-1}))$$ and applying $(2)$ we obtain in each case:
1. ($n-1=2\lambda, S\mas=\P^{\lambda-1},S\menos=\P^{\lambda+1},f_Q,f_P$ fixed generators.) $$h^0(I_{R_{n-1}}(2))_b\geq \lambda (\lambda-1)$$
2. ($n-1=2\lambda, S\mas=\P^{\lambda}_1,S\menos=\P^{\lambda}_2,f_Q,f_P$ invariant (not fixed) generators, with $f_P\cap \P^{\lambda}_i=P_i$, $f_Q\cap P^{\lambda}_i=Q_i$.) $$h^0(I_{R_{n-1}}(2))_b\geq \lambda (\lambda-1)$$
3. $(n-1=2\lambda, S\mas=\P^{\lambda}_1,S\menos=\P^{\lambda}_2,f_Q,f_P$ fixed generators, with $f_P\in \P^{\lambda}_1$, $f_Q\in \P^{\lambda}_2$.) $$h^0(I_{R_{n-1}}(2))_b\geq \lambda (\lambda-1)+1$$
4. $(n-1=2\lambda-1, S\mas=\P^{\lambda-1},S\menos=\P^{\lambda},f_Q$ fixed generator in $\P^{\lambda}$, and $f_q$ invariant generator, with $f_Q\cap P^{\lambda-1}=Q_1$.) $$h^0(I_{R_{n-1}}(2))_b\geq (\lambda-1)^2$$
Now, let us compute the harmonic quadrics.
Let $E_k$ the set of $k+1$ generic points in $D_k$ and $E_l$ the set of $l+1$ generic points on $D_l$. Note that a harmonic quadric that passes through a point of $R_{n-1}$ passes through the image point too. From this, a quadric passing through $E_k (E_l)$ meets $D_k (D_l)$ in $2k+2 (2l+2)$ points because $D_k (D_l)$ is invariant by the involution. Moreover, a harmonic quadric that contains $D_k$ and $D_l$, contains the invariant (not fixed) generators too, because they meet each space base in a point.
Finally, a harmonic quadric containing $D_k$ and $D_l$ and passing through $m$ generic points of $R_{n-1}$ contains their images ($2m$ points) and the corresponding $2m$ generators. Let $E_m$ be $m$ generic points of $R_{n-1}$ and let $E$ be $E_k\cup E_l \cup E_m$.
If $Q$ is a harmonic quadric passing through the points of $E$, then $D_k\cup
D_l\cup \{$invariant generators$\}\cup 2mf\subset Q\cap R_{n-1}$. If $2m>2(n-1)-(k+l)-$number of invariant generators, then $R_{n-1}\subset Q$ and we have the exact sequence: $$0\lrw H^0(I_{R_{n-1}}(2))_h\lrw H^0(I_{\P^{n-1}}(2))_h\lrw H^0(\Te_E(2))$$ From this: $$h^0(I_{R_{n-1}}(2))_h\geq h^0(\Te_{\P^{n-1}}(2))_h-(n+1+m)$$
In each case we obtain:
1. There are not invariant generators. Taking $m=\lambda+1$ we have: $$h^0(I_{R_{n-1}}(2))_h\geq \left(^{\lambda+1}_{\;\;\, 2}\right) + \left(^{\lambda+3}_{\;\;\,
2}\right)-(3\lambda+3)$$
2. There are two invariant generators. Taking $m=\lambda$ we have: $$h^0(I_{R_{n-1}}(2))_h\geq \left(^{\lambda+2}_{\;\;\, 2}\right) + \left(^{\lambda+2}_{\;\;\,
2}\right)-(3\lambda+2)$$
3. There are not invariant generators. Taking $m=\lambda+1$ we have: $$h^0(I_{R_{n-1}}(2))_h\geq \left(^{\lambda+2}_{\;\;\, 2}\right) + \left(^{\lambda+2}_{\;\;\,
2}\right)-(3\lambda+3)$$
4. There is an invariant generator. Taking $m=\lambda$ we have: $$h^0(I_{R_{n-1}}(2))_h\geq \left(^{\lambda+1}_{\;\;\, 2}\right) + \left(^{\lambda+2}_{\;\;\,
2}\right)-(3\lambda+1)$$
We see that the sum of the bounds computed for the harmonic and base quadrics is the quadrics of $H^0(I_{R_{n-1}}(2))$ in all cases, so these bounds are reached in all cases, and we have the number of base quadrics:
\[baseregladaracional\]
Let $R_{n-1}\subset \P^n$ be a rational normal scroll of degree $n-1$. Let $\vhi:R_{n-1}\lrw R_{n-1}$ be a harmonic involution. Then we have the following cases:
1. All the generators are invariant. There are two directrix curves of base point $D_k,D_l$ with $k+l=n-1$. They lay on the base spaces $\P^k$, $\P^l$. $$h^0(I_{R_{n-1}}(2))_b=kl$$
2. There are two invariant (fixed or not):
1. $n-1=2\lambda$ ($n$ even).
1. The base spaces are $\P^{\lambda-1}$,$\P^{\lambda+1}$. There are two fixed generators in $\P^{\lambda+1}$.
$$h^0(I_{R_{n-1}}(2))_b=\lambda(\lambda-1)$$
2. The base spaces are $\P^{\lambda}$,$\P^{\lambda}$. There is a fixed generator in each of them.
$$h^0(I_{R_{n-1}}(2))_b=\lambda(\lambda-1)+1$$
3. The base spaces are $\P^{\lambda}$,$\P^{\lambda}$. There are not fixed generators.
$$h^0(I_{R_{n-1}}(2))_b=\lambda(\lambda-1)$$
2. $n-1=2\lambda-1$ ($n$ odd).
1. The base spaces are $\P^{\lambda-1}$,$\P^{\lambda}$. There is a fixed generator in $\P^{\lambda}$.
$$h^0(I_{R_{n-1}}(2))_b=(\lambda-1)^2$$
Involutions of the canonical curve of genus $\pi>4$. {#involucionesC}
----------------------------------------------------
We have investigated all the possible cases where the quadrics that contain a canonical curve are a base-harmonic system:
\[uno\]
The unique cases where the system of quadrics containing a canonical curve $C_{\K}$ of genus $\pi$, $\pi>4$ are a base-harmonic system respect to base spaces $\P^{g-1}$, $\P^{\pi-g-1}$ ($\pi\geq 2g-1>0
$) with $b$ independent base quadrics are:
1. $b=(g-1) (\pi-g-2)$
1. If $\pi=2g,2g+1,2g+2$ and the curve has a $g^1_3$ or an involution of genus $g$, or both of them; except if $\pi=6$ and $g=2$ when $C_{\K}$ can have a $g^2_5$.
2. If $\pi\neq 2g,2g+1,2g+2$ and the curve $C_{\K}$ has a $\gamma^1_2$ of genus $g$.
2. $b=(g-1)(\pi-g-2)+1$, $\pi=2g-1r$ or $\pi=2g$ and $C_{\K}$ has a $g^1_3$ (not a $\gamma^1_2$).
3. $b=(g-1)(\pi-g-1)$ and $C_{\K}$ has a $g^1_3$ (not a $\gamma^1_2$).
\[coro1\]
The unique involutions on a trigonal canonical curve of genus $\pi$, $\pi>4$ are of genus $\frac{\pi}{2},\frac{\pi-1}{2}$ or $\frac{\pi-2}{2}$.
[**Proof:**]{} Let $C_{\K}\subset \P^n$ be a trigonal canonical curve and let $R_{\pi-2}$ be the ruled surface of trisecants. Suppose that $C_{\K}$ has an involution of genus $g$. Then the system of quadrics $\P(H^0(I_{C_{\K}}(2)))$ is a base-harmonic system with $(g-1)(\pi-g-2)$ independent base quadrics. Since $R_{\pi-2}=\bigcap_{Q\supset C_{\K}}Q$, we have a harmonic involution over $R_{\pi-2}$. By Proposition \[baseregladaracional\] we know that $\pi=2g$, $\pi=2g+1$ or $\pi=2g+2$ and the conclusion follows.
\[coro2\]
The unique involutions on a smooth quintic plane curve are of genus $2$.
\[casoparticular\]
Let $C_{\K}\subset \P^{\pi-1}$ a canonical curve of genus $\p$, $\pi>4$. Then $C_{\K}$ has an involution of genus $1$ if and only if the quadrics of $\P(H^0(I_{C_{\K}}(2)))$ are a base-harmonic system respect to a point and a space $\P^{\pi-2}$ without base quadrics.
[**Proof:**]{} If $C_{\K}$ has an involution of genus $1$, we know that the involution generates an elliptic cone, the system of quadrics $\P(H^0(I_{C_{\K}}(2)))$ is harmonic and it hasn’t base quadrics respect to $\P^0$ and $\P^{\pi-2}$.
Conversely, if the system of quadrics $\P(H^0(I_{C_{\K}}(2)))$ is harmonic respect to $\P^0$ and $\P^{\pi-2}$, necessary it hasn’t base quadrics, because the quadric containing $C_{\K}$ are reducible. If $C_{\K}$ is not trigonal we have an involution of genus $1$ in $C_{\K}$. If $C_{K}$ is trigonal, by Corollary \[coro1\], $\pi=2,3,4$. But we have supposed that $\pi>4$.
\[fundamental\]
Let $C_{\K}\subset \P^{\pi-1}$ be the canonical curve of genus $\pi$, with $\pi>4$. If $C_{\K}$ has an involution of genus $g$ then $\pi\geq 2g-1$ and the quadrics of $\P(H^0(I_{C_{\K}}(2)))$ are a base-harmonic system respect to the base spaces $\P^{g-1}$, $\P^{\pi-g-1}$ that contains $(g-1)(\pi-g-2)$ independent base quadrics. Conversely, these conditions are sufficient to grant the existence of an involution, except when:
1. $\pi=6,g=2$ and $C_{\K}$ has a $g^2_5$; or
2. $\pi=2g$,$2g+1$ or $2g+2$ and $C_{\K}$ is trigonal.
*Let us study what happens at the two exceptions:*
1. Suppose that $C_{\K}$ is a canonical curve of genus $6$ with a $g^2_5$, that is,it is isomorphic to a smooth plane curve of degree $5$. Suppose that the quadrics of $\P(H^0(I_{C_{\K}}(2)))$ are a base-harmonic system respect to the base spaces $\P^1$ and $\P^3$. It induces a harmonic involution on the Veronese surface and then, an involution on the plane. Obviously, the generic plane curve of degree $5$ of the plane is not invariant by this involution. So in this case the hypothesis of the above theorem are not sufficient.
However, there are smooth quintic plane curves invariant by an involution. For example, we can take the quintic curve $f(x_0,x_1)-x_2^4x_0=0$ on the plane, where $f(x_0,x_1)$ is a generic homogeneous polynomial of degree $5$. This curve is smooth an it’s invariant by the involution $$\begin{array}{ccc}
{x_0\lrw x_0;}&{x_1\lrw x_1;}&{x_2\lrw -x_2}\\
\end{array}$$
2. Now, suppose that $C_{\K}$ is trigonal. Then it lies on a rational ruled surface $S_e=\P(\Te_{P^1}\oplus \Te_{P^1}(-e))$ in the linear systems $|3X_0+af|$. The canonical embedding is obtained by the linear system $X_0+(a-e-2)f$ on the ruled surface.
If $\P(H^0(I_{C_{\K}}(2)))$ is a base-harmonic system then it defines a harmonic involution on the ruled surface $S_e$. Moreover, we have an induced involution in the linear system $|3X_0+a f|$. The generic curve of this linear system is not invariant by the involution. We see that the hypothesis of the theorem are not sufficient.
On the other hand, there are smooth curves on these linear systems invariant by the involution. Let us see an example. Consider the rational ruled surface $S_0\cong \P^1\times \P^1$ with coordinates $[(x_0,x_1),(y_0,y_1)]$. We can take the curve on $S_0$ with equation: $$x_0^ny_0^3-x_0^ny_0y_1^2+x_1^ny_1^3+x_1^ny_0^2y_1=0$$ with $n\geq 5$ even. This is a smooth curve of type $(3,n)$ on the linear system $3X_0+nf$. Moreover it is invariant by the involution $$\begin{array}{cccc}
{x_0\lrw x_0;}&{x_1\lrw -x_1;}\\
{y_0\lrw y_0;}&{y_1\lrw -y_1}\\
\end{array}$$
[77]{}
Preprint. math.AG/0006204.
Preprint. math.AG/0107114.
GTM, 52. Springer–Verlag, 1977.
[^1]: Supported by an F.P.U. fellowship of Spanish Government
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Black holes are extreme expressions of gravity$^1$. Their existence is predicted by Einstein’s theory of general relativity and is supported by observations$^{2,3,4}$. Black holes obey quantum mechanics and evaporate spontaneously$^5$. Here it is shown that a mass rate $R_f\sim 3\times 10^{-8} (M_0/M)^{1/2}$ $M_0$ yr$^{-1}$ onto the horizon of a black hole with mass $M$ (in units of solar mass $M_0$) stimulates a black hole into rapid evaporation. Specifically, $\sim 3 M_0$ black holes can emit a large fraction of their mass, and explode, in $M/R_f \sim 3\times 10^7 (M/M_0)^{3/2}$ yr. These stimulated black holes radiate a spectral line power $P \sim 2\times 10^{39} (M_0/M)^{1/2}$ erg s$^{-1}$, at a wavelength $\lambda \sim 3\times 10^5 (M/M_0)$ cm. This prediction can be observationally verified.'
address: 'Kapteyn Astronomical Institute, University of Groningen, P.O. Box 800, 9700 AV Groningen, The Netherlands; [email protected]'
author:
- MARCO SPAANS
title: STIMULATED BLACK HOLE EVAPORATION
---
Quantum Mechanics of Black Holes
================================
Black holes have been studied for about a century, ever since Schwarzschild found the metric that solves the Einstein equation of general relativity for a point mass$^2$. The Schwarzschild metric is appropriate for a black hole of mass M, zero charge and zero spin1. Characteristic of a black hole are its horizon (a trapped surface) and its singularity. Quantum mechanics and thermodynamics suggest that black holes have a temperature $T \sim 10^{-7} (M_0/M)$ K. This allows them to lose mass through Hawking radiation$^4$.
The context of this work is a black hole and the interaction with its environment. This interaction has quantum mechanical significance because a black hole and its surroundings are not disjoint a priori. Instead, together they establish, through observation, whether a horizon is a net absorber or emitter of mass. Useful quantities to recall below are the Planck length, $l_p \sim 2\times 10^{-33}$ cm, the Planck time, $t_p \sim 5\times 10^{-44}$ s, the Planck mass, $m_p \sim 2\times 10^{-5}$ g and the solar mass, $M_0 \sim 2\times 10^{33}$ g. A black hole’s Schwarzschild radius is $R_s \sim 3\times 10^5 (M/M_0)$ cm, $c$ denotes the speed of light, $G$ is the gravitational constant and $h$ is Planck’s constant.
The first point to consider is the black hole singularity$^{6,7}$. Its natural habitat is Planckian space-time, characterized by $l_p$, $t_p$ and $m_p$. To explore the quantum properties of singularities, the minimal self-consistent approach is to endow general relativity’s smooth space-time manifold with the discrete units $l_p$, $t_p$ and $m_p$, and to pursue their consequences$^{6,8}$. The only dynamical quantity that can be constructed from them is the mass rate $m_p/t_p = c^3/G \sim 6\times 10^{12}$ $M_0$ yr$^{-1}$, which is independent of $h$ and huge. However, in four dimensions the volume $V \sim ct_p l_p^3$ of a Planckian singularity is adjusted to $aV = (a^{1/4}l_p)^4$, with $a = R_s/l_p$, as follows. An observer cannot localize a singularity to better than the linear size of a horizon, anywhere along its surface and at any time, since a singularity can be stretched radially to points on the horizon without an observer able to know. Therefore, the local unit $l_p$ acts through the non-local measure $a^{1/4}l_p$, when one considers the dynamics of a Planckian space-time region enclosed by a horizon. The mass rate $m_p/t_p$ scales as length$^{-2}$ and the horizon-limited quantum rate becomes
$$R_f = (4\pi a^{1/2})^{-1}m_p/t_p \sim 3\times 10^{-8} (M_0/M)^{1/2}\ M_0\ {\rm yr}^{-1},$$
for a sphere and proportional to $(hc^{13}/G^5)^{1/4}$. So a black hole endowed with quantum discreteness accommodates mass, by absorption or emission, at a maximum rate $R_f << m_p/t_p$.
The second point to consider is the interplay between $R_f$ and a mass rate onto the horizon, $R$. The mass rates $R$ and $R_f$ can be compared directly when both are measured by the same observer. E.g., one that is in an asymptotically flat region and tracks $M$ as a function of time, $t$, through various astronomical observations. If $dM/dt = R > R_f$ then not all mass should have been added to the black hole, according to the observer. Of course, the flow $R$ is oblivious to this constraint, it obeys space-time curvature and crosses the horizon. Both assertions have to be true, though. They are, in a universe where time is relative, if the black hole eventually loses the excess mass gained by an $R > R_f$ event, provided that the latter event directly causes the required evaporation. This solution implies the existence of a spontaneous and a stimulated evaporation state for a black hole. Stimulated evaporation must proceed at $R_f$, i.e., at a mass rate that reciprocates a maximum absorption trigger with ditto emission. Also, classically the observer detects just black hole mass and has no information on how emitted particles connect to prior accretion. Hence, if activated, stimulated evaporation persists since an observer can only confirm unambiguously that the excess mass acquired by an $R > R_f$ event has been lost if the black hole vanishes.
Define $\psi_+$ to be a conventional black hole and $\psi_-$ to be a stimulated black hole evaporating at $R_f$. The superposition $\psi = (1-\alpha )\psi_+ + \alpha\psi_-$ represents a black hole as a dynamical Planckian space-time region enclosed by a horizon that interacts with an environment. As long as $R-R_f\le 0$ one has $\alpha = 0$, while $R-R_f > 0$ stimulates $\alpha = 0 \rightarrow 1$. A stimulated black hole is not a thermal emitter because power $P = R_fc^2$ scales as $M^{-1/2}$, instead of the usual $M^{-2}$, and is strongly boosted compared to the blackbody rate$^5$. Rather, one has a spectral line with particles of wavelength $\lambda \sim R_s$ across a width $\delta\lambda \sim \lambda /2\pi$, set by horizon geometry as this is the only degree of freedom. The above also pertains to black holes with charge $q$ and angular momentum per unit mass $s$ if the radius $r_+ = M + (M^2 - q^2 - s2)^{1/2}$ is used$^1$. The identity of matter need not be preserved by Planckian space-time, and a traversable wormhole or quantum hair is not part of $\psi$ Indeed, just $l_p$, $t_p$ and $m_p$ are used to describe space-time. Yet, $\alpha = 0 \rightarrow 1$ constitutes an extraction of mass forced by the environment. This enriches the information paradox$^9$.
Implications and Discussion
===========================
Most black holes in the universe comprise stellar origin ones of $\sim 3-30 M_0$. These evaporate in $M/R_f \sim 3\times 10^7 (M/M_0)^{3/2}$ yr if $R > R_f$ holds for brief periods only. But, $R_f \sim 5\times 10^{-9}$ $M_0$ yr$^{-1}$ for a $30 M_0$ black hole and $R > R_f$ is easily achieved, e.g., by an accretion disk$^{8,10}$. The latter is often present and typically radiates at an efficiency $\epsilon \sim 0.1$, with $\epsilon Rc^2$ below the radiation pressure limit of the Eddington luminosity, $L_{edd} \sim 10^{38} (M/M_0)$ erg s$^{-1}$. Also, $M/R_f \sim 5\times 10^9$ yr is cosmologically long and stimulated evaporation is thus a small effect for $\sim 30 M_0$ black holes. Much rarer intermediate mass black holes in globular clusters ($M \sim 10^{2-5} M_0$), and supermassive black holes in centers of galaxies ($M \sim 10^{5-10} M_0$), are also prone to stimulated effects since $R_f$ is even smaller. While $R > R_f$ holds for cosmologically long periods and $M/R_f$ exceeds the age of the universe. Hence, stimulated mass loss is tiny. Primordial black holes of $\sim 10^{15}$ g survive until today. These have $R_f \sim 2$ $M_0$ yr$^{-1}$ and $R > R_f$ appears to be unattainable.
Fortunately, $\sim 3 M_0$ black holes have $R_f \sim 2\times 10^{-8}$ $M_0$ yr$^{-1}$. Rates $R > R_f$ are realisable, but sporadically because $\epsilon R_fc^2 \sim L_{edd}$ and radiation pressure can shut off accretion through a disk$^{10}$. So $R < R_f$ most of the time and $M/R_f \sim 10^8$ yr, which is relatively short. This suggests that a significant fraction of black holes formed at $\sim 3 M_0$ can evaporate, producing a power $P \sim 2\times 10^{39} (M_0/M)^{1/2}$ erg s$^{-1}$. When $M < 10^{-2} M_0$ a stimulated black hole emits photons above the $\sim 10$ MHz cut-off of the earth’s ionosphere. In its final year it converts $M \sim 10^{-5} M_0$ (so $\sim 10^{49}$ erg) into $>10$ GHz photons. The last second produces $P > 10^{44}$ erg s$^{-1}$ at $>10^{15}$ Hz, comparable to an entire galaxy. These explosive final stages should provide valuable probes of the standard model of particle physics.
[0]{}
C. W. Misner, K. P. Thorne and J.A. Wheeler, [*Gravitation*]{} (Freeman 1973).
K. Schwarzschild, [*Sitzber. Deut. Akad. Wiss.*]{} [**1**]{} (1916) 189.
J. Magorrian et al., [*AJ*]{} [**115**]{} (1998) 2286.
B. P. Abbott et al., [*Phys. Rev. Lett.*]{} [**116**]{} (2016) 061102-1.
S. W. Hawking, [*Nature*]{} [**248**]{} (1974) 30.
J. A. Wheeler, [*Geometrodynamics*]{} (Academic Press 1962).
M. Spaans, [*Int. J. Mod. Phys. D*]{} [**22**]{} (2013) 1330022-1.
M. Spaans, [*Quantum Spacetime: Mimicry of Paths and Black Holes*]{} (Nova 2015).
S. W. Hawking, [*Phys. Rev D*]{} [**14**]{} (1976) 2460.
G. B. Rybicki and A. P. Lightman, [*Radiative Processes in Astrophysics*]{} (Wiley-Interscience 1979).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A description is given of all primitive $\d$-series mod $p$ of order $1$ which are eigenvectors of all the Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, and which are $\d$-Fourier expansions of $\d$-modular forms of arbitrary order and weight $w$ with $deg(w)=\k\geq 0$; this set of $\d$-series is shown to be in a natural one-to-one correspondence with the set of series mod $p$ (of order $0$) which are eigenvectors of all the Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$ and which are Fourier expansions of (classical) modular forms of weight $\equiv \k+2$ mod $p-1$.'
address: |
University of New Mexico\
Albuquerque, NM 87131
author:
- Alexandru Buium and Arnab Saha
title: Hecke operators on differential modular forms mod $p$
---
ø § i[\_1]{}
\[section\]
\[section\] \[theorem\][Corollary]{} \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Definition]{} \[theorem\][Remark]{} \[theorem\][**Example**]{}
Introduction
============
This present paper is a direct continuation of [@eigen] and , implicitly, of a series of papers devoted to the study of arithmetic differential equations [@char; @difmod; @Barcau; @book; @dcc; @igusa]; however, for the convenience of the reader, the present paper is written so as to be logically independent of [@eigen] and of the other above cited papers. Rather, we will quickly review the relevant material from some of these papers as needed.
The plan of this Introduction is as follows. We begin by quickly recalling the basic definitions of this theory following [@char; @book]. For more details on some of these definitions we refer to the body of the present paper. Then we will state our main result (Theorem \[maint\]). Finally we will make some comments on the larger picture and motivations beyond this theory.
$\d$-functions [@char; @book]
-----------------------------
A map $\d:A \ra B$ from a ring $A$ into a $p$-torsion free $A$-algebra $B$ is called a [*$p$-derivation*]{} if the map $\phi:A \ra B$, $\phi(x)=x^p+p\d x$, is a ring homomorphism. When $\d$ is given $\phi$ will always have the meaning above. A ring equipped with a $p$-derivation will be refered to as a [*$\d$-ring*]{}. Denote by $R$ the completion of the maximum unramified extension of the ring of $p$-adic integers. Set $k=R/pR$, $K=R[1/p]$, let $\phi:R \ra R$ be the unique ring automorphism lifting the $p$-power Frobenius $F:k \ra k$, and denote by $\d:R \ra R$ the $p$-derivation $\d x=\frac{\phi(x)-x^p}{p}$. This makes $R$ a $\d$-ring and this $\d$-ring structure on $R$ is unique. For any affine smooth scheme $V \subset {\mathbb A}^m$ over $R$ a function $f:V(R)\ra R$ will be called a [*$\d$-function of order $r$*]{} on $V$ [@char] if there exists a restricted power series $\Phi$ in $m(r+1)$ variables, with $R$-coefficients such that $f(a)=\Phi(a,\d a,...,\d^r a)$, for all $a\in V(R)\subset R^m$. (Recall that a power series is called [*restricted*]{} is its coefficients tend to $0$ $p$-adically.) We denote by $\cO^r(V)$ the ring of $\d$-functions of order $r$ on $V$. This concept can be naturally extended to the non-affine case [@char] but we will not need this extension in the present paper.
$\d$-modular forms [@difmod; @book]
-----------------------------------
Let $N>4$ be an integer coprime to $p$ and let $X$ be either the affine modular curve $Y_1(N)$ over $R$ or its ordinary locus $Y_1(N)_{ord}$ (i.e. the locus where the Eisentein form $E_{p-1}$ is invertible). Let $L$ be the line bundle on the complete modular curve $X_1(N)$ over $R$ such that the global sections of the powers $L^{\otimes \k}$, $\k \geq 0$, are the classical modular forms (on $\Gamma_1(N)$) of weight $\k$ over $R$ and let $V\ra X$, $V:=Spec\ \bigoplus_{\k \in {\mathbb Z}}L^{\otimes \k}$, be the natural ${\mathbb G}_m$-torsor associated to the restriction of $L$ to $X$. A [*$\d$-modular function of order $r$*]{} (on $\Gamma_1(N)$) [@difmod; @book] will mean a $\d$-function of order $r$ on $V$, i.e. an element of $\cO^r(V)$. Let $W={\mathbb Z}[\phi]$ be the polynomial ring in the variable $\phi$. Then the multiplicative monoid $W$ naturally acts on $R^{\times}$; for $w \in W$ and $\lambda\in R^{\times}$ we write $(w,\lambda)\mapsto \lambda^w$ for the action. Evaluation at $\phi=1$ defines a ring homomorphism $deg:W={\mathbb Z}[\phi]\ra {\mathbb Z}$. A $\d$-modular function $f\in \cO^r(V)$ will be called a [*$\d$-modular form of weight*]{} $w \in W$ if $f(\lambda \cdot a)=\lambda^w f(a)$ for $a\in V(R)$ and $\lambda \in R^{\times}$, where $\lambda \cdot a$ is defined via the ${\mathbb G}_m$-action on $V$.
$\d$-Fourier expansions
-----------------------
Any $\d$-modular function of order $r$ has a natural [*$\d$-Fourier expansion*]{} in the ring of [*$\d$-series*]{} $R((q))[q',...,q^{(r)}]\h$ where $q,q',...,q^{(r)}$ are variables, $R((q)):=R[[q]][1/q]$, and the upper $\h$ means here (and everywhere later) [*completion in the $p$-adic topology*]{}.
There are unique $p$-derivations $\d$ from $R((q))[q',...,q^{(r)}]\h$ to $R((q))[q',...,q^{(r+1)}]\h$ extending $\d$ on $R$ and such that $\d q=q'$, $\d q'=q''$, etc. The $\d$-Fourier expansion maps are compatible with the classical Fourier expansion maps and commute with $\d$. Recall that for $\k \in {\mathbb Z}_{\geq 0}$ the classical Hecke operators $T_{\k+2}(n)$ (with $n \geq 1$, $(n,p)=1$) and $T_{\k+2}(p)$ act on $R((q))$. We have an induced action of $T_{\k+2}(n)$, $T_{\k+2}(p)$ on $k((q))$; clearly $T_{\k+2}(p)$ on $k((q))$ coincides with Atkin’s operator $U$ on $k((q))$, defined by $U(\sum a_n q^n)=\sum a_{np}q^n$. A series $\varphi \in k((q))$ will called [*primitive*]{} if $U \varphi=0$. A $\d$-series in $k((q))[q',...,q^{(r)}]$ will be called [*primitive*]{} if its image in $k((q))$ under the specialization $q'=...=q^{(r)}=0$ is primitive. One can define Hecke operators $T_{\k}(n)$, $pT_{\k}(p)$ on $R((q))[q',...,q^{(r)}]\h$ (where $pT_{\k}(p)$ is only “partially defined" i.e. defined on an appropriate subspace); cf. Sections 2 and 3 below for all the relevant details. These operators induce operators $T_{\k}(n)$, “$pT_{\k}(p)$" on $k((q))[q',...,q^{(r)}]$ (where “$pT_{\k}(p)$" is only “partially defined" i.e. defined on an appropriate subspace; the “ " signs are meant to remind us that the operator $T_{\k}(p)$ itself is not defined mod $p$).
Main result
-----------
The following is our main result; it is a consequence of Theorems \[dezastru\] and \[converseofdezastru\] in the body of the paper. Assume $X=Y_1(N)_{ord}$ and let $\k \in {\mathbb Z}_{\geq 0}$.
\[maint\] There is a one-to-one correspondence between the following sets of objects:
i\) Series in $qk[[q]]$ which are eigenvectors of all Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$, and which are Fourier expansions of classical modular forms over $k$ of weight $\equiv \k+2$ mod $p-1$;
ii\) Primitive $\d$-series in $k[[q]][q']$ which are eigenvectors of all Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, and which are $\d$-Fourier expansions of $\d$-modular forms of some order $r \geq 0$ and weight $w$ with $deg(w)=\k$.
This correspondence preserves the respective eigenvalues.
1\) As Theorems \[dezastru\] and \[converseofdezastru\] will show the correspondence in Theorem \[maint\] is given, on a computational level, by an entirely explicit formula (but note that the proof that this formula establishes the desired correspondence is [*not*]{} merely computational.) The formula is as follows. If $\varphi=\sum_{m \geq 1} a_m q^m \in k[[q]]$ is a series as in i) of the Theorem then $a_1 \neq 0$ and the corresponding $\d$-series in ii) is given by $$\varphi^{\sharp,2}:=\sum_{(n,p)=1}\frac{a_n}{n} q^n -\frac{a_p}{a_1} \cdot \left(\sum_{m\geq 1}a_m q^{mp}\right) \frac{q'}{q^p}+ e \cdot \left(\sum_{m\geq 1}a_m q^{mp^2}\right) \cdot \left(\frac{q'}{q^p}\right)^p\in k[[q]][q'],$$ where $e$ is $1$ or $0$ according as $\k$ is $0$ or $>0$. (The upper index $2$ in $\varphi^{\sharp,2}$ is meant to reflect the $p^2$ exponent in the right hand side of the above equality; later in the body of the paper we will encounter a $\varphi^{\sharp,1}$ series as well. The $\sharp$ sign is meant to reflect the link between these objects and the objects $f^{\sharp}$ introduced in [@eigen].)
2\) Theorem \[maint\] provides a complete description of primitive $\d$-series mod $p$ of order $1$ which are eigenvectors of all the Hecke operators and which are $\d$-Fourier expansions of $\d$-modular forms of arbitrary order. It would be desirable to have such a description in characteristic zero and/or for higher order $\d$-series. However note that all known examples (so far) of $\d$-modular forms of order $\geq 2$ which are eigenvectors of all Hecke operators have the property that their $\d$-Fourier expansion reduced mod $p$ has order $1$; by the way some of these forms play a key role in [@eigen; @dcc; @BP]. So it is reasonable to ask if it is true that [*any $\d$-modular form of order $\geq 1$ which is an eigenvector of all the Hecke operators must have a $\d$-Fourier expansion whose reduction mod $p$ has order $1$*]{}.
3\) Note that in ii) of the above Theorem one can take the order to be $r=1$ and the weight to be $w=\k$. Also note that the $\d$-modular forms in ii) above have, a priori, “singularities" at the cusps and at the supersingular points. Nevertheless, in the special case when the classical modular forms in i) above come from newforms on $\Gamma_0(N)$ over ${\mathbb Z}$ of weight $2$ one can choose the $\d$-modular forms in ii) of weight $0$, order $2$, and [*without singularities at the cusps or at at the supersingular points*]{}; this was done in [@eigen] where the corresponding $\d$-modular forms were denoted (at least in the “non-CL" case) by $f^{\sharp}$. These $f^{\sharp}$s played, by the way, a key role in the proof of the main results in [@BP] about linear dependence relations among Heegner points. It would be interesting to find analogues of the forms $f^{\sharp}$ in higher weights.
4\) One of the subtleties of the above theory is related to the fact that the operator “$pT_{\k}(p)$" is not everywhere defined. The failure of this operator to be everywhere defined is related to the failure of “the fundamental theorem of symmetric polynomials" in the context of $\d$-functions; cf. [@eigen; @dcc]. The domain of definition of “$pT_{\k}(p)$" will be the space of all $\d$-series for which the analogue of “the fundamental theorem of symmetric polynomials" holds; these $\d$-series will be called [*Taylor $\d-p$-symmetric*]{}. One of our main results will be a complete determination the space of Taylor $\d -p$-symmetric $\d$-series; cf. Theorems \[conv\] and \[infinitesums\].
Comments on $\d$-geometry [@book]
---------------------------------
The present paper fits into a more general program for which we refer to [@book]. Roughly speaking this program proposes to enrich (usual) algebraic geometry by replacing algebraic equations (i.e. expressions of the form $f=0$, $f$ a polynomial function) with arithmetic differential equations (i.e. expressions of the form $f=0$, $f$ a $\d$-function). This enriched geometry can be referred to as [*$\d$-geometry*]{}. One of the main motivations/applications of $\d$-geometry is the construction of certain quotients of (usual) algebraic curves by actions of (usual) correspondences. Such quotients fail to exist within (usual) algebraic geometry in the sense that the corresponding categorical quotients in (usual) algebraic geometry reduce to a point. On the contrary, in $\d$-geometry, one can construct a number of interesting such categorical quotients, e.g. the quotient of the modular curve $Y_1(N)$ by the action of the Hecke correspondences. The construction/underdstanding of the latter is based upon the theory of $\d$-modular forms.
On a more “philosophical" level note that $\d$-geometry and, more generally, $\Lambda$-geometry (which is a several prime generalization of $\d$-geometry) can be viewed as an incarnation of the “geometry over the field with one element"; cf. the Introduction to [@book] for remarks on the single prime case and [@f1] for a systematic explanation of this viewpoint in the several prime case.
On the other hand, from a more “pragmatic" point of view, we note that $\d$-geometry has applications to (usual) arithmetic geometry such as: matters surrounding the Manin-Mumford conjecture [@pjets; @dcc], congruences between (usual) modular forms [@difmod; @Barcau], and linear dependence relations among Heegner points [@BP].
Plan of the paper
-----------------
Sections 2 and 3 introduce Hecke operators $T_{\k}(n)$, $(n,p)=1$ and “$pT_{\k}(p)$" respectively, acting on $\d$-series. Section 4 gives the complete determination of the $\d$-series mod $p$ of order $1$ for which “the analogue of the fundamental theorem of symmetric polynomials" holds. Section 5 gives a multiplicity one theorem for $\d$-series which are eigenvectors of all Hecke operators. Section 6 begins with an overview of $\d$-modular forms [@difmod; @book] and Serre-Katz $p$-adic modular forms [@Katz]; then we use the multiplicity one result plus results in [@difmod; @book] and [@Katz] to prove results implying Theorem \[maint\].
Acknowledgment
--------------
While writing this paper the first author was partially supported by NSF grant DMS-0852591. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.
Hecke operators away from $p$
=============================
Classical Hecke operators
-------------------------
Throughout the paper the divisors of a given non-zero integer are always taken to be positive, the greatest common divisor of two non-zero integers $m,n$ is denoted by $(m,n)$, and we use the convention $(m,n)=n$ for $m=0$, $n \neq 0$. Fix throughout the paper an integer $N \geq 4$ and let $\epsilon:{\mathbb Z}_{> 0}\ra \{0,1\}$ be the “trivial primitive character" mod $N$ defined by $\epsilon(A)=1$ if $(A,N)=1$ and $\epsilon(A)=0$ otherwise.
For each integer $n \geq 1$ and each integer $N \geq 4$ consider the set $$\{(A,B,D); A,B,D \in {\mathbb Z}_{\geq 0}, AD=n, (A,N)=1, B<D\}$$ Triples $A,B,D$ will always be assumed to be in the set above. Recall (cf., say, [@Knapp]) the action of the $n$-th Hecke operator $T_{\k}(n)$ on classical modular forms $f=\sum_{m \geq 0}a_mq^m$ on $\Gamma_0(N)$ of weight $\k\geq 2$ with complex coefficients $a_m \in {\mathbb C}$ given by
$$\begin{array}{rcl}
T_{\k}(n)f & := & n^{\k-1}\sum_{A,B,D}D^{-\k} f(\zeta_D^B q^{A/D})\\
\ & \ & \ \\
& = & \sum_{m \geq 0}\left(
\sum_{A|(n,m)} \epsilon(A)A^{\k-1}a_{\frac{mn}{A^2}}\right) q^m.
\end{array}$$
Here $q=e^{2 \pi \sqrt{-1}z}$, $\zeta_D:=e^{2 \pi \sqrt{-1}/D}$.
Hecke operators $T_{\k}(n)$ on $\d$-series
------------------------------------------
Now assume $n$ and $N$ are coprime to $p$ and assume $q,q',q'',...,q^{(r)},...$ are indeterminates.
For each integer $\k \in {\mathbb Z}$ the [*Hecke*]{} operator $f \mapsto T_{\k}(n)f$ on $R((q))[q',...,q^{(r)}]\h$ is defined as follows. For $f=f(q,q',...,q^{(r)})$, $$\label{eqq}
T_{\k}(n)f:=n^{\k-1}\sum_{A,B,D}D^{-\k} f(\zeta_D^B q^{A/D}, \d(\zeta_D^B q^{A/D}),...,\d^r(\zeta_D^B q^{A/D})).$$
Here $\zeta_D=\zeta_n^{n/D}\in R$ where $\zeta_n \in R$ is a fixed primitive $n$-th root of unity and the right hand side of (\[eqq\]) is a priori in the ring $$\label{rring}
R((q_n))\h[q'_n,...,q_n^{(r)}]\h,\ \ q_n=q^{1/n}.$$ However, by [@book] Proposition 3.13, $$q_n',...,q_n^{(r)} \in R[q,q^{-1},q',...,q^{(r)}]\h$$ hence the ring (\[rring\]) equals $$R((q_n))\h[q',...,q^{(r)}]\h.$$ Since $T_{\k}(n)f$ is invariant under the substitution $q_n^{(i)}\mapsto \d^i(\zeta_n q_n)$ it follows that $T_{\k}(n)f \in R((q))\h[q',...,q^{(r)}]\h$. So the operators $T_{\k}(n)$ send $R((q))\h[q',...,q^{(r)}]\h$ into itself. As we shall see below for $n \geq 2$ the operators $T_{\k}(n)$ do [*not*]{} send $R[[q]][q',...,q^{(r)}]\h$ into itself.
The operators $T_{\k}(n)$ on $R((q))[q',...,q^{(r)}]\h$ induce operators still denoted by $T_{\k}(n)$ on $k((q))[q',...,q^{(r)}]$.
Recall the operator $V$ on $R((q))\h$ defined by $V(\sum a_nq^n)=\sum a_nq^{pn}$. It induces an operator still denoted by $V$ on $k((q))$.
For $r=0$, $T_{\k}(n)$ commute with the operator $V$ on $R((q))\h$.
Order $r=1$
-----------
We have the following formula for the Hecke action on $\d$-series of order $1$:
\[Tf\] Assume that $$\label{f}
f=\sum_{m,m'} a_{m,m'} q^m(q')^{m'}$$ where $m \in \bZ$, $m'\in \bZ_{\geq 0}$. Then we have the following congruence mod $(p)$: $$\label{cuani}
T_{\k}(n)f \equiv \sum_{m,m'} \left( \sum_{A | (n,m)} n^{-m'}\epsilon(A)A^{\k+2m'-1}
a_{\frac{mn}{A^2}-m'p,m'} \right) q^{m-m'p}(q')^{m'}.$$
[*Proof*]{}. Note that $$\label{ord1}
\begin{array}{rcl}
\d(\zeta_D^B q^{A/D}) & = & \frac{1}{p}[\phi(\zeta_D^B q^{A/D})-(\zeta_D^B q^{A/D})^p]\\
\ & \ & \ \\
\ & = & \frac{1}{p} [\zeta_D^{Bp} (q^p+pq')^{A/D}-\zeta_D^{Bp} q^{Ap/D}]\\
\ & \ & \ \\
\ & \equiv & \frac{A}{D}\zeta_D^{Bp} q^{(A-D)p/D} q'\ \ \ mod\ (p).\end{array}$$ Then the formula in the statement of the Proposition follows by a simple computation, using the fact that $$\sum_{B=0}^{D-1} \zeta_D^{m+m'p}$$ is $D$ or $0$ according as $D$ divides or does not divide $m+m'p$.
\[buium\] Let $$\label{theseries}
\overline{f}=\sum_{m'} \overline{f}_{m'}(q) \left( \frac{q'}{q^p} \right)^{m'}\in k((q))[q'],\ \ \ \overline{f}_{m'}(q)\in k((q)).$$ Then for any integer $\k$ and any integer $n \geq 1$ coprime to $p$ we have: $$T_{\k}(n)\overline{f}= \sum_{m'} n^{-m'}(T_{\k+2m'}(n)\overline{f}_{m'}(q))\left( \frac{q'}{q^p} \right)^{m'}.$$ In particular for $\overline{\lambda}_n \in k$ we have $T_{\k}(n)\overline{f}= \overline{\lambda}_n \overline{f}$ if and only if $$T_{\k+2m'} (n) \overline{f}_{m'}= n^{m'}\overline{\lambda}_n \overline{f}_{m'}\ \ \text{for all}\ \ m'\geq 0.$$
[*Proof*]{}. This follows immediately from Proposition \[Tf\].
Let us say that a series in $k((q))[q',...,q^{(r)}]$ is [*holomorphic at infinity*]{} if it belongs to $k[[q]][q',...,q^{(r)}]$. Also denote by $v_p$ the $p$-adic valuation on ${\mathbb Z}$.
\[echelon\] Assume that, for a given $\k\in {\mathbb Z}$ the series $\overline{f}\in k[[q]][q']$ has the property that $T_{\k}(n)\overline{f}$ is holomorphic at infinity for all $n \geq 1$ coprime to $p$. Then $\overline{f}$ has the form $$\label{wash}
\overline{f}(q,q') = \varphi_0(q)+\sum_{m'\geq 1} (V^{v_p(m')+1} (\varphi_{m'}(q)))\left( \frac{q'}{q^p}\right)^{m'},$$ with $$\label{hair}
\varphi_0 \in k[[q]],\ \ \
\varphi_{m'}(q) \in q^{m'/p^{v_p(m')}}k[[q]]\ \ \text{for}\ \ m'\geq 1.$$
[*Proof*]{}. Note that, since $T_{\k}(1)\overline{f}=\overline{f}$, $\overline{f}$ is holomorphic at infinity so equation (\[hair\]) follows from (\[wash\]). Let $\overline{f}$ be the reduction mod $p$ of a series as in (\[f\]). It is enough to show if two integers $m_0 \geq 1$ and $m'\geq 1$ satisfy $v_p(m_0)\leq v_p(m')$ then $\overline{a}_{m_0,m'}=0$. Pick such integers $m_0,m'$ and set $i=v_p(m_0)$, $m_0=p^i\mu$, $m'=p^i \mu'$, $n=\mu+p\mu'$. Clearly $n$ is coprime to $p$. Picking out the coefficient of $q^{p^i-p^{i+1}\mu'}(q')^{p^i\mu'}$ in the equation in Proposition \[Tf\] we get $$\overline{a}_{m_0,m'}=\overline{a}_{p^i n-p^{i+1}\mu',p^i \mu'}= 0$$ and we are done.
\[voce\] Let $\k$ be an integer, let $\overline{f}\in k[[q]][q']$ be holomorphic at infinity, and assume that for any integer $n \geq 1$ coprime to $p$ we are given a $\overline{\lambda}_n \in k$. Then $T_{\k}(n)\overline{f}= \overline{\lambda}_n \overline{f}$ for all $(n,p)=1$ if and only if $\overline{f}$ has the form (\[wash\]) and $$T_{\k+2m'} (n) \varphi_{m'}(q)= n^{m'}\overline{\lambda}_n \varphi_{m'}(q)\ \ \text{for all}\ \ m'\geq 0.$$
[*Proof*]{}. This follows directly from the previous corollaries plus the commutation of $T_{\k}(n)$ and $V$ on $k[[q]]$.
Order $r=2$
-----------
Let us record the formula giving the Hecke action on $\d$-series of order $2$. This formula will not be used in the sequel.
\[ord22\] If $f=\sum_{m,m
,m''} a_{m,m',m''} q^m (q')^{m'}(q'')^{m''}\in R((q))[q',q'']\h$ then we have the following congruence mod $p$: $$\begin{array}{rcl}
T_{\k}(n)f & \equiv & \sum
A^{\k-1} \left(\frac{A}{D}\right)^{m'+m''} \times a_{m,m',m''} \times q^{A(m+m'p+m''p^2)/D} \\
\ & \ & \ \\
\ & \ & \times \left(\frac{q'}{q^p}\right)^{m'} \times \left[ \frac{q''}{q^{p^2}}+\frac{\d(A/D)}{A/D} \cdot \left(\frac{q'}{q^p}\right)^p+
\frac{1}{2}\left( \frac{A}{D}-1\right) \cdot \left( \frac{q'}{q^p}\right)^{2p}\right]^{m''}\end{array}$$ where the sum in the right hand side runs through all $m,m',m'',A,D$ with $A \geq 1, AD=n, (A,N)=1,D|m+m'p+m''p^2$.
[*Proof*]{}. A computation similar to the one in the proof of Proposition \[Tf\].
Note that the formula in Proposition \[ord22\] acquires a simpler form for special $n$s. Indeed assume $n=\ell$ is a prime. If $\ell \equiv 1$ mod $p$ then $\frac{A}{D}-1=0$ in $k$. If $\ell \equiv 1$ mod $p^2$ then $\d(A/D)=0$ in $k$. Finally if $\ell \equiv 1$ mod $p$ but $\ell \not\equiv 1$ mod $p^2$ then $\d(A/D)\neq 0$ in $k$.
Frobenii {#Frobenii}
--------
Consider the ring endomorphisms $F,F_k,F_{/k}$ of $k((q))[q',...,q^{(r)}]$ defined as follows: $F$ is the $p$-power Frobenius (the “absolute Frobenius"); $F_k$ is the ring automorphism that acts as the $p$-power Frobenius on $k$ and is the identity on the variables $q,q',...,q^{(r)}$; $F_{/k}$ is the ring endomorphism that is the identity on $k$ and sends $q,q',...,q^{(r)}$ into $q^p,(q')^p,...,(q^{(r)})^p$ respectively (the “relative Frobenius"). So we have $F=F_k \circ F_{/k}=F_{/k}\circ F_k$. Of course $V=F_{/k}$ on $k((q))$. Also clearly $T_{\k}(n)$ commute with $F$. By Proposition \[Tf\] $T_{\k}(n)$ also commute with $F_{k}$ on $k((q))[q']$; so $T_{\k}(n)$ commute with $F_{/k}$ on $k((q))[q']$.
Hecke operator at $p$
=====================
Taylor and Laurent $\d$-symmetry
--------------------------------
Following [@eigen] we consider the $R-$algebras $$\begin{array}{rcl}
A & := &
R[[s_1,...,s_p]][s_p^{-1}]\h[s_1',...,s_p',...,s_1^{(r)},...,s_p^{(r)}]\h,\\
\ & \ & \ \\
B & := &
R[[q_1,...,q_p]][q_1^{-1}...q_p^{-1}]\h[q_1',...,q_p',...,q_1^{(r)},...,q_p^{(r)}]\h,
\end{array}$$ where $s_1,...,s_p,s'_1,...,s_p',...$ and $q_1,...,q_p,q'_1,...,q_p',...$ are indeterminates. In [@eigen], Lemma 9.10 we proved that the natural algebra map $$A \ra B,\ \ \ s^{(i)}_j
\mapsto \d^iS_j,$$ where $S_1,...,S_p$ are the fundamental symmetric polynomials in $q_1,...,q_p$, is injective with torsion free cokernel. We will view this algebra map as an inclusion.
An element $G \in B$ is called [*Laurent $\d-$symmetric*]{} [@eigen] if it is the image of some element $G_{(p)} \in A$ (which is then unique). An element $f \in R((q))\h[q',...,q^{(r)}]\h$ will be called [*Laurent $\d-p$-symmetric*]{} if $$\Sigma_pf:=\sum_{j=1}^pf(q_j,...,q_j^{(r)}) \in B$$ is Laurent $\d-$symmetric.
In the same way one can consider the algebras $$\begin{array}{rcl}
A & := &
R[[s_1,...,s_p]][s_1',...,s_p',...,s_1^{(r)},...,s_p^{(r)}]\h,\\
\ & \ & \ \\
B & := &
R[[q_1,...,q_p]][q_1',...,q_p',...,q_1^{(r)},...,q_p^{(r)}]\h.
\end{array}$$ As before the natural algebra map $$A \ra B,\ \ \ s^{(i)}_j
\mapsto \d^iS_j,$$ is injective with torsion free cokernel.
An element $G \in B$ will be called [*Taylor $\d-$symmetric*]{} if it is the image of some element $G_{(p)} \in A$ (which is then unique). An element $f \in R[[q]][q',...,q^{(r)}]\h$ will be called [*Taylor $\d-p$-symmetric*]{} if $$\Sigma_pf:=\sum_{j=1}^pf(q_j,...,q_j^{(r)}) \in B$$ is Taylor $\d-$symmetric.
Clearly a Taylor $\d-p$-symmetric series is also Laurent $\d-p$-symmetric.
1\) Any element of $R[[q]]$ (respectively $R((q))$) is Taylor (respectively Laurent) $\d-p$-symmetric.
2\) The Taylor (respectively Laurent) $\d-p$-symmetric elements in $R[[q]][q',...,q^{(r)}]\h$ (respectively $R((q))\h[q',...,q^{(r)}]\h$) form a $p$-adically closed $R$-submodule.
3\) If $f$ is Taylor (respectively Laurent) $\d-p$-symmetric then $\phi(f)$ is Taylor (respectively Laurent) $\d-p$-symmetric.
4\) If $f\in R[[q]][q',...,q^{(r)}]\h$ (respectively $f \in R((q))\h[q',...,q^{(r)}]\h$) and $pf$ is Taylor (respectively Laurent) $\d-p$-symmetric then $f$ is Taylor (respectively Laurent) $\d-p$-symmetric.
5\) By 1)-4) any element $f$ in $R[[q]][q',...,q^{(r)}]\h$ (respectively in $R((q))\h[q',...,q^{(r)}]\h$) of the form $$f=\frac{\sum_{i=0}^m \phi^i(g_i)}{p^{\nu}}$$ where $g_i$ are in $R[[q]]$ (respectively in $R((q))$) is Taylor (respectively Laurent) $\d-p$-symmetric. In particular for any $g$ in $R[[q]]$ (respectively in $R((q))$) we have that $\d g=\frac{\phi(g)-g^p}{p}$, and more generally $\frac{\phi^i(g)-g^{p^i}}{p}$ are Taylor (respectively Laurent) $\d-p$-symmetric.
6\) Let $\cF \in R[[T_1,T_2]]^g$ be a formal group law, and let $\psi \in R[[T]][T,...,T^{(r)}]\h$ be such that $$\psi(\cF(T_1,T_2),...,\d^r\cF(T_1,T_2))=
\psi(T_1,...,T_1^{(r)})+\psi(T_2,...,T_2^{(r)})$$ in the ring $$R[[T_1,T_2]][T'_1,T_2',,...,T_1^{(r)},T_2^{(r)}]\h.$$ (Such a $\psi$ is called a $\d$-[*character*]{} of ${\mathcal F}$.) Let $\varphi(q) \in qR[[q]]$ and let $$f:=\psi(\varphi(q),...,\d^r(\varphi(q))) \in
R[[q]][q',...,q^{(r)}]\h.$$ Then $f$ is Taylor $\d-p$-symmetric. Cf the argument in [@dcc].
Note that if $\cF$ is defined over ${\mathbb Z}_p$ then $\cF$ posses a $\d$-character $\psi$ of order $r$ at most the height of $\cF$ mod $p$ such that $$\psi(T,0,...,0)\in T+T^p{\mathbb Z}_p[[T]];$$ cf. [@book], proof of Proposition 4.26.
Applying the above considerations to the multiplicative formal group we get that for any $\varphi(q)\in qR((q))$ the series $$\frac{1}{p}\log\left(\frac{\phi(\varphi(q)+1)}{(\varphi(q)+1)^p}\right)$$ is Taylor $\d-p$-symmetric. (Here, as usual, $\log(1+T)=T-T^2/2+T^3/3-...$)
7\) The series $$\label{Psi}\Psi=\frac{1}{p}\log \left(\frac{\phi(q)}{q^p}\right)$$ is Laurent $\d-p$-symmetric; cf. [@eigen], proof of Proposition 9.13.
8\) In [@eigen] we also defined the concept of $\d$-symmetric element in $$R[[q_1,...,q_p,...,q_1^{(r)},...,q_p^{(p)}]]$$ (without the qualification “Taylor" or “Laurent"). We will not use this concept in the present paper. But note that if a series is Taylor $\d$-symmetric then it is also $\d$-symmetric in the sense of [@eigen] (and Laurent $\d$-symmetric in the sense of the present paper).
For any Taylor (respectively Laurent) $\d-p$-symmetric $$f\in R[[q]][q',...,q^{(r)}]\h \ \ \text{(respectively
$f \in R((q))\h[q',...,q^{(r)}]\h$)}$$ we define $$U f:=p^{-1}(\Sigma_p f)_{(p)}(0,...,0,q,...,0,...,0,q^{(r)})$$ which is an element in $p^{-1}R[[q]][q',...,q^{(r)}]\h$ (respectively in $p^{-1}R((q))\h[q',...,q^{(r)}]\h$).
The operator $pU$ takes $R[[q]][q',...,q^{(r)}]\h$ (respectively in $R((q))\h[q',...,q^{(r)}]\h$) into $R[[q]][q',...,q^{(r)}]\h$ (respectively in $R((q))\h[q',...,q^{(r)}]\h$). On the other hand the restriction of $U$ to $R((q))\h$ (respectively $R[[q]]$) takes values in $R((q))\h$ (respectively $R[[q]]$) and is equal to the classical $U$-operator $$U(\sum a_mq^m)=\sum a_{mp}q^m.$$
Define for any $f \in R((q))\h[q',...,q^{(r)}]\h$ the series $$Vf := f(q^p,...,\d^r(q^p)) \in R((q))\h[q',...,q^{(r)}]\h.$$ So for any Taylor (respectively Laurent) $\d-p$-symmetric $f$ in $R[[q]][q',...,q^{(r)}]\h$ (respectively in $R((q))\h[q',...,q^{(r)}]\h$) and any $\k\in {\mathbb Z}$ we may define $$pT_{\k}(p)f = pUf+p^{\k}Vf$$ which is an element in $p^{\k}R[[q]][q',...,q^{(r)}]\h$ (respectively in $p^{\k}R((q))\h[q',...,q^{(r)}]\h$).
The restriction of $pT_{\k}(p)$ to $R((q))$ is, of course, $p$ times the “classical" Hecke operator $T_{\k}(p)$ on $R((q))$ defined by $$T_{\k}(p)(\sum a_m q^m)=\sum a_{pm}q^m+p^{\k-1}\sum a_m q^{pm}.$$ Recall:
[@eigen] The series $\Psi$ in (\[Psi\]) satisfies $$pU\Psi=\Psi,\ \ V\Psi=p\Psi.$$
For the next definition recall that the homomorphism $$\overline{A}:=A\otimes_R k \ra \overline{B}:=B \otimes_R k$$ is injective (in both situations described in the beginning of the section).
An element $\overline{G} \in \overline{B}$ is called [*Taylor $\d$-symmetric mod $p$*]{} (respectively [*Laurent $\d-$symmetric mod $p$*]{}) if it is the image of some element $\overline{G}_{(p)} \in \overline{A}$ (which is then unique). An element $\overline{f}\in k[[q]][q',...,q^{(r)}]\h$ (respectively $\overline{f} \in k((q))[q',...,q^{(r)}]$) will be called Taylor (respectively Laurent) $\d-p$-[*symmetric*]{} if $$\Sigma_p\overline{f}:=\sum_{j=1}^p\overline{f}(q_j,...,q_j^{(r)}) \in \overline{B}$$ is Taylor $\d-$symmetric mod $p$ (respectively Laurent $\d$-symmetric mod $p$).
Clearly any Taylor $\d-p$-symmetric series is Laurent $\d-p$-symmetric.
\[multe\] 1) The Taylor (respectively Laurent) $\d-p$-symmetric elements in $k[[q]][q',...,q^{(r)}]$ (respectively in $k((q))[q',...,q^{(r)}]$) form a $k$-subspace closed under $F_k$ and $F$ (hence also under $F_{/k}$).
2\) If $f \in R[[q]][q',...,q^{(r)}]\h$ (respectively $f\in R((q))\h[q',...,q^{(r)}]\h$) is congruent mod $p$ to a Taylor (respectively Laurent) $\d-p$-symmetric element then the image of $\overline{f}$ of $f$ in $k[[q]][q',...,q^{(r)}]$ (respectively in $k((q))[q',...,q^{(r)}]$) Taylor (respectively Laurent) $\d-p$-symmetric.
For any Taylor (respectively Laurent) $\d-p$-symmetric $$\text{$\overline{f}\in k[[q]][q',...,q^{(r)}]\h$ (respectively $k((q))[q',...,q^{(r)}]$)}$$ we may define $$``pU" \overline{f}:=(\Sigma_p \overline{f})_{(p)}(0,...,0,q,...,0,...,0,q^{(r)})$$ which is an element of $k[[q]][q',...,q^{(r)}]\h$ (respectively $k((q))[q',...,q^{(r)}]$).
The operator $``pU"$ clearly commutes with the operators $F$ and $F_k$ and hence it also commutes with the operator $F_{/k}$ (cf. section \[Frobenii\]). If $$\text{$f \in R[[q]][q',...,q^{(r)}]\h$ (respectively $f \in R((q))\h[q',...,q^{(r)}]\h$)}$$ is Taylor (respectively Laurent) $\d-p$-symmetric and $\overline{f}$ is the reduction mod $p$ of $f$ viewed as an element in $k[[q]][q',...,q^{(r)}]$ (respectively in $k((q))[q',...,q^{(r)}]$) then $``pU"\overline{f}$ is the reduction mod $p$ of $pUf$; this justifies the notation in $``pU"\overline{f}$.
Note that the operator $U:R((q))\h\ra R((q))\h$ induces an operator still denoted by $U$, $U:k((q))\ra k((q))$ (which is, of course, the classical $U$-operator $U\overline{f}=
\sum \overline{a}_{mp}q^m$, for $\overline{f}=\sum \overline{a}_m q^m\in k((q))$). On the other hand note that $``pU"\overline{f}=0$ for all $\overline{f}\in k((q))$. Finally note that if $\k \geq 1$ then the operator $T_{\k}(p)$ on $R((q))$ induces an operator $T_{\k}(p)$ on $k((q))$; if $\k \geq 2$ then $T_{\k}(p)$ on $k((q))$ coincides with $U$ on $k((q))$.
Define the ring endomorphism $V$ of $$\text{$k[[q]][q',...,q^{(r)}]$ (respectively $k((q))[q',...,q^{(r)}]$)}$$ as the reduction mod $p$ of the operator $V$ over $R$. (Note that $V(q')=0$ and $F_{/k}(q')=(q')^p$ so in particular $V\neq F_{/k}$ on $k((q))[q']$.) As in the case of characteristic zero, for any $\kappa\in {\mathbb Z}_{\geq 0}$ and any Taylor (respectively Laurent) $\d-p$-symmetric series $\overline{f}$ in $k[[q]][q',...,q^{(r)}]$ (respectively $k((q))[q',...,q^{(r)}]$) we define $$``pT_{\k}(p)"\overline{f}=``pU"\overline{f}+\ee \cdot V\overline{f}$$ which is again an element of $k[[q]][q',...,q^{(r)}]$ (respectively $k((q))[q',...,q^{(r)}]$). (Note that $\ee$ is $0$ or $1$ according as $\k$ is $>0$ or $0$.)
The operator $V$ clearly commutes with $F$ and $F_k$ (and hence also with $F_{/k}$). So the operators $``pT_{\k}(p)"$ commute with $F,F_k,F_{/k}$.
Also for $f$ any Taylor (respectively Laurent) $\d-p$-symmetric series in $R[[q]][q',...,q^{(r)}]\h$ (respectively $R((q))\h[q',...,q^{(r)}]\h$) with reduction mod $p$ $\overline{f}$ we have that $``pT_{\k}(p)"\overline{f}$ is the reduction mod $p$ of $pT_{\k}(p)f$ which, again, justifies our notation.
Structure of Laurent and Taylor $\d-p$-symmetric series
=======================================================
In what follows we address the problem of determining what series are Laurent (respectively Taylor) $\d-p$-symmetric and determining the action of our operators $``pU"$ on them. We will use the following notation: for all $\varphi=\sum \overline{a}_nq^n \in k((q))$ we define $$\label{varphiminunu}\varphi^{(-1)}:=\theta^{p-2}\varphi=\sum_{(n,p)=1} \frac{\overline{a}_n}{n}q^n \in k((q))$$ where $\theta=q\frac{d}{dq}$ is the Serre theta operator.
\[conv\] If an element $\overline{f}\in k[[q]][q']$ is Taylor $\d-p$-symmetric then it has the form $$\label{blabla}
\overline{f}=\varphi_0(q)+\sum_{s\geq 0} (V^{s+1}(\varphi_{p^s}(q)))\left( \frac{q'}{q^p}\right)^{p^s}\in k((q))[q']$$ with $\varphi_0(q)\in k[[q]]$, $\varphi_1(q),\varphi_p(q),\varphi_{p^2}(q),...\in qk[[q]]$
Conversely we will prove:
\[infinitesums\] Any element of the form $$\overline{f}=\varphi_0(q)+\sum_{s\geq 0} (V^{s+1}(\varphi_{p^s}(q)))\left( \frac{q'}{q^p}\right)^{p^s}\in k((q))[q']$$ with $\varphi_0(q),\varphi_1(q),\varphi_p(q),\varphi_{p^2}(q),...\in k((q))$ is Laurent $\d-p$-symmetric and $$``pU" \overline{f}=-\sum_{s\geq 0} V^s(\varphi^{(-1)}_{p^s}(q))+\sum_{s\geq 0} (V^{s+1}(U( \varphi_{p^s}(q))))\left(\frac{q'}{q^p}\right)^{p^s}.$$ If in addition $\overline{f}\in k[[q]][q']$ (i.e. if $
\varphi_0(q) \in k[[q]]$ and $\varphi_1(q),\varphi_p(q),\varphi_{p^2}(q),...\in qk[[q]])$) then $\overline{f}$ is Taylor $\d-p$-symmetric.
\[voce2\] Let $\overline{f}\in k((q))[q']$ be Laurent $\d-p$-symmetric and let $\overline{\lambda}_p\in k$. Then $``pT_{\k}(p)"\overline{f}=\overline{\lambda}_p\cdot \overline{f}$ if and only if:
1\) $U(\varphi_{p^s}(q))=\overline{\lambda}_p \cdot \varphi_{p^s}(q)$ for all $s\geq 0$ and
2\) $\ee\cdot V(\varphi_0(q))-\sum_{s\geq 0} V^s(\varphi_{p^s}^{(-1)}(q))=\overline{\lambda}_p\cdot \varphi_0(q)$.
If $\overline{f}\in k[[q]][q']$ is Taylor $\d-p$-symmetric then the series $``pU"\overline{f}$ and $``pT_{\k}(p)"\overline{f}$ are again Taylor $\d-p$-symmetric.
It is tempting to conjecture that any Taylor $\d-p$-symmetric series in $k[[q]][q',...,q^{(r)}]$ must belong to $k[[q]][q']$.
We will first prove Theorem \[infinitesums\]. The plan will be to first prove this Theorem in case $\overline{f}$ is a monomial in $k[q,q']$; cf. Lemma \[echelon2\] below. This will imply, of course, that Theorem \[infinitesums\] holds in case $\overline{f}$ is a finite sum of monomials. The rest of the proof will be devoted to extending the result from finite to infinite sums of monomials; this will require an analysis of $(s_1,...,s_p)$-adic convergence of certain series.
\[echelon2\] For any $n \in {\mathbb Z}$ and $s \in {\mathbb Z}_{\geq 0}$ the element $$\overline{f}=q^{np^{s+1}}(q')^{p^s}=q^{(n+1)p^{s+1}} \left(\frac{q'}{q^p}\right)^{p^s}\in k((q))[q']$$ is Laurent $\d-p$-symmetric (and actually Taylor $\d-p$-symmetric if $n \geq 0$.) Moreover $$``pU"\overline{f}=\begin{cases}
q^{(n+1)p^s}\left(\frac{q'}{q^p}\right)^{p^s}\ \ \text{if}\ \ p|n+1\\
\ \\
-\frac{q^{(n+1)p^s}}{n+1}\ \ \text{if}\ \ p\not| n+1
\end{cases}$$
[*Proof*]{}. It is enough to consider the case $s=0$; the general case follows by applying the $p$-power Frobenius.
For $n=-1$ note that $$q^{-p}q' \equiv \Psi\ \ mod\ \ (p)$$ and so $q^{-p}q'$ is Laurent $\d-p$-symmetric because $\Psi$ is Laurent $\d-p$-symmetric. Also $``pU" \overline{f}=\overline{f}$ because $pU\Psi=\Psi$.
Assume now $n\neq -1$. We have $$\begin{array}{rcl}
\d(q^{n+1}) & = & \frac{1}{p}[(q^p+pq')^{n+1}-q^{p(n+1)}]\\
\ & \ & \ \\
\ & = & \frac{1}{p}\left[p(n+1) q^{pn} q' +\sum_{j\geq 2} \frac{p^j}{j!} (n+1)...(n-j+2) q^{p(n+1-j)}(q')^j\right]\end{array}$$ For $j \geq 2$ (and since $p \geq 5$) we have $$v_p\left( \frac{p^j}{j!} \right) \geq j-v_p(j!)\geq j-\frac{j}{p-1}>1.$$ It follows that $$\label{deltadeputere}
\d(q^{n+1})=(n+1)[q^{pn}q'+pF_{n+1}(q,q')],\ \ \
F_{n+1}(q,q') \in
R[q,q^{-1},q'].$$ In particular $\d(q^{n+1})$ is divisible by $n+1$ in $R((q))\h[q']\h$ and we have the following congruence in $R((q))\h[q']\h$: $$\label{adele}
\frac{1}{n+1}\d(q^{n+1})\equiv q^{np}q'\ \ mod\ \ (p).$$ By Remark \[multe\], assertions 4) and 5), the left hand side of the latter congruence is Laurent $\d-p$-symmetric (and also Taylor $\d-p$-symmetric if $n \geq 0$) and hence $q^{pn}q'$ is Laurent $\d-p$-symmetric (and also Taylor $\d-p$-symmetric if $n \geq 0$).
To compute $``pU"\overline{f}$ start with the following computation in $R((q))\h[q']\h$: $$\begin{array}{rcl}
p^2(n+1)U\left(\frac{\d(q^{n+1})}{n+1}\right) & = & pU(p\d(q^{n+1})) \\
\ & \ & \ \\
\ & = & pU(\phi(q^{n+1}))-pU(q^{p(n+1)})\\
\ & \ & \ \\
\ & = & \phi(pU(q^{n+1}))-pU(q^{p(n+1)})\\
\ & \ & \ \\
\ & = & \begin{cases}
-pq^{n+1}\ \ \text{if}\ \ p\not| n+1\\
p\phi(q^{\frac{n+1}{p}})- pq^{n+1}\ \ \text{if}\ \ p|n+1\end{cases}\\
\ & \ & \ \\
\ & = &
\begin{cases}
-pq^{n+1}\ \ \text{if}\ \ p\not| n+1\\
p^2\d(q^{\frac{n+1}{p}})\ \ \text{if}\ \ p|n+1\end{cases}\\
\ & \ & \ \\
\ & = & \begin{cases}
-pq^{n+1}\ \ \text{if}\ \ p\not| n+1\\
p^2 \frac{n+1}{p} \left[ q^{p(\frac{n+1}{p}-1)}q'+pF_{\frac{n+1}{p}}(q,q')\right]\ \ \text{if}\ \ p|n+1\end{cases}
\end{array}$$ from which we get the following congruences mod $p$ in $R((q))\h[q']\h$: $$pU(q^{pn}q')\equiv pU \left( \frac{\d(q^{n+1})}{n+1} \right)
\equiv \begin{cases}
-\frac{q^{n+1}}{n+1}\ \ \text{if}\ \ p\not|n+1\\
\ \\
q^{n+1-p}q'\ \ \text{if}\ \ p|n+1.\end{cases}$$ and we are done.
\[ast\] Consider the polynomials $$s_1,...,s_p,s'_1,...,s'_p, D \in k[q_1,...,q_p,q'_1,...,q_p],\ \ D:=\prod_{i<j}(q_i-q_j).$$ Then the polynomials $$D^p q'_1,...,D^p q'_p$$ are linear combinations of $$1,s'_1,...,s'_p$$ with coefficients in $k[q_1,...,q_p]$.
[*Proof*]{}. For $j=1,...,p$ let $s_{ij}$ be obtained from $s_i$ by setting $q_j=0$; so $s_{ij}$ is the $i$th fundamental symmetric polynomial in $\{q_1,...,q_p\}\backslash \{q_j\}$. Taking $\d$ in the equalities $$q_1+...+q_p=s_1,...,q_1...q_p=s_p$$ in $R[q_1,...,q_p,q'_1,...,q'_p]$ and reducing mod $p$ we get the following equalities in $k[q_1,...,q_p,q'_1,...,q'_p]$: $$\begin{array}{rcl}
q'_1+...+q'_p & = & s'_1-\gamma_1\\
\ & \ & \ \\
s_{11}^p q'_1+...+s_{1p}^pq'_p & = & s'_2-\gamma_2\\
\ & \ & \ \\
............................... & \ & \ \\
\ & \ & \ \\
s_{p-1,1}^pq'_1+...+s_{p-1,p}^pq'_p & = & s_p'-\gamma_p
\end{array}$$ for some $\gamma_1,...,\gamma_p \in k[q_1,...,q_p]$. View this as a linear system of equations with unknowns $q'_1,...,q'_p$. We shall be done if we prove that the determinant of the matrix of this system is $\pm D^p$. This follows by taking determinants in the obvious identity of matrices
$$\left( \begin{array}{rrrr}
q_1^{p-1} & -q_1^{p-2} & ... & 1\\
\ & \ & \ & \ \\
q_2^{p-1} & -q_2^{p-2} & ... & 1\\
\ & \ & \ & \ \\
... & \ & \ & \ \\
\ & \ & \ & \ \\
q_p^{p-1} & -q_p^{p-2} & ... & 1 \end{array}\right)
\left( \begin{array}{rrrr}
1 & 1 & ... & 1\\
\ & \ & \ & \ \\
s_{11} & s_{12} & ... & s_{1p}\\
\ & \ & \ & \ \\
... & \ & \ & \ \\
\ & \ & \ & \ \\
s_{p-1,1} & s_{p-1,2} & ... & s_{p-1,p}\end{array}\right)=(D_{ij})$$
where $$D_{ij}=\prod_{s \neq j}(q_i-q_s)$$ and noting that $(D_{ij})$ is a diagonal matrix with determinant $D^2$.
\[asta\] Assume the notation of Lemma \[ast\] and $n \geq 0$. Then the element $$\sum_{i=1}^p q_i^{np}q'_i \in k[[q_1,...,q_p]][q'_1,...,q'_p]$$ is a linear combination of $$1,s'_1,...,s'_p$$ with coefficients in the ideal $$(s_1,...,s_p)^{[(n+1)/p]-1}k[s_1,...,s_p].$$
[*Proof*]{}. By Lemma \[ast\] we can write $$\sum_{i=1}^p q_i^{np}q'_i=A_0+\sum_{j=1}^p A_j s'_j$$ where $A_j \in k[q_1,...,q_p,D^{-1}]$ for $j=0,...,p$. On the other hand, by (\[adele\]) $\sum_{i=1}^p q_i^{np}q'_i$ is the reduction mod $p$ of $$\frac{1}{n+1}\sum_{i=1}^p \d(q_i^{n+1}) \in R[q_1,...,q_p,q'_1,...,q_p].$$ We claim that the following holds: $$\label{claim}
\sum_{i=1}^p \d(q_i^{n+1}) \in (s_1,...,s_p,s'_1,...,s'_p)^{[(n+1)/p]}R[s_1,...,s_p,s'_1,...,s'_p].$$ Assuming (\[claim\]) is true let us show how to conclude the proof of the Lemma. By (\[claim\]) we get that $$\sum_{i=1}^p q_i^{np}q'_i \in (s_1,...,s_p,s'_1,...,s_p)^{[(n+1)/p]}k[s_1,...,s_p,s'_1,...,s'_p].$$ So we have $$\sum_{i=1}^p q_i^{np}q'_i=\sum B_{i_1...i_p}(s'_1)^{i_1}...(s'_p)^{i_p}$$ where $$B_{i_1...i_p} \in (s_1,...,s_p)^{[(n+1)/p]-i_1-...-i_p}k[s_1,...,s_p].$$ Since $s'_1,...,s'_p$ are algebraically independent over $k[q_1,...,q_p]$ we get $$\begin{array}{rcl}
A_0 & = & B_{0...0}\\
\ & \ & \ \\
A_1 & = & B_{10...0}\\
\ & \ & \ \\
A_2 & = & B_{010...0},\ \text{etc}\end{array}$$ hence $$A_j \in (s_1,...,s_p)^{[(n+1)/p]-1}k[s_1,...,s_p],\ \ j=0,...,p$$ which ends the proof of the Lemma.
To check (\[claim\]) above note that $$\sum_{i=1}^p \d(q_i^{n+1})=\d\left(\sum_{i=1}^p q_i^{n+1}\right)+\frac{\left(\sum_{i=1}^p q_i^{n+1}\right)^p-\sum_{i=1}^p q_i^{(n+1)p}}{p}.$$ The second term in the right hand side of the above equation is a homogeneous polynomial in $q_1,...,q_p$ of degree $(n+1)p$ hence it is a weighted homogeneous polynomial in $s_1,...,s_p$ of weight $(n+1)p$ where $s_1,...,s_p$ are given weights $1,...,p$ respectively. Hence this polynomial is a sum of monomials in $s_1,...,s_p$ of degree $\geq n+1$. Similarly $\sum_{i=1}^p q_i^{n+1}$ is a sum of monomials in $s_1,...,s_p$ of degree $\geq [(n+1)/p]$. This implies that $\d(\sum_{i=1}^p q_i^{n+1})$ is a sum of monomials in $s_1,...,s_p,s'_1,...,s'_p$ of degree $\geq [(n+1)/p]$ which proves (\[claim\]).
[*Proof of Theorem \[infinitesums\].*]{} In view of Lemma \[echelon2\] (which treats the case of monomials) we see that in order to prove that $\overline{f}$ in the statement of the Theorem is Laurent (respectively Taylor) $\d-p$-symmetric it is enough to show that any series of the form $$\overline{f}=\sum_{n=0}^{\infty} \overline{c}_n q^{pn}q' \in k[[q]][q']$$ is Taylor $\d-p$-symmetric. By Lemma \[asta\] we may write $$\sum_{i=1}^p q_i^{np}q'_i=G_{0n}+\sum_{j=1}^p G_{jn} s'_j$$ where $$G_{jn}\in (s_1,...,s_p)^{[(n+1)/p]-1}k[s_1,...,s_p],\ \ j=0,...,p.$$ Since $G_j:=\sum_{n=0}^{\infty} \overline{c}_n G_{jn}$ are convergent in $k[[s_1,...,s_p]]$ we have $$\sum_{i=1}^p \overline{f}(q_i)=G_0+\sum_{j=1}^p G_j s'_j \in k[[s_1,...,s_p]][s'_1,...,s'_p]$$ which proves that $\overline{f}$ is Taylor $\d-p$-symmetric. The assertion about $``pU" \overline{f}$ follows from Lemma \[echelon\] by taking limits.
Next we proceed to proving Theorem \[conv\]. We need a preparation. Let $C_p(q_1,q_2):=\frac{q_1^p+q_2^p-(q_1+q_2)^p}{p}\in {\mathbb Z}[q_1,q_2]$. We start with a version of Lemma \[ast\]:
\[shape\] Consider the elements $\sigma=q_1+q_2\in k[q_1,q_2]$ and $\pi=q_1q_2\in k[q_1,q_2]$ and let $\gamma \in k[q_1,q_2]$ be the image of $C_p(q_1,q_2)\in {\mathbb Z}[q_1,q_2]$. Then $$q'_1=\frac{\pi'-q_1^p\sigma'+q_1^p\gamma}{(q_2-q_1)^p},\ \ q'_2=-\frac{\pi'-q_2^p\sigma'+q_2^p\gamma}{(q_2-q_1)^p}$$ in the ring $$k[q_1,q_2,q'_1,q'_2,\frac{1}{q_2-q_1}].$$
[*Proof*]{}. Applying $\d$ to the defining equations of $\sigma$ and $\pi$ we get $$\begin{array}{rcl}
q_1'+q_2' & = & \sigma'-\gamma\\
\ & \ & \ \\
q_2^pq'_1+ q_1^p q'_2 & = & \pi'\\
\end{array}$$ and solve for $q'_1,q_2'$.
For the next Lemma let us denote by $v_{q_2-q_1}:k((q_1,q_2))^{\times}\ra {\mathbb Z}$ the normalized valuation on the fraction field $k((q_1,q_2))$ of $k[[q_1,q_2]]$ attached to the irreducible series $q_2-q_1 \in k[[q_1,q_2]]$; in other words, if $0\neq F(q_1,q_2)\in k[[q_1,q_2]]$ then $v_{q_2-q_1}(F)$ is the maximum integer $i$ such that $(q_2-q_1)^i$ divides $F$ in $k[[q_1,q_2]]$.
\[val\] Let $\Phi(q)=\sum_{m=0}^{\infty} \beta_mq^m \in k[[q]]$, $\Phi \not\in k$, $Supp\ \Phi:=\{m \in {\mathbb Z}_{\geq 0};\beta_m\neq 0\}$. Then $$v_{q_2-q_1}(\Phi(q_2)-\Phi(q_1))=p^{\min\{v_p(m);0\neq m \in Supp\ \Phi\}}.$$
[*Proof*]{}. We have $$\begin{array}{rcl}
\Phi(q_2)-\Phi(q_1) & = & \sum_{(n,p)=1}\sum_{i=0}^{\infty} \beta_{np^i}(q_2^{np^i}-q_1^{np^i})\\
\ & \ & \ \\
\ & = & \sum_{i=0}^{\infty} (q_2-q_1)^{p^i}G(q_1,q_2)\end{array}$$ where $$G_i(q_1,q_2)=\sum_{(n,p)=1} \beta_{np^i}(q_2^{(n-1)p^i}+q_2^{(n-2)p^i}q_1^{p^i}+...+q_1^{(n-1)p^i}).$$ Let $i_0=\min\{v_p(m);0\neq m \in Supp\ \Phi\}$. Then $\beta_{np^{i}}=0$ for all $(n,p)=1$ and $i<i_0$ and there exists $n_0$, $(n_0,p)=1$ such that $\beta_{n_0p^{i_0}}\neq 0$. It is enough to show that $G_{i_0}(q_1,q_2)$ is not divisible by $q_2-q_1$ in $k[[q_1,q_2]]$ equivalently that $G(q,q)\neq 0$. But $$G_{i_0}(q,q)=\sum_{(n,p)=1} n \beta_{np^{i_0}} q^{(n-1)p^{i_0}}\neq 0.$$
[*Proof of Theorem \[conv\]*]{}. We proceed by induction on the degree $\deg(\overline{f})$ of $\overline{f}$ viewed as a polynomial in $q'$ with coefficients in $k[[q]]$. If this degree is $0$ we are done. Assume now the degree is $\geq 1$. We may assume $\overline{f}(0,0)=0$.
By hypothesis, $$\overline{f}(q_1,q'_1)+...+\overline{f}(q_p,q'_p)=G$$ in $k[[q_1,...,q_p]][q'_1,...,q'_p]$, where $G\in k[[s_1,...,s_p]][s'_1,...,s'_p]$. Setting $q_3=...=q_p=0$ and $q'_3=...=q'_p=0$ we get $$\label{cucu}
\overline{f}(q_1,q'_1)+\overline{f}(q_2,q'_2)=G(\sigma,\pi,0,...,0,\sigma',\pi',0,...,0).$$ Note that $k[[q_1,q_2]]$ is a finite $k[[\sigma,\pi]]$-algebra so $\sigma',\pi'$ are algebraically independent over $k((q_1,q_2))$. By Lemma \[shape\] the left hand side of (\[cucu\]) is a polynomial $H$ in $\sigma', \pi'$ with coefficients in $k((q_1,q_2))$. On the other hand since $H$ is in the right hand side of (\[cucu\]) $H$ has coefficients in $k[[q_1,q_2]]$. Hence each non-zero coefficient of the polynomial $H$ has $v_{q_2-q_1}$-adic valuation $\geq 0$. Now write $$\overline{f}(q,q')=\sum_{m'}\Phi_{m'}(q)(q')^{m'},\ \ \ \Phi_{m'} \in k[[q]].$$ Also write each $m'$ as $m'=n'p^{i'}$ with $n'$ not divisible by $p$. Using Lemma \[shape\] we have $H=\sum_{m'} H_{m'}$ where $$\label{FF}
H_{m'}=\frac{F_{m'}}{(q_2-q_1)^{n'p^{i'+1}}}$$ where $F_{m'} \in k((q_1,q_2))[\sigma',\pi']$ is given by $$\begin{array}{rcl}
F_{m'} & = &
\Phi_{m'}(q_1)\left((\pi')^{p^{i'}}-q_1^{p^{i'+1}}(\sigma')^{p^{i'}}+q_1^{p^{i'+1}}\gamma^{p^{i'}}\right)^{n'}\\
\ & \ & \ \\
\ & \ &
+(-1)^{n'}\Phi_{m'}(q_2)
\left((\pi')^{p^{i'}}-q_2^{p^{i'+1}}(\sigma')^{p^{i'}}+q_2^{p^{i'+1}}\gamma^{p^{i'}}\right)^{n'}.\end{array}$$ Note that the coefficient of $(\pi')^{m'}$ in $F_{m'}$ is $$\label{star}
\Phi_{m'}(q_1)+(-1)^{n'}\Phi_{m'}(q_2)$$ while the coefficient of $(\pi')^{m'-p^{i'}}(\sigma')^{p^{i'}}$ in $F_{m'}$ is $$\label{starr}
-n'\left( q_1^{p^{i'+1}}\Phi_{m'}(q_1)+(-1)^{n'}q_2^{p^{i'+1}}\Phi_{m'}(q_2)\right).$$ Let now $m'=\deg(\overline{f})$. If $n'$ is even the polynomial (\[star\]) has $v_{q_2-q_1}$-adic valuation $0$ which contradicts the fact that the non-zero coefficients of $H$ have $v_{q_2-q_1}$-adic valuation $\geq 0$. So $n'$ is odd. By Lemma \[val\] the $v_{q_2-q_1}$-adic valuation of (\[star\]) equals $$p^{\min\{v_p(m);0 \neq m \in Supp\ \Phi_{m'}\}},\ \ \ \text{if}\ \ \Phi_{m'}\not\in k.$$ Also the $v_{q_2-q_1}$-adic valuation of (\[starr\]) equals $$p^{\min\{v_p(m);m\in Supp(q^{p^{i'+1}}\Phi_{m'})\}}=p^{\min\{v_p(m+p^{i'+1});m \in Supp\ \Phi_{m'}\}}.$$ By the fact that the non-zero coefficients of $H$ have $v_{q_2-q_1}$-adic valuation $\geq 0$ we get that $$\label{apple}
p^{\min\{v_p(m);0 \neq m \in Supp\ \Phi_{m'}\}} \geq n'p^{i'+1}\ \ \text{if}\ \ \Phi_{m'}\not\in k$$ and $$\label{applee}
p^{\min\{v_p(m+p^{i'+1});m \in Supp\ \Phi_{m'}\}} \geq n'p^{i'+1}.$$ From (\[apple\]) we get $$\label{conc1}
v_p(m) \geq i'+1\ \ \text{for all} \ \ 0\neq m \in Supp\ \Phi_{m'},\ \ \text{if}\ \ \Phi_{m'}\not\in k.$$ We claim now that $n'=1$. Assume $n'\geq 2$. By (\[apple\]) $$v_p(m)>i'+1\ \ \text{for all}\ \ 0 \neq m \in Supp\ \Phi_{m'},\ \ \text{if}\ \ \Phi_{m'}\not\in k.$$ Hence $$v_p(m+p^{i'+1})=i'+1\ \ \text{for all}\ \ m \in Supp\ \Phi_{m'}.$$ By (\[applee\]) $p^{i'+1}\geq 2p^{i'+1}$, a contradiction. This ends the proof that $n'=1$.
By (\[conc1\]) $$\Phi_{m'}(q)(q')^{m'}=(V^{i'+1}\varphi)(q')^{p^{i'}}$$ for some $\varphi \in k[[q]]$. By Lemma \[echelon2\] $\Phi_{m'}(q)(q')^{m'}$ is Taylor $\d-p$-symmetric hence so is $\overline{f}-\Phi_{m'}(q)(q')^{m'}$ which has smaller degree than $\overline{f}$. We conclude by the induction hypothesis.
Multiplicity one
================
We begin by recalling the well known situation for series in $k[[q]]$. Then we proceed with our main results about $\d$-series in $k[[q]][q']$.
Throughout this section we fix $\k \in {\mathbb Z}_{\geq 0}$.
A series $\varphi\in qk[[q]]$ is said to be an [*eigenvector of all Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$*]{}, with eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p \in k$ if $\varphi\neq 0$ and the following hold: $$\label{eqvarphi}
\begin{cases}
T_{\k+2}(n)\varphi = \overline{\lambda}_n \cdot \varphi, \ \ \ (n,p)=1\\
\ \\
T_{\k+2}(p)\varphi = \overline{\lambda}_p\cdot \varphi.\end{cases}$$ Of course the last equation in (\[eqvarphi\]) is equivalent to $$U\varphi=\overline{\lambda}_p\cdot \varphi.$$
Assume $\varphi\in qk[[q]]$ is an eigenvector of all Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$, with eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p \in k$ Then there exists $\gamma \in k^{\times}$ such that $$\label{defoffi}
\varphi(q):=\gamma \cdot
\sum_{(n,p)=1} \sum_{i \geq 0} \overline{\lambda}_n\overline{\lambda}_p^i \cdot q^{np^i}.$$
[*Proof*]{}. Pick out coefficient of $q$ in the first equation (\[eqvarphi\]) and the coefficient of $q^m$, $m \geq 1$ in the second equation (\[eqvarphi\]). (Here we use the convention that $0^0=1$.)
A $\d$-series $\overline{f}=\overline{f}(q,q')\in k[[q]][q']$ is said to be an [*eigenvector of all Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$*]{}, with eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p \in k$ if $\overline{f}$ is Taylor $\d-p$-symmetric and satisfies $$\label{eqf}
\begin{cases}
nT_{\k}(n)\overline{f} = \overline{\lambda}_n \cdot \overline{f}, \ \ \ (n,p)=1;\\
\ \\
``pT_{\k}(p)"\overline{f} = \overline{\lambda}_p \cdot \overline{f}. \end{cases}$$
\[arnab\] Assume $\overline{f}=\overline{f}(q,q')\in k[[q]][q']$, $\overline{f}\not\in k$, is an eigenvector of all Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, with eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p \in k$. Then there exists $\varphi=\varphi(q)\in qk[[q]]$ and $c, c_i \in k$, $i \geq 0$, with $\ee \cdot c_{i-1}=\overline{\lambda}_p\cdot c_i$ for $i\gg 0$, such that $\varphi$ is an eigenvector of all Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$, with the same eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p$ and such that $$\label{zor}
\begin{array}{rcl}
\overline{f} & = & c+
\left(\sum_{i \geq 0} c_i F_{/k}^i\right)\varphi^{\sharp,2},\\
\ & \ & \ \\
\varphi^{\sharp,2} & := & \varphi^{(-1)}-\overline{\lambda}_p \cdot V(\varphi)
\frac{q'}{q^p}+\ee\cdot V^2(\varphi)\left(\frac{q'}{q^p}\right)^p.\end{array}$$
One can also write $\overline{f}$ in (\[zor\]) as $$\begin{array}{rcl}
\overline{f} & = & c+
\sum_{i \geq 0} c_i \left[ V^i(\varphi^{(-1)})-\overline{\lambda}_p \cdot V^{i+1}(\varphi)
\left(\frac{q'}{q^p}\right)^{p^i}+\ee\cdot V^{i+2}(\varphi)\left(\frac{q'}{q^p}\right)^{p^{i+1}}\right]\\
\ & \ & \ \\
\ & = & c+\left( \sum_{i\geq 0} c_iV^i\right) \varphi^{(-1)} + \sum_{i \geq 0}
(\ee c_{i-1}-\overline{\lambda}_p c_i) V^{i+1}(\varphi) \left(\frac{q'}{q^p}\right)^{p^i},\end{array}$$ where $c_{-1}:=0$. Note that the condition that $\ee \cdot c_{i-1}=\overline{\lambda}_p\cdot c_i$ for $i\gg 0$ insures that the right hand side of the first equation in (\[zor\]) is a polynomial in the variable $q'$.
Looking at the constant terms in (\[eqf\]) one sees that if $c\neq 0$ then $$\label{oprit}
\begin{cases}
\overline{\lambda}_n = n \cdot \sum_{A|n} \epsilon(A)A^{\k-1}, \ \ (n,p)=1;\\
\ \\
\overline{\lambda}_p = \ee. \ \end{cases}$$
Conversely we will prove:
\[curcan\] Let $\k \in {\mathbb Z}_{\geq 0}$. Assume $\varphi=\varphi(q)\in qk[[q]]$ is an eigenvector of all Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$, with eigenvalues $\overline{\lambda}_n,\overline{\lambda}_p \in k$. Let $c_i \in k$ for $i \geq 0$ with $\ee \cdot c_{i-1}=\overline{\lambda}_p\cdot c_i$ for $i\gg 0$. Also let $c$ be an arbitrary element in $k$ or $0$ according as equations (\[oprit\]) hold or fail respectively. Let $\overline{f}\in k[[q]][q']$ be defined by Equation (\[zor\]). Then $\overline{f}$ an eigenvector of all Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, with the same eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p$.
Let $k[F_{/k}]$ be the $k$-algebra generated by $F_{/k}$ which is a commutative polynomial ring in one variable. Note that $k[[q]][q']$ is a $k[F_{/k}]$-module and the $k$-linear space of series $\overline{f}(q,q')\in k[[q]][q']$ with $f(0,0)=0$ is a torsion free $k[F_{/k}]$-submodule. Note also that the ideal $qk[[q]]$ is a torsion free module over the ring $k[[F_{/k}]]$ of power series in $F_{/k}$. Finally recall that a $\d$-series $\overline{f}(q,q')\in k[[q]][q']$ is called [*primitive*]{} if $U(\overline{f}(q,0))=0$. Theorems \[arnab\] and \[curcan\] immediately imply:
Fix $\overline{\lambda}_n \in k$ for $(n,p)=1$ and $\overline{\lambda}_p \in k$. Let ${\mathcal F}$ be the $k$-linear space of all the $\d$-series $\overline{f}=\overline{f}(q,q')\in k[[q]][q']$ with $f(0,0)=0$ which are either $0$ or are eigenvectors of all Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, with eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p \in k$. We have ${\mathcal F}\neq 0$ if and only if there exists an eigenvector $\varphi\in qk[[q]]$ of all Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$, with eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p$. Assume furthermore that this is the case and let $\varphi^{\sharp,2}$ be defined as in (\[zor\]). Then $\varphi^{\sharp,2}$ belongs to ${\mathcal F}$ and is a primitive $\d$-series; also any primitive $\d$-series in ${\mathcal F}$ is a $k$-multiple of $\varphi^{\sharp,2}$. Furthemore the following hold:
1\) If $\k>0$, $\overline{\lambda}_p=0$ then ${\mathcal F}$ is a free $k[[F_{/k}]]$-submodule of $k[[q]]$ of rank $1$ with basis $\varphi^{\sharp,2}=\varphi^{(-1)}$.
2\) If either $\k >0$, $\overline{\lambda}_p\neq 0$ or $\k=0$, $\overline{\lambda}_p=0$ then ${\mathcal F}$ is a free $k[F_{/k}]$-submodule of $k[[q]][q']$ of rank one with basis $\varphi^{\sharp,2}$.
3\) If $\k=0$, $\overline{\lambda}_p\neq 0$ then ${\mathcal F}$ is a free $k[F_{/k}]$-submodule of $k[[q]][q']$ of rank $1$ with basis $$\label{varphisharpunu}
\varphi^{\sharp,1}:=\left(\sum_{i\geq 0} (\overline{\lambda}_p)^{-i} F^i_{/k}\right)\varphi^{\sharp,2}.$$
Note that $$\varphi^{\sharp,1}=\left(\sum_{i\geq 0} (\overline{\lambda}_p)^{-i} V^i\right)\varphi^{(-1)}-\overline{\lambda}_p\cdot V(\varphi)\cdot \frac{q'}{q^p}$$ and also that $\varphi^{\sharp,1}$ is the unique element of $qk[[q]]$ satisfying the equation $$V(\varphi^{\sharp,1})-\overline{\lambda}_p \varphi^{\sharp,1}+\overline{\lambda}_p \varphi^{\sharp,2}=0.$$
[*Proof of Theorem \[arnab\]*]{}. For any series $\beta(q) \in k[[q]]$ write $$\beta(q)=\sum_{m\geq 0} a_m(\beta) q^m.$$ By Theorem \[conv\] and Corollaries \[voce\] and \[voce2\] $\overline{f}$ has the form (\[blabla\]) and
$$\label{urrs}\begin{array}{rcll}
T_{\k}(n)\varphi_0 & = & \frac{\overline{\lambda}_n}{n} \cdot \varphi_0, & (n,p)=1\\
\ & \ & \ & \ \\
T_{\k+2p^s}(n)\varphi_{p^s} & = & \overline{\lambda}_n \cdot \varphi_{p^s}, & (n,p)=1,\ s\geq 0\\
\ & \ & \ & \ \\
U(\varphi_{p^s}) & = & \overline{\lambda}_p \cdot \varphi_{p^s}, & s \geq 0\\
\ & \ & \ & \ \\
\ee\cdot V(\varphi_0) -\sum_{s \geq 0} V^s(\varphi_{p^s}^{(-1)}) & = & \overline{\lambda}_p \cdot \varphi_0. & \ \end{array}$$
In particular the following equalities hold:
$$\label{ocup}\begin{array}{rcll}
a_{np^s}(\varphi_0) & = & \frac{\overline{\lambda}_n}{n} \cdot a_{p^s}(\varphi_0), & (n,p)=1, s\geq 0,\\
\ & \ & \ & \ \\
a_{n}(\varphi_{p^s}) & = & \overline{\lambda}_n \cdot a_1(\varphi_{p^s}), & (n,p)=1, s\geq 0,\\
\ & \ & \ & \ \\
a_{mp}(\varphi_{p^s}) & = & \overline{\lambda}_p \cdot a_m(\varphi_{p^s}), & m\geq 1, s\geq 0,\\
\ & \ & \ & \ \\
\ee\cdot a_{p^{s-1}}(\varphi_0)- a_1(\varphi_{p^s}) & = & \overline{\lambda}_p \cdot a_{p^s}(\varphi_0), & s \geq 0,\end{array}$$
where by convention we set $a_{p^{s-1}}(\varphi_0)=0$ if $s=0$. Let $c=a_0(\varphi_0)$ and $c_i=a_{p^i}(\varphi_0)$ for $i \geq 0$. By (\[ocup\]) we get $$\label{fom}
\begin{array}{rcll}
a_{np^i}(\varphi_0) & = & \frac{\overline{\lambda}_n}{n} \cdot c_i, & (n,p)=1,\ \ i \geq 0\\
\ & \ & \ & \ \\
a_{np^i}(\varphi_{p^s}) & = & \overline{\lambda}_n \overline{\lambda}_p^i \cdot (\ee \cdot c_{s-1}-\overline{\lambda}_p c_s),
& (n,p)=1,\ \ i \geq 0,\ s\geq 0,\end{array}$$ where $c_{-1}:=0$. Define $\varphi$ by the equality (\[defoffi\]) with $\gamma=1$.
Assume first that there is an $s \geq 0$ such that $a_1(\varphi_{p^s}) \neq 0$. Then $\varphi_{p^s}$ is a non-zero multiple of $\varphi$ so (\[eqvarphi\]) follows from (\[urrs\]) and (\[zor\]) follows from (\[fom\]). Since $\overline{f}$ is a polynomial in $q'$ we get that $\ee \cdot c_{s-1}-\overline{\lambda}_p c_s=0$ for $s\gg 0$.
Assume now that $a_1(\varphi_{p^s})=0$ for all $s \geq 0$. Then $\varphi_{p^s}=0$ for all $s \geq 0$ hence $\overline{f}=\varphi_0$. By the last equation in (\[urrs\]) and since $\varphi_0\not\in k$ we get $\ee=\overline{\lambda}_p=0$. Then the right hand side of (\[zor\]) becomes $$\label{hihi}
c+\sum_{i \geq 0} \sum_{(n,p)=1} c_i \frac{\overline{\lambda}_n}{n} q^{np^i}.$$ By the first equation in (\[fom\]) we get that (\[hihi\]) equals $\varphi_0=\overline{f}$; so equation (\[zor\]) holds. Clearly $U\varphi=0$ so the second equality in (\[eqvarphi\]) holds. Finally, since $\varphi_0 \not\in k$ we may write $\varphi_0=F^d_{/k}\tilde{\varphi}_0$ with $\tilde{\varphi}_0\in k[[q]]$ and $d$ maximal with this property; in particular $c_d \neq 0$. Note that $\theta \tilde{\varphi}_0=c_d \varphi$. Also by (\[urrs\]) we have $T_{\k}(n)\tilde{\varphi}_0=\frac{\overline{\lambda}_n}{n} \tilde{\varphi}_0$ for $(n,p)=1$. Hence $$T_{\k+2}(n)\varphi=c_d^{-1} T_{\k+2}(n)\theta \tilde{\varphi}_0=c_d^{-1} n \theta (T_{\k}(n)\tilde{\varphi}_0)=c_d^{-1} \overline{\lambda}_n \theta \tilde{\varphi}_0=\overline{\lambda}_n \varphi$$ and so the first equality in (\[eqvarphi\]) holds. This ends the proof.
[*Proof of Theorem \[curcan\]*]{}. This follows directly from Corollary \[buium\] and Theorem \[infinitesums\] using the following facts (which are direct consequences of the formulae for the Hecke operators acting on Fourier coefficients (\[cuani\])): $$\begin{array}{rcll}
T_{\k+2p^i}(n)\varphi & = & \overline{\lambda}_n \cdot \varphi, & (n,p)=1, \ \ \ i\geq 0\\
\ & \ & \ & \ \\
T_{\k}(n)(\varphi^{(-1)}) & = & \frac{\overline{\lambda}_n}{n} \cdot \varphi, & (n,p)=1. \end{array}$$
$\d$-modular forms
==================
Review of classical modular forms
---------------------------------
Start by recalling some basic facts about modular forms; cf. [@DI]. Let $N>4$ be an integer and let $B$ be a ${\mathbb Z}[1/N,\zeta_N]$-algebra. Let $Y=Y_1(N)$ be the affine modular curve over $B$ classifying pairs $(E,\alpha)$ consisting of elliptic curves $E$ over $B$-algebras plus a level $\Gamma_1(N)$ structure $\alpha:{\mathbb Z}/N{\mathbb Z} \ra E$. Let $Y_{ord}$ be the ordinary locus in $Y$ (i.e. the locus where the Eisenstein form $E_{p-1}$ is invertible). Let $X$ be $Y$ or $Y_{ord}$. Let $L$ be the line bundle on $X$, direct image of the sheaf of relative differentials on the universal elliptic curve over $X$, and let $$\label{ve}
V=Spec\left(\bigoplus_{\k \in {\mathbb Z}}L^{\otimes \k}\right)\ra X$$ be the ${\mathbb G}_m$-torsor associated to $L$.
Set $M=\cO(V)=\bigoplus_{\k \in {\mathbb Z}}L^{\otimes \k}$. Recall that there is a Fourier expansion map $$E:M\ra B((q))$$ defined by the cusp $\Gamma_1(N)\cdot \infty$ [@DI], p. 112. Recall also that $Y$ has a natural compactification, $X_1(N)$, equipped with a natural line bundle, still denoted by $L$, extending the line bundle $L$ on $Y$, such that the space of classical modular forms, $M(\Gamma_1(N),B,\k)\subset L^{\otimes \k}$, on $\Gamma_1(N)$ of weight $\k$, defined over $B$ identifies with $H^0(X_1(N),L^{\otimes \k})$. Recall that the diamond operators act on $M(\Gamma_1(N),B,\k)$; the invariant elements form the space $M(\Gamma_0(N),B,\k)$ of classical modular forms on $\Gamma_0(N)$ of weight $\k$ defined over $B$. Recall the $q$-expansion principle: for any $B$ as above there is an induced injective Fourier expansion map $E:M(\Gamma_1(N),B,\k)\ra B[[q]]$ and if $B' \subset B$ then $M(\Gamma_1(N),B',\k)$ identifies with the group of all $f \in M(\Gamma_1(N),B,\k)$ such that $E(f)\in B'[[q]]$. Recall also the following base change property: if $B'$ is any $B$-algebra and either $B'$ is flat over $B$ or $\k\geq 2$ and $N$ is invertible in $B'$ then the map $M(\Gamma_1(N),B,\k)\otimes_B B'\ra M(\Gamma_1(N),B',\k)$ is an isomorphism; cf. [@DI], p.111.
$\d$-series from classical modular forms
----------------------------------------
\[nopiine\] Let $\k \in {\mathbb Z}_{\geq 0}$ and let $f(q)=\sum_{m\geq 1}a_mq^m\in q{\mathbb Z}_p[[q]]$ be a series satisfying $a_1=1$ and $$\label{knacond}
\begin{cases}
a_{p^i n} = a_{p^i} a_n\ \ \text{for $(n,p)=1,\ i\geq 0$}\\
a_{p^{i-1}}a_p = a_{p^i}+p^{\k+1}a_{p^{i-1}}\ \ \text{for $i \geq 2$}.\end{cases}$$ Let $\varphi:=\overline{f}=\sum_{m \geq 1} \overline{a}_m q^m \in q{\mathbb F}_p[[q]]$ be the reduction mod $p$ of $f(q)$. Then the series $$\label{inpat}
f^{\sharp,2}=f^{\sharp,2}(q,q',q''):=\frac{1}{p}
\cdot \sum_{n \geq 1}\frac{a_n}{n}(p^{\k}\phi^2(q)^n-a_p \phi(q)^n+pq^n) \in {\mathbb Q}_p[[q,q',q'']]$$ belongs to ${\mathbb Z}_p[[q]][q',q'']\h$ and its reduction mod $p$ equals $$\label{baie}
\overline{f^{\sharp,2}}=\overline{f^{\sharp,2}(q,q',q'')}=\varphi^{(-1)}-\overline{a}_p V(\varphi)\frac{q'}{q^p}+ \ee \cdot V^{2}(\varphi)\left(\frac{q'}{q^p}\right)^{p}
\in {\mathbb F}_p[[q]][q'].$$
[*Proof*]{}. For $\k=0$ the argument is in [@BP]; the case $\k>0$ is entirely similar. (Note that the form $f_{[a_p]}^{(0)}$ in [@BP] is congruent mod $p$ to $f$ itself.)
Note that conditions (\[knacond\]) imply that $U \varphi=\overline{a}_p \cdot \varphi$.
\[gura\] Let $\k \in {\mathbb Z}_{\geq 0}$ and let $F\subset {\mathbb C}$ be a number field with ring of integers $\cO_F$. Let $$\label{cameo}
f(q)=\sum_{m\geq 1}a_mq^m\in q\cO_F[[q]]$$ be the Fourier expansion of a cusp form $$f\in M(\Gamma_0(N),\cO_F,\k+2).$$ Assume $a_1=1$ and assume $f(q)$ is an eigenvector for all the Hecke operators $T_{\k+2}(n)$ with $n \geq 1$. Assume $p$ is a rational prime that splits completely in $F$, consider an embedding $\cO_F \subset {\mathbb Z}_p$, view $f(q)$ as an element of $q{\mathbb Z}_p[[q]]$, and let $\varphi:=\overline{f}=\sum_{m \geq 1} \overline{a}_m q^m \in q{\mathbb F}_p[[q]]$ is the reduction mod $p$ of $f(q)$. Then the equalities (\[knacond\]) hold. So by Theorem \[nopiine\] the series $$\label{inpat}
f^{\sharp,2}=f^{\sharp,2}(q,q',q''):=\frac{1}{p}
\cdot \sum_{n \geq 1}\frac{a_n}{n}(p^{\k}\phi^2(q)^n-a_p \phi(q)^n+pq^n) \in {\mathbb Q}_p[[q,q',q'']]$$ belongs to ${\mathbb Z}_p[[q]][q',q'']\h$ and its reduction mod $p$ equals $$\label{baie}
\overline{f^{\sharp,2}}:=\overline{f^{\sharp,2}(q,q',q'')}=\varphi^{(-1)}-\overline{a}_p V(\varphi)\frac{q'}{q^p}+ \ee \cdot V^{2}(\varphi)\left(\frac{q'}{q^p}\right)^{p}
\in {\mathbb F}_p[[q]][q'].$$ Note also that $T_{\k+2}(n)\varphi=\overline{a}_n \cdot \varphi$ for $(n,p)=1$ and $U\varphi=\overline{a}_p \cdot \varphi$. So by Theorem \[curcan\] $\overline{f^{\sharp,2}}=\varphi^{\sharp,2}$ is an eigenvector of the Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, with eigenvalues $\overline{a}_n,\overline{a}_p$. Also, by the same Theorem, if in addition $\overline{a}_p\neq 0$ and $\k=0$, then the series $\varphi^{\sharp,1}$ in (\[varphisharpunu\]) is also an eigenvector of the Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, with eigenvalues $\overline{a}_n,\overline{a}_p$.
\[punga\] Consider the Ramanujan series $$P(q):=E_2(q):=1-24\sum_{m \geq 1}\left(\sum_{d|m}d \right)q^m$$ and assume $N$ is prime. Consider the series $$g(q):=-\frac{1}{24}(P(q)-NP(q^N))=\frac{N-1}{24}+f(q)\in {\mathbb Z}_{(p)}[[q]],$$ where $$\label{acestf}
f(q)=\sum_{m \geq 1}\left(\sum_{A|m}\epsilon(A)A\right)q^m.$$ Then $g(q)$ is the Fourier expansion of a classical modular form in $M(\Gamma_0(N),{\mathbb Z}_{(p)},2)$ which is an eigenvector of the Hecke operators $T_2(n)$ for all $n \geq 1$ with eigenvalues $a_n:=\sum_{A|n}\epsilon(A)A$; cf. [@DI], Example 2.2.6, Proposition 3.5.1, and Remark 3.5.2. Let $\varphi:=\overline{f}=\sum_{m \geq 1} \overline{a}_m q^m \in q{\mathbb F}_p[[q]]$ be the reduction mod $p$ of $f(q)$. By [@Knapp], Theorem 9.17, the equalities (\[knacond\]) hold with $\k=0$. So by Theorem \[nopiine\] the series $$\label{inpat2}
f^{\sharp,2}=f^{\sharp,2}(q,q',q''):=\frac{1}{p}
\cdot \sum_{n \geq 1}\frac{a_n}{n}(\phi^2(q)^n-a_p \phi(q)^n+pq^n) \in {\mathbb Q}_p[[q,q',q'']]$$ belongs to ${\mathbb Z}_p[[q]][q',q'']\h$ and its reduction mod $p$ equals $$\label{baie2}
\overline{f^{\sharp,2}}:=\overline{f^{\sharp,2}(q,q',q'')}=\varphi^{(-1)}-\overline{a}_p V(\varphi)\frac{q'}{q^p}+ V^{2}(\varphi)\left(\frac{q'}{q^p}\right)^{p}
\in {\mathbb F}_p[[q]][q'].$$ Note also that $T_{2}(n)\varphi=\overline{a}_n \cdot \varphi$ for $(n,p)=1$ and $U\varphi=\overline{a}_p \cdot \varphi$. So by Theorem \[curcan\] $\overline{f^{\sharp,2}}=\varphi^{\sharp,2}$ is an eigenvector of the Hecke operators $nT_{0}(n)$, “$pT_{0}(p)$", $(n,p)=1$, with eigenvalues $\overline{a}_n,\overline{a}_p$. Also, by the same Theorem, if in addition $\overline{a}_p\neq 0$ and $\k=0$, then the series $\varphi^{\sharp,1}$ in (\[varphisharpunu\]) is also an eigenvector of the Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, with eigenvalues $\overline{a}_n,\overline{a}_p$. Note that if $N \equiv 1$ mod $p$ then Equations \[oprit\] hold because $$\begin{cases}
a_n=\sum_{A|n}\epsilon(A)A\equiv n \sum_{A|n}\epsilon(A)A^{-1},\ \ \text{mod $p$ for $(n,p)=1$},\\
\ \\
a_p=\sum_{A|p}\epsilon(A)A\equiv 1\ \ \text{mod $p$}.\end{cases}$$ Note also that if $N \equiv 1$ mod $p$ it follows that $f(q)\equiv g(q)$ mod $p$ so $\varphi(q)$ is the Fourier expansion of a modular form in $M(\Gamma_0(N),{\mathbb F}_p,2)$
Review of $\d$-modular forms [@eigen; @igusa]
---------------------------------------------
Let $V$ be an affine smooth scheme over $R$ and fix a closed embedding $V \subset {\mathbb A}^m$ into an affine space over $R$.
A map $f:V(R)\ra R$ is called a [*$\d$-function of order $r$*]{} on $X$ [@char] if there exists a restricted power series $\Phi$ in $m(r+1)$ variables, with $R$-coefficients such that $$f(a)=\Phi(a,\d a,...,\d^r a),$$ for all $a\in V(R)\subset R^m$. We denote by $\cO^r(V)$ the ring of $\d$-functions of order $r$ on $V$.
(Recall that [*restricted*]{} means [*with coefficients converging $p$-adically to $0$*]{}; also the definition above does not depend on the embedding $V \subset {\mathbb A}^m$.) Composition with $\d$ defines $p$-derivations $\d:\cO^r(V)\ra \cO^{r+1}(V)$. The rings $\cO^r(V)$ have the following universality property: for any $R$-algebra homomorphism $u:\cO(V)\ra A$ where $A$ is a $p$-adically complete $\d$-ring there are unique $R$-algebra maps $u^r:\cO^r(V) \ra A$ that commute in the obvious sense with $\d$ and prolong $u$.
Let now $V$ be as in (\[ve\]) with $B=R$ and ${\mathbb Z}[1/N,\zeta_N]\subset R$ a fixed embedding.
[@igusa] A $\d-$[*modular function*]{} of order $r$ (on $\Gamma_1(N)$, holomorphic on $X$) is a $\d$-function $f:V(R)\ra R$ of order $r$.
Let $W:=\bZ[\phi]$ be the ring generated by $\phi$. For $w=\sum a_i\phi^i\in W$ ($a_i \in \bZ$) set $deg(w)=\sum a_i\in \bZ$; for $\lambda\in R^{\times}$ we set $\lambda^w:=\prod \phi^i(\lambda)^{a_i}$.
A [*$\d$-modular form of weight $w$*]{} (of order $r$, on $\Gamma_1(N)$, holomorphic on $X$) is a $\d$-modular function $f:V(R)\ra R$ of order $r$ such that $$f(\lambda \cdot a)=\lambda^w f(a),$$ for all $\lambda\in R^{\times}$ and $a\in V(R)$, where $(\lambda,a)\mapsto \lambda \cdot a$ is the natural action $R^{\times}\times V(R)\ra V(R)$.
We denote by $M^r:=\cO^r(V)$ the ring of all $\d$-modular functions of order $r$ and we set $M^{\infty}:=\bigcup_{r \geq 0} M^r$. We denote by $M^r(w)$ the $R$-module of $\d$-modular forms of order $r$ and weight $w$; cf. [@igusa]. (In [@eigen] the space $M^r(w)$ was denoted by $M^r(\Gamma_1(N),R,w)$ or $M^r_{ord}(\Gamma_1(N),R,w)$ according as $X$ is $Y$ or $Y_{ord}$.) Note that $M^r(0)$ identifies with $\cO^r(X)$ which, in its turn, embeds into $\cO^r(X_1(N))$.
By the universality property of the rings $M^r=\cO^r(V)$ there exists a unique $\d$-ring homomorphism (the [*$\d$-Fourier expansion map*]{}) $$E:M^{\infty}\ra S^{\infty}_{for}:=\bigcup_{r\geq 0}R((q))[q',...,q^{(n)}]\h,\ \ \ E(f)=f(q,q',q'',...),$$ extending the Fourier expansion map $E:M\ra R((q))\h$. We may also consider the composition $$M^{\infty} \ra S^{\infty}_{for}\stackrel{\pi}{\ra} R((q))\h,\ \ \ f \mapsto f(q),$$ where the map $\pi$ sends $q',q'',...$ into $0$; we refer to this composition as the [*Fourier expansion map*]{}.
Recall the “$\d$-expansion principle":
\[princip\] [@eigen] The maps $E:M^r(w)\ra R((q))[q',...,q^{(r)}]\h$ are injective with torsion free cokernel; hence the induced maps $\overline{E}:M^r(w)\otimes k \ra k((q))[q',...,q^{(r)}]$ are injective.
[*Proof*]{}. This is [@eigen], Lemma 6.1.
Recall also the following result:
\[zzz\] [@eigen] If in Example \[gura\] $\k=0$, $F={\mathbb Q}$, and $p\gg 0$ then the series $f^{\sharp,2}(q,q',q'')\in R[[q]][q',q'']\h$ in (\[inpat\]) is the image of a (unique) $\d$-modular form (still denoted by) $f^{\sharp,2}\in \cO^2(X_1(N)) \subset M^2(0)$. If in addition $f$ in Example \[gura\] is of “CL type" then the series $\varphi^{\sharp,1} \in k[[q]][q']$ in that Example is the image of a $\d$-modular form $f^{\sharp,1}\in \cO^1(X_1(N))\subset M^1(0)$.
Here by $f$ being of [*CL type*]{} we mean that the Neron model of the elliptic curve over ${\mathbb Q}$ associated to $f$ via the Eichler-Shimura construction has good ordinary reduction and its base change to $R$ is the canonical lift of this reduction; cf. [@eigen; @BP] for more details.
[*Proof*]{}. Let $f^{\sharp}\in \cO^r(X_1(N))$ be as in [@eigen], Theorems 6.3 and 6.5; cf. also [@BP], Lemma 4.18. So $r$ is $1$ or $2$ according as $f$ is or is not of CL type. Then Theorem \[zzz\] follows from [@eigen], Theorems 6.3 and 6.5, by letting the $\d$-modular form $f^{\sharp,2}$ be defined by $$f^{\sharp,2}:=\begin{cases}
f^{\sharp},\ \ \text{if $f$ is not of CL type},\\
\phi(f^{\sharp})-a_pf^{\sharp},\ \ \ \text{if $f$ is of CL type},\end{cases}$$ and by letting $$f^{\sharp,1}:=f^{\sharp}\ \ \ \text{if $f$ is of CL type}.$$
It is tempting to conjecture that if in Example \[gura\] $\k\geq 0$ is arbitrary, $F={\mathbb Q}$, and $p\gg 0$ then the series $f^{\sharp,2}(q,q',q'')$ is the $\d$-Fourier expansion of a $\d$-modular form $f^{\sharp,2}\in M^r(\k)$ for some $r\geq 2$. An appropriate variant of this should also hold for arbitrary $F$. As we shall see, however, the situation is drastically different with Example \[punga\]; cf. Theorem \[contrast\].
Recall the [*Serre derivation operator*]{} $\partial: M\ra M$ introduced by Serre and Katz [@Katz]. (Cf. also [@book], p.254 for a review). Recall that $\partial(L^{\otimes n}) \subset L^{\otimes (n+2)}$. Recall also that if $X$ is contained in $Y_{ord}$ then one has the Ramanujan form $P \in M^0(2)$. By [@book], Propositions 3.43, 3.45, 3.56, there exists a unique sequence of $R$-derivations $\partial_j:M^{\infty}\ra M^{\infty}$, $j \geq 0$, such that $$\begin{cases}
\partial_j \circ \phi^s = 0 \ \ \text{on $M$ for $j \neq s$}\\
\partial_j \circ \phi^j = p^j \cdot \phi^j \circ \partial\ \ \text{on $M$ for $j \geq 0$}\end{cases}$$ These derivations then also have the property that $$\begin{cases}
\partial_j = 0\ \ \text{on $M^{j-1}$ for $j \geq 1$}\\
\partial_j \circ \d^j=\phi^j \circ \partial\ \ \text{on $M$ for $j \geq 0$}\end{cases}$$ and that $$\label{pastreaza}
\partial_j(M^r(w)) \subset M^r(w+2\phi^j).$$ Recall the Ramanujan theta operator $\theta=q\frac{d}{dq}:R((q))\ra R((q))$. Then by [@book], Lemma 4.18, there is a unique sequence of $R$-derivations $\theta_j:S^{\infty}_{for}\ra S^{\infty}_{for}$ such that $$\label{conju}
\begin{cases}
\theta_j \circ \phi^s = 0 \ \ \text{on $R((q))$ for $j \neq s$}\\
\theta_j \circ \phi^j = p^j \cdot \phi^j \circ \theta\ \ \text{on $R((q))$ for $j \geq 0$};\end{cases}$$ and such that $$\label{onem}
\begin{cases}
\theta_j = 0\ \ \text{on $R((q))[q',...,q^{(j-1)}]\h$ for $j \geq 1$}\\
\theta_j \circ \d^j=\phi^j \circ \theta\ \ \text{on $R((q))$ for $j \geq 0$.}\end{cases}$$
\[barcfor\] For any $w=\sum_{i=0}^r a_i\phi^i\in W$, any $j \geq 0$, and any $f \in M^r(w)$ the following formula holds in $S^{\infty}_{for}$: $$E(\partial_j f)=\theta_j(E(f))-a_jp^jE(f)E(P)^{\phi^j}.$$
[*Proof*]{}. This was proved in [@book], Proposition 8.42 in the case of “$\d$-Serre-Tate expansions"; the case of $\d$-Fourier expansions is entirely similar. (The level $1$ case of this Proposition was proved in [@Barcau] using the structure of the ring of modular forms of level $1$.)
Finally we recall the $\d$-modular forms $f^1$ and $f^{\partial}$ introduced in [@difmod] and [@Barcau] respectively:
\[funu\] [@difmod; @Barcau; @book] For each $r \geq 1$ there exists a unique form $f^r \in M^r(-1-\phi^r)$ such that $$E(f^r)=\Psi^{\phi^{r-1}}+p\Psi^{\phi^{r-2}}+...+p^{r-1}\Psi.$$ In particular $$E(f^r) \equiv \left(\frac{q'}{q^p}\right)^{p^{r-1}}\ \ \ mod\ \ \ p.$$
\[fpartz\] [@Barcau; @book] Assume $X=Y_{ord}$. Then there exists a unique form $f^{\partial} \in M^1(\phi-1)$ such that $E(f^{\partial})=1$. The form $f^{\partial}$ is invertible in the ring $M^1$ and its inverse belongs to $M^1(1-\phi)$. Furthermore the image of $f^{\partial}$ in $M^1 \otimes k$, coincides with the image of the Eisenstein series $E_{p-1} \in M(\Gamma_1(N),R,p-1)$.
Note that Proposition \[funu\] holds, in particular for $X=Y=Y_1(N)$. However Proposition \[fpartz\] fails for $X=Y$: the form $f^{\partial}$ has “singularities" at the supersingular points.
Review of Katz’ generalized $p$-adic modular forms [@Katz; @Gouvea]
-------------------------------------------------------------------
Let $B$ be a $p$-adically complete ring, $p\geq 5$, and let $N$ be an integer coprime to $p$. Consider the functor $$\label{hite}
\{\text{$p$-adically complete $B$-algebras}\} \ra \{\text{sets}\}$$ that attaches to any $A$ the set of isomorphism classes of triples $(E/A,\varphi, \alpha)$, where $E$ is an elliptic curve over $A$, $\varphi$ is a trivialization, and $\alpha$ is an arithmetic level $\Gamma_1(N)$ structure. Recall that a [*trivialization*]{} is an isomorphism between the formal group of $E$ and the formal group of the multiplicative group; an [*arithmetic level $N$ structure*]{} is defined as an inclusion of flat group schemes over $B$, $\alpha: \mu_{N}\ra E$. So if $B$ contains a primitive $N$-th root of unity (which we fix) then an arithmetic level $\Gamma_1(N)$ structure is the same as a level $\Gamma_1(N)$ structure. The functor (\[hite\]) is representable by a $p$-adically complete ring $\bW(B,N)$. The elements of this ring are called by Katz [@Katz] [*generalized $p$-adic modular forms*]{}; an element $f\in \bW(B,N)$ can be identified with a rule that naturally attaches to any test object $(E/A,\varphi,\alpha)$ an element $f(E,\varphi,\alpha)\in A$. Note that $\bW(B,N)=\bW(\bZ_p,N) \widehat{\otimes} B$. Moreover there is a $\bZ_p^{\times}$-action on $\bW(B,N)$, $(\lambda,f) \mapsto \lambda \cdot f$, coming from the action of $\bZ_p^{\times}$ on the formal group of the multiplicative group.
There is a natural [*Fourier expansion map*]{} $E:\bW(B,N) \ra B((q))\hat{\ }$ which is injective and has a flat cokernel over $B$ coming from evaluation on the Tate curve. [*From now on we shall view $\bW(B,N)$ as a subring of $B((q))\h$ via the Fourier expansion map.*]{}
For $X=Y$ or $Y_{ord}$ note that the image of $\cO(V)=\bigoplus L^{\otimes \k} \ra R((q))\h$ is contained in $\bW$ and the morphism $\cO(V)\ra \bW$ is ${\mathbb Z}_p^{\times}$-equivariant with $\lambda \in {\mathbb Z}_p^{\times}$ acting on $\eta \in L^{\otimes \k}$ via $(\lambda, \eta)\mapsto \lambda^{\k}\eta$.
Also $\bW(\bZ_p,N)$ possesses a natural ring endomorphism $Frob$ which reduces modulo $p$ to the $p$-power Frobenius endomorphism of $\bW(\bZ_p,N) \otimes \bZ/p\bZ$. So if $R=\hat{\bZ}_p^{ur}$, as usual, and if $\phi$ is the automorphism of $R$ lifting Frobenius then $Frob \widehat{\otimes} \phi$ is a lift of Frobenius on $$\bW:=\bW(R,N)=\bW(\bZ_p,N) \widehat{\otimes} R$$ which we denote by $\phi_0$. Moreover the homomorphism $\bW(R,N) \ra R((q))\hat{\ }$ commutes with the action of $\phi_0$ where $\phi_0$ on $R((q))\hat{\ }$ is defined by $\phi_0(\sum a_nq^n):= \sum \phi(a_n)q^{np}$. Finally $\phi_0$ commutes with the action of ${\mathbb Z}_p^{\times}$.
Let $\chi:{\mathbb Z}_p^{\times} \ra {\mathbb Z}_p^{\times}$ be a continuous character. An element $f \in \bW$ is said to have weight $\chi$ if $\lambda \cdot f=\chi(\lambda) f$ for all $\lambda\in {\mathbb Z}_p^{\times}$; cf [@Serre; @Gouvea]. We view integers $m \in {\mathbb Z}$ as identified with continuous characters by attaching to $m$ the character $\chi(\lambda)=\lambda^m$. Recall from [@Gouvea], p. 21 that the set of all $f \in \bW(B,N)\cap B[[q]]$ that have weight $\chi$ identifies with the set of [*$p$-adic modular forms of weight $\chi$ defined over $B$*]{} in the sense of Serre [@Serre] i.e. the set of series in $B[[q]]$ which are $p$-adic limits of classical modular forms over $B$ of weights $\k_n\in {\mathbb Z}$ and level $N$ where $\k_n \ra \chi$. Note that since $\phi_0$ commutes with the action of ${\mathbb Z}_p^{\times}$ on $\bW$ it follows that if $f \in \bW$ has weight $\chi$ then so does $\phi_0(f)\in \bW$.
Application to $\d$-eigenforms {#we}
------------------------------
As noted in [@igusa] the image of the Fourier expansion map $M^{\infty}\ra R((q))\h$ is contained in $\bW$; this is by the universality property of $\cO^r(V)$ and by the fact that $\bW$ possesses a lift of Frobenius $\phi_0$ and hence it is naturally a $\d$-subring of $R((q))\h$.
\[nea\] The image of $M^r(w)$ in $\bW$ consists of elements of weight $deg(w)$.
[*Proof*]{}. It is easy to see that one may replace $X$ in the statement above by an open set of it. So one may assume $L$ is free on $X$. Let $x$ be a basis of $L$. Then any element $f \in M^r(w)$ can be written as $f=f_0\cdot x^w$ where $f_0 \in \cO^r(X)$. Now the image of $x$ in $\bW$ has weight $1$. Since $\phi_0$ on $\bW$ preserves the elements of a given weight it follows that the image of $x^w$ in $\bW$ has weight $deg(w)$. On the other hand $f_0$ is a $p$-adic limit of polynomials with $R$-coefficients in elements of the form $\d^i g_0$, where $g_0 \in \cO(X)$. Again, since $\phi_0$ sends elements of weight $0$ in $\bW$ into elements of weight $0$ the same is true for $\d:\bW \ra \bW$. Since the image of $g_0$ in $\bW$ has weight $0$ so does the image of $\d^i g_0$ in $\bW$ and hence so does the image of $f$.
Next we state our main applications to “$\d$-eigenforms" (i.e. $\d$-modular forms whose $\d$-Fourier expansions are “$\d$-eigenseries"). First we will prove:
\[dezastru\] Assume $\overline{f}=\overline{f}(q,q')\in k[[q]][q']$ is not a $p$-th power in $k[[q]][q']$ and assume $\overline{f}$ is the reduction mod $p$ of the $\d$-Fourier expansion of a $\d$-modular form in $M^r(w)$ with $r\geq 0$, $\k: =deg(w)\geq 0$. Assume furthermore that $\overline{f}$ is an eigenvector of all Hecke operators $nT_{\k}(n)$, “$pT_{\k}(p)$", $(n,p)=1$, with eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p \in k$. Then there exists $\varphi=\varphi(q)\in qk[[q]]$ which is the Fourier expansion of a modular form in $M(\Gamma_1(N),k,\k')$, $\k'\geq 0$, $\k' \equiv \k+2$ mod $p-1$, and there exist $c, c_i \in k$, $i \geq 0$, with $\ee \cdot c_{i-1}=\overline{\lambda}_p c_i$ for $i \gg 0$, such that $\varphi$ is an eigenvector of all Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$, with the same eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p$ and such that $\overline{f}$ satisfies (\[zor\]).
Conversely we will prove:
\[converseofdezastru\] Assume $\varphi \in qk[[q]]$ is the Fourier expansion of a modular form in $M(\Gamma_1(N),k,\k')$, $\k' \geq 0$, $\k'\equiv \k+2$ mod $p-1$, and that $\varphi$ is an eigenvector of all Hecke operators $T_{\k+2}(n)$, $T_{\k+2}(p)$, $(n,p)=1$, with eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p \in k$. Assume $X=Y_{ord}$. Consider the series $\overline{f}=\overline{f}(q,q')\in k[[q]][q']$ defined by the formula (\[zor\]) with $c=0, c_i \in k$ for $i\geq 0$, and $c_i=0$ for $i \gg 0$. Then $\overline{f}$ is the $\d$-Fourier expansion of a $\d$-modular form $f \in M^1(\k)$ and (by Theorem \[curcan\]) is an eigenvector of all Hecke operators $nT_{\k}(n)$, “$T_{\k}(p)$", $(n,p)=1$, with the same eigenvalues $\overline{\lambda}_n, \overline{\lambda}_p$.
Note that Theorems \[dezastru\] and \[converseofdezastru\] imply Theorem \[maint\] in the Introduction. The one-to-one correspondence in Theorem \[maint\] is given by $\varphi\mapsto \varphi^{\sharp,2}$ with $\varphi^{\sharp,2}$ defined by (\[zor\]).
[*Proof of Theorem \[dezastru\]*]{}. By Theorem \[arnab\] all we have to show is that $\varphi$ in that Theorem is the Fourier expansion of a modular form in $M(\Gamma_1(N),k,\k')$, $\k' \equiv \k+2$ mod $p-1$. Since $\overline{f}$ is not a $p$-th power we may assume $c_0=1$. Now if $\overline{f}(q,q')$ is the reduction mod $p$ of the $\d$-Fourier expansion $$E(f)=f(q,q',...,q^{(r)})\in S^{\infty}_{for}$$ of a $\d$-modular form $f \in M^r(w)$ then, by Proposition \[barcfor\], and Equations \[zor\] and \[onem\] we have the following congruences mod $p$ in $S^{\infty}_{for}$: $$\begin{array}{rcl}
E(\partial_1 f) & \equiv & \theta_1(E(f))\\
\ & \ & \ \\
\ & \equiv & -\overline{\lambda}_p V(\varphi)q^{-p}\theta_1(\d q)\\
\ & \ & \ \\
\ & \equiv & -\overline{\lambda}_p V(\varphi)q^{-p} \phi(\theta q)\\
\ & \ & \ \\
\ & \equiv & -\overline{\lambda}_p V(\varphi).
\end{array}$$
By Equation (\[pastreaza\]) we have that $\partial_1 f\in M^r(w+2\phi)$. So by Proposition \[nea\] the image $E(\partial_1 f)(q,0,...,0)$ of $E(\partial_1 f)$ in $R((q))\h$ is an element of weight $\k+2$ in $\bW$. So $E(\partial_1 f)(q,0,...,0)$ is congruent mod $p$ to the Fourier expansion of a classical modular form of weight $\k' \equiv \k+2$ mod $p-1$. So $\overline{\lambda}_p V(\varphi)$ is the Fourier expansion of a modular form in $M(\Gamma_1(N),k,\k')$.
If $\overline{\lambda}_p \neq 0$ then $V(\varphi)$ is the Fourier expansion of a modular form in $M(\Gamma_1(N),k,\k')$ hence so is $\varphi=UV\varphi$ (because $U$ preserves the weight [@Gross], p.458).
If $\overline{\lambda}_p =0$ then, by (\[defoffi\]) we have $\varphi=\sum_{(n,p)=1}\overline{\lambda}_n q^n$ so $\varphi=\theta (\varphi^{(-1)})=\theta(\varphi_0)$. Now $\varphi_0$ is the image of $E(f)$ in $k[[q]]$ so, as above, by Proposition \[nea\], $\varphi_0$ is the Fourier expansion of a modular form in $M(\Gamma_1(N),k,\k'')$ where $\k'' \equiv \k$ mod $p-1$. But $\theta$ sends Fourier expansions of modular forms of weight $\k''$ into Fourier expansions of modular forms of weight $\k''+p+1$; cf. [@Gross], p. 458. So $\varphi$ is the Fourier expansion of a modular form in $M(\Gamma_1(N),k,\k''+p+1)$, and we are done because $\k''+p+1 \equiv \k+2$ mod $p-1$.
[*Proof of Theorem \[converseofdezastru\]*]{}. Set $\k'=\k+2+(p-1)\nu$, $\nu\geq 0$. Since $\varphi^{(-1)}(q)=\theta^{p-2} \varphi(q)$ by get that $\varphi^{(-1)}(q)$ is the Fourier expansion of a modular form over $k$ of weight $\k'+(p-2)(p+1)=\k+(p-1)(p+\nu)$ hence $V^i(\varphi^{(-1)}(q))$ is the Fourier expansion of a modular form over $k$ of weight $\k_{0,i}:=p^i(\k+(p-1)(p+\nu))$; the latter lifts to a modular form $\Phi_{0,i}\in M(\Gamma_1(N),R,\k_{0,i})$ which can be viewed as an element in $M^0(\k_{0,i})$. Also $V^{i+1}(\varphi)$ and $V^{i+2}(\varphi)$ are Fourier expansions of modular forms over $k$ of weights $\k_{1,i}:=p^{i+1}\k'$ and $\k_{2,i}:=p^{i+2}\k'$ so they lift to modular forms $\Phi_{i,1}\in M(\Gamma_1(N),R,\k_{1,i})$ and $\Phi_{2,i}\in M(\Gamma_1(N),R,\k_{2,i})$ respectively. The latter can be viewed as elements of $M^0(\k_{1,i})$ and $M^0(\k_{2,i})$ respectively. Finally note that $f^1 \cdot f^{\partial}\in M^1(-2)$ and the Eisenstein form $E_{p-1}$ can be viewed as an element in $M^0(p-1)$; its inverse is an element in $M^0(1-p)$. Let $\lambda_p\in R$ be a lift of $\overline{\lambda}_p$. Note that $\k_{0,i}\equiv \k$ mod $p-1$; set $e_{0,i}:=\frac{\k-\k_{0,i}}{p-1}$. Similarly $\k_{1,i}\equiv \k+2$ mod $p-1$ and $\k_{2,i}\equiv \k+2p$ mod $p-1$; set $e_{1,i}:=\frac{\k+2-\k_{1,i}}{p-1}$ and $e_{2,i}:=\frac{\k+2p-\k_{2,i}}{p-1}$. Then, by Propositions \[funu\] and \[fpartz\] $\overline{f}$ is the $\d$-Fourier expansion of the $\d$-modular form $$\label{celalalt}
\sum_{i\geq 0} c_i\left[E_{p-1}^{e_{0,i}}\cdot \Phi_{0,i}-\lambda_p \cdot E_{p-1}^{e_{1,i}}\cdot \Phi_{1,i} \cdot (f^1 \cdot f^{\partial})+\ee
\cdot E_{p-1}^{e_{2,i}} \cdot \Phi_{2,i} \cdot (f^1 \cdot f^{\partial})^p\right]$$ which is an element of $M^1(\k).$ This ends the proof.
\[oua\] We consider a special case of Example \[gura\]. Let $$f(q)=\sum_{m\geq 1}a_mq^m\in q{\mathbb Z}[[q]]$$ be the Fourier expansion of a cusp form $f\in M(\Gamma_0(N),{\mathbb Z},2)$. Assume $a_1=1$ and assume $f(q)$ is an eigenvector for all the Hecke operators $T_{2}(n)$ with $n \geq 1$. Assume $p$ is a prime and let $\varphi:=\overline{f}=\sum_{m \geq 1} \overline{a}_m q^m \in q{\mathbb F}_p[[q]]$ be the reduction mod $p$ of $f(q)$. Then the equalities (\[knacond\]) hold with $\k=0$. So by Theorem \[nopiine\] the series $$\label{inpatuletz}
f^{\sharp,2}=f^{\sharp,2}(q,q',q''):=\frac{1}{p}
\cdot \sum_{n \geq 1}\frac{a_n}{n}(p^{\k}\phi^2(q)^n-a_p \phi(q)^n+pq^n) \in {\mathbb Q}_p[[q,q',q'']]$$ belongs to ${\mathbb Z}_p[[q]][q',q'']\h$ and its reduction mod $p$ equals $$\label{baiemica}
\overline{f^{\sharp,2}}:=\overline{f^{\sharp,2}(q,q',q'')}=\varphi^{(-1)}-\overline{a}_p V(\varphi)\frac{q'}{q^p}+ V^{2}(\varphi)\left(\frac{q'}{q^p}\right)^{p}
\in {\mathbb F}_p[[q]][q'].$$ Note also that $T_{2}(n)\varphi=\overline{a}_n \cdot \varphi$ for $(n,p)=1$ and $U\varphi=\overline{a}_p \cdot \varphi$. So by Theorem \[curcan\] $\overline{f^{\sharp,2}}$ is an eigenvector of the Hecke operators $nT_0(n)$, “$pT_0(p)$", $(n,p)=1$, with eigenvalues $\overline{a}_n,\overline{a}_p$. In addition, if $p\gg 0$, by Theorem \[zzz\], the series $f^{\sharp,2}(q,q',q'')$ in (\[inpatuletz\]) is the $\d$-Fourier expansion of a $\d$-modular form $f^{\sharp,2}\in \cO^2(X_1(N))\subset M^2(0)$.
On the other hand, as in the proof, of Theorem \[converseofdezastru\], $\varphi^{(-1)}(q)$ is the Fourier expansion of a modular form over $k$ of weight $p^2-p$; the latter lifts to a modular form $\Phi_{0}\in M(\Gamma_1(N),R,p^2-p)$ which can be viewed as an element in $M^0(p^2-p)$. Also $V(\varphi)$ and $V^{2}(\varphi)$ are Fourier expansions of modular forms over $k$ of weights $2p$ and $2p^2$ so they lift to modular forms $\Phi_{1}\in M(\Gamma_1(N),R,2p)$ and $\Phi_{2}\in M(\Gamma_1(N),R,2p^2)$ respectively. The latter can be viewed as elements of $M^0(2p)$ and $M^0(2p^2)$ respectively. Then $\overline{f^{\sharp,2}(q,q',q'')}$ is the $\d$-Fourier expansion of the $\d$-modular form $$\label{celalaltuletz}
f^{!}:=E_{p-1}^{-p}\cdot \Phi_{0}-a_p \cdot E_{p-1}^{-2}\cdot \Phi_1 \cdot (f^1 \cdot f^{\partial})+
\cdot E_{p-1}^{-2p} \cdot \Phi_{2} \cdot (f^1 \cdot f^{\partial})^p\in M^1(0).$$
Note now that $f^{\sharp,2}\in M^2(0)$ and $f^{!}\in M^1(0)$ have the same $\d$-Fourier expansion and the same weight. By Proposition \[princip\] (the “$\d$-expansion principle") we get the following:
In the notation of Example \[oua\] we have the congruence $f^{\sharp,2}\equiv f^{!}$ mod $p$ in $M^2(0)$.
Note that the right hand side of this congruence has order $1$ and has a priori “singularities" both at the cusps of $X_1(N)$ and at the supersingular points. In stark contrast with that, the left hand side of the above congruence has [*no*]{} “singularity" at either the cusps or the supersingular points.
Also in stark contrast with Theorem \[zzz\] we have the following consequence of Theorem \[dezastru\].
\[contrast\] Let $f(q)$ be as in Example \[punga\] and assume $N \not\equiv 1$ mod $p$ (for instance $p\gg 0$). Then the series $\overline{f^{\sharp}(q,q',q'')}$ in (\[baie2\]) is not the image of any element in any space $M^r(w)$ with $r \geq 0$, $deg(w)=0$.
[*Proof*]{}. Assume the notation of Example \[punga\]. By Theorem \[dezastru\] it follows that the image of $f(q)$ in ${\mathbb F}_p[[q]]$ is the Fourier expansion of some modular form $\widehat{f}\in M(\Gamma_1(N),{\mathbb F}_p,2+(p-1)\nu)$, $\nu\geq 0$. On the other hand, by Example \[punga\] we know that the image of $g(q)$ in ${\mathbb F}_p[[q]]$ is the Fourier expansion of a modular form $\widehat{g}\in M(\Gamma_0(N),{\mathbb F}_p,2)$. It follows that the modular form $$\widehat{h}:=E_{p-1}^{\nu}\cdot \widehat{g}-\widehat{f}\in M(\Gamma_1(N),{\mathbb F}_p,2+(p-1)\nu)$$ has Fourier expansion a constant $\gamma:=\frac{N-1}{24} \in {\mathbb F}_p^{\times}$. On the other hand $\gamma$, viewed as an element in $M(\Gamma_0(N),{\mathbb F}_p,0)$ has Fourier expansion $\gamma$. By the Serre and Swinnerton-Dyer Theorem [@Goren], p.140, the difference $\widehat{h}-\gamma$ is divisible by $E_{p-1}-1$ in the ring $\bigoplus_{\k \in {\mathbb Z}}M(\Gamma_1(N),{\mathbb F}_p,\k)$. It follows that the weights $2+(p-1)\nu$ and $0$ are congruent mod $p-1$, a contradiction.
[10]{}
Barcau, M: Isogeny covariant differential modular forms and the space of elliptic curves up to isogeny, Compositio Math., 137 (2003), 237-273.
Borger, J., $\Lambda$-rings and the field with one element, arXiv:math/0906.3146
Buium, A.: Differential characters of Abelian varieties over $p-$adic fields, Invent. Math. 122, 309-340 (1995).
Buium, A.: Geometry of $p$-jets, Duke Math. J., 82, 2, (1996), 349-367.
Buium, A.: Differential modular forms, Crelle J., 520 (2000), 95-167.
Buium, A.: Arithmetic Differential Equations. Math. Surveys and Monographs 118, AMS (2005)
Buium, A.: Differential characters on curves, Serge Lang memorial volume, to appear
Buium A., Differential eigenforms, J. Number Theory 128 (2008), 979-1010.
Buium A., Saha A., The ring of differential Fourier expansions, preprint.
Buium, A., Poonen, B.: Independence of points on elliptic curves arising from special points on modular and Shimura curves, II: local results, Compositio Math., 145 (2009), 566-602.
Deligne, P., Rappoport, M., Schemas de modules de courbes elliptiques, LNM 349, Springer 1973, pp. 143-316.
Diamond, F., and Im, J.: Modular forms and modular curves. In:Seminar on Fermat’s Last Theorem, Conference Proceedings, Volume 17, Canadian Mathematical Society, pp. 39-134 (1995).
Goren, E. Z., Lectures on Hilbert Modular Varieties and Modular Forms, CRM Monograph Series CRMM 14, 2002.
Gouvea, F., Arithmetic of $p$-adic modular forms, Lecture Notes in Math. 1304, Springer, 1985.
Gross, B. H., A tameness criterion for Galois representations associated to modular forms mod $p$, Duke Math. J., 61, 2, 445-517 (1990)
Hida, H., Geometric modular forms and elliptic curves, World Scientific (2000).
Katz, N.: $p-$adic properties of modular schemes and modular forms, LNM 350, Springer, Heidelberg (1973).
Knapp, A.: Elliptic Curves, Math. Notes, Princeton Univ. Press (1992)
Serre, J. P.: Formes modulaires et fonctions zéta $p-$adiques. In: LNM 350 (1973).
| {
"pile_set_name": "ArXiv"
} |
[Solvable and/or integrable many-body models on a circle]{}
$\bigskip $
**Oksana Bihun**$^{\ast 1}$ and **Francesco Calogero**$^{+2}\bigskip $
$^{\ast }$Department of Mathematics, Concordia College at Moorhead, MN, USA
$^{+}$Physics Department, University of Rome La Sapienza" and Istituto Nazionale di Fisica Nucleare, Sezione di Roma
$^{1}[email protected]
$^{2}[email protected], [email protected]
**Abstract**
Various many-body models are treated, which describe $N$ points confined to move on a plane circle. Their Newtonian equations of motion (“accelerations equal forces”) are *integrable*, i. e. they allow the *explicit* exhibition of $N$ *constants of motion* in terms of the dependent variables and their time-derivatives. Some of these models are moreover *solvable* by purely algebraic operations, by (explicitly performable) quadratures and, finally, by functional inversions. The techniques to manufacture these models are not new; some of these models are themselves new; others are reinterpretations of known models.
Introduction
============
The investigation of the time evolution of an arbitrary number $N$ of point-particles the dynamics of which is determined by Newtonian equations of motion (“accelerations equal forces”) is of course a fundamental topic in physics and mathematics. The identification in this context of models *amenable to exact treatments* is a major area of research in mathematical physics and applied mathematics, having a centuries-old history and having been boosted by developments in the last few decades, which also impacted several areas of physics beyond mechanics and many fields of pure mathematics. An interesting related development which is now becoming of interest is the study of such models in which the motion is restricted to lie on an *a priori* prescribed manifold: see for instance [AM1959]{} [@CG1993] [@H2011] **[@ET2012].** In this paper we make some initial, simple steps in this direction by focussing on various many-body models describing the evolution of $N$ points whose positions on a plane are characterized by $N$ *unit* 2-vectors, thereby forcing their motion to be confined to *a circle of unit radius centered at the origin*. All these models are characterized by *Newtonian* equations of motion: accelerations equal forces, which in these models are of *one-body*, *two-body* or, in some cases, *many-body* type, and might depend on the velocities of the moving particles in addition to their positions. All these models are *autonomous:* their equations of motion are time-independent. They are *all amenable to exact treatments*: in particular they *all* allow the explicit identification of $N$ *constants of motion* in terms of the $N$ dependent variables and their $N$ time-derivatives (for terminological simplicity we hereafter call such models *integrable*). In some cases their *initial-value problems* can be moreover *solved* by (explicitly performable) quadratures and subsequent functional inversions, preceded by purely algebraic operations, such as solving systems of *linear* constant-coefficients ODEs, or (equivalently) evaluating the $N$ *eigenvalues* of known (time-dependent) $N\times N$ matrices or (equivalently) the $N$ *zeros* of known (time-dependent) polynomials of degree $N$ (for terminological simplicity we hereafter call such models *solvable*). The techniques to manufacture these models are *not new*; some of these models are themselves *new*; others are essentially *reinterpretations of known models*. The dynamics of these models are not analyzed in detail; but in some cases the main features of their behavior are ascertained, for instance for *isochronous* models the time evolution of which is *isochronous* (i. e., *completely periodic* with a *fixed* period independent of the initial data), or for models *all* motions of which are *multiply periodic*.
The equations of motion of the $N$-body problems treated below are listed with minimal comments in the following Section 2, to facilitate the hasty reader wishing to get an immediate idea of the findings reported in this paper. These results are then proven in the subsequent Section 3: the titles of its subsections indicate case-by-case the techniques employed to arrive at the relevant results. Finally, a terse Section 4 entitled “Outlook” outlines possible developments, to be eventually reported in other papers. Some mathematical details are confined to two Appendices.
Many-body models on a circle amenable to exact treatments
=========================================================
In the following subsections we display, with minimal comments, various $N$-body problems of Newtonian type (“accelerations equal forces”) describing motions on a circle and amenable to exact treatments (detailed in the following Section 3). But we provide firstly a terse subsection devoted to notation.
Notations
---------
The models under consideration generally feature $N$ points moving in a plane. We identify these $N$ points by 3-vectors $\vec{r}_{n}$, $n=1,2,...,N$ for which we use the following 3-dimensional notation:$$\vec{r}_{n}\equiv \left( \cos \theta _{n},~\sin \theta _{n},~0\right) \equiv
\left( x_{n},~y_{n},~0\right) ~. \label{rn}$$
Hereafter $N$ is an *arbitrary positive integer* (generally $N\geq 2$) and indices such as $n,$ $m,$ $\ell $ run over the *positive integers* from $1$ to $N$ (unless otherwise explicitly indicated).
Clearly these vectors $\vec{r}_{n}$ have *unit* length,
$$\vec{r}_{n}\cdot \vec{r}_{n}=1~. \label{rnunit}$$
Throughout this paper the dot sandwiched among two vectors denotes the standard *scalar* product, so that for instance$$\vec{r}_{n}\cdot \vec{r}_{m}=\cos \left( \theta _{n}-\theta _{m}\right) ~.
\label{rndotrm}$$It is moreover convenient to introduce the *unit* vector $\hat{z}$ orthogonal to the $xy$-plane,
$$\hat{z}\equiv \left( 0,~0,~1\right) ~, \label{zhat}$$
and to denote by the “wedge” symbol $\wedge $ the standard (3-dimensional) *vector* product, so that
$$\hat{z}\wedge \vec{r}_{n}=-\vec{r}_{n}\wedge \hat{z}=\left( -\sin \theta
_{n},~\cos \theta _{n},~0\right) ~,$$
$$\left( \hat{z}\wedge \vec{r}_{m}\right) \cdot \vec{r}_{n}=\left( \vec{r}_{m}\wedge \vec{r}_{n}\right) \cdot \hat{z}=\sin \left( \theta _{n}-\theta
_{m}\right) ~. \label{rmwedgerndotzhat}$$
Hereafter we deal with time-dependent vectors
$$\vec{r}_{n}\left( t\right) \equiv \left( \cos \theta _{n}\left( t\right)
,~\sin \theta _{n}\left( t\right) ,~0\right) ~,$$
and superimposed dots indicate derivatives with respect to the time variable $t$ so that, for instance,
$$\overset{\cdot }{\vec{r}}_{n}=\dot{\theta}_{n}~\left( -\sin \theta
_{n},~\cos \theta _{n},~0\right) =\dot{\theta}_{n}~\hat{z}\wedge \vec{r}_{n}~, \label{rndot}$$
$$\begin{aligned}
\overset{\cdot \cdot }{\vec{r}}_{n} &=&\ddot{\theta}_{n}~\left( -\sin \theta
_{n},~\cos \theta _{n},~0\right) -\dot{\theta}_{n}^{2}~\left( \cos \theta
_{n},~\sin \theta _{n},~0\right) \notag \\
&=&\ddot{\theta}_{n}~\hat{z}\wedge \vec{r}_{n}-\dot{\theta}_{n}^{2}~\vec{r}_{n}~. \label{rndotdot}\end{aligned}$$
Note that here we omitted, for notational simplicity, to indicate *explicitly* the time-dependence of the quantities appearing in these $N$ equations; we will often do this below without repeating this warning.
Several other identities are reported in Appendix A: they are useful to obtain the results reported below, but are not necessary to understand the findings reported in the following subsections.
Two models obtained via techniques of generalized Lagrangian interpolation
--------------------------------------------------------------------------
*First model*:
\[ManyBodyForcesModel\] $$\begin{aligned}
&&\mu _{n}~\overset{\cdot \cdot }{\vec{r}}_{n}=-\mu _{n}~\left( \overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{n}\right) ~\vec{r}_{n}
\notag \\
&&+\hat{z}\wedge \vec{r}_{n}~\left\{ \left[ \mu _{n}~\left( \overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{n}\right) +\eta _{n}~\left(
\vec{r}_{n}\wedge \overset{\cdot }{\vec{r}}_{n}\right) \cdot \hat{z}\right]
~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \frac{\left( \vec{r}_{\ell }\cdot
\vec{r}_{n}\right) }{\left( \vec{r}_{\ell }\wedge \vec{r}_{n}\right) \cdot
\hat{z}}\right] \right. \notag \\
&&\left. +\left[ \left( \vec{r}_{n}\wedge \overset{\cdot }{\vec{r}}_{n}\right) \cdot \hat{z}\right] \sum_{\ell =1,~\ell \neq n}^{N}\left[ \frac{\sigma _{n}\left( \underline{\vec{r}}\right) }{\sigma _{\ell }\left(
\underline{\vec{r}}\right) }~\frac{\mu _{\ell }~\left( \vec{r}_{\ell }\wedge
\overset{\cdot }{\vec{r}}_{\ell }\right) \cdot \hat{z}+\eta _{\ell }}{\left(
\vec{r}_{\ell }\wedge \vec{r}_{n}\right) \cdot \hat{z}}\right] \right\} ~,
\label{ManyBodyForcesOnCircle}\end{aligned}$$$$\sigma _{n}\left( \underline{\vec{r}}\right) =\dprod\limits_{\ell =1,~\ell
\neq n}^{N}\left[ \left( \vec{r}_{\ell }\wedge \vec{r}_{n}\right) \cdot \hat{z}\right] ~. \label{sigman}$$
*Second model*:
$$\begin{aligned}
&&\mu _{n}~\overset{\cdot \cdot }{\vec{r}}_{n}==-\mu _{n}~\left( \overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{n}\right) ~\vec{r}_{n}
\notag \\
&&+\sum_{\ell =1,~\ell \neq n}^{N}\left\{ \left[ \left( \vec{r}_{\ell
}\wedge \vec{r}_{n}\right) \cdot \hat{z}\right] ^{-1}~\left\{ \left[ \left(
\vec{r}_{n}\wedge \overset{\cdot }{\vec{r}}_{n}\right) \cdot \hat{z}\right] ~\left[ \mu _{\ell }~\left( \vec{r}_{\ell }\wedge \overset{\cdot }{\vec{r}}_{\ell }\right) \cdot \hat{z}+\eta _{\ell }\right] \right. \right. \notag \\
&&\left. \left. +\left[ \mu _{n}~\left( \vec{r}_{n}\wedge \overset{\cdot }{\vec{r}}_{n}\right) \cdot \hat{z}+\eta _{n}\right] ~\left[ \left( \vec{r}_{\ell }\wedge \overset{\cdot }{\vec{r}}_{\ell }\right) \cdot \hat{z}\right]
\right\} ~\left( \vec{r}_{\ell }\wedge \vec{r}_{n}\right) \right\} ~.
\label{TwoBodyForcesOnCircle}\end{aligned}$$
In these Newtonian equations $\mu _{n}$ and $\eta _{n}$ are $2N$ arbitrary constants, and for the rest of the notation see Subsection 2.1; note in particular the property (\[rnunit\]), implying that the $N$ vectors $\vec{r}_{n}$ have *unit* modulus, hence that the $N$ points whose time evolution is determined by these equations of motion are constrained to move on the circle of *unit* radius centered at the origin of the Cartesian plane.
These equations of motion are *covariant*, implying that the corresponding $N$-body problems are *rotation-invariant*.
These two $N$-body problems are both *integrable*: they possess $N$ *constants of motion*, the explicit expressions of which in terms of the vectors $\vec{r}_{n}$ and their time-derivatives $\overset{\cdot }{\vec{r}}_{n}$ are displayed in the following Subsection 3.1. The equations of motion of the first, (\[ManyBodyForcesOnCircle\]), of these two models feature *many-body* forces due to the presence in their right-hand (“forces”) sides of the quantities $\sigma _{n}\left( \underline{\vec{r}}\right) $, see (\[sigman\]), but *their initial-value problem is* *solvable by purely algebraic operations*; nevertheless their time evolution can be quite complicated (detailed analyses are not performed in this paper; the fact that *solvable* models can exhibit quite complicated dynamics is of course well known, see for instance the papers where a 3-body model is studied the time evolution of which is highly nontrivial in spite of the fact that its Aristotelian equations of motion—“velocity equal forces”—are quite neat and that its initial-value problem can be reduced to solving a single algebraic equation [@CGSS]).* *
Two *solvable* models obtained via a reinterpretation of known models
---------------------------------------------------------------------
The *first model* is merely a transcription of the *solvable* “Sutherland model”, see Subsection 3.2. It reads as follows:$$\overset{\cdot \cdot }{\vec{r}}_{n}=-\left( \overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{n}\right) ~\vec{r}_{n}+g^{2}~\hat{z}\wedge \vec{r}_{n}~\sum_{\ell =1,~\ell \neq n}^{N}\left\{ \frac{\vec{r}_{n}\cdot \vec{r}_{\ell }}{\left[ \left( \vec{r}_{\ell }\wedge \vec{r}_{n}\right) \cdot \hat{z}\right] ^{3}}\right\} ~. \label{2a}$$Here $g$ is an *arbitrary* “coupling constant”, and the rest of the notation is, we trust, clear (see Subsection 2.1).
The *second model* is also merely a transcription of a well-known *solvable* model (“of goldfish type”), see Subsection 3.2. It reads as follows:$$\begin{aligned}
&&\overset{\cdot \cdot }{\vec{r}}_{n}=-\left( \overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{n}\right) ~\vec{r}_{n}+g_{0}~\hat{z}\wedge \vec{r}_{n}+g_{1}~\overset{\cdot }{\vec{r}}_{n} \notag \\
&&+\hat{z}\wedge \vec{r}_{n}~\sum_{\ell =1,~\ell \neq n}^{N}\left\{ \frac{2~\overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{\ell }+g_{2}~\left[ \left( \overset{\cdot }{\vec{r}}_{n}\wedge \vec{r}_{\ell }+\overset{\cdot }{\vec{r}}_{\ell }\wedge \vec{r}_{n}\right) \cdot \hat{z}\right]
+g_{3}~\vec{r}_{n}\cdot \vec{r}_{\ell }}{\left( \vec{r}_{\ell }\wedge \vec{r}_{n}\right) \cdot \hat{z}}\right\} ~. \notag \\
&& \label{2b}\end{aligned}$$Here the $4$ constants $g_{0},$ $g_{1},$ $g_{2}$ and $g_{3}$ are *arbitrary* constants, and the rest of the notation is, we trust, clear (see Subsection 2.1).
These equations of motion are *covariant*, implying that the corresponding $N$-body problems are *rotation-invariant*.
Two $N$-body problems on a circle obtained by changes of dependent variables
----------------------------------------------------------------------------
These two *solvable* models are merely transcriptions of two well-known one-dimensional *solvable* models, see Subsection 3.3. The *first model* reads as follows:
$$\begin{aligned}
&&\overset{\cdot \cdot }{\vec{r}}_{n}=-\left( \overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{n}\right) ~\vec{r}_{n}-\hat{z}\wedge
\vec{r}_{n}~\left\{ 2~\left[ \left( \overset{\cdot }{\vec{r}}_{n}\cdot
\overset{\cdot }{\vec{r}}_{n}\right) ~\frac{y_{n}}{x_{n}}\right] \right.
\notag \\
&&\left. +4~x_{n}~y_{n}-x_{n}^{5}~\sum_{\ell =1,~\ell \neq n}^{N}\left[
\frac{y_{\ell }}{\left( \vec{r}_{\ell }\wedge \vec{r}_{n}\right) \cdot \hat{z}}\right] ^{3}\right\} ~. \label{CalCircle}\end{aligned}$$
Here $x_{n}\equiv \cos \theta _{n}$ and $y_{n}\equiv \sin \theta _{n}$ are the two Cartesian components in the plane of the vector $\vec{r}_{n},$ see (\[rn\]).
This model is *isochronous* with period $\pi $,$$\vec{r}_{n}\left( t\pm \pi \right) =\vec{r}_{n}\left( t\right) ~.$$
The *second model* reads as follows:
$$\begin{aligned}
&&\overset{\cdot \cdot }{\vec{r}}_{n}=-\left( \overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{n}\right) ~\vec{r}_{n}-\hat{z}\wedge
\vec{r}_{n}~\left\{ 2~\left[ \left( \overset{\cdot }{\vec{r}}_{n}\cdot
\overset{\cdot }{\vec{r}}_{n}\right) ~\frac{y_{n}}{x_{n}}\right] \right.
\notag \\
&&\left. +x_{n}~y_{n}-x_{n}~\sum_{\ell =1,~\ell \neq n}^{N}\left\{ \frac{2~+x_{n}^{2}~x_{\ell }^{2}}{x_{\ell }~\left[ \left( \vec{r}_{\ell }\wedge
\vec{r}_{n}\right) \cdot \hat{z}\right] }\right\} \right\} ~.
\label{GoldCircle}\end{aligned}$$
Here $x_{n}\equiv \cos \theta _{n}$ and $y_{n}\equiv \sin \theta _{n}$ are again the two Cartesian components in the plane of the vector $\vec{r}_{n},$ see (\[rn\]).
*All* solutions of this model are *multiply periodic*, see Subsection 3.3.
Note that—in contrast to the equations of motions reported in the two preceding subsections—those displayed herein, (\[CalCircle\]) and ([GoldCircle]{}), are* not* written in *covariant* fashion, i. e. without any explicit appearance of the Cartesian components $x_{n}\equiv
\cos \theta _{n}$ and $y_{n}\equiv \sin \theta _{n}$ of the vector $\vec{r}_{n}$; indeed these equations of motion are *not* rotation-invariant, or equivalently, they are *not* invariant for translations along the circle (on which the motions take place due to the constraint (\[rnunit\])).
Proofs
======
In the following subsections we substantiate the findings reported in the preceding Section 2.
Solvable and integrable models on the circle manufactured via techniques of generalized Lagrangian interpolation
----------------------------------------------------------------------------------------------------------------
In this subsection we employ the technique to manufacture many-body models amenable to exact treatments introduced in [@C2001] (see in particular Chapter 3 of this book, entitled “$N$-body problems treatable via techniques of exact Lagrangian interpolation in spaces of one or more dimensions”). We begin with a terse review of this method, in the specific case of one-dimensional space with an appropriate choice of the set of “seeds” (namely, of the $N$ functions providing the point of departure for the generalized Lagrangian interpolation approach).
The set of seeds we conveniently take as basis for our treatment are the $N$ functions $$\begin{aligned}
&&\left\{ s_{n}\left( \theta \right) \right\} _{n=1}^{N}=\left\{ \exp \left[
i~\left( 2~n-N-1\right) ~\theta \right] \right\} _{n=1}^{N} \notag \\
&=&\{\exp \left[ i~\left( 1-N\right) ~\theta \right] ,~\exp \left[ i~\left(
3-N\right) ~\theta \right] ,~... \notag \\
&&...\exp \left[ i~\left( N-3\right) ~\theta \right] ,~\exp \left[ i~\left(
N-1\right) ~\theta \right] \}~. \label{seeds}\end{aligned}$$
*Remark 3.1.1*. These exponential functions with *imaginary* argument are *complex*, but clearly this set of seeds could be replaced without significant changes by an equivalent set featuring instead sines and cosines of *real* arguments. The use of exponentials merely facilitates some of the following developments. Likewise the factor $2$ in the argument of these functions has been introduced merely to yield neater versions of the equations of motions that will be obtained, see below. The fact that these seeds are invariant under the transformation $\theta
\Rightarrow \theta +2\pi $ suggests to interpret the variable $\theta $ as an *angle* in the plane. $\blacksquare $
We then consider a function $f\left( \theta \right) $ representable as a *linear* superposition of these $N$ seeds,
\[GenLag\] $$f\left( \theta \right) =\sum_{n=1}^{N}\left[ h_{n}~s_{n}\left( \theta
\right) \right] ~, \label{fhn}$$where the $N$ coefficients $h_{n}$ are *a priori* arbitrary numbers. And we denote with $f_{n}$ the $N$ values that this function takes at the $N$ (*arbitrarily assigned*) “nodes” $\theta =\theta _{n}$,$$f_{n}=f\left( \theta _{n}\right) ~; \label{fn}$$and we display the representation of this function in terms of these $N$ values, via the (“generalized Lagrangian interpolation”) formula$$f\left( \theta \right) =\sum_{n=1}^{N}\left[ f_{n}~q^{\left( n\right)
}\left( \theta ~\left\vert \underline{\theta }\right. \right) \right] ~.
\label{fqn}$$The $N$ “interpolational functions” $q^{\left( n\right) }\left( \theta
~\left\vert \underline{\theta }\right. \right) $ depend on the variable $\theta $ and on the $N$ nodes $\theta _{n}$ (hence on the $N$-vector having these nodes as its components, hereafter denoted as $\underline{\theta }\equiv \left( \theta _{1},~\theta _{2},~...,~\theta _{N}\right) $); they are themselves *linear* superpositions of the seeds $s_{n}\left( \theta
\right) $, to insure consistency among (\[fqn\]) and (\[fhn\]); and they feature the property
$$q^{\left( n\right) }\left( \theta _{m}~\left\vert \underline{\theta }\right.
\right) =\delta _{nm} \label{qdeltanm}$$
to insure consistency among (\[fqn\]) and (\[fn\]) (here and hereafter $\delta _{nm}$ is the Kronecker symbol: $\delta _{nm}=1$ if $n=m$, $\delta
_{nm}=0$ if $n\neq m$).
The explicit representation of these interpolational functions $q^{\left(
n\right) }\left( \theta ~\left\vert \underline{\theta }\right. \right) $ in terms of the $N$ seeds $s_{n}\left( \theta \right) $ and the $N$ nodes $\theta _{n}$ reads [@C2001]
\[Reprqn\] $$q^{(n)}(\theta ~\left\vert \underline{\theta }\right. )=\frac{\Delta (\theta
_{1},\ldots ,\theta _{n-1},\theta ,\theta _{n+1},\ldots ,\theta _{N})}{\Delta (\theta _{1},\ldots ,\theta _{N})}~,$$where $$\Delta (\underline{\theta })=\left\vert
\begin{array}{cccc}
s_{1}(\theta _{1}) & s_{2}(\theta _{1}) & \ldots & s_{N}(\theta _{1}) \\
s_{1}(\theta _{2}) & s_{2}(\theta _{2}) & \ldots & s_{N}(\theta _{2}) \\
\vdots & \vdots & \ddots & \vdots \\
s_{1}(\theta _{N}) & s_{2}(\theta _{N}) & \ldots & s_{N}(\theta _{N})\end{array}\right\vert ~.$$This determinant—with the set of seeds (\[seeds\])—is of Vandermonde type hence it can be explicitly evaluated, yielding for the interpolational functions the expression
$$q^{\left( n\right) }\left( \theta ~\left\vert \underline{\theta }\right.
\right) =s_{1}\left( \theta -\theta _{n}\right) ~\dprod\nolimits_{\ell
=1,~\ell \neq n}^{N}\left[ \frac{\exp \left( 2~i~\theta \right) -\exp \left(
2~i~\theta _{\ell }\right) }{\exp \left( 2~i~\theta _{n}\right) -\exp \left(
2~i~\theta _{\ell }\right) }\right] ~. \label{qn}$$
The next step is to introduce the time variable $t$. As in [@C2001], we assume hereafter that the $N$ seeds $s_{n}\left( \theta \right) $ are time-independent; we moreover assume the function $f\left( \theta \right) $ to be also time-independent (thereby simplifying the more general treatment of [@C2001]). A time-dependence is only introduced for the nodes $\theta
_{n}\equiv \theta _{n}\left( t\right) ;$ indeed they shall be the dependent variables of the dynamical systems we manufacture. Of course the fact that the nodes $\theta _{n}\left( t\right) $ evolve over time entails that the values $f_{n}$ taken by the function $f\left( \theta \right) $ at these nodes (see (\[fn\])) also evolve over time: $$f_{n}\equiv f_{n}\left( t\right) =f\left[ \theta _{n}\left( t\right) \right]
~. \label{fnt}$$
We then posit a convenient relation among the time evolution of the $N$ nodes $\theta _{n}\left( t\right) $ and the time evolution of the $N$ quantities $f_{n}\left( t\right) $, by setting$$f_{n}\left( t\right) =\rho _{n}\left[ \underline{\theta }\left( t\right) \right] ~\dot{\theta}_{n}\left( t\right) +\gamma _{n}\left[ \underline{\theta }\left( t\right) \right] ~. \label{qndot}$$Here we introduced the $2N$ functions $\rho _{n}\left( \underline{\theta }\right) $ and $\gamma _{n}\left( \underline{\theta }\right) $ of the $N$ nodes $\theta _{n}$, that will be assigned later at our convenience (but note that we forsake—again, for simplicity—the possibility to assign an *explicit* time-dependence to these functions, in addition to their dependence on the $N$ nodes).
The next step is to ascertain the time dependence of the $N$ nodes $\theta
_{n}\equiv \theta _{n}\left( t\right) $ implied by these assignments. To this end we time-differentiate the relation (\[qndot\]), getting the following expressions for the second time-derivatives of the $N$ nodes $\theta _{n}\equiv \theta _{n}\left( t\right) $:$$\rho _{n}\left( \underline{\theta }\right) ~\ddot{\theta}_{n}=\dot{f}_{n}-\sum_{m=1}^{N}\left\{ \left[ \frac{\partial ~\gamma _{n}\left(
\underline{\theta }\right) }{\partial ~\theta _{m}}+\frac{\partial ~\rho
_{n}\left( \underline{\theta }\right) }{\partial ~\theta _{m}}~\dot{\theta}_{n}\right] ~\dot{\theta}_{m}\right\} ~. \label{thetandotdot}$$
Our next step is to evaluate the quantity $\dot{f}_{n},$ which (see ([fnt]{})) reads$$\dot{f}_{n}=\frac{\partial ~f\left( \theta _{n}\right) }{\partial ~\theta
_{n}}~\dot{\theta}_{n}~.$$To evaluate this quantity we can use the finite-dimensional representation of the differential operator, yielding (for functions which are *linear* superpositions of the seeds $s_{n}\left( \theta \right) $, see ([GenLag]{})), the *exact* formula [@C2001]
$$\frac{\partial ~f\left( \theta _{n}\right) }{\partial ~\theta _{n}}=\sum_{m=1}^{N}\left[ D_{nm}\left( \underline{\theta }\right) ~f_{m}\right]
~,$$
with the $N\times N$ matrix $D$ defined componentwise as follows [@C2001]:$$D_{nm}\left( \underline{\theta }\right) =\frac{\partial ~q^{\left( m\right)
}(\theta ~\left\vert \underline{\theta }\right. )}{\partial ~\theta }~~~\text{evaluated at~~~}\theta =\theta _{n}~,$$hence in our case (see (\[seeds\]) and (\[Reprqn\])) reading
\[MatrixD\] $$D_{nm}\left( \underline{\theta }\right) =\delta _{nm}~\sum_{\ell =1,~\ell
\neq n}^{N}\cot \left( \theta _{n}-\theta _{\ell }\right) +\left( 1-\delta
_{nm}\right) ~\frac{\sigma _{n}\left( \underline{\theta }\right) }{\sigma
_{m}\left( \underline{\theta }\right) }~\frac{1}{\sin \left( \theta
_{n}-\theta _{m}\right) }~, \label{Dnm}$$$$\sigma _{n}\left( \underline{\theta }\right) =\dprod\limits_{\ell =1,~\ell
\neq n}^{N}\left[ \sin \left( \theta _{n}-\theta _{\ell }\right) \right] ~.
\label{sigma}$$Note that this definition coincides, via (\[rmwedgerndotzhat\]), with ([sigman]{}).
We therefore conclude that the system (\[thetandotdot\]) yields the following set of $N$ Newtonian equations of motion for the dependent variables $\theta _{n}\equiv \theta _{n}\left( t\right) $:
$$\begin{aligned}
&&\rho _{n}\left( \underline{\theta }\right) ~\ddot{\theta}_{n}=\dot{\theta}_{n}~\left[ \rho _{n}\left( \underline{\theta }\right) ~\dot{\theta}_{n}+\gamma _{n}\left( \underline{\theta }\right) \right] ~\sum_{\ell
=1,~\ell \neq n}^{N}\left[ \cot \left( \theta _{n}-\theta _{\ell }\right) \right] \notag \\
&&+\dot{\theta}_{n}~\sum_{\ell =1,~\ell \neq n}^{N}\left\{ \frac{\sigma
_{n}\left( \underline{\theta }\right) }{\sigma _{\ell }\left( \underline{\theta }\right) }~\frac{\left[ \rho _{\ell }\left( \underline{\theta }\right) ~\dot{\theta}_{\ell }+\gamma _{\ell }\left( \underline{\theta }\right) \right] }{\sin \left( \theta _{n}-\theta _{\ell }\right) }\right\}
\notag \\
&&-\sum_{m=1}^{N}\left\{ \left[ \frac{\partial ~\rho _{n}\left( \underline{\theta }\right) }{\partial ~\theta _{m}}~\dot{\theta}_{n}+\frac{\partial
~\gamma _{n}\left( \underline{\theta }\right) }{\partial ~\theta _{m}}\right]
~\dot{\theta}_{m}\right\} ~. \label{GenNbody}\end{aligned}$$
Of course to obtain this system of $N$ second-order ODEs we also used ([qndot]{}).
Let us now emphasize that, as a consequence of the way these $N$-body problems have been manufactured, they are *integrable*$.$ It is indeed plain that the time independence of the function $f\left( \theta
\right) $ entails (via (\[fhn\]), (\[fn\]) and (\[qndot\])) the relations
\[FirstOrderODEs\] $$\sum_{m=1}^{N}\left\{ h_{m}~s_{m}\left[ \theta _{n}\left( t\right) \right]
\right\} =\rho _{n}\left[ \underline{\theta }\left( t\right) \right] ~\dot{\theta}_{n}\left( t\right) +\gamma _{n}\left[ \underline{\theta }\left(
t\right) \right] ~. \label{hthetadot}$$Here we have displayed the *time-dependence* of the various quantities, in order to emphasize the *time-independence* of the $N$ coefficients $h_{m}$, which can actually be evaluated by solving this system of $N$ *linear* equations, thereby obtaining (via (\[Reprqn\])) the following formulas:$$h_{m}=q^{\left( m\right) }\left( \vartheta _{m}~\left\vert \underline{\theta
}\right. \right) ~,~~~~\vartheta _{m}\equiv \frac{i~\log \left[ \rho
_{m}\left( \underline{\theta }\right) ~\dot{\theta}_{m}+\gamma _{m}\left(
\underline{\theta }\right) \right] }{2~m-N-1}~, \label{hn}$$where of course the $N$ nodes $\theta _{m}\equiv \theta _{m}\left( t\right) $ and their $N$ time derivatives $\dot{\theta}_{m}\equiv \dot{\theta}_{m}\left( t\right) $ can be evaluated at any arbitrary time $t$. It is thus plain that the $N$-body systems (\[GenNbody\]) are *integrable* for any *arbitrary* assignment of the $2N$ functions $\rho _{m}\left(
\underline{\theta }\right) $ and $\gamma _{m}\left( \underline{\theta }\right) $ of the $N$ dependent variables $\theta _{n}$, with these $N$ quantities $h_{m}$ providing $N$ *constants of motion* given by explicit (generally nontrivial) expressions in terms of the $N$ nodes $\theta _{n}$ and their $N$ time-derivatives $\dot{\theta}_{n}$.
We are still free to assign the $2N$ functions $\rho _{n}\left( \underline{\theta }\right) $ and $\gamma _{n}\left( \underline{\theta }\right) .$ There are two natural choices.
The first one reads simply
$$\rho _{n}\left( \underline{\theta }\right) =\mu _{n}\text{ ,~~~}\gamma
_{n}\left( \underline{\theta }\right) =\eta _{n}~, \label{rhogammaconstants}$$
with $\mu _{n}$ and $\eta _{n}$ arbitrary *constant* parameters. It clearly yields (see (\[GenNbody\])) an $N$-body system characterized by the following set of Newtonian equations of motion:$$\begin{aligned}
&&\mu _{n}~\ddot{\theta}_{n}=\dot{\theta}_{n}~\left( \mu _{n}~\dot{\theta}_{n}+\eta _{n}\right) ~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \cot \left(
\theta _{n}-\theta _{\ell }\right) \right] \notag \\
&&+\dot{\theta}_{n}~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \frac{\sigma
_{n}\left( \underline{\theta }\right) }{\sigma _{\ell }\left( \underline{\theta }\right) }~\frac{\left( \mu _{\ell }~\dot{\theta}_{\ell }+\eta _{\ell
}\right) }{\sin \left( \theta _{n}-\theta _{\ell }\right) }~\right] ~.
\label{WithManyBodyForces}\end{aligned}$$Here the functions $\sigma _{n}\left( \underline{\theta }\right) $ of the $N$ nodes $\theta _{m}$ are of course defined by (\[sigma\]).
The second assignment of the $2N$ functions $\rho _{n}\left( \underline{\theta }\right) $ and $\gamma _{n}\left( \underline{\theta }\right) $ is suggested by the structure of the system (\[GenNbody\]). It reads$$\rho _{n}\left( \underline{\theta }\right) =\mu _{n}\text{~}\sigma
_{n}\left( \underline{\theta }\right) \text{,~\ ~}\gamma _{n}\left(
\underline{\theta }\right) =\eta _{n}~\sigma _{n}\left( \underline{\theta }\right) ~, \label{rhogmmasigma}$$where again $\mu _{n}$ and $\eta _{n}$ are arbitrary *constant* parameters and the functions $\sigma _{n}\left( \underline{\theta }\right) $ are defined as above, see (\[sigma\]), implying (by logarithmic differentiation)
$$\frac{\partial ~\gamma _{n}\left( \underline{\theta }\right) }{\partial
~\theta _{m}}=\gamma _{n}\left( \underline{\theta }\right) ~\left\{ \delta
_{nm}~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \cot \left( \theta _{n}-\theta
_{\ell }\right) \right] -\left( 1-\delta _{nm}\right) ~\cot \left( \theta
_{n}-\theta _{m}\right) \right\} ~,$$
and likewise$$\frac{\partial ~\rho _{n}\left( \underline{\theta }\right) }{\partial
~\theta _{m}}=\rho _{n}\left( \underline{\theta }\right) ~\left\{ \delta
_{nm}~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \cot \left( \theta _{n}-\theta
_{\ell }\right) \right] -\left( 1-\delta _{nm}\right) ~\cot \left( \theta
_{n}-\theta _{m}\right) \right\} ~.$$Thereby the $N$-body system gets characterized by the following, simpler set of Newtonian equations of motion:
$$\mu _{n}~\ddot{\theta}_{n}=\sum_{\ell =1,~\ell \neq n}^{N}\left[ \frac{\dot{\theta}_{n}~\left( \mu _{\ell }~\dot{\theta}_{\ell }+\eta _{\ell }\right)
+\left( \mu _{n}~\dot{\theta}_{n}+\eta _{n}\right) ~\dot{\theta}_{\ell
}~\cos \left( \theta _{n}-\theta _{\ell }\right) }{\sin \left( \theta
_{n}-\theta _{\ell }\right) }\right] ~. \label{WithTwoBodyForces}$$
The differences among these two $N$-body systems, (\[WithManyBodyForces\]) and (\[WithTwoBodyForces\]), deserve to be emphasized: the $N$-body model (\[WithManyBodyForces\]) involves *many-body* forces, due to the presence of the functions $\sigma _{n}\left( \underline{\theta }\right) $ and $\sigma _{\ell }\left( \underline{\theta }\right) $ in its right-hand (“forces”) side; while the $N$-body model (\[WithTwoBodyForces\]) only involves *two-body* forces. Both systems can be integrated once, corresponding to the transition from their $N$ *second-order* Newtonian equations of motion to the corresponding $N$ *first-order* ODEs (\[hthetadot\]). On the other hand, as we show below, only the first of these two *integrable* systems is *solvable*.
Indeed, for the first system (but not for the second!), the $N$ first-order ODEs (\[hthetadot\]) are *uncoupled*, reading simply, via ([rhogammaconstants]{}),
\[ODEsForThetan\] $$\mu _{n}~\dot{\theta}_{n}=-\eta _{n}+\sum_{m=1}^{N}\left[ h_{m}~s_{m}\left(
\theta _{n}\right) \right] ~,$$or, equivalently (see (\[seeds\]))$$\mu _{n}~\exp \left[ \left( N+1\right) ~i~\theta _{n}\right] ~\dot{\theta}_{n}=-\eta _{n}~\exp \left[ \left( N+1\right) ~i~\theta _{n}\right]
+\sum_{m=1}^{N}\left[ h_{m}~\exp \left( 2~m~i~\theta _{n}\right) \right] ~,$$where the $N$ quantities $h_{n}$ are explicitly known in terms of the $2N$ initial data $\theta _{n}\left( 0\right) $, $\dot{\theta}_{n}\left( 0\right)
$ (via (\[hn\]), (\[rhogammaconstants\]) and (\[qn\]): see Appendix B).
These first-order ODEs can be integrated; we confine the relevant developments to Appendix B.
Although the technique to manufacture these two *solvable* and *integrable* $N$-body problems, (\[WithManyBodyForces\]) and ([WithTwoBodyForces]{}), is *not* new [@C2001], these models are, to the best of our knowledge, themselves *new*; hence a detailed discussion of the actual behavior of these systems has not yet been done. In the present paper we limit our consideration to pointing out how these models can be reformulated to describe the evolution of $N$ points whose positions on a plane are characterized by $N$ *unit* 2-vectors $\vec{r}_{n}\left( t\right) $, see the notation introduced in Subsection 2.1. To this end one utilizes the formulas (\[rndotdot\]), (\[rndotrm\]), ([rmwedgerndotzhat]{}) and the relevant ones among those conveniently collected in Appendix A. And it is plain that one thereby obtains the two models ([ManyBodyForcesModel]{}) and (\[TwoBodyForcesOnCircle\]).
Solvable models on the circle manufactured by reinterpreting known solvable models
----------------------------------------------------------------------------------
In this section we tersely indicate how to obtain the two models (\[2a\]) and (\[2b\]).
The *first model* obtains from the $N$-body system characterized by the following Newtonian equations of motion (with velocity-independent two-body forces):
$$\ddot{\theta}_{n}=g^{2}~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \frac{\cos
\left( \theta _{n}-\theta _{\ell }\right) }{\sin ^{3}\left( \theta
_{n}-\theta _{\ell }\right) }\right] ~. \label{Suth}$$
Here $g$ is an arbitrary “coupling constant”, and the rest of the notation is, we trust, clear.
This is a well-known *solvable* many-body problem, generally associated with the name of Bill Sutherland, who was the first to show the possibility to treat this $N$-body problem by exact methods (originally in a quantal context [@S]); its treatment in a classical (Hamiltonian) context is provided in several textbooks, see for instance [@P1990] [C2001]{} [@S2004].
It is plain that the model (\[2a\]) is merely the transcription of this model via the notation of Subsection 2.1.
The *second model* obtains from the $N$-body system characterized by the following Newtonian equations of motion (with velocity-dependent one-body and two-body forces):$$\ddot{\theta}_{n}=g_{0}+g_{1}~\dot{\theta}_{n}+\sum_{\ell =1,~\ell \neq
n}^{N}\left\{ \left[ 2~\dot{\theta}_{n}~\dot{\theta}_{\ell }+g_{2}~\left(
\dot{\theta}_{n}+\dot{\theta}_{\ell }\right) +g_{3}\right] ~\cot \left(
\theta _{n}-\theta _{\ell }\right) \right\} ~.$$Here $g_{0},$ $g_{1},$ $g_{2}$ and $g_{3}$ are $4$ arbitrary coupling constants, and we again trust the rest of the notation to be clear.
This is also a well known *solvable* model, see for instance eq. (2.3.5-12) on page 199 of [@C2001].
And it is again plain that the model (\[2b\]) is merely the transcription of this model via the notation of Subsection 2.1 and Appendix A.
How to manufacture $N$-body problems with *angles* as dependent variables
-------------------------------------------------------------------------
In the preceding subsection we have shown how certain $N$-body models with dependent variables naturally interpretable as *angles* can be reformulated as $N$-body models describing the time evolution on a plane of particles *constrained to move on a circle*. In this subsection we indicate how, via a simple change of dependent variables, essentially *any* $N$-body model can be reformulated so that its dependent variables can be interpreted as *angles*, hence subsequently it can also be reformulated (in fact in many ways) so that it describes the time evolution of particles *constrained to move on a plane circle*.
The trick to achieve this goal is quite elementary and general; we illustrate it below via two examples.
Consider an $N$-body model in which the positions of the $N$ point-particles—moving in one-dimensional space—are identified by $N$ coordinates $z_{n}\equiv z_{n}\left( t\right) ,$ and perform the change of dependent variables by positing, say,$$z_{n}\left( t\right) =\tan \left[ \theta _{n}\left( t\right) \right] ~.
\label{ZnTanThetan}$$
*Remark 3.3.1*. Of course this assignment defines $\theta _{n}\left(
t\right) $ only $\func{mod}\left( \pi \right) $; and clearly many other assignments could be instead made—different but having an analogous effect, such as $z_{n}=1/\sin \left( 2\theta _{n}\right) $, or $z_{n}=\tan
^{3}\theta _{n}$, etc. . $\blacksquare $
In the *first example* we take as point of departure the $N$-body problem characterized by the Newtonian equations of motion
$$\ddot{z}_{n}=-4~z_{n}+g^{2}~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \left(
z_{n}-z_{\ell }\right) ^{-3}\right] ~. \label{Cal}$$
Here $g$ is an arbitrary (real) coupling constant. This is a well-known *solvable* model (see for instance [@C2001]); it is *isochronous*, all its solutions being *completely periodic with period* $\pi $, $$z_{n}\left( t\pm \pi \right) =z_{n}\left( t\right) ~. \label{iso}$$
Via the change of dependent variables (\[ZnTanThetan\]) the equations of motion (\[Cal\]) become (as the diligent reader will easily verify, utilizing if need be the identities reported in the last part of Appendix A)
$$\begin{aligned}
&&\ddot{\theta}_{n}=-2~\dot{\theta}_{n}^{2}~\tan \theta _{n}-4~\sin \theta
_{n}~\cos \theta _{n} \notag \\
&&+g^{2}~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \frac{\cos ^{5}\theta
_{n}~\sin ^{3}\theta _{\ell }}{\sin ^{3}\left( \theta _{n}-\theta _{\ell
}\right) }\right] ~. \label{3a}\end{aligned}$$
*Remark 3.3.2*. This model of course hereditates the property of *isochrony* of the model (\[Cal\]) it has been obtained from:$$\theta _{n}\left( t\pm \pi \right) =\theta _{n}\left( t\right) ~~~\func{mod}\left( \pi \right) ~.~\blacksquare$$
The next task is to transform these equations of motion, (\[3a\]), into equations of motion for points moving in the plane but constrained to stay on a *circle* of *unit* radius centered at the origin. To realize this goal one may now use the change of dependent variables from the *angles* $\theta _{n}$ to the *vectors* $\vec{r}_{n}$ described in Subsection 2.1, using if need be the identities reported in the first part of Appendix A. And it is plain that in this manner one arrives at the equations of motion (\[CalCircle\]).
In the *second example* we take as point of departure the well-known *solvable* $N$-body problem characterized by the following Newtonian equations of motion (see eq. (2.3.4.2-1) on page 188 of [@C2001]):
$$\ddot{z}_{n}=-z_{n}+\sum_{\ell =1,~\ell \neq n}^{N}\left( \frac{2~\dot{z}_{n}~\dot{z}_{\ell }+1}{z_{n}-z_{\ell }}\right) ~. \label{Gold}$$
*All* solutions of this model are *multiply periodic*, being (generally nonlinear) superpositions of the $N$ functions $b_{m}\left(
t\right) =\cos \left( \sqrt{m}~t+\beta _{m}\right) ,$ $m=1,...,N$ (with the $N$ phases $\beta _{m}$ depending on the initial data); for special initial data only functions $b_{m}\left( t\right) $ with $m$ a *squared-integer* contribute, yielding solutions *completely periodic* with period $2\pi $. [@C2001]
Via the change of dependent variables (\[ZnTanThetan\]) equations of motion (\[Gold\]) become (as the diligent reader will easily verify, utilizing again, if need be, the identities reported in the last part of Appendix A)$$\begin{aligned}
&&\ddot{\theta}_{n}=-2~\dot{\theta}_{n}^{2}~\tan \theta _{n}-\sin \theta
_{n}~\cos \theta _{n} \notag \\
&&+\cos \theta _{n}~\sum_{\ell =1,~\ell \neq n}^{N}\left[ \frac{2~\dot{\theta}_{n}~\dot{\theta}_{\ell }+\cos ^{2}\theta _{n}~\cos ^{2}\theta _{\ell }}{\cos \theta _{n}~\sin \left( \theta _{n}-\theta _{\ell }\right) }\right] ~.\end{aligned}$$Then we transform these equations of motion into equations of motion for points moving in the plane but constrained to stay on a *circle* of *unit* radius centered at the origin, by using again the change of dependent variables from the *angles* $\theta _{n}$ to the *vectors* $\vec{r}_{n}$ described in Subsection 2.1 via—if need be—the identities reported in the first part of Appendix A. And it is plain that in this manner one arrives at the equations of motion (\[GoldCircle\]).
Outlook
=======
Our original motivation to undertake this line of research was the intention to manufacture $N$-body problems amenable to exact treatments describing motions on a sphere, or more generally on manifolds. We consider the results reported in this paper as a modest first step in that direction. We also believe that the actual behavior of the *new* models reported in this paper—see (\[ManyBodyForcesModel\]) and (\[TwoBodyForcesOnCircle\])—shall eventually deserve a more detailed scrutiny than that provided in Subsection 3.1.
Appendix A: identities
======================
It is plain that the notation introduced in Subsection 2.1 entails the following additional identities:
$$\overset{\cdot }{\vec{r}}_{n}\cdot \vec{r}_{n}=0~,~~~\overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{n}=\dot{\theta}_{n}^{2},~~~\left( \vec{r}_{n}\wedge \overset{\cdot }{\vec{r}}_{n}\right) \cdot \hat{z}=\dot{\theta}_{n}~, \label{Ardotscalar}$$
$$\overset{\cdot \cdot }{\vec{r}}_{n}\cdot \vec{r}_{n}=-\dot{\theta}_{n}^{2}~,~~~\overset{\cdot \cdot }{\vec{r}}_{n}\cdot \left( \hat{z}\wedge
\vec{r}_{n}\right) =\ddot{\theta}_{n}~, \label{Ardotdotscalar}$$
$$\overset{\cdot }{\vec{r}}_{n}\cdot \vec{r}_{m}=-\dot{\theta}_{n}~\sin \left(
\theta _{n}-\theta _{m}\right) ~, \label{Arndotrmscal}$$
$$\overset{\cdot }{\vec{r}}_{n}\cdot \overset{\cdot }{\vec{r}}_{m}=\dot{\theta}_{n}~\dot{\theta}_{m}~\cos \left( \theta _{n}-\theta _{m}\right) ~;
\label{Arndotrmdotscal}$$
$$\hat{z}\wedge \overset{\cdot }{\vec{r}}_{n}=-\dot{\theta}_{n}~\vec{r}_{n}~,
\label{Azhatvectrdot}$$
$$\hat{z}\wedge \overset{\cdot \cdot }{\vec{r}}_{n}=-\ddot{\theta}_{n}~\vec{r}_{n}-\dot{\theta}_{n}^{2}~\hat{z}\wedge \vec{r}_{n}~;
\label{Azhatvectrdotdot}$$
$$\left( \overset{\cdot }{\vec{r}}_{n}\wedge \vec{r}_{m}\right) \cdot \hat{z}=-\dot{\theta}_{n}~\cos \left( \theta _{n}-\theta _{m}\right) ~,
\label{Arndotrmvect}$$
$$\left( \overset{\cdot }{\vec{r}}_{n}\wedge \overset{\cdot }{\vec{r}}_{m}\right) \cdot \hat{z}=-\dot{\theta}_{n}~\dot{\theta}_{m}~\sin \left(
\theta _{n}-\theta _{m}\right) ~. \label{Arndotrmdotvect}$$
We also display here some relations among the time-dependent “coordinates”
$$z_{n}\equiv z_{n}\left( t\right) =\tan \theta _{n}\left( t\right) ~,
\label{Axn}$$
and the “angles” $\theta _{n}\equiv \theta _{n}\left( t\right) $:$$z_{n}-z_{m}=\frac{\sin \left( \theta _{n}-\theta _{m}\right) }{\cos \theta
_{n}~\cos \theta _{m}}~,~~~\frac{1}{z_{n}-z_{m}}=\frac{\cos \theta _{n}~\cos
\theta _{m}}{\sin \left( \theta _{n}-\theta _{m}\right) }~;
\label{Aznminuszm}$$
$$\dot{z}_{n}=\frac{\dot{\theta}_{n}}{\cos ^{2}\theta _{n}}~,~~~\dot{z}_{n}~z_{m}=\frac{\dot{\theta}_{n}~\sin \theta _{m}}{\cos ^{2}\theta
_{n}~\cos \theta _{m}}~,~~~\dot{z}_{n}~\dot{z}_{m}=\frac{\dot{\theta}_{n}~\dot{\theta}_{m}}{\cos ^{2}\theta _{n}~\cos ^{2}\theta _{m}}~;
\label{Azndot}$$
\[Azndotzm\] $$\frac{\dot{z}_{n}+\dot{z}_{m}}{z_{n}-z_{m}}=\frac{\dot{\theta}_{n}~\cos
^{2}\theta _{m}+\dot{\theta}_{m}~\cos ^{2}\theta _{n}}{\cos \theta _{n}~\cos
\theta _{m}~\sin \left( \theta _{n}-\theta _{m}\right) }~,$$$$\frac{\dot{z}_{n}~z_{m}+\dot{z}_{m}~z_{n}}{z_{n}-z_{m}}=\frac{\dot{\theta}_{n}~\sin \theta _{m}~\cos \theta _{m}+\dot{\theta}_{m}~\sin \theta
_{n}~\cos \theta _{n}}{\cos \theta _{n}~\cos \theta _{m}~\sin \left( \theta
_{n}-\theta _{m}\right) }~,$$$$\frac{\dot{z}_{n}~\dot{z}_{m}}{z_{n}-z_{m}}=\frac{\dot{\theta}_{n}~\dot{\theta}_{m}}{\cos \theta _{n}~\cos \theta _{m}~\sin \left( \theta
_{n}-\theta _{m}\right) }~;$$
$$\ddot{z}_{n}=\frac{\ddot{\theta}_{n}}{\cos ^{2}\theta _{n}}+\frac{2~\dot{\theta}_{n}^{2}~\sin \theta _{n}~}{\cos ^{3}\theta _{n}}=\frac{\ddot{\theta}_{n}+2~\dot{\theta}_{n}^{2}~\tan \theta _{n}}{\cos ^{2}\theta _{n}}~.
\label{Azndotdot}$$
Appendix B: solution of the system (\[ODEsForThetan\])
======================================================
In this Appendix we indicate how the initial-value problem of the system of $N$ (decoupled) first-order ODEs (\[ODEsForThetan\]) is solved.
Let us, for notational convenience, make here the following change of variables:
$$\zeta _{n}\left( t\right) =\exp \left[ i~\theta _{n}\left( t\right) \right]
~, \label{Bzn}$$
entailing$$\dot{\zeta}_{n}\left( t\right) =i~\dot{\theta}_{n}\left( t\right) ~\exp \left[ i~\theta _{n}\left( t\right) \right] ~.$$
We then use the relation (\[Bzn\]) to rewrite the equations of motion ([ODEsForThetan]{}) as follows:
$$\mu ~\zeta ^{N}~\dot{\zeta}=i~\left[ -\eta ~\zeta
^{N+1}+\sum_{m=1}^{N}\left( h_{m}~\zeta ^{2m}\right) \right] ~. \label{ODE}$$
*Remark B.1*. Let us emphasize that, in the last formula and below (in this Appendix B), as a notational simplification, we *omit* to indicate explicitly the time-dependence of the dependent variable $\zeta
_{n}\equiv \zeta _{n}\left( t\right) $, as well as its dependence on the index $n$; and likewise the dependence on this index $n$ of the parameters $\mu _{n}$ and $\eta _{n}$. $\blacksquare $
The ODE (\[ODE\]) can clearly be solved by the following quadrature:$$\dint\limits_{\zeta \left( 0\right) }^{\zeta \left( t\right) }d\xi ~\xi
^{N-2}~\left\{ -\eta ~\xi ^{N-1}+\sum_{m=1}^{N}\left[ h_{m}~\xi ^{2\left(
m-1\right) }\right] \right\} ^{-1}=\frac{i~t}{\mu }~. \label{zt}$$
To perform the integration it is convenient to introduce the $2\left(
N-1\right) $ zeros $\xi _{j}$ of the polynomial of degree $2\left(
N-1\right) $ appearing in the denominator of the integrand,
$$-\eta ~\xi ^{N-1}+\sum_{m=1}^{N}\left[ h_{m}~\xi ^{2\left( m-1\right) }\right] =h_{N}~\dprod\limits_{j=1}^{2\left( N-1\right) }\left( \xi -\xi
_{j}\right) ~,$$
and then the $2\left( N-1\right) $ “residues” $\phi _{j}$ defined by setting$$\left\{ -\eta ~\xi ^{N-1}+\sum_{m=1}^{N}\left[ h_{m}~\xi ^{2\left(
m-1\right) }\right] \right\} ^{-1}=h_{N}^{-1}~\dsum\limits_{j=1}^{2\left(
N-1\right) }\left( \frac{\phi _{j}}{\xi -\xi _{j}}\right) ~. \label{Res}$$Note that these formulas imply that the computation of, firstly, the $2\left( N-1\right) $ zeros $\xi _{j},$ and, secondly, the $2\left(
N-1\right) $ residues $\phi _{j},$ is a purely *algebraic* task (although not one that can be analytically performed for $N\geq 3$); hence these quantities can in principle be considered known functions of the parameter $\eta $ (from which they inherit a dependence on the index $n$, see *Remark B.1*) and of the $N$ constants of motion $h_{m}$. As for these $N$ quantities $h_{m}$ (which are of course independent of the index $n
$) they are—in the context of the *initial-value* problem for the dynamical system (\[WithManyBodyForces\])—explicitly given by the formulas (\[hn\]) at $t=0$ (let us reiterate that these expressions of the $N$ constants of motion $h_{m}$ are valid throughout the time evolution, and of course, in particular, at the *initial* time $t=0$).
The final step is to perform the integration in the left-hand side of ([zt]{}). Via (\[Res\]) the key ingredient to do so is the formula
$$\begin{aligned}
&&\dint\limits_{\zeta _{0}}^{\zeta }d\xi ~\frac{\xi ^{N-2}}{\xi -\xi _{0}}=\dint\limits_{\zeta _{0}-\xi _{0}}^{\zeta -\xi _{0}}d\xi ~\frac{\left( \xi
+\xi _{0}\right) ^{N-2}}{\xi } \notag \\
&=&\dint\limits_{\zeta _{0}-\xi _{0}}^{\zeta -\xi _{0}}d\xi ~\sum_{k=0}^{N-2}
\left[ \binom{N-2}{k}~\xi ^{k-1}~\xi _{0}^{N-2-k}\right] \notag \\
&=&\xi _{0}^{N-2}~\log \left( \frac{\zeta -\xi _{0}}{\zeta _{0}-\xi _{0}}\right) +\sum_{k=1}^{N-2}\left\{ \binom{N-2}{k}~\frac{\xi _{0}^{N-2-k}}{k}~\left[ \left( \zeta -\xi _{0}\right) ^{k}-\left( \zeta _{0}-\xi _{0}\right)
^{k}\right] \right\} ~. \notag \\
&&\end{aligned}$$
[9]{} L. Auslander and L. Markus, “Classical differential equations on manifolds”, Trans. Amer. Math. Soc. **91**, 113-128 (1959).
P. E. Crouch and R. Grossman, “Numerical integration of ordinary differential equations on manifolds”, J. Nonlinear Sci. **3**, 1-33 (1993).
E. Hairer, *Solving Differential Equations on Manifolds*, Lecture Notes, Université de Genève, 2011.
F. Erman and O. T. Turgut, “A many-body problem with point-interactions on two-dimensional manifolds”, arXiv:1204.2171v2\[math-ph\]12Apr2012.
F. Calogero, *Classical Many-Body Problems Amenable to Exact Treatments*, Lecture Notes in Physics Monographs **m66**, Springer, Berlin, 2001.
F. Calogero, D. Gómez-Ullate, P. M. Santini and M. Sommacal, The transition from regular to irregular motions, explained as travel on Riemann surfaces, J. Phys. A: Math. Gen. **38**, 8873-8896 (2005); Towards a theory of chaos explained as travel on Riemann surfaces, J. Phys. A.: Math. Theor. **42**, 015205 (26 pages) (2009); D. Gómez-Ullate, P. M. Santini, M. Sommacal and F. Calogero, Understanding complex dynamics by means of an associated Riemann surface, Physica D **241**, 1291-1305 (2012).
B. Sutherland, “Exact results for a quantum many-body problem in one dimension”, Phys. Rev. **A4**, 2019-2021 (1971); “Exact results for a quantum many-body problem in one dimension. II”, Phys. Rev. **A5**, 1372-1376 (1972).
A. M. Perelomov, *Integrable systems of classical mechanics and Lie algebras*, Birkhäuser, Basel, 1990.
B. Sutherland, *Beautiful models - 70 years of exactly solved quantum many-body problems*, World Scientific, Singapore, 2004.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present some new analytical tools for the error analysis of hybridizable discontinuous Galerkin (HDG) method for linear elasticity. These tools allow us to analyze more variants of HDG method using the projection-based approach, which renders the error analysis simple and concise. The key result is a tailored projection for the Lehrenfeld-Schöberl type HDG (HDG+ for simplicity) methods. By using the projection we recover the error estimates of HDG+ for steady-state and time-harmonic elasticity in a simpler analysis. We also present a semi-discrete (in space) HDG+ method for transient elastic waves and prove it is uniformly-in-time optimal convergent by using the projection-based error analysis. Numerical experiments supporting our analysis are presented at the end.'
address:
- University of Delaware
- University of Delaware
author:
- Shukai Du
- 'Francisco-Javier Sayas'
bibliography:
- 'references.bib'
title: 'New analytical tools for HDG in elasticity, with applications to elastodynamics'
---
Introduction
============
The paper is devoted to present some new techniques for the a priori error analysis of a new class of hybridizable discontinuous Galerkin methods. The methods in this class use a special type of stabilization function that was first introduced by Lehrenfeld and Schöberl in [@LeSc:2010]. We will call them HDG+ methods for simplicity. Instead of attempting to reach for maximal generality, we will focus on linear elasticity on tetrahedral meshes.
We begin by reviewing some existing works. The first HDG method for linear elasticity was proposed in [@SoCoSt:2009]. The method enforces the symmetry of the stress strongly and uses order $k$ polynomial spaces for all variables. It was then proved in [@FuCoSt:2015] that the method is optimal for displacement but only suboptimal for stress (order of $k+\frac{1}{2}$); they also showed the order is sharp on triangular meshes in the numerical experiments. To recover optimal convergence (based on which the superconvergence by post-processing is possible), there are mainly three approaches. The first approach relaxes the strong symmetry condition to weak symmetry [@CoSh:2013]. In general, mixed finite element methods based on weak symmetric stress formulations are relatively easier to implement but use more degrees of freedom and therefore can be more costly to compute. The second and the third approaches are all based on strong symmetric stress formulations, where the conservation of angular momentum is automatically preserved. The second approach [@CoFu:2018] uses $M$-decomposition [@CoFuSa:2017] to enrich the approximation space for stress by adding some basis functions. The approach recovers optimal convergence and also provides an associated tailor projection as an useful tool for error analysis. However, the added basis can be rational functions instead of polynomials and therefore can lead to some difficulties in implementation.
The focus of this paper is the third approach, which is to use HDG+ method for linear elasticity. The method was originally proposed in [@LeSc:2010] for diffusion problems, then applied to steady-state linear elasticity in [@QiShSh:2018]. It uses only polynomial basis functions, achieves one order higher convergence rate for the displacement without post-processing, and its computational complexity is the same to the standard HDG method (order $k$ for both stress and displacement) for their global systems. However, the existing error analysis of HDG+ methods are all based on using orthogonal projections [@HuPrSa:2017; @Oi:2015; @QiShSh:2018; @QiSh:2016; @QiSh2:2016], which make the analysis slightly more complicated (it requires a bootstrapping argument to prove convergence of all variables, as opposed to consecutive energy and duality proofs), and detached from the existing projection-based error analysis of HDG methods [@ChCo:2012; @CoFuHuJiSaMaSa:2018; @CoGoSa:2010; @CoMu:2015; @CoQiSh2012; @CoQu:2014; @CoSh:2013; @GrMo:2011], where specifically constructed projections are used to make the analysis simple and concise. This motivates us to find a new kind of projection for HDG+ for elasticity. The goal of the projection is twofold: first of all, it takes care of all the off-diagonal terms in the matrix form of the equations (except for a $2\times 2$ block which is considered as a single diagonal term), and allows us to do a simple energy estimate for some of the variables; second, it facilitates a duality argument where the adjoint equation is fed with the missing error terms to estimate, by using the adjoint projection (consisting of a simple change of sign in the stabilization parameter $\boldsymbol\tau$). This has been done in [@DuSa:2019] for diffusion problems. A novelty of this paper is the fact that we complement the projection with an error term that does not affect the error bounds or the simplicity of their proofs. We have attached this error term to the projection to make the arguments simpler.
In summary, we have devised a projection for the HDG+ methods for linear elasticity. The projection enables us to: (1) recycle existing projection-based error analysis techniques for the analysis of HDG+ methods; (2) make the error analyis simple and concise; (3) build connections between $M$-decompositions [@CoFu:2018; @CoFuSa:2017] and HDG+ methods. To be more specific, we present a semi-discrete HDG+ method for transient elastic waves that has a uniform-in-time optimal convergence. We show that the proof for optimal convergence can be easily obtained by using the new projection and some existing techniques in traditional HDG methods for evolutionary equations [@CoFuHuJiSaMaSa:2018]. Moreover, we recover the error estimates for steady-state elasticity [@QiShSh:2018] and frequency domain elastodynamics [@HuPrSa:2017] by using the projection-based analysis, and we show that the analysis can be simplified in both cases. Since the construction of the HDG+ projection involves first constructing a projection associated to an $M$-decomposition, it also shed some light upon the connections between these two kinds of methods.
To provide a more intuitive view, we put the main procedures of constructing the projection in the flow chart Figure \[fig:flow\_chart\].
(init) at (0,3) [${\mathcal}P_k^\mathrm{sym}\times {\mathcal}P_k$]{}; (Mdecomp) at (0,2) [$({\mathcal}P_k^\mathrm{sym}\oplus\Sigma_\mathrm{fillM})\times {\mathcal}P_k$]{}; (weakcom) at (0,1) [$({\mathcal}P_k^\mathrm{sym}\oplus\Sigma_\mathrm{fillM})\times {\mathcal}P_{k+1}$]{}; (HDGplus) at (0,0) [${\mathcal}P_k^\mathrm{sym}\times {\mathcal}P_{k+1}$]{}; (bdrem) at (2.25,0) [${\mathcal}R_k$]{}; (HDGplus\_p) at (0,-1) [${\mathcal}P_k^\mathrm{sym}(K)\times {\mathcal}P_{k+1}(K)$]{}; (bdrem\_p) at (2.25,-1) [${\mathcal}R_k({\partial}K)$]{};
(init) to node \[right\] [$M$-decomposition (Section \[sec:proj\_ref\].1)]{} (Mdecomp); (Mdecomp) to node \[right\] [weak commutativity (Section \[sec:proj\_ref\].2)]{} (weakcom); (weakcom) to node \[left\] [$L^2$ projection (Section \[sec:proj\_ref\].3)]{} (HDGplus); (weakcom) to node \[right\] [boundary remainder (Section \[sec:proj\_ref\].3)]{} (bdrem); (HDGplus) to node \[left\] [push forward (Section \[sec:proj\_phy\])]{} (HDGplus\_p); (bdrem) to node \[right\] [push forward (Section \[sec:proj\_phy\])]{} (bdrem\_p);
The associated boundary remainder term related to the projection behaves like interpolation error and it depends only on the local projection. As we will see later in the applications, the boundary remainder together with the stabilization parameter play a key role in the optimal convergence of the HDG+ methods, allowing us more flexibility, since now we only need to find a projection such that its associated boundary remainder is small enough to guarantee optimal convergence, instead of enforcing it to vanish, which is the case of the standard projection.
The rest of the paper is organized as follows. In Section \[sec:proj\], we present the main theorem about the projection and its properties. In Section \[sec:proj\_ref\], we show the procedures for constructing the projection on the reference element. In Section \[sec:proj\_phy\], we develop a systematic approach of changes of variables to obtain the projection on the physical element. In Section \[sec:steady\_state\] and \[sec:freq\], we recover the error estimates in steady-state elasticity [@QiShSh:2018] and elasto-dynamics [@HuPrSa:2017] using projection-based analysis. In Section \[sec:trans\_elas\], we present a semi-discrete HDG method for transient elastic waves and prove it is optimally convergent, uniformly in the time variable. Finally, we give some numerical experiments to support our analysis.
The projection {#sec:proj}
==============
Since the main tool and one of the principal novelties of this article is the new HDG projection for elasticity, we first introduce its main properties in Theorem \[th:PROJ\]. The reader just interested in the applications can skip the sections devoted to its construction and analysis (Section \[sec:proj\_ref\] and \[sec:proj\_phy\]) and jump directly to how it is used (Section \[sec:steady\_state\], \[sec:freq\] and \[sec:trans\_elas\]). To speed up the introduction of the projection we give a quick notation list to be used throughout the article.
- For a domain $\mathcal O\subset {{\mathbb R}}^3$, the respective inner products of $L^2(\mathcal O;{{\mathbb R}}^3)$ and $L^2(\mathcal O;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ will be denoted $$({\boldsymbol{u}},{\boldsymbol{v}})_{\mathcal O}:=\int_{\mathcal O} {\boldsymbol{u}}\cdot{\boldsymbol{v}},
\qquad
({\boldsymbol{\sigma}},{\boldsymbol{\rho}})_{\mathcal O}:=\int_{\mathcal O}
{\boldsymbol{\sigma}} :{\boldsymbol{\rho}},$$ where in the latter the colon denotes the Frobenius product of matrices and ${{\mathbb R^{3\times 3}_{\mathrm{sym}}}}$ is the space of symmetric $3\times 3$ matrices. The norm of both spaces will be denoted $\|\cdot\|_{\mathcal O}$.
- For a Lipschitz domain $\mathcal O$, the inner product in $L^2(\partial \mathcal O;{{\mathbb R}}^3)$ will be denoted $$\langle {\boldsymbol{\mu}},{\boldsymbol{\eta}}\rangle_{\partial\mathcal O}:=\int_{\partial\mathcal O}
{\boldsymbol{\mu}}\cdot{\boldsymbol{\eta}}$$ and the associated norm will be denoted $\|\cdot\|_{\partial\mathcal O}$.
- The Sobolev seminorm in $H^m(\mathcal O;{{\mathbb R}}^3)$ and $H^m(\mathcal O;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ will be denoted $|\cdot|_{m,\mathcal O}$.
- The symmetric gradient operator (linearized strain) is given by $${\boldsymbol{\varepsilon}}({\boldsymbol{u}}):=\tfrac12 (\mathrm D{\boldsymbol{u}}+(\mathrm D{\boldsymbol{u}})^\top),$$ and the divergence operator $\mathrm{div}$ will be applied to symmetric-matrix-valued functions by acting on their rows, outputting a column vector-valued function.
- When $\mathcal O$ is a Lipschitz domain, we will consider the space $$H(\mathcal O,\mathrm{div};{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}):=
\{ {\boldsymbol{\sigma}}\in L^2(\mathcal O;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\,:\,
\mathrm{div}\,{\boldsymbol{\sigma}}\in L^2(\mathcal O;{{\mathbb R}}^3)\},$$ and the normal traction operator $
\gamma_n:H(\mathcal O,\mathrm{div};{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\to H^{-1/2}(\partial\mathcal O;{{\mathbb R}}^3),
$ defined by Betti’s formula $$\langle \gamma_n {\boldsymbol{\sigma}},\gamma{\boldsymbol{v}}\rangle_{\partial\mathcal O}
:=({\boldsymbol{\sigma}},{\boldsymbol{\varepsilon}}({\boldsymbol{v}}))_{\mathcal O}
+(\mathrm{div}\,{\boldsymbol{\sigma}},{\boldsymbol{v}})_{\mathcal O}
\qquad \forall {\boldsymbol{v}}\in H^1(\mathcal O;{{\mathbb R}}^3).$$ Here $\gamma$ is the trace operator, $H^{-1/2}(\partial\mathcal O;{{\mathbb R}}^3)$ is the dual space of $H^{1/2}(\partial\mathcal O;{{\mathbb R}}^3)$, and the angled bracket denotes their duality product that extends the $L^2(\partial\mathcal O;{{\mathbb R}}^3)$ inner product.
For discretization we will consider a sequence of tetrahedral meshes $\mathcal T_h$ and the following notation:
- $K$ is a tetrahedron, of diameter $h_K$ and inradius at most $c\,h_K$ for a fixed shape-regularity constant $c>0$.
- $\mathcal F(K)$ is the set of faces of $K$.
- $\mathcal P_k(K;X)$ is the space of $X$-valued polynomial functions of degree up to $k\ge 0$, where $X\in \{{{\mathbb R}}^3,{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}\}$.
- $\mathcal R_k(\partial K;X):=\prod_{F\in \mathcal F(K)} \mathcal P_k(F;X)$ is the space of piecewise polynomial functions on the boundary of $K$, with $X$ as above.
- ${\mathrm}P_k:L^2(K;X)\rightarrow {\mathcal}P_k(K;X)$ is the orthogonal projection onto the image space, with $X$ as above.
- $\mathrm P_M:L^2(\partial K;{{\mathbb R}}^3)\to \mathcal R_k(\partial K;{{\mathbb R}}^3)$ is the orthogonal projection onto the image space. This operator will often be applied on the trace of a function in $H^1(K;{{\mathbb R}}^3)$ on $\partial K$.
- ${\boldsymbol{\tau}}\in \mathcal R_0(\partial K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ is a so-called stabilization function satisfying $$\label{eq:tauhm1}
C_1 h_K^{-1} \|{\boldsymbol{\mu}}\|_{\partial K}^2
\le \langle{\boldsymbol{\tau}}{\boldsymbol{\mu}},{\boldsymbol{\mu}}\rangle_{\partial K}
\le C_2 h_K^{-1} \|{\boldsymbol{\mu}}\|_{\partial K}^2
\qquad\forall {\boldsymbol{\mu}}\in L^2(\partial K;{{\mathbb R}}^3),$$ for fixed positive constants $C_1$ and $C_2$, independent of $h$ (i.e., of the particular mesh).
- The wiggled inequality sign $a\lesssim b$ hides a constant $a\le C\, b$ that is independent of $h$, while $a \approx b$ means $a\lesssim b\lesssim a$.
\[th:PROJ\] For $k\ge 1$, there exists a family of projections and associated boundary remainder operators $$\begin{aligned}
{6}
\Pi:H^1(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\times H^1(K;{{\mathbb R}}^3)
& \to \mathcal P_k(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}) \times \mathcal P_{k+1}(K;{{\mathbb R}}^3),\\
\mathrm R: H^1(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\times H^1(K;{{\mathbb R}}^3) & \to \mathcal R_k(\partial K;{{\mathbb R}}^3),\end{aligned}$$ depending on ${\boldsymbol{\tau}}\in \mathcal R_0(\partial K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$, where ${\boldsymbol{\tau}}$ is a constant positive definite matrix on each face of $K$, and if $({\boldsymbol{\sigma}}_K,{\boldsymbol{u}}_K)=\Pi({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}})$ and ${\boldsymbol{\delta}}_K=\mathrm R({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}})$, the following conditions hold:
\[eq:tauProp\] $$\begin{aligned}
{6}\label{eq:tauProp_a}
({\boldsymbol{u}}_K-{\boldsymbol{u}},{\boldsymbol{v}})_K &=0 &\quad &
\forall {\boldsymbol{v}}\in \mathcal P_{k-1}(K;{{\mathbb R}}^3),\\
\label{eq:tauProp_b}
-(\mathrm{div}\,({\boldsymbol{\sigma}}_K-{\boldsymbol{\sigma}}),{\boldsymbol{w}})_K
+\langle {\boldsymbol{\tau}} \mathrm P_M({\boldsymbol{u}}_K-{\boldsymbol{u}}),{\boldsymbol{w}}\rangle_{\partial K} &=
\langle {\boldsymbol{\delta}}_K,{\boldsymbol{w}}\rangle_{\partial K}
&& \forall {\boldsymbol{w}}\in \mathcal P_{k+1}(K;{{\mathbb R}}^3),\\
\label{eq:tauProp_c}
-\langle ({\boldsymbol{\sigma}}_K-{\boldsymbol{\sigma}}){\boldsymbol{n}}
-{\boldsymbol{\tau}}({\boldsymbol{u}}_K-{\boldsymbol{u}}),{\boldsymbol{\mu}}\rangle_{\partial K} &=
\langle {\boldsymbol{\delta}}_K,{\boldsymbol{\mu}}\rangle_{\partial K}
&& \forall {\boldsymbol{\mu}}\in \mathcal R_k(\partial K;{{\mathbb R}}^3).\end{aligned}$$
Moreover, if ${\boldsymbol{\tau}}$ satisfies and $({\boldsymbol{\sigma}},{\boldsymbol{u}})\in H^m(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\times H^{m+1}(K;{{\mathbb R}}^3)$ with $1\le m\le k+1$, then we have the estimates: $$\label{eq:1.3a}
\| {\boldsymbol{\sigma}}_K-{\boldsymbol{\sigma}}\|_K+ h_K^{-1}\|{\boldsymbol{u}}_K-{\boldsymbol{u}}\|_K
+h_K^{1/2}\|{\boldsymbol{\delta}}_K\|_{\partial K}
\le C h_K^m (|{\boldsymbol{\sigma}}|_{m,K}+|{\boldsymbol{u}}|_{m+1,K}).$$ The constant $C$ depends only on the polynomial degree $k$, the constants $C_1$ and $C_2$ in and the shape-regularity constant $c$. Finally, the ‘adjoint’ projection can be defined as $$\label{eq:1.4}
\Pi({\boldsymbol{\sigma}},{\boldsymbol{u}};-{\boldsymbol{\tau}}):=({\boldsymbol{\sigma}}_K,{\boldsymbol{u}}_K),
\qquad\mbox{where}\qquad
({\boldsymbol{\sigma}}_K,-{\boldsymbol{u}}_K):=\Pi({\boldsymbol{\sigma}},-{\boldsymbol{u}};{\boldsymbol{\tau}}).$$ This projection satisfies the properties with the same ${\boldsymbol{\delta}}_K$, if we substitute ${\boldsymbol{\tau}}$ by $-{\boldsymbol{\tau}}$.
Note that conditions are not enough to define the projection $\Pi$ but are exactly the ones that will be needed for the applications. As a final note, notice that combining and , we have $$\begin{aligned}
({\boldsymbol{\sigma}}_K-{\boldsymbol{\sigma}},{\boldsymbol{\varepsilon}}({\boldsymbol{w}}))_K=0\qquad
\forall {\boldsymbol{w}}\in{\mathcal}P_k(K;{\mathbb}R^3).\end{aligned}$$ Since $k\ge1$ and ${\mathcal}P_0(K;{\mathbb}R_{\mathrm}{sym}^{3\times3})\subset {\boldsymbol{\varepsilon}}({\mathcal}P_1(K;{\mathbb}R^3))$, we have $$\label{eq:rmproj}
({\boldsymbol{\sigma}}_K-{\boldsymbol{\sigma}},{\boldsymbol{\theta}})_K=0\qquad
\forall {\boldsymbol{\theta}}\in{\mathcal}P_0(K;{\mathbb}R_{\mathrm}{sym}^{3\times3}).$$
The projection in the reference element {#sec:proj_ref}
=======================================
Preparatory work on the reference element {#sec:Prep}
-----------------------------------------
In this section we will work on the reference tetrahedron $\widehat K:=\{ (x_1,x_2,x_3)\in {{\mathbb R}}^3\,:\, x_1,x_2,x_3,1-x_1-x_2-x_3>0\}$. The trace for vector-valued functions on the reference element will be denoted $\widehat\gamma$, and the normal traction operator on the reference element will be denoted $\widehat\gamma_n$. We will use shortened notation for the following spaces:
\[eq:2.1\] $$\begin{aligned}
{6}
& \widehat{{\boldsymbol{V}}}:=\mathcal P_{k}(\widehat K;{{\mathbb R}}^3),
&\qquad &
\widehat{{\boldsymbol{V}}}_-:=\mathcal P_{k-1}(\widehat K;{{\mathbb R}}^3),
\qquad
\widehat{{\boldsymbol{V}}}_+:=\mathcal P_{k+1}(\widehat K;{{\mathbb R}}^3),\\
& \widehat\Sigma :=\mathcal P_k(\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}),
&\qquad &
\widehat\Sigma_-:=\{ {\boldsymbol{\sigma}}\in\widehat\Sigma\,:\,
\mathrm{div}\,{\boldsymbol{\sigma}}=0, \quad \widehat\gamma_n{\boldsymbol{\sigma}}=0\}
\oplus {\boldsymbol{\varepsilon}}(\widehat{{\boldsymbol{V}}}).\end{aligned}$$
A new space $\widehat\Sigma_+$ will be defined once we have introduced some tools for it. Note that these constructions can be done directly on any tetrahedron $K$ [@CoFu:2018].
The first of these constructions is a lifting of the traction operator. It will act on the space $$L^2_{\mathcal M}(\partial\widehat K;{{\mathbb R}}^3)
:=\{{\boldsymbol{\mu}}\in L^2(\partial\widehat K;{{\mathbb R}}^3)\,:\,
\langle{\boldsymbol{\mu}},\widehat\gamma{\boldsymbol{m}}\rangle_{\partial\widehat K}=0
\quad\forall{\boldsymbol{m}}\in\mathcal M\},$$ where $\mathcal M$ is the six-dimensional space of infinitesimal rigid motions $$\begin{aligned}
{6}
\mathcal M:=& \{{\boldsymbol{m}}(\mathbf x):=\mathbf b+A\mathbf x\,:\,
A^\top =-A,\quad \mathbf b\in {{\mathbb R}}^3\}\\
=&\{{\boldsymbol{m}}\in H^1(\widehat K;{{\mathbb R}}^3)\,:\, {\boldsymbol{\varepsilon}}({\boldsymbol{m}})=0\}.\end{aligned}$$ Note that the trace space of $\mathcal M$ is also six-dimensional, i.e., if ${\boldsymbol{m}}\in \mathcal M$ vanishes on the boundary of $\widehat K$, then ${\boldsymbol{m}}={\boldsymbol{0}}$. We thus define the operator $\widehat\gamma_n^+:L^2_{\mathcal M}(\partial\widehat K;{{\mathbb R}}^3)\to H(\widehat K,\mathrm{div};{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ by
\[eq:gmp\] $$\begin{aligned}
{6}
\widehat\gamma_n^+{\boldsymbol{\mu}}:={\boldsymbol{\varepsilon}}({\boldsymbol{u}}),
\quad\mbox{where}\quad
& {\boldsymbol{u}}\in H^1(\widehat K;{{\mathbb R}}^3)/\mathcal M,\\
& ({\boldsymbol{\varepsilon}}({\boldsymbol{u}}),{\boldsymbol{\varepsilon}}({\boldsymbol{v}}))_{\widehat K}
=\langle{\boldsymbol{\mu}},\widehat\gamma{\boldsymbol{v}}\rangle_{\partial\widehat K}
\quad\forall {\boldsymbol{v}}\in H^1(\widehat K;{{\mathbb R}}^3),\end{aligned}$$
or equivalently
\[eq:gmp2\] $$\begin{aligned}
{6}
\widehat\gamma_n^+{\boldsymbol{\mu}}:={\boldsymbol{\sigma}},
\quad\mbox{where}\quad
&{\boldsymbol{\sigma}}\in {\boldsymbol{\varepsilon}}(H^1(\widehat K;{{\mathbb R}}^3)),\\
& ({\boldsymbol{\sigma}},{\boldsymbol{\varepsilon}}({\boldsymbol{v}}))_{\widehat K}
=\langle{\boldsymbol{\mu}},\widehat\gamma{\boldsymbol{v}}\rangle_{\partial\widehat K}
\quad\forall {\boldsymbol{v}}\in H^1(\widehat K;{{\mathbb R}}^3).\end{aligned}$$
The definition is correct since it involves the solution of a coercive variational problem on a quotient space, due to Korn’s Second Inequality. From it is clear that $$\label{eq:6}
\mathrm{div}\,\widehat\gamma_n^+{\boldsymbol{\mu}}=0,
\qquad
\widehat\gamma_n\widehat\gamma_n^+{\boldsymbol{\mu}}={\boldsymbol{\mu}}
\qquad \forall {\boldsymbol{\mu}}\in L^2_{\mathcal M}(\partial K;{{\mathbb R}}^3).$$ We next consider the spaces $$\begin{aligned}
{6}
&\widehat{{\boldsymbol{M}}}:=\mathcal R_k(\partial\widehat K;{{\mathbb R}}^3),
\qquad
\widehat\Sigma_S := \{{\boldsymbol{\sigma}}\in\widehat\Sigma\,:\, \mathrm{div}\,{\boldsymbol{\sigma}}=0\},\\
& \widehat\Theta:=\{{\boldsymbol{\mu}}\in \widehat{{\boldsymbol{M}}}\,:\,
\langle\widehat\gamma_n{\boldsymbol{\sigma}}
+\widehat\gamma{\boldsymbol{m}},{\boldsymbol{\mu}}\rangle_{\partial\widehat K}=0
\quad\forall ({\boldsymbol{\sigma}},{\boldsymbol{m}})\in \widehat\Sigma_S\times \mathcal M \},\end{aligned}$$ and $$\widehat\Sigma_{\mathrm{fill}}:=\widehat\gamma_n^+\widehat\Theta.$$
\[th:2.1\] The following properties hold:
- $\mathrm{div}\,{\boldsymbol{\sigma}}=0$ for all ${\boldsymbol{\sigma}}\in \widehat\Sigma_{\mathrm{fill}}$,
- $\widehat\gamma_n^+$ is an isomorphism between $\widehat\Theta$ and $\widehat\Sigma_{\mathrm{fill}}$, and its inverse is $\widehat\gamma_n$,
- $\widehat\Sigma_{\mathrm{fill}}\cap \widehat\Sigma=\{0\}$,
- $\widehat{{\boldsymbol{M}}}=\widehat\Theta\oplus\widehat\gamma_n\widehat\Sigma_S\oplus \gamma\mathcal M$ with orthogonal sum.
Properties (a) and (b) are easy consequences of . By (a), $\widehat\Sigma_{\mathrm{fill}}\cap \widehat\Sigma=\widehat\Sigma_{\mathrm{fill}}\cap \widehat\Sigma_S$ and, therefore, if ${\boldsymbol{\sigma}}\in \widehat\Sigma_{\mathrm{fill}}\cap \widehat\Sigma_S$, then $\widehat\gamma_n{\boldsymbol{\sigma}}\in \widehat\Theta\cap \widehat\gamma_n\widehat\Sigma_S=\{0\}$ (the latter two sets are orthogonal to each other by definition of $\widehat\Theta$), but then (b) proves that ${\boldsymbol{\sigma}}=0$, which proves (c). Finally, if ${\boldsymbol{\sigma}}\in \widehat\Sigma_S$ and ${\boldsymbol{m}}\in \mathcal M$, then $$\langle\widehat\gamma_n{\boldsymbol{\sigma}},\gamma{\boldsymbol{m}}\rangle_{\partial\widehat K}
=(\mathrm{div}\,{\boldsymbol{\sigma}},{\boldsymbol{m}})_{\widehat K}
+({\boldsymbol{\sigma}},{\boldsymbol{\varepsilon}}({\boldsymbol{m}}))_{\widehat K}=0,$$ which shows that the sum $\widehat\gamma_n\widehat\Sigma_S+\widehat\gamma\mathcal M$ is orthogonal. Since $\widehat\Theta$ is the orthogonal complement of the latter set, (d) is proved.
The Cockburn-Fu discrete pair for elasticity [@CoFu:2018] is given by the spaces $$\widehat\Sigma_+:=\widehat\Sigma\oplus\widehat\Sigma_{\mathrm{fill}}$$ (the sum is direct because of Theorem \[th:2.1\]) and $\widehat{{\boldsymbol{V}}}$. In their context of $M$-decompositions for elasticity, the following result is a key one, that we will need to work with our extended pair $\widehat\Sigma_+\times \widehat{{\boldsymbol{V}}}_+$.
\[th:2.2\] The following properties hold:
- ${\boldsymbol{\varepsilon}}(\widehat{{\boldsymbol{V}}})\subset \widehat\Sigma_-$,
- $\mathrm{div}\,\widehat\Sigma_+\subset \widehat{{\boldsymbol{V}}}_-$,
- $\widehat\gamma_n\widehat\Sigma_+
+ \widehat\gamma\widehat{{\boldsymbol{V}}}\subset \widehat{{\boldsymbol{M}}}$,
- $\mathrm{dim}\,\widehat{{\boldsymbol{M}}}
=\mathrm{dim}\,\widehat\Sigma_-^\perp+\mathrm{dim}\,\widehat{{\boldsymbol{V}}}_-^\perp,$ where the orthogonal complement $\widehat\Sigma_-^\perp$ is taken in $\widehat\Sigma_+$, while $\widehat{{\boldsymbol{V}}}_-^\perp$ is taken in $\widehat{{\boldsymbol{V}}}$,
- $\widehat{{\boldsymbol{M}}}=\widehat\gamma_n\widehat\Sigma_-^\perp
\oplus \gamma \widehat{{\boldsymbol{V}}}_-^\perp$, with orthogonal sum,
- The map $
\widehat\Sigma_-^\perp\times\widehat{{\boldsymbol{V}}}_-^\perp\ni ({\boldsymbol{\sigma}},{\boldsymbol{u}})
\longmapsto
\widehat\gamma_n{\boldsymbol{\sigma}}+\widehat\gamma{\boldsymbol{u}}\in \widehat{{\boldsymbol{M}}}
$ is an isomorphism.
Property (a) follows by definition and (b) is a simple consequence of Theorem \[th:2.1\](a). To show (c), note simply that $\widehat\gamma_n\widehat\Sigma_{\mathrm{fill}}=\widehat\Theta\subset \widehat{{\boldsymbol{M}}}$, while all other elements are polynomials of degree less than or equal to $k$.
To prove (d) to (f), it will be convenient to identify the set $$\label{eq:2.5}
\widehat\Sigma_S^\circ:=
\{ {\boldsymbol{\sigma}}\in \widehat\Sigma\,:\,
\mathrm{div}\,{\boldsymbol{\sigma}}=0,\quad\widehat\gamma_n{\boldsymbol{\sigma}}=0\}
=\ker \widehat\gamma_n|_{\widehat\Sigma_S},$$ which appeared in the definition of $\widehat\Sigma_-$ .
By Theorem \[th:2.1\](d) and , we have $$\begin{aligned}
{6}
\mathrm{dim}\,\widehat{{\boldsymbol{M}}}
&=\mathrm{dim}\,\widehat\Theta
+\mathrm{dim}\,\widehat\gamma_n\widehat\Sigma_S
+\mathrm{dim}\,\gamma\mathcal M
=\mathrm{dim}\,\widehat\Sigma_{\mathrm{fill}}
+\mathrm{dim}\,\widehat\Sigma_S-\mathrm{dim}\,\widehat\Sigma_S^\circ
+\mathrm{dim}\,\mathcal M.
\label{eq:2.6}\end{aligned}$$ Using the definitions of $\widehat\Sigma_\pm$ we have $$\begin{aligned}
{6}
\nonumber
\mathrm{dim}\,\widehat\Sigma_-^\perp
&=\mathrm{dim}\,\widehat\Sigma_+-\mathrm{dim}\,\widehat\Sigma_-
=\mathrm{dim}\,\widehat\Sigma+\mathrm{dim}\,\widehat\Sigma_{\mathrm{fill}}
-(\mathrm{dim}\,\widehat\Sigma_S^\circ
+\mathrm{dim}\,{\boldsymbol{\varepsilon}}(\widehat{{\boldsymbol{V}}}))\\
&=\mathrm{dim}\,\widehat{{\boldsymbol{V}}}_-+\mathrm{dim}\,\widehat\Sigma_S
+\mathrm{dim}\,\widehat\Sigma_{\mathrm{fill}}
-(\mathrm{dim}\,\widehat\Sigma_S^\circ+\mathrm{dim}\,\widehat{{\boldsymbol{V}}}
-\mathrm{dim}\,\mathcal M),
\label{eq:2.7}\end{aligned}$$ where in the last equality we have applied that $\mathrm{div}:\widehat\Sigma\to\widehat{{\boldsymbol{V}}}_-$ is onto (this is easy to prove), its kernel is $\widehat\Sigma_S$, and the kernel of ${\boldsymbol{\varepsilon}}:\widehat{{\boldsymbol{V}}}\to {\boldsymbol{\varepsilon}}(\widehat{{\boldsymbol{V}}})$ is $\mathcal M$. The equalities and prove (d).
Due to parts (a) and (b) of this theorem, we have $$\langle\widehat\gamma_n{\boldsymbol{\sigma}},\widehat\gamma{\boldsymbol{u}}\rangle_{\partial\widehat K}
=0 \qquad
\forall {\boldsymbol{\sigma}}\in \widehat\Sigma_-^\perp,
\quad{\boldsymbol{u}} \in \widehat{{\boldsymbol{V}}}_-^\perp,$$ which proves that the sum of $\widehat\gamma_n \widehat\Sigma_-^\perp$ and $\widehat\gamma \widehat{{\boldsymbol{V}}}_-^\perp$ is orthogonal. Since $\widehat\gamma:\widehat{{\boldsymbol{V}}}_-^\perp\to \widehat{{\boldsymbol{M}}}$ is injective (this result is known; see, for instance, [@sayas2013raviart] and [@DuSa:2019]), the property (e) will follow from having proved that $\widehat\gamma_n:\widehat\Sigma_-^\perp\to \widehat{{\boldsymbol{M}}}$ is injective.
We first prove the following technical result: if ${\boldsymbol{\sigma}}\in \widehat\Sigma_+$ satisfies $\mathrm{div}\,{\boldsymbol{\sigma}}=0$ and $\widehat\gamma_n{\boldsymbol{\sigma}}=0$, then ${\boldsymbol{\sigma}}\in \widehat\Sigma$, i.e., the component in $\widehat\Sigma_{\mathrm{fill}}$ vanishes. To do that, take ${\boldsymbol{\sigma}}={\boldsymbol{\sigma}}_1+{\boldsymbol{\sigma}}_2\in \widehat\Sigma+\widehat\Sigma_{\mathrm{fill}}$ and note that Theorem \[th:2.1\](a) shows that if $\mathrm{div}\,{\boldsymbol{\sigma}}=0$, then $\mathrm{div}\,{\boldsymbol{\sigma}}_1=0$. Since $$0=\widehat\gamma_n{\boldsymbol{\sigma}}=
\widehat\gamma_n{\boldsymbol{\sigma}}_1
+\widehat\gamma_n{\boldsymbol{\sigma}}_2\in \widehat\gamma_n\widehat\Sigma_S
\oplus\widehat\Theta,$$ by Theorem \[th:2.1\](d), it follows that $\widehat\gamma_n{\boldsymbol{\sigma}}_1=
\widehat\gamma_n{\boldsymbol{\sigma}}_2=0$, which proves that ${\boldsymbol{\sigma}}_2=0$ by Theorem \[th:2.1\](b).
The proof of injectivity of $\widehat\gamma_n|_{\widehat\Sigma_-^\perp}$ is then simple. If ${\boldsymbol{\sigma}}\in \widehat\Sigma_-^\perp$ satisfies $\widehat\gamma_n{\boldsymbol{\sigma}}=0$, then by part (a) $$(\mathrm{div}\,{\boldsymbol{\sigma}},{\boldsymbol{v}})_{\widehat K}
=-({\boldsymbol{\sigma}},{\boldsymbol{\varepsilon}}({\boldsymbol{v}}))_{\widehat K}=0
\qquad\forall {\boldsymbol{v}}\in \widehat{{\boldsymbol{V}}}$$ and, taking ${\boldsymbol{v}}=\mathrm{div}\,{\boldsymbol{\sigma}}$ (by part (b)), we prove that $\mathrm{div}\,{\boldsymbol{\sigma}}=0$. Therefore, ${\boldsymbol{\sigma}}\in \widehat\Sigma_S^\circ\subset\widehat\Sigma_-$ and hence ${\boldsymbol{\sigma}}=0$. This completes the proof of (e) and (f).
For the rest of this section we fix the stabilization parameter ${\boldsymbol{\tau}}\in \mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ such that $$\label{eq:3.0}
{\boldsymbol{\tau}}|_F \mbox{ is positive definite } \quad\forall F\in \mathcal F(\widehat K).$$
A projection on an extended space
---------------------------------
Given $({\boldsymbol{\sigma}},{\boldsymbol{u}})\in H^1(\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\times H^1(\widehat K,{{\mathbb R}}^3)$, we look for
\[eq:3.1\] $$\widehat\Pi_0({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}}):=
({\boldsymbol{\sigma}}_{\widehat K},{\boldsymbol{u}}_{\widehat K})
\in \widehat\Sigma_+\times \widehat{{\boldsymbol{V}}}_+$$ satisfying $$\begin{aligned}
{6}
\label{eq:3.1b}
({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}},{\boldsymbol{v}})_{\widehat K} = 0
& \qquad && \forall {\boldsymbol{v}}\in \widehat{{\boldsymbol{V}}}_-,\\
\label{eq:3.1c}
({\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}},{\boldsymbol{\theta}})_{\widehat K}=0
&&&\forall {\boldsymbol{\theta}}\in \widehat\Sigma_-,\\
\label{eq:3.1d}
\langle \widehat\gamma_n({\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}})
-{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma ({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}}),
{\boldsymbol{\mu}}\rangle_{\partial\widehat K} =0
&&&\forall {\boldsymbol{\mu}}\in \widehat{{\boldsymbol{M}}},\\
\label{eq:3.1e}
-(\mathrm{div}\,({\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}}),{\boldsymbol{w}})_{\widehat K}
+\langle{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma ({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}}),
\widehat\gamma {\boldsymbol{w}}\rangle_{\partial\widehat K} =0
&&&\forall {\boldsymbol{w}}\in \widehat{{\boldsymbol{V}}}^\perp,\end{aligned}$$
where $\widehat{{\boldsymbol{V}}}^\perp$ is the orthogonal complement of $\widehat{{\boldsymbol{V}}}$ in $\widehat{{\boldsymbol{V}}}_+$. Note that is equivalent to $$({\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}},{\boldsymbol{\varepsilon}}({\boldsymbol{w}}))_{\widehat K}
-\langle \widehat\gamma_n({\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}})
-{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma ({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}}),
\widehat\gamma{\boldsymbol{w}}\rangle_{\partial\widehat K} =0
\qquad \forall {\boldsymbol{w}}\in \widehat{{\boldsymbol{V}}}^\perp,$$ and, since ${\boldsymbol{\varepsilon}}(\widehat{{\boldsymbol{V}}})\subset \widehat\Sigma_-$ and $\widehat\gamma\widehat{{\boldsymbol{V}}}\subset\widehat{{\boldsymbol{M}}}$, equations and imply that we can substitute by the condition $$\label{eq:3.2}
-(\mathrm{div}\,({\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}}),{\boldsymbol{w}})_{\widehat K}
+\langle{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma ({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}}),
\widehat\gamma {\boldsymbol{w}}\rangle_{\partial\widehat K} =0\qquad
\forall {\boldsymbol{w}}\in \widehat{{\boldsymbol{V}}}_+,$$ which contains redundant restrictions already imposed in the other equations. Note also that the projection $\mathrm P_{\widehat M}$ can be eliminated in but not in or in , and that the bilinear form $$\langle{\boldsymbol{\tau}}\mathrm P_{\widehat M}{\boldsymbol{\mu}},{\boldsymbol{\eta}}\rangle_{\partial\widehat K}$$ is symmetric, bounded, and positive semi-definite in $L^2(\partial\widehat K;{{\mathbb R}}^3)$. We next prove that $\widehat\Pi_0$ is actually a well-defined projection onto $\widehat\Sigma_+\times\widehat{{\boldsymbol{V}}}_+$.
\[prop:3.1\] The process of defining the projection $\widehat\Pi_0$ in is equivalent to a square invertible linear system.
Note first that by Theorem \[th:2.2\](d) $$\begin{aligned}
{6}
\mathrm{dim}\,\widehat{{\boldsymbol{V}}}_-+\mathrm{dim}\,\widehat\Sigma_-
+\mathrm{dim}\,\widehat{{\boldsymbol{M}}}+\mathrm{dim}\,\widehat{{\boldsymbol{V}}}^\perp
=&\mathrm{dim}\,\widehat{{\boldsymbol{V}}}_-+\mathrm{dim}\,\widehat\Sigma_-
+\mathrm{dim}\,\widehat{{\boldsymbol{V}}}_-^\perp+\mathrm{dim}\,\widehat\Sigma_-^\perp \\
& +\mathrm{dim}\,\widehat{{\boldsymbol{V}}}_+-\mathrm{dim}\,\widehat{{\boldsymbol{V}}}\\
=&\mathrm{dim}\,\widehat\Sigma_++\mathrm{dim}\,\widehat{{\boldsymbol{V}}}_+,\end{aligned}$$ which proves that is equivalent to a linear system with as many equations as unknowns. We thus only need to prove uniqueness of solution.
A homogeneous solution of is a pair $({\boldsymbol{\sigma}}_{\widehat K},{\boldsymbol{u}}_{\widehat K})\in \widehat\Sigma_+\times \widehat{{\boldsymbol{V}}}_+$ satisfying (recall how is a consequence of )
$$\begin{aligned}
{6}
\label{eq:3.3a}
({\boldsymbol{u}}_{\widehat K},{\boldsymbol{v}})_{\widehat K} = 0 &
\qquad && \forall {\boldsymbol{v}}\in \widehat{{\boldsymbol{V}}}_-,\\
\label{eq:3.3b}
({\boldsymbol{\sigma}}_{\widehat K}\,,{\boldsymbol{\theta}})_{\widehat K}=0
&&&\forall {\boldsymbol{\theta}}\in \widehat\Sigma_-,\\
\label{eq:3.3c}
\langle \widehat\gamma_n{\boldsymbol{\sigma}}_{\widehat K}
-{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma {\boldsymbol{u}}_{\widehat K},
{\boldsymbol{\mu}}\rangle_{\partial\widehat K} =0
&&&\forall {\boldsymbol{\mu}}\in \widehat{{\boldsymbol{M}}},\\
\label{eq:3.3d}
-(\mathrm{div}\,{\boldsymbol{\sigma}}_{\widehat K},{\boldsymbol{w}})_{\widehat K}
+\langle{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma {\boldsymbol{u}}_{\widehat K},
\widehat\gamma {\boldsymbol{w}}\rangle_{\partial\widehat K} =0&
&&\forall {\boldsymbol{w}}\in \widehat{{\boldsymbol{V}}}_+.\end{aligned}$$
We now take ${\boldsymbol{w}}={\boldsymbol{u}}_{\widehat K}\in \widehat{{\boldsymbol{V}}}_-^\perp$ (with the complement in $\widehat{{\boldsymbol{V}}}_+$; also see ) in , recalling that $\mathrm{div}\,{\boldsymbol{\sigma}}_{\widehat K}\in \widehat{{\boldsymbol{V}}}_-$ (cf. Theorem \[th:2.2\](b)) and obtain $$\langle{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma{\boldsymbol{u}}_{\widehat K},
\widehat\gamma{\boldsymbol{u}}_{\widehat K}\rangle_{\partial\widehat K}
=\langle{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma{\boldsymbol{u}}_{\widehat K},
\mathrm P_{\widehat M}\widehat\gamma{\boldsymbol{u}}_{\widehat K}\rangle_{\partial\widehat K}
=0.$$ This argument uses that ${\boldsymbol{\tau}}$ is piecewise constant, so that multiplication by ${\boldsymbol{\tau}}$ is an endomorphism in $\widehat{{\boldsymbol{M}}}$. Since ${\boldsymbol{\tau}}$ is positive definite on each face, this proves that $\mathrm P_{\widehat M}\widehat\gamma{\boldsymbol{u}}_{\widehat K}={\boldsymbol{0}}$.
Using the above conclusion and taking ${\boldsymbol{\mu}}=\widehat\gamma_n {\boldsymbol{\sigma}}_{\widehat K}\in \widehat{{\boldsymbol{M}}}$ in (cf. Theorem \[th:2.2\](c)), it follows that $\widehat\gamma_n{\boldsymbol{\sigma}}_{\widehat K}={\boldsymbol{0}}$. Given the fact that ${\boldsymbol{\sigma}}_{\widehat K}\in \widehat\Sigma_-^\perp$ (by ) and Theorem \[th:2.2\](f), this proves that ${\boldsymbol{\sigma}}_{\widehat K}={\boldsymbol{0}}$.
Note finally that $\mathrm{div}\, {\boldsymbol{\varepsilon}}({\boldsymbol{u}}_{\widehat K})\in \widehat{{\boldsymbol{V}}}_-$ (Theorem \[th:2.2\](b)) and $\widehat\gamma_n{\boldsymbol{\varepsilon}}({\boldsymbol{u}}_{\widehat K})\in
\widehat{{\boldsymbol{M}}}$ (Theorem \[th:2.2\](c)), so that $$\begin{aligned}
{6}
({\boldsymbol{\varepsilon}}({\boldsymbol{u}}_{\widehat K}),{\boldsymbol{\varepsilon}}({\boldsymbol{u}}_{\widehat K}))_{\widehat K}
=& -(\mathrm{div}\,{\boldsymbol{\varepsilon}}({\boldsymbol{u}}_{\widehat K}),{\boldsymbol{u}}_{\widehat K})_{\widehat K}
+\langle \widehat\gamma_n {\boldsymbol{\varepsilon}}({\boldsymbol{u}}_{\widehat K}),
\widehat\gamma{\boldsymbol{u}}_{\widehat K}\rangle_{\partial\widehat K}\\
=&\langle \widehat\gamma_n {\boldsymbol{\varepsilon}}({\boldsymbol{u}}_{\widehat K}),
\mathrm P_{\widehat M}\widehat\gamma{\boldsymbol{u}}_{\widehat K}\rangle_{\partial\widehat K}=0,\end{aligned}$$ and therefore ${\boldsymbol{u}}_{\widehat K}\in \mathcal M$. This implies that $\widehat\gamma{\boldsymbol{u}}_{\widehat K}=\mathrm P_{\widehat M}\widehat\gamma{\boldsymbol{u}}_{\widehat K}={\boldsymbol{0}}$ and, therefore, ${\boldsymbol{u}}_{\widehat K}={\boldsymbol{0}},$ which completes the proof.
Looking at the proof of Proposition \[prop:3.1\], it is clear that we could have also defined the projection for any ${\boldsymbol{\tau}}\in \mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ such that ${\boldsymbol{\tau}}|_F$ is negative definite on each face.
\[prop:3.2\] For any $\tau_{\max}\ge \tau_{\min}>0$ and $k\ge 1$, there exists $C=C(\tau_{\max},\tau_{\min},k)$ such that if $({\boldsymbol{\sigma}}_{\widehat K},{\boldsymbol{u}}_{\widehat K})=\widehat\Pi_0({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}})$, then $$\label{eq:3.55}
\|{\boldsymbol{\sigma}}_{\widehat K}\|_{\widehat K}
+\|{\boldsymbol{u}}_{\widehat K}\|_{\widehat K}
\le C (\|{\boldsymbol{\sigma}}\|_{1,\widehat K}+\|{\boldsymbol{u}}\|_{1,\widehat K})
\quad \forall ({\boldsymbol{\sigma}},{\boldsymbol{u}})\in H^1(\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}\times {{\mathbb R}}^3),$$ whenever ${\boldsymbol{\tau}}\in \mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ satisfies $$\label{eq:3.5}
\tau_{\min}\le {\boldsymbol{\mu}}\cdot ({\boldsymbol{\tau}}|_F{\boldsymbol{\mu}}) \le \tau_{\max}
\qquad
\forall{\boldsymbol{\mu}}\in {{\mathbb R}}^3, |{\boldsymbol{\mu}}|=1, \quad \forall F\in \mathcal F(\widehat K).$$
Numbering the faces of $\widehat K$ and the entries of a symmetric matrix with indices from one to six, we can identify $\mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\equiv ({{\mathbb R^{3\times 3}_{\mathrm{sym}}}})^4\equiv {{\mathbb R}}^{24}$. We can thus make the identification $$\{{\boldsymbol{\tau}}\in \mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\,:\,
{\boldsymbol{\tau}}\mbox{ satisfies \eqref{eq:3.0}}\} \equiv \mathbb O \subset {{\mathbb R}}^{24},$$ where $\mathbb O$ is an open set. We can also identify $$\{{\boldsymbol{\tau}}\in \mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\,:\,
{\boldsymbol{\tau}}\mbox{ satisfies \eqref{eq:3.5}}\} \equiv \mathbb K
\subset {{\mathbb R}}^{24},$$ where $\mathbb K\subset\mathbb O$ is compact. To be more specific, we can write the identification ${\boldsymbol{\tau}}\equiv \mathbf x=(x_1,\ldots,x_{24})$ as follows: $$\begin{aligned}
{\boldsymbol{\tau}}=\sum_{i=1}^{24}x_i{\boldsymbol{\chi}}_i,\end{aligned}$$ where ${\boldsymbol{\chi}}_i\in{\mathcal}R_0({\partial}\widehat{K};{\mathbb}R_{\mathrm}{sym}^{3\times3})$ satisfies that ${\boldsymbol{\chi}}_{6(p-1)+q}$ is supported on the $p$-th face ($p=1\rightarrow 4$) and its $q$-th entry ($q=1\rightarrow 6$) in ${\mathbb}R_{\mathrm}{sym}^{3\times3}$ is equal to $1$ and the rest of the entries are all equal to $0$.
Consider now the continuous trial and test spaces $$\begin{aligned}
{6}
\mathcal U:= & H^1(\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\times H^1(\widehat K;{{\mathbb R}}^3),\\
\mathcal V:= &
L^2(\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}) \times L^2(\widehat K;{{\mathbb R}}^3)
\times L^2(\partial\widehat K;{{\mathbb R}}^3)\times H^1(\widehat K;{{\mathbb R}}^3), \end{aligned}$$ and their discrete counterparts $$\mathcal U_k := \widehat\Sigma_+\times \widehat{{\boldsymbol{V}}}_+,
\qquad
\mathcal V_k :=
\widehat{{\boldsymbol{V}}}_- \times \widehat\Sigma_-
\times \widehat{{\boldsymbol{M}}} \times \widehat{{\boldsymbol{V}}}^\perp.$$ We can define bilinear forms $$a_j:\mathcal U\times \mathcal V\to \mathbb R \qquad j=0,\ldots,24$$ as follows: $$\begin{aligned}
a_0(({\boldsymbol{\sigma}},{\boldsymbol{u}}),({\boldsymbol{v}},{\boldsymbol{\theta}},{\boldsymbol{\mu}},{\boldsymbol{w}}))&:=
({\boldsymbol{\sigma}},{\boldsymbol{\theta}})_{\widehat{K}}+({\boldsymbol{u}},{\boldsymbol{v}})_{\widehat{K}}
+{\langle \widehat{\gamma}_n{\boldsymbol{\sigma}},{\boldsymbol{\mu}}\rangle}_{{\partial}\widehat{K}}
-({\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}},{\boldsymbol{w}})_{\widehat{K}},\\
a_j(({\boldsymbol{\sigma}},{\boldsymbol{u}}),({\boldsymbol{v}},{\boldsymbol{\theta}},{\boldsymbol{\mu}},{\boldsymbol{w}}))&:=
-{\langle {\boldsymbol{\chi}}_j{\mathrm}P_{\widehat{M}}\widehat{\gamma}{\boldsymbol{u}},{\boldsymbol{\mu}}\rangle}_{{\partial}\widehat{K}}
+{\langle {\boldsymbol{\chi}}_j{\mathrm}P_{\widehat{M}}\widehat{\gamma}{\boldsymbol{u}},\widehat{\gamma}{\boldsymbol{w}}\rangle}_{{\partial}\widehat{K}}.\end{aligned}$$ We can see that $a_j$ are all bounded ${\boldsymbol{\tau}}$-independent (but dependent on $k$ through the operator $\mathrm P_{\widehat M}$) bilinear forms. Then equations can be rephrased in the following form: given $U\in \mathcal U$, find $\widehat\Pi_0 U\in \mathcal U_k$ such that $$\label{eq:3.7}
a_0(\widehat\Pi_0 U-U,V)+\sum_{j=1}^{24} x_j a_j(\widehat\Pi_0 U-U,V)=0
\qquad\forall V\in \mathcal V_k,$$ Equations are uniquely solvable for every $\mathbf x\in \mathbb O$ (this is a restatement of Proposition \[prop:3.1\]) and define a bounded linear operator $T_k(\mathbf x): \mathcal U \to \mathcal U_k.$ The function $T_k$, from $\mathbb O$ to the space of bounded linear operators from $\mathcal U$ to $\mathcal U_k$, is rational and therefore bounded on the compact set $\mathbb K$. We can thus bound $$\label{eq:3.8}
\| T_k(\mathbf x) U\|_\star \le C(k,\mathbb K,\star) \| U\|_{\mathcal U}
\qquad\forall U\in \mathcal U,$$ where $\|\cdot\|_\star$ is any norm we choose in $\mathcal U_k$. If we select the norm $$\|({\boldsymbol{\sigma}},{\boldsymbol{u}})\|_\star:=
\|{\boldsymbol{\sigma}}\|_{\widehat K}
+\|{\boldsymbol{u}}\|_{\widehat K},$$ in $\mathcal U_k$, then becomes .
A simple argument shows that algebraic condition is equivalent to asking that the spectrum of ${\boldsymbol{\tau}}|_F$ is contained in $[\tau_{\min},\tau_{\max}]$ for all $F\in \mathcal F(\widehat K)$ and also to the inequality $$\label{eq:3.55a}
\tau_{\min}\|{\boldsymbol{\mu}}\|_{\partial\widehat K}^2
\le \langle{\boldsymbol{\tau}}\,{\boldsymbol{\mu}},{\boldsymbol{\mu}}\rangle_{\partial\widehat K}
\le \tau_{\max}\|{\boldsymbol{\mu}}\|_{\partial\widehat K}^2
\qquad
\forall{\boldsymbol{\mu}}\in L^2(\partial\widehat K;{{\mathbb R}}^3).$$
The projection and the remainder
--------------------------------
Let now $\mathrm P_{\widehat\Sigma}:L^2(\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\to \widehat\Sigma$ be the orthogonal projection onto $\widehat\Sigma$. For $({\boldsymbol{\sigma}},{\boldsymbol{u}})\in H^1(\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\times H^1(\widehat K;{{\mathbb R}}^3)$ we define
\[eq:3.10\] $$\label{eq:3.10a}
\widehat\Pi({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}}):=({\boldsymbol{\sigma}}_{\widehat K}^c,{\boldsymbol{u}}_{\widehat K})
\in \widehat\Sigma \times \widehat{{\boldsymbol{V}}}_+,
\qquad
\widehat{\mathrm R}({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}}):=
\widehat\gamma_n({\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}}_{\widehat K}^c)
\in \widehat{{\boldsymbol{M}}},$$ where $$\label{eq:3.10b}
{\boldsymbol{\sigma}}_{\widehat K}^c:=\mathrm P_{\widehat\Sigma}{\boldsymbol{\sigma}}_{\widehat K},
\qquad
({\boldsymbol{\sigma}}_{\widehat K},{\boldsymbol{u}}_{\widehat K})=\widehat\Pi_0({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}}).$$
Since $\widehat\Sigma\times \widehat{{\boldsymbol{V}}}_+\subset \widehat\Sigma_+\times \widehat{{\boldsymbol{V}}}_+$, it follows that $\widehat\Pi_0({\boldsymbol{\theta}},{\boldsymbol{v}};{\boldsymbol{\tau}})=({\boldsymbol{\theta}},{\boldsymbol{v}})$ for all $({\boldsymbol{\theta}},{\boldsymbol{v}})\in \widehat\Sigma\times \widehat{{\boldsymbol{V}}}_+$ and therefore $$\label{eq:3.111}
\widehat\Pi({\boldsymbol{\theta}},{\boldsymbol{v}};{\boldsymbol{\tau}})=({\boldsymbol{\theta}},{\boldsymbol{v}}),
\qquad
\widehat{\mathrm R}({\boldsymbol{\theta}},{\boldsymbol{v}};{\boldsymbol{\tau}})=0
\qquad
\forall ({\boldsymbol{\theta}},{\boldsymbol{v}})\in \widehat\Sigma\times \widehat{{\boldsymbol{V}}}_+.$$ In particular, $\widehat\Pi$ is a projection onto $\widehat\Sigma\times \widehat{{\boldsymbol{V}}}_+$. Note that we do not give a set of equations to define $\widehat\Pi$, which is given as the application of $\widehat\Pi_0$ followed by an orthogonal projection applied to the first component of the output. However, the following equations relate the projection $\widehat\Pi$ and the associated remainder $\widehat{\mathrm R}$.
\[prop:3.3\] Let ${\boldsymbol{\tau}}\in \mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ satisfy , then $({\boldsymbol{\sigma}}_{\widehat K}^c,{\boldsymbol{u}}_{\widehat K}):=\widehat\Pi({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}})$ and ${\boldsymbol{\delta}}_{\widehat K}:=\widehat{\mathrm R}({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}})$ satisfy
$$\begin{aligned}
{6}
\label{eq:3.11a}
({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}},\mathrm{div}\,{\boldsymbol{\theta}})_{\widehat K} &= 0
\qquad && \forall {\boldsymbol{\theta}} \in \widehat\Sigma,\\
\label{eq:3.11c}
-(\mathrm{div}\,({\boldsymbol{\sigma}}_{\widehat K}^c-{\boldsymbol{\sigma}}),{\boldsymbol{w}})_{\widehat K}
+\langle{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma ({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}}),
\widehat\gamma {\boldsymbol{w}}\rangle_{\partial\widehat K} &=
\langle{\boldsymbol{\delta}}_{\widehat K},\widehat\gamma {\boldsymbol{w}}\rangle_{\partial\widehat K}
&\quad&\forall {\boldsymbol{w}}\in \widehat{{\boldsymbol{V}}}_+,\\
\label{eq:3.11b}
-\langle \widehat\gamma_n({\boldsymbol{\sigma}}_{\widehat K}^c-{\boldsymbol{\sigma}})
-{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma ({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}}),
{\boldsymbol{\mu}}\rangle_{\partial\widehat K} &=
\langle{\boldsymbol{\delta}}_{\widehat K}, {\boldsymbol{\mu}} \rangle_{\partial\widehat K}
&&\forall {\boldsymbol{\mu}}\in \widehat{{\boldsymbol{M}}}.\end{aligned}$$
Following the definition , we introduce the intermediate projection $({\boldsymbol{\sigma}}_{\widehat K},{\boldsymbol{u}}_{\widehat K})=\widehat\Pi_0({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}})$ so that ${\boldsymbol{\sigma}}_{\widehat K}^c=\mathrm P_{\widehat\Sigma}{\boldsymbol{\sigma}}_{\widehat K}$ and ${\boldsymbol{\delta}}_{\widehat K}=\widehat\gamma_n({\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}}_{\widehat K}^c)$.
Since $\mathrm{div}\,\widehat\Sigma\subset \widehat{{\boldsymbol{V}}}_-$ (this is a direct consequence of the definitions), is a consequence of . Note also that ${\boldsymbol{\varepsilon}}(\widehat{{\boldsymbol{V}}}_+)\subset \widehat\Sigma$ (again by definition) and therefore $$\label{eq:3.12}
({\boldsymbol{\sigma}}_{\widehat K}^c-{\boldsymbol{\sigma}}_{\widehat K},
{\boldsymbol{\varepsilon}}({\boldsymbol{w}}))_{\widehat K}
=0 \quad \forall {\boldsymbol{w}}\in \widehat{{\boldsymbol{V}}}_+.$$ The identity is a direct consequence of . Finally, by and $$\begin{aligned}
{6}
-(\mathrm{div}\,({\boldsymbol{\sigma}}_{\widehat K}^c-{\boldsymbol{\sigma}}),{\boldsymbol{w}})_{\widehat K}
+\langle{\boldsymbol{\tau}} \mathrm P_{\widehat M}\widehat\gamma ({\boldsymbol{u}}_{\widehat K}-{\boldsymbol{u}}),
\widehat\gamma {\boldsymbol{w}}\rangle_{\partial\widehat K}
=& -(\mathrm{div}\,({\boldsymbol{\sigma}}_{\widehat K}^c-{\boldsymbol{\sigma}}_{\widehat K}),
{\boldsymbol{w}})_{\widehat K}\\
=&-\langle\widehat\gamma_n ({\boldsymbol{\sigma}}_{\widehat K}^c-{\boldsymbol{\sigma}}_{\widehat K}),
\widehat\gamma {\boldsymbol{w}}\rangle_{\partial\widehat K},\end{aligned}$$ which proves .
\[prop:3.4\] For any $\tau_{\max}\ge \tau_{\min}>0$, and integers $k\ge 1$, $1\le m\le k+1$, there exists $C=C(\tau_{\max},\tau_{\min},k,m)$ such that if $$\begin{aligned}
{6}
({\boldsymbol{\sigma}}_{\widehat K}^c,{\boldsymbol{u}}_{\widehat K})&=
\widehat\Pi({\boldsymbol{\sigma}},{\boldsymbol{u}};\eta{\boldsymbol{\tau}}),
\qquad ({\boldsymbol{\sigma}},{\boldsymbol{u}})\in H^m(\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})
\times H^{m+1}(\widehat K;{{\mathbb R}}^3)
\qquad 0\neq \eta\in {{\mathbb R}},\\
{\boldsymbol{\delta}}_{\widehat K}&=\widehat{\mathrm R}({\boldsymbol{\sigma}},{\boldsymbol{u}};\eta{\boldsymbol{\tau}}),\end{aligned}$$ with ${\boldsymbol{\tau}}\in \mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ satisfying , then $$\|{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_{\widehat K}^c\|_{\widehat K}
+|\eta|\, \|{\boldsymbol{u}}-{\boldsymbol{u}}_{\widehat K}\|_{\widehat K}
+\|{\boldsymbol{\delta}}_{\widehat K}\|_{\partial\widehat K}
\le C \left( |{\boldsymbol{\sigma}}|_{m,\widehat K}+|\eta|\, |{\boldsymbol{u}}|_{m+1,\widehat K}\right).$$
Let $({\boldsymbol{\sigma}}_{\widehat K},{\boldsymbol{u}}_{\widehat K})=\widehat\Pi_0({\boldsymbol{\sigma}},{\boldsymbol{u}};\eta\,{\boldsymbol{\tau}})$. Note that $({\boldsymbol{\sigma}}_{\widehat K},\eta\,{\boldsymbol{u}}_{\widehat K})=\widehat\Pi_0({\boldsymbol{\sigma}},\eta\,{\boldsymbol{u}};{\boldsymbol{\tau}})$ (this is obvious from the equations that define $\widehat\Pi_0$, namely ) and therefore $$({\boldsymbol{\sigma}}_{\widehat K}^c,\eta\,{\boldsymbol{u}}_{\widehat K})
=\widehat\Pi({\boldsymbol{\sigma}},\eta\,{\boldsymbol{u}};{\boldsymbol{\tau}}),
\qquad
{\boldsymbol{\delta}}_{\widehat K}=\widehat{\mathrm R}({\boldsymbol{\sigma}},\eta\,{\boldsymbol{u}};{\boldsymbol{\tau}}).$$ By Proposition \[prop:3.2\], we have $$\|{\boldsymbol{\sigma}}_{\widehat K}\|_{\widehat K}
+|\eta|\, \|{\boldsymbol{u}}_{\widehat K}\|_{\widehat K}
\le C(\|{\boldsymbol{\sigma}}\|_{1,\widehat K}+|\eta|\,\|{\boldsymbol{u}}\|_{1,\widehat K}),$$ for some constant $C$ depending only on $\tau_{\min}$, $\tau_{\max}$ and $k$. Since $\widehat\gamma_n:\widehat\Sigma_+\to \widehat{{\boldsymbol{M}}}$ is bounded, there exists a constant $D=D(k)$ such that $$\|{\boldsymbol{\delta}}_{\widehat K}\|_{\partial\widehat K}
\le D\,\|{\boldsymbol{\sigma}}_{\widehat K}-{\boldsymbol{\sigma}}_{\widehat K}^c\|_{\widehat K}
\le 2 D\, \|{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_{\widehat K}\|_{\widehat K}
+D\|{\boldsymbol{\sigma}}-{\mathrm}P_{\widehat\Sigma}{\boldsymbol{\sigma}}\|_{\widehat K}.$$ Taking now ${\boldsymbol{\theta}}\in \widehat\Sigma$ and ${\boldsymbol{v}}\in \widehat{{\boldsymbol{V}}}_+$, and applying to the pair $({\boldsymbol{\theta}},\eta\,{\boldsymbol{v}})$, we have $$\|{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_{\widehat K}\|_{\widehat K}
+|\eta|\, \|{\boldsymbol{u}}-{\boldsymbol{u}}_{\widehat K}\|_{\widehat K}
\le (1+C) (\|{\boldsymbol{\sigma}}-{\boldsymbol{\theta}}\|_{1,\widehat K}
+|\eta|\, \|{\boldsymbol{u}}-{\boldsymbol{v}}\|_{1,\widehat K}),$$ and $$\|{\boldsymbol{\delta}}_{\widehat K}\|_{\partial\widehat K}
\le 2 D (1+C) (\|{\boldsymbol{\sigma}}-{\boldsymbol{\theta}}\|_{1,\widehat K}+|\eta|\, \|{\boldsymbol{u}}-{\boldsymbol{v}}\|_{1,\widehat K})+D\|{\boldsymbol{\sigma}}-{\mathrm}P_{\widehat{\Sigma}}{\boldsymbol{\sigma}}\|_{\widehat K}.$$ Note that ${\mathrm}P_{\widehat{\Sigma}}$ is the $L_2$ projection onto $\widehat\Sigma$. Therefore, there exists a constant $C'=C'(\tau_{\min},\tau_{\max},k)$ such that $$\|{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_{\widehat K}\|_{\widehat K}
+|\eta|\, \|{\boldsymbol{u}}-{\boldsymbol{u}}_{\widehat K}\|_{\widehat K}
+\|{\boldsymbol{\delta}}_{\widehat K}\|_{\partial\widehat K}
\le
C' \left(\inf_{{\boldsymbol{\theta}}\in \widehat\Sigma}\|{\boldsymbol{\sigma}}-{\boldsymbol{\theta}}\|_{1,\widehat K}
+|\eta|\, \inf_{{\boldsymbol{v}}\in \widehat{{\boldsymbol{V}}}_+}\|{\boldsymbol{u}}-{\boldsymbol{v}}\|_{1,\widehat K}\right).$$ Finally, notice that ${\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_{\widehat{K}}^c={\boldsymbol{\sigma}}-{\mathrm}P_{\widehat{\Sigma}}{\boldsymbol{\sigma}}+{\mathrm}P_{\widehat{\Sigma}}({\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_{\widehat{K}})$. The result follows now by a compactness argument (Bramble-Hilbert lemma).
The projection in the physical elements {#sec:proj_phy}
=======================================
Pull-backs and push-forwards
----------------------------
In this section we derive a systematic approach to changes of variables for vector- and matrix-valued functions from a general shape-regular tetrahedron to the reference element. The language mimics that of [@sayas2013raviart] (or [@DuSa:2019]). Let $K$ be a tetrahedron and $F:\widehat K\to K$ be an invertible affine map from the reference element to $K$. We will denote $\mathrm B:=\mathrm DF$ and $J:=\det\mathrm B$. We also consider the piecewise constant function $a:\partial\widehat K\to (0,\infty)$ such that for all integrable $\phi$, $$\int_{\partial K} \phi=\int_{\partial\widehat K} |a| \, \phi\circ F.$$ The trace and normal trace operators on $K$ will be denoted $\gamma$ and $\gamma_n$. Given $${\boldsymbol{u}},{\boldsymbol{u}}^*:K\to {{\mathbb R}}^3,
\quad
{\boldsymbol{\sigma}},{\boldsymbol{\sigma}}^*:K\to {{\mathbb R^{3\times 3}_{\mathrm{sym}}}},
\quad
{\boldsymbol{\mu}},{\boldsymbol{\mu}}^*:\partial K \to {{\mathbb R}}^3,$$ we define $$\begin{aligned}
{6}
\widehat{{\boldsymbol{u}}} & := \mathrm B^\top {\boldsymbol{u}}\circ F & & :\widehat K\to {{\mathbb R}}^3,
& \qquad &
\widecheck{{\boldsymbol{u}}}^* &:= |J|\mathrm B^{-1} {\boldsymbol{u}}^*\circ F & :\widehat K\to {{\mathbb R}}^3,\\
\widehat{{\boldsymbol{\sigma}}} &:= |J|\mathrm B^{-1}({\boldsymbol{\sigma}}\circ F)\mathrm B^{-\top}
& & : \widehat K \to {{\mathbb R^{3\times 3}_{\mathrm{sym}}}},
& &
\widecheck{{\boldsymbol{\sigma}}}^* &:= \mathrm B^\top ({\boldsymbol{\sigma}}^*\circ F)\mathrm B
& : \widehat K \to {{\mathbb R^{3\times 3}_{\mathrm{sym}}}},\\
\widehat{{\boldsymbol{\mu}}} &:= \mathrm B^\top {\boldsymbol{\mu}}\circ F &&:\partial\widehat K \to {{\mathbb R}}^3,
& &
\widecheck{{\boldsymbol{\mu}}}^* &:=|a| \mathrm B^{-1}{\boldsymbol{\mu}}^*\circ F
&:\partial\widehat K \to {{\mathbb R}}^3,\end{aligned}$$ where $\mathrm B^{-\top}=(\mathrm B^\top)^{-1}$. The following properties are easy to prove.
\[prop:4.1\] We have the following identities
$$\begin{aligned}
{6}
\label{eq:4.1a}
({\boldsymbol{u}},{\boldsymbol{u}}^*)_K
&=(\widehat{{\boldsymbol{u}}},\widecheck{{\boldsymbol{u}}}^*)_{\widehat K}
& \qquad & \forall {\boldsymbol{u}},{\boldsymbol{u}}^*\in L^2(K;{{\mathbb R}}^3),\\
\label{eq:4.1b}
({\boldsymbol{\sigma}},{\boldsymbol{\sigma}}^*)_K
&=(\widehat{{\boldsymbol{\sigma}}},\widecheck{{\boldsymbol{\sigma}}}^*)_{\widehat K}
& & \forall {\boldsymbol{\sigma}},{\boldsymbol{\sigma}}^*\in L^2(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}),\\
\label{eq:4.1c}
\langle{\boldsymbol{\mu}},{\boldsymbol{\mu}}^*\rangle_{\partial K}
&=\langle\widehat{{\boldsymbol{\mu}}},\widecheck{{\boldsymbol{\mu}}}^*\rangle_{\partial\widehat K}
& & \forall {\boldsymbol{\mu}},{\boldsymbol{\mu}}^*\in L^2(\partial K;{{\mathbb R}}^3).\end{aligned}$$
The next group of results about changes of variables involve the interaction of integrals with differential operators or trace operators.
\[prop:4.2\] We have the following indentities:
$$\begin{aligned}
{6}
\label{eq:4.2a}
({\boldsymbol{\sigma}},{\boldsymbol{\varepsilon}}({\boldsymbol{u}}))_K
&=(\widehat{{\boldsymbol{\sigma}}},\widehat{{\boldsymbol{\varepsilon}}}(\widehat{{\boldsymbol{u}}}))_{\widehat K}
& \qquad & \forall {\boldsymbol{\sigma}}\in L^2(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}),
&&\forall{\boldsymbol{u}}\in H^1(K;{{\mathbb R}}^3),\\
\label{eq:4.2b}
({\boldsymbol{u}},\mathrm{div}\,{\boldsymbol{\sigma}})_K
&=(\widehat{{\boldsymbol{u}}},\widehat{\mathrm{div}}\,\widehat{{\boldsymbol{\sigma}}})_{\widehat K}
& & \forall {\boldsymbol{\sigma}}\in H^1(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}),
&&\forall{\boldsymbol{u}}\in L^2(K;{{\mathbb R}}^3),\\
\label{eq:4.2co}
\langle \gamma_n{\boldsymbol{\sigma}},\gamma{\boldsymbol{u}}\rangle_{\partial K}
& =\langle\widehat\gamma_n \widehat{{\boldsymbol{\sigma}}},
\widehat\gamma\widehat{{\boldsymbol{u}}}\rangle_{\partial\widehat K}
& & \forall {\boldsymbol{\sigma}}\in H^1(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}),
&&\forall{\boldsymbol{u}}\in H^1(K;{{\mathbb R}}^3),\\
\label{eq:4.2c}
\langle \gamma_n{\boldsymbol{\sigma}},{\boldsymbol{\mu}}\rangle_{\partial K}
& =\langle\widehat\gamma_n \widehat{{\boldsymbol{\sigma}}},
\widehat{{\boldsymbol{\mu}}}\rangle_{\partial\widehat K}
& & \forall {\boldsymbol{\sigma}}\in H^1(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}),
&&\forall{\boldsymbol{\mu}}\in L^2(\partial K;{{\mathbb R}}^3),\\
\label{eq:4.2d}
\langle {\boldsymbol{\tau}}\,\gamma{\boldsymbol{u}},{\boldsymbol{\mu}}\rangle_{\partial K}
& =\langle\widecheck{{\boldsymbol{\tau}}}\,\widehat\gamma \widehat{{\boldsymbol{u}}},
\widehat{{\boldsymbol{\mu}}}\rangle_{\partial\widehat K}
&& \forall {\boldsymbol{u}}\in H^1(K;{{\mathbb R}}^3),
&&\forall{\boldsymbol{\mu}}\in L^2(\partial K;{{\mathbb R}}^3),\\
\label{eq:4.2e}
\langle {\boldsymbol{\tau}}\,\gamma{\boldsymbol{u}},\gamma {\boldsymbol{v}}\rangle_{\partial K}
& =\langle\widecheck{{\boldsymbol{\tau}}}\,\widehat\gamma \widehat{{\boldsymbol{u}}},
\widehat\gamma \widehat{{\boldsymbol{v}}}\rangle_{\partial\widehat K}
&& \forall {\boldsymbol{u}}, {\boldsymbol{v}}\in H^1(K;{{\mathbb R}}^3).\end{aligned}$$
Here ${\boldsymbol{\tau}}\in \mathcal R_0(\partial K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ and $$\label{eq:taucheck}
\widecheck{{\boldsymbol{\tau}}}:=|a| \mathrm B^{-1}({\boldsymbol{\tau}}\circ F)\mathrm B^{-\top}
\in \mathcal R_0(\partial\widehat K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}).$$
Using the definitions, it is easy to prove that $$\label{eq:4.44}
\widecheck{{\boldsymbol{\varepsilon}}({\boldsymbol{u}})}=\widehat{{\boldsymbol{\varepsilon}}}(\widehat{{\boldsymbol{u}}}),
\qquad
\widehat{\gamma{\boldsymbol{u}}}=\widehat\gamma\widehat{{\boldsymbol{u}}},
\qquad
\widecheck{{\boldsymbol{\tau}}{\boldsymbol{\mu}}}=\widecheck{{\boldsymbol{\tau}}}\,\widehat{{\boldsymbol{\mu}}}.$$ Then , , and are easy consequences of Proposition \[prop:4.1\]. Let now $\mathrm B=(b_{ij})_{i,j=1}^3$ and $\mathrm A=(a_{ij})_{i,j=1}^3=\mathrm B^{-1}$. Using implicit summation for repeated indices, we have for each $i$ $$\begin{aligned}
{6}
(\widehat{\mathrm{div}}\,\widehat{{\boldsymbol{\sigma}}})_i
&=\partial_{\widehat x_k}\widehat{{\boldsymbol{\sigma}}}_{ki}
=|J| a_{kl}a_{im}\partial_{\widehat x_k}(\sigma_{lm}\circ F)\\
&=|J| a_{kl} a_{im} b_{jk} (\partial_{x_j}\sigma_{lm}) \circ F
=|J| \delta_{jl} a_{im} (\partial_{x_j}\sigma_{lm}) \circ F\\
&=|J| a_{im} (\partial_{x_j}\sigma_{jm})\circ F
=|J| a_{im} (\mathrm{div}\,{\boldsymbol{\sigma}})_m\circ F\\
&=|J| (\mathrm B^{-1}\,\mathrm{div}\,{\boldsymbol{\sigma}})_i\circ F
=(\widecheck{\mathrm{div}\,{\boldsymbol{\sigma}}})_i\end{aligned}$$ and therefore $\widehat{\mathrm{div}}\,\widehat{{\boldsymbol{\sigma}}}=\widecheck{\mathrm{div}\,{\boldsymbol{\sigma}}}$. This and prove . Note next that, using identities we have already proved and Proposition \[prop:4.1\], we have
\[eq:5.2\] $$\begin{aligned}
{6}
\langle\widecheck{\gamma_n{\boldsymbol{\sigma}}},
\widehat\gamma\widehat{{\boldsymbol{u}}}\rangle_{\partial\widehat K}
=&\langle\widecheck{\gamma_n{\boldsymbol{\sigma}}},
\widehat{\gamma {\boldsymbol{u}}}\rangle_{\partial\widehat K}
= \langle\gamma_n{\boldsymbol{\sigma}},\gamma{\boldsymbol{u}}\rangle_{\partial K}
=(\mathrm{div}\,{\boldsymbol{\sigma}},{\boldsymbol{u}})_K
+({\boldsymbol{\sigma}},{\boldsymbol{\varepsilon}}({\boldsymbol{u}}))_K\\
=&(\widehat{\mathrm{div}}\,\widehat{{\boldsymbol{\sigma}}},\widehat{{\boldsymbol{u}}})_{\widehat K}
+(\widehat{{\boldsymbol{\sigma}}},\widehat{{\boldsymbol{\varepsilon}}}(\widehat{{\boldsymbol{u}}}))_{\widehat K}
=\langle\widehat\gamma_n\widehat{{\boldsymbol{\sigma}}},
\widehat\gamma\widehat{{\boldsymbol{u}}}\rangle_{\partial \widehat K},\end{aligned}$$
and is thus proved. Since $H^{1/2}(\partial K;{{\mathbb R}}^3)$ is dense in $L^2(\partial K;{{\mathbb R}}^3)$, follows from .
The final result of this section contains all scaling inequalities. Shape-regularity can be rephrased as the asymptotic equivalences (recall that we are in three dimensions) $$\label{eq:4.3}
\|\mathrm B\|\approx h_K,
\qquad
\|\mathrm B^{-1}\|\approx h_K^{-1},
\qquad
|J| \approx h_K^3,
\qquad
|a| \approx h_K^2.$$ Therefore $$\label{eq:4.4}
\|\phi\|_K \approx h_K^{3/2} \|\phi\circ F\|_{\widehat K},
\qquad
\|\phi\|_{\partial K} \approx h_K \|\phi\circ F\|_{\partial\widehat K}.$$
For the hat transformations, we have the following scaling rules, which can be easily proved using , , and the chain rule.
\[prop:4.3\] With hidden constants depending on shape-regularity and on $m\ge 0$, we have
$$\begin{aligned}
{6}
|\widehat{{\boldsymbol{u}}}|_{m,\widehat K}
& \approx h_K^{m-1/2}|{\boldsymbol{u}}|_{m,K}
& \qquad & \forall {\boldsymbol{u}}\in H^m(K;{{\mathbb R}}^3),\\
|\widehat{{\boldsymbol{\sigma}}}|_{m,\widehat K}
& \approx h_K^{m-1/2} |{\boldsymbol{\sigma}}|_{m,K}
& & \forall {\boldsymbol{\sigma}}\in H^m(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}),\\
\|\widecheck{{\boldsymbol{\mu}}}\|_{\partial\widehat K}\approx
\|\widehat{{\boldsymbol{\mu}}}\|_{\partial\widehat K}
& \approx \|{\boldsymbol{\mu}}\|_{\partial K}
& & \forall {\boldsymbol{\mu}}\in L^2(\partial K;{{\mathbb R}}^3). \end{aligned}$$
The projection and the remainder on $K$
---------------------------------------
Consider the spaces $$\begin{aligned}
{6}
\Sigma(K):=&
\{ {\boldsymbol{\sigma}}\in L^2(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\,:\,\widehat{{\boldsymbol{\sigma}}}\in \widehat\Sigma\}
=\mathcal P_k(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}}),\\
{\boldsymbol{V}}_+(K) :=& \{{\boldsymbol{u}}\in L^2(K;{{\mathbb R}}^3)\,:\,\widehat{{\boldsymbol{u}}}\in \widehat{{\boldsymbol{V}}}_+\}
=\mathcal P_{k+1}(K;{{\mathbb R}}^3),\\
{\boldsymbol{M}}(\partial K):=&
\{{\boldsymbol{\mu}}\in L^2(\partial K;{{\mathbb R}}^3)\,:\,\widehat{{\boldsymbol{\mu}}}\in \widehat{{\boldsymbol{M}}}\}\\
=&\{{\boldsymbol{\mu}}\in L^2(\partial K;{{\mathbb R}}^3)\,:\,\widecheck{{\boldsymbol{\mu}}}\in \widehat{{\boldsymbol{M}}}\}
=\mathcal R_k(\partial K;{{\mathbb R}}^3).\end{aligned}$$ The projection and the remainder are defined by a pull-back process: given $({\boldsymbol{\sigma}},{\boldsymbol{u}})\in H^1(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})\times H^1(K;{{\mathbb R}}^3)$, we define
\[eq:5.3\] $$\Pi({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}}):=({\boldsymbol{\sigma}}_K,{\boldsymbol{u}}_K)
\in \Sigma(K)\times {\boldsymbol{V}}_+(K),
\qquad
\mathrm R({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}}):={\boldsymbol{\delta}}_K\in {\boldsymbol{M}}(\partial K),$$ by the relations $$\begin{aligned}
{6}
\label{eq:4.9bbb}
(\widehat{{\boldsymbol{\sigma}}_K},\widehat{{\boldsymbol{u}}_K})
&=\widehat\Pi(\widehat{{\boldsymbol{\sigma}}},\widehat{{\boldsymbol{u}}};\widecheck{{\boldsymbol{\tau}}}),
\qquad \widecheck{{\boldsymbol{\tau}}}:=|a| \mathrm B^{-1}({\boldsymbol{\tau}}\circ F)\mathrm B^{-\top},\\
\widecheck{{\boldsymbol{\delta}}_K}
&=\widehat{\mathrm R}(\widehat{{\boldsymbol{\sigma}}},\widehat{{\boldsymbol{u}}};\widecheck{{\boldsymbol{\tau}}}).\end{aligned}$$
We now prove Theorem \[th:PROJ\]. We start by proving a technical lemma, continue showing that equations hold (we present this as Proposition \[prop:5.2\]), and finish by proving the estimates .
\[lemma:5.1\] If $\mathrm P_M$ is the $L^2(\partial K;{{\mathbb R}}^3)$ orthogonal projector onto ${\boldsymbol{M}}(\partial K)$, then $$\label{eq:5.1}
\widehat{\mathrm P_M{\boldsymbol{\mu}}}=\mathrm P_{\widehat M}\widehat{{\boldsymbol{\mu}}}
\qquad
\forall{\boldsymbol{\mu}} \in L^2(\partial K;{{\mathbb R}}^3).$$ Therefore, if ${\boldsymbol{\tau}}\in \mathcal R_0(\partial K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$, we have $$\begin{aligned}
{6}
\langle {\boldsymbol{\tau}}\,\mathrm P_M\gamma{\boldsymbol{u}},{\boldsymbol{\mu}}\rangle_{\partial K}
& =\langle\widecheck{{\boldsymbol{\tau}}}\,
\mathrm P_{\widehat M}\widehat\gamma \widehat{{\boldsymbol{u}}},
\widehat{{\boldsymbol{\mu}}}\rangle_{\partial\widehat K}
&\qquad & \forall {\boldsymbol{u}}\in H^1(K;{{\mathbb R}}^3),
&&\forall{\boldsymbol{\mu}}\in L^2(\partial K;{{\mathbb R}}^3),\\
\langle {\boldsymbol{\tau}}\,\mathrm P_M\gamma{\boldsymbol{u}},\gamma {\boldsymbol{v}}\rangle_{\partial K}
& =\langle\widecheck{{\boldsymbol{\tau}}}\,
\mathrm P_{\widehat M}\widehat\gamma \widehat{{\boldsymbol{u}}},
\widehat\gamma \widehat{{\boldsymbol{v}}}\rangle_{\partial\widehat K}
&& \forall {\boldsymbol{u}}, {\boldsymbol{v}}\in H^1(K;{{\mathbb R}}^3),\end{aligned}$$ with $\widecheck{{\boldsymbol{\tau}}}$ defined in .
It follows from the definitions that $\widehat{\mathrm P_M{\boldsymbol{\mu}}}\in \widehat{{\boldsymbol{M}}}$ and that ${\boldsymbol{\mu}}^*\in {\boldsymbol{M}}(K)$ if and only if $\widecheck{{\boldsymbol{\mu}}}^*\in \widehat{{\boldsymbol{M}}}$. Therefore, $$\langle\widehat{\mathrm P_M{\boldsymbol{\mu}}},\widecheck{{\boldsymbol{\mu}}}^*\rangle_{\partial\widehat K}
=\langle\mathrm P_M{\boldsymbol{\mu}},{\boldsymbol{\mu}}^*\rangle_{\partial K}
=\langle{\boldsymbol{\mu}},{\boldsymbol{\mu}}^*\rangle_{\partial K}
=\langle \widehat{{\boldsymbol{\mu}}},\widecheck{{\boldsymbol{\mu}}}^*\rangle_{\partial\widehat K}
=\langle \mathrm P_{\widehat M}\widehat{{\boldsymbol{\mu}}},
\widecheck{{\boldsymbol{\mu}}}^* \rangle_{\partial\widehat K}
\qquad\forall{\boldsymbol{\mu}}^*\in \widehat{{\boldsymbol{M}}},$$ and is proved. The next two identities in the statement follow from and Proposition \[prop:4.1\] and .
\[prop:5.2\] Assume that ${\boldsymbol{\tau}}\in \mathcal R_0(\partial K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ satisfies $$\langle{\boldsymbol{\tau}}{\boldsymbol{\mu}},{\boldsymbol{\mu}}\rangle_{\partial K}>0 \qquad
\forall {\boldsymbol{\mu}} \in L^2(\partial K;{{\mathbb R}}^3), \quad{\boldsymbol{\mu}}\neq{\boldsymbol{0}}.$$ If $({\boldsymbol{\sigma}}_K,{\boldsymbol{u}}_K)=\Pi({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}})$ and ${\boldsymbol{\delta}}_K=\mathrm R({\boldsymbol{\sigma}},{\boldsymbol{u}};{\boldsymbol{\tau}})$, then the following equations hold:
$$\begin{aligned}
{6}
\label{eq:5.33a}
({\boldsymbol{u}}_K-{\boldsymbol{u}},\mathrm{div}\,{\boldsymbol{\theta}})_{K} &= 0
\qquad && \forall {\boldsymbol{\theta}} \in \Sigma(K),\\
\label{eq:5.33b}
-(\mathrm{div}\,({\boldsymbol{\sigma}}_K-{\boldsymbol{\sigma}}),{\boldsymbol{w}})_K
+\langle{\boldsymbol{\tau}} \mathrm P_M\gamma ({\boldsymbol{u}}_K-{\boldsymbol{u}}),
\gamma {\boldsymbol{w}}\rangle_{\partial K} &=
\langle{\boldsymbol{\delta}}_K,\gamma {\boldsymbol{w}}\rangle_{\partial K}
&\,\,&\forall {\boldsymbol{w}}\in {\boldsymbol{V}}_+(K),\\
\label{eq:5.33c}
-\langle \gamma_n({\boldsymbol{\sigma}}_K-{\boldsymbol{\sigma}})
-{\boldsymbol{\tau}} \mathrm P_M \gamma ({\boldsymbol{u}}_K-{\boldsymbol{u}}),
{\boldsymbol{\mu}}\rangle_{\partial K} &=
\langle{\boldsymbol{\delta}}_K, {\boldsymbol{\mu}} \rangle_{\partial K}
&&\forall {\boldsymbol{\mu}}\in {\boldsymbol{M}}(\partial K).\end{aligned}$$
The equality follows from changing variables to the reference element, applying and . Similarly, follows from using , Lemma \[lemma:5.1\], and . Finally, follows from applying , Lemma \[lemma:5.1\], and .
Assume now that ${\boldsymbol{\tau}}$ is of order $h_K^{-1}$, as expressed in . By and Proposition \[prop:4.3\] we can write (take ${\boldsymbol{\eta}}=\widecheck{{\boldsymbol{\mu}}}$ for ${\boldsymbol{\mu}}\in L^2(\partial K;{{\mathbb R}}^3)$) $$\label{eq:5.5}
\alpha_1 C_1 h_K^{-1}\|{\boldsymbol{\eta}}\|_{\partial\widehat K}^2
\le \langle \widecheck{{\boldsymbol{\tau}}}\, {\boldsymbol{\eta}},
{\boldsymbol{\eta}}\rangle_{\partial\widehat K}
\le \alpha_2 C_2 h_K^{-1}\|{\boldsymbol{\eta}}\|_{\partial\widehat K}^2
\qquad
\forall {\boldsymbol{\eta}}\in L^2(\partial\widehat K;{{\mathbb R}}^3),$$ where $\alpha_1$ and $\alpha_2$ are constants related to shape-regularity of $K$. We are now ready to apply Proposition \[prop:3.4\] with $\tau_{\min}=\alpha_1 C_1$, $\tau_{\max}=\alpha_2C_2$ and $\eta=h_K^{-1}$ (compare with ). Using the definition of the projection , Proposition \[prop:4.3\] for the scaling properties, and Proposition \[prop:3.4\] for the estimates, we have $$\begin{aligned}
{6}
\|{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_K\|_K+h_K^{-1}\|{\boldsymbol{u}}-{\boldsymbol{u}}_K\|_K
+ h_K^{1/2}\|{\boldsymbol{\delta}}_K\|_K
\approx & h_K^{1/2}
(\|\widehat{{\boldsymbol{\sigma}}}-\widehat{{\boldsymbol{\sigma}}_K}\|_{\widehat K}
+h_K^{-1} \|\widehat{{\boldsymbol{u}}}-\widehat{{\boldsymbol{u}}_K}\|_{\widehat K}
+\|\widecheck{{\boldsymbol{\delta}}_K}\|_{\partial\widehat K})\\
\lesssim & h_K^{1/2} (|\widehat{{\boldsymbol{\sigma}}}|_{m,\widehat K}
+h_K^{-1} |\widehat{{\boldsymbol{u}}}|_{m+1,\widehat K})\\
\approx & h_K^m (|{\boldsymbol{\sigma}}|_{m,K}+ |{\boldsymbol{u}}|_{m+1,K}),\end{aligned}$$ and is thus proved.
Steady-state elasticity {#sec:steady_state}
=======================
Method and convergence estimates {#sec:st_main}
--------------------------------
From now on, we shift our attention from the construction of the projection to its applications. We begin by introducing more notation for the rest of the paper. Let $\Omega$ be a Lipschitz polyhedral domain in ${\mathbb}R^3$. We denote the compliance tensor by ${\mathcal}A\in L^\infty(\Omega;{\mathcal}B({\mathbb}R_\mathrm{sym}^{3\times3}))$, where ${\mathcal}B({{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ is the space of linear maps from ${{\mathbb R^{3\times 3}_{\mathrm{sym}}}}$ to itself. We assume $\rho\ge\rho_0$ for some positive constant $\rho_0$, and ${\mathcal}A$ is uniformly symmetric and positive almost everywhere on $\Omega$, i.e., there exists $C_0>0$ such that for almost all $\mathbf x\in \Omega$, $$\begin{aligned}
\left.\begin{array}{l}
({\mathcal}A(\mathbf x){\boldsymbol{\xi}}):{\boldsymbol{\chi}} = {\boldsymbol{\xi}}:({\mathcal}A(\mathbf x){\boldsymbol{\chi}})\\[5pt]
({\mathcal}A(\mathbf x){\boldsymbol{\xi}}):{\boldsymbol{\xi}} \ge C_0 {\boldsymbol{\xi}}:{\boldsymbol{\xi}}
\end{array}
\right\}
\quad\forall{\boldsymbol{\xi}},{\boldsymbol{\chi}}\in{\mathbb}R_{\mathrm}{sym}^{3\times3}.\end{aligned}$$
Let ${\mathcal}T_h$ be a family of conforming tetrahedral partitions of $\Omega$, which we assume to be shape-regular. Namely, there exists a fixed constant $c_0>0$, such that $\frac{h_K}{\rho_K}\le c_0$ for all $K\in{\mathcal}T_h$, where $\rho_K$ denotes the inradius of $K$. We denote $h:=\max_{K\in{\mathcal}T_h}h_K$ as the mesh size and set the following discrete spaces: $$\begin{aligned}
{\boldsymbol{V}}_h:=\prod_{K\in{\mathcal}T_h}\mathcal P_{k}(K;\mathbb R_{\mathrm}{sym}^{3\times3}),\quad
{\boldsymbol{W}}_h:=\prod_{K\in{\mathcal}T_h}\mathcal P_{k+1}(K;\mathbb R^3),\quad
{\boldsymbol{M}}_h:=\prod_{K\in{\mathcal}T_h}\mathcal R_k({\partial}K;\mathbb R^3).\end{aligned}$$ The related discrete inner products are denoted as $$\begin{aligned}
({\boldsymbol{\sigma}},{\boldsymbol{\tau}})_{{\mathcal}T_h}:=\sum_{K\in{\mathcal}T_h}({\boldsymbol{\sigma}},{\boldsymbol{\tau}})_K,\quad
({\boldsymbol{u}},{\boldsymbol{v}})_{{\mathcal}T_h}:=\sum_{K\in{\mathcal}T_h}({\boldsymbol{u}},{\boldsymbol{v}})_K,\quad
{\langle {\boldsymbol{\mu}},{\boldsymbol{\lambda}}\rangle}_{{\partial}{\mathcal}T_h}
:=\sum_{K\in{\mathcal}T_h}{\langle {\boldsymbol{\mu}},{\boldsymbol{\lambda}}\rangle}_{{\partial}K}.\end{aligned}$$ Finally, we use the following notation for the discrete and the weighted discrete norms: $$\begin{aligned}
{5}
&\|\cdot\|_{{\mathcal}T_h}:=(\cdot,\cdot)_{{\mathcal}T_h}^{1/2},
&\qquad &\|\cdot\|_{{\partial}{\mathcal}T_h}:={\langle \cdot,\cdot\rangle}_{{\partial}{\mathcal}T_h}^{1/2},\\
&\|\cdot\|_{*_1}:=(*_1\,\cdot,\cdot)_{{\mathcal}T_h}^{1/2},
&&\|\cdot\|_{*_2}:={\langle *_2\,\cdot,\cdot\rangle}_{{\partial}{\mathcal}T_h}^{1/2},\end{aligned}$$ where $*_1={\mathcal}A$ or $\rho$, and $*_2={\boldsymbol{\tau}}$ or ${\boldsymbol{\tau}}^{-1}$. Here ${\boldsymbol{\tau}}\in \prod_{K\in{\mathcal}T_h}{\mathcal}R_0({\partial}K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$ is the stabilization function which we assume to satisfy on every element $K$.
In this section, we give a projection-based error analysis to the HDG+ method introduced in [@QiShSh:2018]. We will show that the analysis can be simplified by using Theorem \[th:PROJ\]. To begin with, we review the steady-state linear elasticity equations:
\[eq:st\_pde\] $$\begin{aligned}
{5}
{\mathcal}A{\boldsymbol{\sigma}} - {\boldsymbol{\varepsilon}}({\boldsymbol{u}}) &= {\boldsymbol{0}} &\qquad&{\mathrm}{in}\ \Omega,\\
-{\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}} &= {\boldsymbol{f}} && {\mathrm}{in}\ \Omega,\\
\gamma{\boldsymbol{u}}&= {\boldsymbol{g}} &&{\mathrm}{on}\ \Gamma:={\partial}\Omega,\end{aligned}$$
where ${\boldsymbol{f}}\in L^2(\Omega;{\mathbb}R^3)$ and ${\boldsymbol{g}}\in H^{1/2}(\Gamma;{\mathbb}R^3)$. The HDG+ method for is: find $({\boldsymbol{\sigma}}_h,{\boldsymbol{u}}_h,\widehat{{\boldsymbol{u}}}_h)\in {\boldsymbol{V}}_h\times{\boldsymbol{W}}_h\times{\boldsymbol{M}}_h$ such that
\[eq:st\_HDG\] $$\begin{aligned}
{5}
({\mathcal}A{\boldsymbol{\sigma}}_h,{\boldsymbol{\theta}})_{{\mathcal}T_h}
+ ({\boldsymbol{u}}_h,{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}
-{\langle \widehat{{\boldsymbol{u}}}_h,{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}&
=0&\quad& \forall {\boldsymbol{\theta}}\in{\boldsymbol{V}}_h,\\
-({\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}}_h,{\boldsymbol{w}})_{{\mathcal}T_h}
+{\langle {\boldsymbol{\tau}}{\mathrm}P_M({\boldsymbol{u}}_h-\widehat{{\boldsymbol{u}}}_h),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
&=({\boldsymbol{f}},{\boldsymbol{w}})_{{\mathcal}T_h}&\quad & \forall{\boldsymbol{w}}\in{\boldsymbol{W}}_h,\\
{\langle
{\boldsymbol{\sigma}}_h{\boldsymbol{n}}
-{\boldsymbol{\tau}}{\mathrm}P_M({\boldsymbol{u}}_h-\widehat{{\boldsymbol{u}}}_h),{\boldsymbol{\mu}}
\rangle}_{\partial{\mathcal}T_h\backslash\Gamma}
&=0&\quad & \forall{\boldsymbol{\mu}}\in{\boldsymbol{M}}_h,\\
{\langle \widehat{{\boldsymbol{u}}}_h,{\boldsymbol{\mu}}\rangle}_\Gamma &= {\langle {\boldsymbol{g}},{\boldsymbol{\mu}}\rangle}_{\Gamma}
&\quad & \forall{\boldsymbol{\mu}}\in{\boldsymbol{M}}_h.\end{aligned}$$
Since we will use a duality argument to estimate the convergence of ${{\boldsymbol{u}}}_h$, here we write down the adjoint equations for :
\[eq:st\_adpde\] $$\begin{aligned}
{5}
{\mathcal}A{\boldsymbol{\Psi}} + {\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}}) &= 0 &\qquad& {\mathrm}{in}\ \Omega,\\
{\mathrm{div}\hspace{1pt}}{\boldsymbol{\Psi}} &= {\boldsymbol{\Theta}} && {\mathrm}{in}\ \Omega,\\
\gamma{\boldsymbol{\Phi}}&=0 && {\mathrm}{on}\ \Gamma,\end{aligned}$$
with input data ${\boldsymbol{\Theta}}\in L^2(\Omega;{\mathbb}R^3)$, and we assume the additional elliptic regularity estimate $$\begin{aligned}
\label{eq:ell_reg}
\|{\boldsymbol{\Psi}}\|_{1,\Omega}+\|{\boldsymbol{\Phi}}\|_{2,\Omega}\le C_{\mathrm}{reg}\|{\boldsymbol{\Theta}}\|_\Omega,\end{aligned}$$ where $C_{\mathrm}{reg}$ is a constant depending only on ${\mathcal}A$ and $\Omega$. The following convergence theorem will be proved in Subsection \[sec:st\_prf\].
\[th:st\] For $k\ge1$ and $1\le m\le k+1$, we have $$\begin{aligned}
\|{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_h\|_{{\mathcal}A}\le C_1h^m(\|{\boldsymbol{\sigma}}\|_{m,\Omega}+\|{\boldsymbol{u}}\|_{m+1,\Omega}).\end{aligned}$$ If holds, then we also have $$\begin{aligned}
\|{\boldsymbol{u}}-{\boldsymbol{u}}_h\|_\Omega\le C_2 h^{m+1}(\|{\boldsymbol{\sigma}}\|_{m,\Omega}+\|{\mathbf}u\|_{m+1,\Omega}).\end{aligned}$$ Here, $C_1$ depends only on the polynomial degree $k$ and the shape-regularity of ${\mathcal}T_h$, while $C_2$ depends also on ${\mathcal}A$ and $C_{\mathrm}{reg}$.
Note that in [@QiShSh:2018 Theorem 2.1] the meshes ${\mathcal}T_h$ are assumed to be quasi-uniform, whereas here we only require ${\mathcal}T_h$ to be shape-regular.
Proof of Theorem \[th:st\] {#sec:st_prf}
--------------------------
To begin with, we use the element-by-element projection defined in (we will only use Theorem \[th:PROJ\], not the definition) on the solution $({\boldsymbol{\sigma}},{\boldsymbol{u}})$: $$\begin{aligned}
(\Pi{\boldsymbol{\sigma}},\Pi{\boldsymbol{u}})
:=\prod_{K\in{\mathcal}T_h}\Pi({\boldsymbol{\sigma}}\big|_K,{\boldsymbol{u}}\big|_K;{\boldsymbol{\tau}}\big|_{{\partial}K}),\quad
{\boldsymbol{\delta}}
:= \prod_{K\in{\mathcal}T_h}\mathrm R({\boldsymbol{\sigma}}\big|_K,{\boldsymbol{u}}\big|_K;{\boldsymbol{\tau}}\big|_{{\partial}K}),\end{aligned}$$ and define the error terms and the approximation terms: $$\begin{aligned}
{5}
{\boldsymbol{\varepsilon}}_h^u&:=\Pi{\boldsymbol{u}}-{\boldsymbol{u}}_h,&\quad
{\boldsymbol{\varepsilon}}_h^\sigma&:=\Pi{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_h,\quad
{\boldsymbol{\varepsilon}}_h^{\widehat{u}}:={\mathrm}P_M{\boldsymbol{u}}-\widehat{{\boldsymbol{u}}}_h,\\
{\boldsymbol{e}}_\sigma&:=\Pi{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}},&
{\boldsymbol{e}}_u&:=\Pi{\boldsymbol{u}}-{\boldsymbol{u}}.\end{aligned}$$ We aim to control the terms ${\boldsymbol{\varepsilon}}_h^*$ ($*={\boldsymbol{u}},{\boldsymbol{\sigma}},\widehat{{\boldsymbol{u}}}$) by the terms ${\boldsymbol{e}}_*$ ($*={\boldsymbol{\sigma}},{\boldsymbol{u}}$) and ${\boldsymbol{\delta}}$.
\[prop:energy\_id\] The following energy identity holds $$\begin{aligned}
\label{eq:st_id}
({\mathcal}A{\boldsymbol{\varepsilon}}_h^\sigma,{\boldsymbol{\varepsilon}}_h^\sigma)_{{\mathcal}T_h}
+{\langle {\boldsymbol{\tau}} {\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}})
,({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}})\rangle}_{{\partial}{\mathcal}T_h}
=({\mathcal}A{\boldsymbol{e}}_\sigma,{\boldsymbol{\varepsilon}}_h^\sigma)_{{\mathcal}T_h}+{\langle {\boldsymbol{\delta}},{\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\rangle}_{{\partial}{\mathcal}T_h}.\end{aligned}$$ Consequently, $$\begin{aligned}
\label{eq:st_est}
\|{\boldsymbol{\varepsilon}}_h^\sigma\|_{{\mathcal}A}^2+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat u})\|_{{\boldsymbol{\tau}}}^2
\le
\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}A}^2+\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}^2.\end{aligned}$$
The proof here will be similar to the proof of [@CoGoSa:2010 Lemma 3.1 & 3.2]. By Theorem \[th:PROJ\], we first obtain a set of projection equations satisfied by $\Pi{\boldsymbol{\sigma}}$, $\Pi{\boldsymbol{u}}$ and ${\boldsymbol{\delta}}$. We then subtract from the projection equations and obtain the following error equations:
\[eq:st\_ereq\] $$\begin{aligned}
\label{eq:st_ereq_1}
({\mathcal}A{\boldsymbol{\varepsilon}}_h^\sigma,{\boldsymbol{\theta}})_{{\mathcal}T_h}
+ ({\boldsymbol{\varepsilon}}_h^u,{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}
-{\langle {\boldsymbol{\varepsilon}}_h^{\widehat{u}},{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}
&=({\mathcal}A{\boldsymbol{e}}_\sigma,{\boldsymbol{\theta}})_{{\mathcal}T_h},\\
\label{eq:st_ereq_2}
-({\mathrm{div}\hspace{1pt}}{\boldsymbol{\varepsilon}}_h^\sigma,{\boldsymbol{w}}) _{{\mathcal}T_h}
+ {\langle {\boldsymbol{\tau}} {\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
&={\langle {\boldsymbol{\delta}},{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h},\\
\label{eq:st_ereq_3}
{\langle {\boldsymbol{\varepsilon}}_h^\sigma{\boldsymbol{n}}
-{\boldsymbol{\tau}} {\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}),{\boldsymbol{\mu}}\rangle}
_{\partial{\mathcal}T_h\backslash\Gamma}
&=-{\langle {\boldsymbol{\delta}},{\boldsymbol{\mu}}\rangle}_{\partial{\mathcal}T_h\backslash\Gamma},\\
\label{eq:st_ereq_4}
{\langle {\boldsymbol{\varepsilon}}_h^{\widehat{u}},{\boldsymbol{\mu}}\rangle}_\Gamma &= 0.\end{aligned}$$
By and we have $$\begin{aligned}
\label{eq:st_ereq_3a}
{\langle {\boldsymbol{\varepsilon}}_h^\sigma{\boldsymbol{n}}
-{\boldsymbol{\tau}} {\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}),
{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\rangle}
_{\partial{\mathcal}T_h}
&=-{\langle {\boldsymbol{\delta}},
{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\rangle}_{\partial{\mathcal}T_h}.\end{aligned}$$ Now taking ${\boldsymbol{\theta}} = {\boldsymbol{\varepsilon}}_h^\sigma$ in , ${\boldsymbol{w}}={\boldsymbol{\varepsilon}}_h^u$ in , and then adding , and , we obtain the energy identity , from which the estimate follows easily.
\[prop:st\_dual\] If $k\ge1$ and holds, then $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}\le hC
\left(
\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}+\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}
\right).\end{aligned}$$ Here, $C$ is independent of $h$, but depends on ${\mathcal}A$ and $C_{\mathrm}{reg}$.
The proof here will be similar to the proof of [@CoGoSa:2010 Lemma 4.1 & Theorem 4.1]. Consider the adjoint equations . We take ${\boldsymbol{\Theta}}={\boldsymbol{\varepsilon}}_h^u$ as the input data and apply the projection on $({\boldsymbol{\Psi}},{\boldsymbol{\Phi}})$ with $-{\boldsymbol{\tau}}$ as the stabilization function. Namely, $$\begin{aligned}
(\Pi{\boldsymbol{\Psi}},\Pi{\boldsymbol{\Phi}}):=
\prod_{K\in{\mathcal}T_h}\Pi({\boldsymbol{\Psi}}\big|_K,{\boldsymbol{\Phi}}\big|_K;-{\boldsymbol{\tau}}\big|_{{\partial}K}),\quad
{\boldsymbol{\Delta}}
:= \prod_{K\in{\mathcal}T_h}\mathrm R({\boldsymbol{\Psi}}\big|_K,{\boldsymbol{\Phi}}\big|_K;-{\boldsymbol{\tau}}\big|_{{\partial}K}).\end{aligned}$$ By Theorem \[th:PROJ\], $\Pi{\boldsymbol{\Psi}}$, $\Pi{\boldsymbol{\Phi}}$ and ${\boldsymbol{\Delta}}$ satisfy $$\begin{aligned}
({\mathcal}A\Pi{\boldsymbol{\Psi}},{\boldsymbol{\theta}})_{{\mathcal}T_h} - (\Pi{\boldsymbol{\Phi}},{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}
+{\langle {\mathrm}P_M{\boldsymbol{\Phi}},{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}
&=({\mathcal}A(\Pi{\boldsymbol{\Psi}}-{\boldsymbol{\Psi}}),{\boldsymbol{\theta}})_{{\mathcal}T_h},\\
({\mathrm{div}\hspace{1pt}}\Pi{\boldsymbol{\Psi}},{\boldsymbol{w}})_{{\mathcal}T_h} + {\langle {\boldsymbol{\tau}} {\mathrm}P_M(\Pi{\boldsymbol{\Phi}}-{\boldsymbol{\Phi}}),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
&=({\boldsymbol{\varepsilon}}_h^u,{\boldsymbol{w}})_{{\mathcal}T_h}-{\langle {\boldsymbol{\Delta}},{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h},\\
-{\langle \Pi{\boldsymbol{\Psi}}{\boldsymbol{n}} +{\boldsymbol{\tau}} {\mathrm}P_M(\Pi{\boldsymbol{\Phi}}-{\boldsymbol{\Phi}}),{\boldsymbol{\mu}}\rangle}
_{\partial{\mathcal}T_h\backslash\Gamma}
&={\langle {\boldsymbol{\Delta}},{\boldsymbol{\mu}}\rangle}_{\partial{\mathcal}T_h\backslash\Gamma},\\
{\langle {\mathrm}P_M {\boldsymbol{\Phi}},{\boldsymbol{\mu}}\rangle}_\Gamma &= 0.\end{aligned}$$ We now test the above equations with ${\boldsymbol{\theta}}={\boldsymbol{\varepsilon}}_h^\sigma$, ${\boldsymbol{w}}={\boldsymbol{\varepsilon}}_h^u$ and ${\boldsymbol{\mu}}={\boldsymbol{\varepsilon}}_h^{\widehat{u}}$, then test the error equations with ${\boldsymbol{\theta}}=\Pi{\boldsymbol{\Psi}}$, ${\boldsymbol{w}}=\Pi{\boldsymbol{\Phi}}$ and ${\boldsymbol{\mu}}={\mathrm}P_M{\boldsymbol{\Phi}}$. By comparing the two sets of equations, we obtain $$\begin{aligned}
({\mathcal}A(\Pi{\boldsymbol{\Psi}}-{\boldsymbol{\Psi}}),{\boldsymbol{\varepsilon}}_h^\sigma)_{{\mathcal}T_h}
+\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}^2
-{\langle {\boldsymbol{\Delta}},{\mathrm}P_M{\boldsymbol{\varepsilon}}_h^u
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\rangle}_{{\partial}{\mathcal}T_h}
=({\mathcal}A{\boldsymbol{e}}_\sigma,\Pi{\boldsymbol{\Psi}})_{{\mathcal}T_h}
+{\langle {\boldsymbol{\delta}},{\mathrm}P_M(\Pi{\boldsymbol{\Phi}}-{\boldsymbol{\Phi}})\rangle}_{{\partial}{\mathcal}T_h}.\end{aligned}$$ After rearranging terms, we have $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}^2
&=-({\mathcal}A(\Pi{\boldsymbol{\Psi}}-{\boldsymbol{\Psi}}),{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_h)_{{\mathcal}T_h}-(\Pi{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}},
{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}}))_{{\mathcal}T_h}\\
&\quad\, +{\langle {\boldsymbol{\Delta}},{\mathrm}P_M{\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\rangle}_{{\partial}{\mathcal}T_h}
+{\langle {\boldsymbol{\delta}},{\mathrm}P_M(\Pi{\boldsymbol{\Phi}}-{\boldsymbol{\Phi}})\rangle}_{{\partial}{\mathcal}T_h}.\end{aligned}$$ By Theorem \[th:PROJ\] (with $m=1$) we have $$\begin{aligned}
\|\Pi{\boldsymbol{\Psi}}-{\boldsymbol{\Psi}}\|_{{\mathcal}T_h}+\|{\boldsymbol{\Delta}}\|_{{\boldsymbol{\tau}}^{-1}}
\lesssim h(\|{\boldsymbol{\Psi}}\|_{1,\Omega} + \|{\boldsymbol{\Phi}}\|_{2,\Omega}).\end{aligned}$$ By we have $$\begin{aligned}
(\Pi{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}},
{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}}))_{{\mathcal}T_h}
=(\Pi{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}},
{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}})-{\mathrm}P_0{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}}))_{{\mathcal}T_h}
\lesssim h\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}\|{\boldsymbol{\Phi}}\|_{2,\Omega}.\end{aligned}$$ Finally, implies $$\begin{aligned}
\|{\mathrm}P_M(\Pi{\boldsymbol{\Phi}}-{\boldsymbol{\Phi}})\|_{{\boldsymbol{\tau}}}\approx \left(\sum_{K\in{\mathcal}T_h}\|h_K^{-1/2}{\mathrm}P_M(\Pi{\boldsymbol{\Phi}}-{\boldsymbol{\Phi}})\|_{\partial K}^2\right)^{1/2}
\lesssim h(\|{\boldsymbol{\Psi}}\|_{1,\Omega} + \|{\boldsymbol{\Phi}}\|_{2,\Omega}).\end{aligned}$$ We next use to control the term $(\|{\boldsymbol{\Psi}}\|_{1,\Omega} + \|{\boldsymbol{\Phi}}\|_{2,\Omega})$ and use Proposition \[prop:energy\_id\] to control the terms $\|{\mathcal}A({\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_h)\|_{{\mathcal}T_h}$ and $\|{\mathrm}P_M{\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\|_{{\boldsymbol{\tau}}}$. The proof is thus completed.
Combing Proposition \[prop:energy\_id\], Proposition \[prop:st\_dual\] and Theorem \[th:PROJ\], Theorem \[th:st\] follows readily.
Frequency-domain elastodynamics {#sec:freq}
===============================
Method and convergence estimates {#method-and-convergence-estimates}
--------------------------------
In this section we give new proofs of error estimates to the second HDG+ method studied in [@HuPrSa:2017] by using projection based analysis for the following time-harmonic linear elasticity problem
\[eq:fq\_pde\] $$\begin{aligned}
{5}
{\mathcal}A{\boldsymbol{\sigma}} - {\boldsymbol{\varepsilon}}({\boldsymbol{u}}) &={\boldsymbol{0}} &\qquad& {\mathrm}{in}\ \Omega,\\
-{\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}} - \kappa^2\rho{\boldsymbol{u}}&= {\boldsymbol{f}} && {\mathrm}{in}\ \Omega,\\
\gamma{\boldsymbol{u}}&= {\boldsymbol{g}} && {\mathrm}{on}\ \Gamma:={\partial}\Omega,\end{aligned}$$
where ${\boldsymbol{f}}\in L^2(\Omega;{\mathbb}C^3)$, ${\boldsymbol{g}}\in H^{1/2}(\Gamma;{\mathbb}C^3)$, and $\rho\in W^{1,\infty}(\Omega;{\mathbb}R)$ denotes the density function. We assume that the wave number $\kappa>0$ and $\kappa^2$ is not a Dirichlet eigenvalue so that is well-posed. Note that because of the different proof techniques, the dependence on the wave numbers of the error estimates will be different and not easy to compare. The emphasis here will be the simplified error analysis thanks to the introduction of the projection.
The HDG+ method for is: find $({\boldsymbol{\sigma}}_h,{\boldsymbol{u}}_h,\widehat{{\boldsymbol{u}}}_h)\in {\boldsymbol{V}}_h\times{\boldsymbol{W}}_h\times{\boldsymbol{M}}_h$ such that
\[eq:fq\_hdg\] $$\begin{aligned}
({\mathcal}A{\boldsymbol{\sigma}}_h,{\boldsymbol{\theta}})_{{\mathcal}T_h}
+ ({\boldsymbol{u}}_h,{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}
-{\langle \widehat{{\boldsymbol{u}}}_h,{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}
&=0,\\
-({\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}}_h,{\boldsymbol{w}})_{{\mathcal}T_h}
+ {\langle {\boldsymbol{\tau}} {\mathrm}P_M({\boldsymbol{u}}_h-\widehat{{\boldsymbol{u}}}_h),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
-(\kappa^2\rho {\boldsymbol{u}}_h,{\boldsymbol{w}})_{{\mathcal}T_h}
&=({\boldsymbol{f}},{\boldsymbol{w}})_{{\mathcal}T_h},\\
{\langle {\boldsymbol{\sigma}}_h{\boldsymbol{n}} -{\boldsymbol{\tau}} {\mathrm}P_M({\boldsymbol{u}}_h-\widehat{{\boldsymbol{u}}}_h),{\boldsymbol{\mu}}\rangle}
_{\partial{\mathcal}T_h\backslash\Gamma}
&=0,\\
{\langle \widehat{{\boldsymbol{u}}}_h,{\boldsymbol{\mu}}\rangle}_\Gamma &= {\langle {\boldsymbol{g}},{\boldsymbol{\mu}}\rangle}_{\Gamma},\end{aligned}$$
for all ${\boldsymbol{\theta}}\in{\boldsymbol{V}}_h$, ${\boldsymbol{w}}\in{\boldsymbol{W}}_h$ and ${\boldsymbol{\mu}}\in{\boldsymbol{M}}_h$. Here, the spaces ${\boldsymbol{V}}_h$, ${\boldsymbol{W}}_h$ and ${\boldsymbol{M}}_h$ are the same as those defined in Section \[sec:st\_main\] except now the functions contained in these spaces take complex values.
For ${\boldsymbol{\Theta}}\in L^2(\Omega;{\mathbb}R^3)$, we assume that the solution to the adjoint equations for
\[eq:fq\_adpde\] $$\begin{aligned}
{5}
{\mathcal}A{\boldsymbol{\Psi}}+ {\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}}) &={\boldsymbol{0}} &\qquad & {\mathrm}{in}\ \Omega,\\
{\mathrm{div}\hspace{1pt}}{\boldsymbol{\Psi}} - \kappa^2\rho{\boldsymbol{\Phi}}&= {\boldsymbol{\Theta}} && {\mathrm}{in}\ \Omega,\\
\gamma{\boldsymbol{\Phi}}&= {\boldsymbol{0}} &&{\mathrm}{on}\ \Gamma,\end{aligned}$$
satisfy $$\begin{aligned}
\label{eq:fq_reg}
\|{\boldsymbol{\Phi}}\|_{2,\Omega} + \|{\boldsymbol{\Psi}}\|_{1,\Omega}
\le C_\kappa\|{\boldsymbol{\Theta}}\|_\Omega,\end{aligned}$$ where we make the dependence on $\kappa$ explicit for the constant $C_\kappa$. For the rest of this section, the wiggles sign $\lesssim$ will hide constants independent of $h$ and $\kappa$. We aim to prove the following theorem.
\[th:fq\] Suppose $k\ge1$ and holds. If $(h^2\kappa^2+h\kappa)C_\kappa$ is small enough, then $$\begin{aligned}
\|{\boldsymbol{u}}-{\boldsymbol{u}}_h\|_{\Omega}&\lesssim
\left(
(1+C_\kappa)h^{m+1} + C_\kappa(\kappa^2+\kappa) h^{m+2}
\right)
\left(
\|{\boldsymbol{\sigma}}\|_{m,\Omega}+\|{\boldsymbol{u}}\|_{m+1,\Omega}
\right),\\
\|{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_h\|_{\Omega}&\lesssim
\left(
h^m+(1+C_\kappa)\kappa h^{m+1} + C_\kappa(\kappa^3+\kappa^2)h^{m+2}
\right)
\left(
\|{\boldsymbol{\sigma}}\|_{m,\Omega}+\|{\boldsymbol{u}}\|_{m+1,\Omega}
\right),\end{aligned}$$ with $1\le m\le k+1$.
Proof of Theorem \[th:fq\]
--------------------------
Since the solution $({\boldsymbol{\sigma}},{\boldsymbol{u}})$ in can take complex values, applying directly Theorem \[th:PROJ\] is not feasible. However, this can be easily fixed by defining a new complex projection based on the original one: for $({\boldsymbol{\sigma}},{\boldsymbol{u}})\in H^1(K;{\mathbb}C_{\mathrm}{sym}^{3\times3})\times H^1(K;{\mathbb}C^3)$, we define $$\begin{aligned}
\label{def:com_proj}
\Pi({\boldsymbol{\sigma}},{\boldsymbol{u}}):=\Pi({\mathrm}{Re}\,{\boldsymbol{\sigma}},{\mathrm}{Re}\,{\boldsymbol{u}})+i\Pi({\mathrm}{Im}\,{\boldsymbol{\sigma}},{\mathrm}{Im}\,{\boldsymbol{u}}).\end{aligned}$$ It is easy to show that this complex projection also satisfies and (the only difference is that now the test functions ${\boldsymbol{\theta}},{\boldsymbol{w}},{\boldsymbol{\mu}}$ in can take complex values).
Similar to the beginning of Section \[sec:st\_prf\], we first define the projections $(\Pi{\boldsymbol{\sigma}},\Pi{\boldsymbol{u}})$, the remainder term ${\boldsymbol{\delta}}$, the error ${\boldsymbol{\varepsilon}}_h^u$, ${\boldsymbol{\varepsilon}}_h^\sigma$, ${\boldsymbol{\varepsilon}}^{\widehat{u}}$, and approximation terms ${\boldsymbol{e}}_\sigma$, ${\boldsymbol{e}}_u$.
\[prop:fq\_id\] The following energy identity holds $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^\sigma\|_{{\mathcal}A}^2
+\|{\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}})\|_{{\boldsymbol{\tau}}}^2
-\kappa^2\|{\boldsymbol{\varepsilon}}_h^u\|_\rho^2
&=({\mathcal}A\overline{{\boldsymbol{e}}_\sigma},{\boldsymbol{\varepsilon}}_h^\sigma)_{{\mathcal}T_h}
+{\langle \overline{{\boldsymbol{\delta}}},{\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\rangle}_{{\partial}{\mathcal}T_h}
-\kappa^2(\rho\overline{{\boldsymbol{e}}_u},{\boldsymbol{\varepsilon}}_h^u)_{{\mathcal}T_h}.\end{aligned}$$ Consequently, $$\begin{aligned}
\label{eq:fq_ene}
\|{\boldsymbol{\varepsilon}}_h^\sigma\|_{{\mathcal}T_h} + \|{\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}})\|_{{\boldsymbol{\tau}}}\lesssim \kappa\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}+\kappa\|{\boldsymbol{e}}_u\|_{{\mathcal}T_h}
+\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}+\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}.\end{aligned}$$
By similar ideas used in the proof of Proposition \[prop:energy\_id\], we obtain the following error equations:
\[eq:fq\_erreq\] $$\begin{aligned}
{5}
({\mathcal}A{\boldsymbol{\varepsilon}}_h^\sigma,{\boldsymbol{\theta}})_{{\mathcal}T_h} + ({\boldsymbol{\varepsilon}}_h^u,{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}-{\langle {\boldsymbol{\varepsilon}}_h^{\widehat{u}},{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}&=({\mathcal}A{\boldsymbol{e}}_\sigma,{\boldsymbol{\theta}})_{{\mathcal}T_h},\\
-({\mathrm{div}\hspace{1pt}}{\boldsymbol{\varepsilon}}_h^\sigma,{\boldsymbol{w}})_{{\mathcal}T_h} + {\langle {\boldsymbol{\tau}} {\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
-(\kappa^2\rho {\boldsymbol{\varepsilon}}_h^u,{\boldsymbol{w}})_{{\mathcal}T_h}
&={\langle {\boldsymbol{\delta}},{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}-(\kappa^2\rho {\boldsymbol{e}}_u,{\boldsymbol{w}})_{{\mathcal}T_h},\\
{\langle {\boldsymbol{\varepsilon}}_h^\sigma{\boldsymbol{n}} -{\boldsymbol{\tau}} {\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}),{\boldsymbol{\mu}}\rangle}
_{\partial{\mathcal}T_h\backslash\Gamma}
&=-{\langle {\boldsymbol{\delta}},{\boldsymbol{\mu}}\rangle}_{\partial{\mathcal}T_h\backslash\Gamma},\\
{\langle {\boldsymbol{\varepsilon}}_h^{\widehat{u}},{\boldsymbol{\mu}}\rangle}_\Gamma &= 0,\end{aligned}$$
for all ${\boldsymbol{\theta}}\in{\boldsymbol{V}}_h$, ${\boldsymbol{w}}\in{\boldsymbol{W}}_h$ and ${\boldsymbol{\mu}}\in{\boldsymbol{M}}_h$. Taking ${\boldsymbol{\theta}}=\overline{{\boldsymbol{\varepsilon}}_h^\sigma}$, ${\boldsymbol{w}}=\overline{{\boldsymbol{\varepsilon}}_h^u}$ and ${\boldsymbol{\mu}}=\overline{{\boldsymbol{\varepsilon}}_h^{\widehat{u}}}$, then adding the equations, we have the energy identity.
Denoting $A^2:=\|{\boldsymbol{\varepsilon}}_h^\sigma\|_{{\mathcal}T_h}^2 + \|{\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}})\|_{{\boldsymbol{\tau}}}^2$, and applying the Cauchy-Schwarz inequality on the identity, we have $$A^2\lesssim \kappa^2\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}^2 + \kappa^2\|{\boldsymbol{e}}_u\|_{{\mathcal}T_h}^2
+A(\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}+\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}).$$ The estimate now follows by using Young’s inequality.
\[prop:fq\_est\] Suppose $k\ge1$ and holds. If $(h^2\kappa^2+h\kappa)C_\kappa$ is small enough, then $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}&\lesssim h(\kappa^2+\kappa)C_\kappa\|{\boldsymbol{e}}_u\|_{{\mathcal}T_h}
+h C_\kappa(\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}+\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}),\\
\|{\boldsymbol{\varepsilon}}_h^\sigma\|_{{\mathcal}T_h} + \|{\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}})\|_{{\boldsymbol{\tau}}}&\lesssim
(1+h\kappa C_\kappa)(\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}+\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}) + (\kappa+h C_\kappa(\kappa^3+\kappa^2))\|{\boldsymbol{e}}_u\|_{{\mathcal}T_h}.\end{aligned}$$
Consider the adjoint equations and take ${\boldsymbol{\Theta}} = {\boldsymbol{\varepsilon}}_h^u$. Following similar ideas in the proof of Proposition \[prop:st\_dual\], we first apply the projection on $({\boldsymbol{\Psi}},{\boldsymbol{\Phi}})$ with stabilization function $-{\boldsymbol{\tau}}$ to obtain the projection equations (about $(\Pi{\boldsymbol{\Psi}},\Pi{\boldsymbol{\Phi}},{\boldsymbol{\Delta}})$). We next test the projection equations with ${\boldsymbol{\varepsilon}}_h^\sigma$, ${\boldsymbol{\varepsilon}}_h^u$ and ${\boldsymbol{\varepsilon}}_h^{\widehat{u}}$, test the error equations with $\overline{\Pi{\boldsymbol{\Psi}}}$, $\overline{\Pi{\boldsymbol{\Psi}}}$ and $\overline{{\mathrm}P_M{\boldsymbol{\Phi}}}$, and compare the two sets of equations. Then we obtain (define ${\boldsymbol{e}}_\Psi:=\Pi{\boldsymbol{\Psi}}-{\boldsymbol{\Psi}}$ and ${\boldsymbol{e}}_\Phi:=\Pi{\boldsymbol{\Phi}}-{\boldsymbol{\Phi}}$ for convenience) $$\begin{aligned}
&({\mathcal}A{\boldsymbol{e}}_\sigma,\overline{\Pi{\boldsymbol{\Psi}}})_{{\mathcal}T_h} + {\langle {\boldsymbol{\delta}},\overline{\Pi{\boldsymbol{\Phi}}}-\overline{{\mathrm}P_M{\boldsymbol{\Phi}}}\rangle}_{{\partial}{\mathcal}T_h}+\kappa^2(\rho({\boldsymbol{u}}-{\boldsymbol{u}}_h),\overline{\Pi{\boldsymbol{\Phi}}})_{{\mathcal}T_h}\\
&\qquad\qquad=({\mathcal}A\overline{{\boldsymbol{e}}_\Psi}, {\boldsymbol{\varepsilon}}_h^\sigma)_{{\mathcal}T_h}+\kappa^2(\rho\overline{{\boldsymbol{\Phi}}},{\boldsymbol{\varepsilon}}_h^u)_{{\mathcal}T_h}+\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}^2-{\langle \overline{{\boldsymbol{\Delta}}},{\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\rangle}_{{\partial}{\mathcal}T_h}.\end{aligned}$$ After rearranging terms, we have $$\begin{aligned}
{6}\nonumber
\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}^2
&=\kappa^2(\overline{{\boldsymbol{e}}_\Phi},\rho({\boldsymbol{u}}-{\boldsymbol{u}}_h))_{{\mathcal}T_h} + \kappa^2(\rho\overline{{\boldsymbol{\Phi}}},{\boldsymbol{u}}-\Pi{\boldsymbol{u}})_{{\mathcal}T_h}
+({\mathcal}A({\boldsymbol{\sigma}}_h-{\boldsymbol{\sigma}}),\overline{{\boldsymbol{e}}_\Psi})_{{\mathcal}T_h} - ({\boldsymbol{e}}_\sigma,\overline{{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}})})_{{\mathcal}T_h}\\
\label{eq:6.99}
&\quad +{\langle {\boldsymbol{\delta}},\overline{\Pi{\boldsymbol{\Phi}}}-\overline{\mathrm{P}_M{\boldsymbol{\Phi}}}\rangle}_{{\partial}{\mathcal}T_h}
+{\langle \overline{{\boldsymbol{\Delta}}},{\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\rangle}_{{\partial}{\mathcal}T_h}.\end{aligned}$$ By Theorem \[th:PROJ\] (with $m=1$), we have $$h^{-1}\|{\boldsymbol{e}}_\Phi\|_{{\mathcal}T_h}+\|{\boldsymbol{e}}_\Psi\|_{{\mathcal}T_h}+\|\Pi{\boldsymbol{\Phi}}-{\mathrm}P_M{\boldsymbol{\Phi}}\|_{{\boldsymbol{\tau}}}
+\|{\boldsymbol{\Delta}}\|_{{\boldsymbol{\tau}}^{-1}}
\lesssim h(\|{\boldsymbol{\Psi}}\|_{1,\Omega}+\|{\boldsymbol{\Phi}}\|_{2,\Omega}),$$ and also $$\begin{aligned}
\kappa^2(\rho\overline{{\boldsymbol{\Phi}}},{\boldsymbol{u}}-\Pi{\boldsymbol{u}})_{{\mathcal}T_h}
=\kappa^2(\rho\overline{{\boldsymbol{\Phi}}}-{\mathrm}P_0(\rho\overline{{\boldsymbol{\Phi}}}),{\boldsymbol{u}}-\Pi{\boldsymbol{u}})_{{\mathcal}T_h}\lesssim h\kappa^2\|\rho\|_{W^{1,\infty}(\Omega)}\|{\boldsymbol{\Phi}}\|_{1,\Omega}\|{\boldsymbol{e}}_u\|_{{\mathcal}T_h}.\end{aligned}$$ By we have $({\boldsymbol{e}}_\sigma,{\boldsymbol{\varepsilon}}(\overline{{\boldsymbol{\Phi}}}))_{{\mathcal}T_h}
=({\boldsymbol{e}}_\sigma,{\boldsymbol{\varepsilon}}(\overline{{\boldsymbol{\Phi}}})-{\mathrm}P_0{\boldsymbol{\varepsilon}}(\overline{{\boldsymbol{\Phi}}}))_{{\mathcal}T_h}
\lesssim h\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}\|{\boldsymbol{\Phi}}\|_{2,\Omega}$. Now we use in to obtain $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}
&\lesssim h^2\kappa^2C_\kappa\|{\boldsymbol{u}}-{\boldsymbol{u}}_h\|_{{\mathcal}T_h}+ h\kappa^2C_\kappa\|\Pi{\boldsymbol{u}}-{\boldsymbol{u}}\|_{{\mathcal}T_h}
+hC_\kappa\|{\boldsymbol{\sigma}}-{\boldsymbol{\sigma}}_h\|_{{\mathcal}T_h}+hC_\kappa\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}\\
&\quad +hC_\kappa\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}+hC_\kappa\|{\mathrm}P_M{\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\|_{{\boldsymbol{\tau}}},\\
&\lesssim h^2\kappa^2C_\kappa\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}+ h\kappa^2C_\kappa\|{\boldsymbol{e}}_u\|_{{\mathcal}T_h}
+hC_\kappa\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}\\
&\quad +hC_\kappa\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}
+hC_\kappa(\|{\boldsymbol{\varepsilon}}_h^\sigma\|_{{\mathcal}T_h}+\|{\mathrm}P_M{\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}\|_{{\boldsymbol{\tau}}}).\end{aligned}$$ Next we use and bound $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}\lesssim
(h^2\kappa^2+h\kappa)C_\kappa\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}+(h\kappa+h\kappa^2)C_\kappa\|{\boldsymbol{e}}_u\|_{{\mathcal}T_h} + hC_\kappa(\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}+\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}).\end{aligned}$$ Therefore, when $(h^2\kappa^2+h\kappa)C_\kappa$ is small enough, we have $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u\|_{{\mathcal}T_h}\lesssim
(h\kappa+h\kappa^2)C_\kappa\|{\boldsymbol{e}}_u\|_{{\mathcal}T_h} + hC_\kappa(\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}T_h}+\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}).\end{aligned}$$ The first inequality is thus proved. The second inequality can be proved by combining and the above inequality.
Theorem \[th:fq\] now follows easily from Proposition \[prop:fq\_est\] and Theorem \[th:PROJ\].
Transient elastodynamics {#sec:trans_elas}
========================
The semi-discrete HDG+ method
-----------------------------
In this section, we present a semi-discrete (in space) HDG+ method for transient elastic waves and prove it is uniformly-in-time optimal convergent. The equations we consider are
\[eq:elwv\_pde\] $$\begin{aligned}
{5}
{\mathcal}A{\boldsymbol{\sigma}}(t)
- {\boldsymbol{\varepsilon}}({\boldsymbol{u}}(t)) &=0
&\qquad&\text{in}\ \Omega\times[0,T],\\
\rho\ddot{{\boldsymbol{u}}}(t)
- {\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}}(t) &= {\boldsymbol{f}}(t)
&&\text{in}\ \Omega\times[0,T],\\
\gamma{\boldsymbol{u}}(t)&={\boldsymbol{g}}(t)
&&\text{on}\ \Gamma\times[0,T],\\
{\boldsymbol{u}}(0)&= {\boldsymbol{u}}_0
&&\text{on}\ \Omega,\\
\dot{{\boldsymbol{u}}}(0)&= {\boldsymbol{v}}_0
&&\text{on}\ \Omega,\end{aligned}$$
where ${\boldsymbol{f}}\in C([0,\infty);L^2(\Omega;{\mathbb}R^3))$ and ${\boldsymbol{g}}\in C([0,\infty);H^{1/2}(\Gamma;{\mathbb}R^3))$. For the initial conditions, we assume ${\boldsymbol{u}}_0,{\boldsymbol{v}}_0\in L^2(\Omega;{\mathbb}R^3)$.
The HDG+ method for looks for $$\begin{aligned}
{\boldsymbol{\sigma}}_h,{\boldsymbol{u}}_h,\widehat{{\boldsymbol{u}}}_h:
[0,\infty)\rightarrow {\boldsymbol{V}}_h\times{\boldsymbol{W}}_h\times{\boldsymbol{M}}_h,\end{aligned}$$ such that for all $t\ge0$
\[eq:elwv\_HDG\] $$\begin{aligned}
\label{eq:elwv_HDGa}
({\mathcal}A{\boldsymbol{\sigma}}_h(t),{\boldsymbol{\theta}})_{{\mathcal}T_h} + ({\boldsymbol{u}}_h(t),{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}
-{\langle \widehat{{\boldsymbol{u}}}_h(t),{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}
&=0,\\
\label{eq:elwv_HDGb}
(\rho\ddot{{\boldsymbol{u}}}_h(t),{\boldsymbol{w}})_{{\mathcal}T_h} - ({\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}}_h(t),{\boldsymbol{w}})_{{\mathcal}T_h}
+{\langle {\boldsymbol{\tau}}\mathrm P_M({\boldsymbol{u}}_h(t)-\widehat{{\boldsymbol{u}}}_h(t)),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
&=({\boldsymbol{f}}(t),{\boldsymbol{w}})_{{\mathcal}T_h},\\
\label{eq:elwv_HDGc}
{\langle {\boldsymbol{\sigma}}_h(t){\boldsymbol{n}}-{\boldsymbol{\tau}}{\mathrm}P_M({\boldsymbol{u}}_h(t)-\widehat{{\boldsymbol{u}}}_h(t)),{\boldsymbol{\mu}}\rangle}_{{\partial}{\mathcal}T_h\backslash\Gamma}
&= 0,\\
\label{eq:elwv_HDGd}
{\langle \widehat{{\boldsymbol{u}}}_h(t),{\boldsymbol{\mu}}\rangle}_\Gamma
&= {\langle {\boldsymbol{g}}(t),{\boldsymbol{\mu}}\rangle}_{\Gamma},\end{aligned}$$
for all ${\boldsymbol{\theta}}\in{\boldsymbol{V}}_h$, ${\boldsymbol{w}}\in{\boldsymbol{W}}_h$ and ${\boldsymbol{\mu}}\in{\boldsymbol{M}}_h$. For the initial conditions of the method, we use ideas from [@CoFuHuJiSaMaSa:2018]. The initial velocity $\dot{{\boldsymbol{u}}}_h(0)$ is defined by using the projection in Theorem \[th:PROJ\] $$\begin{aligned}
\label{eq:elwv_iniv}
(\times,\dot{{\boldsymbol{u}}}_h(0))&= \Pi({\mathcal}A^{-1}{\boldsymbol{\varepsilon}}({\boldsymbol{v}}_0),{\boldsymbol{v}}_0;{\boldsymbol{\tau}}),\end{aligned}$$ and the initial displacement ${\boldsymbol{u}}_h(0)$ is defined to be the solution of the HDG+ discretization of the steady-state system
\[eq:elwv\_pdet0\] $$\begin{aligned}
{5}
{\mathcal}A{\boldsymbol{\sigma}}(0) - {\boldsymbol{\varepsilon}}({\boldsymbol{u}}(0))&=0
&\qquad &\text{in}\ \Omega,\\
-{\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}}(0) &=
-{\mathrm{div}\hspace{1pt}}({\mathcal}A^{-1}{\boldsymbol{\varepsilon}}({\boldsymbol{u}}_0))
&&\text{in}\ \Omega,\\
{\boldsymbol{u}}(0) &= g(0)
&&\text{on}\ \Gamma.\end{aligned}$$
Namely, we find $({\boldsymbol{\sigma}}_h(0),{\boldsymbol{u}}_h(0), \widehat{{\boldsymbol{u}}}_h(0))\in {\boldsymbol{V}}_h\times{\boldsymbol{W}}_h\times{\boldsymbol{M}}_h$ such that
\[eq:elwv\_HDGt0\] $$\begin{aligned}
\label{eq:elwv_HDGt0a}
({\mathcal}A{\boldsymbol{\sigma}}_h(0),{\boldsymbol{\theta}})_{{\mathcal}T_h} + ({\boldsymbol{u}}_h(0),{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}
-{\langle \widehat{{\boldsymbol{u}}}_h(0),{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}
&=0,\\
\label{eq:elwv_HDGt0b}
- ({\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}}_h(0),{\boldsymbol{w}})_{{\mathcal}T_h}
+{\langle {\boldsymbol{\tau}}\mathrm P_M({\boldsymbol{u}}_h(0)-\widehat{{\boldsymbol{u}}}_h(0)),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
&=(-{\mathrm{div}\hspace{1pt}}({\mathcal}A^{-1}{\boldsymbol{\varepsilon}}({\boldsymbol{u}}_0)),{\boldsymbol{w}})_{{\mathcal}T_h},\\
\label{eq:elwv_HDGt0c}
{\langle {\boldsymbol{\sigma}}_h(0){\boldsymbol{n}}-{\boldsymbol{\tau}}{\mathrm}P_M({\boldsymbol{u}}_h(0)-\widehat{{\boldsymbol{u}}}_h(0)),{\boldsymbol{\mu}}\rangle}_{{\partial}{\mathcal}T_h\backslash\Gamma}
&= 0,\\
\label{eq:elwv_HDGt0d}
{\langle \widehat{{\boldsymbol{u}}}_h(0),{\boldsymbol{\mu}}\rangle}_\Gamma
&= {\langle {\boldsymbol{g}}(0),{\boldsymbol{\mu}}\rangle}_{\Gamma},\end{aligned}$$
for all $({\boldsymbol{\theta}},{\boldsymbol{w}},{\boldsymbol{\mu}})\in {\boldsymbol{V}}_h\times {\boldsymbol{W}}_h \times{\boldsymbol{M}}_h$. For notational convenience, we define the following space-time norms $$\begin{aligned}
\vvvert\cdot\vvvert_{*,p}^{[0,T]}:=
\left(\int_0^T \|\cdot(t)\|_*^p\,\mathrm dt\right)^{1/p},\quad\end{aligned}$$ where $*$ can be replaced by $H^m(\Omega)$, $\Omega$, ${\mathcal}T_h$, ${\mathcal}A$, $\rho$, ${\partial}{\mathcal}T_h$, ${\boldsymbol{\tau}}$ or ${\boldsymbol{\tau}}^{-1}$. The parameter $p$ takes values in $\{1,2,\infty\}$, and when $p=\infty$, we consider the supreme norm in time instead of $L^p$ integration. Now we define the projections and the remainder terms for all $t\ge0$: $$\begin{aligned}
(\Pi{\boldsymbol{\sigma}}(t),\Pi{\boldsymbol{u}}(t))
:= \prod_{K\in{\mathcal}T_h}\Pi({\boldsymbol{\sigma}}(t)\big|_K,{\boldsymbol{u}}(t)\big|_K;{\boldsymbol{\tau}}\big|_{{\partial}K}),
\quad
{\boldsymbol{\delta}}(t):=\prod_{K\in{\mathcal}T_h}\mathrm R({\boldsymbol{\sigma}}(t)\big|_K,{\boldsymbol{u}}(t)\big|_K;{\boldsymbol{\tau}}\big|_{{\partial}K}).\end{aligned}$$ The related error and approximation terms (for all $t\ge0$) are denoted $$\begin{aligned}
{4}
{\boldsymbol{\varepsilon}}_h^\sigma(t) &:= \Pi{\boldsymbol{\sigma}}(t)-{\boldsymbol{\sigma}}_h(t),&\quad
{\boldsymbol{\varepsilon}}_h^u(t)&:=\Pi{\boldsymbol{u}}(t)-{\boldsymbol{u}}_h(t),\quad
{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(t)
:=\mathrm P_M{\boldsymbol{u}}(t)-\widehat{{\boldsymbol{u}}}_h(t),\\
{\boldsymbol{e}}_\sigma(t)&:=\Pi{\boldsymbol{\sigma}}(t)-{\boldsymbol{\sigma}}(t),&
{\boldsymbol{e}}_u(t)&:=\Pi{\boldsymbol{u}}(t)-{\boldsymbol{u}}(t).\end{aligned}$$ For the rest of this section, the wiggles sign $\lesssim$ will hide constants independent of $h$ and $T$. The main results in this section are Theorems \[th:elwv\_th1\] and \[th:elwv\_spconv\].
\[th:elwv\_th1\] For $k\ge1$, the following estimates hold $$\begin{aligned}
&\|{\boldsymbol{\varepsilon}}_h^\sigma(T)\|_{{\mathcal}A}
+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(T)
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(T))\|_{{\boldsymbol{\tau}}}
+\|\dot{{\boldsymbol{\varepsilon}}}_h^u(T)\|_\rho\\
&\qquad\qquad\lesssim
\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}A}
+\vvvert\dot{{\boldsymbol{e}}}_\sigma\vvvert_{{\mathcal}A,1}^{[0,T]}
+
\vvvert\ddot{{\boldsymbol{e}}}_u\vvvert_{\rho,1}^{[0,T]}
+
\vvvert{\boldsymbol{\delta}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}
+\vvvert\dot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},1}^{[0,T]}
,\\
&\|\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(T)\|_{{\mathcal}A}
+\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(T)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(T))\|_{{\boldsymbol{\tau}}}
+\|\ddot{{\boldsymbol{\varepsilon}}}_h^u(T)\|_\rho\\
&\qquad\qquad\lesssim
\|\dot{{\boldsymbol{e}}}_\sigma(0)\|_{{\mathcal}A}
+\|\ddot{{\boldsymbol{e}}}_u(0)\|_\rho
+\vvvert\ddot{{\boldsymbol{e}}}_\sigma\vvvert_{{\mathcal}A,1}^{[0,T]}
+\vvvert\dddot{{\boldsymbol{e}}}_u\vvvert_{\rho,1}^{[0,T]}
+\vvvert\dot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}
+\vvvert\ddot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},1}^{[0,T]}
.\end{aligned}$$
For the next theorem, we need to assume that the elliptic regularity condition $$\begin{aligned}
\label{eq:elwv_ellreg}
\|{\boldsymbol{\Phi}}\|_{2,\Omega}+\|{\mathcal}A^{-1}{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}})\|_{1,\Omega}\le C_{\mathrm}{reg}\|{\mathrm{div}\hspace{1pt}}({\mathcal}A^{-1}{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}}))\|,\end{aligned}$$ holds for any ${\boldsymbol{\Phi}}\in H^1(\Omega;{\mathbb}R^3)$ such that the right hand side of the above inequality is finite. Note that is the same as . We rephrase it here to have it in the form we will use it.
\[th:elwv\_spconv\] If $k\ge1$ and holds, then $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{\Omega}&\lesssim
h (1+T)^2
\bigg(
\|{\boldsymbol{e}}_\sigma(0)\|_{\Omega}
+\vvvert\dot{{\boldsymbol{e}}}_\sigma\vvvert_{\Omega,\infty}^{[0,T]}
+\vvvert\ddot{{\boldsymbol{e}}}_\sigma\vvvert_{\Omega,\infty}^{[0,T]}\\
&\phantom{\lesssim h(1+T)^2\bigg(}+\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}
+\vvvert\dot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}
+\vvvert\ddot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}\bigg)\\
&\quad +(1+T)^2\bigg(\|{\boldsymbol{e}}_u(0)\|_{\Omega}
+\vvvert\ddot{{\boldsymbol{e}}}_u\vvvert_{\Omega,\infty}^{[0,T]}
+\vvvert\dddot{{\boldsymbol{e}}}_u\vvvert_{\Omega,\infty}^{[0,T]}\bigg).\end{aligned}$$ Therefore, for $1\le m\le k+1$, $$\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{\Omega}\lesssim
h^{m+1} (1+T)^2\sum_{i=0}^3 \bigg(\vvvert{\boldsymbol{\sigma}}^{(i)}\vvvert_{H^m(\Omega),\infty}^{[0,T]}
+\vvvert{\boldsymbol{u}}^{(i)}\vvvert_{H^{m+1}(\Omega),\infty}^{[0,T]}\bigg),$$
We next give the proofs for the above two theorems in the following two subsections respectively.
Energy estimates
----------------
In this subsection, we give a proof to Theorem \[th:elwv\_th1\]. We first present two lemmas, which give the error equations when $t\ge0$ and the error equations when $t=0$, respectively.
For all $t\ge0$, the following error equations
\[eq:elwv\_ereq\] $$\begin{aligned}
\label{eq:elwv_ereqa}
({\mathcal}A{\boldsymbol{\varepsilon}}_h^\sigma(t),{\boldsymbol{\theta}})_{{\mathcal}T_h}
+ ({{\boldsymbol{\varepsilon}}_h^u}(t),{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}
-{\langle {\boldsymbol{\varepsilon}}_h^{\widehat u}(t),{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}
&=({\mathcal}A{\boldsymbol{e}}_\sigma(t),{\boldsymbol{\theta}})_{{\mathcal}T_h},\\
(\rho\ddot{{\boldsymbol{\varepsilon}}}_h^u(t),{\boldsymbol{w}})_{{\mathcal}T_h}
- ({\mathrm{div}\hspace{1pt}}{\boldsymbol{\varepsilon}}_h^\sigma(t),{\boldsymbol{w}})_{{\mathcal}T_h}
+{\langle {\boldsymbol{\tau}}\mathrm P_M({{\boldsymbol{\varepsilon}}_h^u}(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
&=(\rho\ddot{{\boldsymbol{e}}}_u(t),{\boldsymbol{w}})_{{\mathcal}T_h}\nonumber \\
\label{eq:elwv_ereqb}
&\quad +{\langle {\boldsymbol{\delta}}(t),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h},\\
\label{eq:elwv_ereqc}
{\langle {\boldsymbol{\varepsilon}}_h^\sigma(t){\boldsymbol{n}}
-{\boldsymbol{\tau}}({{\boldsymbol{\varepsilon}}_h^u}(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)),{\boldsymbol{\mu}}\rangle}
_{{\partial}{\mathcal}T_h\backslash\Gamma}
&= -{\langle {\boldsymbol{\delta}}(t),{\boldsymbol{\mu}}\rangle}_{{\partial}{\mathcal}T_h\backslash\Gamma},\\
\label{eq:elwv_ereqd}
{\langle {\boldsymbol{\varepsilon}}_h^{\widehat u}(t),{\boldsymbol{\mu}}\rangle}_\Gamma
&= 0,\\
\label{eq:elwv_ereqe}
{\boldsymbol{\varepsilon}}_h^u(0) &= \Pi{\boldsymbol{u}}_0-{{\boldsymbol{u}}}_h(0),\\
\label{eq:elwv_ereqf}
\dot{{\boldsymbol{\varepsilon}}}_h^u(0) &= {\boldsymbol{0}},\end{aligned}$$
hold for all $({\boldsymbol{\theta}},{\boldsymbol{w}},{\boldsymbol{\mu}})\in {\boldsymbol{V}}_h\times {\boldsymbol{W}}_h \times{\boldsymbol{M}}_h$.
Use , , , , and Theorem \[th:PROJ\].
The following error equations
\[eq:elwv\_ereqt0\] $$\begin{aligned}
\label{eq:elwv_ereqt0a}
({\mathcal}A{\boldsymbol{\varepsilon}}_h^\sigma(0),{\boldsymbol{\theta}})_{{\mathcal}T_h}
+ ({{\boldsymbol{\varepsilon}}_h^u}(0),{\mathrm{div}\hspace{1pt}}{\boldsymbol{\theta}})_{{\mathcal}T_h}
-{\langle {\boldsymbol{\varepsilon}}_h^{\widehat u}(0),{\boldsymbol{\theta}}{\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}
&=({\mathcal}A{\boldsymbol{e}}_\sigma(0),{\boldsymbol{\theta}})_{{\mathcal}T_h},\\
\label{eq:elwv_ereqt0b}
- ({\mathrm{div}\hspace{1pt}}{\boldsymbol{\varepsilon}}_h^\sigma(0),{\boldsymbol{w}})_{{\mathcal}T_h}
+{\langle {\boldsymbol{\tau}}\mathrm P_M({{\boldsymbol{\varepsilon}}_h^u}(0)
-{\boldsymbol{\varepsilon}}_h^{\widehat u}(0)),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h}
&={\langle {\boldsymbol{\delta}}(0),{\boldsymbol{w}}\rangle}_{{\partial}{\mathcal}T_h},\\
\label{eq:elwv_ereqt0c}
{\langle {\boldsymbol{\varepsilon}}_h^\sigma(0){\boldsymbol{n}}
-{\boldsymbol{\tau}}({{\boldsymbol{\varepsilon}}_h^u}(0)
-{\boldsymbol{\varepsilon}}_h^{\widehat u}(0)),{\boldsymbol{\mu}}\rangle}_{{\partial}{\mathcal}T_h\backslash\Gamma}
&= -{\langle {\boldsymbol{\delta}}(0),{\boldsymbol{\mu}}\rangle}_{{\partial}{\mathcal}T_h\backslash\Gamma},\\
\label{eq:elwv_ereqt0d}
{\langle {\boldsymbol{\varepsilon}}_h^{\widehat u}(0),{\boldsymbol{\mu}}\rangle}_\Gamma
&= 0,\end{aligned}$$
hold for all $({\boldsymbol{\theta}},{\boldsymbol{w}},{\boldsymbol{\mu}})\in {\boldsymbol{V}}_h\times {\boldsymbol{W}}_h \times{\boldsymbol{M}}_h$.
Use , , and Theorem \[th:PROJ\].
The next proposition gives estimates to the error terms when $t=0$.
\[prop:ene\_id\_t0\] The following estimates hold: $$\begin{aligned}
\label{eq:prp_enid_t0_a}
\|{\boldsymbol{\varepsilon}}_h^\sigma(0)\|_{{\mathcal}A}^2
+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(0)-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(0))\|_{{\boldsymbol{\tau}}}^2
&\le
\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}A}^2
+\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2,\\
\label{eq:prp_enid_t0_b}
\|\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(0)\|_{{\mathcal}A}^2
+\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(0)-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(0))\|_{{\boldsymbol{\tau}}}^2
&\le
\|\dot{{\boldsymbol{e}}}_\sigma(0)\|_{{\mathcal}A}^2
+\|\dot{{\boldsymbol{\delta}}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2,\\
\label{eq:prp_enid_t0_c}
\|\ddot{{\boldsymbol{\varepsilon}}}_h^u(0)\|_\rho
&\le \|\ddot{{\boldsymbol{e}}}_u(0)\|_\rho.\end{aligned}$$
Taking ${\boldsymbol{\theta}} = {\boldsymbol{\varepsilon}}_h^\sigma(0)$, ${\boldsymbol{w}}={\boldsymbol{\varepsilon}}_h^u(0)$ and ${\boldsymbol{\mu}} = {\boldsymbol{\varepsilon}}_h^{\widehat u}(0)$ in the error equations and adding the equations, we have $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^\sigma(0)\|_{{\mathcal}A}^2
+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(0)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(0))\|_{{\boldsymbol{\tau}}}^2
=({\mathcal}A{\boldsymbol{e}}_\sigma(0),{\boldsymbol{\varepsilon}}_h^\sigma(0))_{{\mathcal}T_h}
+{\langle {\boldsymbol{\delta}}(0),{\boldsymbol{\varepsilon}}_h^u(0)
- {\boldsymbol{\varepsilon}}_h^{\widehat u}(0)\rangle}_{{\partial}{\mathcal}T_h}.\end{aligned}$$ The first estimate then follows from the latter identity.
Consider the error equations . We take the first order derivative of - and test the equations with ${{\boldsymbol{\theta}}} = \dot{{\boldsymbol{\varepsilon}}}_h^\sigma$, ${\boldsymbol{w}}=\dot{{\boldsymbol{\varepsilon}}}_h^u$, ${\boldsymbol{\mu}} = \dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u}$. We next add the equations, evaluate them at $t=0$, and use the fact that $\dot{{\boldsymbol{\varepsilon}}}_h^u(0)=0$. Then $$\begin{aligned}
\|\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(0)\|_{{\mathcal}A}^2
+\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(0)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u}(0))\|_{{\boldsymbol{\tau}}}^2
=({\mathcal}A\dot{{\boldsymbol{e}}}_\sigma(0),\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(0))_{{\mathcal}T_h}
+{\langle \dot{{\boldsymbol{\delta}}}(0),\dot{{\boldsymbol{\varepsilon}}}_h^u (0)
- \dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u}(0)\rangle}_{{\partial}{\mathcal}T_h},\end{aligned}$$ from which the second estimate follows.
Finally, taking $t=0$ in and subtracting , then taking ${\boldsymbol{w}}=\ddot{{\boldsymbol{\varepsilon}}}_h^u(0)$, we have $$\begin{aligned}
(\ddot{{\boldsymbol{\varepsilon}}}_h^u(0),\ddot{{\boldsymbol{\varepsilon}}}_h^u(0))_\rho
=(\ddot{{\boldsymbol{e}}}_u(0),\ddot{{\boldsymbol{\varepsilon}}}_h^u(0))_\rho,\end{aligned}$$ which implies .
Note that the boundary remainder operator $\mathrm R$ defined in Theorem \[th:PROJ\] is linear, thus commuting with the time derivative $$\begin{aligned}
\dot{{\boldsymbol{\delta}}}(t) =
\mathrm R(\dot{{\boldsymbol{\sigma}}}(t),\dot{{\boldsymbol{u}}}(t);{\boldsymbol{\tau}}).\end{aligned}$$ This commutativity holds for the projection $\Pi$ as well for similar reasons.
\[prop:ene\_id\_tge0\] For $t\ge0$, we have $$\begin{aligned}
&\frac{1}{2}\frac{\mathrm d}{\mathrm dt}
\left(
\|{\boldsymbol{\varepsilon}}_h^\sigma(t)\|_{{\mathcal}A}^2
+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(t))\|_{{\boldsymbol{\tau}}}^2
+\|\dot{{\boldsymbol{\varepsilon}}}_h^u(t)\|_\rho^2
\right)\nonumber\\
\label{eq:prp_id_tge0_a}
&\qquad\qquad =(\rho\ddot{{\boldsymbol{e}}}_u(t),\dot{{\boldsymbol{\varepsilon}}}_h^u(t))
+({\mathcal}A\dot{{\boldsymbol{e}}}_\sigma(t),{\boldsymbol{\varepsilon}}_h^\sigma(t))
+{\langle {\boldsymbol{\delta}}(t),\dot{{\boldsymbol{\varepsilon}}}_h^u(t)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u}(t)\rangle},\\
&\frac{1}{2}\frac{\mathrm d}{\mathrm dt}
\left(
\|\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(t)\|_{{\mathcal}A}^2
+\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(t)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(t))\|_{{\boldsymbol{\tau}}}^2
+\|\ddot{{\boldsymbol{\varepsilon}}}_h^u(t)\|_\rho^2
\right)\nonumber\\
\label{eq:prp_id_tge0_b}
&\qquad\qquad=(\rho\dddot{{\boldsymbol{e}}}_u(t),\ddot{{\boldsymbol{\varepsilon}}}_h^u(t))
+({\mathcal}A\ddot{{\boldsymbol{e}}}_\sigma(t),\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(t))
+{\langle \dot{{\boldsymbol{\delta}}}(t),\ddot{{\boldsymbol{\varepsilon}}}_h^u(t)
-\ddot{{\boldsymbol{\varepsilon}}}_h^{\widehat u}(t)\rangle}.\end{aligned}$$
Taking the first order derivative of and testing it with ${\boldsymbol{\theta}}={\boldsymbol{\varepsilon}}_h^\sigma(t)$, then choosing ${\boldsymbol{w}}=\dot{{\boldsymbol{\varepsilon}}}_h^u(t)$ in and ${\boldsymbol{\mu}}=\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(t)$ in , and finally taking the first order derivative of then adding the equations, we obtain .
Taking the second order derivative of and testing it with ${\boldsymbol{\theta}}=
\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(t)$, then taking the first order derivative of and testing it with ${\boldsymbol{w}}=\ddot{{\boldsymbol{\varepsilon}}}_h^u(t)$, taking the first order derivative of and testing it with ${\boldsymbol{\mu}} = \ddot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(t)$, and finally taking the second order derivative of and adding the equations, we obtain .
The final ingredient we need is a Grönwall type inequality.
\[lm:int\_ineq\] Suppose $\phi,\beta,l$ are continuous and positive functions defined on $[0,\infty)$ and $r$ is a constant. If $$\begin{aligned}
\phi^2(t)\le r + 2\int_0^t\phi(s)\beta(s)\mathrm d s
+\phi(t)l(t),\end{aligned}$$ then $$\begin{aligned}
\phi^2(t)\le 2 r
+\left(
2\int_0^t\beta(s)\mathrm ds
+\sup_{s\in[0,t]}l(s)
\right)^2.\end{aligned}$$
Consider the interval $[0,t]$ and let $\phi(t^*)$ maximize $\phi$ in the interval. Then we have $$\begin{aligned}
\phi^2(t^*)
&\le r+\phi(t^*)\left(2\int_0^t\beta(s)\mathrm ds + \sup_{s\in[0,t]}l(s)\right)\\
&\le r + \frac{1}{2}
\left(
\phi^2(t^*)+\left(2\int_0^t\beta(s)\mathrm ds + \sup_{s\in[0,t]}l(s)\right)^2
\right).\end{aligned}$$
With Proposition \[prop:ene\_id\_t0\], Proposition \[prop:ene\_id\_tge0\] and Lemma \[lm:int\_ineq\], we are ready to prove Theorem \[th:elwv\_th1\].
Integrating from $0$ to $T$, we have $$\begin{aligned}
\frac{1}{2}
\bigg(&
\|{\boldsymbol{\varepsilon}}_h^\sigma(t)\|_{{\mathcal}A}^2
+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(t))\|_{{\boldsymbol{\tau}}}^2
+\|\dot{{\boldsymbol{\varepsilon}}}_h^u(t)\|_\rho^2
\bigg)\bigg|_{t=0}^{t=T}\\
=&
{\langle {\boldsymbol{\delta}}(t),\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(t))\rangle}_{{\partial}{\mathcal}T_h}\,\bigg|_{t=0}^{t=T}\\
&+\int_0^T \left(
\left(\rho\ddot{{\boldsymbol{e}}}_u(t),\dot{{\boldsymbol{\varepsilon}}}_h^u(t)\right)_{{\mathcal}T_h}
+\left({\mathcal}A\dot{{\boldsymbol{e}}}_\sigma(t),{\boldsymbol{\varepsilon}}_h^\sigma(t)\right)_{{\mathcal}T_h}
-{\langle \dot{{\boldsymbol{\delta}}}(t),\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(t)-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(t))\rangle}_{{\partial}{\mathcal}T_h}
\right)\mathrm dt\\
\le &
\|{\boldsymbol{\delta}}(T)\|_{{\boldsymbol{\tau}}^{-1}}
\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(T)
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(T))\|_{{\boldsymbol{\tau}}}
+\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}
\left(\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}A}^2
+\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2\right)^{1/2}\\
& +
\int_0^T
\left(\|\dot{{\boldsymbol{\varepsilon}}}_h^u(t)\|_\rho^2
+\|{\boldsymbol{\varepsilon}}_h^\sigma(t)\|_{{\mathcal}A}^2
+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(t))\|_{{\boldsymbol{\tau}}}^2
\right)^{1/2}\\
& \qquad \left(
\|\ddot{{\boldsymbol{e}}}_u(t)\|_\rho^2+\|\dot{{\boldsymbol{e}}}_\sigma(t)\|_{{\mathcal}A}^2
+\|\dot{{\boldsymbol{\delta}}}(t)\|_{{\boldsymbol{\tau}}^{-1}}^2
\right)^{1/2}\mathrm dt,\end{aligned}$$ where we used to estimate $\|{\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u(0)
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(0))\|_{{\boldsymbol{\tau}}}$ in the last step.
Now we define $$\begin{aligned}
\phi^2(t)&:=\|\dot{{\boldsymbol{\varepsilon}}}_h^u(t)\|_\rho^2+
\|{\boldsymbol{\varepsilon}}_h^\sigma(t)\|_{{\mathcal}A}^2
+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(t))\|_{{\boldsymbol{\tau}}}^2,\\
\beta^2(t)&:=
\|\ddot{{\boldsymbol{e}}}_u(t)\|_\rho^2+\|\dot{{\boldsymbol{e}}}_\sigma(t)\|_{{\mathcal}A}^2
+\|\dot{{\boldsymbol{\delta}}}(t)\|_{{\boldsymbol{\tau}}^{-1}}^2,\\
l(t)&:=2\|{\boldsymbol{\delta}}(t)\|_{{\boldsymbol{\tau}}^{-1}},\\
r&:=2\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}A}^2+3\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2.\end{aligned}$$ Note that by and , we have $
\phi^2(0)
\le \|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}A}^2
+\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2,
$ and therefore, $$\begin{aligned}
\phi^2(T)
\le r+ 2 \int_0^T \phi(t)\beta(t){\mathrm}{d}t +\phi(T)l(T).\end{aligned}$$ By Lemma \[lm:int\_ineq\], we have $$\begin{aligned}
\phi^2(T)&\le 2 r
+\left(
2\int_0^T\beta(t)\mathrm dt
+\sup_{t\in[0,T]}l(t)
\right)^2
\lesssim r +\sup_{t\in[0,T]}l^2(t)+ \left(\int_0^T\beta(t){\mathrm}{d}t\right)^2\\
&\lesssim \|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2 + \|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}A}^2
+\left(\vvvert{\boldsymbol{\delta}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}\right)^2\\
&\quad\,+\left(\int_0^T\left(
\|\ddot{{\boldsymbol{e}}}_u(t)\|_\rho+\|\dot{{\boldsymbol{e}}}_\sigma(t)\|_{{\mathcal}A}
+\|\dot{{\boldsymbol{\delta}}}(t)\|_{{\boldsymbol{\tau}}^{-1}}\right)
\mathrm{d}t\right)^2\\
&\lesssim \left(\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}A}
+\vvvert{\boldsymbol{\delta}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}
+\vvvert\ddot{{\boldsymbol{e}}}_u\vvvert_{\rho,1}^{[0,T]}
+ \vvvert\dot{{\boldsymbol{e}}}_\sigma\vvvert_{{\mathcal}A,1}^{[0,T]}+\vvvert\dot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},1}^{[0,T]}\right)^2.\end{aligned}$$
Similarly, for the second estimate, we integrate from $0$ to $T$, then use to estimate $\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(0)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(0))\|_{{\boldsymbol{\tau}}}$, and obtain $$\begin{aligned}
&\frac{1}{2}
\left(
\|\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(t)\|_{{\mathcal}A}^2
+\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(t)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(t))\|_{{\boldsymbol{\tau}}}^2
+\|\ddot{{\boldsymbol{\varepsilon}}}_h^u(t)\|_\rho^2
\right)\bigg|_{t=0}^{t=T}\\
&\le
\|\dot{{\boldsymbol{\delta}}}(T)\|_{{\boldsymbol{\tau}}^{-1}}
\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(T)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(T))\|_{{\boldsymbol{\tau}}}
+\|\dot{{\boldsymbol{\delta}}}(0)\|_{{\boldsymbol{\tau}}^{-1}}
\left(\|\dot{{\boldsymbol{e}}}_\sigma(0)\|_{{\mathcal}A}^2
+\|\dot{{\boldsymbol{\delta}}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2\right)^{1/2}\\
&\quad\, +
\int_0^T\hspace{-4pt}
\left(\|\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(t)\|_{{\mathcal}A}^2
\!+\!\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(t)
\!-\!\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(t))\|_{{\boldsymbol{\tau}}}^2
\!+\!\|\ddot{{\boldsymbol{\varepsilon}}}_h^u(t)\|_\rho^2\right)^{\frac12}
\left(
\|\dddot{{\boldsymbol{e}}}_u(t)\|_\rho^2\!+\!\|\ddot{{\boldsymbol{e}}}_\sigma(t)\|_{{\mathcal}A}^2
\!+\!\|\ddot{{\boldsymbol{\delta}}}(t)\|_{{\boldsymbol{\tau}}^{-1}}^2
\right)^{\frac12}\hspace{-4pt}\mathrm dt.\end{aligned}$$
Now we define $$\begin{aligned}
\phi^2(t)&:=
\|\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(t)\|_{{\mathcal}A}^2
+\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(t)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}(t))\|_{{\boldsymbol{\tau}}}^2
+\|\ddot{{\boldsymbol{\varepsilon}}}_h^u(t)\|_\rho^2,\\
\beta^2(t)&:=
\|\dddot{{\boldsymbol{e}}}_u(t)\|_\rho^2+\|\ddot{{\boldsymbol{e}}}_\sigma(t)\|_{{\mathcal}A}^2
+\|\ddot{{\boldsymbol{\delta}}}(t)\|_{{\boldsymbol{\tau}}^{-1}}^2,\\
l(t)&:=2\|\dot{{\boldsymbol{\delta}}}(t)\|_{{\boldsymbol{\tau}}^{-1}},\\
r&:=2\|\dot{{\boldsymbol{e}}}_\sigma(0)\|_{{\mathcal}A}^2
+3\|\dot{{\boldsymbol{\delta}}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2
+\|\ddot{{\boldsymbol{e}}}_u(0)\|_{\rho}^2.\end{aligned}$$ Since by and we have $
\phi^2(0)
\le
\|\dot{{\boldsymbol{e}}}_\sigma(0)\|_{{\mathcal}A}^2
+\|\dot{{\boldsymbol{\delta}}}(0)\|_{{\boldsymbol{\tau}}^{-1}}^2
+\|\ddot{{\boldsymbol{e}}}_u(0)\|_\rho^2,
$ and it follows that $$\begin{aligned}
\phi^2(T)
\le r+ 2 \int_0^T \phi(t)\beta(t){\mathrm}{d}t +\phi(T)l(T).\end{aligned}$$ Using Lemma \[lm:int\_ineq\] again we obtain the second estimate.
Duality argument
----------------
In this subsection, we give a proof to Theorem \[th:elwv\_spconv\] by using the duality argument. To begin with, we consider the adjoint equations of :
\[eq:elwv\_adpde\] $$\begin{aligned}
{5}\label{eq:elwv_adpdea}
{\mathcal}A{\boldsymbol{\Psi}} + {\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}}) &= 0&\qquad& {\mathrm}{in}\ \Omega\times[0,T],\\
\label{eq:elwv_adpdeb}
\rho\ddot{{\boldsymbol{\Phi}}}+{\mathrm{div}\hspace{1pt}}{\boldsymbol{\Psi}} &= 0 && {\mathrm}{in}\ \Omega\times[0,T],\\
\label{eq:elwv_adpdec}
\gamma{\boldsymbol{\Phi}} &= 0 && {\mathrm}{on}\ \Gamma\times[0,T],\\
\label{eq:elwv_adpded}
{\boldsymbol{\Phi}}(T) &= 0 && {\mathrm}{on}\ \Omega,\\
\label{eq:elwv_adpdee}
\dot{{\boldsymbol{\Phi}}}(T) &= {\boldsymbol{\varepsilon}}_h^u(T) &&{\mathrm}{on}\ \Omega.\end{aligned}$$
For a time-dependent function $f:[0,\infty)\rightarrow X$, we write $$\begin{aligned}
\underline{f}(t):=\int_t^T f(s){\mathrm}{d}s.\end{aligned}$$ The following proposition allows us to control the solution of in certain energy norms. Similar results can be found in [@CoQu:2014].
\[prop:elwv\_reg\] The following inequality holds $$\begin{aligned}
\label{eq:elwv_reg}
\vvvert{\boldsymbol{\Phi}}\vvvert_{H^1(\Omega),\infty}^{[0,T]}+\vvvert\dot{{\boldsymbol{\Phi}}}\vvvert_{\Omega,\infty}^{[0,T]}
\le C
\|{\boldsymbol{\varepsilon}}_h^u(T)\|_\Omega.\end{aligned}$$ If holds, then $$\begin{aligned}
\label{eq:elwv_mrreg}
\vvvert\underline{{\boldsymbol{\Psi}}}\vvvert_{H^1(\Omega),\infty}^{[0,T]}
+\vvvert\underline{{\boldsymbol{\Phi}}}\vvvert_{H^2(\Omega),\infty}^{[0,T]}
\le C\|{\boldsymbol{\varepsilon}}_h^u(T)\|_\Omega.\end{aligned}$$
By conservation of energy we have $$\begin{aligned}
\|\dot{{\boldsymbol{\Phi}}}(t)\|_\rho^2 + \|{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}})(t)\|_{{\mathcal}A^{-1}}^2 = \|{\boldsymbol{\varepsilon}}_h^u(T)\|_\rho^2,\end{aligned}$$ for all $t\in[0,T]$. Now follows by using Korn’s second inequality.
Integrating and from $t$ to $T$, we have $$\begin{aligned}
{\mathcal}A\underline{{\boldsymbol{\Psi}}}(t) + {\boldsymbol{\varepsilon}}(\underline{{\boldsymbol{\Phi}}}(t))=0,\qquad
{\mathrm{div}\hspace{1pt}}\underline{{\boldsymbol{\Psi}}}(t)=\rho\dot{{\boldsymbol{\Phi}}}(t)-\rho\dot{{\boldsymbol{\Phi}}}(T),\end{aligned}$$ for all $t\in[0,T]$. Combining the latter equations with and , we obtain .
Now we define the dual projections and boundary remainder terms for the adjoint problem (for all $t\in[0,T]$) $$\begin{aligned}
(\Pi{\boldsymbol{\Psi}}(t),\Pi{\boldsymbol{\Phi}}(t))
&:= \prod_{K\in{\mathcal}T_h}\Pi({\boldsymbol{\Psi}}(t)\big|_K,{\boldsymbol{\Phi}}(t)\big|_K;-{\boldsymbol{\tau}}\big|_{{\partial}K}),\\
{\boldsymbol{\Delta}}(t)&:=\prod_{K\in{\mathcal}T_h}\mathrm R({\boldsymbol{\Psi}}(t)\big|_K,{\boldsymbol{\Phi}}(t)\big|_K;-{\boldsymbol{\tau}}\big|_{{\partial}K}).\end{aligned}$$ We also define the corresponding approximation terms $$\begin{aligned}
{\boldsymbol{e}}_\Phi(t) := \Pi{\boldsymbol{\Phi}}(t)-{\boldsymbol{\Phi}}(t),\qquad
{\boldsymbol{e}}_\Psi(t) := \Pi{\boldsymbol{\Psi}}(t) - {\boldsymbol{\Psi}}(t).\end{aligned}$$
\[prop:wav\_dual\_id\] The following identity holds $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{\rho}^2 =
\sum_{i=1}^7 T_i,\end{aligned}$$ where $$\begin{aligned}
{5}
T_1&:=(\rho\dot{{\boldsymbol{\Phi}}}(0),\Pi{\boldsymbol{u}}_0-{\boldsymbol{u}}_h(0))_{{\mathcal}T_h}, &\qquad\quad
T_2&:=-\int_0^T{\langle {\boldsymbol{\delta}}(t),{\boldsymbol{e}}_\Phi(t)\rangle}_{{\partial}{\mathcal}T_h}\mathrm dt,\\
T_3&:=-\int_0^T{\langle {\boldsymbol{\Delta}}(t),{\boldsymbol{\varepsilon}}_h^u(t)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)\rangle}_{{\partial}{\mathcal}T_h}\mathrm dt, &
T_4&:=\int_0^T({\mathcal}A{\boldsymbol{e}}_\Psi(t),{\boldsymbol{\sigma}}(t)-{\boldsymbol{\sigma}}_h(t))_{{\mathcal}T_h}\mathrm dt,\\
T_5&:=-\int_0^T({\mathcal}A{\boldsymbol{e}}_\sigma(t),{\boldsymbol{\Psi}}(t))_{{\mathcal}T_h}\mathrm dt,&
T_6&:=-\int_0^T(\rho{\boldsymbol{\Phi}}(t),\ddot{{\boldsymbol{e}}}_u(t))_{{\mathcal}T_h}\mathrm dt,\\
T_7&:=\int_0^T(\rho(\ddot{{\boldsymbol{u}}}(t)-\ddot{{\boldsymbol{u}}}_h(t)),{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}\mathrm dt.\end{aligned}$$
From the adjoint equations tested with $({\boldsymbol{\varepsilon}}_h^\sigma(t),{\boldsymbol{\varepsilon}}_h^u(t),{\boldsymbol{\varepsilon}}_h^{\widehat u}(t))$, and from the properties of the adjoint projection (see Theorem \[th:PROJ\]), we obtain $$\begin{aligned}
({\mathcal}A\Pi{\boldsymbol{\Psi}}(t),{\boldsymbol{\varepsilon}}_h^\sigma(t))_{{\mathcal}T_h}
- (\Pi{\boldsymbol{\Phi}}(t),{\mathrm{div}\hspace{1pt}}{\boldsymbol{\varepsilon}}_h^\sigma(t))_{{\mathcal}T_h}
+ {\langle \mathrm P_M{\boldsymbol{\Phi}}(t),{\boldsymbol{\varepsilon}}_h^\sigma(t){\boldsymbol{n}}\rangle}_{{\partial}{\mathcal}T_h}&
=({\mathcal}A{\boldsymbol{e}}_\Psi(t),{\boldsymbol{\varepsilon}}_h^\sigma(t))_{{\mathcal}T_h},\\
({\mathrm{div}\hspace{1pt}}\Pi{\boldsymbol{\Psi}}(t),{\boldsymbol{\varepsilon}}_h^u(t))_{{\mathcal}T_h}
+{\langle {\boldsymbol{\tau}}\mathrm P_M(\Pi{\boldsymbol{\Phi}}(t)-{\boldsymbol{\Phi}}(t)),{\boldsymbol{\varepsilon}}_h^u(t)\rangle}_{{\partial}{\mathcal}T_h} &=
-{\langle {\boldsymbol{\Delta}}(t),{\boldsymbol{\varepsilon}}_h^u(t)\rangle}_{{\partial}{\mathcal}T_h}\\
& \phantom{=}-(\rho\ddot{{\boldsymbol{\Phi}}}(t),{\boldsymbol{\varepsilon}}_h^u(t))_{{\mathcal}T_h},\\
-{\langle \Pi{\boldsymbol{\Psi}}(t){\boldsymbol{n}}+{\boldsymbol{\tau}}(\Pi{\boldsymbol{\Phi}}(t)-{\boldsymbol{\Phi}}(t)),{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)\rangle}
_{{\partial}{\mathcal}T_h\backslash\Gamma}&=
{\langle {\boldsymbol{\Delta}},{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)\rangle}_{{\partial}{\mathcal}T_h\backslash\Gamma},\\
{\langle \mathrm P_M{\boldsymbol{\Phi}}(t),
{\boldsymbol{\varepsilon}}_h^\sigma(t){\boldsymbol{n}}
-{\boldsymbol{\tau}}{\mathrm}P_M({\boldsymbol{\varepsilon}}_h^u(t)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t))
\rangle}_\Gamma&=0.\end{aligned}$$ (Note that ${\boldsymbol{\Phi}}\equiv 0$ on $\Gamma$.) Taking now ${\boldsymbol{\theta}} = \Pi{\boldsymbol{\Psi}}(t)$, ${\boldsymbol{w}}=\Pi{\boldsymbol{\Phi}}(t)$, ${\boldsymbol{\mu}}=\mathrm P_M{\boldsymbol{\Phi}}(t)$ and ${\boldsymbol{\mu}}=\Pi{\boldsymbol{\Psi}}(t){\boldsymbol{n}}+{\boldsymbol{\tau}}{\mathrm}P_M(\Pi{\boldsymbol{\Phi}}(t)-{\boldsymbol{\Phi}}(t))$ in the error equations to , and comparing the two sets of equations, we have $$\begin{aligned}
&-{\langle {\boldsymbol{\Delta}}(t),{\boldsymbol{\varepsilon}}_h^u(t)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)\rangle}_{{\partial}{\mathcal}T_h}
+({\mathcal}A{\boldsymbol{e}}_\Psi(t),{\boldsymbol{\varepsilon}}_h^\sigma(t)) _{{\mathcal}T_h}
- (\rho\ddot{{\boldsymbol{\Phi}}}(t),{\boldsymbol{\varepsilon}}_h^u(t))_{{\mathcal}T_h}\\
&\qquad\qquad=({\mathcal}A{\boldsymbol{e}}_\sigma(t),\Pi{\boldsymbol{\Psi}}(t))_{{\mathcal}T_h}
+{\langle {\boldsymbol{\delta}}(t),\Pi{\boldsymbol{\Phi}}(t)-{\boldsymbol{\Phi}}(t)\rangle}_{{\partial}{\mathcal}T_h}
+(\rho(\ddot{{\boldsymbol{u}}}_h(t)-\ddot{{\boldsymbol{u}}}(t)),\Pi{\boldsymbol{\Phi}}(t))_{{\mathcal}T_h}.\end{aligned}$$ After rearranging terms, we have $$\begin{aligned}
\label{eq:7.200}
\nonumber
(\rho\ddot{{\boldsymbol{\Phi}}}(t),{\boldsymbol{\varepsilon}}_h^u(t))_{{\mathcal}T_h}=
&-{\langle {\boldsymbol{\Delta}}(t),{\boldsymbol{\varepsilon}}_h^u(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)\rangle}_{{\partial}{\mathcal}T_h}
-{\langle {\boldsymbol{\delta}}(t),\Pi{\boldsymbol{\Phi}}(t)-{\boldsymbol{\Phi}}(t)\rangle}_{{\partial}{\mathcal}T_h}\\
&-(\rho(\ddot{{\boldsymbol{u}}}_h(t)-\ddot{{\boldsymbol{u}}}(t)),\Pi{\boldsymbol{\Phi}})_{{\mathcal}T_h}
+({\mathcal}A{\boldsymbol{e}}_\Psi(t),{\boldsymbol{\sigma}}(t)-{\boldsymbol{\sigma}}_h(t))_{{\mathcal}T_h}\\
\nonumber
&-({\mathcal}A{\boldsymbol{e}}_\sigma(t),{\boldsymbol{\Psi}}(t))_{{\mathcal}T_h}.\end{aligned}$$
Defining now $$\begin{aligned}
\eta(t):=
(\rho\dot{{\boldsymbol{\Phi}}}(t),{\boldsymbol{\varepsilon}}_h^u(t))_{{\mathcal}T_h}
-(\rho{\boldsymbol{\Phi}}(t),\dot{{\boldsymbol{\varepsilon}}}_h^u(t))_{{\mathcal}T_h},\end{aligned}$$ which satisfies $$\begin{aligned}
\eta(T) = (\rho{\boldsymbol{\varepsilon}}_h^u(T),{\boldsymbol{\varepsilon}}_h^u(T))_{{\mathcal}T_h},\quad
\eta(0) = (\rho\dot{{\boldsymbol{\Phi}}}(0),\Pi{\boldsymbol{u}}_0-{\boldsymbol{u}}_h(0))_{{\mathcal}T_h},\end{aligned}$$ due to ${\boldsymbol{\Phi}}(T)=0$, $\dot{{\boldsymbol{\Phi}}}(T)={\boldsymbol{\varepsilon}}_h^u(T)$, ${\boldsymbol{\varepsilon}}_h^u(0)=\Pi{\boldsymbol{u}}_0-{\boldsymbol{u}}_h(0)$ and $\dot{{\boldsymbol{\varepsilon}}}_h^u(0)=0$ (see and ). By , we have $$\begin{aligned}
\dot{\eta}(t) =& (\rho\ddot{{\boldsymbol{\Phi}}}(t),{\boldsymbol{\varepsilon}}_h^u(t))_{{\mathcal}T_h}
-(\rho{\boldsymbol{\Phi}}(t),\ddot{{\boldsymbol{\varepsilon}}}_h^u(t))_{{\mathcal}T_h}\\
=&
-{\langle {\boldsymbol{\Delta}}(t),{\boldsymbol{\varepsilon}}_h^u(t)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)\rangle}_{{\partial}{\mathcal}T_h}
-{\langle {\boldsymbol{\delta}}(t),{\boldsymbol{e}}_\Phi(t)\rangle}_{{\partial}{\mathcal}T_h}\\
&+({\mathcal}A{\boldsymbol{e}}_\Psi(t),{\boldsymbol{\sigma}}(t)-{\boldsymbol{\sigma}}_h(t))_{{\mathcal}T_h}
-({\mathcal}A{\boldsymbol{e}}_\sigma(t),{\boldsymbol{\Psi}}(t))_{{\mathcal}T_h}\\
&+(\rho(\ddot{{\boldsymbol{u}}}(t)-\ddot{{\boldsymbol{u}}}_h(t)),{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}
-(\rho{\boldsymbol{\Phi}}(t),\ddot{{\boldsymbol{e}}}_u(t))_{{\mathcal}T_h}.\end{aligned}$$ Therefore $$\begin{aligned}
\|\rho^{1/2}{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}^2 = \int_0^T\dot{\eta}(t)\mathrm dt+
(\rho\dot{{\boldsymbol{\Phi}}}(0),\Pi{\boldsymbol{u}}_0-{\boldsymbol{u}}_h(0))_{{\mathcal}T_h},\end{aligned}$$ and the proof is completed.
The next proposition gives an estimate for the term $T_7$ in Proposition \[prop:wav\_dual\_id\].
\[prop:elwv\_tricky\] If holds, then $$\begin{aligned}
\left|T_7\right|
&\lesssim h
\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}
\bigg(
\|{\boldsymbol{\varepsilon}}_h^\sigma(0)\|_{{\mathcal}T_h}
+T\vvvert\dot{{\boldsymbol{\varepsilon}}}_h^{\sigma}\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}\\
&\quad\,
+\|{\mathrm}P_M{\boldsymbol{\varepsilon}}_h^u(0)-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(0)\|_{{\boldsymbol{\tau}}}
+T\vvvert{\mathrm}P_M\dot{{\boldsymbol{\varepsilon}}}_h^u-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}\vvvert_{{\boldsymbol{\tau}},\infty}^{[0,T]}\\
&\quad\,
+h\|\ddot{{\boldsymbol{e}}}_u(0)\|_{{\mathcal}T_h}
+hT\vvvert\dddot{{\boldsymbol{e}}}_u\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}
+\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}+T\vvvert\dot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}
\bigg).\end{aligned}$$
In the coming arguments, for the sake of shortening some estimates, we will prove bounds in terms of the quantity $$\begin{aligned}
\label{eq:def_theta}
{\boldsymbol{\Theta}}(T):=\sup_{t\in[0,T]}|\underline{{\boldsymbol{\Phi}}}(t)|_{2,\Omega}
+\sup_{t\in[0,T]}|\underline{{\boldsymbol{\Psi}}}(t)|_{1,\Omega},\end{aligned}$$ which we have shown in Proposition \[prop:elwv\_reg\] that, assuming , we have the estimate $$\begin{aligned}
\label{eq:thtabnd}
{\boldsymbol{\Theta}}(T)\lesssim \|{\boldsymbol{\varepsilon}}_h^u(T)\|_{\Omega}.\end{aligned}$$ Note that $$\begin{aligned}
\label{eq:elwv_trick_decom}
T_7
=\int_0^T(\rho\ddot{{\boldsymbol{\varepsilon}}}_h^u(t),{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}\mathrm dt - \int_0^T (\rho\ddot{{\boldsymbol{e}}}_u(t),{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}\mathrm dt.\end{aligned}$$ For the second term of , we have $$\begin{aligned}
\left|\int_0^T (\rho \ddot{{\boldsymbol{e}}}_u(t),{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}\mathrm dt\right|
&=\left|(\rho\ddot{{\boldsymbol{e}}}_u(0),{\boldsymbol{e}}_{\underline{\Phi}}(0))_{{\mathcal}T_h}
+\int_0^T (\rho\dddot{{\boldsymbol{e}}}_u(t), {\boldsymbol{e}}_{\underline{\Phi}}(t))_{{\mathcal}T_h}\mathrm dt\right|\\
&\lesssim h^2\left(\|\ddot{{\boldsymbol{e}}}_u(0)\|_{{\mathcal}T_h}+T\vvvert\dddot{{\boldsymbol{e}}}_u\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}\right){\boldsymbol{\Theta}}(T).\end{aligned}$$ We next estimate the remaining term in . Since $\ddot{{\boldsymbol{\varepsilon}}}_h^u(t)\big|_K\in{\mathcal}P_{k+1}(K;{\mathbb}R^3)$ for all $K$, we have $$\begin{aligned}
\int_0^T(\rho\ddot{{\boldsymbol{\varepsilon}}}_h^u(t),{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}\mathrm dt
=\int_0^T(\rho\ddot{{\boldsymbol{\varepsilon}}}_h^u(t),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}\mathrm dt,\end{aligned}$$ where, ${\mathrm}P_{k+1}^\rho$ is the $\rho$-weighted $L^2$ projection onto ${\mathcal}P_{k+1}(K;{\mathbb}R^3)$. Testing the second error equation with ${\boldsymbol{w}}={\mathrm}P_{k+1}^\rho {\boldsymbol{e}}_\Phi(t)$ we have $$\begin{aligned}
(\rho\ddot{{\boldsymbol{\varepsilon}}}_h^u(t),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}
&= ({\mathrm{div}\hspace{1pt}}{\boldsymbol{\varepsilon}}_h^\sigma(t),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}
-{\langle {\boldsymbol{\tau}}\mathrm P_M({{\boldsymbol{\varepsilon}}_h^u}(t)
-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_\Phi(t)\rangle}_{{\partial}{\mathcal}T_h}\\
&\quad\,
+(\rho\ddot{{\boldsymbol{e}}}_u(t),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_\Phi(t))_{{\mathcal}T_h}\nonumber+{\langle {\boldsymbol{\delta}}(t),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_\Phi(t)\rangle}_{{\partial}{\mathcal}T_h}\\
&=: Q_1(t)+Q_2(t)+Q_3(t)+Q_4(t).\end{aligned}$$ Now we use integration by part and obtain $$\begin{aligned}
\int_0^T Q_1(t)\mathrm dt
&= ({\mathrm{div}\hspace{1pt}}{{\boldsymbol{\varepsilon}}}_h^\sigma(0),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(0))_{{\mathcal}T_h}
+ \int_0^T ({\mathrm{div}\hspace{1pt}}\dot{{\boldsymbol{\varepsilon}}}_h^\sigma(t),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(t))_{{\mathcal}T_h}\mathrm dt,\\
\int_0^T Q_2(t)\mathrm dt
&=-{\langle {\boldsymbol{\tau}}\mathrm P_M({{\boldsymbol{\varepsilon}}_h^u}(0)
-{\boldsymbol{\varepsilon}}_h^{\widehat u}(0)),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(0)\rangle}_{{\partial}{\mathcal}T_h}\\
&\quad\, -\int_0^T {\langle {\boldsymbol{\tau}}\mathrm P_M({\dot{{\boldsymbol{\varepsilon}}}_h^u}(t)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u}(t)),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(t)\rangle}_{{\partial}{\mathcal}T_h}\mathrm dt,\\
\int_0^T Q_3(t)\mathrm dt
&=(\rho\ddot{{\boldsymbol{e}}}_u(0),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(0))_{{\mathcal}T_h}
+\int_0^T(\rho\dddot{{\boldsymbol{e}}}_u(t),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(t))_{{\mathcal}T_h}\mathrm dt,\\
\int_0^T Q_4(t)\mathrm dt
&={\langle {\boldsymbol{\delta}}(0),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(0)\rangle}_{{\partial}{\mathcal}T_h}
+\int_0^T{\langle \dot{{\boldsymbol{\delta}}}(t),{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(t)\rangle}_{{\partial}{\mathcal}T_h}\mathrm dt.\end{aligned}$$ Note that $$\begin{aligned}
{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(t) = {\mathrm}P_{k+1}^\rho(\Pi\underline{{\boldsymbol{\Phi}}}(t)-\underline{{\boldsymbol{\Phi}}}(t))
= \Pi\underline{{\boldsymbol{\Phi}}}(t) - \underline{{\boldsymbol{\Phi}}}(t) - {\mathrm}P_{k+1}^\rho\underline{{\boldsymbol{\Phi}}}(t) + \underline{{\boldsymbol{\Phi}}}(t).\end{aligned}$$ Combining the above with the convergence properties about $\Pi{\boldsymbol{\Psi}}(t)$ and $\Pi{\boldsymbol{\Phi}}(t)$ (see Theorem \[th:PROJ\]) we have $$\begin{aligned}
\|{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(t)\|_{K}
&\lesssim h_K^2 (|\underline{{\boldsymbol{\Phi}}}(t)|_{2,K} + |\underline{{\boldsymbol{\Psi}}}(t)|_{1,K}),\\
\|{\boldsymbol{\tau}}^{1/2}{\mathrm}P_{k+1}^\rho{\boldsymbol{e}}_{\underline{\Phi}}(t)\|_{{\partial}K}
&\lesssim h_K (|\underline{{\boldsymbol{\Phi}}}(t)|_{2,K} + |\underline{{\boldsymbol{\Psi}}}(t)|_{1,K}).\end{aligned}$$ Now back to the estimate of $\int_0^TQ_i(t)\mathrm dt$, we have $$\begin{aligned}
\left|\int_0^TQ_1(t)\mathrm dt\right|&\lesssim
h(\|{\boldsymbol{\varepsilon}}_h^\sigma(0)\|_{{\mathcal}T_h}+T\vvvert\dot{{\boldsymbol{\varepsilon}}}_h^{\sigma}\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}){\boldsymbol{\Theta}}(T),\\
\left|\int_0^T Q_2(t)\mathrm dt\right|&\lesssim
h(\|{\mathrm}P_M{\boldsymbol{\varepsilon}}_h^u(0)-{\boldsymbol{\varepsilon}}_h^{\widehat{u}}(0)\|_{{\boldsymbol{\tau}}}
+T\vvvert{\mathrm}P_M\dot{{\boldsymbol{\varepsilon}}}_h^u-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat{u}}\vvvert_{{\boldsymbol{\tau}},\infty}^{[0,T]})
{\boldsymbol{\Theta}}(T),\\
\left|\int_0^T Q_3(t)\mathrm dt\right| &\lesssim h^2 (\|\ddot{{\boldsymbol{e}}}_u(0)\|_{{\mathcal}T_h}+T\vvvert\dddot{{\boldsymbol{e}}}_u\vvvert_{{\mathcal}T_h,\infty}^{[0,T]})
{\boldsymbol{\Theta}}(T),\\
\left|\int_0^T Q_4(t)\mathrm dt\right|
&\lesssim h (\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}+T\vvvert\dot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]})
{\boldsymbol{\Theta}}(T),\end{aligned}$$ where we used the fact that $\|{\mathrm{div}\hspace{1pt}}{\boldsymbol{\sigma}}\|_K\lesssim h_K^{-1}\|{\boldsymbol{\sigma}}\|_K$ for any ${\boldsymbol{\sigma}}\in{\mathcal}P_k(K;{{\mathbb R^{3\times 3}_{\mathrm{sym}}}})$. Finally, we use to bound ${\boldsymbol{\Theta}}(T)$ and the proof is completed.
Now we are ready to prove Theorem \[th:elwv\_spconv\].
Consider Proposition \[prop:wav\_dual\_id\]. We will give estimates for the terms $T_i$ for $i=1\rightarrow 7$.
For $T_1$, by we have $$\begin{aligned}
\left| T_1\right|=\left|(\rho\dot{{\boldsymbol{\Phi}}}(0),\Pi{\boldsymbol{u}}_0-{\boldsymbol{u}}_h(0))_{{\mathcal}T_h}\right|
\lesssim \|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}(\|\Pi{\boldsymbol{u}}_0-{\boldsymbol{u}}_0\|_{{\mathcal}T_h}+\|{\boldsymbol{u}}_0 - {\boldsymbol{u}}_h(0)\|_{{\mathcal}T_h}).\end{aligned}$$ Note that ${\boldsymbol{u}}_h(0)$ is the solution of the HDG+ scheme . By Theorem \[th:st\] in Section \[sec:st\_main\], we have $$\begin{aligned}
\|{\boldsymbol{u}}_h(0) - {\boldsymbol{u}}_0\|_{{\mathcal}T_h}\lesssim
h(\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}T_h}+\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}).\end{aligned}$$ Therefore $$\begin{aligned}
\left|(\rho\dot{{\boldsymbol{\Phi}}}(0),\Pi{\boldsymbol{u}}_0-{\boldsymbol{u}}_h(0))_{{\mathcal}T_h}\right|
\lesssim \|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}(\|{\boldsymbol{e}}_{u}(0)\|_{{\mathcal}T_h}+h\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}T_h}
+h\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}).\end{aligned}$$
For $T_2$, we have $$\begin{aligned}
\left| T_2\right|&=\left|\int_0^T{\langle {\boldsymbol{\delta}}(t),\Pi{\boldsymbol{\Phi}}(t)-{\boldsymbol{\Phi}}(t)\rangle}_{{\partial}{\mathcal}T_h}\mathrm{d}t\right|\\
&=\left|{\langle {\boldsymbol{\delta}}(0),\Pi\underline{{\boldsymbol{\Phi}}}(0)
-\underline{{\boldsymbol{\Phi}}}(0)\rangle}_{{\partial}{\mathcal}T_h}
+\int_0^T
{\langle \dot{{\boldsymbol{\delta}}}(t),\Pi\underline{{\boldsymbol{\Phi}}}(t)
-\underline{{\boldsymbol{\Phi}}}(t)\rangle}_{{\partial}{\mathcal}T_h}\mathrm{d}t\right|\\
&\le
\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}\|\Pi\underline{{\boldsymbol{\Phi}}}(0)-\underline{{\boldsymbol{\Phi}}}(0)\|_{{\boldsymbol{\tau}}}
+\int_0^T\|\dot{{\boldsymbol{\delta}}}(t)\|_{{\boldsymbol{\tau}}^{-1}}\|\Pi\underline{{\boldsymbol{\Phi}}}(t)-\underline{{\boldsymbol{\Phi}}}(t)\|_{{\boldsymbol{\tau}}}\,\mathrm{d}t\\
&\lesssim h\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}
\left(
\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}
+T\vvvert\dot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}
\right),\end{aligned}$$ where we used the convergence properties about $\Pi\underline{{\boldsymbol{\Phi}}}$ (by Theorem \[th:PROJ\]) and then equation to bound ${\boldsymbol{\Theta}}(T)$ (see for the definition).
Using similar ideas, for $T_3$ we have $$\begin{aligned}
\left| T_3\right|&=\left| \int_0^T
{\langle {\boldsymbol{\Delta}}(t),{\boldsymbol{\varepsilon}}_h^u(t)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(t)\rangle}_{{\partial}{\mathcal}T_h}\mathrm{d}t\right|\\
&=\left|{\langle \underline{{\boldsymbol{\Delta}}}(0),{\boldsymbol{\varepsilon}}_h^u(0)
-{\boldsymbol{\varepsilon}}_h^{\widehat u}(0)\rangle}_{{\partial}{\mathcal}T_h}
+\int_0^T{\langle \underline{{\boldsymbol{\Delta}}}(t),\dot{{\boldsymbol{\varepsilon}}}_h^u(t)
-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u}(t)\rangle}_{{\partial}{\mathcal}T_h}\mathrm{d}t\right|\\
&\le
\|\underline{{\boldsymbol{\Delta}}}(0)\|_{{\boldsymbol{\tau}}^{-1}}
\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(0)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(0))\|_{{\boldsymbol{\tau}}}
+\int_0^T
\|\underline{{\boldsymbol{\Delta}}}(t)\|_{{\boldsymbol{\tau}}^{-1}}
\|\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u(t)-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u}(t))\|_{{\boldsymbol{\tau}}}\mathrm{d}t
\\
&\lesssim
h\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}
\left(
\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(0)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(0))\|_{{\boldsymbol{\tau}}}
+T\vvvert\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u})\vvvert_{{\boldsymbol{\tau}},\infty}^{[0,T]}
\right),\end{aligned}$$ and for $T_4$, we have $$\begin{aligned}
\left| T_4\right|&=\left|\int_0^T({\mathcal}A{\boldsymbol{e}}_\Psi(t),{\boldsymbol{\sigma}}(t)-{\boldsymbol{\sigma}}_h(t))_{{\mathcal}T_h}\mathrm{d}t\right|\\
&\le \left| ({\mathcal}A(\Pi\underline{{\boldsymbol{\Psi}}}(0)-\underline{{\boldsymbol{\Psi}}}(0))
,{\boldsymbol{\sigma}}(0)-{\boldsymbol{\sigma}}_h(0))_{{\mathcal}T_h}\right|\\
&\quad\,+\left|\int_0^T
({\mathcal}A(\Pi\underline{{\boldsymbol{\Psi}}}(t)-\underline{{\boldsymbol{\Psi}}}(t))
,\dot{{\boldsymbol{\sigma}}}(t)-\dot{{\boldsymbol{\sigma}}}_h(t))_{{\mathcal}T_h}\mathrm{d}t\right|\\
&\lesssim
h\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}
\left(
\|{\boldsymbol{\sigma}}(0)-{\boldsymbol{\sigma}}_h(0)\|_{{\mathcal}T_h}
+T\vvvert\dot{{\boldsymbol{\sigma}}}-\dot{{\boldsymbol{\sigma}}}_h\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}
\right).\end{aligned}$$
For $T_5$, by and , we have $$\begin{aligned}
\left|T_5\right|&=\left|\int_0^T({\boldsymbol{e}}_\sigma(t),{\boldsymbol{\varepsilon}}({\boldsymbol{\Phi}})(t))_{{\mathcal}T_h}\mathrm{d}t\right|\\
&\le \left|(\Pi{\boldsymbol{\sigma}}(0)-{\boldsymbol{\sigma}}(0),{\boldsymbol{\varepsilon}}(\underline{{\boldsymbol{\Phi}}})(0)
-\mathrm P_0{\boldsymbol{\varepsilon}}(\underline{{\boldsymbol{\Phi}}})(0))_{{\mathcal}T_h}\right|\\
&\quad\, +\left|\int_0^T(\Pi\dot{{\boldsymbol{\sigma}}}(s)-\dot{{\boldsymbol{\sigma}}}(t),
{\boldsymbol{\varepsilon}}(\underline{{\boldsymbol{\Phi}}})(t)
-\mathrm P_0{\boldsymbol{\varepsilon}}(\underline{{\boldsymbol{\Phi}}})(t))_{{\mathcal}T_h}\mathrm{d}t\right|\\
&\lesssim
h\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}
\left(
\|\Pi{\boldsymbol{\sigma}}(0)-{\boldsymbol{\sigma}}(0)\|_{{\mathcal}T_h}
+T\vvvert\Pi\dot{{\boldsymbol{\sigma}}}-\dot{{\boldsymbol{\sigma}}}\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}
\right).\end{aligned}$$
For $T_6$, we simply use and obtain $$\begin{aligned}
\left| T_6\right|=\left|\int_0^T(\rho{\boldsymbol{\Phi}},\ddot{{\boldsymbol{e}}}_u)_{{\mathcal}T_h}\right|
\lesssim \vvvert{\boldsymbol{\Phi}}\vvvert_{\Omega,\infty}^{[0,T]}\vvvert\ddot{{\boldsymbol{e}}}_u\vvvert_{\Omega,1}^{[0,T]}
\lesssim T\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}
\vvvert\ddot{{\boldsymbol{e}}}_u\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}.\end{aligned}$$
Now we use Proposition \[prop:elwv\_tricky\] to estimate $T_7$, and finally obtain $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^u(T)\|_{{\mathcal}T_h}
\lesssim\
&(\|{\boldsymbol{e}}_{u}(0)\|_{{\mathcal}T_h}+h\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}T_h}
+h\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}})\\
&+h\left(
\|{\boldsymbol{\delta}}(0)\|_{{\boldsymbol{\tau}}^{-1}}
+T\vvvert\dot{{\boldsymbol{\delta}}}\vvvert_{{\boldsymbol{\tau}}^{-1},\infty}^{[0,T]}
\right)\\
&+h\left(
\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u(0)-{\boldsymbol{\varepsilon}}_h^{\widehat u}(0))\|_{{\boldsymbol{\tau}}}
+T\vvvert\mathrm P_M(\dot{{\boldsymbol{\varepsilon}}}_h^u-\dot{{\boldsymbol{\varepsilon}}}_h^{\widehat u})\vvvert_{{\boldsymbol{\tau}},\infty}^{[0,T]}
\right)\\
&+h\left(
\|{\boldsymbol{\varepsilon}}_h^\sigma(0)\|_{{\mathcal}T_h}
+T\vvvert\dot{{\boldsymbol{\varepsilon}}}_h^\sigma\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}
\right)\\
&+h\left(
\|{\boldsymbol{e}}_\sigma(0)\|_{{\mathcal}T_h}
+T\vvvert\dot{{\boldsymbol{e}}}_\sigma\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}
\right)+T\vvvert\ddot{{\boldsymbol{e}}}_u\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}\\
&+h^2
\bigg(
\|\ddot{{\boldsymbol{e}}}_u(0)\|_{{\mathcal}T_h}
+T\vvvert\dddot{{\boldsymbol{e}}}_u\vvvert_{{\mathcal}T_h,\infty}^{[0,T]}
\bigg).\end{aligned}$$ Combing the above estimates with Proposition \[prop:ene\_id\_t0\] and Theorem \[th:elwv\_th1\], the proof is completed.
Numerical experiments
=====================
In this section, we present some numerical experiments to support our error estimates in Section \[sec:trans\_elas\]. Note that there are experiments for the steady-state and time-harmonic cases in [@QiShSh:2018] and [@HuPrSa:2017] respectively.
![Visualization of meshes used for error tests.[]{data-label="fig:meshvis"}](cube4ref "fig:") ![Visualization of meshes used for error tests.[]{data-label="fig:meshvis"}](chimney4ref "fig:")
[**Convergence test.**]{} The experiments in this part are carried out on a cubic domain $\Omega = (0,1)^3$ and a non-convex polyhedral domain (we will refer to it as the chimney; see Figure \[fig:meshvis\]), the time interval is $[0,T]$ with $T=\frac{3}{2}$, and we aim to estimate the relative $L^2$ errors $$E_\sigma:=\frac{\|{\boldsymbol{\sigma}}_h(T)-{\boldsymbol{\sigma}}(T)\|_\Omega}{\|{\boldsymbol{\sigma}}(T)\|_\Omega}, \qquad
E_u:=\frac{\|{\boldsymbol{u}}_h(T)-{\boldsymbol{u}}(T)\|_\Omega}{\|{\boldsymbol{u}}(T)\|_\Omega}.$$ We consider a non-homogeneous isotropic material $$\begin{aligned}
{\mathcal}A{\boldsymbol{\sigma}}
= \frac{1}{2\mu}{\boldsymbol{\sigma}} -\frac{\lambda}{2\mu(2\mu+3\lambda)} {\mathrm}{tr}\,{\boldsymbol{\sigma}}\,{\mathbf}I,\end{aligned}$$ with Lamé parameters $$\begin{aligned}
\lambda = \frac{2+x^2+y^2+z^2}{1+x^2+y^2+z^2},\qquad\mu = 3+\cos(xyz),\end{aligned}$$ and the mass density is constant $\rho\equiv1$. As exact solution we take ${\boldsymbol{u}}({\boldsymbol{x}},t):={\boldsymbol{U}}({\boldsymbol{x}})H(t)$, where $$\begin{aligned}
{\boldsymbol{U}}=(\cos(\pi x)\sin(\pi y)\cos(\pi z),\ 5x^2yz+4xy^2z+3xyz^2+17,\ \cos(2x)\cos(3y)\cos(z)),\end{aligned}$$ and the temporal part is $H(t)=t^3(1-t)^2$. The input data ${\boldsymbol{f}}$ and ${\boldsymbol{g}}$ are chosen so that is satisfied. Note that we have chosen the exact solution ${\boldsymbol{u}}$ that has vanishing initial conditions. This simplifies the calculations of ${\boldsymbol{u}}_h(0)$ and $\dot{{\boldsymbol{u}}}_h(0)$ (they are automatically ${\boldsymbol{0}}$ by and ). When ${\boldsymbol{u}}$ does not have vanishing initial conditions, the calculation of $\dot{{\boldsymbol{u}}}_h(0)$ involves projecting $\dot{{\boldsymbol{u}}}(0)$ to a space enriched by the $M$-decomposition spaces (the Cockburn-Fu filling), which we have not found an easy way to implement. Finding easier ways of calculating $\dot{{\boldsymbol{u}}}_h(0)$ for non-vanishing initial conditions will constitute our future works.
For the numerical schemes, we use the HDG+ method for space discretization and Trapezoidal Rule Convolution Quadrature (TRCQ) (see [@Ba:2010; @Lu:1988]) for time integration. This is equivalent to using Trapezoidal Rule time-stepping in the semidiscrete system. The time interval $[0,T]$ is equally divided and each timestep is of length $\kappa$. Since the error from the TRCQ is ${\mathcal}O(\kappa^2)$, we choose $\kappa \approx h^{(k+2)/2}$ so that the error from the time discretization does not pollute the order of convergence of the space discretization.
From Figure \[fig:ord\_conv\] or Table \[tab:ord\_conv\], we observe that the orders of convergence for ${\boldsymbol{\sigma}}_h(T)$ and ${\boldsymbol{u}}_h(T)$ are ${\mathcal}O(h^{k+1})$ and ${\mathcal}O(h^{k+2})$ respectively, agreeing the estimates in Theorem \[th:elwv\_th1\] and Theorem \[th:elwv\_spconv\].
----- -------- --------- ------- --------- ------- --------- ------- --------- ------- -- --
$k$ $h$ Error Order Error Order Error Order Error Order
1.6329 2.82E-1 - 4.32E-2 - 1.27E-1 - 3.93E-2 -
1 0.8164 7.06E-2 2.00 6.16E-3 2.81 3.12E-2 2.02 5.36E-3 2.87
0.4082 1.93E-2 1.87 5.44E-4 3.50 8.16E-3 1.93 4.92E-4 3.44
0.2041 4.79E-3 2.01 5.65E-5 3.27 1.99E-3 2.03 5.09E-5 3.28
1.6329 1.30E-1 - 1.84E-2 - 6.44E-2 - 1.68E-2 -
2 0.8164 1.72E-2 2.92 1.52E-3 3.59 6.83E-3 3.24 1.23E-3 3.78
0.4082 2.31E-3 2.89 5.92E-5 4.68 8.96E-4 2.93 4.74E-5 4.70
0.2041 2.86E-4 3.02 2.52E-6 4.55 - - - -
1.6329 4.36E-2 - 9.50E-3 - 2.18E-2 - 7.74E-3 -
3 0.8164 3.90E-3 3.48 3.35E-4 4.83 1.47E-3 3.89 2.59E-4 4.90
0.4082 2.56E-4 3.93 6.34E-6 5.72 9.54E-5 3.94 4.88E-6 5.73
----- -------- --------- ------- --------- ------- --------- ------- --------- ------- -- --
: History of convergence for ${\boldsymbol{\sigma}}_h(T)$ and ${\boldsymbol{u}}_h(T)$ with sequence of uniform refinements in space and over-refinements in time. []{data-label="tab:ord_conv"}
![History of convergence for ${\boldsymbol{\sigma}}_h(T)$ and ${\boldsymbol{u}}_h(T)$ with sequence of uniform refinements in space and over-refinements in time. Top two figures for cubic meshes and bottom two for chimney.[]{data-label="fig:ord_conv"}](ord_conv "fig:") ![History of convergence for ${\boldsymbol{\sigma}}_h(T)$ and ${\boldsymbol{u}}_h(T)$ with sequence of uniform refinements in space and over-refinements in time. Top two figures for cubic meshes and bottom two for chimney.[]{data-label="fig:ord_conv"}](ord_conv_chim "fig:")
[**Locking test.**]{} Note that the HDG+ method was shown to be free from volumetric locking for steady state system in [@QiShSh:2018]. We here conduct some locking experiments for elastic waves. Most of the experiment settings will be the same as the convergence test. Let us just mention the differences. We shall conduct two experiments (denoted by A and B) on the cubic domain where the Lamé parameters are chosen as $(\lambda,\mu)=(1.5\times10^2,3)$ for test A and $(\lambda,\mu)=(1.5\times10^4,3)$ for test B. Their corresponding Poisson’s ratios can be easily calculated: $\nu\approx0.49$ for test A and $\nu\approx0.4999$ for test B. For the exact solutions, the temporal part $H(t)$ is unchanged while the spacial part is changed to $$\begin{aligned}
{\boldsymbol{U}}=\left(-x^2(x-1)^2y(y-1)(2y-1)z(1-z),\ y^2(y-1)^2x(x-1)(2x-1)z(1-z),\ 0\right).\end{aligned}$$ This choice of the exact solution is a simple 3D adaptation of those used in [@BeLi:1979; @QiShSh:2018] for locking experiments in 2D. We collect the history of convergence for ${\boldsymbol{\sigma}}_h(T)$ and ${\boldsymbol{u}}_h(T)$ in Table \[tab:ord\_conv\_locking\].
----- ---------- ---------- ------- ---------- ------- ---------- ------- ---------- ------- -- --
$k$ $h$ Error Order Error Order Error Order Error Order
1.63e+00 8.21e-01 - 2.29e+00 - 8.25e-01 - 2.29e+00 -
1 8.16e-01 4.66e-01 0.82 4.47e-01 2.36 4.69e-01 0.82 4.46e-01 2.36
4.08e-01 1.78e-01 1.39 5.78e-02 2.95 1.79e-01 1.39 5.75e-02 2.95
2.04e-01 4.52e-02 1.98 7.43e-03 2.96 4.54e-02 1.98 7.41e-03 2.96
1.63e+00 5.19e-01 - 1.36e+00 - 5.20e-01 - 1.36e+00 -
2 8.16e-01 1.98e-01 1.39 1.24e-01 3.45 2.00e-01 1.38 1.24e-01 3.45
4.08e-01 3.70e-02 2.42 7.61e-03 4.03 3.73e-02 2.42 7.58e-03 4.03
2.04e-01 4.79e-03 2.95 4.33e-04 4.14 4.82e-03 2.95 4.31e-04 4.14
----- ---------- ---------- ------- ---------- ------- ---------- ------- ---------- ------- -- --
: History of convergence for ${\boldsymbol{\sigma}}_h(T)$ and ${\boldsymbol{u}}_h(T)$ with sequence of uniform refinements in space and over-refinements in time. All experiments are conducted on the unit cube.[]{data-label="tab:ord_conv_locking"}
From Table \[tab:ord\_conv\_locking\], we observe no degeneration of the convergence rates for ${\boldsymbol{\sigma}}_h(T)$ and ${\boldsymbol{u}}_h(T)$ as the Poisson’s ratio $\nu$ approaches the incompressible limit $0.5$. This supports that the HDG+ method is volumetric locking free for elastic waves.
Extensions and conclusion
=========================
For the sake of conciseness, we have limited the discussion to the setting of elastic problems on simplicial meshes. However, the tools we introduce here can be extended to construct HDG projections in a much wider setting. We next discuss three possible extensions.
1. [**Elasticity on polyhedral meshes.**]{} The HDG+ projection for elasticity can be extended to a projection on polyhedral elements. One way to achieve this is to construct the projection directly on the physical element, instead of first constructing the projection on the reference element and then using a push-forward operator (this is what we did in this paper). This alternative approach is feasible since the $M$-decomposition can be applied on general polyhedral elements (see [@CoFuSa:2017; @CoFu:2018]).
2. [**HDG+ for elliptic diffusion.**]{} The HDG+ projection can be constructed for steady-state diffusion. We have explored this in [@DuSa:2019] for simplicial meshes. For general polyhedral meshes, the projection can be obtained by following a similar procedure as demonstrated in Figure \[fig:flow\_chart\]. It can be summarized in three steps: (1) Enrich the approximation space for the flux so that the $M$-decomposition is achieved; (2) Define an extended projection by enforcing the weak-commutativity property on the homogeneous polynomial space of order $k+1$ (similar to ); (3) Define a composite projection and collect the remainder term on the boundary of the element.
3. [**Standard HDG for elasticity.**]{} We can also construct a projection for the standard HDG method for elasticity (where polynomial spaces of order $k$ are used for both the stress and the displacement). This is achieved by defining the composite projection and the boundary reminder before constructing the extended projection. To be more specific, suppose $$\begin{aligned}
{\boldsymbol{\Pi}}^M: H^1(K;{\mathbb}R_{\mathrm}{sym}^{3\times3})\times H^1(K;{\mathbb}R^3)&\rightarrow {\mathcal}P_k(K;{\mathbb}R_{\mathrm}{sym}^{3\times3})\oplus\Sigma_{\mathrm}{fill}(K)\times {\mathcal}P_{k}(K;{\mathbb}R^3)\\
({\boldsymbol{\sigma}},{\boldsymbol{u}})&\mapsto ({\boldsymbol{\Pi}}_\sigma^M({\boldsymbol{\sigma}},{\boldsymbol{u}}),{\boldsymbol{\Pi}}_u^M({\boldsymbol{\sigma}},{\boldsymbol{u}}))\end{aligned}$$ is the $M$-decomposition associated projection. We then define $$\begin{aligned}
{\boldsymbol{\Pi}}({\boldsymbol{\sigma}},{\boldsymbol{u}})&:=({\mathrm}P_k{\boldsymbol{\Pi}}_\sigma^M({\boldsymbol{\sigma}},{\boldsymbol{u}}),{\boldsymbol{\Pi}}_u^M({\boldsymbol{\sigma}},{\boldsymbol{u}})),\\
{\boldsymbol{\delta}}&:={\boldsymbol{\Pi}}_\sigma^M({\boldsymbol{\sigma}},{\boldsymbol{u}})\cdot{\mathbf}n-{\mathrm}P_k{\boldsymbol{\Pi}}_\sigma^M({\boldsymbol{\sigma}},{\boldsymbol{u}})\cdot{\mathbf}n.\end{aligned}$$ This completes the definition of the projection (and the associate boundary remainder) for the standard HDG method for elasticity. The rest of the error analysis follows the exact same procedure we have discussed in this paper. For instance, for the steady-state problem, we obtain the same energy estimate , namely, $$\begin{aligned}
\|{\boldsymbol{\varepsilon}}_h^\sigma\|_{{\mathcal}A}^2+\|\mathrm P_M({\boldsymbol{\varepsilon}}_h^u-{\boldsymbol{\varepsilon}}_h^{\widehat u})\|_{{\boldsymbol{\tau}}}^2
\le
\|{\boldsymbol{e}}_\sigma\|_{{\mathcal}A}^2+\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}^2.\end{aligned}$$ In this case, the term $\|{\boldsymbol{\delta}}\|_{{\boldsymbol{\tau}}^{-1}}$ has an ${\mathcal}O(h^{k+1/2})$ convergence rate because ${\boldsymbol{\tau}}={\mathcal}O(1)$. We thus recover the existing suboptimal estimates obtained in [@FuCoSt:2015] in a unified way by using the same arguments. The only difference here is a simple change of the projection.
To conclude, we have proposed some new mathematical tools for the error analysis of HDG methods. The two most important ones are: (1) the extended projection constructed by enforcing the weak commutativity on a higher order polynomial space (see ); (2) the boundary remainder reflecting the discrepancy between the normal traces of the $M$-decomposition associate projection and a composite projection (see ). These tools allow us to flexibly devise projections for more variants of HDG methods. We have demonstrated this by constructing the projection for the Lehrenfeld-Schöberl HDG (HDG+) method for elasticity. By using the projection, we are able to recover the existing error estimates in a more concise analysis for the steady-state and the time-harmonic elastic problems. For elastic waves, we have successfully used the projection to devise a semi-discrete HDG+ scheme (the initial velocity of the semi-discrete scheme is defined by using the HDG+ projection) and prove its uniformly-in-time optimal convergence. Improving the generality of the tools will constitute the future works.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We performed low-resolution spectroscopy for the red giant stars in the Galactic globular cluster (GC) NGC 5286, which is known to show intrinsic heavy element abundance variations. We found that the observed stars in this GC are clearly divided into three subpopulations by CN index (CN-weak, CN-intermediate, and CN-strong). The CN-strong stars are also enhanced in the calcium HK$''$ (7.4$\sigma$) and CH (5.1$\sigma$) indices, while the CN-intermediate stars show no significant difference in the strength of HK$''$ index with CN-weak stars. From the comparison with high-resolution spectroscopic data, we found that the CN- and HK$''$-strong stars are also enhanced in the abundances of Fe and $s$-process elements. It appears, therefore, that these stars are later generation stars affected by some supernovae enrichment in addition to the asymptotic giant branch ejecta. In addition, unlike normal GCs, sample stars in NGC 5286 show the CN-CH positive correlation, strengthening our previous suggestion that this positive correlation is only discovered in GCs with heavy element abundance variations such as M22 and NGC 6273.'
author:
- 'Dongwook Lim, Seungsoo Hong, and Young-Wook Lee'
title: 'The CN-CH positive correlation in the globular cluster NGC 5286'
---
Introduction {#intro}
============
During the last two decades, an increasing number of observations have shown that most of the Milky Way globular clusters (GCs) host multiple stellar populations, each of which has different chemical properties [e.g., @Lee99; @Car09; @Gra12; @Pio15 and references therein]. These GCs share similar characteristics, such as light element abundance variations and a central concentration of later generation stars [e.g., @Car09; @Lar11], although some exceptional cases are also reported. The abundance variations in the light elements, discovered in most GCs, are explained as an enrichment and/or pollution by intermediate-mass asymptotic giant branch (AGB) stars [@DC04; @Dan16], massive interacting binary stars [@de09; @Bas13], and fast-rotating massive stars (FRMSs; @PC06 [@Dec07]). Several GCs with heavy element abundance variations, including $\omega$-Centauri and M22 [@Lee99; @jwlee09; @Mar09; @JP10], however, show evidence of supernovae (SNe) enrichment, which suggests that they were massive enough in the past to retain SNe ejecta [@Bau08; @ST17]. In the hierarchical merging paradigm, they would have contributed to the formation of the Milky Way, since these GCs could be former nuclei of dwarf galaxies [see @Lee07; @Han15; @Da16], and therefore, would help to solve the “missing satellites problem” [@Kly99; @Moo99; @Moo06].
A direct way to find these unique GCs is a measurement of heavy element abundance of stars in a GC using high-resolution spectroscopy [e.g., @Da09; @Yong14; @Joh15]. However, our previous studies have demonstrated that the low-resolution spectroscopy for the calcium HK$'$ index can be used to more effectively detect the heavy element variations in a GC [@Lim15; @Han15]. Interestingly, we found that these GCs also show the CN-CH positive correlation among red giant branch (RGB) stars, unlike the CN-CH anticorrelation generally observed in “normal” GCs [@SS91; @sglee05; @Kay08; @Pan10; @Smo11]. If confirmed, this would imply that the CN-CH positive correlation can be used as a probe for the GCs with heavy element variations.
In order to further confirm our conjecture, we have performed low-resolution spectroscopy for the RGB stars in NGC 5286. This GC is relatively poorly studied, although @Mar15 [hereafter ] recently showed some abundance variations in Fe and $s$-process elements among RGB stars from high-resolution spectroscopy. The purpose of this paper is to report that RGB stars in this GC are clearly divided into three subpopulations by CN index, and CN-strong stars are also enhanced in the calcium HK$'$ and CH indices, indicating the CN-CH positive correlation.
Observations and Data reduction {#obs}
===============================
Our observations were performed with the du Pont 2.5m telescope at Las Campanas Observatory (LCO) during four nights from April to June 2016. We have used multi-object spectroscopy mode of Wide Field Reimaging CCD Camera (WFCCD) with HK grism, which provides a dispersion of 0.8 [Å]{}/pixel and a central wavelength of 3700 [Å]{}. Spectroscopic target stars are selected from the 2MASS All-Sky Point Source Catalog. In particular, we have included a number of stars observed by for the comparison with high-resolution spectroscopy. For these observations, three multi-slit masks, each of which contains about 25 slits of 1.2 width, were designed. We had obtained four 1500-second science exposures, three flats, and an arc lamp frame for each mask. The data reduction was performed with IRAF[^1] and the modified version of the WFCCD reduction package, following @Lim15 and @Pro06. The radial velocity (RV) of each star was measured using $rvidlines$ task in the IRAF RV package, and the signal-to-noise ratio (S/N) was estimated at $\sim$3900 [Å]{}. After the rejection of non-member stars (RV $>$ 2.0$\sigma$ of the mean velocity of the GC) and low S/N ($<$ 8.0) spectra, 44 stars are finally used for our analysis. The median RV for these stars is 52.5 km/s, which is comparable to but somewhat smaller than the value of 57.4 km/s reported by @Har10 and 61.5 km/s estimated by . Compared to the RVs measured from high-resolution spectroscopy, the typical uncertainty of those from our low-resolution measurement appears to be quite large ($\sim$ 20km/s). Figure \[fig\_cmd\] shows the selected sample stars on the color-magnitude diagram (CMD), for which the photometry was obtained at the LCO 2.5m du Pont telescope. Fifteen target stars, however, are outside the field-of-view (FOV) of the photometry (8.85 $\times$ 8.85) and therefore not plotted in this CMD. A detailed description of this photometry can be found in @Lim15.
{width="48.00000%"}
Finally, we measured the S(3839) index for CN band, the HK$'$ index for Ca II H and K lines, and the CH4300 index for CH band of each target star in NGC 5286, following @Lim15. The definitions for these indices are $$\begin{aligned}
{\rm HK'} & = & -2.5 \log{\frac{F_{3916-3985}}{2F_{3894-3911}+F_{3990-4025}}} , \\
{\rm CN}(3839) & = & -2.5 \log{\frac{F_{3861-3884}}{F_{3894-3910}}} , \\
{\rm CH4300} & = & -2.5 \log{\frac{F_{4285-4315}}{0.5F_{4240-4280}+0.5F_{4390-4460}}} ,\end{aligned}$$ where $F_{3916-3985}$, for example, is the integrated flux from 3916 to 3985 [Å]{}. All of these indices are defined as the ratio of the absorption strength to nearby continuum strength. The measurement error for each sample was estimated from Poisson statistics in the flux measurements [@VE06]. In addition, we measured the delta indices ($\delta$CN, $\delta$HK$'$, and $\delta$CH) to compare the chemical abundances of stars without the effect of magnitude in the same manner as in previous studies [e.g., @Nor81; @Har03]. These $\delta$-indices are calculated as the difference between the original index for each star and the least-squares fitting of the full sample (black solid lines in the left panels of Figure \[fig\_index\]) in a GC. The measured indices and errors are listed in Table \[tab\_index\].
Multiple stellar populations in NGC 5286 {#spec}
========================================
{width="70.00000%"}
{width="45.00000%"}
Figure \[fig\_index\] shows the measured spectral indices of stars as functions of $K$ magnitude, obtained from 2MASS catalog. The CN, HK$'$, and CH indices increase with decreasing magnitude because the brighter RGB stars have lower temperatures and the strengths of these molecular bands generally increase with decreasing temperature. Therefore the chemical abundances of stars are compared on the $\delta$-index versus magnitude diagrams. It is important to note that the observed stars show a large spread in $\delta$-index that is at least several times larger than the measurement error. The standard deviations for all sample stars are 0.23 for CN, 0.07 for HK$'$, and 0.07 for CH. In particular, the CN index distribution shows the largest spread. Note that a bimodality or a large spread in CN distribution is generally observed in most GCs [@Nor81; @Nor87; @Bri92; @Har03; @Kay08; @SMar08][^2]. Therefore, we have divided subpopulations of RGB stars in NGC 5286 on the histogram of the $\delta$CN index (see Figure \[fig\_hist\]). It is clear from this histogram that RGB stars are divided into three subpopulations: CN-weak ($\delta$CN $<$ -0.2; blue circles), CN-intermediate (-0.2 $\leq$ $\delta$CN $<$ 0.1; green circles), and CN-strong (0.1 $\leq$ $\delta$CN; red circles). The distribution of CN index into three or more subpopulations is similar to that reported in NGC 1851 [@Camp12; @Lim15; @Sim17]. This is also consistent with the recent results from population models and spectroscopic observations which show that most GCs host three or more subpopulations [see, e.g., @Jang14; @Car15]. The presence of multiple populations is also observed from recent photometry using UV filters, which are mainly sensitive to N abundance [@Mil15; @Pio15]. In this regard, further observations are required to see that the trimodal CN distribution, observed in NGC 1851 and NGC 5286, is a ubiquitous feature in other GCs as well.
As shown in the right panels of Figure \[fig\_index\], the CN-strong subpopulation is significantly enhanced also in the $\delta$HK$'$ and $\delta$CH indices. The differences between CN-strong and CN-weak subpopulations are 0.537 for $\delta$CN, 0.123 for $\delta$HK$'$, and 0.094 for $\delta$CH, which are significant at the levels of 20.7$\sigma$, 7.4$\sigma$, and 5.1$\sigma$, respectively, compared to the standard deviation of the mean. The CN-intermediate stars, however, show no clear difference in the strength of $\delta$HK$'$ index with CN-weak stars ($\Delta\delta$HK$'$ = 0.02). Therefore, three subpopulations in the NGC 5286 can be characterized as CN-weak/HK$'$-weak, CN-intermediate/HK$'$-weak, and CN-strong/HK$'$-strong. In particular, the difference in calcium abundance (HK$'$) suggests that this GC also belongs to the group of unique GCs showing intrinsic dispersion in heavy element abundance, in agreement with a result by based on Fe and $s$-process elements[^3]. The fact that CN-strong stars are also enhanced in the CH band implies a presence of CN-CH positive correlation in this GC (see Section \[CN\_CH\] below).
{width="42.00000%"}
In order to see whether the CN- and HK$'$-strong stars in our study are also enhanced in Fe and $s$-process elements, we have compared our results with high-resolution spectroscopy by . In Figure \[fig\_comp\], our $\delta$CN and $\delta$HK$'$ indices are plotted with \[Na/Fe\] and \[Ca/Fe\] abundances, respectively, for 33 common stars. In general, the strength of CN band is correlated with the N and Na abundances, while the CH band is affected by C abundance [@Sne92; @Smi96; @Mar08]. The upper panel of Figure \[fig\_comp\] also shows a strong correlation between \[Na/Fe\] and $\delta$CN index, which is in good agreement with previous studies [@Sne92; @Lim16][^4]. The CN-weak (blue) and CN-strong (red) subpopulations are almost identical to the $s$-poor (triangles) and $s$-rich (squares) groups, respectively. In addition, the $\delta$HK$'$ index is understandably correlated with the \[Ca/Fe\] abundance with a few exceptions (see the lower panel of Figure \[fig\_comp\]). According to this comparison, the difference in $\delta$HK$'$ index between CN-weak and CN-strong stars ($\sim$ 0.094) is equivalent to 0.09 dex in $\Delta$\[Ca/Fe\] and 0.15 dex in $\Delta$\[Fe/H\]. These comparisons confirm that our results from low-resolution spectroscopy are consistent with those from high-resolution spectroscopy by . Consequently, the later generation stars in NGC 5286 show the enhancements not only in light elements (CN) but also in heavy elements (Fe and Ca) and $s$-process elements, although the presence of Fe spread requires further investigations [see @Muc15; @jwlee16].
The CN-CH positive correlation in globular clusters with heavy element variations {#CN_CH}
=================================================================================
As described above, the CN-CH anticorrelation is one of the typical features in the low-resolution spectroscopic studies of GCs [@SS91; @Kra94; @Har03; @sglee05; @Pan10; @Smo11 and references therein]. This feature is most likely due to the anticorrelation between C and N abundances [see, e.g., @Smi96; @Coh05]. In the multiple population paradigm, the mechanism responsible for the Na-O anticorrelation would also produce C-N anticorrelation [@Ven13; @Di16]. Our previous studies, however, found a significant CN-CH positive correlation, instead of an anticorrelation, among RGB stars in M22 and NGC 6273 [@Han15; @Lim15]. Interestingly, both GCs are known to host multiple stellar populations with different heavy element abundances [see also @Mar11; @Joh17]. Since NGC 5286 is also one of the GCs showing spreads in the abundances of heavy elements, we would expect a similar positive correlation between CN and CH indices. As expected, Figure \[fig\_cnch\] shows that $\delta$CN and $\delta$CH indices are positively correlated similarly to the cases of M22 and NGC 6273.
{width="48.00000%"}
{width="100.00000%"}
In order to establish how the heavy element abundance variations would affect the CN-CH relation of GCs, we have plotted in Figure \[fig\_cnch\_all\] the $\delta$CN versus $\delta$CH diagrams for six GCs (NGC 288, NGC 6723, NGC 1851, NGC 6273, M22, and NGC 5286), together with the $\delta$HK$'$ distributions. The spectroscopic and photometric data are taken from @Han15 and @Lim15 [@Lim16]. The slope of the CN-CH relation for each GC is estimated by maximum likelihood, and the values are listed in the upper panels of Figure \[fig\_cnch\_all\]. We can see from this figure that normal GCs without a difference in Ca abundance, NGC 288 and NGC 6723, show the conventional CN-CH anticorrelation. On the contrary, GCs with the difference in HK$'$ index between the two subpopulations, NGC 6273, M22, and NGC 5286, show the CN-CH positive correlation. In the case of NGC 1851, the difference in HK$'$ index is relatively small, and the CN-CH relation seems to be flat[^5]. In order to test the significance of the correlation, we have calculated the Spearman’s rank correlation coefficient for each GC. The obtained correlation coefficient are -0.52 and -0.62 for NGC 288 and NGC 6723, -0.05 for NGC 1851, and +0.37, +0.61, and +0.52 for NGC 6273, M22, and NGC 5286, respectively. The $p$-values are very small (7.3 $\times$ $10^{-7}$, 1.4 $\times$ $10^{-4}$, 7.3 $\times$ $10^{-4}$, 2.9 $\times$ $10^{-11}$, and 2.9 $\times$ $10^{-4}$ for NGC 288, NGC 6723, NGC 6273, M22, and NGC 5286), confirming that the correlations are statistically significant, except for NGC 1851 ($p$-value = 0.72). Because the negative Spearman coefficient (-1.0 $\sim$ 0.0) indicates anticorrelation while the positive coefficient (0.0 $\sim$ +1.0) is for positive correlation, this result confirms the systematic variation in the CN-CH correlation among sample GCs. Therefore, the origin of the CN-CH positive correlation appears to be explicitly relevant to the heavy element abundance variations. In this respect, it would be interesting to measure \[C/Fe\], \[N/Fe\], and the C+N+O sum from high-resolution spectroscopy.
As discussed in @Lim15, the origin of the CN-CH positive correlation, as well as of the heavy element abundance variations, is most likely because the later generation stars are enriched by some SNe in addition to the intermediate-mass AGB stars and/or FRMSs. Unlike the intermediate-mass AGB stars, which are suggested to be mainly responsible for the N-enhancement and C-depletion of later generation stars, the SNe ejecta would supply both N and C elements together with other heavy elements. Our result for the CN-CH positive correlation in NGC 5286 (Figure \[fig\_cnch\]) appears to be similar to those of M22 and NGC 6273. Interestingly, inspection of Figure \[fig\_index\] shows that the observed stars in NGC 5286 can be divided into three subpopulations: CN-weak/HK$'$-weak stars (first generation; G1), CN-intermediate/HK$'$-weak stars (second generation; G2), and CN-strong/HK$'$-strong stars (third generation; G3). These differences in chemical properties imply that the SNe enrichment played a role only in the formation of G3 stars, whereas it had almost no impact on the formation of G2 stars. Although the origin of this complex chemical enrichment requires further investigation, one possibility is a time-dependent gas removal of SNe ejecta in a proto-GC. For example, a recent hydrodynamical simulation by @Cap17 shows that the gas removal in a dwarf galaxy was more efficient in the first 600 Myr, while most of the gas could be retained later when the type II SNe rate was significantly decreased. Similar to this, the SNe ejecta could have fully escaped from the proto-NGC 5286 in the early phase, while some of them could have been retained later with decreasing SNe rate. This would explain the absence and presence of some SNe enrichment in G2 and G3, respectively.
We are grateful to the anonymous referee for a number of helpful suggestions. We also thank Sang-Il Han for providing photometric data for Figure \[fig\_cmd\]. Support for this work was provided by the National Research Foundation of Korea to the Center for Galaxy Evolution Research, and through the grant programs No. 2017R1A6A3A11031025 and No. 2017R1A2B3002919.
Bastian, N., Lamers, H. J. G. L. M., de Mink, S. E., et al. 2013, , 436, 2398 Baumgardt, H., Kroupa, P., & Parmentier, G. 2008, , 384, 1231 Briley, M. M., Smith, G. H., Bell, R. A., Oke, J. B., & Hesser, J. E. 1992, , 387, 612 Campbell, S. W., Yong, D., Wylie-de Boer, E. C., et al. 2012, , 761, L2 Caproni, A., Amaral Lanfranchi, G., Campos Baio, G. H., Kowal, G., & Falceta-Gon[ç]{}alves, D. 2017, , 838, 99 Carretta, E. 2015, , 810, 148 Carretta, E., Bragaglia, A., Gratton, R. G., et al. 2009, , 505, 117 Cohen, J. G., Briley, M. M., & Stetson, P. B. 2005, , 130, 1177 Da Costa, G. S. 2016, The General Assembly of Galaxy Halos: Structure, Origin and Evolution, 317, 110 Da Costa, G. S., Held, E. V., Saviane, I., & Gullieuszik, M. 2009, , 705, 1481 D’Antona, F., & Caloi, V. 2004, , 611, 871 D’Antona, F., Vesperini, E., D’Ercole, A., et al. 2016, , 458, 2122 Decressin, T., Meynet, G., Charbonnel, C., Prantzos, N., & Ekstr[ö]{}m, S. 2007, , 464, 1029 de Mink, S. E., Pols, O. R., Langer, N., & Izzard, R. G. 2009, , 507, L1 Di Criscienzo, M., Ventura, P., Garc[í]{}a-Hern[á]{}ndez, D. A., et al. 2016, , 462, 395 Gratton, R. G., Carretta, E., & Bragaglia, A. 2012, , 20, 50 Han, S.-I., Lim, D., Seo, H., & Lee, Y.-W. 2015, , 813, L43 Harbeck, D., Smith, G. H., & Grebel, E. K. 2003, , 125, 197 Harris, W. E. 2010, arXiv:1012.3224 Jang, S., Lee, Y.-W., Joo, S.-J., & Na, C. 2014, , 443, L15 Johnson, C. I., Caldwell, N., Rich, R. M., et al. 2017, , 836, 168 Johnson, C. I., & Pilachowski, C. A. 2010, , 722, 1373 Johnson, C. I., Rich, R. M., Pilachowski, C. A., et al. 2015, , 150, 63 Kayser, A., Hilker, M., Grebel, E. K., & Willemsen, P. G. 2008, , 486, 437 Klypin, A., Kravtsov, A. V., Valenzuela, O., & Prada, F. 1999, , 522, 82 Kraft, R. P. 1994, , 106, 553 Lardo, C., Bellazzini, M., Pancino, E., et al. 2011, , 525, A114 Lee, J.-W. 2016, , 226, 16 Lee, J.-W., Kang, Y.-W., Lee, J., & Lee, Y.-W. 2009, , 462, 480 Lee, S.-G. 2005, Journal of Korean Astronomical Society, 38, 23 Lee, Y.-W., Gim, H. B., & Casetti-Dinescu, D. I. 2007, , 661, L49 Lee, Y.-W., Joo, J.-M., Sohn, Y.-J., et al. 1999, , 402, 55 Lim, D., Han, S.-I., Lee, Y.-W., et al. 2015, , 216, 19 Lim, D., Lee, Y.-W., Pasquato, M., Han, S.-I., & Roh, D.-G. 2016, , 832, 99 Marino, A. F., Milone, A. P., Karakas, A. I., et al. 2015, , 450, 815 (M15) Marino, A. F., Milone, A. P., Piotto, G., et al. 2009, , 505, 1099 Marino, A. F., Sneden, C., Kraft, R. P., et al. 2011, , 532, A8 Marino, A. F., Villanova, S., Piotto, G., et al. 2008, , 490, 625 Martell, S. L., Smith, G. H., & Briley, M. M. 2008, , 120, 7 Milone, A. P., Marino, A. F., Piotto, G., et al. 2015, , 808, 51 Moore, B., Diemand, J., Madau, P., Zemp, M., & Stadel, J. 2006, , 368, 563 Moore, B., Ghigna, S., Governato, F., et al. 1999, , 524, L19 Mucciarelli, A., Lapenna, E., Massari, D., et al. 2015, , 809, 128 Norris, J. 1987, , 313, L65 Norris, J., Cottrell, P. L., Freeman, K. C., & Da Costa, G. S. 1981, , 244, 205 Pancino, E., Rejkuba, M., Zoccali, M., & Carrera, R. 2010, , 524, A44 Piotto, G., Milone, A. P., Bedin, L. R., et al. 2015, , 149, 91 Prantzos, N., & Charbonnel, C. 2006, , 458, 135 Prochaska, J. X., Weiner, B. J., Chen, H.-W., & Mulchaey, J. S. 2006, , 643, 680 Silich, S., & Tenorio-Tagle, G. 2017, , 465, 1375 Simpson, J. D., Martell, S. L., & Navin, C. A. 2017, , 465, 1123 Smith, G. H., Shetrone, M. D., Bell, R. A., Churchill, C. W., & Briley, M. M. 1996, , 112, 1511 Smolinski, J. P., Martell, S. L., Beers, T. C., & Lee, Y. S. 2011, , 142, 126 Sneden, C., Kraft, R. P., Prosser, C. F., & Langer, G. E. 1992, , 104, 2121 Suntzeff, N. B., & Smith, V. V. 1991, , 381, 160 Sweigart, A. V., & Mengel, J. G. 1979, , 229, 624 Ventura, P., Di Criscienzo, M., Carini, R., & D’Antona, F. 2013, , 431, 3642 Vollmann, K., & Eversberg, T. 2006, Astronomische Nachrichten, 327, 862 Yong, D., Roederer, I. U., Grundahl, F., et al. 2014, , 441, 3396
ID R.A. Decl. $K$ CN errCN $\delta$CN HK$'$ errHK$'$ $\delta$HK$'$ CH errCH $\delta$CH ID$_{M15}$
------------ ----------- ----------- --------- --------- -------- ------------ -------- ---------- --------------- -------- -------- ------------ ------------
N5286-1001 206.39160 -51.45328 11.6980 0.2202 0.0225 0.1524 0.5628 0.0180 0.0622 1.1472 0.0126 0.0190 –
N5286-1004 206.42943 -51.41643 11.0020 0.3421 0.0153 0.2331 0.6209 0.0128 0.0770 1.2116 0.0090 0.0663 –
N5286-1005 206.36223 -51.34178 12.5070 0.2922 0.0586 0.2722 0.5396 0.0499 0.0892 1.1720 0.0340 0.0637 –
N5286-1006 206.36823 -51.34811 12.2170 0.2265 0.0503 0.1894 0.6185 0.0390 0.1502 1.1377 0.0285 0.0222 –
N5286-1007 206.36032 -51.39484 12.0630 0.2115 0.0215 0.1653 0.5421 0.0173 0.0642 1.1304 0.0121 0.0111 –
N5286-1008 206.48909 -51.42993 10.6760 -0.1508 0.0214 -0.2790 0.4899 0.0139 -0.0743 1.0571 0.0098 -0.0962 –
N5286-1045 206.74561 -51.36780 12.8490 0.2049 0.0527 0.2050 0.5290 0.0426 0.1000 1.1699 0.0289 0.0699 –
N5286-1046 206.74057 -51.37590 12.8780 -0.1537 0.0395 -0.1518 0.3734 0.0276 -0.0539 1.0392 0.0183 -0.0601 –
N5286-1067 206.79630 -51.30748 12.9280 0.1277 0.0775 0.1326 0.5415 0.0590 0.1173 1.1379 0.0411 0.0399 –
N5286-1068 206.78522 -51.40542 12.7240 0.1445 0.0438 0.1373 0.5134 0.0344 0.0765 1.1465 0.0234 0.0435 –
N5286-1077 206.39413 -51.31052 10.9470 0.4391 0.0276 0.3269 0.6283 0.0244 0.0810 1.1521 0.0178 0.0054 –
N5286-3002 206.58405 -51.35508 10.7360 0.2799 0.0316 0.1552 0.6655 0.0246 0.1051 1.2687 0.0172 0.1169 527G
N5286-3004 206.60188 -51.36125 10.3860 -0.0429 0.0236 -0.1882 0.3426 0.0182 -0.2396 1.1449 0.0111 -0.0155 757G
N5286-3005 206.59503 -51.32511 13.5190 -0.3595 0.0544 -0.3197 0.2316 0.0360 -0.1559 1.0126 0.0221 -0.0710 697G
N5286-3007 206.57454 -51.34800 13.0210 0.3054 0.0401 0.3157 0.4591 0.0358 0.0407 1.1705 0.0232 0.0748 399G
N5286-3009 206.63637 -51.34433 10.4220 -0.2349 0.0278 -0.3781 0.5173 0.0167 -0.0627 1.0921 0.0118 -0.0674 1297G
N5286-3010 206.71512 -51.35806 13.8950 -0.0398 0.0529 0.0222 0.3885 0.0396 0.0244 1.0101 0.0269 -0.0642 1767G
N5286-3011 206.63579 -51.32022 13.6950 0.1239 0.0440 0.1741 0.3852 0.0370 0.0087 1.0791 0.0241 -0.0001 1269G
N5286-3012 206.61157 -51.33022 13.7600 0.1798 0.0452 0.2338 0.4184 0.0386 0.0459 1.0567 0.0261 -0.0209 939G
N5286-3014 206.61841 -51.31781 13.2930 -0.2311 0.0479 -0.2047 0.3945 0.0312 -0.0070 0.9885 0.0216 -0.1005 1057G
N5286-3015 206.65620 -51.33216 13.2050 -0.3211 0.0605 -0.2999 0.3699 0.0375 -0.0371 1.0883 0.0241 -0.0029 1547G
N5286-3017 206.71629 -51.38522 13.0530 -0.3447 0.0471 -0.3324 0.3577 0.0290 -0.0587 0.9224 0.0204 -0.1726 5441G
N5286-3020 206.68457 -51.34958 13.4260 0.2628 0.0394 0.2971 0.4285 0.0353 0.0353 1.2015 0.0221 0.1157 1729G
N5286-3021 206.67270 -51.34514 10.9670 0.0882 0.0180 -0.0228 0.5405 0.0134 -0.0056 1.0781 0.0097 -0.0680 1687G
N5286-3022 206.75072 -51.35933 11.6600 -0.2314 0.0456 -0.3015 0.5102 0.0276 0.0072 1.1262 0.0190 -0.0030 5541G
N5286-3023 206.68832 -51.36242 12.1110 0.0271 0.0305 -0.0163 0.4513 0.0230 -0.0237 1.0054 0.0164 -0.1127 1737G
N5286-3024 206.65457 -51.34297 13.4610 -0.0498 0.0513 -0.0134 0.3564 0.0390 -0.0346 1.0702 0.0251 -0.0148 1537G
N5286-3033 206.64291 -51.42208 12.5850 0.4697 0.0388 0.4543 0.4815 0.0384 0.0360 1.1566 0.0255 0.0501 1369G
N5286-3034 206.65366 -51.39170 14.1630 0.2136 0.0601 0.2914 0.3737 0.0531 0.0263 1.0459 0.0351 -0.0219 1529G
N5286-3035 206.61162 -51.43067 13.4270 -0.3975 0.0786 -0.3631 0.2833 0.0489 -0.1098 1.0653 0.0301 -0.0205 947G
N5286-3037 206.60625 -51.40703 14.0950 -0.3722 0.0762 -0.2984 0.3193 0.0470 -0.0323 1.0287 0.0302 -0.0406 827G
N5286-3039 206.69041 -51.41572 12.7890 -0.0149 0.0483 -0.0182 0.4006 0.0365 -0.0322 1.0765 0.0241 -0.0249 1747G
N5286-3040 206.61467 -51.41781 13.1280 -0.1347 0.0411 -0.1180 0.3533 0.0294 -0.0584 1.0432 0.0192 -0.0499 996G
N5286-3042 206.58870 -51.42358 12.0010 -0.0442 0.0330 -0.0941 0.4462 0.0237 -0.0356 1.2179 0.0148 0.0971 587G
N5286-3043 206.52225 -51.38350 13.6400 -0.1130 0.0706 -0.0661 0.3493 0.0516 -0.0306 1.1905 0.0308 0.1099 29G
N5286-3044 206.57805 -51.39339 13.6610 -0.2677 0.0531 -0.2195 0.3291 0.0352 -0.0495 1.0672 0.0223 -0.0128 437G
N5286-3045 206.51570 -51.37811 13.4250 -0.0339 0.0397 0.0003 0.3704 0.0302 -0.0228 1.0367 0.0200 -0.0491 17G
N5286-3047 206.55225 -51.38778 13.7410 -0.0787 0.0528 -0.0258 0.3818 0.0387 0.0082 1.1475 0.0242 0.0694 169G
N5286-3048 206.50739 -51.35966 13.4060 -0.3709 0.0473 -0.3378 0.3359 0.0290 -0.0586 1.0584 0.0185 -0.0280 7G
N5286-3049 206.56929 -51.40445 13.7600 0.2490 0.0424 0.3030 0.4271 0.0377 0.0547 1.1281 0.0245 0.0505 289G
N5286-3050 206.57391 -51.36686 12.6600 0.1835 0.0351 0.1725 0.4306 0.0297 -0.0102 1.0876 0.0199 -0.0170 379G
N5286-3051 206.58330 -51.38422 13.0020 -0.0815 0.0586 -0.0723 0.3927 0.0426 -0.0269 1.1958 0.0261 0.0995 509G
N5286-3053 206.59665 -51.40594 14.1720 -0.0480 0.0580 0.0304 0.3662 0.0437 0.0193 1.0408 0.0288 -0.0267 707G
N5286-3055 206.55904 -51.34925 14.0610 -0.2148 0.0740 -0.1430 0.3444 0.0504 -0.0093 1.1065 0.0315 0.0363 207G
[^1]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA) under a cooperative agreement with the National Science Foundation.
[^2]: Although the evolutionary mixing effect can also contribute to the large spread in CN index distribution among bright RGB stars [@SM79], this effect alone cannot explain a discrete distribution and a wide spread in the unevolved stars [see, e.g., @Kay08].
[^3]: @Muc15 questioned the presence of an intrinsic Fe spread in some GCs, including M22. They found no obvious Fe spread when Fe abundance is measured from Fe II line with photometric gravity. On the other hand, @jwlee16 has refuted this claim from the independent spectroscopic analysis. In any case, the presence of apparent Ca spread in these GCs would suggest that they were affected by SNe enrichment.
[^4]: Careful inspection of the upper panel of Figure \[fig\_comp\] also shows the possibility that the $s$-poor and $s$-rich groups are probably separated on this diagram, with the $s$-rich stars more enhanced in both \[Na/Fe\] and $\delta$CN. This would imply that the variations in light elements would be present in each group with different $s$-process elements abundances, which has already been found in other GCs with $s$-process element and Fe variations, such as M2 and M22 [@Mar11; @Yong14]. More sample of stars, however, are needed to confirm this trend in NGC 5286.
[^5]: Recently, @Sim17 discovered seven stars strongly enhanced in both CN and CH indices in NGC 1851.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.